Paper deep dive

OrganLens: Organ-Specific Representation Learning for CT Foundation Models

Zhixuan Ge, Anqi Li, Sadeer Al-Kindi, Hanwen Xu, Wei Qiu

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 85

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%

Last extracted: 8/1/2026, 11:21:02 AM

Summary

OrganLens is a framework for organ-specific representation learning in CT foundation models. It uses a shared encoder conditioned on organ identity, combined with self-distillation and anatomy-mask supervision, to produce 11 distinct organ-specific representations from a single CT volume without external segmentation masks. The model outperforms baselines like DINOv2 in tasks such as cardiomegaly detection and lung-cancer mortality prediction.

Entities (14)

Heart · anatomical-structure · 98%Lung · anatomical-structure · 98%OrganLens · modelframework · 98%Cardiomegaly · medical-condition · 95%DinoV2 · model · 95%Lung Cancer Mortality · medical-conditionendpoint · 95%CT-RATE · dataset · 92%INSPECT · dataset · 92%NLST · dataset · 92%RAD-ChestCT · dataset · 92%CT-CLIP · model · 90%GigaHeart · model · 90%Merlin · model · 90%SPECTRE · model · 90%

Relation Signals (12)

OrganLens → evaluatedon → INSPECT

confidence 95% · We evaluate OrganLens on ... INSPECT

OrganLens → evaluatedon → NLST

confidence 95% · We evaluate OrganLens on ... NLST

OrganLens → evaluatedon → CT-RATE

confidence 95% · We evaluate OrganLens on CT-RATE

OrganLens → evaluatedon → RAD-ChestCT

confidence 95% · We evaluate OrganLens on ... RAD-ChestCT

OrganLens → improvesmetricfor → Lung Cancer Mortality

confidence 95% · lung representations improve the Harrell C-index for NLST lung-cancer mortality by 14.2%

OrganLens → improvesmetricfor → Cardiomegaly

confidence 95% · heart representations raise CT-RATE cardiomegaly AUROC from 0.910 to 0.953

OrganLens → outperforms → DinoV2

confidence 95% · Relative to CT-pretrained DINOv2, heart representations raise CT-RATE cardiomegaly AUROC from 0.910 to 0.953

OrganLens → uses → DinoV2

confidence 95% · built on a shared CT-pretrained DINOv2 encoder

CT-CLIP → aligns →

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:A CT examination captures multiple organs, but many biomedical questions concern abnormalities, prognosis, or longitudinal change in a specific organ. These questions require a separate representation for each organ within the same CT volume. Existing CT foundation models commonly produce a single volume-level representation, while recent anatomy-aware methods either encode pre-separated organ volumes or explicitly disentangle images into organ token groups. The former may remove clinically relevant surrounding context, while the latter does not condition a shared encoder on a selected organ before its features are formed. We introduce OrganLens for organ-specific representation learning through self-supervision. An organ identity conditions a shared CT encoder, while organ-specific distillation and anatomy-mask supervision shape features for anatomy-weighted pooling into organ-specific representations. At inference, the shared model produces 11 organ-specific representations without external segmentation masks. We evaluate OrganLens on CT-RATE, RAD-ChestCT, INSPECT, and NLST across diverse acquisitions and downstream evaluations. Relative to CT-pretrained DINOv2, heart representations raise CT-RATE cardiomegaly AUROC from 0.910 to 0.953, while lung representations improve the Harrell C-index for NLST lung-cancer mortality by 14.2\%. The global representation reaches INSPECT Recall@10 of 33.09\% and 32.04\% for text-to-image and image-to-text retrieval, respectively. Across organ-related tasks, anatomically matched representations provide stronger task-relevant signal, while the global representation retains broad utility. OrganLens offers a scalable approach to organ-specific CT representation learning with a shared encoder. More broadly, it provides the medical research community with a reusable framework for studying organ-specific disease across cohorts and clinical endpoints.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

83,430 characters extracted from source content.

Expand or collapse full text

OrganLens: Organ-Specific Representation Learning for CT Foundation Models Zhixuan Ge zg33@rice.edu , Anqi Li al318@rice.edu Rice UniversityHoustonTexasUSA , Sadeer Al-Kindi Department of CardiologyDeBakey Heart and Vascular CenterHoustonTexasUSA sal-kindi@houstonmethodist.org , Hanwen Xu University of WashingtonSeattleWashingtonUSA xuhw@cs.washington.edu and Wei Qiu Rice UniversityHoustonTexasUSA wq8@rice.edu (2027) Abstract. A CT examination captures multiple organs, but many biomedical questions concern abnormalities, prognosis, or longitudinal change in a specific organ. These questions require a separate representation for each organ within the same CT volume. Existing CT foundation models commonly produce a single volume-level representation, while recent anatomy-aware methods either encode pre-separated organ volumes or explicitly disentangle images into organ token groups. The former may remove clinically relevant surrounding context, while the latter does not condition a shared encoder on a selected organ before its features are formed. We introduce OrganLens for organ-specific representation learning through self-supervision. An organ identity conditions a shared CT encoder, while organ-specific distillation and anatomy-mask supervision shape features for anatomy-weighted pooling into organ-specific representations. At inference, the shared model produces 11 organ-specific representations without external segmentation masks. We evaluate OrganLens on CT-RATE, RAD-ChestCT, INSPECT, and NLST across diverse acquisitions and downstream evaluations. Relative to CT-pretrained DINOv2, heart representations raise CT-RATE cardiomegaly AUROC from 0.910 to 0.953, while lung representations improve the Harrell C-index for NLST lung-cancer mortality by 14.2%. The global representation reaches INSPECT Recall@10 of 33.09% and 32.04% for text-to-image and image-to-text retrieval, respectively. Across organ-related tasks, anatomically matched representations provide stronger task-relevant signal, while the global representation retains broad utility. OrganLens offers a scalable approach to organ-specific CT representation learning with a shared encoder. More broadly, it provides the medical research community with a reusable framework for studying organ-specific disease across cohorts and clinical endpoints. Code and pretrained models are available at https://github.com/gezhixuan/OrganLens. CT foundation models, representation learning, medical image analysis, organ-specific modeling †copyright: none†journalyear: 2027†conference: KDD 2027 AI for Sciences Track; August 2027; San Jose, CA, USA†ccs: Computing methodologies Machine learning approaches†ccs: Applied computing Health informatics†ccs: Computing methodologies Computer vision representations 1. Introduction CT captures multiple organ systems and their spatial relationships in a single examination, supporting analyses of abnormalities, prognosis, and longitudinal change that often target a specific organ. Such analyses call for organ-specific representations, as disease processes and clinically relevant phenotypes may vary substantially across organs within the same individual. Recent CT foundation models learn transferable representations across visual and vision–language tasks (Hamamci et al., 2026; Blankemeier et al., 2026; Yang et al., 2025). However, their general-purpose representations are typically defined at the volume level rather than conditioned on a selected organ. Because a volume-level representation must integrate organs that differ in size, appearance, and relevance to a given endpoint, it dilutes organ-specific evidence. Existing anatomy-aware approaches obtain organ specificity in three ways. Encoding pre-segmented organ inputs isolates anatomical structures (Yamamoto and Kikuchi, 2026), but may discard surrounding context relevant to abnormality detection. Disentangling images into organ-wise token groups recovers organ-specific components (Song et al., 2025), but without explicit conditioning on a selected organ, the learned groups may retain irrelevant features and dilute organ-specific signal. Localizing anatomy through text-derived supervision (Lin et al., 2024; Shui et al., 2025; You et al., 2026) depends on report–region alignment, which may lack the spatial precision needed to distinguish adjacent organs. A separate encoder for each organ would also scale poorly and prevent representation sharing. The goal is therefore not to locate or isolate an organ, but to learn a shared encoder that can be conditioned on a specific organ, adaptively select which features belong to it, and produce organ-specific representations. This raises a central question: can one CT encoder adapt its representation to a selected organ while preserving surrounding anatomical context and producing anatomy-weighted representations without manual segmentation at inference? To answer this question, we introduce OrganLens, an organ-specific representation learning framework built on a shared CT-pretrained DINOv2 encoder (Oquab et al., 2024; Xu et al., 2025). OrganLens combines three components. Organ-identity conditioning adapts encoder features to each of 11 organs. Organ-guided cropping retains the selected anatomy in self-distillation views. Anatomy-mask supervision trains a spatial decoder to predict an organ-specific anatomical mask. At inference, this mask provides soft weights for pooling patch features without external segmentation masks. OrganLens also retains regions surrounding the target organ, preserving useful anatomical context. The model produces 11 organ-specific representations for downstream analyses. We evaluate these representations on CT-RATE (Hamamci et al., 2026), RAD-ChestCT (Draelos et al., 2021), INSPECT (Huang et al., 2023), and the National Lung Screening Trial (NLST) (National Lung Screening Trial Research Team, 2011). These cohorts span unenhanced CT, CT pulmonary angiography, and low-dose screening and support four task families: abnormality detection, prognostic prediction, longitudinal analysis, and image–text retrieval. Across these evaluations, OrganLens generally outperformed the competing methods, including CT-pretrained DINOv2 and GigaHeart (Xu et al., 2025). Anatomically matched representations obtained the highest macro-AUROC on both abnormality detection cohorts, demonstrating the benefit of organ-specific representation learning. For prognosis, anatomically matched representations ranked first among OrganLens representations for 16 of 17 endpoints, and OrganLens achieved the highest C-index for every endpoint across visits in the longitudinal analysis. Together, these results indicate that predictive signal was concentrated in anatomically matched organs rather than distributed uniformly across the entire volume. Beyond organ-specific analyses, the global OrganLens representation performed best in image–text retrieval on CT-RATE and INSPECT, suggesting that the combined organ-specific representations anchor image–text correspondence to individual organs rather than coarse anatomical regions. These findings suggest that OrganLens could provide a foundation for future clinical applications involving organ-specific disease characterization, risk stratification, and longitudinal monitoring. Our contributions are: • We introduce OrganLens, which learns organ-specific features while retaining surrounding anatomical context and produces 11 organ-specific representations. • We develop a self-distillation framework that combines organ-identity conditioning, organ-guided cropping, and anatomy-mask supervision to produce organ-specific representations through anatomy-weighted pooling. • We demonstrate the clinical value of learning multiple organ-specific representations from the same CT examination: anatomically matched representations provide finer-grained, task-relevant features for organ-focused downstream applications, while their global aggregation retains broad utility for whole-volume tasks. 2. Related Work Figure 1. Overview of OrganLens. (a) Organ-identity conditioning, organ-guided cropping, organ-specific distillation, and anatomy-mask supervision produce anatomy-weighted organ-specific representations. (b) The representations are reused across downstream clinical applications. 2.1. CT and Volumetric Foundation Models Foundation models provide reusable representations across endpoints (Bommasani et al., 2021). CT-specific self-supervision spans 3D image restoration in Models Genesis, hierarchical pretext tasks in Swin UNETR, unified 2D–3D training in UniMiSS, and subvolume context modeling in VoCo (Zhou et al., 2021; Tang et al., 2022; Xie et al., 2022; Wu et al., 2024; Yang et al., 2025). DINO and DINOv2 learn transferable ViT features through self-distillation, masked patch prediction, and representation regularization (Caron et al., 2021; Zhou et al., 2022; Sablayrolles et al., 2019; Oquab et al., 2024). CT-CLIP aligns CT-RATE scans and reports for zero-shot detection and retrieval (Hamamci et al., 2026). Merlin scales CT vision–language learning using a large clinical resource (Blankemeier et al., 2026), while FM-HCT transfers a self-supervised head CT encoder across detection tasks (Zhu et al., 2026). SPECTRE combines DINO-style self-supervision with report-level vision–language alignment in a hierarchical volumetric transformer (Claessens et al., 2025). SigVLP uses axial rotary position embeddings for variable-depth volumes and aligns 3D chunks with organ-wise report observations (Wang et al., 2026). These systems provide local or volume-level features and include both general and anatomy-specific models. Recent CT models include staged anatomical and semantic pretraining in FlexiCT, volumetric DINO self-distillation in CoralBay, and lung-focused pretraining in UCLIF. LCTfound uses diffusion pretraining with imaging and clinical information for lung CT (Li et al., 2026b; Gatopoulos et al., 2026; Li et al., 2026a; Gao et al., 2026). TANGERINE uses computationally efficient 3D masked-autoencoder pretraining for volumetric thoracic CT (McConnell et al., 2026). The CIPHER preprint adapts contrastive masked-autoencoder pretraining to pretreatment pneumonitis-risk prediction (Muneer et al., 2026). Across this literature, representations are not indexed by multiple organ identities applied to the same examination. 2.2. Anatomical Localization and Promptable Segmentation Anatomical segmentation provides explicit spatial priors for focused CT analysis. TotalSegmentator delineates many organs and structures in CT; MedSAM, SegVol, and ONCOPILOT provide promptable segmentation for medical images, 3D volumes, and solid tumors, respectively (Wasserthal et al., 2023; Ma et al., 2024; Du et al., 2024; Machado et al., 2025). Language-driven universal segmentation further uses semantic organ and tumor labels to parameterize a shared CT model (Liu et al., 2024). For end-to-end abdominal CT classification, ORACLE-CT uses multi-organ segmentations to constrain label-specific attention pooling over encoder features (Dahal et al., 2026). These anatomical masks support organ cropping, regional pooling, and localized supervision. These workflows leave feature formation independent of the selected identity. Cropping narrows the available context, post hoc pooling selects already-formed features, and external inference masks introduce preprocessing and segmentation domain shift. OrganLens uses masks during pretraining and predicts its own spatial mask at inference. 2.3. Anatomy-Aware and Multi-Organ Representation Learning Anatomical priors can also shape the representation itself. Adam-v2 (Taher et al., 2024) models part–whole relations through objectives for localizability, composability, and decomposability. DrasCLR organizes lung CT features by anatomical location (Yu et al., 2024), GigaHeart specializes features for heart-transplantation assessment (Xu et al., 2025), and BoneCoT learns across skeletal sites and bone-metastasis tasks (Zhao et al., 2026). These methods cover relational, location-aware, cardiac, and skeletal representations with fixed anatomical scopes or representation schemes. Beyond CT-specific models, Oh et al. condition ultrasound transformer features on anatomical context, while AG-SSD combines anatomy-consistent cropping and self-distillation for CT and MRI (Oh et al., 2026; Yu and Dai, 2026). Both show that anatomical context can guide feature formation and self-distillation across imaging modalities. Coda models variation and consistency across 23 radiographic regions (Hosseinzadeh Taher et al., 2025). OWT disentangles CT or MRI into composable organ-token groups, whereas TotalFM aligns segmented organ volumes with report sentences (Song et al., 2025; Yamamoto and Kikuchi, 2026). Pan-FM jointly pretrains across seven organ and tissue MRI inputs under missing-organ settings (Wu et al., 2026). Their anatomical structure arises from cross-region training, token-group disentanglement, separated inputs, or multi-organ fusion. OrganLens addresses a different setting: one shared encoder maps the same CT examination to representations indexed by a selected organ identity. 2.4. Localized and Query-Based Vision–Language Learning Medical vision–language models localize features through semantic queries. GLoRIA aligns report words with image regions (Huang et al., 2021), MedKLIP associates medical entities and knowledge descriptions with spatial regions (Wu et al., 2023), and KAD conditions radiograph models on disease queries (Zhang et al., 2023). CT-GLIP constructs organ-level image–text pairs, fVLM aligns anatomical regions with report descriptions, MG-3D combines global alignment with local reconstruction, and OKA-CT derives organ-hierarchical report knowledge for organ-specific supervision (Lin et al., 2024; Shui et al., 2025; Ni et al., 2026; You et al., 2026). EXACT jointly learns organ segmentation and report-supervised anomaly localization, producing disease-specific voxel maps constrained to the corresponding anatomy (Bai et al., 2026). Their objectives center on text alignment, zero-shot diagnosis, retrieval, localization, or anomaly mapping. Taken together, prior methods obtain anatomical specificity from masks, isolated inputs, fixed-domain encoders, token decomposition, multi-organ fusion, or report queries. OrganLens uses a selected organ to condition a shared image-only encoder while the full slice remains visible. A predicted spatial mask provides weights for feature pooling without requiring external masks at inference. 3. Method 3.1. Overview and Problem Formulation Let V=xss=1SV=\x_s\_s=1^S denote a CT volume represented at inference by S uniformly sampled slices that retain surrounding anatomical context, and let o∈o denote an organ identity. We seek a shared foundation model that maps the same CT volume to an organ-specific representation (V,o)z(V,o). During pretraining, each CT slice x is paired with organ identity o and its binary mask yoy^o. We initialize the shared encoder from a DINOv2 vision transformer (Oquab et al., 2024) pretrained on CT images. As shown in Figure 1a, OrganLens learns organ-specific representations through three key components. Organ-identity conditioning adds a learned organ embedding to the CLS token, adapting encoder features to the selected organ. Organ-guided crop sampling constructs global and local views that retain the selected anatomy for organ-conditioned student–teacher distillation. Anatomy-mask supervision trains an auxiliary decoder to predict a spatial mask for the selected organ, which weights patch features within each slice and slice representations across the CT volume. The ground-truth mask yoy^o is used only for crop sampling and spatial supervision and is not passed to the encoder. At inference, only the sampled CT slices and organ identity are required. Figure 1b summarizes the downstream applications. 3.2. Organ-Identity Conditioning The “organ bank” and “tokenizer” in Figure 1a denote a learned embedding lookup that maps a discrete organ identity to an additive CLS offset. Let oe_o be the learned embedding for organ o, α a trainable scalar, and c the learned CLS token. For input patch embeddings jj=1N\u_j\_j=1^N, conditioning and encoding are written together as (1) [clso,(jo)j=1N]=fθ([+αo,(j)j=1N]).[h_cls^o,(h_j^o)_j=1^N]=f_θ([c+ _o,(u_j)_j=1^N]). Initializing α to zero preserves the CT-pretrained model at the first forward pass; subsequent optimization learns the strength of the organ offset. The “organ token” in Figure 1 denotes an additive offset rather than an extra sequence token, leaving the backbone architecture unchanged. Every augmented view and both networks receive the same organ identity, so student and teacher representations are compared under identical conditioning. 3.3. Organ-Specific Distillation Stages 2 and 4 of Figure 1a adapt DINOv2 student–teacher training (Oquab et al., 2024) to the organ identity. Each crop is constructed jointly from a CT slice and the mask corresponding to its organ identity. Geometric transformations are applied to both, whereas intensity augmentation affects only the image. Candidate global and local crops must retain foreground from the anatomy specified by the organ identity. This organ-guided cropping ensures that both student and teacher views contain the anatomy specified by the identity. The student processes global and local crops; the teacher receives only global crops. The teacher parameters are updated as an exponential moving average of the corresponding student parameters, including the backbone, organ conditioner, projection heads, and anatomy decoder. Gradients propagate only through the student. Following DINOv2, we combine image-level DINO distillation (Caron et al., 2021), patch-level iBOT prediction (Zhou et al., 2022), and KoLeo regularization (Sablayrolles et al., 2019; Oquab et al., 2024). Let tg(o)q_t^g(o) and sv(o)p_s^v(o) denote teacher and student prototype distributions from global view g and student view v under the same organ identity. For a masked global view g, let ℳgM_g denote its masked patch positions, and let k index the K prototypes. The image- and patch-level losses are (2) ℒDINO _DINO =−∑k=1Kqt,kg(o)log⁡ps,kv(o), =- _k=1^Kq_t,k^g(o) p_s,k^v(o), ℒiBOT _iBOT =−1|ℳg|∑j∈ℳg∑k=1Kqt,j,kg(o)log⁡ps,j,kg(o). =- 1|M_g| _j _g _k=1^Kq_t,j,k^g(o) p_s,j,k^g(o). The DINO loss is averaged over each teacher global view and all student views other than the matching global view. The iBOT loss is averaged over masked global views. KoLeo acts on normalized student CLS features and discourages representation collapse. Temperature scaling, teacher centering, and loss aggregation otherwise follow DINOv2. 3.4. Anatomy-Mask Supervision Self-distillation alone does not explicitly encourage patch features to localize the queried organ. The anatomy-mask supervision stage in Figure 1a addresses this by predicting organ-specific spatial mask logits from student patch features for each global crop. The patch sequence is reshaped to its two-dimensional grid and processed with progressive convolution and bilinear upsampling. For sample i and global view g, let yigoy_ig^o be the aligned binary anatomical target, ℓigo _ig^o the predicted logits, and πigo=σ(ℓigo) _ig^o=σ( _ig^o) the predicted probability map. We combine pixel-wise binary cross-entropy with a differentiable Dice-based loss (Milletari et al., 2016) to address foreground imbalance. For notational brevity, let π=πigoπ= _ig^o and y=yigoy=y_ig^o within one sample–view pair, and let P denote the number of pixels. The component losses and their aggregation are (3) ℒBCE(π,y)=−1P∑r=1P[yrlog⁡πr+(1−yr)log⁡(1−πr)],ℒDice(π,y)=1−2∑r=1Pπryr+ϵD∑r=1Pπr+∑r=1Pyr+ϵD,ℒanatomy=meani,g[ℒBCE(πigo,yigo)+ℒDice(πigo,yigo)]. array[]@l@\;=\;l@L_BCE(π,y)& - 1P _r=1^P [y_r _r+(1-y_r) (1- _r) ],\\ L_Dice(π,y)& 1- 2 _r=1^P _ry_r+ _D _r=1^P _r+ _r=1^Py_r+ _D,\\ L_anatomy& i,gmean [L_BCE( _ig^o,y_ig^o)+L_Dice( _ig^o,y_ig^o) ]. array For numerical stability, BCE is evaluated directly from logits. Dice is computed per sample and global view and then averaged, with ϵD=10−6 _D=10^-6. The DINO, iBOT, and KoLeo terms follow DINOv2, giving the complete pretraining objective (4) ℒ=ℒDINO+ℒiBOT+λKoLeoℒKoLeo+ℒanatomy.L=L_DINO+L_iBOT+ _KoLeoL_KoLeo+L_anatomy. We set λKoLeo=0.1 _KoLeo=0.1. Backpropagation through the student’s anatomy decoder shapes the patch features during pretraining. At inference, the teacher’s anatomy decoder supplies the soft weights used for anatomy-weighted pooling. Figure 2. Organ-specific abnormality detection on CT-RATE and RAD-ChestCT. (a) AUROC ranks of the 11 organ-specific representations for each CT-RATE abnormality, where rank 1 indicates the best-performing organ representation. Black outlines mark the predefined anatomically matched organ for each abnormality. (b) AUROC by abnormality on CT-RATE and (c) RAD-ChestCT, comparing the anatomically matched OrganLens representation with baseline models using volume-level representations. Label-outline colors indicate the organ matched to each abnormality. Panel a shows a heatmap of the ranks of 11 organ-specific representations across 18 CT-RATE abnormality labels, with outlined cells marking the anatomically matched representations. Panels b and c show radar plots comparing anatomically matched OrganLens representations with CT-pretrained DINOv2, GigaHeart, and Merlin on CT-RATE and RAD-ChestCT. Figure 3. Organ-specific prognostic evaluation on INSPECT and NLST. (a,b) Test Harrell C-indices for OrganLens, CT-pretrained DINOv2, GigaHeart, Merlin, and SPECTRE. OrganLens results use the anatomically matched representation for each endpoint. Error bars are 95% confidence intervals from 1,000 patient-level bootstrap resamples. (c) Kaplan–Meier curves compare the lowest and highest quartiles of held-out Cox log-hazard scores, excluding the middle 50%. Figure 4. Longitudinal NLST assessment using OrganLens. (a–d) Mean anatomically matched Cox log-hazard scores at T0–T2 among participants who experienced four mortality endpoints. (e–h) Serial coronal CT images and 5-year lung-cancer-mortality estimates for a case and comparison participant. The red marker denotes a radiologist-identified T2 abnormality. Panels a through d show increasing mean Cox log-hazard scores across three NLST screening time points for four cardiopulmonary mortality endpoints. Panels e and f show serial CT images and increasing model-derived 5-year lung-cancer-mortality estimates for a case, with a radiologist-marked abnormality at T2. Panels g and h show a comparison participant with stable low estimates. Figure 5. Image–text retrieval on CT-RATE (a,b) and INSPECT (c,d). Bars show Recall@K with 95% confidence intervals from 1,000 bootstrap resamples with candidate sets held fixed. Four grouped bar plots compare OrganLens, CT-pretrained DINOv2, GigaHeart, Merlin, CT-CLIP, and SPECTRE for image-to-text and text-to-image retrieval on CT-RATE and INSPECT. 3.5. Organ-Specific Representation At inference, each slice xsx_s is encoded under organ identity o. The encoder returns an organ-specific CLS feature s,clsoh_s,cls^o and patch features sjoj=1N\h_sj^o\_j=1^N. The anatomy decoder produces spatial mask logits from the patch features. The CLS feature carries the organ-identity conditioning and supports image-level self-distillation and KoLeo regularization during pretraining. The final organ-specific representation is constructed from spatially pooled patch features. We apply a sigmoid to the logits and average-pool the probabilities to the patch grid, obtaining soft weights msjo∈[0,1]m_sj^o∈[0,1]. For every sampled slice, the slice representation and predicted organ area are (5) so=∑j=1Nmsjosjo∑j=1Nmsjo,aso=1N∑j=1Nmsjo.z_s^o= _j=1^Nm_sj^oh_sj^o _j=1^Nm_sj^o, a_s^o= 1N _j=1^Nm_sj^o. We retain all S sampled slices and aggregate their representations in proportion to predicted organ area: (6) (V,o)=∑s=1Sasoso∑s=1Saso.z(V,o)= _s=1^Sa_s^oz_s^o _s=1^Sa_s^o. Applying this procedure to each organ identity produces one organ-specific representation from the same CT volume using shared model parameters. Downstream tasks may use any organ-specific representation or the global representation defined below. For tasks requiring whole-volume information, let =o1,…,o11O=\o_1,…,o_11\ denote the 11 organ identities used by OrganLens: heart, lung, aorta, liver, spleen, kidneys, pancreas, stomach, intestine, esophagus, and trachea. We define the global representation by concatenating all 11 organ-specific representations: (7) global(V)=(V,o1)‖⋯‖(V,o11),z_global(V)=z(V,o_1)\,\|\,·s\,\|\,z(V,o_11), where ∥\| denotes feature concatenation. This construction adds no pretraining module. The downstream head operates on the concatenated features. The retrieval experiments use all 11 organ identities so that CT-RATE and INSPECT use the same image representation. 4. Data and Experimental Setup 4.1. Study Design and Cohorts For OrganLens, CT-RATE is the sole adaptation cohort and also supports internal evaluation. RAD-ChestCT, INSPECT, and NLST are used only for downstream transfer. Across the four cohorts, abnormality labels, clinical outcomes, serial scans, and reports support the four evaluation task families. CT-RATE. CT-RATE includes data from 21,304 patients and a total of 50,188 non-contrast 3D chest CT volumes, including alternative reconstructions. The images support pretraining, while radiology reports, multi-abnormality labels, and acquisition metadata support abnormality detection and image–text retrieval (Hamamci et al., 2026). RAD-ChestCT. The public RAD-ChestCT release used here contains 3,630 scans with 23 abnormality labels automatically extracted from reports (Draelos et al., 2021, 2022). It contains 2,286 training, 984 validation, and 360 held-out test volumes and evaluates supervised cross-cohort transfer rather than zero-shot generalization. INSPECT. INSPECT contains 23,248 CT pulmonary angiography studies from 19,402 patients, with 3D images, report impressions, longitudinal electronic health records, and diagnostic and prognostic labels (Huang et al., 2023). It tests transfer beyond non-contrast CT. Paired reports and clinical outcomes support retrieval and prognostic prediction, respectively. NLST. NLST enrolled 53,454 participants at high risk for lung cancer and offered a baseline plus two annual low-dose CT or chest-radiography screenings (National Lung Screening Trial Research Team, 2011). We use serial CT examinations and linked cancer and mortality records for time-to-event prediction and longitudinal analysis. 4.2. Preprocessing and Organ Definitions Following the CT-CLIP-style pipeline (Hamamci et al., 2026), images are resampled to 11\,m isotropic spacing, clamped to [−1000,1000][-1000,1000] HU, and scaled to [−1,1][-1,1]. Pretraining uses aligned axial, coronal, and sagittal image–mask pairs. Following GigaHeart (Xu et al., 2025), we represent each downstream volume with S=64S=64 uniformly sampled axial slices, each center cropped to 224×224224× 224. Supplementary Section A.2 provides the interpolation, crop–padding, channel-conversion, and normalization details. Each pretraining unit contains a CT slice, an organ identity, and its binary mask. We apply TotalSegmentator (Wasserthal et al., 2023) to CT-RATE and use the resulting pseudo-label masks to define 11 targets: heart, lung, aorta, liver, spleen, kidneys, pancreas, stomach, intestine, esophagus, and trachea. Bilateral kidneys, lung lobes, and intestinal substructures are merged before target selection. 4.3. Pretraining and Implementation Details We initialize the student and teacher ViT-L/16 backbones (Dosovitskiy et al., 2021) from the CT-pretrained checkpoint released with GigaHeart (Xu et al., 2025), which was trained on an external chest CT collection using the DINOv2 framework (Oquab et al., 2024). Each sample produces two 224×224224× 224 global crops and eight 96×9696× 96 local crops using organ-guided sampling. Pretraining uses AdamW for three complete passes and 525,000 optimization iterations with a global batch size of 1,024 across eight GPUs and a base learning rate of 2×10−42× 10^-4. Supplementary Section A.3 provides the crop-acceptance and fallback rules, iBOT masking settings, optimization schedules, decoder configuration, and numerical details. Generative AI usage is described in Section 8. 5. Downstream Evaluation and Results Supplementary Section A.4 provides downstream architectures, optimization settings, and model-selection procedures. Across downstream tasks, we compare OrganLens with CT-pretrained DINOv2, GigaHeart (Xu et al., 2025), Merlin (Blankemeier et al., 2026), and SPECTRE (Claessens et al., 2025). Image–text retrieval additionally includes CT-CLIP (Hamamci et al., 2026). Available baselines follow the same task-specific data splits and downstream protocols. SPECTRE’s pretraining data included the CT-RATE training split, INSPECT, and NLST. Its evaluations on these cohorts are therefore pretraining-exposed rather than fully external transfer (Claessens et al., 2025). For analyses requiring a global OrganLens representation, the main results use concatenation. Corresponding mean aggregation results are reported in the supplementary tables. 5.1. Abnormality Detection We evaluated abnormality detection across 16 CT-RATE abnormalities and 22 RAD-ChestCT abnormalities. To assess the quality of the learned representations, we extracted representations using frozen encoders and trained a separate lightweight one-hidden-layer MLP probe for each abnormality. The CT-RATE probes were evaluated on the validation split, whereas RAD-ChestCT performance was reported on the held-out test split. For nearly all abnormalities, the anatomically matched representation achieved the highest performance among the OrganLens representations (Figure 2a). Supplementary Table S1 reports the complete matrices. Across 16 CT-RATE abnormalities, anatomically matched OrganLens representations achieved a macro-AUROC of 0.856 (Figure 2b), compared with 0.788, 0.792, 0.780, and 0.836 for CT-pretrained DINOv2, GigaHeart, Merlin, and SPECTRE, respectively. OrganLens achieved the highest AUROC for 12 of the 16 abnormalities. For cardiomegaly and pericardial effusion, the heart representation reached AUROCs of 0.953 and 0.925, outperforming the strongest baselines, GigaHeart at 0.917 and SPECTRE at 0.860, respectively. The esophagus representation reached 0.849 for hiatal hernia, compared with 0.750 for SPECTRE, while the lung representation reached 0.906 for mosaic attenuation pattern, compared with 0.853 for CT-pretrained DINOv2. On the external RAD-ChestCT dataset, anatomically matched OrganLens representations achieved a macro-AUROC of 0.767 (Figure 2c), compared with 0.684, 0.694, 0.688, and 0.721 for CT-pretrained DINOv2, GigaHeart, Merlin, and SPECTRE, respectively. OrganLens achieved the highest AUROC for 19 of 22 abnormalities. Its advantages were preserved for cardiomegaly, reaching 0.946 compared with 0.903 for GigaHeart, and for pericardial effusion, reaching 0.791 compared with 0.714 for Merlin. OrganLens performed strongly on abnormalities not included in CT-RATE evaluation, including interstitial lung disease, for which the lung representation reached 0.842 compared with 0.820 for SPECTRE. These results support the cross-cohort generalizability of organ-specific representations. 5.2. Prognostic Prediction We assessed prognostic prediction across eight INSPECT endpoints and nine NLST endpoints. To assess the quality of the learned representations, we extracted representations using frozen encoders and trained a lightweight MLP Cox model for each endpoint. INSPECT used its predefined splits, whereas NLST used a patient-level 70%/10%/20% split. Cardiovascular endpoints used the heart representation, pulmonary and lung-cancer endpoints used lung representation, and broader outcomes used global representations. Supplementary Table S2 reports all results. Across eight INSPECT endpoints, the anatomically matched representation achieved the highest C-index among all OrganLens representations for seven endpoints and a mean C-index of 0.741 (Figure 3a), compared with 0.715 for CT-pretrained DINOv2, 0.716 for GigaHeart, 0.697 for Merlin, and 0.721 for SPECTRE. The largest gains over the strongest endpoint-specific baseline were 0.037 for cardiomegaly, 0.026 for pulmonary hypertension, and 0.015 for atelectasis. Across nine NLST endpoints, matched OrganLens representations achieved the highest C-index for seven and a mean C-index of 0.789 (Figure 3b). CT-pretrained DINOv2, GigaHeart, Merlin, and SPECTRE averaged 0.686, 0.729, 0.731, and 0.772. SPECTRE was pretrained on the full NLST cohort (Claessens et al., 2025). For lung-cancer, respiratory, and COPD mortality, the lung representation reached C-indices of 0.780, 0.935, and 0.960, corresponding to relative improvements of 14.2%, 8.7%, and 6.9% over CT-pretrained DINOv2. Kaplan–Meier curves compared the lowest and highest risk quartiles (Figure 3c), with log-rank p values reported in each panel. Across the displayed endpoints, the high-risk quartiles showed consistently lower event-free survival, supporting the prognostic stratification captured by anatomically matched representations. These curves assess stratification, not calibration. 5.3. Longitudinal Analysis The longitudinal analysis used serial CT examinations and linked mortality follow-up from NLST. Participants with fewer than three complete screening time points formed the training and validation cohorts, which used all of their available scans. Participants with CT examinations available at all three time points (T0, T1, and T2) comprised the longitudinal test cohort. At each test visit, we used the scan from that visit to estimate mortality risk over the following five years. Each visit was evaluated independently, and the Cox risk score was converted to an absolute risk using the baseline hazard estimated from the training cohort. Across the common longitudinal test cohort, OrganLens achieved the strongest overall C-index performance among all evaluated models across the T0, T1, and T2 screening visits. Comparisons for each endpoint across the T0, T1, and T2 visits are provided in Supplementary Figure S1. Among participants who later experienced each endpoint, mean anatomically matched Cox scores increased monotonically from T0 to T2 (Figure 4a–d). Heart representation scores at T0, T1, and T2 were 0.928, 1.013, and 1.200 for ischemic-heart-disease mortality (n=150n=150), and 0.507, 0.589, and 0.685 for cardiovascular mortality (n=278n=278). Lung representation scores were 2.006, 2.358, and 2.684 for respiratory mortality (n=120n=120), and 2.944, 3.745, and 4.600 for COPD mortality (n=79n=79). The four endpoint-specific trajectories were not on a common score scale. For a participant who died of lung cancer on day 1,590, model-derived 5-year mortality estimates were 9.9%, 10.7%, and 39.0% on days 0, 374, and 977 (Figure 4e,f). A radiologist-marked T2 abnormality coincided with the sharp final-visit risk increase. For comparison, a participant who did not die from lung cancer had estimated risks of 0.16%, 0.17%, and 0.17% on days 31, 361, and 747 (Figure 4g,h). Each estimate used the scan from the corresponding visit and was not accumulated from T0. 5.4. Image–Text Retrieval Retrieval used the scan-level global representation defined in Section 3.5. For every model evaluated, the image encoder, ClinicalBERT text encoder (Huang et al., 2019), and projection heads were jointly fine-tuned for five epochs using the symmetric CLIP objective (Radford et al., 2021). CT-RATE text inputs combined the report findings and impression, whereas INSPECT text inputs used the report impression. We evaluated Recall@K in both directions with one paired item per identity. The official evaluation splits contained 1,564 CT-RATE and 3,212 INSPECT image–text pairs. All methods used the same fixed candidate set within each cohort. OrganLens outperformed all evaluated baselines at every K in both text-to-image and image-to-text retrieval on both cohorts (Figure 5). On CT-RATE, text-to-image and image-to-text Recall@5 values were 36.45% and 37.08%, while Recall@10 values were 46.55% and 46.42%. On INSPECT, the corresponding values were 23.88% and 23.19% at Recall@5 and 33.09% and 32.04% at Recall@10. 5.5. Ablation Study Figure 6. Pretraining-component and representation-pooling ablations on (a) CT-RATE and (b) RAD-ChestCT. Points show macro-AUROC for each variant. OrganLens (Full), highlighted in green, is the common reference for both ablations. Two horizontal lollipop plots show macro-AUROC for three pretraining-component ablations and four representation-pooling variants on CT-RATE and RAD-ChestCT. OrganLens Full is highlighted in green. We ablated organ-guided cropping and anatomy-mask supervision in a two-by-two design. Variants without anatomy-mask supervision used mean CLS pooling, whereas variants with it used anatomy-weighted pooling. Removing both components yielded macro-AUROCs of 0.798 on CT-RATE and 0.711 on RAD-ChestCT. Removing anatomy-mask supervision yielded 0.809 and 0.723. Removing organ-guided crops yielded 0.853 and 0.761. OrganLens (Full) reached 0.856 and 0.767 (Figure 6a,b). Anatomy-mask supervision accounted for the larger gain, with a smaller contribution from cropping. We next ablated the two anatomy-weighted pooling stages while retaining the organ-specific encoder. Mean CLS pooling averaged organ-specific slice CLS features, while uniform patch pooling averaged unweighted patch features. The variant without inter-slice weighting retained predicted within-slice spatial mask but averaged slices uniformly, whereas OrganLens (Full) weighted slices by predicted organ area. Complete label-wise results for matched, mean, and concatenated representations are reported in Supplementary Tables S4 and S5. Mean CLS pooling achieved macro-AUROCs of 0.802 on CT-RATE and 0.712 on RAD-ChestCT. Uniform patch pooling increased them to 0.832 and 0.722. Removing inter-slice weighting yielded 0.851 and 0.748. OrganLens (Full) reached 0.856 and 0.767 (Figure 6a,b), improving 12 of 16 CT-RATE abnormalities, with two ties, and 19 of 22 RAD-ChestCT abnormalities relative to uniform slice weighting. Overall, applying predicted spatial mask within slices and weighting slices by predicted organ area both contributed, with larger gains from the former on CT-RATE and from both stages on RAD-ChestCT. 6. Conclusion OrganLens addresses the mismatch between volume-level CT representations and organ-focused biomedical questions through a shared encoder conditioned on an organ identity. Anatomy-mask supervision shapes patch features for pooling with a predicted spatial mask, removing the need for external segmentation masks at inference. Across diverse cohorts and task families, anatomically matched representations generally improved abnormality detection, prognostic prediction, and longitudinal risk assessment over volume-level baselines. The global representation preserved broader transfer and achieved the strongest image–text retrieval performance on both evaluated cohorts. Consistent cross-cohort performance further supports the generalizability of OrganLens across acquisition settings and downstream tasks. OrganLens therefore offers a scalable framework for studying organ-specific phenotypes, longitudinal change, and clinical risk across CT cohorts. 7. Limitations and Ethical Considerations Although OrganLens was retrospectively evaluated across four chest CT cohorts, OrganLens adaptation was limited to CT-RATE and may therefore reflect cohort-specific biases. Its prospective performance and transfer beyond chest CT remain untested. TotalSegmentator pseudo-label errors may propagate into the predicted spatial mask. Two-dimensional encoding of 64 sampled slices does not model volumetric continuity and may miss small or sparsely sampled findings. Reported discrimination and retrieval metrics do not establish clinical benefit. The prognostic models were not assessed for calibration, and longitudinal predictions were repeated cross-sectional estimates rather than dynamic updates. Data were deidentified and accessed under the dataset agreements, and no reidentification was attempted. OrganLens should not guide care without prospective validation, uncertainty assessment, and clinician oversight. 8. Generative AI Usage Generative AI tools assisted with experimental code development and debugging, data analysis, figure and table preparation, literature organization, and manuscript editing. All AI-assisted outputs were reviewed and verified by the authors, who retain full responsibility for the work. References X. Bai, M. Liu, T. Song, Y. Chen, H. Yang, K. Anmahapong, Z. Li, Y. Zhou, and Q. Tian (2026) EXACT: an explainable anomaly-aware vision foundation model for analysis of 3D chest CT. arXiv preprint arXiv:2604.24146. Cited by: §2.4. L. Blankemeier, A. Kumar, J. P. Cohen, J. Liu, L. Liu, D. Van Veen, S. J. S. Gardezi, H. Yu, M. Paschali, Z. Chen, et al. (2026) Merlin: a computed tomography vision–language foundation model and dataset. Nature 652 (8112), p. 1318–1328. Cited by: §1, §2.1, §5. R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskill, et al. (2021) On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258. Cited by: §2.1. M. Caron, H. Touvron, I. Misra, H. Jégou, J. Mairal, P. Bojanowski, and A. Joulin (2021) Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 9650–9660. External Links: Link Cited by: §2.1, §3.3. C. Claessens, C. Viviers, G. D’Amicantonio, E. Bondarev, and F. van der Sommen (2025) Scaling self-supervised and cross-modal pretraining for volumetric CT transformers. arXiv preprint arXiv:2511.17209. Cited by: §2.1, §5.2, §5. L. Dahal, Y. Bhandari, G. Rubin, and J. Y. Lo (2026) ORACLE-CT: anatomy-aware support pooling for CT classification. arXiv preprint arXiv:2606.05460. External Links: Document, Link Cited by: §2.2. A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby (2021) An image is worth 16x16 words: transformers for image recognition at scale. In International Conference on Learning Representations, External Links: Link Cited by: §4.3. R. L. Draelos, D. Dov, M. A. Mazurowski, J. Y. Lo, R. Henao, G. D. Rubin, and L. Carin (2021) Machine-learning-based multiple abnormality prediction with large-scale chest computed tomography volumes. Medical image analysis 67, p. 101857. Cited by: §1, §4.1. R. L. Draelos, D. Dov, M. A. Mazurowski, J. Y. Lo, R. Henao, G. D. Rubin, and L. Carin (2022) RAD-ChestCT: radiology reports and CT imaging dataset. Zenodo. External Links: Document, Link Cited by: §4.1. Y. Du, F. Bai, T. Huang, and B. Zhao (2024) SegVol: universal and interactive volumetric medical image segmentation. Advances in Neural Information Processing Systems 37, p. 110746–110783. Cited by: §2.2. Z. Gao, G. Zhang, H. Liang, et al. (2026) A lung CT vision foundation model facilitating disease diagnosis and medical imaging. Nature Communications 17 (1), p. 35. External Links: Document, Link Cited by: §2.1. I. Gatopoulos, N. Känzig, S. Otálora, and F. Tang (2026) CoralBay: a self-supervised CT foundation model. arXiv preprint arXiv:2606.03888. External Links: Link Cited by: §2.1. I. E. Hamamci, S. Er, C. Wang, F. Almas, A. G. Simsek, S. N. Esirgun, I. Dogan, O. F. Durugol, B. Hou, S. Shit, et al. (2026) Generalist foundation models from a multimodal dataset for 3D computed tomography. Nature Biomedical Engineering. External Links: Document, Link Cited by: §1, §1, §2.1, §4.1, §4.2, §5. M. R. Hosseinzadeh Taher, J. Hong, R. Soni, and G. Avinash (2025) Learning foundation models from multi-organ medical images by capturing consistency and diversity of anatomical structures. In International Conference on Medical Image Computing and Computer-Assisted Intervention, p. 294–304. Cited by: §2.3. K. Huang, J. Altosaar, and R. Ranganath (2019) ClinicalBERT: modeling clinical notes and predicting hospital readmission. arXiv:1904.05342. Cited by: §5.4. S. Huang, Z. Huo, E. Steinberg, C. Chiang, C. Langlotz, M. Lungren, S. Yeung, N. Shah, and J. Fries (2023) INSPECT: a multimodal dataset for patient outcome prediction of pulmonary embolisms. In Advances in Neural Information Processing Systems, Vol. 36. External Links: Link Cited by: §1, §4.1. S. Huang, L. Shen, M. P. Lungren, and S. Yeung (2021) GLoRIA: a multimodal global-local representation learning framework for label-efficient medical image recognition. In Proceedings of the IEEE/CVF international conference on computer vision, p. 3942–3951. Cited by: §2.4. J. Li, Y. Xing, X. Gao, Z. Ye, M. Wang, and F. Song (2026a) A self-supervised foundation model based on three-dimensional chest CT scans for lung cancer diagnosis and prognosis prediction. Radiology: Imaging Cancer 8 (2), p. e250360. External Links: Document, Link Cited by: §2.1. Y. Li, Y. Gao, H. Dong, Y. Lai, S. Wang, M. Safari, J. E. Baciak, and X. Yang (2026b) Universal CT representations from anatomy to disease phenotype through agglomerative pretraining. arXiv preprint arXiv:2605.21906. External Links: Link Cited by: §2.1. J. Lin, Y. Xia, J. Zhang, K. Yan, L. Lu, J. Luo, and L. Zhang (2024) CT-GLIP: 3d grounded language–image pretraining with CT scans and radiology reports for full-body scenarios. arXiv preprint arXiv:2404.15272. Cited by: §1, §2.4. J. Liu, Y. Zhang, K. Wang, M. C. Yavuz, X. Chen, Y. Yuan, H. Li, Y. Yang, A. Yuille, Y. Tang, and Z. Zhou (2024) Universal and extensible language–vision models for organ segmentation and tumor detection from abdominal computed tomography. Medical Image Analysis 97, p. 103226. External Links: Document Cited by: §2.2. J. Ma, Y. He, F. Li, L. Han, C. You, and B. Wang (2024) Segment anything in medical images. Nature communications 15 (1), p. 654. Cited by: §2.2. L. Machado, L. Alberge, H. Philippe, et al. (2025) A promptable CT foundation model for solid tumor evaluation. npj Precision Oncology 9 (1), p. 121. External Links: Document Cited by: §2.2. N. McConnell, P. Vasudev, D. Yamada, D. Cheng, M. Azimbagirad, J. McCabe, S. Aslani, A. H. Shahin, Y. Zhou, et al. (2026) A computationally frugal, open-source chest CT foundation model for thoracic disease detection in lung cancer screening programmes. Communications Medicine 6, p. 83. External Links: Document, Link Cited by: §2.1. F. Milletari, N. Navab, and S. Ahmadi (2016) V-net: fully convolutional neural networks for volumetric medical image segmentation. In Fourth International Conference on 3D Vision, 3DV 2016, Stanford, CA, USA, October 25-28, 2016, p. 565–571. External Links: Link, Document Cited by: §3.4. A. Muneer, E. Showkatian, Y. Kitsel, et al. (2026) CT-based deep foundation model for predicting immune checkpoint inhibitor-induced pneumonitis risk in lung cancer. medRxiv. External Links: Document, Link Cited by: §2.1. National Lung Screening Trial Research Team (2011) The National Lung Screening Trial: overview and study design. Radiology 258 (1), p. 243–253. External Links: Document, Link Cited by: §1, §4.1. X. Ni, L. Wu, J. Zhuang, Q. Wang, M. Wu, V. Vardhanabhuti, L. Zhang, H. Gao, and H. Chen (2026) MG-3D: multi-grained knowledge-enhanced vision–language pre-training for 3D medical image analysis. Medical Image Analysis 111, p. 104027. External Links: ISSN 1361-8415, Document, Link Cited by: §2.4. S. Oh, M. Kim, G. Jung, H. Lee, Y. Kim, S. Kim, H. Kwon, and H. Bae (2026) Anatomy-aware representation learning for medical ultrasound. In The Fourteenth International Conference on Learning Representations, External Links: Link Cited by: §2.3. M. Oquab, T. Darcet, T. Moutakanni, H. Vo, M. Szafraniec, V. Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, et al. (2024) DINOv2: learning robust visual features without supervision. Transactions on Machine Learning Research. External Links: Link Cited by: §1, §2.1, §3.1, §3.3, §3.3, §4.3. A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever (2021) Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, M. Meila and T. Zhang (Eds.), Proceedings of Machine Learning Research, Vol. 139, p. 8748–8763. External Links: Link Cited by: §5.4. A. Sablayrolles, M. Douze, C. Schmid, and H. Jégou (2019) Spreading vectors for similarity search. In International Conference on Learning Representations, External Links: Link Cited by: §2.1, §3.3. Z. Shui, J. Zhang, W. Cao, S. Wang, R. Guo, L. Lu, L. Yang, X. Ye, T. Liang, Q. Zhang, and L. Zhang (2025) Large-scale and fine-grained vision–language pre-training for enhanced CT image understanding. In The Thirteenth International Conference on Learning Representations, Cited by: §1, §2.4. S. Song, S. Yoon, P. Jin, S. Kim, M. Tivnan, Y. Oh, R. Meng, L. Chen, Z. Lyu, D. Wu, et al. (2025) OWT: a foundational organ-wise tokenization framework for medical imaging. arXiv preprint arXiv:2505.04899. Cited by: §1, §2.3. M. R. H. Taher, M. B. Gotway, and J. Liang (2024) Representing part-whole hierarchies in foundation models by learning localizability composability and decomposability from anatomy via self supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 11269–11281. Cited by: §2.3. Y. Tang, D. Yang, W. Li, H. R. Roth, B. Landman, D. Xu, V. Nath, and A. Hatamizadeh (2022) Self-supervised pre-training of Swin transformers for 3D medical image analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 20730–20740. Cited by: §2.1. J. Wang, H. Reynaud, I. E. Hamamci, S. Er, S. Shit, B. Menze, and B. Kainz (2026) SigVLP: sigmoid volume–language pre-training for self-supervised CT-volume adaptive representation learning. arXiv preprint arXiv:2602.21735. Cited by: §2.1. J. Wasserthal, H. Breit, M. T. Meyer, M. Pradella, D. Hinck, A. W. Sauter, T. Heye, D. T. Boll, J. Cyriac, S. Yang, et al. (2023) TotalSegmentator: robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence 5 (5), p. e230024. Cited by: §2.2, §4.2. C. Wu, X. Zhang, Y. Zhang, Y. Wang, and W. Xie (2023) MedKLIP: medical knowledge enhanced language-image pre-training for X-Ray diagnosis. In Proceedings of the IEEE/CVF international conference on computer vision, p. 21372–21383. Cited by: §2.4. L. Wu, J. Zhuang, and H. Chen (2024) VoCo: a simple-yet-effective volume contrastive learning framework for 3D medical image analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 22873–22882. Cited by: §2.1. Q. Wu, G. McIlvain, Z. Yu, and J. Wen (2026) Pan-FM: a pan-organ foundation model with saliency-guided masking for missing robustness. arXiv preprint arXiv:2605.07055. Cited by: §2.3. Y. Xie, J. Zhang, Y. Xia, and Q. Wu (2022) UniMiSS: universal medical self-supervised learning via breaking dimensionality barrier. In European Conference on Computer Vision, p. 558–575. Cited by: §2.1. H. Xu, A. Woicik, S. Asadian, J. Shen, Z. Zhang, A. Nabipoor, J. P. Musi, J. Keenan, M. Khorsandi, B. Al-Alao, I. Dimarakis, H. Chalian, Y. Lin, D. Fishbein, J. Pal, S. Wang, and S. Lin (2025) A cardiac-specific CT foundation model for heart transplantation. medRxiv. External Links: Document, Link Cited by: §A.3, §1, §1, §2.3, §4.2, §4.3, §5. K. Yamamoto and T. Kikuchi (2026) TotalFM: an organ-separated 3D-CT foundation model leveraging large-scale routine clinical radiology data. arXiv preprint arXiv:2601.00260. Cited by: §1, §2.3. J. Yang, D. Cai, J. Liu, Z. Zhuang, Y. Zhao, F. Wang, C. Li, C. Hu, B. Gai, Y. Chen, Y. Li, L. Wang, F. Gao, and X. Wu (2025) CRCFound: a colorectal cancer CT image foundation model based on self-supervised learning. Advanced Science 12 (41), p. e07339. External Links: Document, Link Cited by: §1, §2.1. G. You, H. Li, Y. Zhang, and Y. Fan (2026) Learning anatomy-grounded CT vision–language representations with organ-hierarchical report knowledge. arXiv preprint arXiv:2607.10953. Cited by: §1, §2.4. H. Yu and Q. Dai (2026) Anatomy-guided self-supervised distillation learning for medical image analysis. IEEE Transactions on Medical Imaging 45 (5), p. 1787–1801. External Links: Document, Link Cited by: §2.3. K. Yu, L. Sun, J. Chen, M. Reynolds, T. Chaudhary, and K. Batmanghelich (2024) DrasCLR: a self-supervised framework of learning disease-related and anatomy-specific representation for 3d lung ct images. Medical Image Analysis 92, p. 103062. Cited by: §2.3. X. Zhang, C. Wu, Y. Zhang, W. Xie, and Y. Wang (2023) Knowledge-enhanced visual-language pre-training on chest radiology images. Nature Communications 14 (1), p. 4542. Cited by: §2.4. H. Zhao, R. Zhang, Z. Wang, Y. Gu, S. Xu, S. Wang, and Y. Li (2026) BoneCoT: multicentre validation of a whole-body skeleton foundation model for bone metastases guided by clinician-derived chain of thought. Nature Biomedical Engineering. External Links: Document, Link Cited by: §2.3. J. Zhou, C. Wei, H. Wang, W. Shen, C. Xie, A. Yuille, and T. Kong (2022) iBOT: image BERT pre-training with online tokenizer. In International Conference on Learning Representations, External Links: Link Cited by: §2.1, §3.3. Z. Zhou, V. Sodha, J. Pang, M. B. Gotway, and J. Liang (2021) Models Genesis. Medical image analysis 67, p. 101840. Cited by: §2.1. W. Zhu, H. Huang, H. Tang, R. Musthyala, B. Yu, L. Chen, E. Vega, T. O’Donnell, R. Hayek, L. Kuohn, S. Dehkharghani, J. A. Frontera, A. V. Masurkar, K. Melmed, and N. Razavian (2026) 3D foundation model for generalizable disease detection in head computed tomography. Nature Biomedical Engineering. External Links: Document, Link Cited by: §2.1. Appendix A Additional Reproducibility Details A.1. Shared Downstream Protocol Unless stated otherwise, downstream evaluations used frozen teacher features, with all examinations from each patient assigned to one split. Hyperparameters were selected using only training and validation data, and available baselines followed the same splits and downstream protocol. Every organ-specific representation was computed from 64 uniformly sampled axial slices. A.2. Detailed Image and Mask Preprocessing Images and segmentation masks are transformed on the same spatial grid using data-appropriate interpolation. Images are resampled to 11\,m isotropic spacing with trilinear interpolation, center cropped or padded to 5123512^3 voxels, clamped to [−1000,1000][-1000,1000] HU, and scaled to [−1,1][-1,1]. Padded voxels take the value −1-1. Segmentation volumes use nearest-neighbor interpolation and zero padding to preserve discrete labels. We extract aligned axial, coronal, and sagittal image–mask pairs at 256×256256× 256 resolution, map each image slice to three-channel grayscale, and apply DINO normalization after spatial cropping. Downstream slices are center cropped to 224×224224× 224 before normalization. A.3. Pretraining Hyperparameters OrganLens was adapted on the CT-RATE training split using CT images and TotalSegmentator pseudo-labels. Only the ViT-L/16 backbones of the student and teacher were initialized from the CT-pretrained DINOv2 checkpoint released with GigaHeart (Xu et al., 2025). The shared DINO/iBOT projection head, organ-conditioning module, and anatomy decoder were initialized from scratch. The two 224×224224× 224 global crops and eight 96×9696× 96 local crops use scale ranges of [0.32,1.0][0.32,1.0] and [0.05,0.32][0.05,0.32], respectively. A candidate global crop is accepted when it contains at least 32 organ-mask pixels, 1% foreground, and 75% of the organ bounding box. A candidate local crop requires at least 8 organ-mask pixels and 3% foreground. If repeated sampling fails, the crop is drawn around the organ bounding box or a foreground point. For iBOT, masking is applied to 50% of samples, with the patch ratio drawn from [0.1,0.5][0.1,0.5]. Each of the three complete pretraining passes contains 175,000 iterations. The learning-rate schedule includes a 12,500-iteration warmup. Weight decay increases from 0.04 to 0.4, and teacher momentum increases from 0.992 to 1.0. The anatomy decoder uses progressive convolutional blocks with channel widths of 256, 128, 64, and 32, followed by bilinear upsampling and a 1×11× 1 prediction layer. Training uses mixed precision and fully sharded data parallelism with a random seed of 0. Numerical stabilizers are 10−810^-8 for KoLeo and 10−610^-6 for Dice. Pretraining was performed on eight NVIDIA RTX PRO 6000 GPUs. A.4. Downstream Model Training Each abnormality head is a one-hidden-layer, 512-unit MLP. The CT-RATE heads are jointly trained for five epochs with unweighted binary cross-entropy. They use AdamW with a learning rate of 2×10−32× 10^-3, weight decay of 0.05, a batch size of 16, one warmup epoch, and no dropout. The RAD-ChestCT heads use the same architecture, optimizer, learning rate, and weight decay, with a batch size of 64 and class-weighted binary cross-entropy. They are trained for 100 epochs, with checkpoint selection performed separately for each label using validation AUROC. The OrganLens, CT-pretrained DINOv2, GigaHeart, and Merlin Cox models standardize their input features and use a one-hidden-layer MLP with 512 ReLU units, no dropout, and one scalar log-hazard output. Full-batch optimization uses AdamW with a learning rate of 2×10−32× 10^-3, weight decay of 0.05, five warmup epochs, and cosine decay. Training runs for up to 200 epochs and stops after 20 consecutive epochs without validation Harrell C-index improvement. These settings are fixed across representation–endpoint pairs rather than selected through a separate hyperparameter search. For jointly fine-tuned image–text retrieval models, the image encoder, ClinicalBERT text encoder, and linear projection heads are optimized for five epochs using AdamW and symmetric contrastive loss. CT-RATE uses learning rates of 10−510^-5 for both the encoders and projection heads, whereas INSPECT uses 10−510^-5 for the encoders and 10−410^-4 for the projection heads. Both use a weight decay of 0.01. Both runs use four GPUs and a micro-batch size of 4. CT-RATE accumulates 16 micro-batches per rank and INSPECT accumulates 32, producing local contrastive pools of 64 and 128 pairs, respectively, before gradient averaging across ranks. INSPECT selects the checkpoint with the highest validation text-to-image Recall@10, whereas CT-RATE uses the fifth-epoch checkpoint for evaluation on its validation candidate set. When multiple volume files correspond to the same CT-RATE scan, we retain the first volume. Appendix B Complete Organ-Specific Abnormality Detection Results Supplementary Table S1 reports AUROCs for every organ-specific OrganLens representation and abnormality on the CT-RATE validation split and RAD-ChestCT held-out test split. The table also includes volume-level CT-pretrained DINOv2, GigaHeart, Merlin, and SPECTRE baselines. CT-pretrained DINOv2 and GigaHeart use mean-pooled slice-level CLS features, Merlin uses its global image embedding, and SPECTRE uses its scan-level CLS feature. All OrganLens representations are computed from 64 uniformly sampled axial slices and use anatomy-weighted pooling. Supplementary Table S1. Complete organ-specific abnormality detection results. Panel A reports validation AUROCs on CT-RATE, and Panel B reports held-out test AUROCs on RAD-ChestCT. Best and second-best values within each abnormality are shown in bold and underlined, respectively. Ties after rounding to three decimal places receive the same formatting. A dagger marks the anatomically matched OrganLens representation used in Figure 2. Concat and mean aggregate all 11 organ-specific representations. CT-pretrained DINOv2 and GigaHeart use mean-pooled slice-level CLS features, Merlin uses its global image embedding, and SPECTRE uses its scan-level CLS feature. SPECTRE was trained on the CT-RATE training split. OrganLens organ-specific representations OrganLens global aggregations Volume-level baselines Abnormality Heart Lung Aorta Liver Spleen Kidneys Pancreas Stomach Intestine Esophagus Trachea Concat Mean CT-pretrained DINOv2 GigaHeart Merlin SPECTRE (A) CT-RATE validation Medical material 0.863 0.852 0.885 0.822 0.849 0.857 0.811 0.824 0.818 0.854 0.847 0.887 0.874 0.792 0.804 0.864 0.873 Arterial wall calcification 0.919 0.915 0.928† 0.909 0.911 0.919 0.906 0.902 0.910 0.919 0.915 0.921 0.922 0.885 0.888 0.890 0.922 Cardiomegaly 0.953† 0.947 0.947 0.919 0.938 0.938 0.916 0.916 0.908 0.942 0.934 0.950 0.949 0.910 0.917 0.902 0.909 Pericardial effusion 0.925† 0.880 0.889 0.844 0.877 0.857 0.839 0.864 0.841 0.871 0.856 0.910 0.897 0.818 0.823 0.831 0.860 Coronary artery wall calcification 0.905† 0.899 0.911 0.898 0.897 0.902 0.893 0.890 0.894 0.904 0.897 0.905 0.909 0.881 0.876 0.881 0.916 Hiatal hernia 0.769 0.763 0.790 0.788 0.789 0.770 0.822 0.814 0.771 0.849† 0.825 0.846 0.833 0.728 0.745 0.730 0.750 Lymphadenopathy 0.739 0.753 0.744 0.727 0.731 0.727 0.714 0.720 0.727 0.749 0.751 0.756 0.753 0.709 0.717 0.719 0.725 Emphysema 0.799 0.819† 0.797 0.791 0.787 0.788 0.783 0.775 0.789 0.801 0.801 0.804 0.807 0.761 0.757 0.760 0.798 Atelectasis 0.758 0.782† 0.747 0.743 0.752 0.728 0.711 0.727 0.728 0.749 0.732 0.765 0.763 0.703 0.707 0.716 0.787 Lung nodule 0.687 0.696† 0.685 0.675 0.676 0.658 0.642 0.656 0.670 0.691 0.679 0.647 0.694 0.638 0.637 0.642 0.685 Lung opacity 0.781 0.839† 0.777 0.726 0.753 0.723 0.681 0.716 0.742 0.777 0.765 0.824 0.816 0.725 0.726 0.699 0.867 Pulmonary fibrotic sequela 0.694 0.714† 0.694 0.674 0.686 0.677 0.674 0.668 0.670 0.695 0.680 0.700 0.694 0.656 0.675 0.655 0.696 Pleural effusion 0.961 0.974† 0.961 0.960 0.962 0.950 0.939 0.952 0.934 0.963 0.958 0.972 0.967 0.928 0.941 0.952 0.972 Mosaic attenuation pattern 0.874 0.906† 0.881 0.861 0.867 0.854 0.821 0.839 0.852 0.882 0.872 0.897 0.885 0.853 0.844 0.799 0.847 Peribronchial thickening 0.804 0.823† 0.808 0.791 0.798 0.796 0.776 0.785 0.789 0.812 0.810 0.825 0.825 0.773 0.779 0.750 0.799 Consolidation 0.811 0.887† 0.823 0.793 0.803 0.794 0.755 0.789 0.801 0.827 0.822 0.854 0.857 0.785 0.776 0.765 0.910 Bronchiectasis 0.781 0.800† 0.781 0.744 0.759 0.750 0.730 0.730 0.743 0.776 0.774 0.792 0.787 0.723 0.740 0.685 0.786 Interlobular septal thickening 0.871 0.892† 0.871 0.848 0.840 0.847 0.803 0.820 0.842 0.862 0.868 0.887 0.887 0.835 0.846 0.823 0.873 (B) RAD-ChestCT held-out test Arterial wall calcification 0.717 0.704 0.682† 0.648 0.646 0.670 0.680 0.635 0.641 0.698 0.675 0.680 0.690 0.648 0.668 0.671 0.632 Atelectasis 0.655 0.758† 0.680 0.668 0.677 0.661 0.639 0.630 0.663 0.659 0.666 0.726 0.710 0.654 0.654 0.649 0.733 Bronchiectasis 0.774 0.796† 0.759 0.732 0.805 0.704 0.648 0.675 0.695 0.770 0.715 0.773 0.769 0.753 0.745 0.718 0.784 Cardiomegaly 0.946† 0.917 0.929 0.907 0.924 0.915 0.893 0.901 0.845 0.892 0.858 0.943 0.923 0.901 0.903 0.893 0.822 Consolidation 0.709 0.776† 0.726 0.688 0.700 0.695 0.714 0.630 0.704 0.704 0.748 0.729 0.737 0.678 0.682 0.687 0.814 Coronary artery wall calcification 0.787† 0.755 0.780 0.750 0.744 0.772 0.746 0.731 0.712 0.749 0.762 0.766 0.768 0.707 0.751 0.779 0.709 Emphysema 0.857 0.909† 0.869 0.828 0.800 0.835 0.771 0.743 0.768 0.837 0.813 0.882 0.890 0.819 0.805 0.729 0.836 Hiatal hernia 0.756 0.753 0.770 0.795 0.770 0.733 0.810 0.833 0.747 0.851† 0.799 0.842 0.823 0.700 0.690 0.764 0.654 Septal thickening 0.764 0.801† 0.726 0.693 0.653 0.664 0.618 0.614 0.537 0.678 0.674 0.701 0.701 0.678 0.644 0.685 0.752 Nodule >1>1 cm 0.585 0.645† 0.653 0.611 0.586 0.568 0.516 0.576 0.549 0.635 0.533 0.524 0.594 0.581 0.550 0.550 0.614 Opacity 0.647 0.711† 0.676 0.603 0.613 0.630 0.622 0.641 0.591 0.677 0.661 0.661 0.663 0.602 0.620 0.676 0.690 Lymphadenopathy 0.692 0.726 0.707 0.697 0.633 0.690 0.713 0.644 0.589 0.743 0.741 0.710 0.723 0.683 0.685 0.673 0.678 Bronchial wall thickening 0.643 0.696† 0.677 0.629 0.617 0.601 0.596 0.578 0.556 0.670 0.550 0.652 0.659 0.613 0.644 0.512 0.584 Pericardial effusion 0.791† 0.686 0.725 0.721 0.702 0.699 0.653 0.693 0.673 0.685 0.650 0.749 0.718 0.677 0.670 0.714 0.665 Pleural effusion 0.921 0.938† 0.914 0.911 0.918 0.897 0.885 0.867 0.893 0.898 0.872 0.939 0.924 0.853 0.887 0.902 0.922 Fibrosis 0.803 0.835† 0.799 0.767 0.770 0.753 0.665 0.728 0.717 0.807 0.768 0.829 0.815 0.825 0.829 0.759 0.861 Ground-glass opacity 0.596 0.670† 0.597 0.555 0.533 0.569 0.537 0.486 0.513 0.552 0.525 0.585 0.574 0.552 0.597 0.549 0.648 Lung cancer 0.769 0.820† 0.771 0.716 0.750 0.767 0.675 0.659 0.721 0.753 0.724 0.786 0.786 0.687 0.724 0.655 0.841 Lung infection 0.557 0.622† 0.591 0.592 0.564 0.525 0.621 0.599 0.554 0.585 0.584 0.612 0.598 0.531 0.582 0.593 0.608 Lung scarring 0.693 0.705† 0.694 0.666 0.680 0.690 0.637 0.662 0.649 0.642 0.655 0.696 0.697 0.612 0.665 0.629 0.655 Interstitial lung disease 0.771 0.842† 0.781 0.755 0.736 0.747 0.675 0.709 0.693 0.737 0.753 0.794 0.789 0.773 0.783 0.765 0.820 Lung calcification 0.631 0.647† 0.607 0.600 0.571 0.568 0.559 0.589 0.634 0.571 0.569 0.591 0.610 0.581 0.571 0.633 0.608 Aortic atherosclerosis 0.642 0.660 0.640† 0.607 0.635 0.624 0.640 0.567 0.579 0.627 0.575 0.621 0.643 0.624 0.613 0.613 0.612 Appendix C Complete Prognostic Prediction Results Supplementary Table S2 reports the complete test-set Harrell C-indices for all organ-specific OrganLens representations, their concatenated and mean aggregations, and the volume-level baselines on INSPECT and NLST. Daggered OrganLens entries correspond to Figure 3. Supplementary Table S2. Complete prognostic prediction results. Panel A reports test-set Harrell C-indices on INSPECT, and Panel B reports test-set Harrell C-indices on NLST. Best and second-best values within each endpoint are shown in bold and underlined, respectively. Ties after rounding to three decimal places receive the same formatting. A dagger marks the anatomically matched OrganLens representation used in Figure 3. Concat and mean aggregate all 11 organ-specific representations. SPECTRE was trained on INSPECT and NLST. Representation Mortality Readmission Pulmonary Hypertension Pleural Effusion Consolidation Edema Atelectasis Cardiomegaly (A) INSPECT OrganLens organ-specific representations Heart 0.783 0.593 0.742† 0.818 0.749 0.661 0.738 0.745† Spleen 0.768 0.586 0.702 0.801 0.713 0.639 0.714 0.703 Kidneys 0.774 0.589 0.695 0.807 0.734 0.655 0.727 0.704 Liver 0.767 0.607 0.687 0.818 0.705 0.654 0.722 0.692 Pancreas 0.783 0.601 0.702 0.819 0.757 0.663 0.741 0.711 Lung 0.795 0.591 0.736 0.836† 0.778† 0.665† 0.756† 0.735 Intestine 0.764 0.578 0.694 0.795 0.719 0.645 0.699 0.693 Aorta 0.784 0.597 0.700 0.813 0.756 0.663 0.743 0.716 Stomach 0.765 0.572 0.689 0.794 0.697 0.651 0.703 0.701 Esophagus 0.786 0.598 0.701 0.819 0.759 0.664 0.744 0.699 Trachea 0.786 0.597 0.697 0.819 0.760 0.664 0.745 0.696 OrganLens global aggregations Concat 0.794† 0.608† 0.732 0.832 0.775 0.656 0.751 0.736 Mean 0.784 0.590 0.711 0.822 0.753 0.663 0.738 0.715 Volume-level baselines CT-pretrained DINOv2 0.784 0.592 0.692 0.802 0.746 0.659 0.741 0.709 GigaHeart 0.782 0.602 0.709 0.811 0.741 0.642 0.735 0.706 Merlin 0.759 0.586 0.676 0.806 0.718 0.654 0.708 0.671 SPECTRE 0.767 0.611 0.717 0.821 0.777 0.635 0.736 0.704 Representation LC inc. All-cause LC death CV IHD Resp. COPD Any cancer Non-LC (B) NLST OrganLens organ-specific representations Heart 0.648 0.732 0.686 0.753† 0.752† 0.886 0.935 0.663 0.738 Spleen 0.663 0.716 0.713 0.662 0.625 0.840 0.817 0.667 0.719 Kidneys 0.671 0.726 0.723 0.683 0.678 0.892 0.929 0.667 0.705 Liver 0.635 0.713 0.684 0.675 0.638 0.851 0.863 0.668 0.727 Pancreas 0.682 0.737 0.737 0.716 0.723 0.910 0.947 0.681 0.733 Lung 0.709† 0.748 0.780† 0.708 0.743 0.935† 0.960† 0.696 0.731 Intestine 0.652 0.711 0.691 0.691 0.665 0.864 0.879 0.643 0.703 Aorta 0.681 0.732 0.735 0.726 0.726 0.898 0.934 0.675 0.735 Stomach 0.662 0.705 0.705 0.705 0.710 0.824 0.813 0.664 0.713 Esophagus 0.686 0.738 0.746 0.718 0.731 0.908 0.949 0.687 0.737 Trachea 0.683 0.737 0.741 0.717 0.727 0.908 0.950 0.684 0.735 OrganLens global aggregations Concat 0.683 0.748† 0.755 0.741 0.749 0.928 0.951 0.707† 0.756† Mean 0.686 0.740 0.742 0.734 0.744 0.901 0.935 0.698 0.732 Volume-level baselines CT-pretrained DINOv2 0.640 0.701 0.683 0.638 0.488 0.860 0.898 0.597 0.674 GigaHeart 0.647 0.723 0.695 0.672 0.695 0.886 0.911 0.632 0.697 Merlin 0.639 0.697 0.676 0.747 0.734 0.860 0.873 0.640 0.713 SPECTRE 0.778 0.721 0.777 0.737 0.766 0.900 0.930 0.630 0.705 Appendix D Longitudinal Baseline Comparison Supplementary Figure S1 compares visit-specific Harrell C-indices across the common NLST longitudinal test cohort. OrganLens used the heart representation for cardiovascular endpoints and the lung representation for respiratory endpoints. It produced the highest point estimates at T0–T2 for cardiovascular, ischemic-heart-disease, and respiratory mortality. For COPD mortality, OrganLens and SPECTRE were comparable at T0, while OrganLens was higher at T1 and T2. Supplementary Figure S1. Longitudinal prognostic comparison on NLST. Visit-specific Harrell C-indices at T0–T2 are shown for (a) ischemic-heart-disease mortality, (b) cardiovascular mortality, (c) respiratory mortality, and (d) COPD mortality. OrganLens uses the heart representation for cardiovascular endpoints and the lung representation for respiratory endpoints. All methods are evaluated on the common longitudinal test cohort. Four line plots compare the visit-specific Harrell C-indices of OrganLens, CT-pretrained DINOv2, GigaHeart, Merlin, and SPECTRE across T0, T1, and T2 for two cardiovascular and two respiratory mortality endpoints. Appendix E Complete Image–Text Retrieval Results Supplementary Table S3 reports Recall@K in both retrieval directions for every evaluated model. Within each cohort, all methods use the same fixed candidate set. For OrganLens, both concatenated and mean aggregations of the 11 organ-specific representations are reported. The OrganLens representations are computed from 64 uniformly sampled axial slices. Supplementary Table S3. Complete bidirectional image–text retrieval results. Values are Recall@K (%). Best and second-best values within each cohort and direction are shown in bold and underlined, respectively. Ties after rounding to two decimal places receive the same formatting. SPECTRE was trained on the CT-RATE training split and INSPECT. The main OrganLens representation concatenates all 11 organ-specific representations. Mean aggregation averages them instead. Text-to-image Image-to-text Method R@5 R@10 R@50 R@100 R@5 R@10 R@50 R@100 CT-RATE OrganLens Concat 36.45 46.55 69.18 78.52 37.08 46.42 69.05 79.09 Mean 37.47 46.93 68.93 79.41 36.25 45.72 69.18 78.45 CT-pretrained DINOv2 27.88 36.32 62.66 74.49 26.21 36.32 62.28 75.38 GigaHeart 28.58 37.79 66.43 78.32 26.85 36.38 64.83 76.85 Merlin 11.00 16.43 42.14 57.23 11.32 17.07 41.43 56.14 CT-CLIP 2.62 4.54 18.16 31.01 2.43 5.05 17.90 31.27 SPECTRE 22.44 31.52 56.59 67.33 22.12 31.39 57.74 67.97 INSPECT OrganLens Concat 23.88 33.09 58.37 68.90 23.19 32.04 57.25 68.43 Mean 22.63 31.66 57.35 68.46 22.73 31.29 56.76 68.06 CT-pretrained DINOv2 15.29 22.38 47.60 61.08 13.61 21.26 45.14 58.00 GigaHeart 11.61 18.15 41.81 55.32 11.74 17.81 40.78 53.89 Merlin 10.37 16.16 39.79 53.77 9.71 15.57 38.89 52.77 CT-CLIP 4.92 8.75 25.37 38.45 6.01 10.18 26.87 38.67 SPECTRE 15.47 23.85 51.77 67.12 15.82 23.57 51.62 66.44 Appendix F Complete Ablation Results Supplementary Table S4 compares the two-by-two combinations of organ-guided cropping and anatomy-mask supervision, and Supplementary Table S5 compares four pooling strategies. Both report the 16 CT-RATE and 22 RAD-ChestCT abnormalities with predefined anatomical matches, including the matched organ-specific representation and the mean and concatenated aggregations of all 11 organ-specific representations. Supplementary Table S4. Complete pretraining-component ablation results. Panels A and B report AUROCs on the CT-RATE validation split and RAD-ChestCT held-out test split for the 16 and 22 abnormalities with predefined anatomical matches, respectively. Matched uses the organ-specific representation assigned to each abnormality, Mean averages the 11 organ-specific representations, and Concat concatenates them. The variants form a two-by-two comparison of organ-guided cropping and anatomy-mask supervision. Variants without anatomy-mask supervision use mean CLS pooling; variants with it use anatomy-weighted pooling. The best and second-best values within each abnormality are shown in bold and underlined, respectively. Each style marks one value; ties after rounding to three decimal places favor OrganLens (Full) when applicable. No organ-guided crops or anatomy-mask supervision No anatomy-mask supervision No organ-guided crops OrganLens (Full) Abnormality Matched Mean Concat Matched Mean Concat Matched Mean Concat Matched Mean Concat (A) CT-RATE validation Arterial wall calcification 0.892 0.888 0.901 0.896 0.895 0.906 0.929 0.920 0.923 0.928 0.922 0.921 Cardiomegaly 0.926 0.925 0.917 0.940 0.939 0.933 0.954 0.950 0.950 0.953 0.949 0.950 Pericardial effusion 0.813 0.797 0.822 0.834 0.842 0.852 0.920 0.882 0.907 0.925 0.897 0.910 Coronary artery wall calcification 0.869 0.868 0.874 0.876 0.876 0.890 0.903 0.905 0.909 0.905 0.909 0.905 Hiatal hernia 0.728 0.744 0.746 0.754 0.747 0.750 0.845 0.820 0.838 0.849 0.833 0.846 Emphysema 0.760 0.751 0.773 0.766 0.765 0.792 0.818 0.806 0.803 0.819 0.807 0.804 Atelectasis 0.718 0.715 0.728 0.725 0.723 0.743 0.777 0.761 0.768 0.782 0.763 0.765 Lung nodule 0.642 0.650 0.662 0.652 0.653 0.672 0.696 0.696 0.692 0.696 0.694 0.647 Lung opacity 0.779 0.767 0.783 0.791 0.792 0.790 0.832 0.799 0.814 0.839 0.816 0.824 Pulmonary fibrotic sequela 0.651 0.655 0.660 0.658 0.659 0.672 0.712 0.691 0.699 0.714 0.694 0.700 Pleural effusion 0.962 0.963 0.961 0.965 0.963 0.966 0.971 0.969 0.971 0.974 0.967 0.972 Mosaic attenuation pattern 0.847 0.839 0.861 0.874 0.870 0.883 0.903 0.884 0.892 0.906 0.885 0.897 Peribronchial thickening 0.770 0.751 0.796 0.773 0.769 0.802 0.825 0.812 0.814 0.823 0.825 0.825 Consolidation 0.824 0.821 0.828 0.835 0.834 0.839 0.888 0.843 0.854 0.887 0.857 0.854 Bronchiectasis 0.738 0.720 0.742 0.744 0.741 0.763 0.788 0.776 0.782 0.800 0.787 0.792 Interlobular septal thickening 0.843 0.842 0.859 0.858 0.852 0.861 0.889 0.876 0.879 0.892 0.887 0.887 (B) RAD-ChestCT held-out test Arterial wall calcification 0.691 0.686 0.676 0.690 0.693 0.702 0.678 0.666 0.677 0.682 0.690 0.680 Atelectasis 0.708 0.707 0.707 0.720 0.716 0.719 0.755 0.708 0.695 0.758 0.710 0.726 Bronchiectasis 0.766 0.768 0.755 0.779 0.781 0.781 0.804 0.785 0.766 0.796 0.769 0.773 Cardiomegaly 0.922 0.918 0.925 0.935 0.930 0.926 0.947 0.932 0.939 0.946 0.923 0.943 Consolidation 0.745 0.736 0.722 0.745 0.743 0.731 0.764 0.747 0.739 0.776 0.737 0.729 Coronary artery wall calcification 0.743 0.742 0.760 0.779 0.778 0.777 0.786 0.764 0.758 0.787 0.768 0.766 Emphysema 0.861 0.861 0.826 0.821 0.837 0.837 0.910 0.876 0.878 0.909 0.890 0.882 Hiatal hernia 0.704 0.699 0.676 0.756 0.736 0.756 0.838 0.774 0.828 0.851 0.823 0.842 Septal thickening 0.668 0.670 0.635 0.722 0.722 0.682 0.809 0.721 0.680 0.801 0.701 0.701 Nodule >1>1 cm 0.545 0.539 0.582 0.575 0.538 0.568 0.614 0.600 0.625 0.645 0.594 0.524 Opacity 0.629 0.638 0.644 0.670 0.674 0.651 0.724 0.636 0.637 0.711 0.663 0.661 Bronchial wall thickening 0.599 0.598 0.600 0.625 0.626 0.686 0.669 0.733 0.649 0.696 0.659 0.652 Pericardial effusion 0.689 0.692 0.692 0.702 0.685 0.698 0.796 0.738 0.738 0.791 0.718 0.749 Pleural effusion 0.911 0.911 0.912 0.921 0.928 0.916 0.936 0.928 0.927 0.938 0.924 0.939 Fibrosis 0.850 0.852 0.813 0.844 0.844 0.839 0.848 0.857 0.817 0.835 0.815 0.829 Ground-glass opacity 0.580 0.581 0.557 0.596 0.596 0.568 0.650 0.548 0.556 0.670 0.574 0.585 Lung cancer 0.764 0.767 0.764 0.774 0.771 0.758 0.823 0.765 0.772 0.820 0.786 0.786 Lung infection 0.545 0.592 0.567 0.568 0.570 0.548 0.612 0.627 0.577 0.622 0.598 0.612 Lung scarring 0.666 0.642 0.632 0.680 0.711 0.678 0.696 0.695 0.678 0.705 0.697 0.696 Interstitial lung disease 0.818 0.818 0.813 0.815 0.818 0.814 0.836 0.802 0.773 0.842 0.789 0.794 Lung calcification 0.601 0.601 0.594 0.597 0.599 0.598 0.620 0.592 0.527 0.647 0.610 0.591 Aortic atherosclerosis 0.630 0.621 0.584 0.592 0.591 0.579 0.624 0.645 0.631 0.640 0.643 0.621 Supplementary Table S5. Complete pooling ablation results. Panels A and B report AUROCs on the CT-RATE validation split and RAD-ChestCT held-out test split for the 16 and 22 abnormalities with predefined anatomical matches, respectively. Matched uses the organ-specific representation assigned to each abnormality, Mean averages the 11 organ-specific representations, and Concat concatenates them. Mean CLS Pooling averages organ-conditioned slice CLS features. Uniform Patch Pooling averages all patch features. The w/o Inter-slice Weighting variant weights patches by predicted spatial support but averages slices uniformly. Full additionally weights slices by predicted organ area. All variants retain all sampled slices. The Full group reproduces the complete-method results in Supplementary Table S1. The best and second-best values within each abnormality are shown in bold and underlined, respectively. Each style marks one value. Ties after rounding to three decimal places favor OrganLens (Full) when applicable. OrganLens w/ Mean CLS Pooling OrganLens w/ Uniform Patch Pooling OrganLens w/o Inter-slice Weighting OrganLens (Full) Abnormality Matched Mean Concat Matched Mean Concat Matched Mean Concat Matched Mean Concat (A) CT-RATE validation Arterial wall calcification 0.895 0.894 0.898 0.914 0.913 0.913 0.925 0.923 0.921 0.928 0.922 0.921 Cardiomegaly 0.924 0.927 0.921 0.940 0.934 0.935 0.949 0.947 0.948 0.953 0.949 0.950 Pericardial effusion 0.813 0.804 0.830 0.861 0.854 0.857 0.906 0.892 0.894 0.925 0.897 0.910 Coronary artery wall calcification 0.877 0.875 0.880 0.901 0.899 0.891 0.908 0.908 0.908 0.905 0.909 0.905 Hiatal hernia 0.743 0.736 0.740 0.763 0.762 0.767 0.836 0.826 0.836 0.849 0.833 0.846 Emphysema 0.771 0.781 0.787 0.815 0.808 0.806 0.822 0.809 0.804 0.819 0.807 0.804 Atelectasis 0.719 0.729 0.726 0.748 0.739 0.735 0.769 0.760 0.763 0.782 0.763 0.765 Lung nodule 0.640 0.646 0.638 0.691 0.683 0.572 0.696 0.694 0.692 0.696 0.694 0.647 Lung opacity 0.759 0.760 0.775 0.814 0.788 0.784 0.834 0.816 0.825 0.839 0.816 0.824 Pulmonary fibrotic sequela 0.675 0.666 0.677 0.696 0.690 0.691 0.709 0.700 0.703 0.714 0.694 0.700 Pleural effusion 0.958 0.960 0.959 0.964 0.960 0.960 0.972 0.964 0.969 0.974 0.967 0.972 Mosaic attenuation pattern 0.853 0.868 0.875 0.879 0.876 0.876 0.900 0.886 0.892 0.906 0.885 0.897 Peribronchial thickening 0.770 0.762 0.775 0.811 0.804 0.807 0.823 0.820 0.822 0.823 0.825 0.825 Consolidation 0.820 0.821 0.832 0.839 0.812 0.807 0.878 0.861 0.861 0.887 0.857 0.854 Bronchiectasis 0.747 0.752 0.770 0.790 0.783 0.783 0.799 0.794 0.796 0.800 0.787 0.792 Interlobular septal thickening 0.864 0.856 0.857 0.880 0.865 0.873 0.888 0.883 0.883 0.892 0.887 0.887 (B) RAD-ChestCT held-out test Arterial wall calcification 0.647 0.641 0.639 0.682 0.685 0.698 0.695 0.688 0.695 0.682 0.690 0.680 Atelectasis 0.685 0.696 0.723 0.695 0.679 0.673 0.735 0.703 0.706 0.758 0.710 0.726 Bronchiectasis 0.800 0.793 0.775 0.765 0.747 0.728 0.797 0.767 0.777 0.796 0.769 0.773 Cardiomegaly 0.917 0.894 0.879 0.933 0.906 0.910 0.940 0.920 0.924 0.946 0.923 0.943 Consolidation 0.729 0.749 0.700 0.698 0.692 0.695 0.746 0.725 0.722 0.776 0.737 0.729 Coronary artery wall calcification 0.766 0.766 0.745 0.778 0.772 0.770 0.783 0.776 0.769 0.787 0.768 0.766 Emphysema 0.838 0.834 0.842 0.882 0.875 0.880 0.908 0.864 0.888 0.909 0.890 0.882 Hiatal hernia 0.745 0.721 0.734 0.764 0.753 0.620 0.835 0.789 0.822 0.851 0.823 0.842 Septal thickening 0.711 0.686 0.651 0.711 0.664 0.659 0.757 0.708 0.715 0.801 0.701 0.701 Nodule >1>1 cm 0.546 0.555 0.546 0.602 0.550 0.447 0.589 0.609 0.550 0.645 0.594 0.524 Opacity 0.624 0.656 0.645 0.629 0.646 0.641 0.681 0.648 0.671 0.711 0.663 0.661 Bronchial wall thickening 0.659 0.644 0.654 0.606 0.544 0.536 0.659 0.631 0.635 0.696 0.659 0.652 Pericardial effusion 0.690 0.656 0.669 0.725 0.709 0.724 0.766 0.752 0.745 0.791 0.718 0.749 Pleural effusion 0.912 0.909 0.903 0.935 0.931 0.931 0.942 0.934 0.938 0.938 0.924 0.939 Fibrosis 0.819 0.820 0.855 0.824 0.825 0.826 0.828 0.813 0.811 0.835 0.815 0.829 Ground-glass opacity 0.580 0.575 0.548 0.593 0.605 0.580 0.638 0.558 0.569 0.670 0.574 0.585 Lung cancer 0.745 0.758 0.759 0.791 0.788 0.790 0.805 0.787 0.791 0.820 0.786 0.786 Lung infection 0.584 0.600 0.580 0.557 0.538 0.521 0.611 0.618 0.625 0.622 0.598 0.612 Lung scarring 0.652 0.665 0.653 0.682 0.669 0.667 0.687 0.695 0.698 0.705 0.697 0.696 Interstitial lung disease 0.830 0.837 0.823 0.804 0.805 0.797 0.816 0.800 0.776 0.842 0.789 0.794 Lung calcification 0.567 0.598 0.593 0.607 0.599 0.614 0.597 0.597 0.603 0.647 0.610 0.591 Aortic atherosclerosis 0.608 0.605 0.620 0.629 0.639 0.626 0.638 0.636 0.615 0.640 0.643 0.621

Instant research discovery