Paper deep dive

Temporal Inversion for Learning Interval Change in Chest X-Rays

Hanbin Ko, Kyungmin Jeon, Doowoong Choi, Chang Min Park

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 80

Intelligence

Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 93%

Last extracted: 4/10/2026, 2:52:14 AM

Summary

TILA (Temporal Inversion-aware Learning and Alignment) is a framework designed to improve temporal reasoning in medical vision-language models by using temporal inversion as a supervisory signal. By reversing image pairs and enforcing consistency in pretraining, fine-tuning, and inference, TILA enhances the sensitivity of models to directional changes in chest X-rays, outperforming existing baselines in progression classification and temporal embedding alignment.

Entities (5)

TILA · framework · 100%ALTA · model-architecture · 95%BioViL-T · model-architecture · 95%MS-CXR-T · dataset · 95%SigLIP · loss-functionmethod · 90%

Relation Signals (3)

TILA → improves → BioViL-T

confidence 95% · We validate the framework on two representative temporal VLP models, BioViL-T [1] and ALTA [13], demonstrating that the same principles consistently enhance directional reasoning

TILA → improves → ALTA

confidence 95% · We validate the framework on two representative temporal VLP models, BioViL-T [1] and ALTA [13], demonstrating that the same principles consistently enhance directional reasoning

TILA → utilizes → MS-CXR-T

confidence 90% · We introduce MS-CXR-T retrieval for standardized evaluation of temporal reasoning.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Recent advances in vision--language pretraining have enabled strong medical foundation models, yet most analyze radiographs in isolation, overlooking the key clinical task of comparing prior and current images to assess interval change. For chest radiographs (CXRs), capturing interval change is essential, as radiologists must evaluate not only the static appearance of findings but also how they evolve over time. We introduce TILA (Temporal Inversion-aware Learning and Alignment), a simple yet effective framework that uses temporal inversion, reversing image pairs, as a supervisory signal to enhance the sensitivity of existing temporal vision-language models to directional change. TILA integrates inversion-aware objectives across pretraining, fine-tuning, and inference, complementing conventional appearance modeling with explicit learning of temporal order. We also propose a unified evaluation protocol to assess order sensitivity and consistency under temporal inversion, and introduce MS-CXR-Tretrieval, a retrieval evaluation set constructed through a general protocol that can be applied to any temporal CXR dataset. Experiments on public datasets and real-world hospital cohorts demonstrate that TILA consistently improves progression classification and temporal embedding alignment when applied to multiple existing architectures.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

79,880 characters extracted from source content.

Expand or collapse full text

Temporal Inversion for Learning Interval Change in Chest X-Rays Hanbin Ko 1,2 Kyungmin Jeon 1,2 Doowoong Choi 3 Chang Min Park 3,4,5 1 Interdisciplinary Program in Bioengineering, Seoul National University Graduate School 2 Integrated Major in Innovative Medical Science, Seoul National University Graduate School 3 Dept. of Radiology, Seoul National University Hospital 4 Seoul National University College of Medicine 5 Inst. of Medical and Biological Engineering, Seoul National University Medical Research Center lucasko1994@snu.ac.kr, kmjeon@snu.ac.kr, justicejustice314@snu.ac.kr, morphius@snu.ac.kr Abstract Recent advances in vision–language pretraining have en- abled strong medical foundation models, yet most analyze radiographs in isolation, overlooking the key clinical task of comparing prior and current images to assess interval change. For chest radiographs (CXRs), capturing interval change is essential, as radiologists must evaluate not only the static appearance of findings but also how they evolve over time. We introduce TILA (Temporal Inversion-aware Learning and Alignment), a simple yet effective framework that uses temporal inversion, reversing image pairs, as a su- pervisory signal to enhance the sensitivity of existing tem- poral vision–language models to directional change. TILA integrates inversion-aware objectives across pretraining, fine-tuning, and inference, complementing conventional ap- pearance modeling with explicit learning of temporal or- der. We also propose a unified evaluation protocol to as- sess order sensitivity and consistency under temporal inver- sion, and introduce MS-CXR-T retrieval , a retrieval evalua- tion set constructed through a general protocol that can be applied to any temporal CXR dataset. Experiments on pub- lic datasets and real-world hospital cohorts demonstrate that TILA consistently improves progression classification and temporal embedding alignment when applied to multi- ple existing architectures. 1. Introduction Vision language pretraining (VLP) models such as CLIP [17] and SigLIP [26] have substantially advanced representation learning by aligning visual and textual in- formation, achieving strong performance in retrieval and zero-shot classification. However, extending these methods to medical imaging is nontrivial, as clinical data are het- erogeneous, reports are unstructured, and labeled image– text pairs are comparatively scarce [4, 12]. These factors limit the direct transfer of general-domain VLPs to robust medical-image understanding tasks. Recent medical VLPs have begun to bridge this gap by learning visual–textual alignment within the clinical do- main, enabling disease classification and localization [2, 7, 20, 21, 29]. Yet most existing models still analyze sin- gle CXRs in isolation, overlooking the core clinical task of comparing a patient’s current study with its most recent prior to assess interval change [30]. In routine interpre- tation, radiologists describe findings as improving, stable, worsening, resolving, or new, emphasizing that temporal comparison is fundamental to diagnostic reasoning [5]. Our focus aligns with standard radiology workflow, where each new examination is explicitly compared with the most re- cent prior. To address this need, recent studies have introduced multi-image encoders for paired CXR analysis. BioViL- T [1] incorporates prior images to represent temporal con- text, ALTA [13] adapts masked vision models for efficient inter-study alignment, and TempA-VLP [24] employs con- trastive objectives to learn dynamic changes across image pairs. Other approaches, including SDPL [31], CheXRel- Net [10], CheXRelFormer [15], and Med-ST [23], ex- plore fine-grained or architectural mechanisms to capture anatomical and temporal patterns. Beyond image encoders, several methods leverage historical reports for temporal rea- soning [6]; while effective, these approaches depend on prior text availability and are thus complementary to image- based modeling. Notably, even the coarser task of binary interval-change screening, determining whether any change has occurred between studies, carries substantial clinical value, as it can prioritize changed cases for detailed review and reduce the reporting burden in high-volume longitudi- nal settings [25]. However, even with these temporal encoders, evaluation remains limited. Most studies rely on a single progres- arXiv:2604.04563v2 [cs.CV] 7 Apr 2026 Figure 1. TILA Framework: Temporal Inversion-aware Learning and Alignment. The framework comprises three stages: pretraining, fine-tuning, and inference. (a–b) Pretraining: paired CXRs are encoded in both original and reversed orders; the Change-aware Sigmoid Loss aligns unchanged cases and separates changed ones. (c) Fine-tuning: the Bidirectional Cross-Entropy (BiCE) enforces label inversion, while the Temporal Consistency Loss (TCL) aligns probability distributions under reversal. (d) Inference: forward and reversed predictions are fused via inversion-aware scoring to enhance robustness and order consistency. sion label, offering little insight into whether models truly capture the direction of change. Directionality is central to diagnostic reasoning, yet progression labels are inher- ently noisy because temporal change is annotated without standardized criteria, making “stable” and borderline cases especially inconsistent. As a result, moderate classifica- tion scores do not necessarily indicate genuine temporal understanding and may reflect reliance on entity presence or chance agreement. To more directly probe directional reasoning, we examine model behavior under temporal in- version: by reversing the image order and swapping cor- responding labels, we can test whether predictions remain logically consistent across time directions. While such in- version is not always clinically symmetric, since not all recovery precisely mirrors worsening, many radiographic findings, particularly those involving changes in size, den- sity, or extent, exhibit approximately reversible visual pat- terns (e.g., effusion volume, pneumothorax size, consolida- tion extent). This makes temporal inversion a useful an- alytical stress test for order sensitivity, providing stronger evidence of temporal reasoning beyond label priors. To address these limitations, we propose TILA, a uni- fied framework that leverages temporal inversion as a su- pervisory signal for paired CXR analysis.TILA inte- grates inversion-aware objectives across pretraining, fine- tuning, and inference to learn directional and order-sensitive interval-change representations, and introduces MS-CXR- T retrieval for standardized evaluation of temporal reasoning. Our work makes the following contributions: • We present TILA, a unified framework that incorpo- rates temporal inversion throughout training and infer- ence to learn order-aware interval-change representations in paired CXRs. • We propose an inversion-aware evaluation protocol to quantify order sensitivity, inversion consistency, and bidirectional robustness, and introduce a reproducible pipeline for constructing progression-aware retrieval evaluation sets, instantiated as MS-CXR-T retrieval . • We demonstrate through extensive experiments that TILA improves progression classification and temporal embed- ding alignment across multiple architectures and datasets, indicating generality and temporal robustness. • We further show that TILA’s temporal representations transfer to binary interval-change screening, a clinically important triage task. 2. Method We present TILA, a general framework that introduces order-awareness into temporal CXR analysis.Rather than modifying network architectures, TILA employs lightweight loss functions and an inversion-aware infer- ence scheme that can be integrated into any paired-image vision–language backbone. We validate the framework on two representative temporal VLP models, BioViL-T [1] and ALTA [13], demonstrating that the same principles consis- tently enhance directional reasoning across distinct archi- tectures. The core idea is to leverage temporal inversion as a supervisory signal throughout training and inference, encouraging models to more explicitly capture directional change and strengthen sensitivity to temporal order. Formally, given a paired-image encoder f θ and a text encoder g φ , TILA introduces inversion-aware objectives at three stages: pretraining, fine-tuning, and inference, as illustrated in Fig. 1. 2.1. Pretraining: Temporal Inversion with SigLIP Objective. The goal of pretraining is to learn temporally aware image–text representations by leveraging inversion as a contrastive signal. TILA introduces a Change-aware Sig- moid loss, built upon the SigLIP formulation [26], which distinguishes whether a report describes interval change and whether new or resolved findings appear between exams. Thus, the alignment between an image pair and its associ- ated report depends on the presence, absence, or direction of temporal change. The key principle is that image–text alignment should behave differently depending on the underlying temporal relationship: • When the report indicates no change, both the original and inverted image pairs should align with the report, as the description remains valid regardless of temporal or- der. • When the report indicates a change, the inverted pair should not align because the temporal direction is re- versed and the described progression no longer holds true. This design enables the model to capture the directionality of progression and distinguish genuine temporal changes from stability. Even in clinically irreversible cases (e.g., post-surgical or fibrotic findings), the inverted pair should not align with the report, since the reversal no longer repre- sents a valid or clinically meaningful relationship between the images and the described change. ImageandTextEmbedding. Let B= (x prev i ,x cur i ,r i ,c i ) |B| i=1 denote a mini-batch of paired CXRs, their corresponding reports, and binary change labels c i ∈0, 1, where c i =1 indicates change and c i =0 indicates no change. The image encoder f and text encoder g produce: v i = f(x prev i ,x cur i ), v swap i = f(x cur i ,x prev i ), t i = g(r i ), where v i and v swap i represent embeddings from the origi- nal and inverted image orders, and t i is the report embed- ding. The binary label c i is automatically assigned using an LLM that compares each follow-up report with its corre- sponding prior to identify interval changes, detecting both progression-related phrases (e.g., “improved”, “worsened”) and appearance or resolution of findings (e.g., “new opac- ity”, “no longer seen”). Change-aware Sigmoid Loss. TILA builds upon the SigLIP pairwise sigmoid contrastive loss [26], which aligns matched image–text pairs while repelling unmatched ones using learnable temperature and bias terms. We extend this with a Change-aware Sigmoid Loss applied to inverted image pairs (x cur ,x prev ), reweighting alignments according to the change label: L change =− 1 |B| |B| X i=1 |B| X j=1 log σ z swap ij (τ swap v swap⊤ i t j −b swap ) , (1) with z swap ij = ( +1, (i=j) & (c i =0), −1, (i̸=j) or (c i =1), and learnable parameters τ swap , b swap . The key difference from standard SigLIP is thatz swap ij =+1 only when the exam- ple matches its own report (i=j) and the case is no change (c i =0); all other pairs—including matched pairs describing change—are treated as negatives. Total Loss. The overall pretraining objective integrates the base and change-aware contrastive terms: L total =L siglip + W L change ,(2) where L siglip is the base pairwise sigmoid loss applied to original-order pairs [26] and W controls the relative contri- bution of the change-aware component. This joint objective yields direction-sensitive representations: stable cases align in both orders, whereas changed cases are discouraged from aligning under reversal. 2.2. Fine-tuning: Enforcing Temporal Inversion Awareness Objective. Fine-tuning enforces temporal consistency during interval-change classification. Each paired CXR is labeled as improved, stable, or worsened. When the temporal order is reversed, the label should invert accord- ingly (e.g., improved ↔ worsened), while stable remains unchanged. Bidirectional Cross-Entropy Loss (BiCE). Let y ∈ improved,stable,worsened denote the ground- truth progression label for a paired input (x prev ,x cur ). We define the inversion mappingI(y) as I(y) =      worsened, if y = improved, stable,if y = stable, improved, if y = worsened. The Bidirectional Cross-Entropy loss averages the classifi- cation losses from the forward and reversed orders: L BiCE = 1 2 h CE f θ (x prev ,x cur ),y + CE f θ (x cur ,x prev ),I(y) i . (3) This enforces label-level inversion consistency, treating progression as a directional continuum rather than a set of independent classes. Temporal Consistency Loss (TCL). While BiCE en- forces label-level symmetry, the Temporal Consistency Loss ensures that the probability distributions themselves transform consistently to inversion. Let p fwd and p bwd de- note the predicted class probabilities for the forward and reversed orders, respectively.We define a transforma- tion S that swaps probabilities between improved and worsened while keeping stable fixed: S(p)[:, 0] = p[:, 2], S(p)[:, 1] = p[:, 1], S(p)[:, 2] = p[:, 0]. (4) The TCL objective is then L TCL = 1 |B| |B| X i=1 p (i) fwd −S(p (i) bwd ) 2 .(5) This encourages the model to assign mirrored probabilities under inversion, enforcing probability-level symmetry be- tween forward and reversed predictions. For instance, a case predicted as highly improved in the forward order should correspondingly yield a high worsened probabil- ity in the reversed order. Total Fine-tuning Loss. The final fine-tuning objective combines classification and probabilistic consistency: L total =L BiCE + λL TCL ,(6) where λ balances accuracy and temporal coherence. Jointly optimizing BiCE and TCL enables the model to maintain order-aware and temporally coherent predictions across all interval-change categories. 2.3. Inference: Inversion-aware Scoring Objective. During inference, TILA enhances robustness by aggregating predictions from both temporal directions. Given a paired input (x prev ,x cur ), the final prediction is obtained by averaging the forward prediction with the inversion-adjusted reversed prediction: score = 1 2 p f θ (x prev ,x cur ) +S p(f θ (x cur ,x prev )) , (7) where p(·) denotes the predicted class probabilities, and S swaps the probabilities for improved and worsened while keeping stable unchanged. This inversion-aware fusion reduces order bias and pre- diction variance, producing more consistent and temporally robust interval-change assessments. 3. Experiments We comprehensively evaluate the proposed TILA frame- work across multiple temporal vision–language tasks, including retrieval, zero-shot classification, supervised fine-tuning, and interval-change screening, benchmarking against state-of-the-art temporal VLPs and our retrained baselines. We also introduce MS-CXR-T retrieval , a evalu- ation set constructed through a reproducible pipeline, and evaluate generalizability on external datasets, including CheXpert [3], RexGradient [28], and a private cohort. 3.1. Model Implementation TILA is implemented on two representative temporal vision–language backbones, BioViL-T [1] and ALTA [13]. To ensure fair comparison and improved batch efficiency, we retrain both base architectures using the SigLIP loss [26] instead of the original CLIP loss, following the same pre- training settings reported in their respective works. This re- trained variant (SigLIP) serves as our baseline, upon which TILA introduces inversion-aware objectives without mod- ifying the underlying architectures. The text encoder g φ is initialized from pretrained CXR-BERT [1]. Training pro- ceeds in two stages, with inversion-aware objectives acti- vated after an initial warm-up phase; introducing them from the start can cause the model to collapse toward predicting all cases as stable, since this label is trivially consistent under inversion: • Pretraining: Models are trained for 10 epochs using the standard SigLIP loss, followed by 20 epochs adding the Change-aware Sigmoid Loss (30 epochs total). We set W = 1 (equal weighting for the change-aware term), use a learning rate of 1× 10 −4 , batch size 144, and the AdamW optimizer. • Fine-tuning: For progression classification, models are trained for 20 epochs with the BiCE loss and 30 additional epochs with the Temporal Consistency Loss (50 epochs total). In full fine-tuning, all parameters are updated. We set λ = 50 to balance the scale between the Temporal Consistency and cross-entropy losses. AdamW is used for the optimizer. Full hyperparameter and training details are provided in Appendix Sec. 6. 3.2. Dataset Description Training Dataset. We train primarily on MIMIC- CXR [8], excluding images without prior counterparts and Table 1. Single vs. paired-image retrieval on MIMIC and CheXpert. Retrieval performance (Recall@k for I2T/T2I) and TEM score. Single-image models process only the current image, whereas paired-image models encode (prev, cur). For each backbone, SigLIP denotes retraining with the sigmoid contrastive loss (baseline), and TILA adds the Change-aware Sigmoid Loss. CheXpert results are reported as mean and 95% CI across ten random subsamples of 3,000 paired CXRs. Model MIMIC RetrievalCheXpert Retrieval I2T2ITEMI2T2ITEM @1@5@10@1@5@10@1@5@10@1@5@10 Single-image VLPs Ko & Park. [11]6.623.233.88.424.534.77.46.2± 0.1 16.1± 0.2 24.7± 0.1 6.6± 0.2 15.9± 0.2 23.0± 0.25.7± 0.4 BioViL-T (single image) [1]3.110.717.93.711.318.43.52.4± 0.18.9± 0.2 14.8± 0.2 2.6± 0.27.6± 0.1 13.2± 0.14.2± 0.3 ALTA (single image) [13]6.518.227.211.028.741.3 9.06.3± 0.1 17.9± 0.2 26.1± 0.2 8.2± 0.2 22.7± 0.1 32.1± 0.110.2± 0.3 BiomedCLIP [27]0.42.03.10.62.33.21.60.2± 0.11.7± 0.22.8± 0.30.6± 0.22.2± 0.23.3± 0.31.9± 0.3 Paired-image VLPs BioViL-T [1]5.217.125.15.516.324.012.13.1± 0.19.8± 0.1 15.2± 0.1 2.8± 0.19.2± 0.2 14.2± 0.113.7± 0.2 BioViL-T SigLIP 11.435.146.111.834.246.915.711.7± 0.231.0± 0.340.6± 0.2 10.1± 0.126.4± 0.235.8± 0.218.5± 0.2 BioViL-T TILA 12.834.845.412.033.047.317.112.5± 0.2 30.2± 0.3 40.2± 0.310.7± 0.2 26.3± 0.3 35.4± 0.220.8± 0.1 ALTA [13]9.124.234.911.832.143.811.68.5± 0.2 23.3± 0.1 32.8± 0.2 9.2± 0.2 24.6± 0.2 34.5± 0.117.6± 0.1 ALTA SigLIP 13.634.946.713.434.948.014.78.4± 0.3 23.4± 0.3 33.4± 0.2 8.2± 0.3 23.8± 0.1 34.9± 0.117.8± 0.2 ALTA TILA 13.834.445.813.834.147.216.08.7± 0.2 24.2± 0.2 34.1± 0.2 9.3± 0.2 24.9± 0.2 35.2± 0.219.2± 0.2 those overlapping with MS-CXR-T [1]. Binary change/no- change labels for pretraining are generated using Gem- ini 2.0 Flash [19]. For classification, we use progression labels from Chest ImaGenome [22] to perform 3-class tem- poral classification, following official splits for consistency across training, validation, and testing. MS-CXR-T Dataset. MS-CXR-T contains paired CXRs from MIMIC annotated by radiologists as improved, stable, or worsened. It serves as the main benchmark for interval-change modeling. MS-CXR-T retrieval Construction Protocol. We describe a general pipeline for constructing progression-aware re- trieval evaluation sets from any temporal CXR dataset with annotated findings. Given a paired CXR case, the pipeline reformulates its report into a set of directional variants for a specific finding, while neutralizing unrelated find- ings to isolate the target progression direction. For exam- ple, for a “stable” pneumothorax case with the original re- port “Right pneumothorax is unchanged. Consolidation is worsening.”, the pipeline generates three directional vari- ants: “Right pneumothorax is improved. Consolidation is present.”, “Right pneumothorax is unchanged. Consolida- tion is present.”, and “Right pneumothorax is worsened. Consolidation is present.” This process isolates the target finding and its progression direction while preserving the natural co-occurrence of other findings, enabling realistic yet controlled evaluation of temporal reasoning. We in- stantiate this protocol on MS-CXR-T to produce MS-CXR- T retrieval . External Validation Sets. We evaluate generalization on three external datasets: (i) paired CXRs from CheXpert [3] for retrieval and temporal embedding alignment, (i) a pri- vate tertiary hospital cohort annotated using correspond- ing radiology reports, covering four findings for progres- sion classification and binary interval-change screening, and (i) paired CXRs from RexGradient [28] for binary interval-change screening. Detailed data curation and la- beling protocols along with data ethics are provided in Ap- pendix Sec. 7, and the binary screening task setup is de- scribed in Appendix Sec. 8. 3.3. Evaluation Protocol We evaluate retrieval performance using Recall@k for image-to-text (I2T) and text-to-image (T2I) tasks, and re- port the Temporal Embedding Matching (TEM) score [1], computed as the F1 overlap of temporal terms between the reference and retrieved reports. For progression classification (zero-shot and fully fine- tuned), we adopt four complementary evaluation settings: • Standard: Macro-accuracy on original image pairs us- ing the ground-truth labels improved, stable, and worsened. This serves as the primary evaluation re- flecting clinical interpretation. • Reversed: Accuracy after reversing image order and cor- responding labels, used to assess the model’s directional sensitivity. • Combined: Average of forward and reversed predictions using inversion-aware scoring (Sec. 2.3), providing a ro- bustness check through bidirectional reasoning. • Consistency: A prediction is considered correct only if the model outputs the correct label in both directions, in- dicating temporal reliability and reducing chance agree- ment. While the Standard setting represents the clinically rel- evant metric, the additional evaluations (Reversed, Com- bined, and Consistency) serve as analytical tools to quan- tify order sensitivity and stability.The clinical utility of these extended evaluations is further discussed in Ap- pendix 10.1. Table 2. Interval-change classification results on MS-CXR-T and a private cohort. We report macro-accuracy (%) for five find- ings—CON: consolidation, PE: pleural effusion, PNE: pneumonia, PTX: pneumothorax, and EDE: edema—together with the average across findings for each evaluation protocol (Standard, Reversed, Combined, Consistency; see Section 3.3). For retrieval and zero-shot classification, TILA models use only the Change-aware Sigmoid pretraining loss, while in supervised settings they additionally apply inversion-aware fine-tuning (BiCE and TCL). SigLIP denotes our re-trained baselines using the sigmoid contrastive loss, and TILA refers to the corresponding models augmented with inversion-aware objectives. ModelStandardReversedCombinedConsistency CONPEPNEPTXEDEAvgCONPEPNEPTXEDEAvgCONPEPNEPTXEDEAvgCONPEPNEPTXEDEAvg MS-CXR-T retrieval BioViL-T [1]49.142.838.333.151.843.142.851.745.136.248.244.853.754.350.837.159.851.223.421.921.710.129.221.3 BioViL-T SigLIP 48.958.355.634.453.650.251.753.261.038.150.751.055.660.064.539.058.155.532.835.340.114.241.032.7 BioViL-T TILA 55.358.758.140.358.0 54.155.959.061.537.056.554.058.862.465.946.562.559.339.839.943.022.844.037.9 ALTA [13]46.833.936.031.140.437.744.247.443.137.941.542.954.054.460.935.451.351.218.411.09.34.710.510.8 ALTA SigLIP 45.743.142.729.145.441.245.247.846.735.544.043.954.555.859.333.253.151.221.923.425.57.631.522.0 ALTA TILA 49.245.945.535.446.444.549.850.248.938.150.747.656.557.262.342.155.854.830.326.426.815.238.627.5 Zero-shot (MS-CXR-T) BioViL-T [1]34.937.434.537.042.837.437.338.641.033.142.038.448.653.453.834.854.149.010.510.710.99.916.811.8 BioViL-T SigLIP 42.750.142.729.155.444.047.243.849.735.551.045.546.155.851.333.258.849.121.923.427.57.639.524.0 BioViL-T TILA 46.546.351.938.852.347.246.246.349.738.652.446.756.758.059.142.260.055.227.424.630.833.236.130.5 ALTA [13]44.641.541.832.744.441.043.744.050.838.740.343.550.248.553.436.856.249.120.419.721.19.017.717.6 ALTA SigLIP 45.245.845.134.149.644.045.448.248.838.545.245.351.854.956.838.256.151.621.420.818.210.320.618.3 ALTA TILA 45.646.750.139.150.146.446.150.152.139.251.347.854.257.658.943.259.454.726.423.828.229.034.628.4 Supervised (MS-CXR-T) CheXRelFormer [15]51.156.037.134.159.047.542.550.736.034.658.844.640.854.244.733.356.445.928.049.210.729.552.334.0 CNN-TF [1]46.655.241.335.261.2 47.937.839.432.122.747.135.948.357.144.037.960.149.527.425.227.614.034.125.7 CNN-TF TILA 48.655.842.937.560.749.143.649.738.537.661.746.352.257.849.239.762.452.336.140.544.226.652.840.1 MLRG [14]61.365.761.744.867.560.251.355.250.237.163.151.459.164.764.645.567.560.338.249.344.725.258.143.1 BioViL-T [1]58.165.764.742.968.960.147.562.353.536.964.953.159.264.266.344.568.460.638.246.646.822.756.342.2 BioViL-T SigLIP 61.367.564.243.069.361.152.661.456.132.663.853.360.465.663.042.267.059.741.340.745.216.653.639.5 BioViL-T TILA 63.366.767.152.370.964.164.565.866.651.270.363.763.766.967.651.068.863.660.258.357.843.167.457.4 ALTA [13]62.266.262.545.569.261.253.457.149.837.365.352.660.265.665.846.266.861.039.249.645.424.357.243.2 ALTA SigLIP 63.265.463.446.070.161.752.455.952.138.066.353.061.764.566.244.968.561.235.748.447.123.060.142.9 ALTA TILA 64.168.366.249.370.163.660.562.358.345.866.758.862.366.167.648.468.462.655.457.958.138.762.954.6 Supervised (Private) CNN-TF [1]41.047.535.931.2-38.933.840.430.631.1-34.043.749.036.132.3-40.321.922.613.812.9-17.8 MLRG [14]56.259.358.852.1-56.644.449.249.456.2-49.858.257.152.155.2-55.725.329.229.824.2-27.2 CheXRelFormer [15]47.948.441.338.9-44.241.849.236.233.0-40.149.748.341.734.6-43.626.543.419.026.2-28.8 BioViL-T [1]55.057.160.349.7-55.644.848.150.253.4-49.257.656.754.553.7-55.729.834.633.325.1-30.7 BioViL-T SigLIP 53.959.351.356.6-55.345.653.148.848.1-48.956.059.354.952.0-55.630.435.634.723.7-31.1 BioViL-T TILA 67.970.263.063.3-66.161.565.755.262.3-61.264.766.858.362.2-63.058.455.253.451.4-54.6 ALTA [13]59.261.361.548.5-57.745.147.651.344.9-47.358.556.255.951.1-55.539.249.645.424.3-39.7 ALTA SigLIP 61.160.762.251.3-58.952.250.253.244.3-50.061.558.957.452.6-57.645.751.347.426.0-42.6 ALTA TILA 65.868.364.260.1-64.660.264.255.856.3-59.262.565.959.259.3-61.855.753.354.248.8-53.0 4. Results 4.1. Experimental Results Retrieval Performance. As shown in Tab. 1, TILA yields comparable Recall to the base models while consistently improving TEM scores on both MIMIC and CheXpert, demonstrating that temporal inversion supervision enhances temporal alignment without compromising retrieval accu- racy. Single-image baselines achieve similar Recall but substantially lower TEM (e.g., BioViL-T: 12.1 vs. 3.5 for the single-image), confirming that paired encoding provides critical temporal context that TILA further strengthens. Interval-Change Retrieval Performance. On MS-CXR- T retrieval (Tab. 2), TILA improves the average Standard accuracy over SigLIP baselines for both backbones and yields consistent gains across the Reversed, Combined, and Consistency protocols, with particularly evident improve- ments for consolidation and pneumothorax. The inversion- aware combined scoring strategy provides modest addi- tional boosts, likely by mitigating label ambiguities; this effect is examined further in the score analysis. Zero-Shot and Supervised Performance. Across both zero-shot and fully supervised settings (Tab. 2), TILA con- sistently improves Standard Avg accuracy over the SigLIP baselines for both BioViL-T and ALTA. The gains are more pronounced in direction-sensitive protocols, with no- tably higher Consistency averages (e.g., zero-shot BioViL- T: 24.0 % → 30.5 %; supervised: 39.5 % → 57.4 %), in- dicating that inversion-aware supervision strengthens tem- poral coherence while maintaining or improving standard classification accuracy across architectures. Figure 2. Score Distribution Analysis for Pleural Effusion (MS-CXR-T). Distribution of prediction scores (scaled by 10) for each pleural effusion progression label, comparing baseline and TILA models in the zero-shot setting. Each boxplot separates cases by progression label (improved, stable, worsened) and label quality (consensus vs. disagreement), shown for both standard and inversion-aware (combined) scoring. Key trends illustrated: i) Changes in scores for each label (improved, stable, worsened). i) Impact of applying inversion-aware scoring (combined). i) Differences in scores based on label quality (consensus vs. disagreement). External Validation. On the external validation set from a private tertiary hospital (Tab. 2), the Avg columns show that models trained with TILA outperform their SigLIP baselines across Standard, Reversed, Combined, and Con- sistency protocols. These results suggest that inversion- aware supervision generalizes effectively beyond the train- ing domain, contributing to more robust and temporally consistent performance in real-world clinical data. 4.2. Ablation Table 3. Ablation of temporal inversion components. Each cell reports the mean macro-accuracy (%) across the five findings (CON, PE, PNE, PTX, EDE) for the indicated evaluation protocol. VariantStandardReversedCombinedConsistency BioViL-T (CLIP)60.153.060.542.1 + SigLIP (Baseline)61.153.359.639.5 + Change-aware SigLIP (Ours)64.658.363.445.4 + BiCE65.262.063.453.1 + BiCE + TCL64.163.763.657.4 We perform ablation studies using BioViL-T as the back- bone to isolate the contribution of each temporal inversion component. While the effects of inversion-aware pretrain- ing on retrieval and zero-shot performance are shown in Ta- bles 1 and 2, here we focus on the supervised setting, where inversion-aware fine-tuning can be applied in addition to the Change-aware Sigmoid pretraining (Tab. 3). Replacing the CLIP objective with the SigLIP loss pro- duces modest shifts in the averaged metrics: Standard ac- curacy increases slightly (60.1 % → 61.1 %), while Com- bined and Consistency averages change minimally. Adding the Change-aware Sigmoid pretraining yields a larger im- provement across all four protocols (Standard: 64.6 %; Re- versed: 58.3 %; Combined: 63.4 %; Consistency: 45.4 %), indicating that inversion at pretraining provides a stronger initialization for downstream temporal classification. Building on this initialization, incorporating BiCE dur- ing fine-tuning primarily boosts direction-sensitive metrics, raising the Reversed and Consistency averages while main- taining the Combined average and slightly improving the Standard average. Adding TCL on top trades a small reduc- tion in Standard for further gains in Reversed, Combined, and especially Consistency, reinforcing temporal symmetry at the probability level. Additional hyperparameter analysis is provided in Ap- pendix 9.2. 4.3. Score Analysis To examine how TILA affects model behavior be- yond classification accuracy, we analyze effusion pre- dictions from MS-CXR-T with respect to label qual- ity.Labels are categorized as multiple experts, one-expert, or disagreement.A multiple experts label indicates consensus among radiologists, whereas disagreement reflects variability in which one reader may assign worsened or improved while another assigns stable. These categories help contextualize ex- pected uncertainty in the ground truth. As shown in Fig. 2, inversion-aware scoring (Sec. 2.3) consistently shifts predictions toward the correct label, im- proving separation between classes and pushing incorrect scores further away. This effect is most pronounced in consensus cases, suggesting that integrating both tempo- ral directions reduces ambiguity and sharpens calibration. For cases with disagreement, TILA produces smaller shifts, reflecting more conservative behavior when the un- derlying labels are uncertain—mirroring how radiologists temper confidence in ambiguous settings. A notable improvement appears in distinguishing stable from improved and worsened. The base- line model frequently assigns low confidence to stable, causing overlap between progression categories.TILA widens this margin through inversion-aware scoring, yield- ing more reliable and well-separated stability scores. Clin- ically, stronger discrimination of stable cases is important for reducing unnecessary interventions and supporting more dependable longitudinal assessment. Table 4. Binary interval-change screening on two external test sets. AUC is reported for fine-tuning and linear probing modes. ModelMode Test Set AUC PrivateRexGradient BioViL-TFine-tune0.7340.565 BioViL-TLinear probe0.6930.545 BioViL-T TILA Fine-tune0.7650.702 BioViL-T TILA Linear probe0.7560.681 4.4. Binary Interval-Change Screening We evaluate whether TILA’s representations generalize to binary change detection, determining whether any interval change occurred (Tab. 4). BioViL-T TILA improves AUC over the baseline on both test sets and in both evaluation modes. In fine-tuning mode, TILA increases AUC from 0.734 to 0.765 on the Private test set and from 0.565 to 0.702 on RexGradient, demonstrating consistently better performance than the baseline across external datasets. 5. Discussion and Conclusion Modeling interval change in paired CXRs remains a chal- lenging problem.We introduced TILA, a framework that incorporates temporal inversion into pretraining, fine- tuning, and inference to enhance direction-sensitive reason- ing. Across retrieval, zero-shot, and supervised evaluations, TILA improves temporal alignment metrics such as TEM and order-consistency metrics, while also yielding higher Standard accuracy for most findings and settings. The inversion-aware scoring strategy further stabilizes predic- tions, and TILA’s modular design enables seamless integra- tion with existing vision–language pipelines. Limitations. Inter-reader variability remains a fundamen- tal challenge in interval-change labeling, particularly for subtle or borderline changes [9, 16, 18]. A further diffi- culty is the lack of unified criteria for defining progression across different findings, as visual manifestations of change vary widely and are not consistently standardized in clini- cal practice. Developing consensus definitions, along with curated multi-expert labels, would provide more reliable ground truth and strengthen the robustness of future tem- poral modeling approaches. Asymmetry of recovery. Temporal inversion is not meant to represent a clinically symmetric reversal of disease, and its validity differs across TILA’s stages. During pretrain- ing, swapped pairs with change are used only as negative supervision, making the objective valid regardless of clini- cal reversibility. For fine-tuning and inference, label inver- sion assumes improved ↔ worsened symmetry, which is not universally met; cautious interpretation is warranted for findings with inherently asymmetric trajectories. Nonethe- less, inversion-aware supervision improves Standard accu- racy, the clinically relevant metric, for most findings, indi- cating that it reinforces rather than disrupts forward-order predictions. Conclusion. TILA provides a simple,architecture- agnostic approach for enhancing temporal sensitivity in paired CXR interpretation. By leveraging temporal inver- sion across training and inference, the framework improves directional alignment and order stability while maintaining or improving Standard accuracy in most settings. TILA of- fers a principled building block for future temporal vision– language models and may support more reliable interval- change assessment when combined with standardized labels and richer temporal context. Acknowledgements We thank H.Y. Cho and K.J. Cho for their valuable advice on this work. This research was supported by a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (grant number: RS-2025-02213531, NTIS number: 2460004432), and by the National Research Foundation of Korea (NRF) grants funded by the Ministry of Science and ICT (MSIT) (Grant No. RS-2024-00354666). References [1] Shruthi Bannur, Stephanie Hyland, Qianchu Liu, Fernando Perez-Garcia, Maximilian Ilse, Daniel C Castro, Benedikt Boecking, Harshita Sharma, Kenza Bouzid, Anja Thieme, et al. Learning to exploit temporal structure for biomed- ical vision-language processing.In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15016–15027, 2023. 1, 2, 4, 5, 6 [2] Benedikt Boecking, Naoto Usuyama, Shruthi Bannur, Daniel C Castro, Anton Schwaighofer, Stephanie Hyland, Maria Wetscherek, Tristan Naumann, Aditya Nori, Javier Alvarez-Valle, et al. Making the most of text semantics to improve biomedical vision–language processing. In Euro- pean conference on computer vision, pages 1–21. Springer, 2022. 1 [3] Pierre Chambon, Jean-Benoit Delbrouck, Thomas Sounack, Shih-Cheng Huang, Zhihong Chen, Maya Varma, Steven QH Truong, Chu The Chuong, and Curtis P Langlotz. Chexpert plus: Hundreds of thousands of aligned radiology texts, im- ages and patients. arXiv preprint arXiv:2405.19538, 2024. 4, 5 [4] Yifan Du, Zikang Liu, Junyi Li, and Wayne Xin Zhao. A survey of vision-language pre-trained models. arXiv preprint arXiv:2202.10936, 2022. 1 [5] Wenjun Hou, Yi Cheng, Kaishuai Xu, Wenjie Li, and Jiang Liu. Recap: Towards precise radiology report generation via dynamic disease progression reasoning. arXiv preprint arXiv:2310.13864, 2023. 1 [6] Haoxu Huang, Cem M Deniz, Kyunghyun Cho, Sumit Chopra, and Divyam Madaan. Hist-aid: Leveraging histori- cal patient reports for enhanced multi-modal automatic diag- nosis. arXiv preprint arXiv:2411.10684, 2024. 1 [7] Shih-Cheng Huang, Liyue Shen, Matthew P Lungren, and Serena Yeung. Gloria: A multimodal global-local represen- tation learning framework for label-efficient medical image recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3942–3951, 2021. 1 [8] Alistair EW Johnson, Tom J Pollard, Lu Shen, Li-wei H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G Mark. Mimic-i, a freely accessible critical care database. Scientific data, 3(1):1–9, 2016. 4 [9] Cholmin Kang, Chunggi Lee, Heon Song, Minuk Ma, and S ́ ergio Pereira. Variability matters: Evaluating inter-rater variability in histopathology for robust cell detection. In European Conference on Computer Vision, pages 552–565. Springer, 2022. 8 [10] Gaurang Karwande, Amarachi B Mbakwe, Joy T Wu, Leo A Celi, Mehdi Moradi, and Ismini Lourentzou. Chexrelnet: An anatomy-aware model for tracking longitudinal relationships between chest x-rays. In International Conference on Med- ical Image Computing and Computer-Assisted Intervention, pages 581–591. Springer, 2022. 1 [11] Hanbin Ko and Chang-Min Park. Bringing clip to the clinic: Dynamic soft labels and negation-aware learning for medical analysis. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 25897–25906, 2025. 5 [12] Hanbin Ko, Gihun Cho, Inhyeok Baek, Donguk Kim, Joon- beom Koo, Changi Kim, Dongheon Lee, and Chang Min Park.Exploring the capabilities of llm encoders for image-text retrieval in chest x-rays.arXiv preprint arXiv:2509.15234, 2025. 1 [13] Chenyu Lian, Hong-Yu Zhou, Dongyun Liang, Jing Qin, and Liansheng Wang. Efficient medical vision-language align- ment through adapting masked vision models. IEEE Trans- actions on Medical Imaging, 2025. 1, 2, 4, 5, 6 [14] Kang Liu, Zhuoqi Ma, Xiaolu Kang, Yunan Li, Kun Xie, Zhicheng Jiao, and Qiguang Miao. Enhanced contrastive learning with multi-view longitudinal data for chest x-ray re- port generation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 10348–10359, 2025. 6 [15] Amarachi B Mbakwe, Lyuyang Wang, Mehdi Moradi, and Ismini Lourentzou. Hierarchical vision transformers for dis- ease progression detection in chest x-ray images. In In- ternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 685–695. Springer, 2023. 1, 6 [16] Brennan Nichyporuk, Jillian Cardinell, Justin Szeto, Raghav Mehta, Jean-Pierre R Falet, Douglas L Arnold, Sotirios A Tsaftaris, and Tal Arbel. Rethinking generalization: The impact of annotation style on medical image segmentation. arXiv preprint arXiv:2210.17398, 2022. 8 [17] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, pages 8748–8763. PMLR, 2021. 1 [18] Marcel P Schilling, Tim Scherr, Friedrich R M ̈ unke, Oliver Neumann, Mark Schutera, Ralf Mikut, and Markus Reischl. Automated annotator variability inspection for biomedical image segmentation. IEEE access, 10:2753–2765, 2022. 8 [19] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean- Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5 [20] Fuying Wang, Yuyin Zhou, Shujun Wang, Varut Vardhanab- huti, and Lequan Yu. Multi-granularity cross-modal align- ment for generalized medical visual representation learn- ing. Advances in Neural Information Processing Systems, 35:33536–33549, 2022. 1 [21] Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medi- cal images and text. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. Con- ference on Empirical Methods in Natural Language Process- ing, page 3876, 2022. 1 [22] Joy T Wu, Nkechinyere N Agu, Ismini Lourentzou, Arjun Sharma, Joseph A Paguio, Jasper S Yao, Edward C Dee, William Mitchell, Satyananda Kashyap, Andrea Giovannini, et al. Chest imagenome dataset for clinical reasoning. arXiv preprint arXiv:2108.00316, 2021. 5 [23] Jinxia Yang, Bing Su, Wayne Xin Zhao, and Ji-Rong Wen. Unlocking the power of spatial and temporal infor- mation in medical multimodal pre-training. arXiv preprint arXiv:2405.19654, 2024. 1 [24] Zhuoyi Yang and Liyue Shen. Tempa-vlp: Temporal-aware vision-language pretraining for longitudinal exploration in chest x-ray image. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4625– 4634. IEEE, 2025. 1 [25] Jihye Yun, Yura Ahn, Kyungjin Cho, Sang Young Oh, Sang Min Lee, Namkug Kim, and Joon Beom Seo. Deep learning for automated triaging of stable chest radiographs in a follow-up setting. Radiology, 309(1):e230606, 2023. 1 [26] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975–11986, 2023. 1, 3, 4 [27] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. A multimodal biomedical foun- dation model trained from fifteen million image–text pairs. NEJM AI, 2(1):AIoa2400640, 2025. 5 [28] Xiaoman Zhang, Juli ́ an N Acosta, Josh Miller, Ouwen Huang, and Pranav Rajpurkar. Rexgradient-160k: A large- scale publicly available dataset of chest radiographs with free-text reports. arXiv preprint arXiv:2505.00228, 2025. 4, 5 [29] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D Manning, and Curtis P Langlotz. Contrastive learning of medical visual representations from paired images and text. In Machine Learning for Healthcare Conference, pages 2– 25. PMLR, 2022. 1 [30] Qingqing Zhu, Tejas Sudharshan Mathai, Pritam Mukherjee, Yifan Peng, Ronald M Summers, and Zhiyong Lu. Utiliz- ing longitudinal chest x-rays and reports to pre-fill radiology reports. In International Conference on Medical Image Com- puting and Computer-Assisted Intervention, pages 189–198. Springer, 2023. 1 [31] Ye Zhu, Jingwen Xu, Fei Lyu, and Pong C Yuen. Symp- tom disentanglement in chest x-ray images for fine-grained progression learning. In International Conference on Med- ical Image Computing and Computer-Assisted Intervention, pages 598–607. Springer, 2024. 1 Temporal Inversion for Learning Interval Change in Chest X-Rays Supplementary Material 6. Model This section provides additional details on implementation, augmentations, hyperparameters, and model selection. 6.1. Implementation Details For pretraining, we use the AdamW optimizer with a co- sine learning-rate schedule and a 100-step warm-up. The learning rate is set to 1 × 10 −4 with a batch size of 144, and all parameters are trained in bfloat16. Pretraining is conducted on three NVIDIA A6000 GPUs for 30 epochs (approximately 54 GPU-hours). Fine-tuning also uses AdamW with cosine scheduling and a 5% warm-up. Models are trained end-to-end for 50 epochs using a learning rate of 1× 10 −5 and a batch size of 128. These experiments are performed on a single NVIDIA A6000 GPU. Projection layers for both image and text en- coders have dimension 128, and the CXR-BERT text en- coder is configured with a maximum token length of 256. 6.2. Augmentation Image augmentations follow the protocols described in BioViL-T [1] and ALTA [13] and are applied consistently across both pretraining and fine-tuning. 6.3. Hyperparameters Logit scales τ and τ swap are initialized to log 10, with bias −10 (SigLIP’s logit-shift parameter). We set W = 1 for the Change-aware Sigmoid loss during pretraining and choose λ = 50 for fine-tuning to balance the magnitude of the Tem- poral Consistency Loss relative to the cross-entropy term. The Change-aware Sigmoid loss is activated after 10 pre- training epochs, and TCL is applied after 20 fine-tuning epochs to prevent premature convergence toward predicting stable for most pairs. 6.4. Compared Models We compare only to models with publicly available code to ensure consistent training and reproducible evaluation. The CNN+Transformer baseline follows the BioViL-T [1] implementation with ImageNet initialization.Methods without publicly available code or reproducible training pipelines were excluded. Temporal CXR baselines remain limited relative to single-image CXR methods. Few existing approaches ex- plicitly model interval change, and the field lacks standard- ized benchmarks, unified data splits, and publicly released training workflows. To avoid data leakage and ensure fair comparison, we include only models we can reproduce end- to-end and evaluate under our inversion-aware protocol. Where available, we also include models with released pre- trained weights, even if full training code is not provided. Recent studies similarly compare against a narrow set of baselines and often rely on reported numbers rather than independent reproduction. Given our focus on temporal ro- bustness—not only raw accuracy—and the need to strictly avoid any MS-CXR-T overlap, controlled and fully repro- ducible evaluation was essential. 7. Data This section describes dataset splits, label generation pro- tocols, and ethical considerations, including specific details on the MS-CXR-T retrieval benchmark. 7.1. Dataset Splits We first exclude chest X-ray images without available prior images. The sample counts for each split are presented in Tab. 5. For CheXpert, since the official validation and test splits lack corresponding reports or prior images, we use the CheXpert training split as an external pool for retrieval evaluation. To minimize sampling bias, we repeatedly sam- ple 3,000 image pairs from this pool 10 times and report the mean and 95% confidence interval across these subsam- ples. For our private dataset, we collected CXR pairs from 2010–2020, filtering for reports and images containing tem- poral keywords (e.g., “improved”, “worsened”, “stable”). Labeling was performed by a researcher with four years of experience in chest X-ray interpretation, and these labels were used as the reference for external validation during fine-tuning. TrainValidationTest MIMIC183,3021,3302,871 CheXpert--123,374 Private--2,233 Table 5. Data Distribution for MIMIC, CheXpert, and Private 7.2. Dataset Approvals and Ethics All LLM-assisted label extraction for MIMIC and the con- struction of MS-CXR-T retrieval were conducted in accor- dance with PhysioNet guidelines for responsible LLM us- age (https://physionet.org/news/post/gpt-responsible-use). All preprocessing for MS-CXR-T retrieval was performed prior to September 2025, during a period in which de- identified report processing on Google Cloud Vertex AI was permitted under the platform’s data-handling policies. The private clinical dataset was collected under institu- tional IRB approval, and all researchers accessing the data were formally registered, credentialed, and authorized for handling de-identified records. 7.3. Change Label Generation Change/no-change labels for pretraining were generated us- ing Gemini 2.0 Flash. The model compares each follow- up report with its corresponding prior (when available) and identifies interval changes, including improved/stable/wors- ened and new/resolved findings. For compatibility with the Change-aware Sigmoid Loss, we invert the binary output (1→0, 0→1), and exclude uncertain cases (“-1”) from train- ing. Sampled-scale validation of LLM choices. To verify that Gemini is not strictly required and to assess the re- liability of alternative LLMs, we conducted a small-scale validation on 200 randomly sampled report pairs. For each pair, we validated the LLM-derived change/no-change la- bel against a manually reviewed label provided by a trained annotator. Agreement rates (with 95% Wilson confidence intervals) were: • Gemini 2.0 Flash: 192/200 (96.0%; 92.3–97.9) • Qwen3-14B: 179/200 (89.5%; 84.5–93.1) These results show that smaller LLMs can also produce reasonable labels, although Gemini exhibited the highest agreement in this pilot study. The prompt used for label generation is provided below: You are given two chest X-ray (CXR) reports: a **previous** CXR report and a **follow-up** CXR report. Your task is to analyze both reports and determine if there are any changes in the follow-up CXR compared to the previous one. ### Instructions: 1. Compare the findings in both reports. 2. If there are **any new, worsening, or improving conditions**, return ‘1‘. 3. If the reports state **”no interval change”** or findings are **stable**, return ‘0‘. 4. If unsure, return ’-1’. 5. Ensure the output is strictly ‘0‘, ‘1‘, ’-1’ without additional text. ### Input Format: - **Previous:** [Insert previous report text] - **Follow-up:** [Insert follow-up report text] ### Output Format: Return only: - ‘0‘ if no changes are detected. - ‘1‘ if any changes (new, improved, or worsened findings) are detected. - ‘-1‘ if uncertain, or any of the previous report or followup report is not a chest X-ray report. ### Few-shot Examples: #### **Example 1: No Changes (Output: 0)** - **Previous:** ”Left pleural effusion is noted. The cardiac silhouette is normal. No acute abnormalities.” - **Follow-up:** ”No interval change.” - **Output:** ‘0‘ #### **Example 2: New Finding (Output: 1)** - **Previous:** ”The lungs are clear. No pleural effusion or pneumothorax. No focal con- solidation. The cardiac silhouette is normal. No acute abnormalities.” - **Follow-up:** ”A new left lower lobe consolidation is noted, con- cerning for pneumonia. No pleural effusion or pneumothorax. The cardiac silhouette is normal.” - **Output:** ‘1‘ #### **Example 3: Improvement in Findings (Output: 1)** - **Pre- vious:** ”Patchy bilateral infiltrates consistent with pneumonia. No pleural effusion. The heart size is within normal limits.” - **Follow- up:** ”Bilateral infiltrates have significantly improved. No pleural ef- fusion. The heart size remains within normal limits. - **Output:** ‘1‘ #### **Example 4: Stable Findings (Output: 0)** - **Previous:** ”Mild left basilar atelectasis. No pneumothorax or pleural effusion. No acute cardiopulmonary abnormalities.” - **Follow-up:** ”Mild left basilar atelectasis remains unchanged. No pneumothorax or pleural effusion.” - **Output:** ‘0‘ 7.4. MS-CXR-T retrieval We describe here the construction process of MS-CXR- T retrieval , our benchmark for evaluating temporal reasoning in CXR report retrieval. The creation pipeline consists of three main stages: Stage 1: Splitting Reports. We first extract the find- ings section and the corresponding report for each image. Reports are split so that each sentence describes a single ra- diological finding. This reduces noise in later stages when manipulating progression labels. The prompt used for this sentence-level splitting is as follows: Analyze the given radiology report text and split it into sentences, each describing a single radiological finding or view position information. Include both positive and negative findings, as well as view position details, but exclude other non-finding information. Follow these guide- lines: 1. Each sentence should contain only one of the following: (exclude the sentence with view position information) a) A clear radiological finding b) The absence of a specific condition (negative finding) 2. Treat each negative finding (absence of a condition) as a separate observation and split it into its own sentence. 3. Keep sentences that describe view position information, but separate them from findings if they appear in the same original sentence. 4. Exclude sentences that do not describe actual radiological find- ings or view positions, such as: - Procedural details - General com- ments about image quality - Patient positioning information (unless it’s specifically about the view position) 5. Maintain the meaning and context of the original findings and view positions while splitting. 6. Minor sentence structure changes or addition of necessary words are allowed to ensure clarity. 7. Remove any redundant information and express each finding or view position concisely. 8. Each split sentence should be understandable independently. 9. Avoid using lists or enumerations within a single sentence; instead, create separate sentences for each item. Example of splitting, including view position, and excluding non- findings: Original Report: ’1. A single frontal view of the chest is provided. 2. No consolidation, pleural effusion, or pneumothorax is observed in both lungs. 3. The heart size is normal.’ Output: ’No con- solidation is observed in both lungs. No pleural effusion is observed in both lungs. No pneumothorax is observed in both lungs. The heart size is normal.’ Process the input text according to these guidelines and return the rel- evant radiological findings and view position information and do not attach any additional text except the split sentences. Stage 2: Prior Reference Omission. We remove all references to prior studies from each finding. This neutral- izes progression status for unrelated findings and allows our benchmark to focus evaluation on the target finding(s) de- fined in MS-CXR-T. This omission step also enables evalu- ation in reversed or swapped settings. The prompt used for removing prior references is: Figure 3. Example workflow for constructing MS-CXR-T retrieval from the original benchmark. You are an expert chest X-ray (CXR) radiologist familiar with radio- logic reports. Your task is to rewrite the given radiology reports by re- moving all references to prior reports or comparisons, while preserving the original structure as much as possible. Input: A radiology report for a chest X-ray (CXR). Output: A revised CXR report focusing solely on current medical findings, excluding references to prior reports, compar- isons, and irrelevant details. Guidelines: Remove Comparisons: Elim- inate any terms or phrases that suggest a comparison, such as ”com- pared to,” ”in comparison with,” ”change”, ”cleared”, ”constant”, ”de- crease”, ”elevate”, ”expand”, ”improve”, ”decrease”, ”increase”, ”per- sistent”, ”reduce”, ”remove”, ”resolve”, ”stable”, ”worse”, ”new”, etc. Focus on Current Findings: Ensure the report only describes the cur- rent state of the patient’s lungs and related structures. Preserve Medical Context: Maintain the original medical terminology and descriptions of abnormalities. Retain Negations: Keep any negative statements about the absence of abnormalities. Example: Original Report: The left apex has not been included on this radiograph. The ET tube terminates 3.9 cm above the carina. The NG tube terminates in the stomach. Surgical clips and a faint metallic coil project over the chest. A left PICC terminates in the mid SVC. EKG leads overlie the chest wall. The lung volumes are low. There are persistent bilateral mid and lower zone hazy opacities. There are persistent bilateral hilar and perihilar linear opacities. No significant interval change is observed in the lung opacities. Bilateral pleural effu- sions are present. The right pleural effusion is greater than the left. No pneumothorax is observed on the right. No cardiomegaly is present. No interval change is observed in the mediastinal silhouette. No sig- nificant interval change is observed in the bony thorax. Output: The left apex has not been included on this radiograph. The ET tube termi- nates 3.9 cm above the carina. The NG tube terminates in the stomach. Surgical clips and a faint metallic coil project over the chest. A left PICC terminates in the mid SVC. EKG leads overlie the chest wall. The lung volumes are low. There are persistent bilateral mid and lower zone hazy opacities. There are bilateral hilar and perihilar linear opac- ities. Bilateral pleural effusions are present. The right pleural effusion is greater than the left. No pneumothorax is observed on the right. No cardiomegaly is present. Stage 3: Creating Progression-Specific Reports. For each predefined finding of interest, we generate three re- ports—one each for ”improved,” ”stable,” and ”worsened” progression. If a specific finding does not exist in the original report, we synthesize the corresponding sentence. The prompt used to generate these progression-specific sen- tences is: **Role:** You are an expert assistant specialized in processing medi- cal text, specifically Chest X-Ray (CXR) reports. **Task:** Given a CXR report text and a specific clinical ’finding’, perform the following steps: 1. **Preprocessing:** * Review the input CXR report. * Remove any sentences that *solely* describe the view position (e.g., ”PA and lateral views were obtained.”, ”AP portable view.”, ”Single frontal view provided.”). Do *not* remove view information if it’s integrated into a sentence describing a finding (though this is less common). * Retain all sentences describing clinical observations, findings, comparisons, and impressions. Avoid removing other general ”unnecessary” words; focus primarily on removing dedicated view position sentences. 2. **Modification based on Finding:** * Identify if the preprocessed report text contains mentions of the provided ’finding’. * **If the finding is mentioned:** * Locate the primary sentence(s) describing the status or appearance of the ’finding’. * Create three versions of the preprocessed report: * **Improved:** Modify the relevant sen- tence(s) minimally to indicate the finding has ’improved’, ’decreased’, ’resolved’, or similar positive change. * **Stable:** Modify the rel- evant sentence(s) minimally to indicate the finding is ’stable’, ’un- changed’, or ’similar’. If the original text already implies stability, ensure this version reflects that clearly. * **Worsened:** Modify the relevant sentence(s) minimally to indicate the finding has ’worsened’, ’increased’, become ’more severe’, or similar negative change. * Make *only the minimal changes* necessary to the specific part about the finding’s status. Keep the rest of the report text identical to the prepro- cessed version. * **If the finding is NOT mentioned:** * Create three versions of the report by appending a new, concise sentence to the end of the preprocessed report text: * **Improved:** Append a sentence like: ”The [finding] shows improvement.” or ”[Finding] is improved.” * **Stable:** Append a sentence like: ”The [finding] appears stable.” or ”[Finding] is stable.” * **Worsened:** Append a sentence like: ”The [finding] has worsened.” or ”There is worsening of the [finding].” or ”[Finding] has increased.” * Use the original preprocessed report text for the beginning of each version. 3. **Output:** * Format the final output as a single JSON object string. * The JSON object must have exactly three keys: ‘”improved”‘, ‘”sta- ble”‘, and ‘”worsened”‘. * The value for each key should be the full text of the corresponding modified report generated in Step 2. Ensure the output is valid JSON. **Input Format Reminder:** The user will provide input in the follow- ing format: ‘finding‘: [The specific clinical finding] ‘report‘: [The full text of the CXR report] **Example 1 (Illustrative - do not repeat in output):** ‘finding‘: pleu- ral effusion ‘report‘: ”When compared to the prior study, the left-sided pleural effusion appears stable. Left consolidation appears relatively stable. No pneumothoraces are seen. The rest of the support lines and tubes are unchanged in position. PA and lateral views were obtained.” **Expected Output Example 1 (Illustrative - do not repeat in output):** “‘json ”improved”: ”When compared to the prior study, the left-sided pleural effusion is improved. Left consolidation appears relatively sta- ble. No pneumothoraces are seen. The rest of the support lines and tubes are unchanged in position.”, ”stable”: ”When compared to the prior study, the left-sided pleural effusion appears stable. Left consoli- dation appears relatively stable. No pneumothoraces are seen. The rest of the support lines and tubes are unchanged in position.”, ”worsened”: ”When compared to the prior study, the left-sided pleural effusion is worsened. Left consolidation appears relatively stable. No pneumoth- oraces are seen. The rest of the support lines and tubes are unchanged in position.” **Example 2 (Illustrative - do not repeat in output):** ‘finding‘: pneu- mothorax ‘report‘: ”Interval improved aeration is noted at both lung bases. Residual patchy and linear left lower lobe atelectasis remains. A small left pleural effusion is present. Single frontal view.” **Expected Output Example 2 (Illustrative - do not repeat in output):** “‘json ”improved”: ”Interval improved aeration is noted at both lung bases. Residual patchy and linear left lower lobe atelectasis remains. A small left pleural effusion is present. The Pneumothorax is im- proved.”, ”stable”: ”Interval improved aeration is noted at both lung bases. Residual patchy and linear left lower lobe atelectasis remains. A small left pleural effusion is present. The Pneumothorax is stable.”, ”worsened”: ”Interval improved aeration is noted at both lung bases. Residual patchy and linear left lower lobe atelectasis remains. A small left pleural effusion is present. The Pneumothorax has worsened.” Now process the following input: 8. Binary Interval-Change Dataset We construct binary interval-change labels (change vs. no change) from radiology report impressions across four datasets. Cases are labeled as no change if the impres- sion contains the phrase “no interval change.” Cases are labeled as change if the impression contains any of the following progression-related keywords: • Worsening: aggravated, exacerbated, increase, worsen, progression, enlarged • Improving:improve, decrease, diminished, reduce, regress, resolve, disappear • New/Developing: new, newly, developed, developing, re- cur, recurrence • Interval changes: interval decrease, interval increase, in- terval improvement, interval worsening Reports matching neither criterion are excluded. Label counts per dataset are summarized in Tab. 6. Table 6. Binary interval-change label distribution. Label 0 denotes no change and Label 1 denotes change. DatasetLabel 0Label 1Total CheXpert49,59153,505103,096 MIMIC22,52745,37067,897 RexGradient8,90210,92119,823 Private5,0005,00010,000 CheXpert and MIMIC exhibit moderate class imbal- ance toward change, reflecting the higher prevalence of progression-related language in follow-up reports. The pri- vate hospital cohort (SNU2) is balanced by randomly sam- pling 5,000 cases per class from the full dataset. 9. Experiment 9.1. Zero-Shot Prompt Design For each finding and progression class, we design 12–17 distinct prompts to capture the diverse phrasing typically found in radiology reports. Using multiple prompts per class helps reduce score variance, as relying on a single tem- plate can lead to unstable results. During zero-shot classifi- cation, we compute the cosine similarity between the image representation and each prompt corresponding to a specific progression label, and average these scores to obtain the fi- nal prediction. all_prompt= ’pneumothorax’: ’improving’:["Improved right pneumothorax.", "Decreased size of pneumothorax compared to prior.", "Interval improvement in pneumothorax.", "Partial resolution of left pneumothorax.", "Pneumothorax has decreased in size.", "Improvement in previously noted pneumothorax.", "Marked reduction in size of pneumothorax.", "Smaller right apical pneumothorax noted today.", "Improved pneumothorax, no acute findings.", "Reduction in pneumothorax volume.", "Improved appearance of left apical pneumothorax.", "Pneumothorax is resolving.", "Pneumothorax shows interval decrease."], ’stable’ : [ "Stable small right pneumothorax.", "No increase in size of pneumothorax.", "Pneumothorax appears unchanged from prior.", "Small left pneumothorax, no acute findings.", "No evidence of expanding pneumothorax.", "Pneumothorax is stable with no tension physiology.", "Minimal pneumothorax, no intervention needed.", "Persistent small pneumothorax without progression.", "Pneumothorax noted, patient remains stable clinically.", "No signs of worsening pneumothorax.", "Pneumothorax is stable in appearance and size.", "No interval change in pneumothorax.", "Left apical pneumothorax stable compared to prior.", "Pneumothorax remains small and non-tension."], ’worsening’ : [ "Worsening right pneumothorax.", "Increased size of pneumothorax compared to prior.", "Interval increase in pneumothorax.", "Progression of left pneumothorax.", "Pneumothorax has enlarged.", "Worsening left apical pneumothorax.", "Marked increase in pneumothorax size.", "Pneumothorax increasing, consider intervention.", "Expansion of previously noted pneumothorax.", "Pneumothorax now involves greater lung volume.", "New increase in size of right pneumothorax.", "Pneumothorax shows interval worsening.", "Pneumothorax progressing compared to previous imaging.", "Enlarging pneumothorax noted on follow-up."] , ’pleural_effusion’: "improving": [ "Improved right pleural effusion.", "Decreased size of pleural effusion compared to prior.", "Interval improvement in pleural effusion.", "Partial resolution of left pleural effusion.", "Reduction in pleural effusion volume.", "Pleural effusion has decreased in size.", "Marked reduction in right pleural effusion.", "Pleural effusion is resolving.", "Improved appearance of left pleural effusion.", "Improvement in previously noted pleural effusion.", "Less fluid seen in pleural space than before.", "Pleural effusion shows interval decrease.", "Decreased right basilar pleural effusion.", ], "stable": [ "Stable small right pleural effusion.", "No increase in size of pleural effusion.", "Pleural effusion appears unchanged from prior.", "Small left pleural effusion, no acute findings.", "No evidence of expanding pleural effusion.", "Pleural effusion is stable in appearance and size.", "Minimal pleural effusion, no intervention needed.", "Persistent small pleural effusion without progression.", "Pleural effusion noted, patient remains stable clinically.", "No signs of worsening pleural effusion.", "No interval change in pleural effusion.", "Left basilar pleural effusion stable compared to yesterday.", "Pleural effusion remains small and unchanged.", "Stable bilateral pleural effusions." ], "worsening": [ "Worsening right pleural effusion.", "Increased size of pleural effusion compared to prior.", "Interval increase in pleural effusion.", "Progression of left pleural effusion.", "Pleural effusion has enlarged.", "Worsening left basilar pleural effusion.", "Marked increase in pleural effusion size.", "Expansion of previously noted pleural effusion.", "Pleural effusion increasing, consider intervention.", "New increase in size of pleural effusion.", "Pleural effusion now causes greater lung compression.", "Pleural effusion shows interval worsening.", "Pleural fluid accumulation appears progressive.", "Enlarging pleural effusion noted on follow-up." ] , ’consolidation’: "improving": [ "Improved right lower lobe consolidation.", "Decreased area of consolidation.", "Interval improvement in consolidation.", "Consolidation has partially resolved.", "Marked reduction in pulmonary consolidation.", "Clearing of previously noted consolidation.", "Consolidation is less extensive than prior.", "Improved left basilar consolidation.", "Reduction in airspace consolidation.", "Consolidation resolving on follow-up imaging.", "Less dense consolidation compared to prior.", "Airspace opacity improving.", "Consolidation has diminished since prior study.", "Patchy consolidation appears improved.", "Fading consolidation with treatment." ], "stable": [ "Stable consolidation in right lower lobe.", "No significant change in consolidation.", "Persistent left basilar consolidation.", "Consolidation appears unchanged from prior.", "Airspace opacity remains stable.", "No interval change in consolidation.", "Chronic consolidation with no acute findings.", "Stable patchy consolidation.", "Consolidation noted without progression.", "No new consolidation identified.", "Findings consistent with stable consolidation.", "Consolidation remains unchanged.", "Stable appearance of parenchymal consolidation.", "Consolidation is chronic and stable.", "No worsening of consolidation." ], "worsening": [ "Worsening right upper lobe consolidation.", "Increased area of consolidation.", "Consolidation more extensive than prior.", "Interval progression of consolidation.", "New or expanding consolidation noted.", "Consolidation has worsened.", "Marked increase in pulmonary consolidation.", "Confluent consolidation involving multiple lobes.", "Airspace consolidation increasing.", "Progressive dense consolidation.", "Patchy consolidation more pronounced.", "Worsening consolidation despite treatment.", "New left basilar consolidation with progression.", "Expanding area of alveolar consolidation.", "Increased opacification consistent with worsening consolidation." ] , ’edema’: "improving": [ "Improved pulmonary edema.", "Decreased pulmonary vascular congestion.", "Edema appears less prominent than prior.", "Interval improvement in pulmonary edema.", "Partial resolution of interstitial edema.", "Reduction in alveolar edema.", "Pulmonary edema has decreased in extent.", "Marked reduction in pulmonary edema.", "Clearing of previously seen pulmonary edema.", "Improved vascular congestion.", "Improved interstitial markings.", "Pulmonary edema resolving with treatment.", "Decreased perihilar opacities.", "Edema improving compared to previous study.", "Less pulmonary edema seen on current film." ], "stable": [ "Stable pulmonary edema.", "No significant change in pulmonary edema.", "Pulmonary edema appears unchanged from prior.", "Edema remains stable in extent.", "Persistent mild pulmonary edema.", "Pulmonary vascular congestion unchanged.", "Interstitial markings stable.", "No interval change in pulmonary edema.", "Pulmonary edema noted without progression.", "No worsening of pulmonary edema.", "Edema appears chronic and stable.", "Stable vascular congestion.", "No new signs of fluid overload.", "Pulmonary edema similar to previous exam.", "Mild pulmonary edema, no acute change." ], "worsening": [ "Worsening pulmonary edema.", "Increased pulmonary vascular congestion.", "Edema appears more prominent than prior.", "Interval increase in pulmonary edema.", "Progressive alveolar edema.", "Pulmonary edema has worsened.", "Marked increase in pulmonary edema.", "Expansion of interstitial edema.", "New or worsening bilateral pulmonary edema.", "Pulmonary edema now more confluent.", "Increasing perihilar opacities.", "Worsening interstitial markings.", "Pulmonary congestion progressing.", "Increased fluid overload signs on imaging.", "Diffuse worsening of pulmonary edema pattern." ], ’pneumonia’: "improving": [ "Improved right lower lobe pneumonia.", "Decreased consolidation in left lung.", "Pneumonia shows interval improvement.", "Clearing of previously seen infiltrates.", "Reduction in airspace opacity.", "Partial resolution of pneumonia.", "Consolidation is less extensive than prior.", "Improved left basilar pneumonia.", "Decreased right middle lobe opacities.", "Pneumonia improving with antibiotic therapy.", "Pulmonary infiltrates have diminished.", "Improved patchy opacities.", "Interval decrease in parenchymal opacities.", "Airspace disease appears less prominent.", "Pneumonia resolving compared to previous imaging." ], "stable": [ "Stable right lower lobe pneumonia.", "Pneumonia appears unchanged from prior.", "Persistent left basilar consolidation.", "No interval change in airspace disease.", "Patchy infiltrates stable in appearance.", "Consolidation remains without significant change.", "No progression of pneumonia.", "Airspace opacity unchanged.", "Stable pneumonia on follow-up imaging.", "No new consolidation identified.", "Chronic-appearing infiltrates, no acute change.", "Stable bilateral patchy opacities.", "No worsening of pneumonia noted.", "Findings consistent with prior pneumonia, stable.", "Stable airspace disease, no interval change." ], "worsening": [ "Worsening right upper lobe pneumonia.", "Increased consolidation in left lower lobe.", "Pneumonia appears more extensive than prior.", "Interval progression of pneumonia.", "New or worsening bilateral infiltrates.", "Airspace opacities have increased.", "Expansion of previously seen pneumonia.", "Pneumonia worsening despite treatment.", "More confluent consolidation noted today.", "Increasing parenchymal opacities.", "Marked progression of airspace disease.", "Increased patchy opacities compared to prior.", "Pulmonary infiltrates have progressed.", "Worsening left lower lobe pneumonia.", "New consolidation suggestive of worsening pneumonia." ] 9.2. Ablation for Hyperparameters We ablate the pretraining weight W for the Change-aware Sigmoid loss and the fine-tuning weight λ for the Temporal Consistency Loss (TCL). Table 7 reports average macro- accuracy on MS-CXR-T retrieval when varying W . Although W = 0.5 achieves the highest Standard accuracy, W = 1 provides the best overall trade-off, yielding the strongest Reversed, Combined, and Consistency scores. Setting W = 2 places too much emphasis on the inverted pairs and de- grades performance across all protocols. Table 8 summarizes the effect of the fine-tuning weight λ on supervised MS-CXR-T classification. Introducing BiCE alone (λ = 0) already results in a substantial improvement over the baseline across all evaluation protocols, demon- strating that enforcing label inversion provides a strong di- rectional signal. Adding TCL with a small weight (λ = 1) produces performance similar to λ = 0, reflecting the scale difference between the cross-entropy and TCL objectives. Increasing the TCL weight to λ = 50 yields the best bal- anced performance, with the high Reversed, Combined, and Consistency scores and only a minor reduction relative to the BiCE-only setting in the Standard metric. At λ = 100, directional metrics remain strong, but Standard and Com- bined accuracies begin to drop, indicating that overly large TCL weights may over-regularize the model. Based on these trends, we adopt λ = 50 for all main experiments. Table 7. Effect of pretraining weight W on MS-CXR-T retrieval . We report average macro-accuracy (%) across all findings for each evaluation protocol. W Average Accuracy (%) StandardReversedCombinedConsistency 050.250.955.432.7 0.554.953.259.136.6 154.154.059.237.9 249.751.254.333.1 Also, in practice, we do not apply the inversion-aware objectives from the very beginning of training. If BiCE or TCL is introduced too early, the model can quickly con- verge to a degenerate solution that predicts stable for most pairs, which locally minimizes the bidirectional losses but is not clinically meaningful. To avoid this shortcut, we first warm up the model with standard objectives and then enable the Change-aware Sigmoid loss after 10 pretraining epochs, and TCL after 20 epochs of fine-tuning. This stag- ing allows the backbone to learn non-trivial temporal struc- Table 8. Effect of fine-tuning weight λ on MS-CXR-T supervised classification. We report average macro-accuracy (%) across all findings for each evaluation protocol. λ Average Accuracy (%) StandardReversedCombinedConsistency Base61.153.359.639.5 065.262.063.453.1 165.061.263.153.3 5064.163.763.657.3 10062.763.861.455.6 ture before inversion-aware regularization is applied. 10. Additional Information 10.1. Clinical Utility of Reversed and Combined Evaluations Radiologists typically assess temporal progression only in the forward (standard) direction. The Reversed and Com- bined evaluations are therefore introduced as analytical tools to rigorously validate the reliability of forward pre- dictions. The Reversed evaluation tests whether a model truly understands temporal progression. For instance, a model reaching 70% accuracy in the forward direction but only 40% in the reversed direction likely exploits spurious cues rather than genuine temporal reasoning. Such discrepancies may undermine the trustworthiness of forward predictions. The Combined evaluation complements this by enforc- ing consistency across both directions. By aggregating re- sults from forward and reversed pairs, it exposes and helps mitigate directional biases. Together, these settings pro- vide crucial analytical validation, ensuring that predictions in standard clinical scenarios are both reliable and inter- pretable. Implications for Temporal Inversion. While reversible scenarios predominate across these findings, temporal in- version is not intended to model exact biological recov- ery. Instead, it provides a controlled perturbation that tests whether models capture the direction of temporal change. Radiologic follow-up primarily evaluates changes in le- sion size, burden, or conspicuity—features that exhibit ap- proximate reversibility and therefore lend themselves to inversion-based analysis.Our experiments Our experi- ments indicate that including inverted pairs improves direc- tional sensitivity without compromising forward-order pre- dictions, supporting temporal inversion as a practical stress test for assessing order-aware interval-change modeling. 10.2. Use of Large Language Models We disclose the use of large language models (LLMs) in this work.LLMs were used in three ways: (i) to re- fine the clarity and presentation of writing; (i) to assist in generating change/no-change labels from radiology reports (see Sec. 7.3); and (i) to construct the MS-CXR-T retrieval benchmark by modifying radiology reports under controlled prompts (see Sec. 7.4). All LLM usage was limited to these supporting roles and did not alter the core experimental re- sults or conclusions.

Instant research discovery