{"id":"79491e69-43e0-499f-b55f-843bc4cd3a32","arxiv_id":"2412.04030","paper_version":3,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"CNNs trained on PadChest chest X-rays classify above chance even when the lungs are masked out, indicating shortcut learning, while glaucoma models on Chaksu rely more on expected anatomy.","lead":"This paper tests whether medical image classifiers still get high scores when the clinically relevant parts of the X-ray or eye image are hidden. It finds that models trained on chest X-rays keep diagnosing above chance without the lungs, suggesting they rely on spurious clues in the image.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"For chest X-rays, 'without lungs' still contains clinically relevant anatomy (heart, mediastinum, pleura); high Full-model AUC there may reflect valid features, not shortcuts, for cardiomegaly and effusion.","rationale":"I read this as an empirical shortcut-detection study whose central claim is that CNNs can classify chest X-rays from outside-lung regions, implying reliance on non-clinical features. The experiments are well-executed and code is provided, but the conclusion depends on the assumption that the lung mask is the full clinically relevant ROI for all five conditions. That assumption is demonstrably false for cardiomegaly (heart outside lung mask) and probably for effusion (pleural space/costophrenic angles). The paper is partly self-aware: §3.2 discusses the heart with the bounding box, and the expert study found a cardiomegaly case in no-lungs images, but the 'consistent over all classes' interpretation in §4.1 does not follow. A matched-area distribution-shift control would also clarify whether the low only-lungs AUC reflects shortcut use or an unfamiliar input transform. The Chaksu analysis and the OOD results are less affected by this concern, so I would not reject the paper; I would require reanalysis or reframing of the chest-X-ray claim before accepting it as stated. This is why I propose CONDITIONAL rather than UNCHANGED.","tokens_in":15142,"tokens_out":7274,"duration_ms":80335,"concrete_test":"Re-run the PadChest analysis using disease-specific ROIs. For cardiomegaly, use a heart segmentation (or heart bounding box) as the ROI and evaluate Full models on 'no heart' images; for effusion, mask the costophrenic angles/pleural space rather than only lungs. Also run a matched-area control: black out random rectangular regions of the same total area as the 'Only lungs' mask, and evaluate the Full model; if AUC on this control is comparable to 'Only lungs,' then the low only-lungs AUC is partly a distribution-shift artifact. If Full-model AUC on 'no heart' or 'no costophrenic angles' stays near chance, the chest-X-ray shortcut claim for those classes is unsupported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper equates 'outside the CheXmask lung segmentation' with 'clinically irrelevant.' That equation fails for at least two of the five PadChest classes. Cardiomegaly is defined by heart size (cardiothoracic ratio), and the heart lies outside the lung mask; 'No lungs' images preserve the heart, while 'Only lungs' images remove it. Pleural effusion is often most visible at the costophrenic angles/pleural space, which may be outside or only partially inside the lung segmentation. Thus §4.1's headline observation—that Full models score higher on without-lungs than on only-lungs images—is exactly what a clinically correct model should do for these classes; it does not by itself demonstrate use of non-clinical shortcuts. The paper itself notes in §3.2 that the bounding box variant 'include[s] or exclude[s] relevant parts like the heart for the cardiomegaly condition,' and §4.5 reports the radiology resident correctly identified a cardiomegaly case in a no-lungs image. Additionally, comparing a model trained on full images and tested on heavily blacked 'Only lungs' images confounds content removal with a large distribution shift; the dilation experiment does not isolate the shortcut mechanism. A model that learned no lung features would likely also fail on only-lungs images for this reason. So the central generalization 'models rely on non-relevant parts' is overstrong for chest X-rays; it is supported mainly for lung-confined conditions (e.g., pneumonia, atelectasis), not uniformly across the five classes.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies shortcut learning in medical image classifiers by training DenseNet-121 models on chest X-rays (PadChest) and eye fundus images (Cháks.u) under five masking strategies: full images, images with the clinically relevant region (lungs or optic disc) removed using precise masks, the same with bounding boxes, and images retaining only the ROI via masks or bounding boxes. The main claims are that all chest X-ray models achieve above-random AUC regardless of masking, that models trained on full images perform better on images without the ROI than on images with only the ROI, that out-of-distribution performance is generally poor, and that a radiology resident found it nearly impossible to diagnose without-ROI images. The paper also reports embedding similarities, SHAP visualizations, and a disc-size dilation experiment for glaucoma that suggests reliance on optic disc size rather than cup-to-disc ratio. The authors conclude that models are affected by spurious correlations, with chest X-rays showing stronger evidence than fundus images.","tokens_in":15364,"tokens_out":6119,"duration_ms":57653,"significance":"If the conclusions hold, the paper provides a useful cautionary demonstration that CNNs can achieve non-trivial AUC even when clinically relevant structures are removed, and it offers a multi-pronged methodology (masking variants, OOD evaluation, embeddings, SHAP, expert study) for auditing shortcut behavior. The glaucoma disc-size experiment is a particularly nice falsifiable test, and the release of code supports reproducibility. However, the central chest X-ray claim is currently overstated: the 'without lungs' condition does not remove all clinically relevant anatomy for cardiomegaly (heart) and effusion (costophrenic angles), so the headline comparison between 'No lungs' and 'Only lungs' does not cleanly separate clinical features from non-clinical shortcuts. The paper's value lies in its methodology and per-class observations, but the interpretation needs substantial refinement before the general conclusion about non-relevant cues is justified.","major_comments":[{"comment":"The headline finding that full-image models score higher on 'No lungs' than on 'Only lungs' images is presented as evidence of reliance on non-clinical shortcuts, but this comparison is confounded for at least two of the five PadChest classes. Cardiomegaly is diagnosed from heart size (cardiothoracic ratio), and the heart lies outside the CheXmask lung segmentation, so the 'No lungs' condition preserves the clinically decisive organ while 'Only lungs' removes it; for pleural effusion, the costophrenic angles may be partially outside the lung mask. The paper itself acknowledges this in §3.2 ('include or exclude relevant parts like the heart for the cardiomegaly condition') and in §4.5 reports that the radiology resident correctly identified a cardiomegaly case in a no-lungs image. The claim that high without-ROI performance reflects spurious correlation should be restricted to classes whose relevant anatomy is fully contained in the lung mask (e.g., pneumonia, atelectasis) or re-analyzed per class with the bounding-box conditions separated from the mask conditions.","section":"§4.1, Fig. 5"},{"comment":"The dilation experiment does not isolate the shortcut mechanism. Applying a model trained on full images to heavily blacked-out 'Only lungs' images induces a large distribution shift, because most pixels are replaced by a constant value; the observed drop in AUC may reflect the model's sensitivity to the masking operation itself rather than reliance on non-lung features. Without a control model trained on images with the same distribution of masks (e.g., random masking or train-time dilation), the conclusion that the model 'uses elements near the lungs' is not uniquely supported. The authors should either add such a control or temper the interpretation of Fig. 5.","section":"§4.1, Fig. 5"},{"comment":"The radiologist study is used to conclude that models rely on non-relevant features because the resident found only 2 of 35 conditions in without-ROI images. This inference does not follow: the resident was deprived of projection information (as the paper notes) and saw only a single image per case, while the model may exploit subtle but clinically valid cues such as heart size for cardiomegaly. The study is also underpowered (one expert, 75 images) and the image selection by model probability may bias the difficulty. The conclusion should be tempered to state that the models' decision cues are not readily usable by a human expert under the same information restrictions, rather than claiming that the resident's difficulty proves the model used non-clinical features.","section":"§4.5, §3.5"}],"minor_comments":[{"comment":"The AUC heatmaps show only mean values across the five folds; adding per-fold values or standard deviations would help assess variability, especially since the significance conclusions rely on a 3-of-5 folds rule.","section":"Fig. 4"},{"comment":"The sentence about the bounding box 'include or exclude relevant parts like the heart for the cardiomegaly condition' is stated in the methods but its implications for interpreting the mask-based results are not carried through in the discussion; consider moving or expanding it in the results section.","section":"§3.2"},{"comment":"The name of the eye fundus dataset is spelled inconsistently ('Cháks.u' vs 'Cháks.u'); please standardize the spelling throughout.","section":"Abstract and §3.1"},{"comment":"The OOD evaluation is on full images; clarify in the caption or text that the 'No ROI' rows refer to models trained on masked images and evaluated on full images, to avoid confusion with the in-distribution masked evaluation.","section":"Table 2"},{"comment":"The cosine similarity results are acknowledged as limited and not aligning with AUC; consider reporting them in the appendix or adding a short interpretation of why shared background pixels dominate the embeddings.","section":"§4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the empirical setup is transparent, with code, OOD evaluation, and a domain-expert study. The main concern is the overgeneralized interpretation of the chest X-ray masking results; this is fixable with a per-class analysis and a more controlled dilation experiment. No citation or novelty concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this. It does what it says: trains DenseNet-121s on PadChest and Chaksu under five masking strategies, measures AUC, adds SHAP and embeddings, and brings in a radiology resident. The Chaksu results are genuinely interesting, especially the optic-disc-size dependence that survives even in 'Only disc' models, and the OOD table is a useful caution. Code is there. The limitations section is candid, including the mask-quality issue and the single-resident study.\n\nThe soft spot is the interpretation of the chest X-ray results. The headline claim—that models rely on non-relevant parts because they score well on 'without lungs' images—is only partly supported. For cardiomegaly and effusion, the anatomy that matters (heart, costophrenic angles, pleural space) is partly or wholly outside the CheXmask lung segmentation. A clinically correct model should still do reasonably well on 'No lungs' images for those classes, and the comparison against 'Only lungs' images is not a clean shortcut test for them. The paper itself notes this in §3.2 and §4.5, but the abstract and §4.1 present the across-class pattern as uniform evidence of shortcuts. That overreaches.\n\nThe second issue is distribution shift. Evaluating a Full model on images with most of the image blacked out is not just removing the ROI; it is a large input transformation. A model that never learned any non-lung shortcut could still perform worse on 'Only lungs' images simply because the input is far outside its training distribution. The dilation experiment helps but does not fully isolate a shortcut mechanism.\n\nWhere does that leave the paper? The Chaksu analysis, the OOD table, and the careful embedding discussion are solid contributions. The chest X-ray result is real but needs re-framing: above-random performance without the lung field is suggestive of shortcuts for lung-confined conditions like pneumonia and atelectasis, but for cardiomegaly and effusion it is confounded by anatomy. The resident study is a nice complement, though 75 images and one reader limits it.\n\nThis deserves a serious referee and, in my view, major revision rather than acceptance as-is—the abstract should be toned down and the per-class anatomy acknowledged. It is honest work, clearly reported, and worth engaging with.","headline":"Solid, honest empirical study; central 'shortcut' claim is overstrong for chest X-rays because 'without lungs' still contains clinically relevant anatomy.","tokens_in":15975,"tokens_out":2291,"would_cite":true,"duration_ms":21589,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Masking the lungs does not stop chest X-ray classifiers from scoring above chance, and full-image models score higher without the lungs than with only the lungs.","keywords":["shortcut learning","spurious correlation","chest X-ray classification","glaucoma classification","medical image masking","model robustness","explainability","out-of-distribution evaluation"],"falsifier":"Evaluate the no-lungs and no-lungs-bounding-box models on images in which everything outside the lung mask is replaced by class-matched noise while the lung region is blackened; if the above-chance AUC persists, non-clinical peripheral cues drive the predictions, but if it falls to 0.5, the apparent shortcut is leakage of lung pixels through imperfect masks.","tokens_in":14917,"feed_emoji":"🩻","tokens_out":5378,"duration_ms":48695,"temperature":0.7,"pith_summary":"Masking experiments on chest X-rays and eye fundus images show that a standard CNN keeps classifying above chance even when the clinically relevant anatomy is removed, and that full-image chest models actually score higher on images without the lungs than on images with only the lungs. The paper argues this is evidence of shortcut learning: models exploit non-clinical cues in the periphery rather than the anatomy clinicians use. On a glaucoma dataset, performance looks unbiased at first, but a mask-dilation experiment reveals that optic disc size alone drives predictions. A radiology resident could rarely diagnose the same masked chest images, and external validation shows poor transfer of all masked models. The work matters because benchmark AUC can overstate the true clinical capability of medical imaging models.","feed_headline":"Chest X-ray AI still diagnoses with lungs masked out","feed_subtitle":"Models score above chance without the clinical region, a sign they rely on shortcuts rather than anatomy.","key_machinery":"The controlling mechanism is a systematic occlusion protocol built from anatomical segmentation masks. For lungs, the CheXmask masks define the region of interest; for fundus images, the Cháks.u optic disc and cup masks do. Each image is presented in five versions—full, outside the mask, outside the bounding box, inside the mask, inside the bounding box—and one DenseNet-121 model is trained per version, then cross-evaluated by AUC with DeLong significance tests, mask-dilation sweeps, SHAP attributions, t-SNE and cosine embedding comparisons, and a radiology-resident reading study. The bounding-box variants are the key control: they remove shape and location information of the ROI without removing nearby structures, which is what reveals reliance on peripulmonary or boundary cues.","core_discovery":"The central claim is that convolutional classifiers trained on PadChest chest X-rays rely on spurious correlations: every model, no matter which masking strategy it was trained on, achieves an AUC above random in matched evaluation, and models trained on full images perform better on images without the lungs than on images containing only the lungs. For atelectasis, effusion, and pneumonia, models evaluated on only-lung images fall near or below chance, while bounding-box versions recover performance, pointing to usage of peripulmonary structures. On the Cháks.u glaucoma dataset, full-image models reach 0.68 AUC while no-disc models fall to near chance, but the dilation experiment shows optic disc size alone drives predictions, which is a shortcut masquerading as clinically sensible focus.","pith_inferences":["A similar train-and-evaluation masking audit could be run cheaply on other pathologies and architectures as a pre-deployment shortcut screen, since the released code supports swapping models.","The disc-size result predicts measurable subgroup performance gaps across populations with different optic disc size distributions; this is directly testable by stratifying AUC by disc size in external data.","The embedding analysis shows that penultimate-feature cosine similarity and t-SNE can rank masking conditions differently from AUC, so representations before the classification head are weak evidence about shortcut use without interventional masking.","Perturbing mask boundaries and measuring the AUC slope would quantify how much of the no-lungs performance is an artifact of imperfect masks versus genuine peripheral shortcut use."],"forward_implications":["Benchmark AUCs on PadChest overstate real capability: models can score well without ever seeing the lungs.","Masking the ROI is not a sufficient robustness audit; only cross-evaluation across masking conditions reveals shortcut reliance.","Peripheral and boundary cues such as pacemakers, image edges, and scanner-specific artifacts are learnable shortcuts that survive targeted anatomical removal.","Models trained only on the ROI, or only outside it, transfer poorly to external datasets, so shortcut reliance limits deployment across hospitals.","For glaucoma, models can fixate on a clinically plausible region while using the wrong feature, so explainability maps alone cannot certify correct reasoning."],"supporting_citations":[{"why":"PadChest chest X-ray dataset, the source of training and in-distribution test images for the five studied conditions.","marker":"[40]"},{"why":"CheXmask supplies the lung segmentation masks used to build the masked versions of PadChest images.","marker":"[43]"},{"why":"Cháks.u fundus dataset with expert glaucoma labels and optic disc and cup masks, with STAPLE-fused masks defining the ROI.","marker":"[45]"},{"why":"ChestX-ray14 test set used as external out-of-distribution evaluation for the chest models.","marker":"[42]"},{"why":"AIROGS glaucoma screening dataset used as external out-of-distribution evaluation for the fundus models.","marker":"[10]"},{"why":"Reverse Classification Accuracy threshold (Dice RCA above 0.7) filters low-quality lung masks, grounding the mask-reliability assumption.","marker":"[44]"},{"why":"DenseNet-121 architecture pre-trained on ImageNet, the single model class studied, with all but the last dense block frozen.","marker":"[47]"},{"why":"SHAP occlusion-based explainability method chosen to attribute model decisions on masked images.","marker":"[54]"},{"why":"DeLong test for correlated AUC comparison, used across five folds to justify claims of statistically significant differences.","marker":"[49]"}],"fun_headline_variants":["Chest X-ray AI diagnoses with lungs masked away","Masked lungs don't stop X-ray AI from diagnosing","X-ray AI beats random even without the region of interest","AI uses shortcuts, not anatomy, to read chest X-rays","Lungs not needed: X-ray AI still diagnoses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the assumption that the segmentation masks accurately identify all clinically relevant parts of the image; if the masks are systematically wrong, the 'without-ROI' images still contain the relevant anatomy or the 'only-ROI' images exclude it, and the shortcut conclusion collapses.","fun_headline_variants_meta":{"raw":{"variants":["Chest X-ray AI diagnoses with lungs masked away","Masked lungs don't stop X-ray AI from diagnosing","X-ray AI beats random even without the region of interest","AI uses shortcuts, not anatomy, to read chest X-rays","Lungs not needed: X-ray AI still diagnoses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000757,"raw_usage":{"total_tokens":3360,"prompt_tokens":936,"completion_tokens":2424,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":2344}},"tokens_in":552,"tokens_out":2424,"duration_ms":16683,"temperature":1.0,"reasoning_tokens":2344,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:49:45.639887+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the no-lungs and no-lungs-bounding-box models on images in which everything outside the lung mask is replaced by class-matched noise while the lung region is blackened; if the above-chance AUC persists, non-clinical peripheral cues drive the predictions, but if it falls to 0.5, the apparent shortcut is leakage of lung pixels through imperfect masks.","supporting_citations":[{"cited_title":"Padchest: A large chest x-ray image dataset with multi-label annotated reports","cited_arxiv_id":null,"evidence_quote":"PadChest chest X-ray dataset, the source of training and in-distribution test images for the five studied conditions."},{"cited_title":"Chexmask: a large-scale dataset of anatomical segmentation masks for multi-center chest x-ray images","cited_arxiv_id":null,"evidence_quote":"CheXmask supplies the lung segmentation masks used to build the masked versions of PadChest images."},{"cited_title":"Cháks.u: A glaucoma specific fundus image database.Scientific data, 10(1):70, 2023","cited_arxiv_id":null,"evidence_quote":"Cháks.u fundus dataset with expert glaucoma labels and optic disc and cup masks, with STAPLE-fused masks defining the ROI."},{"cited_title":"Chestx-ray8: Hospital-scale chest x-ray database and bench- marks on weakly-supervised classification and localization of common thorax dis- eases","cited_arxiv_id":null,"evidence_quote":"ChestX-ray14 test set used as external out-of-distribution evaluation for the chest models."},{"cited_title":"Airogs: artificial intelligence for robust glaucoma screening chal- lenge","cited_arxiv_id":null,"evidence_quote":"AIROGS glaucoma screening dataset used as external out-of-distribution evaluation for the fundus models."},{"cited_title":"Reverse classi- fication accuracy: predicting segmentation performance in the absence of ground truth","cited_arxiv_id":null,"evidence_quote":"Reverse Classification Accuracy threshold (Dice RCA above 0.7) filters low-quality lung masks, grounding the mask-reliability assumption."},{"cited_title":"A unified approach to interpreting model pre- dictions","cited_arxiv_id":null,"evidence_quote":"SHAP occlusion-based explainability method chosen to attribute model decisions on masked images."}],"review_version":1}