Paper deep dive
PANDA - Prototype-Anchored Alignment for Partially Unpaired Multimodal Learning, with Applications to Alzheimers MRI and TCGA Pathology
Sheethal Bhat, Mahfuzur Rahman Chowdhury, Paula Andrea Perez-Toro, Stephan Wunderlich, Rose Dawn Bharat, Siming Bayer, Andreas Maier
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 8/29/2026, 3:43:21 AM
Summary
The paper introduces PANDA (Prototype-Anchored Data Alignment), a two-stage framework for partially unpaired multimodal learning in medical imaging. PANDA transfers auxiliary information to a primary-modality model without requiring auxiliary inputs at inference. It learns shared embeddings from paired data, estimates class prototypes from auxiliary modalities, and then aligns the primary encoder to these frozen prototypes. The method is evaluated on Alzheimer's Disease (AD) classification using MRI with auxiliary modalities (tabular scores, FDG-PET, handwriting kinematics) and on TCGA-Lung survival prediction using Whole-Slide Images (WSI) with RNA-seq. PANDA improves performance over unimodal baselines and handles arbitrary pairing rates, including zero overlap.
Entities (12)
Relation Signals (7)
PANDA ā appliedto ā Alzheimerās Disease Neuroimaging Initiative
confidence 95% Ā· We evaluate PANDA on two applications. On a 1,021-subject multi-scanner ADNI cohort...
PANDA ā appliedto ā The Cancer Genome Atlas
confidence 95% Ā· On TCGA-Lung survival prediction... PANDA improves over WSI-only...
PANDA ā handles ā Partially Unpaired Data
confidence 95% Ā· Multimodal medical prediction often faces incomplete pairing... PANDA accommodates arbitrary pairing rates
PANDA ā uses ā Prototype Anchoring
confidence 95% Ā· PANDA (Prototype-Anchored Data Alignment), a two-stage framework... estimates class prototypes from auxiliary modalities
PANDA ā improves ā AUC
confidence 90% Ā· Relative to the same-backbone MRI-only baseline, PANDA attains AUC 0.868 +-0.020 (+7.9pp)
RNA-seq ā isauxiliaryto ā Whole-slide Images
confidence 90% Ā· ...using RNA-seq as auxiliary data... on TCGA-Lung survival prediction from whole-slide images
MRI ā isprimarymodalityfor ā Alzheimer's Disease Classification
confidence 90% Ā· Structural MRI is a natural primary modality for the AD study...
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Multimodal medical prediction often faces incomplete pairing: auxiliary modalities with complementary signal are available for only a subset of subjects (or none) and cannot be assumed at deployment. We introduce PANDA (Prototype Anchored Data Alignment), a two-stage framework that transfers auxiliary information to a primary-modality model without auxiliary inputs at inference. Stage 1 learns a shared embedding from the paired subset and estimates class prototypes from auxiliary modalities; Stage 2 trains the primary encoder on all subjects using cross-entropy plus alignment to the frozen prototypes. Because supervision is defined at the class-prototype level, PANDA accommodates arbitrary pairing rates, including zero subject overlap. We evaluate PANDA on two applications. On a 1,021-subject multi-scanner ADNI cohort, we perform AD/CN classification with three auxiliary modalities at distinct pairing rates: tabular scores (44.8%), FDG-PET (18.7%), and external handwriting kinematics (0% overlap). Relative to the same-backbone MRI-only baseline, PANDA attains AUC 0.868 +-0.020 (+7.9pp) and reduces 1.5T CN false positives by 24.3pp; on a fully trainable Conv5-FC3 backbone it reaches AUC 0.893 (best overall). A pairing-rate ablation shows that the joint anchor remains within seed noise from 75% to 5% pairing. On TCGA-Lung survival prediction from whole-slide images with RNA-seq as auxiliary data, PANDA improves over WSI-only on 2-year OS (AUC +3.5pp) and Cox PH (C-index +9.0pts) and outperforms full-fusion training, which underperforms WSI-only, while requiring no RNA at inference; wide confidence intervals on this smaller cohort keep the gains below conventional significance. Overall, PANDA provides a deployment-oriented mechanism for leveraging incomplete auxiliary modalities to improve primary-modality prediction.
Tags
Links
- Source: https://arxiv.org/abs/2608.25970v1
- Canonical: https://arxiv.org/abs/2608.25970v1
Trouble viewing inline? Open PDF directly ā
Full Text
108,146 characters extracted from source content.
Expand or collapse full text
PANDA: Prototype-Anchored Alignment for Partially Unpaired Multimodal Learning, with Applications to Alzheimerās MRI and TCGA Pathology Sheethal Bhat Email: sheethal.bhat@fau.de Affiliation: Pattern Recognition Lab, Friedrich-Alexander-UniversitƤt, Erlangen-NĆ¼rnberg, Germany Mahfuzur Rahman Chowdhury Affiliation: Pattern Recognition Lab, Friedrich-Alexander-UniversitƤt, Erlangen-NĆ¼rnberg, Germany Paula Andrea PĆ©rez-Toro Affiliation: Pattern Recognition Lab, Friedrich-Alexander-UniversitƤt, Erlangen-NĆ¼rnberg, Germany Stephan Wunderlich Affiliation: Department of Neurology, Klinikum NĆ¼rnberg, Paracelsus Medical University, NĆ¼rnberg, Germany Affiliation: Department of Radiology, LMU University Hospital, LMU Medizin, Ludwig-Maximilians-UniversitƤt MĆ¼nchen, Munich, Germany Rose Dawn Bharat Affiliation: National Institute of Mental Health and Neurosciences (NIMHANS), Bengaluru, India Siming Bayer Affiliation: Pattern Recognition Lab, Friedrich-Alexander-UniversitƤt, Erlangen-NĆ¼rnberg, Germany Andreas Maier Affiliation: Pattern Recognition Lab, Friedrich-Alexander-UniversitƤt, Erlangen-NĆ¼rnberg, Germany Abstract Multimodal medical prediction frequently contends with incomplete pairing: auxiliary modalities with complementary signal are available for only a subset of subjects (or none) and cannot be assumed at deployment. We introduce PANDA (Prototype-Anchored Data Alignment), a two-stage framework that transfers auxiliary information to a primary-modality model without auxiliary inputs at inference. Two-stage training first derives class-specific auxiliary prototypes from the shared embedding, then aligns the primary encoder to these frozen prototypes. Because supervision is defined at the class-prototype level, PANDA accommodates arbitrary pairing rates, including zero subject overlap. We evaluate PANDA on two applications. On a 1,021-subject multi-scanner ADNI cohort, we perform AD/CN classification with three auxiliary modalities at distinct pairing rates: tabular scores (44.8%), FDG-PET (18.7%), and external handwriting kinematics (0% overlap). Relative to the same-backbone MRI-only baseline, PANDA improves AUC on both encoders, with a larger gain on the weaker one: on a transfer-pretrained MedicalNet backbone it attains AUC 0.868Ā±0.0200.868Ā± 0.020 (+7.9+7.9 p over the MRI-only baseline) and reduces 1.5 T CN false positives by 25.2 p, while on a stronger, fully trainable Conv5-FC3 backbone it reaches a SOTA AUC of 0.8930.893. On TCGA-Lung survival prediction from whole-slide images, using RNA-seq as auxiliary data, PANDA improves over WSI-only on 2-year OS (AUC +3.5+3.5 p) and Cox PH (C-index +9.0+9.0 pts). It also outperforms full-fusion training, which underperforms WSI-only. PANDA requires no RNA at inference; wide confidence intervals on this smaller cohort keep the gains below conventional significance. Overall, PANDA provides a deployment-oriented mechanism for leveraging incomplete auxiliary modalities to improve primary-modality prediction. keywordsPANDA, Alzheimerās disease, MRI classification, unpaired multimodal learning, prototype alignment, cross-domain generalisation. 1 Introduction Multimodal learning has become a standard approach in the general medical domain for prediction tasks BaltruÅ”aitis et al. (2018); Zhang et al. (2011); Suk et al. (2014); however, most existing methods either assume that all modalities are jointly observed for every subject or rely on imputation and generative modelling to synthesize unobserved modalities. In practical settings, this assumption is rarely met. Auxiliary measurements that carry complementary signal, such as additional imaging, molecular assays, clinical scores, or behavioural markers, are typically available for only a subset of subjects in a large cohort. Moreover, their availability varies from one modality to the next, and they are frequently absent altogether at deployment. Discarding unpaired subjects or imputing missing modalities, the latter either by generative synthesis Sharma and Hamarneh (2020); Havaei et al. (2016); Hu et al. (2020) or by reconstructing missing-modality features through Bayesian meta-learning Ma et al. (2021), are the usual responses, but both waste the large pool of primary-modality data that is available, either shrinking the training set or injecting errors during imputation. In this study, we investigate how a classifier constrained to a single primary modality at inference time can nonetheless leverage auxiliary modalities that are only partially paired with the primary modality during training, if paired at all. To this end, we shift supervision from per-subject correspondences to class-level geometry Snell et al. (2017); Radford et al. (2021): auxiliary modalities are used solely to define fixed class prototypes, and the primary-modality encoder is aligned to these prototypes using all subjects, whether paired or unpaired. We subsequently interpret this alignment in information-theoretic terms (Section 4.3): it maximizes a lower bound on the mutual information between the primary encoder and the auxiliary modalities, thereby providing insight into why reduced pairing density need not degrade performance. The approach is developed for medical imaging and evaluated in depth on structural MRI Alzheimerās disease (AD) classification, the primary application, enabling analyses of scanner robustness and disease-severity structure. Structural MRI is a natural primary modality for the AD study because it is non-invasive, broadly available, and sensitive to the hippocampal and cortical atrophy that characterises disease progression Weiner et al. (2017); Chen et al. (2019); Wen et al. (2020). In the MRI/AD setting, real-world deployment additionally entails two interrelated challenges that are rarely addressed jointly. Transfer is further demonstrated on a structurally unrelated second application: whole-slide-image cancer survival prediction with a genomic auxiliary modality Ilse et al. (2018); Chen et al. (2021). A first challenge is scanner heterogeneity. Clinical neuroimaging datasets are acquired across sites using scanners that differ in field strength (1.5 T vs. 3 T), manufacturer, and acquisition protocol. Such variability induces systematic image-level domain shifts that can degrade classification performance Johnson et al. (2007); Fortin et al. (2017); Glocker et al. (2019). Scanner heterogeneity is frequently circumvented rather than addressed: evaluations are often restricted to a single field strength or protocol-matched cohorts, and head-to-head cross-scanner validation is uncommon, leaving scanner robustness insufficiently characterized Glocker et al. (2019); Leming et al. (2022); Song et al. (2022). In the ADNI Weiner et al. (2017) cohort, an MRI-only baseline performs substantially worse at 1.5 T than at 3 T, with elevated false-positive rates among 1.5 T control subjects. Although domain adaptation and harmonization methods can partially mitigate these shifts, they typically require scanner labels during training or paired acquisitions that are often unavailable in retrospective multi-site studies. A second challenge is incomplete multimodal coverage. Alongside MRI, clinical cohorts routinely collect auxiliary measurements, including neuropsychological scores (MMSE, CDR, FAQ), FDG-PET, and digital biomarkers, each of which provides complementary disease-relevant information. However, complete modality coverage is often unavailable at the subject level, and the paired subsets vary across modalities. In ADNI Weiner et al. (2017), cognitive scores and FDG-PET are each available for only a fraction of subjects. Moreover, some biomarker sources (e.g., digital handwriting kinematic assessments) are collected in entirely separate cohorts with no subject overlap. Standard multimodal fusion methods either require complete pairing at both training and inference Zhang et al. (2011); Suk et al. (2014) or impute missing modalities Sharma and Hamarneh (2020); Havaei et al. (2016); both are impractical at deployment and can introduce imputation-induced inductive bias. These considerations motivate the hypothesis that, despite heterogeneity in raw feature space, the modalities encode a shared disease-discriminative structure, such that AD-versus-CN class separation is relatively modality-invariant. Consequently, the class geometry induced by an auxiliary modality can serve as a supervisory target for the primary-modality encoder. Aligning the primary encoder to this shared class-level geometry, rather than to per-subject correspondences, allows auxiliary modalities to shape the MRI representation for all subjects, whether paired or unpaired. Specifically, this work introduces PANDA (Prototype-Anchored Data Alignment)11 1 A preliminary version of this work was presented in abstract form at BAIOSPHERE MEDICAL 2026 Bhat et al. (2026)., a two-stage framework illustrated in Fig. 2. In Stage 1, the MRI encoder and the auxiliary encoders are trained jointly on the available paired data to form a shared embedding, from which the auxiliary class prototypes are computed and then fixed (Fig. 2, top). In Stage 2, the MRI encoder is trained on all subjects to align with these fixed prototypes (middle). At inference time, the auxiliary encoders are discarded and each scan is classified from MRI alone (bottom). This formulation accommodates three qualitatively different pairing regimes within a single training procedure: subject-level tabular pairing (44.8 %), partial PET pairing (18.7 %), and an external handwriting-kinematics prototype with no ADNI Weiner et al. (2017) overlap (0 %). On the ADNI Weiner et al. (2017) AD/CN cohort, the full model substantially outperforms an MRI-only baseline. When the same prototype alignment is applied on a stronger, fully-trainable encoder, it attains the highest overall performance on this cohort (AUC 0.893), surpassing the strongest unimodal baseline (see Section 6.4). A scanner-stratified analysis indicates that tabular alignment primarily reduces false-positive rates at 1.5 T, whereas PET alignment improves discrimination at 3 T. A pairing-rate ablation yields a notable result: the joint Tab+PET anchor does not degrade as pairing is reduced, indicating that full pairing is unnecessary. In addition, a zero-shot analysis on held-out MCI subjects shows that an MRI-only baseline does not reliably preserve the ordinal severity continuum, whereas the joint prototype-aligned model does. A simple linear severity head, trained only on graded CN/AD scores and never on MCI, further exposes an explicit monotone CNā severity axis without reducing AUC. Finally, cross-domain validation on TCGA-Lung (whole-slide images with RNA-seq) indicates that the mechanism is not neuroimaging-specific, with directionally consistent gains that the smaller cohort leaves below significance. Thus, the main contributions of this paper are: 1. A framework for partially unpaired multimodal learning in medical imaging. PANDA is a two-stage, prototype-anchored alignment framework that enables a primary-modality classifier to leverage auxiliary modalities across pairing regimes (subject-level, partial, or external-cohort with zero overlap) within a single training procedure, while requiring only the primary modality at inference. We demonstrate the framework on two medical-imaging domainsābrain MRI (AD/CN classification) and lung-pathology whole-slide imaging (survival)āand scope the claim to these settings rather than asserting generality beyond the domains evaluated here. 2. Full pairing is not required. A pairing-rate ablation showing that sparse heterogeneous pairing matches full pairing (within seed noise) for the joint anchor, with a mechanistic explanation based on class-prototype separation, a finding with direct implications for how multimodal cohorts are best collected. 3. Alzheimerās classification (primary application). A systematic scanner-stratified evaluation on ADNI Weiner et al. (2017) showing that different auxiliary modalities target different sources of scanner bias (clinical-score prototypes reduce 1.5 T false positives while PET prototypes improve 3 T discrimination), and a zero-shot analysis on held-out MCI subjects showing that the joint prototype-aligned model preserves a monotone CNā severity ordering that a single-modality baseline does not. A severity-head extension makes this axis explicit using graded CN/AD scores only; MCI never contributes a training gradient and binary AUC is unchanged. 4. Cancer survival (second, cross-domain application). Validation on TCGA-Lung showing that the same prototype-transfer mechanism carries from 3D neuroimaging to gigapixel whole-slide imaging, exceeding full-fusion joint training despite using no RNA at inference. 2 Related Work 2.1 MRI-Based AD Classification and Scanner Robustness Deep convolutional networks operating on volumetric T1-weighted MRI have demonstrated high AD/CN discrimination performance, although such results are typically reported under constrained evaluation conditions. Moreover, a substantial share of reported gains reflects methodological artefacts rather than genuine discrimination such as, subject-level data leakage between training and test partitions Wen et al. (2020); Chowdhury et al. (2026). Despite these caveats, transfer learning from large-scale medical imaging pretraining corpora (e.g., MedicalNet Chen et al. (2019)) remains a dominant state-of-the-art (SOTA) approach in many AD/CN MRI benchmarks, as it lowers the sample complexity of AD-specific fine-tuning. Regardless of architecture, a consistent limitation remains: sensitivity to scanner domain. Models trained on 3 T ADNI scans degrade on 1.5 T acquisitions, where lower SNR and contrast differences systematically shift feature distributions Glocker et al. (2019). Two classes of approaches explicitly target this distribution shift, each with distinct assumptions and supervision requirements. Batch-effect correction methods such as ComBat Johnson et al. (2007); Fortin et al. (2017) and deep variants model scanner variation as a nuisance covariate and remove it via regression; they can be effective when scanner labels are available, but implicitly assume that biological effects are separable from scanner-induced variation, an assumption that can fail when the two are entangled. Domain-adaptation methods instead reduce sourceātarget discrepancy via adversarial training Ganin et al. (2016) or distribution matching Lee et al. (2019), but require access to target-domain scans during training and therefore do not naturally generalise to scanners unseen at development time. Both families achieve robustness by introducing an explicit scanner model and relying on privileged information (scanner labels or target-domain data). PANDA adopts an alternative strategy: instead of explicitly modelling scanner effects, it aligns the MRI representation to auxiliary modalities that are invariant to scanner acquisition by construction (e.g., cognitive scores, PET metabolism, or handwriting kinematics). This induces suppression of scanner-correlated variability as an implicit consequence of the alignment objective, without requiring scanner labels or access to target-domain scans during training or inference. Empirically, tabular-prototype alignment substantially reduces the false-positive rate on 1.5 T scans. 2.2 Multimodal Fusion and Missing Modality Multimodal fusion for AD has been studied extensively, ranging from early (feature-concatenation) to late (decision-level) fusion of MRI, PET, and clinical scores Zhang et al. (2011); Suk et al. (2014). Recent methods align multiple modalities for multiclass diagnosis, for example by combining MRI, tau-PET, diffusion MRI, and cognitive scores via contrastive alignment and a tabular foundation model Huang et al. (2025). Such methods typically assume that all modalities are available at inference, an assumption that is often violated in clinical cohorts. Missing-modality methods relax this requirement via modality-dropout training Neverova et al. (2016), generative imputation of absent modalities Sharma and Hamarneh (2020); Havaei et al. (2016), knowledge distillation from a complete-modality teacher Hu et al. (2020); Wang et al. (2020), or Bayesian meta-learning of absent-modality features Ma et al. (2021). In particular, Wang et al. Wang et al. (2020) address incomplete modalities via knowledge distillation from a complete-modality teacher; our distillation baseline follows this paradigm, whereas PANDAās prototype anchoring offers an alternative that does not require a fully paired teacher. However, these approaches still require the auxiliary modality for at least a subset of subjects (or access to a complete-modality teacher), and therefore do not apply when the auxiliary cohort has zero subject overlap, as in our handwriting regime (rHW=0r_HW=0). Prototype anchoring, by contrast, depends only on class-level geometry and can accommodate disjoint auxiliary modalities. In the federated setting, Le et al. Le et al. (2025) propose cross-modal prototype-based alignment to handle severely missing modalities across clients; we adapt a similar prototype-anchored principle to the centralised, partially unpaired setting, in which auxiliary modalities can be sparsely paired or drawn from an entirely external cohort rather than distributed across federated clients. Closer to our setting, recent AD-specific methods handle incomplete data without explicit imputation. Ou et al. Ou et al. (2024) learn a graph-embedded latent space over all subjects, including those missing PET, and Liu et al. Liu et al. (2025) align features across acquisition stages so that costly later-stage modalities need not be acquired at test time. PANDA shares this deployment objective but aligns the primary encoder to frozen class-level prototypes rather than to a fused latent space or per-subject stage features, enabling the use of an entirely external auxiliary cohort. A separate line of work addresses partial pairing by restricting multimodal training to the fully paired subset and discarding subjects missing any modality Suk et al. (2014). We show that this strategy is detrimental: a joint model trained only on the paired subset underperforms the unimodal baseline on both ADNI and TCGA (Tables 2, 6), because the paired subset is a smaller and biased sample of the training distribution. PANDA avoids this failure mode by separating alignment (Stage 1, paired subjects) from classification (Stage 2, all subjects). 2.3 Contrastive and Prototype Learning Self-supervised contrastive learning is widely used for representation learning, ranging from instance-discrimination frameworks Chen et al. (2020) to supervised contrastive objectives that encourage within-class compactness in the embedding space Khosla et al. (2020). Cross-modal contrastive learning based on the InfoNCE objective, exemplified by CLIP Radford et al. (2021) for imageātext pairs, has also been applied in medical imaging for modality alignment Zhang et al. (2022) and cross-modal retrieval. In parallel, prototypical networks Snell et al. (2017) classify samples by their distance to class-mean prototypes in embedding space, and have been adapted to few-shot medical classification Ouyang et al. (2020). PANDA departs from these lines of work in two key aspects. First, its prototypes are computed from the auxiliary-modality encoders rather than from the primary (MRI) encoder, injecting auxiliary-modality class geometry into the MRI representation. Second, although Stage 1 trains the encoders jointly, the prototypes themselves are fixed as class means at the end of Stage 1 and are not learnable during Stage 2 fine-tuning; this avoids the circularity that arises when a primary encoder is optimised against prototypes that are simultaneously updated from that same encoder. Empirically, making the prototypes learnable during MRI fine-tuning causes them to collapse onto the MRI encoderās class means, eroding their value as an independent supervisory signal. 2.4 Survival Prediction from Pathology and Genomics Attention-based multiple instance learning (ABMIL) Ilse et al. (2018) is widely used for survival prediction from gigapixel whole-slide images (WSI), where patch-level attention aggregation yields interpretable slide-level representations. Joint WSI + genomics models, including recent mixture-of-experts fusion Xiong et al. (2024), can improve upon WSI-only survival prediction Chen et al. (2022); Chen et al. (2021); however, they typically require RNA or mutation measurements at inference and are commonly trained only on the fully paired subset. Recent pathology foundation models (e.g., UNI2 Chen et al. (2024) and CONCH Lu et al. (2024)) provide high-quality patch embeddings that reduce labeled-data requirements for slide-level tasks. In our TCGA validation, we use UNI2-h ABMIL features as the WSI representation and apply the same two-stage prototype framework with RNA-seq as the auxiliary modality, demonstrating transfer from neuroimaging to computational pathology without architectural modification. 3 Dataset 3.1 ADNI AD/CN Cohort ADCN00200200400400600600178178570570119119154154(a) Field strength3 T1.5 T50ā5960ā6465ā6970ā7475ā7980ā8485+0010010020020013132323292967677373616131313737454515815821021015215280804242(b) Age distributionADCNSiemensGEPhilips00200200400400600600171171939333335255251131138585(c) ManufacturerADCN Figure 1: Cohort composition of the 1,021-subject ADNI AD/CN dataset. (a) Class distribution by field strength (AD: 297, CN: 724); AD subjects are proportionally more represented at 1.5 T (40.1%) than CN (21.2%). (b) Age distribution in 5-year bins; AD subjects are on average 2.6 years older (75.6Ā±7.875.6Ā± 7.8 vs. 73.0Ā±7.373.0Ā± 7.3 years). (c) Manufacturer distribution; Siemens dominates both groups. Data were obtained from the Alzheimerās Disease Neuroimaging Initiative (ADNI) Weiner et al. (2017), a longitudinal multi-site study that acquires structural MRI, PET imaging, and standardised neuropsychological assessments across the diagnostic spectrum. For binary AD/CN classification, we constructed a cohort of 1,021 subjects (AD: 297, CN: 724) drawn from ADNI-1/GO, ADNI-2, and ADNI-3. To avoid subject-level leakage, we selected one T1-weighted MRI per subject (earliest available visit; highest-quality acquisition) and excluded multi-channel or anomalous scans. We then created a fixed split, stratified by diagnosis and scanner field strength prior to any modelling: ā¢ Training/validation: n=844n=844 (AD: 249, CN: 595). ā¢ Held-out test: n=177n=177 (AD: 48, CN: 129). Scanner heterogeneity is a defining feature of ADNI: 273 scans (26.7%) were acquired at 1.5 T and 748 (73.3%) at 3 T (test set: 61 at 1.5 T, 116 at 3 T). Throughout this work, 1.5 T is defined as protocol_field_strength ā¤2.0ā¤ 2.0 and 3 T as protocol_field_strength >2.0>2.0. Importantly, field strength is class-imbalanced (AD: 40.1% at 1.5 T vs. CN: 21.3%; Fig. 1a), which can act as a systematic confound and motivates the scanner-stratified analyses in Section 6.2. Mean age at scan is 75.6Ā±7.875.6Ā± 7.8 years for AD and 73.0Ā±7.373.0Ā± 7.3 years for CN (two-sample t-test: t=4.55t=4.55, p<0.001p<0.001). The cohort is 53.7% female (548/1,021); the AD group is slightly male-dominated (M: 165, F: 132; 55.6% male), whereas the CN group is predominantly female (F: 416, M: 308; 57.5% female), consistent with prior ADNI analyses Weiner et al. (2017). The overall class distribution is AD: 29.1% and CN: 70.9%. Scanner manufacturer distribution is Siemens 696 (68.2%), GE 206 (20.2%), and Philips 118 (11.6%). Fig. 1 summarises these distributions. 3.2 Auxiliary Modalities Clinical tabular scores. Table 1 summarises the availability of the clinical tabular variables in the 844-subject training/validation cohort. Standardised neuropsychological and demographic variables were extracted from the ADNI repository, and we retained three key features: Mini-Mental State Examination (MMSE), Clinical Dementia Rating global score (CDR), and Functional Activities Questionnaire total score (FAQ). Cognitive and functional assessments exhibit substantial missingness, with each feature available for only about half of subjects (FAQ: 51.5%, MMSE: 45.1%, CDR: 44.8% available; Table 1), consistent with variable assessment schedules across ADNI phases. We define the fully paired subset pairedD_paired as the set of subjects for whom all three tabular features are simultaneously observed. This yields |paired|=378|D_paired|=378 training subjects (AD: 170, CN: 208), corresponding to a tabular pairing rate of rtab=44.8%r_tab=44.8\%. All tabular-prototype experiments use pairedD_paired as the Stage 1 anchor pool; feature scaling is fit independently within each training fold to prevent information leakage. Table 1: Tabular feature availability in the 844-subject training/validation cohort (AD: 249, CN: 595). The fully paired subset pairedD_paired comprises subjects for whom all three features are simultaneously available. Feature Available Missing Available (%) FAQ total score 435 409 51.5 MMSE score 381 463 45.1 CDR global score 378 466 44.8 pairedD_paired (all 3 features) 378 466 44.8 FDG-PET volumes. FDG-PET scans were obtained from the ADNI repository for a subset of the AD/CN cohort. Following DICOM-to-NIfTI conversion, 248 FDG-PET volumes were available, spanning both static and dynamic acquisitions under standard ADNI protocols. Within each training fold, approximately 158 subjects (AD: ā 70, CN: ā 88) were matched to a corresponding MRI scan; we denote this PET-paired subset by PETS_PET, yielding a pairing rate of rPETā18.7%r_PETā 18.7\%. PET availability is concentrated in later ADNI phases and is therefore predominantly associated with 3 T imaging, which makes PET anchors relatively sparse for the 1.5 T scans where scanner-induced bias is most pronounced. Handwriting kinematics (DARWIN). The DARWIN (Diagnosis AlzheimeR WIth haNdwriting) dataset Cilia et al. (2022) is a publicly available handwriting kinematics benchmark comprising 174 participants (AD and healthy controls) acquired using a digitising tablet. Participants completed 25 standardised handwriting and drawing tasks. For each task, 18 kinematic descriptors (e.g., velocity, pressure, and pen-up/pen-down ratio) were extracted, yielding a 25Ć18=45025Ć 18=450-dimensional feature vector per subject. DARWIN shares no subjects, sites, or acquisition hardware with the ADNI cohort; consequently, the handwriting pairing rate is rHW=0r_HW=0. DARWIN class prototypes are computed on the DARWIN cohort and transferred directly into Stage 2 of ADNI training without any shared subjects. 3.3 MCI Cohort for Zero-Shot Severity Evaluation We additionally define an independent cohort of 147 subjects with mild cognitive impairment (MCI) and resolvable MRI in the same image-quality-filtered pool as the AD/CN test set (EMCI: 49, MCI: 79, LMCI: 19). This cohort is withheld from all training stages, and no MCI scans, labels, or clinical scores are used for model fitting. Because these subjects are drawn from ADNI using the same sites and hardware as the AD/CN cohort, they are subject to the same 1.5 T/3 T field-strength imbalance. We use this cohort exclusively for post hoc evaluation of whether the learned MRI representations encode an ordinal disease-severity gradient beyond binary AD/CN separation. Severity proxies include MMSE, CDR global score, FAQ total score, and NPI total score. Ordinal consistency is assessed using KruskalāWallis tests and one-sided MannāWhitney tests across EMCI/MCI/LMCI, together with a monotonicity check against the CN/AD reference groups (Section 6.5). 3.4 TCGA-Lung: Cross-Domain Validation To assess whether the proposed prototype-alignment mechanism transfers beyond neuroimaging, we evaluate PANDA on a cross-domain histopathology survival prediction task using TCGA-Lung (TCGA-LUAD + TCGA-LUSC). This setting instantiates the same training constraint as ADNI, but with whole-slide imaging (WSI) features as the primary modality and bulk RNA-seq as the auxiliary modality. Modalities and availability. Diagnostic H&E WSI constitute the primary input, and bulk RNA-seq (downloaded from the GDC portal) serves as the auxiliary modality. We consider two endpoints: (i) binary 2-year overall survival (n=594n=594) and (i) Cox proportional-hazards survival modelling (n=853n=853). WSI is available for all subjects, whereas RNA-seq is available for 591/594 (99.5%) and 849/853 (99.5%), respectively, corresponding to a near-complete pairing regime. 4 Method Figure 2: Overview of the two-stage prototype-anchored alignment framework. Stage 1 jointly trains the MRI encoder together with the auxiliary encoders (tabular, PET, external handwriting) under a single combined objective (per-modality cross-entropy and pairwise cross-modal contrastive losses), then computes frozen class prototypes c(k) Ī¼_c^(k) from the auxiliary embeddings. Stage 2 trains the MRI encoder gMRIg_MRI on all N=844N=844 subjects with the combined objective ā=āCE+ākĪ»kāāproto(k)L=L_CE+ _k _kL^(k)_proto (Ļ=0.07Ļ=0.07), pulling the normalised MRI projection ip_i toward the frozen prototypes. At inference all auxiliary encoders are discarded and the class score is the convex blend of the linear-head and prototype-cosine softmaxes (Eq. 5); only MRI is required. A standard baseline in multimodal learning assumes per-subject pairing and improves performance either by fusing modalities in a joint model or by enforcing cross-modal alignment with paired samples (e.g., contrastive objectives). These designs typically require auxiliary modalities to be present for a substantial subset of subjects and, in many cases, at deployment. In our setting, auxiliary measurements are heterogeneous and can be partially paired, sparsely paired, or available only in an external cohort with no subject overlap, while the deployed model must operate on the primary modality alone. PANDA addresses this mismatch by using auxiliary data only to estimate class-level anchors. In Stage 1, auxiliary-modality encoders are used to compute class prototypes, which are then frozen; in Stage 2, the primary encoder is trained to align its representations to these fixed prototypes for all subjects. Fig. 2 summarises the overall procedure. 4.1 Framework Problem Formulation Let =(i,yi)i=1ND=\(x_i,y_i)\_i=1^N denote a dataset of N subjects, where ix_i is a 3-D T1-weighted MRI volume and yiā0,1y_iā\0,1\ is the binary AD/CN label. Every subject has an MRI scan; the primary classifier f:ā¦y^f:x y must operate on MRI alone at inference. In addition, K auxiliary modalities are available, each observed for a subset of subjects. For modality kā1,ā¦,Kkā\1,ā¦,K\, let kā1,ā¦,NS_k \1,ā¦,N\ denote the paired subset, with pairing rate rk=|k|/Nr_k=|S_k|/N. In our study, K=3K=3 auxiliary modalities are considered: ā¢ Tabular clinical scores (MMSE, CDR, FAQ): rtabā0.45r_tabā 0.45 (ntab=378n_tab=378 of 844 training subjects). ā¢ FDG-PET volumes: rPETā0.19r_PETā 0.19 (nPETā158n_PETā 158 per training fold). ā¢ Handwriting kinematics (DARWIN dataset Cilia et al. (2022)): rHW=0r_HW=0, with no subject overlap with the ADNI cohort. Prototypes are derived from the external cohort and transferred directly. The goal is to train f such that (i) it exploits all available auxiliary supervision at training time, (i) it requires no auxiliary input at inference, and (i) its learned representation is robust to scanner field-strength variation (1.5 T vs. 3 T). Encoder Architecture MRI backbone. The primary encoder gMRIg_MRI is a 3-D ResNet-18 initialised from MedicalNet Chen et al. (2019), pretrained on a large corpus of volumetric segmentation tasks. It maps each 1283128^3-voxel input to a 512-dimensional feature vector iāā512h_i ^512. A two-layer MLP projection head ĻMRI _MRI reduces ih_i to a 256-dimensional embedding ip_i, which is ā2 _2-normalised to the unit hypersphere: i=ĻMRIā(i)āĻMRIā(i)ā2.p_i= _MRI(h_i)\| _MRI(h_i)\|_2. (1) A small constant is added to norm denominators (here and in the prototype normalisation) for numerical stability. The classification head is a single linear layer applied to the 512-dimensional backbone features ih_i (before projection), so that classification and representation alignment are trained with separate parameter paths. Auxiliary encoders. Each auxiliary modality k has its own encoder gkg_k that maps its modality-specific input to a 256-dimensional ā2 _2-normalised embedding i(k)z_i^(k) in the same shared hypersphere as ip_i: ā¢ Tabular encoder: a three-layer MLP (dtabā128ā256d_tabā 128ā 256) with LayerNorm and dropout (0.3) after each hidden layer. Subjects with missing values in any tabular input are excluded from tabular-encoder training and from the prototype anchor pool, ensuring that prototypes are computed from complete feature vectors. ā¢ PET encoder: a 3-D ResNet-10 (a lightweight MedicalNet variant Chen et al. (2019)) applied to 1283128^3-voxel intensity-normalised PET volumes, followed by a two-layer MLP projection head matching the MRI head. ā¢ Handwriting encoder: a two-layer MLP (450ā256ā256450ā 256ā 256) applied to the concatenated DARWIN handwriting kinematics vector (25 tasks Ć 18 features). This encoder is trained exclusively on the external DARWIN cohort; its outputs are used only to compute class prototypes, and its parameters remain frozen throughout Stage 2. Class Prototype Construction For each auxiliary modality k and class cā0,1cā\0,1\, the class prototype c(k) Ī¼_c^(k) is the ā2 _2-normalised mean embedding over all paired training subjects of that class: c(k)=Normalizeā”(1|kc|āāiākci(k)),kc=iāk:yi=c. Ī¼_c^(k)=Normalize\! ( 1|S_k^c| _i _k^cz_i^(k) ), _k^c=\i _k:y_i=c\. (2) Prototypes are computed once from the jointly-trained auxiliary encoders at the end of Stage 1 and remain frozen for the remainder of training. Freezing prevents prototype drift and ensures that the MRI encoder optimises toward stable geometric targets. For the handwriting modality (rHW=0r_HW=0), the paired subset HWS_HW contains no ADNI subjects. We therefore compute handwriting class prototypes on the external DARWIN cohort and transfer them directly as 0HW,1HW\ Ī¼_0^HW, Ī¼_1^HW\. This external prototype transfer regime evaluates whether class-level geometry learned in an entirely separate cohort can act as an effective anchor for MRI representation learning. Moreover, it serves as a control against the hypothesis that prototype alignment requires subject-level correspondence between modalities: any observed benefit in this regime must arise from the transferred class separation in prototype space rather than from per-subject cross-modal pairing. 4.2 Training and Inference scheme Stage 1: joint multi-modal training and frozen prototypes. Stage 1 jointly trains the MRI encoder and all auxiliary encoders using a single optimiser and a unified objective. The loss comprises (i) per-modality cross-entropy terms for modalities with AD/CN labels and (i) supervised cross-modal contrastive terms over the available paired subsets (MRIātabular, MRIāPET, and tabularāPET), yielding a shared ā2 _2-normalised embedding space. The handwriting encoder is trained on the external DARWIN cohort and contributes only class-level prototypes (no subject overlap with ADNI). After Stage 1, all auxiliary encoders are frozen and class prototypes c(k)\ Ī¼_c^(k)\ are computed once via (2) and held fixed for Stage 2. Stage 2: prototype-anchored MRI training. The MRI encoder is then trained on the full dataset D, regardless of auxiliary modality availability, with the combined loss ā=āCE+āk=1KĪ»kāāproto(k),L=L_CE+ _k=1^K _kL^(k)_proto, (3) where āCEL_CE is the cross-entropy loss on AD/CN logits computed from backbone features ih_i, each Ī»k _k weights modality k, and each āproto(k)L^(k)_proto is a temperature-scaled prototype cross-entropy Snell et al. (2017) that pulls the normalised MRI projection ip_i toward the frozen prototype of its correct class: ā(k)proto=ā1|ā¬|āiāā¬logexpā”(iā yi(k)/Ļ)ācexpā”(iā c(k)/Ļ),L^(k)_proto=- 1|B| _i \! (p_iĀ· Ī¼^(k)_y_i/Ļ ) _c \! (p_iĀ· Ī¼^(k)_c/Ļ ), (4) with ā¬B a training mini-batch, Ļ the temperature parameter Radford et al. (2021), and cā0,1cā\0,1\ ranging over both classes. The defining property of this objective is that the same prototype yi(k) Ī¼^(k)_y_i is used for every subject i, whether or not iāki _k: paired and unpaired subjects receive identical geometric targets for a given modality, so no subject is excluded from alignment supervision. This is what allows the MRI encoder to learn from the entire training cohort even though the auxiliary signal is defined on the paired subset alone. Inference At test time, all auxiliary encoders are discarded. Classification of a new subject uses only the MRI backbone gMRIg_MRI and the linear classification head, incurring no additional computational cost relative to the MRI-only baseline. At inference the class score for a test subject is the weighted average of the linear-head softmax, whose logits are āā”(i) (h_i), and the prototype cosine softmax, p^ā(cā£i) p(c _i) =Ī±0āsoftmaxcā(āā”(i))+Ī±1āsoftmaxcā(ākāTab,PET,HWiā c(k)/Ļ), = _0\,softmax_c ( (h_i) )+ _1\,softmax_c ( _kā\Tab,PET,HW\p_iĀ· Ī¼^(k)_c/Ļ ), (5) Ī±0,Ī±1ā[0,1],Ī±0+Ī±1=1, _0, _1ā[0,1], _0+ _1=1, with y^=argā”maxcāp^ā(cā£i) y= _c p(c _i). The weight (Ī±0,Ī±1)( _0, _1) is fit on the validation fold (Implementation Details, Section 5); the reported pADp_AD (including the zero-shot MCI analysis of Section 6.5) is p^ā(ADā£i) p(AD _i). When Ī±1=0 _1=0 this reduces to the MRI-only linear classifier. The prototype-cosine term sums the raw cosine logits against all three frozen prototype sets (kāTab,PET,HWkā\Tab,PET,HW\) with equal weight and a single shared temperature Ļ=0.07Ļ=0.07, followed by a single softmax; since the prototypes c(k) Ī¼^(k)_c are frozen constants, no auxiliary input is required at inference and the classifier remains MRI-only for every test subject. Severity-Axis Extension The base framework is trained for binary CN/AD discrimination and does not explicitly enforce an ordinal severity structure (Section 6.5). We therefore augment Stage 2 with a linear severity head hsev:ā512āāh_sev:R^512 applied to the 512512-dimensional backbone features ih_i (the pre-projection encoder output), so that the severity signal is read off features independent of the tabular/PET/hand-crafted-aligned projection space, and trained by masked Huber regression against graded clinical-severity composites computed on the CN/AD training pool only: āsev=1||āāiāHuberā”(hsevā(i),si),L_sev= 1|S| _i Huber (h_sev(h_i),\,s_i ), (6) where S denotes CN/AD training subjects with complete coverage for the chosen composite and sis_i is the corresponding z-scored value. We consider two composites: sorths_orth (mean of z-scored FAQ and NPI; n=265n=265 of 844 CN/AD training subjects) and sdiags_diag (mean of z-scored CDR and negated z-scored MMSE; n=378n=378). Unless otherwise stated, we use sorths_orth because FAQ/NPI reflect functional and neuropsychiatric status rather than directly restating the diagnostic criteria used to assign CN/AD labels. In the CN/AD training pool, FAQ is observed for 435435 subjects and NPI for 265265; because every NPI-observed subject also has FAQ, NPI is the limiting factor and the complete-coverage set for sorths_orth is exactly these n=265n=265 subjects. Each component is z-scored independently over all CN/AD subjects for which it is observed (FAQ over 435435, NPI over 265265) before averaging. Missing values are handled by exclusion rather than imputation: the masked Huber loss requires both raw FAQ and NPI to be present, so only the 265265 complete-coverage subjects contribute gradients (sdiags_diag is defined analogously, requiring both CDR and MMSE, n=378n=378). This complete-coverage subset is also field-strengthāskewed (208 subjects at ā¤2.0ā¤ 2.0 T vs. 57 at >2.0>2.0 T), so the learned severity axis is confounded by lower-field scanner characteristics; we therefore treat it as an exploratory readout and interpret absolute severity values with corresponding caution (see also Section 7). The severity head is optimised jointly with the Stage 2 objective, āCE+0.3ā(ātab+āPET+āHW)+Ī»sevāāsevL_CE+0.3(L_tab+L_PET+L_HW)+ _sevL_sev with Ī»sev=0.3 _sev=0.3; equivalently, this is the general Stage-2 objective of (3) with Ī»k=0.3 _k=0.3 for all prototype terms, augmented by the severity term. Stage 1 is reused from the frozen joint checkpoint (i.e., not retrained), isolating the effect of the severity head from changes in the Stage 1 alignment geometry. MCI subjects contribute no training gradients at any stage: the composite is defined and standardised using CN/AD subjects only, and the head is evaluated on MCI in a zero-shot manner. 4.3 Information-Theoretic Motivation We provide a brief information-theoretic motivation for prototype anchoring and for the empirical observation that reduced pairing need not degrade performance. These arguments are heuristic, intended to build intuition rather than to constitute rigorous proofs, and rely on simplifying assumptions that we state where used. Let Yā0,1Yā\0,1\ denote the class label, (k)z^(k) the Stage 1 auxiliary embedding for modality k, and p the ā2 _2-normalised MRI projection in (1). Alignment transfers class information. The Stage 1 cross-modal InfoNCE objective provides a variational lower bound on the mutual information between paired MRI and auxiliary embeddings, Iā”(,(k))ā„logā”Nā¬āāNCEI(p;z^(k))ā„ N_B-L_NCE Poole et al. (2019). This bound pertains to the InfoNCE term alone; the full Stage 1 objective, which also includes per-modality cross-entropy terms, does not directly inherit it. The class prototype c(k)=Normalizeā”(ā”[(k)ā£Y=c]) Ī¼^(k)_c=Normalize (E[z^(k) Y=c] ) preserves class-discriminative directions of the auxiliary embedding. Minimising the prototype cross-entropy in (4) therefore encourages p to align with the auxiliary-induced class geometry Boudiaf et al. (2020). Because the target prototypes are fixed and class-level, this supervisory signal is available to all subjects during Stage 2, including those without paired auxiliary measurements. Prototype separation as an anchor-quality proxy. Consider the cosine score sc=ā c(k)/Ļs_c=pĀ· Ī¼^(k)_c/Ļ under a two-class Gaussian-channel approximation with equal class-conditional variances. Then Iā”(s,Y)I(s;Y) is monotone in the deflection coefficient J=ā„1(k)ā0(k)ā„2/(Ļ02+Ļ12)J= Ī¼^(k)_1- Ī¼^(k)_0 ^2/( _0^2+ _1^2) (assuming Ļ02+Ļ12>0 _0^2+ _1^2>0), so larger inter-prototype separation yields a stronger anchor. Since prototypes are finite-sample estimates of class means, we have āā„^1ā^0ā„2=ā„1ā0ā„2+trā”Varā”(^1ā^0)E Ī¼_1- Ī¼_0 ^2= Ī¼_1- Ī¼_0 ^2+tr\,Var( Ī¼_1- Ī¼_0), which implies that subsampling can increase empirical separation while also increasing estimator variance. This identity assumes unnormalised class means; for the ā2 _2-normalised prototypes used here it holds only approximately, and the bias term is also affected by the normalisation. Consequently, moderate reductions in pairing can leave the empirical anchor quality unchanged (or slightly improved) until the class-mean estimates become too noisy (empirically, around ā40ā 40 paired AD subjects per fold; Section 6.3). Importantly, this is an estimator effect on empirical prototypes rather than an increase in the population mutual information Iā”((k),Y)I(z^(k);Y). We test this prediction by measuring inter-prototype separation as a function of pairing rate in Section 6.3. 5 Experimental Setup Implementation Details MRI volumes were resampled to 1 m isotropic in RAS orientation and cropped to 1283128^3 voxels using MONAI Cardoso et al. (2022). Intensity normalisation (zero mean, unit variance over non-background voxels) and training-time augmentation (random affine: Ā±10āĀ± 10 , Ā±10%Ā± 10\% scale, Ā±10Ā± 10 m translation) were applied. FDG-PET volumes were preprocessed to 1283128^3 voxels at 2 m isotropic resolution with 99th-percentile intensity normalisation. All experiments use three random but fixed seeds under 5-fold stratified cross-validation on the training set; the test set is evaluated once per seed using the best validation-AUC checkpoint. Reported metrics are mean Ā± SD across the three seeds. Stage 1 uses a single AdamW optimiser (weight decay 10ā410^-4) with fixed learning rates and no scheduler: MRI parameters (layer3/layer4 and the projection head) use lr =10ā4=10^-4 and the auxiliary encoders (tabular, PET, handwriting) use lr =5Ć10ā5=5Ć 10^-5, trained for 50 epochs (handwriting: 100). Tabular feature scaling is fit on the training fold only. We set Ļ=0.07Ļ=0.07 and Ī»tab=0.5 _tab=0.5, Ī»PET=0.3 _PET=0.3, Ī»HW=0.5 _HW=0.5 on validation AUC at 100% pairing, held fixed across pairing-rate ablations. Stage 2 then fine-tunes the MRI encoder alone, also with AdamW but at a lower backbone rate (10ā510^-5; projection head 10ā410^-4), for up to 40 epochs with early stopping (patience =10=10) on validation AUC and ReduceLROnPlateau (factor =0.5=0.5, patience =3=3). All runs execute on 2ĆA100 GPUs. The inference blend weights (Ī±0,Ī±1)( _0, _1) in (5) are fit independently for each Stage-2 fold (and seed) on that foldās held-out validation splitāthe same split used for checkpoint selection, with no separate nested cross-validationāby minimising the negative log-likelihood of the blended probability against the validation labels. The weights are not grid-searched: (Ī±0,Ī±1)=softmaxā”()( _0, _1)=softmax( Īø) is parameterised by a single learnable 22-vector Īø (initialised at zero) and optimised with Adam (learning rate 10ā310^-3) for 200200 full-batch epochs, retaining the best-validation-loss epoch. At test time the five folds are combined by soft-voting (averaging the blended p p). Across the 1515 fold/seed fits the objective is bimodal rather than a single stable optimum: roughly half the folds converge near (0.40,0.60)(0.40,0.60) and the remainder to the mirror configuration (0.59,0.41)(0.59,0.41) that up-weights the linear head, giving a mean of (Ī±0,Ī±1)=(0.487Ā±0.091, 0.513Ā±0.091)( _0, _1)=(0.487Ā± 0.091,\,0.513Ā± 0.091) across folds. The soft-voting ensemble averages over this per-fold variation, and the near-equal mean indicates the two heads contribute comparably overall. Baselines Unless otherwise stated, ADNI baselines are trained on the full 844-subject training/validation set and evaluated on the held-out test set (n=177n=177) with MRI-only inputs at inference. Auxiliary modalities, when used, are provided during training only. We compare against 13 ADNI baselines and 2 TCGA baselines spanning (i) standard multimodal fusion, (i) missing-modality training strategies, (i) teacherāstudent distillation, and (iv) recent published methods. For each comparator, Appendix A (Table 7) specifies the training-time auxiliary inputs, inference-time inputs, and any protocol adaptations required for fair comparison. Pairing-Rate Ablation Protocol To quantify sensitivity to reduced auxiliary pairing, we subsample the paired pools kS_k prior to Stage 2. For the joint Tab+PET anchor we consider fractions rā1.0,0.75,0.5,0.25,0.15,0.10,0.05rā\1.0,0.75,0.5,0.25,0.15,0.10,0.05\, and for the tabular-only anchor we use rā1.0,0.5,0.25rā\1.0,0.5,0.25\. Subsampling is performed independently per modality and stratified by class to preserve class proportions within each prototype pool. For each fraction, we reuse the jointly trained Stage 1 encoder weights and recompute only the prototype vectors c(k) Ī¼_c^(k) on the subsampled pools. Stage 2 training (MRI encoder on all N=844N=844 subjects) and inference are otherwise unchanged. This protocol isolates the effect of prototype estimation quality from changes in the size of the MRI training set. Cross-Domain Evaluation Setup To assess domain generality, we apply PANDA, unchanged, to a histopathology survival task on TCGA-Lung (LUAD + LUSC). Here the primary modality is whole-slide images (WSI): diagnostic H&E slides are tiled at 20Ć magnification, encoded with the UNI2-h pathology foundation model Chen et al. (2024), and aggregated per slide by attention-based multiple-instance learning (ABMIL) into a 1536-dimensional bag embedding. The auxiliary modality is bulk RNA-seq (ā 99% pairing): raw counts are reduced to the top-2000 most variable genes, log1p-transformed, and z-scored per gene on the training fold. We evaluate two survival endpoints. For binary 2-year overall survival, subjects are labelled as deceased within 730 days (11) or surviving beyond 730 days (00); censored cases with follow-up ā¤730ā¤ 730 days are excluded. This yields n=594n=594 subjects (test =119=119) with a label-stratified split (random_state=0=0). For Cox proportional-hazards modelling, we include all subjects with valid follow-up and treat censored cases through the partial likelihood, yielding n=853n=853 (test =171=171) with stratification by event indicator. All TCGA experiments use the same three random seeds as the ADNI setup. To probe sensitivity to reduced auxiliary coverage, we subsample the RNA-paired pool at rRNAā0.25,0.50,1.00r_RNAā\0.25,0.50,1.00\. We report AUC, balanced accuracy, and macro-F1 for the binary endpoint and Harrellās C-index for the Cox model; all other architectural and optimisation settings are unchanged. This cross-domain evaluation tests whether the proposed prototype-alignment mechanism extends beyond neuroimaging and beyond sparsely paired auxiliary modalities. 6 Results 6.1 Main Results Table 2: Held-out test performance. Training-set size (number of subjects) is given in parentheses after each method: the top block trains on the full 844-subject pool (exploiting unpaired subjects), the bottom block on the 378 tab-complete subjects only. Training lists the modalities available during training. Unless noted, evaluation is on the held-out test set (n=177n=177; AD =48=48, CN =129=129) with MRI-only inputs at inference. Mean Ā± SD across 3 seeds. Bold indicates the best value per column. Bootstrap 95 % CIs for primary comparisons vs. MRI-only are reported in the text (10 000 iterations on seed-averaged probabilities). Significance of the AUC gain vs. MRI-only (bootstrap with HolmāBonferroni correction): āp<0.05^*p<0.05, pāā£ā<0.01^**p<0.01, āp<0.001^***p<0.001. ā Suk et al. is evaluated on the paired test subset (n=75n=75) rather than the full n=177n=177 set, as its late-fusion design requires both modalities at inference; it is therefore not directly comparable to the other rows. The final block lists independently-published architectures with different backbones, run on our protocol for reference and excluded from the per-column bold. Method (Training size) Training AUC ā BalAcc ā MacroF1 ā F1-AD ā Rec-AD ā Rec-CN ā PANDA (844) MRI+Tab 0.790Ā±.0120.790_Ā±.012 0.692Ā±.0210.692_Ā±.021 0.689Ā±.0190.689_Ā±.019 0.550Ā±.0310.550_Ā±.031 0.563Ā±.0450.563_Ā±.045 0.822Ā±.0110.822_Ā±.011 PANDA (844) MRI+Tab+PET 0.851Ā±.0100.851_Ā±.010 0.763Ā±.0360.763_Ā±.036 0.733Ā±.0370.733_Ā±.037 0.638Ā±.0460.638_Ā±.046 0.750Ā±.0740.750_Ā±.074 0.775Ā±.0550.775_Ā±.055 PANDA (844) MRI+Tab+PET+HW 0.868Ā±.020ā0.868_Ā±.020^* 0.772Ā±.0410.772_Ā±.041 0.767Ā±.0330.767_Ā±.033 0.662Ā±.0540.662_Ā±.054 0.681Ā±.0870.681_Ā±.087 0.863Ā±.0070.863_Ā±.007 MRI-only Baseline (844) MRI 0.789Ā±.0160.789_Ā±.016 0.691Ā±.0190.691_Ā±.019 0.670Ā±.0090.670_Ā±.009 0.546Ā±.0260.546_Ā±.026 0.632Ā±.0800.632_Ā±.080 0.749Ā±.0450.749_Ā±.045 Knowledge distillation Hu et al. (2020) (844) MRI+Tab+PET 0.777Ā±.0160.777_Ā±.016 0.682Ā±.0080.682_Ā±.008 0.669Ā±.0060.669_Ā±.006 0.537Ā±.0100.537_Ā±.010 0.597Ā±.0390.597_Ā±.039 0.767Ā±.0290.767_Ā±.029 Modality dropout (844) MRI+Tab 0.800Ā±.0060.800_Ā±.006 0.701Ā±.0050.701_Ā±.005 0.682Ā±.0060.682_Ā±.006 0.561Ā±.0060.561_Ā±.006 0.639Ā±.0100.639_Ā±.010 0.762Ā±.0130.762_Ā±.013 Modality dropout (844) MRI+Tab+PET 0.799Ā±.0200.799_Ā±.020 0.693Ā±.0170.693_Ā±.017 0.669Ā±.0080.669_Ā±.008 0.550Ā±.0200.550_Ā±.020 0.653Ā±.0550.653_Ā±.055 0.734Ā±.0200.734_Ā±.020 HeMIS Havaei et al. (2016) (844) MRI+Tab 0.775Ā±.0080.775_Ā±.008 0.688Ā±.0320.688_Ā±.032 0.699Ā±.0310.699_Ā±.031 0.542Ā±.0530.542_Ā±.053 0.486Ā±.0690.486_Ā±.069 0.889Ā±.0100.889_Ā±.010 Graph-smoothness reg. Weston et al. (2008) (844) MRI+Tab 0.753Ā±.0140.753_Ā±.014 0.667Ā±.0200.667_Ā±.020 0.650Ā±.0290.650_Ā±.029 0.520Ā±.0230.520_Ā±.023 0.597Ā±.0200.597_Ā±.020 0.736Ā±.0580.736_Ā±.058 Graph-SLC Ou et al. (2024) (844) MRI+Tab 0.802Ā±.0160.802_Ā±.016 0.700Ā±.0260.700_Ā±.026 0.695Ā±.0130.695_Ā±.013 0.559Ā±.0380.559_Ā±.038 0.576Ā±.0980.576_Ā±.098 0.824Ā±.0480.824_Ā±.048 Fusion (pairs only) (378) MRI+Tab 0.750Ā±.0220.750_Ā±.022 0.653Ā±.0020.653_Ā±.002 0.654Ā±.0030.654_Ā±.003 0.493Ā±.0050.493_Ā±.005 0.486Ā±.0260.486_Ā±.026 0.819Ā±.0220.819_Ā±.022 Suk et al. Suk et al. (2014) (378)ā MRI+Tab 0.689Ā±.0630.689_Ā±.063 0.628Ā±.0880.628_Ā±.088 0.590Ā±.1400.590_Ā±.140 0.610Ā±.0350.610_Ā±.035 0.710Ā±.1210.710_Ā±.121 0.545Ā±.2920.545_Ā±.292 DiaMond Li et al. (2025) MRI+PET 0.641Ā±.0230.641_Ā±.023 0.601Ā±.0190.601_Ā±.019 0.602Ā±.0270.602_Ā±.027 0.415Ā±.0180.415_Ā±.018 0.403Ā±.0550.403_Ā±.055 0.798Ā±.0840.798_Ā±.084 Wang et al. Wang et al. (2024) MRI 0.752Ā±.0710.752_Ā±.071 0.677Ā±.0670.677_Ā±.067 0.670Ā±.0540.670_Ā±.054 0.521Ā±.0920.521_Ā±.092 0.542Ā±.1480.542_Ā±.148 0.811Ā±.0160.811_Ā±.016 HyperFusion Duenias et al. (2025) MRI+Tab 0.750Ā±.0100.750_Ā±.010 0.592Ā±.0510.592_Ā±.051 0.586Ā±.0640.586_Ā±.064 0.344Ā±.1340.344_Ā±.134 0.292Ā±.1530.292_Ā±.153 0.892Ā±.0510.892_Ā±.051 IC-MKD Kwak et al. (2025) MRI+PET 0.763Ā±.0040.763_Ā±.004 0.648Ā±.0120.648_Ā±.012 0.654Ā±.0100.654_Ā±.010 0.483Ā±.0210.483_Ā±.021 0.451Ā±.0350.451_Ā±.035 0.845Ā±.0110.845_Ā±.011 Table 2 summarises held-out performance on the 177-subject ADNI Weiner et al. (2017) test set. The MRI-only baseline attains AUC =0.789Ā±0.016=0.789Ā± 0.016. Training a conventional fusion model on the 378-subject fully paired subset (Fusion, pairs only) does not improve AUC, indicating that restricting training to paired subjects is suboptimal. With the same tabular pairing rate, PANDA (MRI+Tab) yields comparable AUC (0.790Ā±0.0120.790Ā± 0.012; padj=1.00p_adj=1.00 vs. MRI-only) while improving balanced accuracy (0.692) and macro-F1 (0.689), with higher CN recall (0.822 vs. 0.749). Adding PET prototypes (PANDA: MRI+Tab+PET, rPET=0.19r_PET=0.19) increases AUC to 0.851Ā±0.0100.851Ā± 0.010 (not significant after HolmāBonferroni, padj=0.119p_adj=0.119) and improves AD recall from 0.632 to 0.750. Incorporating handwriting prototypes from the external DARWIN cohort Cilia et al. (2022) (PANDA: MRI+Tab+PET+HW, rHW=0r_HW=0) further increases AUC to 0.868Ā±0.0200.868Ā± 0.020 (+7.9+7.9 p vs. MRI-only; padj=0.011p_adj=0.011) and raises CN recall to 0.863, demonstrating effective transfer of class geometry from a cohort with zero subject overlap. Among baselines, Graph-SLC (a graph-embedded latent-space clustering method) Ou et al. (2024) achieves AUC =0.802Ā±0.016=0.802Ā± 0.016, and modality-dropout variants are similar (0.800 and 0.799; p=0.608p=0.608 vs. MRI-only). Although these methods slightly exceed PANDA (MRI+Tab) in AUC, they do not mitigate scanner-associated error (1.5 T CN false-positive rate 41.441.4ā47.8%47.8\% vs. 52.2%52.2\% for MRI-only), whereas PANDA with tabular anchoring reduces it to 27.9%27.9\% (Table 3). Knowledge distillation Hu et al. (2020), HeMIS Havaei et al. (2016), and graph-smoothness regularisation Weston et al. (2008) perform at or below the MRI-only baseline. Additional published architectures evaluated under our protocol (DiaMond, a bi-modal MRIāPET vision transformer Li et al. (2025); the diagnosis network of Wang et al. Wang et al. (2024); HyperFusion, a tabular-conditioned hypernetwork Duenias et al. (2025); and IC-MKD, incomplete cross-modal mutual knowledge distillation Kwak et al. (2025)) also underperform PANDA in AUC. Fig. 3 provides a qualitative visualisation of representation changes. Figure 3: Held-out test-set MRI embeddings (seed 42, fold 0) under PCA and t-SNE, for MRI-only, MRI+Tab+PET, and the full MRI+Tab+PET+HW model. F is the Fisher discriminability ratio of CN vs. AD on the primary axis of each projection. Progressive auxiliary alignment reorganises the unsupervised embedding geometry into a more clearly class-separated arrangement: F rises from 0.010.01 for MRI-only to 1.141.14 (PCA) / 0.640.64 (t-SNE) for MRI+Tab+PET and 0.960.96 (PCA) / 1.011.01 (t-SNE) for the full MRI+Tab+PET+HW model. The projections illustrate this qualitative reorganisation; quantitative discriminability is reported by test AUC in Table 2. 6.2 Scanner-Stratified Analysis Table 3: Scanner field-strength stratification. 1.5 T: n=61n=61 (AD =24=24, CN =37=37); 3 T: n=116n=116 (AD =24=24, CN =92=92). FP % = CN misclassified as AD. Mean Ā± SD across 3 seeds. Bold: best per column. ā Suk et al. is evaluated on the paired subset: 1.5 T is still n=61n=61, but its 3 T group is n=14n=14 (not 116), so its 3 T column is not directly comparable to the other rows. The final block lists independently-published architectures with different backbones, shown for reference and excluded from the per-column bold. Method 1.5T AUC ā 1.5T FP % ā 3T AUC ā 3T FP % ā PANDA: MRI+Tab 0.703Ā±.0090.703_Ā±.009 27.9Ā±3.427.9_Ā± 3.4 0.804Ā±.0120.804_Ā±.012 13.8Ā±2.213.8_Ā± 2.2 PANDA: MRI+Tab+PET 0.736Ā±.0030.736_Ā±.003 39.6Ā±4.539.6_Ā± 4.5 0.886Ā±.0200.886_Ā±.020 15.6Ā±5.915.6_Ā± 5.9 PANDA: MRI+Tab+PET+HW 0.783Ā±.0130.783_Ā±.013 27.0Ā±6.627.0_Ā± 6.6 0.894Ā±.0190.894_Ā±.019 8.3Ā±1.88.3_Ā± 1.8 MRI-only Baseline 0.695Ā±.0090.695_Ā±.009 52.2Ā±11.452.2_Ā± 11.4 0.814Ā±.0220.814_Ā±.022 14.1Ā±2.314.1_Ā± 2.3 Knowledge distillation Hu et al. (2020) 0.678Ā±.0290.678_Ā±.029 41.4Ā±1.341.4_Ā± 1.3 0.800Ā±.0060.800_Ā±.006 15.9Ā±4.615.9_Ā± 4.6 Modality dropout (Tab) 0.699Ā±.0120.699_Ā±.012 41.4Ā±2.541.4_Ā± 2.5 0.825Ā±.0130.825_Ā±.013 16.7Ā±1.416.7_Ā± 1.4 Modality dropout (Tab+PET) 0.717Ā±.0120.717_Ā±.012 47.8Ā±8.947.8_Ā± 8.9 0.816Ā±.0230.816_Ā±.023 18.1Ā±1.018.1_Ā± 1.0 HeMIS Havaei et al. (2016) 0.665Ā±.0190.665_Ā±.019 17.1Ā±4.617.1_Ā± 4.6 0.823Ā±.0050.823_Ā±.005 8.7Ā±1.88.7_Ā± 1.8 Graph-smoothness reg. Weston et al. (2008) 0.653Ā±.0220.653_Ā±.022 55.0Ā±11.355.0_Ā± 11.3 0.762Ā±.0180.762_Ā±.018 14.9Ā±3.614.9_Ā± 3.6 Graph-SLC Ou et al. (2024) 0.703Ā±.0320.703_Ā±.032 34.2Ā±7.134.2_Ā± 7.1 0.829Ā±.0290.829_Ā±.029 10.9Ā±4.110.9_Ā± 4.1 Suk et al. Suk et al. (2014)ā 0.682Ā±.0770.682_Ā±.077 45.0Ā±29.445.0_Ā± 29.4 0.707Ā±.1080.707_Ā±.108 47.6Ā±29.447.6_Ā± 29.4 DiaMond Li et al. (2025) 0.494Ā±.0390.494_Ā±.039 38.7Ā±12.538.7_Ā± 12.5 0.682Ā±.0250.682_Ā±.025 12.7Ā±8.112.7_Ā± 8.1 Wang et al. Wang et al. (2024) 0.667Ā±.0980.667_Ā±.098 44.1Ā±5.644.1_Ā± 5.6 0.762Ā±.0820.762_Ā±.082 8.7Ā±3.98.7_Ā± 3.9 HyperFusion Duenias et al. (2025) 0.630Ā±.0110.630_Ā±.011 25.2Ā±10.025.2_Ā± 10.0 0.798Ā±.0110.798_Ā±.011 5.1Ā±3.15.1_Ā± 3.1 IC-MKD Kwak et al. (2025) 0.672Ā±.0420.672_Ā±.042 46.8Ā±3.446.8_Ā± 3.4 0.784Ā±.0050.784_Ā±.005 2.9Ā±0.52.9_Ā± 0.5 Table 3 reports performance stratified by scanner field strength. The MRI-only baseline exhibits pronounced field-strength sensitivity, with AUC =0.695=0.695 at 1.5 T versus 0.814 at 3 T, and a 1.5 T CN false-positive rate of 52.2%. This behaviour is consistent with class-imbalanced acquisition (38.6% of AD vs. 20.7% of CN acquired at 1.5 T in the training cohort), which permits a field-strength shortcut. Tabular prototype alignment (PANDA: MRI+Tab) primarily reduces scanner-associated false positives, lowering the 1.5 T FP % from 52.2% to 27.9% with minimal change in AUC. Adding PET prototypes (PANDA: MRI+Tab+PET) increases discrimination, particularly at 3 T (AUC 0.886), but yields a higher 1.5 T FP % (39.6%), consistent with sparse PET pairing limiting its influence on the 1.5 T subgroup. The full model (MRI+Tab+PET+HW) achieves the strongest overall stratified profile (1.5 T AUC =0.783=0.783, FP % =27.0=27.0; 3 T AUC =0.894=0.894, FP % =8.3=8.3). HeMIS attains a lower 1.5 T FP % (17.1%) but does so with substantially reduced AD sensitivity (Rec-AD =0.486=0.486; Table 2). The handwriting anchor, despite having no ADNI subjects, provides complementary scanner-bias mitigation to the PET-driven AUC gains, yielding the most balanced trade-off. Fig. 4 plots these four columns directly; Fig. 5 shows the corresponding t-SNE embeddings coloured by field strength rather than by diagnosis, making the 1.5 T/3 T mixing visible per method. Figure 4: Scanner-stratified AUC and CN false-positive rate across methods (3 seeds, mean Ā± std). The full model (MRI+Tab+PET+HW) attains the highest AUC at both field strengths and the lowest 3T FP%, reducing 1.5T FP% to 27.0%, without using scanner labels during training. Figure 5: t-SNE of held-out test-set MRI embeddings (seed 42, fold 0) coloured by scanner field strength (1.5 T vs. 3 T) rather than by diagnosis; marker shape indicates CN/AD. Progressive auxiliary alignment (MRI-only ā MRI+Tab+PET ā MRI+Tab+PET+HW) visibly interleaves the two scanner populations, consistent with the false-positive-rate reduction in Fig. 4. 6.3 Ablations Pairing Rate Sensitivity Fig. 6 reports held-out test AUC as the paired fraction r is reduced from 100% to 5%. Figure 6: Test AUC vs. pairing fraction r for Tab-only and Tab+PET alignment (mean Ā± std). Dashed line: MRI-only baseline (0.7890.789). Tab+PET AUC is flat within seed noise across r; Tab-only degrades below 25%25\%. For tabular-only alignment, AUC is stable down to r=50%r=50\% (0.790 to 0.785) but decreases at r=25%r=25\% (0.758; ā95ā 95 paired subjects), which corresponds to approximately 40 AD subjects per fold for prototype estimation. For the joint Tab+PET anchor, AUC shows no systematic dependence on pairing rate: across rā100%,ā¦,5%rā\100\%,ā¦,5\%\ it remains within 0.849ā0.867, with overlapping seed variance. Thus, reduced pairing does not improve performance, but demonstrates that full pairing is not required: rā5rā 5ā10%10\% matches the fully paired configuration within seed noise. This behaviour is consistent with an anchor-quality effect under finite-sample prototype estimation. For tabular prototypes, the inter-class cosine similarity (AD vs. CN) becomes more negative as pairing decreases (mean cosine: ā0.649-0.649 at 100%, ā0.731-0.731 at 25%, ā0.760-0.760 at 10%), indicating increased empirical class separation. PET prototypes remain maximally anti-parallel by construction (cosine =ā1.000=-1.000) and therefore do not contribute to this trend. A plausible explanation is that the full paired pool (n=378n=378) contains a wide severity spectrum, including borderline cases that shift class means toward each other. Subsampling can preferentially yield more homogeneous within-class subsets, increasing empirical prototype separation and providing a more stable alignment target during Stage 2. Optimal-Transport Pseudo-Pairing Stage 1 contrastive alignment uses only genuinely paired subjects (tabular: 378/844; PET: ā 158/844), while handwriting contributes external class prototypes without per-subject pairing. To test whether unpaired MRI subjects can benefit from an additional alignment signal, we introduce an intermediate step that periodically (every 5 epochs) computes a Sinkhorn optimal-transport coupling between the MRI embedding distribution (all 844 subjects) and each auxiliary embedding distribution using cost 1ācosā”(ā ,ā )1- (Ā·,Ā·). The resulting soft couplings are used as additional (pseudo-paired) InfoNCE supervision, thereby achieving an AUC score of 0.862Ā±0.0140.862Ā± 0.014, which is statistically indistinguishable from the baseline 0.868Ā±0.0200.868Ā± 0.020 and within our equivalence band. We therefore treat optimal-transport pseudo-pairing as a neutral ablation under the pairing densities considered, indicating that class-level prototype anchoring captures most of the transferable cross-modal geometry in this setting. 6.4 Encoder Trainability and Backbone Generalizability Our primary MRI backbone is a MedicalNet ResNet-18 in which the first two residual stages are frozen at segmentation-pretrained weights. As a stronger fully trainable alternative, we consider the Conv5-FC3 encoder of Wen et al. Wen et al. (2020) (autoencoder warm start on the same ADNI pool; end-to-end training). On the same n=177n=177 evaluation protocol, Conv5-FC3 improves the MRI-only baseline AUC from 0.789Ā±0.0160.789Ā± 0.016 to 0.881Ā±0.0090.881Ā± 0.009. The largest difference occurs at 1.5 T (AUC 0.6950.695 vs. 0.800; CN FP % 52.352.3 vs. 33.3), consistent with limited trainability increasing reliance on scanner-correlated features. As presented in Table 4, adding tabular alignment leaves overall AUC essentially unchanged (0.881 to 0.883) but improves scanner robustness, reducing CN FP % at 1.5 T from 33.3 to 19.8 and at 3 T from 8.7 to 4.3. Incorporating the full four-way anchor further improves discrimination (AUC =0.893Ā±0.003=0.893Ā± 0.003) and yields the lowest false-positive rates among the evaluated configurations. These results indicate that prototype alignment transfers across encoder architectures and that additional auxiliary anchors can yield complementary gains. Table 4: Backbone generalizability (n=177n=177). Prototype alignment re-run with the fully-trainable Conv5-FC3 encoder Wen et al. (2020). Mean Ā± SD across 3 seeds; bold: best per column. ā MedicalNet four-way with second stage checkpoints selected based on lowest validation loss with highest validation AUC. Model AUC 1.5T AUC 1.5T FP % 3T AUC 3T FP % Conv5-FC3 (MRI-only) Wen et al. (2020) 0.881Ā±.0090.881_Ā±.009 0.800Ā±.0070.800_Ā±.007 33.3Ā±7.733.3_Ā± 7.7 0.911Ā±.0100.911_Ā±.010 8.7Ā±1.58.7_Ā± 1.5 PANDA + Conv5-FC3 (MRI+Tab) 0.883Ā±.0040.883_Ā±.004 0.799Ā±.0110.799_Ā±.011 19.8Ā±9.219.8_Ā± 9.2 0.912Ā±.0080.912_Ā±.008 4.3Ā±0.94.3_Ā± 0.9 PANDA, MedicalNet (MRI+Tab+PET+HW)ā 0.875Ā±.0150.875_Ā±.015 0.789Ā±.0170.789_Ā±.017 24.3Ā±2.224.3_Ā± 2.2 0.899Ā±.0120.899_Ā±.012 7.2Ā±2.07.2_Ā± 2.0 PANDA + Conv5-FC3 (MRI+Tab+PET+HW)ā 0.893Ā±.0030.893_Ā±.003 0.816Ā±.0100.816_Ā±.010 18.0Ā±7.118.0_Ā± 7.1 0.916Ā±.0060.916_Ā±.006 4.0Ā±1.44.0_Ā± 1.4 6.5 Zero-Shot MCI Severity Inference The held-out MCI cohort comprises 147 subjects (EMCI: 49, MCI: 79, LMCI: 19). For anchored models, applying the prototype-distance rule to each subjectās MRI projection yields a continuous pAD(i)p_AD^(i); for the severity-head extension (Section 4.2), we additionally compute sevā_āscore(i)sev\_score^(i) via hsevh_sev. We report two complementary statistic families (Table 5). Primary analyses include a 3-group KruskalāWallis test over EMCI/MCI/LMCI and one-sided MannāWhitney tests on adjacent stage pairs. We also report ordering_intact, a monotone-median check over CN << EMCI << MCI << LMCI << AD (computed with CN and AD included), which evaluates whether MCI falls at the expected ordinal position rather than only whether sub-stages differ. Fig. 7 shows the full distributions. Table 5: Zero-shot MCI ordinal analysis under the primary-analysis hierarchy (primary: 3-group KW restricted to EMCI/MCI/LMCI, and one-sided adjacent-pair MannāWhitney; ordering_intact: full monotone chain CN<<EMCI<<MCI<<LMCI<<AD on medians, computed over all 177 test subjects plus the 147 MCI subjects). ns: pā„0.05pā„ 0.05. Method Score KW p (MCI-only) MW p EMCI<<MCI MW p MCI<<LMCI ordering_intact MRI-only baseline pADp_AD 0.00230.0023 <0.001<0.001 0.7880.788 (ns) FAIL PANDA: MRI+Tab+PET+HW pADp_AD 0.0600.060 (ns) 0.0070.007 0.6770.677 (ns) PASS PANDA + severity head pADp_AD 0.0600.060 (ns) 0.0070.007 0.6770.677 (ns) PASS PANDA + severity head sev_score 0.0880.088 (ns) 0.0100.010 0.6370.637 (ns) PASS Figure 7: Zero-shot MCI ordinal structure across the full CNā chain (not just EMCI/MCI/LMCI), for the MRI-only baseline, MRI+Tab+PET+HW (the full model), and the severity-head extension on both pADp_AD and sev_score. The connecting line through group medians makes ordering_intact visible directly: the MRI-only panel does not rise at the AD step (AD median 0.5730.573 ties the MCI median), failing the monotone chain, whereas the joint MRI+Tab+PET+HW model and both severity-head panels rise monotonically from CN to AD. The MRI-only baseline separates EMCI/MCI/LMCI on the primary 3-group test (KruskalāWallis p=0.0023p=0.0023) but fails ordering_intact: the AD median ties the MCI median (0.5730.573), yielding a non-monotone chain (CN/EMCI/MCI/LMCI medians 0.297<0.346<0.573<0.5970.297<0.346<0.573<0.597, with AD at 0.5730.573). In contrast, the four-way joint model (MRI+Tab+PET+HW; Table 2) passes ordering_intact with monotone medians 0.027<0.125<0.436<0.462<0.7200.027<0.125<0.436<0.462<0.720 across CN/EMCI/MCI/LMCI/AD, despite only weak evidence for MCI sub-stage separation (KW p=0.060p=0.060). The severity-head extension (Section 4.2) adds an explicit severity readout while leaving test AUC unchanged (0.868Ā±0.0200.868Ā± 0.020; Section 6.1). On this cohort, ordering_intact passes for both pADp_AD (KW p=0.060p=0.060) and sev_score (KW p=0.088p=0.088). Neither KW test reaches conventional significance; accordingly, we do not claim sharp separation of MCI sub-stages. The supported conclusion is that the CNā median ordering is monotone for the joint and severity-head models, whereas it is not for the MRI-only baseline. sev_score correlates strongly with pADp_AD (Pearson r=0.941r=0.941) but is not a trivial reparameterization (rā 1rā 1; Section 7). We next evaluate the generality of prototype alignment under a shift in both data domain and objective, moving from ADNI diagnosis to TCGA-Lung survival prediction. 6.6 Second Application: Cross-Domain Generalisation to TCGA-Lung Survival Table 6: TCGA-Lung generalisation. Binary 2yr OS: test n=119n=119. Cox PH: test n=171n=171. Mean Ā± SD across 3 seeds. Bold: best per column. ā RNA used at inference. Method Binary AUC Binary F1 Cox C-index WSI-only 0.593Ā±.0200.593_Ā±.020 0.389Ā±.0210.389_Ā±.021 0.460Ā±.0030.460_Ā±.003 Full Fusion (WSI+RNA)ā 0.591Ā±.0080.591_Ā±.008 0.348Ā±.0320.348_Ā±.032 0.453Ā±.0020.453_Ā±.002 Paired-only 50% (no proto. transfer) 0.579Ā±.0300.579_Ā±.030 0.338Ā±.0230.338_Ā±.023 0.527Ā±.0290.527_Ā±.029 Paired-only 25% (no proto. transfer) 0.576Ā±.0380.576_Ā±.038 0.376Ā±.0560.376_Ā±.056 0.513Ā±.0230.513_Ā±.023 PANDA 25% RNA 0.593Ā±.0260.593_Ā±.026 0.336Ā±.0780.336_Ā±.078 0.529Ā±.0180.529_Ā±.018 PANDA 50% RNA 0.612Ā±.0180.612_Ā±.018 0.380Ā±.0530.380_Ā±.053 0.519Ā±.0380.519_Ā±.038 PANDA 100% RNA 0.628Ā±.0070.628_Ā±.007 0.495Ā±.0150.495_Ā±.015 0.550Ā±.0160.550_Ā±.016 Figure 8: TCGA-Lung generalisation (3 seeds, mean Ā± std). Left: binary 2-year overall survival AUC. Right: Cox proportional-hazards C-index. PANDA (prototype alignment) at 100% RNA pairing outperforms both WSI-only and Full Fusion on both tasks without using RNA at inference; Paired-only baselines at matched RNA fractions fall below PANDA at the same fraction, isolating the gain to the prototype-transfer mechanism rather than RNA availability alone. Table 6 and Fig. 8 summarise TCGA results. The WSI-only baseline attains AUC =0.593Ā±0.020=0.593Ā± 0.020 (95 % CI [0.492, 0.689]) for binary 2-year OS and C-index =0.460Ā±0.003=0.460Ā± 0.003 (95 % CI [0.383, 0.537]) for Cox PH, i.e., near-chance survival discrimination from H&E morphology alone. Full Fusion (joint WSI + RNA ABMIL; RNA used at inference) does not improve over WSI-only on either endpoint (AUC =0.591=0.591, C-index =0.453=0.453), indicating unstable optimisation when trained directly on the partially paired pool (ā470ā 470 subjects). Restricting Stage 2 to the paired subset (Paired-only baselines) likewise matches or underperforms WSI-only on binary AUC (0.576ā0.579), indicating that discarding unpaired subjects is detrimental. In contrast, prototype-based contrastive transfer using 100% RNA pairing improves both endpoints without RNA at inference (AUC =0.628Ā±0.007=0.628Ā± 0.007, 95 % CI [0.521, 0.733]; C-index =0.550Ā±0.016=0.550Ā± 0.016, 95 % CI [0.482, 0.616]), corresponding to absolute gains of +3.5+3.5 p AUC and +9.0+9.0 C-index points vs. WSI-only. Bootstrap tests vs. WSI-only yield p=0.146p=0.146 (AUC) and p=0.059p=0.059 (C-index); with ntest=119n_test=119 (binary) and 171 (Cox), CIs remain wide and the improvements do not reach Ī±=0.05Ī±=0.05, although the direction is consistent across seeds and tasks. Matched-fraction Paired-only comparisons support a prototype-transfer effect beyond pairing rate: at 25% RNA, PANDA achieves C-index =0.529=0.529 vs. 0.513 for Paired-only 25% (+1.6+1.6 pts) with identical RNA training data. 7 Discussion Across ADNI and TCGA-Lung, prototype alignment consistently converts partially paired auxiliary modalities into improvements in the primary-modality model (better calibration and scanner robustness) without requiring auxiliary inputs at inference. The effect is mediated by geometry: each auxiliary anchor reshapes the MRI/WSI embedding space, and the three anchors studied target complementary geometric failure modes. Tabular alignment provides the most direct example of geometric regularisation. It leaves AUC approximately unchanged but improves CN recall and macro-F1 (Table 2), consistent with reduced within-class scatter around clinically defined cluster centres rather than added boundary signal. This restructuring also explains the large 1.5 T false-positive reduction (Table 3): borderline 1.5 T controls are pulled toward a clinically grounded CN anchor instead of remaining near the decision boundary. PET exhibits the complementary behaviour: it yields the largest single AUC gain but does not improve 1.5 T robustness (Tables 2, 3). A likely explanation is limited coverage: at 19% pairing, the PET prototype is estimated from ā 30 paired AD subjects per fold and is dominated by high-quality 3 T scans from later ADNI phases, providing limited constraint on the borderline 1.5 T controls that drive false positives. Increasing PET pairing density, or adding an explicit 1.5 T-robust regulariser, remains future work. This modality-specific division of labour is also reflected in baseline behaviour. Modality-dropout baselines benefit little from PET (Table 2: MRI+Tab 0.800 vs. MRI+Tab+PET 0.799), whereas prototype alignment improves substantially (0.790ā 0.851). We do not isolate the mechanism experimentally, but hypothesise that sparse-modality utilisation differs: concatenate-and-mask training exposes the fusion head to real PET only on the paired fraction (19% pairing, further reduced by dropout) and cannot propagate PET-derived structure to the unpaired majority, whereas prototype alignment estimates class-level PET geometry from paired subjects and propagates it to all MRI embeddings via the alignment loss. In addition, the dropout head is trained on mixed present/zero-filled auxiliary patterns that differ from deployment, while PANDAās classifier consumes MRI features consistently at train and test. We will test these hypotheses in future work by stratifying performance on PET-paired versus PET-unpaired test subjects. The handwriting anchor isolates the minimal requirement for cross-cohort transfer. With rHW=0r_HW=0, no ADNI subject contributes to the DARWIN prototype, yet handwriting yields the strongest scanner-stratified performance (Table 3). This indicates that transfer does not require semantic modality alignment; it requires prototype discriminability. A well-separated class geometry in one domain (handwriting kinematics) can constrain a structurally unrelated domain (3D MRI) when labels are shared. The pairing-rate ablation (Figure 6) further supports deployability: full pairing is unnecessary. The Tab+PET anchor remains within seed noise from 75% to 5% pairing, whereas tabular-only alignment degrades only when the paired pool falls below ā 95 subjects per fold, where class-mean estimates become unstable. We attribute the Tab+PET stability to prototype quality: the fully paired pool spans a broad severity spectrum, including borderline subjects that dilute class means; random subsampling more often excludes these ambiguous cases, yielding tighter and more separated prototypes (mirrored by increasing ADāCN prototype separation as pairing decreases). Under resource constraints, sampling a modest number of pairs per modality can be preferable to maximising pairing completeness. The TCGA-Lung results are consistent with the same decoupling of alignment and classification: effects are directionally consistent across seeds and tasks but underpowered at the available test sizes (Table 6). Binary 2-year survival is difficult to power because censoring and thresholding discard time-to-event information, whereas the Cox C-index (risk ranking) shows a numerically larger gain (+9.0+9.0 points; 0.460ā0.5500.460ā 0.550) that did not reach the conventional Ī±=0.05Ī±=0.05 threshold (p=0.059p=0.059; Table 6); the per-arm 95%95\% CIs overlap (WSI-only [0.383,0.537][0.383,0.537] vs. PANDA [0.482,0.616][0.482,0.616]), so we report it as a consistent directional trend rather than a clinically established effect. Full Fusion falling to or below WSI-only on Cox is consistent with early-fusion failure under small paired pools BaltruÅ”aitis et al. (2018): the ā 470 paired subjects are insufficient to learn a reliable RNA-to-survival mapping while maintaining WSI attention quality. PANDA avoids this by separating alignment (Stage 1, paired subjects) from prediction (Stage 2, all subjects), ensuring the survival head is trained on the full WSI set. Finally, the aligned representation encodes structure beyond binary objectives. In zero-shot MCI inference (Section 6.5), unimodal CN/AD separation does not reliably embed MCI along the severity axis (the ordinal ordering is not preserved), whereas prototype alignment stabilises the CNā ordering, and an explicit severity headātrained without any MCI supervisionārecovers the same monotone ordering without AUC cost. These results indicate that severity information is already present in the MRI representation and primarily requires an appropriate readout objective. Not all approaches to exploit this structure were supported. A geometry-transfer variant (matching the encoderās inter-prototype angle to auxiliary pre-computed angles) achieved the highest raw metric (0.877Ā±0.0180.877Ā± 0.018), but a four-arm ablation (real, identity-shuffled, label-shuffled correspondence, and a gradient-severed no-op) could not distinguish it from the no-op control on downstream AUC. The null result is confined to classification accuracy: the geometry pathway itself behaved as designed (values are meanĀ±s.d. across three seeds). In the real arm the inter-prototype angle converged toward the auxiliary target, with cosine moving from ā0.597Ā±0.020-0.597Ā± 0.020 to ā0.769Ā±0.019-0.769Ā± 0.019 and the geometry loss falling ā4.7Ćā4.7Ć (0.412Ā±0.035ā0.089Ā±0.0120.412Ā± 0.035ā 0.089Ā± 0.012). In the gradient-severed no-op neither moved toward the target: the cosine stayed positive and drifted away from the (negative) auxiliary targets (+0.227Ā±0.092ā+0.370Ā±0.092+0.227Ā± 0.092ā+0.370Ā± 0.092), and the geometry loss did not decrease. Reshaping this angle therefore did not, on its own, translate into an AUC gain over the no-opānot that the loss had no effect on the representation. A plausible explanation is a training-exposure artefact of the dual-loader control: at every step both arms draw an additional, larger batch through the shared MRI encoder and projection head (matching compute and BatchNorm running statistics), and only the geometry-loss coefficient is zeroed in the no-op; the shared extra exposure through BatchNorm statistics can account for most of any regularisation benefit, making AUC an insensitive readout for whether the geometry loss itself matters. We did not log per-pathway gradient norms to confirm this directly, so we report the exposure-artefact account as an inference from the controlās design rather than a measured quantity (Section 7). Limitations and future work. The evaluation is confined to ADNI for AD/CN classification; external MRI cohorts (e.g., OASIS-3 LaMontagne et al. (2019), UK Biobank Sudlow et al. (2015)) are required to establish generalisability across sites and scanners. The persistent 1.5 T/3 T gap (27 p FP % vs. 8 p at 3 T, even for the best model) indicates that prototype alignment mitigates but does not eliminate scanner-related bias; deployment in 1.5 T-dominant settings will likely require dedicated harmonisation. One contributing factor is the restricted trainability of the encoder: freezing the early backbone layers limits adaptability, while making the entire encoder trainable boosts the unimodal baseline and yields the strongest aligned performance (Section 6.4). However, backbone design itself was not systematically optimized in the main study. PET pairing at 19% is also too sparse to enforce scanner-robust alignment at 1.5 T; denser PET pairing (or amyloid PET) is a natural extension. The severity-head extension (Section 4.2) is trained on sorths_orth, available for only 265/844 training subjects, and this subset is scanner-skewed (208 at ā¤2.0ā¤ 2.0 T vs. 57 at >2.0>2.0 T). Denser and more scanner-balanced severity coverage would strengthen conclusions beyond the current single-composite result. A within-class pairwise rank loss, explored as an alternative approach to ordinal structure, did not improve sub-stage separation and degraded AD-adjacent ordering (MCIā ) in 3-seed evaluation; we therefore do not pursue it further. This loss is a RankNet-style pairwise logistic term (binary cross-entropy on the predicted severity gap s^iās^j s_i- s_j for each non-tied within-class pair, with no margin parameter), applied only to the AD sugnnore tbset (n=120n=120; the CN population showed no usable within-class severity spread), with weight Ī»wc=0.3 _wc=0.3 on sorths_orth and pairs drawn from a dedicated severity loader. Compared against the severity-head baseline (Ī»wc=0 _wc=0) using the primary adjacent-pair tests, it flipped the ordering_intact check from pass to fail on the severity readout, with the LMCI median falling below the MCI median. The geometry-transfer variant (Section 7) was indistinguishable from a no-op control in downstream AUCādespite measurably reshaping the encoderās inter-prototype geometryāand we did not log per-pathway gradient norms to establish whether the dual-loader exposure artefact or the loss itself explains this; we therefore treat it as unconfirmed and flag this outcome for others using similar dual-loader training. Finally, loss weights (Ī»tab _tab, Ī»PET _PET, Ī»HW _HW) were tuned at 100% pairing, and a systematic sweep is deferred to revision. Three directions follow directly from these findings. Multi-cohort generalisation. The core claimārelaxing full pairing to exploit larger training cohorts improves the primary-modality classifierāis demonstrated on ADNI and TCGA-Lung. The next step is evaluation across additional external AD cohorts (OASIS-3 LaMontagne et al. (2019), AIBL Ellis et al. (2009), UK Biobank Sudlow et al. (2015)) and other partially paired benchmarks, to characterise when the approach succeeds as a general property rather than a dataset-specific effect. Non-adversarial scanner harmonisation. The residual 1.5 T/3 T gap (Section 7) and the failure of distribution-matching baselines (adversarial scanner discrimination and sliced-Wasserstein alignment on PET embeddings Lee et al. (2019); Ganin et al. (2016)) motivate prototype-space alternatives. Aligning per-scanner class centroids avoids fitting a discriminator and requires only class and domain labels, not dense per-domain distribution estimates. A power analysis linking harmonisation performance to the number of paired scanner samples would quantify the additional data requirements and inform future multi-site collection. Richer anchor geometry and severity supervision. The severity head currently relies on a single label-orthogonal composite available for 265 CN/AD subjects; denser, scanner-balanced graded scores, and anchors that represent within-class heterogeneity (mixture/distributional prototypes rather than a single class mean), are promising extensions given substantial within-AD severity variation observed in our data. 8 Conclusion We introduced PANDA, a two-stage prototype-anchored alignment framework for multimodal AD classification with partial pairing. A single training procedure accommodates heterogeneous pairing densities (44.8% tabular, 18.7% PET, and 0% external handwriting) while requiring only MRI at inference. On ADNI, the full model achieves AUC =0.868=0.868 (+7.9+7.9 p vs. MRI-only) and reduces the 1.5 T CN false-positive rate by >24>24 p. The reported gains are relative to same-backbone MRI-only baselines. Conversely, a fully trainable Conv5-FC3 encoder attains AUC =0.881=0.881 without alignment, and PANDA improves it further to AUC =0.893=0.893 while approximately halving the 1.5 T false-positive rate. Thus, across backbones, prototype alignment improves the corresponding baseline, with larger AUC gains for weaker encoders and larger robustness gains once AUC saturates. A pairing-rate ablation shows that full pairing is not required for the joint anchor: 55ā10%10\% pairing matches the fully paired configuration within seed noise, implying that collecting additional modalities at low overlap can be preferable to assembling a small fully paired cohort. Cross-domain evaluation on TCGA-Lung indicates that the benefits are not neuroimaging-specific. Zero-shot analysis on held-out MCI subjects further characterises the representation: the unimodal MRI baseline does not preserve the CNā ordinal chain, whereas the prototype-aligned model maintains monotone group medians. A linear severity head, trained on graded CN/AD scores without any MCI gradients, exposes this axis explicitly with no AUC cost, indicating that severity information is present in the aligned embedding and can be read out without MCI-specific supervision. In contrast, a class-prototype geometry-transfer variant was not supported: despite the highest raw metric, a four-arm ablation (real, identity-shuffled, label-shuffled correspondence, and a gradient-severed no-op with matched training exposure) could not distinguish it from the no-op control, consistent with training-exposure confounding rather than a geometry-transfer effect. Overall, these results support partially paired prototype alignment as a deployment-oriented strategy for leveraging incomplete auxiliary data to improve accuracy and scanner robustness without auxiliary inputs at test time. Code Availability The full training, evaluation, and analysis code for PANDA will be made publicly available upon acceptance. The datasets used are accessible through their respective repositories under their own data-use agreements: ADNI (https://adni.loni.usc.edu), TCGA via the GDC Data Portal (https://portal.gdc.cancer.gov), and the DARWIN handwriting dataset Cilia et al. (2022). Acknowledgements The authors gratefully acknowledge the scientific support and HPC resources provided by the Erlangen National High Performance Computing Center (NHR@FAU) of the Friedrich-Alexander-UniversitƤt Erlangen-NĆ¼rnberg (FAU). The hardware is partially funded by the German Research Foundation (DFG). Data used in preparation of this article were obtained from the Alzheimerās Disease Neuroimaging Initiative (ADNI) database (https://adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at https://adni.loni.usc.edu/wp-content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf. Appendix A Baseline Configurations Table 7: Baselines and their configuration. āAux (train)ā: auxiliary modalities used during training only; āInferā: inputs required at test time. Unless stated, ADNI baselines train on all 844 subjects and are evaluated on the full test set (n=177n=177). Method Aux (train) Infer Key configuration / adaptation ADNI (AD/CN classification) MRI-only ā MRI MedicalNet ResNet-18, cross-entropy only MRI+Tab (pairs) Tab MRI+Tab Feature concatenation; trained on the 378 paired subjects only Suk (SAE fusion) Suk et al. (2014) Tab MRI+Tab Stacked-autoencoder late fusion; both modalities at test; paired test (n=75n=75) Knowledge distill. Hu et al. (2020) Tab+PET MRI MRI student matches MRI+Tab+PET prototype teacher (KL, T=4T=4) Modality dropout Tab / Tab+PET MRI Random branch dropout (p=0.5p=0.5); two configurations HeMIS Havaei et al. (2016) Tab+PET MRI Hetero-modal mean/variance pooling over present modalities Graph-smoothness Weston et al. (2008) ā MRI Manifold regularisation; k=10k=10 N graph, pairwise L2L_2 penalty Graph-SLC Ou et al. (2024) Tab MRI Graph-embedded latent; decoders reconstruct missing; tabular zeroed at test DiaMond Li et al. (2025) PET MRI Bi-modal MRI+PET ViT (bi-attention), run MRI-only Wang et al. Wang et al. (2024) PET MRI Diagnosis network; PET-synthesis branch omitted HyperFusion Duenias et al. (2025) Tab MRI Hypernetwork conditions MRI on tabular (age, sex, APOE) IC-MKD Kwak et al. (2025) PET MRI MRI+PET teacher ā MRI student; reimplemented (no official code) CCSW Lee et al. (2019) Tab MRI Contrastive class-conditional sliding-window alignment TCGA-Lung (survival) Full Fusion RNA WSI+RNA Joint WSI+RNA ABMIL with RNA at inference (complete-data upper bound) Paired-only RNA WSI Same architecture restricted to the paired subset (Stages 1ā2) References BaltruÅ”aitis et al. (2018) BaltruÅ”aitis, T., Ahuja, C., Morency, L.-P.: Multimodal machine learning: A survey and taxonomy. IEEE Transactions on Pattern Analysis and Machine Intelligence 41(2), 423ā443 (2018) https://doi.org/10.1109/TPAMI.2018.2798607 Zhang et al. (2011) Zhang, D., Wang, Y., Zhou, L., Yuan, H., Shen, D., Alzheimerās Disease Neuroimaging Initiative: Multimodal classification of Alzheimerās disease and mild cognitive impairment. NeuroImage 55(3), 856ā867 (2011) https://doi.org/10.1016/j.neuroimage.2011.01.008 Suk et al. (2014) Suk, H.-I., Lee, S.-W., Shen, D., Alzheimerās Disease Neuroimaging Initiative: Hierarchical feature representation and multimodal fusion with deep learning for AD/MCI diagnosis. NeuroImage 101, 569ā582 (2014) https://doi.org/10.1016/j.neuroimage.2014.06.077 Sharma and Hamarneh (2020) Sharma, A., Hamarneh, G.: Missing MRI pulse sequence synthesis using multi-modal generative adversarial network. IEEE Transactions on Medical Imaging 39(4), 1170ā1183 (2020) https://doi.org/10.1109/TMI.2019.2945521 Havaei et al. (2016) Havaei, M., Guizard, N., Chapados, N., Bengio, Y.: HeMIS: Hetero-modal image segmentation. In: Medical Image Computing and Computer-Assisted Intervention (MICCAI), p. 469ā477. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-46723-8_54 Hu et al. (2020) Hu, M., Maillard, M., Zhang, Y., Ciceri, T., La Barbera, G., Bloch, I., Gori, P.: Knowledge distillation from multi-modal to mono-modal segmentation networks. In: Medical Image Computing and Computer-Assisted Intervention (MICCAI), p. 772ā781. Springer, Cham (2020). https://doi.org/10.1007/978-3-030-59710-8_75 Ma et al. (2021) Ma, M., Ren, J., Zhao, L., Tulyakov, S., Wu, C., Peng, X.: SMIL: Multimodal learning with severely missing modality. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, p. 2302ā2310 (2021). https://doi.org/10.1609/aaai.v35i3.16330 Snell et al. (2017) Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. In: Advances in Neural Information Processing Systems (NeurIPS), vol. 30 (2017) Radford et al. (2021) Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (ICML), p. 8748ā8763 (2021). PMLR Weiner et al. (2017) Weiner, M.W., Veitch, D.P., Aisen, P.S., Beckett, L.A., Cairns, N.J., Green, R.C., Harvey, D., Jack, C.R., Jagust, W., Morris, J.C., et al.: The Alzheimerās disease neuroimaging initiative 3: Continued innovation for clinical trial improvement. Alzheimerās & Dementia 13(4), 561ā571 (2017) https://doi.org/10.1016/j.jalz.2016.10.006 Chen et al. (2019) Chen, S., Ma, K., Zheng, Y.: Med3D: Transfer learning for 3D medical image analysis. arXiv preprint arXiv:1904.00625 (2019) https://doi.org/10.48550/arXiv.1904.00625 Wen et al. (2020) Wen, J., Thibeau-Sutre, E., Diaz-Melo, M., Samper-GonzĆ”lez, J., Routier, A., Bottani, S., Dormont, D., Durrleman, S., Burgos, N., Colliot, O.: Convolutional neural networks for classification of Alzheimerās disease: Overview and reproducible evaluation. Medical Image Analysis 63, 101694 (2020) https://doi.org/10.1016/j.media.2020.101694 Ilse et al. (2018) Ilse, M., Tomczak, J., Welling, M.: Attention-based deep multiple instance learning. In: International Conference on Machine Learning (ICML), p. 2127ā2136 (2018). PMLR Chen et al. (2021) Chen, R.J., Lu, M.Y., Weng, W.-H., Chen, T.Y., Williamson, D.F.K., Manz, T., Shady, M., Mahmood, F.: Multimodal co-attention transformer for survival prediction in gigapixel whole slide images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), p. 4015ā4025 (2021). https://doi.org/10.1109/ICCV48922.2021.00398 Johnson et al. (2007) Johnson, W.E., Li, C., Rabinovic, A.: Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics 8(1), 118ā127 (2007) https://doi.org/10.1093/biostatistics/kxj037 Fortin et al. (2017) Fortin, J.-P., Parker, D., TunƧ, B., Watanabe, T., Elliott, M.A., Ruparel, K., Roalf, D.R., Satterthwaite, T.D., Gur, R.C., Gur, R.E., et al.: Harmonization of multi-site diffusion tensor imaging data. NeuroImage 161, 149ā170 (2017) https://doi.org/10.1016/j.neuroimage.2017.08.047 Glocker et al. (2019) Glocker, B., Robinson, R., Castro, D.C., Dou, Q., Konukoglu, E.: Machine learning with multi-site imaging data: An empirical study on the impact of scanner effects. In: Medical Imaging Meets NeurIPS Workshop (2019). https://doi.org/10.48550/arXiv.1910.04597 Leming et al. (2022) Leming, M., Das, S., Im, H.: Construction of a confounder-free clinical MRI dataset in the mass general brigham system for classification of Alzheimerās disease. Artificial Intelligence in Medicine 129, 102309 (2022) https://doi.org/10.1016/j.artmed.2022.102309 Song et al. (2022) Song, Y.-H., Yi, J.-Y., Noh, Y., Jang, H., Seo, S.W., Na, D.L., Seong, J.-K.: On the reliability of deep learning-based classification for Alzheimerās disease: Multi-cohorts, multi-vendors, multi-protocols, and head-to-head validation. Frontiers in Neuroscience 16, 851871 (2022) https://doi.org/10.3389/fnins.2022.851871 Bhat et al. (2026) Bhat, S., Bayer, S., PĆ©rez-Toro, P.A., Maier, A.: Exploiting partially overlapping auxiliary modalities for Alzheimerās MRI classification via prototype alignment. In: BAIOSPHERE MEDICAL 2026 ā Book of Abstracts, p. 220. FAU University Press, Erlangen, Germany (2026). Abstract. https://open.fau.de/items/2b0cedf6-52f8-4140-a159-a47b05ae5e84 Chowdhury et al. (2026) Chowdhury, M.R., Bayer, S., PĆ©rez-Toro, P.A., Maier, A., Bhat, S.: Robust deep learning for Alzheimerās MRI classification in heterogeneous ADNI acquisitions. In: BAIOSPHERE MEDICAL 2026 ā Book of Abstracts, p. 171. FAU University Press, Erlangen, Germany (2026). Abstract. https://open.fau.de/items/2b0cedf6-52f8-4140-a159-a47b05ae5e84 Ganin et al. (2016) Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., Lempitsky, V.: Domain-adversarial training of neural networks. Journal of Machine Learning Research 17(59), 1ā35 (2016) Lee et al. (2019) Lee, C.-Y., Batra, T., Baig, M.H., Ulbricht, D.: Sliced Wasserstein discrepancy for unsupervised domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 10285ā10295 (2019). https://doi.org/10.1109/CVPR.2019.01053 Huang et al. (2025) Huang, S., Zhong, L., Shi, Y.: Multistage alignment and fusion for multimodal multiclass Alzheimerās disease diagnosis. In: Medical Image Computing and Computer Assisted Intervention (MICCAI). Lecture Notes in Computer Science, vol. 15974, p. 375ā385. Springer, Cham (2025) Neverova et al. (2016) Neverova, N., Wolf, C., Taylor, G., Nebout, F.: ModDrop: Adaptive multi-modal gesture recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 38(8), 1692ā1706 (2016) https://doi.org/10.1109/TPAMI.2015.2461544 Wang et al. (2020) Wang, Q., Zhan, L., Thompson, P., Zhou, J.: Multimodal learning with incomplete modalities by knowledge distillation. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD). ACM, ??? (2020). https://doi.org/10.1145/3394486.3403234 Le et al. (2025) Le, H.Q., Thwal, C.M., Qiao, Y., Tun, Y.L., Nguyen, M.N.H., Huh, E.-N., Hong, C.S.: Cross-modal prototype based multimodal federated learning under severely missing modality. Information Fusion 122, 103219 (2025) https://doi.org/10.1016/j.inffus.2025.103219 Ou et al. (2024) Ou, Z., Jiang, C., Liu, Y., Zhang, Y., Cui, Z., Shen, D.: A graph-embedded latent space learning and clustering framework for incomplete multimodal multiclass Alzheimerās disease diagnosis. In: Medical Image Computing and Computer Assisted Intervention (MICCAI), p. 45ā55. Springer, Cham (2024) Liu et al. (2025) Liu, Y., Liu, M., Zhang, Y., Sun, K., Shen, D.: A progressive single-modality to multi-modality classification framework for Alzheimerās disease sub-type diagnosis. In: Machine Learning in Clinical Neuroimaging (MLCN), MICCAI Workshop, p. 123ā133. Springer, Cham (2025) Chen et al. (2020) Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International Conference on Machine Learning (ICML), p. 1597ā1607 (2020). PMLR Khosla et al. (2020) Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: Advances in Neural Information Processing Systems (NeurIPS), vol. 33, p. 18661ā18673 (2020) Zhang et al. (2022) Zhang, Y., Jiang, H., Miura, Y., Manning, C.D., Langlotz, C.P.: Contrastive learning of medical visual representations from paired images and text. In: Machine Learning for Healthcare Conference (MLHC), p. 2ā25 (2022). PMLR Ouyang et al. (2020) Ouyang, C., Biffi, C., Chen, C., Kart, T., Qiu, H., Rueckert, D.: Self-supervision with superpixels: Training few-shot medical image segmentation without annotation. In: European Conference on Computer Vision (ECCV), p. 762ā780. Springer, Cham (2020). https://doi.org/10.1007/978-3-030-58526-6_45 Xiong et al. (2024) Xiong, C., Chen, H., Zheng, H., Wei, D., Zheng, Y., Sung, J.J.Y., King, I.: MoME: Mixture of multimodal experts for cancer survival prediction. In: Medical Image Computing and Computer Assisted Intervention ā MICCAI 2024. Lecture Notes in Computer Science, vol. 15004. Springer, ??? (2024). https://doi.org/10.1007/978-3-031-72083-3_30 Chen et al. (2022) Chen, R.J., Lu, M.Y., Williamson, D.F.K., Chen, T.Y., Lipkova, J., Noor, Z., Shaban, M., Shady, M., Williams, M., Joo, B., Mahmood, F.: Pan-cancer integrative histology-genomic analysis via multimodal deep learning. Cancer Cell 40(8), 865ā878 (2022) https://doi.org/10.1016/j.ccell.2022.07.004 Chen et al. (2024) Chen, R.J., Ding, T., Lu, M.Y., Williamson, D.F.K., Jaume, G., Song, A.H., Chen, B., Zhang, A., Shao, D., Shaban, M., et al.: Towards a general-purpose foundation model for computational pathology. Nature Medicine 30(3), 850ā862 (2024) https://doi.org/10.1038/s41591-024-02857-3 Lu et al. (2024) Lu, M.Y., Chen, B., Williamson, D.F.K., Chen, R.J., Liang, I., Ding, T., Jaume, G., Odintsov, I., Le, L.P., Gerber, G., et al.: A visual-language foundation model for computational pathology. Nature Medicine 30(3), 863ā874 (2024) https://doi.org/10.1038/s41591-024-02856-4 Cilia et al. (2022) Cilia, N.D., De Gregorio, G., De Stefano, C., Fontanella, F., Marcelli, A., Parziale, A.: Diagnosing Alzheimerās disease from on-line handwriting: A novel dataset and performance benchmarking. Engineering Applications of Artificial Intelligence 111, 104822 (2022) https://doi.org/10.1016/j.engappai.2022.104822 Poole et al. (2019) Poole, B., Ozair, S., van den Oord, A., Alemi, A.A., Tucker, G.: On variational bounds of mutual information. In: International Conference on Machine Learning (ICML) (2019) Boudiaf et al. (2020) Boudiaf, M., Rony, J., Ziko, I.M., Granger, E., Pedersoli, M., Piantanida, P., Ben Ayed, I.: A unifying mutual information view of metric learning: Cross-entropy vs. pairwise losses. In: European Conference on Computer Vision (ECCV) (2020) Cardoso et al. (2022) Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., Murrey, B., Myronenko, A., Zhao, C., Yang, D., et al.: MONAI: An open-source framework for deep learning in healthcare. arXiv preprint arXiv:2211.02701 (2022) Weston et al. (2008) Weston, J., Ratle, F., Collobert, R.: Deep learning via semi-supervised embedding. In: Proceedings of the 25th International Conference on Machine Learning (ICML), p. 1168ā1175 (2008) Li et al. (2025) Li, Y., Ghahremani, M., Wally, Y., Wachinger, C.: DiaMond: Dementia diagnosis with multi-modal vision transformers using MRI and PET. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), p. 107ā116 (2025) Wang et al. (2024) Wang, C., Piao, S., Huang, Z., Gao, Q., Zhang, J., Li, Y., Shan, H.: Joint learning framework of cross-modal synthesis and diagnosis for Alzheimerās disease by mining underlying shared modality information. Medical Image Analysis 91, 103032 (2024) https://doi.org/10.1016/j.media.2023.103032 Duenias et al. (2025) Duenias, D., Nichyporuk, B., Arbel, T., Riklin Raviv, T.: HyperFusion: A hypernetwork approach to multimodal integration of tabular and medical imaging data for predictive modeling. Medical Image Analysis 102 (2025) https://doi.org/10.1016/j.media.2025.103503 Kwak et al. (2025) Kwak, M.G., Mao, L., Zheng, Z., Su, Y., Lure, F., Li, J.: A cross-modal mutual knowledge distillation framework for Alzheimerās disease diagnosis: Addressing incomplete modalities. IEEE Transactions on Automation Science and Engineering 22, 14218ā14233 (2025) https://doi.org/10.1109/TASE.2025.3556290 LaMontagne et al. (2019) LaMontagne, P.J., Benzinger, T.L.S., Morris, J.C., Keefe, S., Hornbeck, R., Xiong, C., Grant, E., Hassenstab, J., Moulder, K., Vlassenko, A.G., Raichle, M.E., Cruchaga, C., Marcus, D.: OASIS-3: Longitudinal neuroimaging, clinical, and cognitive dataset for normal aging and Alzheimer disease. medRxiv (2019) https://doi.org/10.1101/2019.12.13.19014902 Sudlow et al. (2015) Sudlow, C., Gallacher, J., Allen, N., Beral, V., Burton, P., Danesh, J., Downey, P., Elliott, P., Green, J., Landray, M., Liu, B., Matthews, P., Ong, G., Pell, J., Silman, A., Young, A., Sprosen, T., Peakman, T., Collins, R.: UK Biobank: An open access resource for identifying the causes of a wide range of complex diseases of middle and old age. PLoS Medicine 12(3), 1001779 (2015) https://doi.org/10.1371/journal.pmed.1001779 Ellis et al. (2009) Ellis, K.A., Bush, A.I., Darby, D., De Fazio, D., Foster, J., Hudson, P., Lautenschlager, N.T., Lenzo, N., Martins, R.N., Maruff, P., Masters, C., Milner, A., Pike, K., Rowe, C., Savage, G., Szoeke, C., Taddei, K., Villemagne, V., Woodward, M., Ames, D.: The Australian Imaging, Biomarkers and Lifestyle (AIBL) study of aging: methodology and baseline characteristics of 1112 individuals recruited for a longitudinal study of Alzheimerās disease. International Psychogeriatrics 21(4), 672ā687 (2009) https://doi.org/10.1017/S1041610209009405