{"id":"7134a8ca-29c1-4f6a-811e-bd7469bf45b0","arxiv_id":"2506.04058","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Concept vectors in an autoencoder's latent space produce stable counterfactual explanations for large chest X-ray pathologies, but the method does not beat the Latent Shift baseline.","lead":"This paper applies Concept Activation Vectors to the latent space of a chest X-ray autoencoder, then adds or subtracts those vectors to create counterfactual images that exaggerate or reduce pathologies. The approach does not beat the Latent Shift baseline in overlap with doctor-labeled regions, but cardiomegaly explanations are competitive and concept vectors are stable across two large X-ray datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Explanation-quality claim lacks a random-direction control: IoU gains for cardiomegaly may reflect global reconstruction artifacts rather than concept-specificity, so the central claim is not yet fully supported.","rationale":"I read the paper as an honest, clearly labeled preliminary study. It does not claim to beat the baseline overall, it reports limitations openly, and it provides code and public datasets for re-implementation. The concept-vector stability evidence is at least internally coherent: cosine similarity to random vectors is near zero, while cross-dataset similarity is 0.406, roughly ten standard deviations away from the random baseline. The reader's CONDITIONAL verdict is appropriate. My concern is not about fraud or hidden errors; it is about a missing control that is load-bearing for the central interpretability claim. The paper shows that CAVs are non-random directions and that averaging them improves IoU for cardiomegaly, but it never shows that random directions fail the same IoU test. Without that control, the visual explanation results could be explained by the autoencoder's reconstruction behavior rather than by concept encoding. This is a concrete, addressable gap rather than a fatal flaw. I would keep the verdict CONDITIONAL, adding the requirement for a random-direction IoU baseline in revision. I partially agree with the reader's weakest assumption: the reader emphasized dataset artifacts and reconstruction fidelity, which are related, but the sharper issue is the absence of a null-model for the explanation metric itself.","tokens_in":6900,"tokens_out":3401,"duration_ms":37049,"concrete_test":"Run the same evaluation pipeline with K=100 random latent directions (same dimensionality, same step size of 10, same autoencoder, same attribution computation) on the same test images and bounding boxes used for Tables 2-3. Compute mean IoU per pathology for single random directions and for the averaged random vector, and compare against the NIH CAV mean IoU values (e.g., 0.363 for Cardiomegaly). If random-direction IoU on Cardiomegaly is comparable to or above 0.363, the concept-specificity claim is unsupported; if random-direction IoU is near zero and clearly below the CAV IoU, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two quantitative pillars: cosine similarity of concept vectors across datasets (Table 1) and IoU of attribution maps against bounding boxes (Tables 2-3). The first pillar only shows that the CAV is not a random direction in latent space; it does not establish that traversing along the CAV changes pixels selectively in the concept region. The second pillar has no null-model control: the paper never reports IoU for random-vector traversals under the same step size, reconstruction, and attribution-mapping protocol. Because the attribution map is computed as the absolute difference between traversed reconstructions and the original image (Section 3.2), any latent direction that produces global intensity or contrast changes will have high overlap with a large central bounding box such as cardiomegaly, independent of whether the direction encodes the clinical concept. The reported 0.363 IoU for the averaged NIH CAV versus 0.286 for Latent Shift is therefore not sufficient to establish concept-specific explanation quality unless random directions are shown to perform substantially worse. The class-label proxy (Section 3.1) reinforces this concern: a linear separator trained on label-defined positive and negative sets may recover any label-correlated visual feature, and the modest cross-dataset cosine similarity of 0.406, while far from random, does not by itself show that the direction is clinically specific.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes using Concept Activation Vectors (CAVs) in the latent space of a pre-trained reconstruction autoencoder to produce counterfactual visual explanations for chest X-rays. Concepts are defined by class labels as proxies (Section 3.1), linear classifiers are trained in the flattened latent space to obtain concept directions, and explanations are generated by adding/subtracting the CAV to image embeddings and taking absolute differences between reconstructions. The authors evaluate concept-vector stability via cosine similarity across NIH and CheXpert datasets, and explanation quality via IoU against bounding boxes, comparing with the Latent Shift baseline. They report that the approach is preliminary, does not outperform the baseline overall, works better for large pathologies such as cardiomegaly, and struggles on smaller pathologies such as atelectasis. The paper includes a link to code and appendices with hyperparameters and example attributions.","tokens_in":7148,"tokens_out":2392,"duration_ms":22872,"significance":"If the central claim were fully supported—that class-label-derived concept directions in a reconstruction autoencoder are stable across datasets and produce concept-specific, clinically meaningful visual explanations—the work would be a useful contribution to interpretable medical imaging and would align with the growing interest in concept-based XAI. The authors are honest about the preliminary nature of the results, and the public code link is a strength, as is the direct comparison with the Latent Shift baseline. However, the current evidence does not yet establish concept-specificity: the explanation-evaluation pipeline lacks any null-model or random-direction control, and the label-free framing in the abstract conflicts with the actual use of class labels to train the CAV classifiers. Because the interpretation of Tables 2 and 3 as evidence of clinically meaningful explanations depends on this control, the central claim is not yet fully supported.","major_comments":[{"comment":"The explanation-quality evaluation lacks a random-direction control. The 2D attribution map is defined as the absolute difference between a traversed reconstruction and the original image (Section 3.2), so any latent direction that produces global intensity or contrast changes will yield high overlap with a large, centrally located bounding box such as the cardiomegaly ground truth, irrespective of whether the direction encodes the clinical concept. The reported IoU of 0.363 for the averaged NIH CAV versus 0.286 for Latent Shift therefore does not establish concept-specific explanation quality unless the same traversal protocol with random latent directions is shown to produce substantially lower IoU. Please add this control, with the same step size, averaging procedure, and attribution-mapping pipeline, and report the distribution over multiple random directions.","section":"Section 3.2 / Section 5.2 / Tables 2-3"},{"comment":"The abstract and conclusion state that concepts are linked to image-level features 'without requiring explicit training on class labels' and 'without explicit label training,' yet Section 3.1 explicitly says 'we use class labels as proxies for the concepts' and the CAV is obtained by training a linear classifier on positive samples labeled with the target class and negative samples without it. This is a direct internal inconsistency and is load-bearing for the framing of the contribution. Please either change the claim to acknowledge the label-based supervision used to define concepts, or provide a genuinely label-free construction of the concept sets.","section":"Abstract / Section 3.1"},{"comment":"The number of concept vectors used for averaging is inconsistent. Section 4.2 states 'we generate explanations using 4 sets of concept vectors for each dataset and then average those vectors,' while Appendix C, Table 5 reports 'Number of concept vectors for averaging: 10.' The reported mean values and standard deviations in Tables 1-3 depend on this number, and the discrepancy prevents exact reproducibility. Please unify the text and tables and state the exact number of concept-vector replicates used for each reported statistic.","section":"Section 4.2 / Appendix C, Table 5"},{"comment":"The interpretation of the cross-dataset cosine similarity is not adequately supported. The paper states that 'in high-dimensional latent space even the smallest observed value (0.1011 for NIH to CheXpert similarity) is very high,' but the absolute value is low, and no statistical test or comparison against a baseline direction (e.g., random vectors with the same norm or a label-agnostic PCA direction) is provided beyond the raw random-vector similarity shown in Table 1. A cosine similarity of 0.1011 that is several standard deviations from the random baseline still represents weak absolute alignment, and it does not by itself establish that the direction is clinically specific rather than merely dataset-correlated. Please add a significance test or a meaningful reference point for 'high' similarity, and avoid over-interpreting low absolute values.","section":"Section 5.1 / Table 1"}],"minor_comments":[{"comment":"There is a typo, 'the the model predictions,' and the phrase 'CA Vs' uses inconsistent spacing throughout; please standardize to 'CAVs.'","section":"Introduction"},{"comment":"The paper says 'we compute the similarity of concept vectors with random vectors' and later refers to 'inter and intra dataset' experiments, but the terminology is applied interchangeably; please clarify which of the three experiments corresponds to Table 1 and which to the appendix.","section":"Section 4.1"},{"comment":"The sentence 'In out case, Intersection over Union provides a measure' contains a typo ('out' should be 'our'), and the metric definition would benefit from a precise formula or reference to the exact equation used.","section":"Section 4.2"},{"comment":"The sentence 'Despite the approach does not beat the baseline in its current form' is grammatically incorrect; consider 'Although the approach does not beat the baseline in its current form.'","section":"Abstract / Conclusion"},{"comment":"The tables report standard deviations for individual CAV results but not for the averaged-vector results; adding confidence intervals or a small-number-of-vectors caveat would help readers judge the stability of the mean IoU values.","section":"Tables 2-3"}],"recommendation":"major_revision","confidential_remarks":"The absence of a random-direction control in the IoU evaluation is the most consequential issue; adding this control is feasible and directly addresses the central claim. The label-free/label-proxy inconsistency is also easy to fix textually. I would not recommend rejection, as the authors are appropriately cautious and the preliminary framing is honest, but the paper needs a substantive experimental addition before the stability and explanation-quality claims can be taken as established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimate, clearly-written preliminary study combining concept activation vectors with latent-space traversal for chest X-ray counterfactuals. The new bits are the cross-dataset CAV similarity measurements (Table 1), the IoU comparisons for five pathologies against the Latent Shift baseline, and the observation that larger pathologies work better than smaller ones. The code is public, the data are standard, and the authors are upfront that the method does not beat the baseline overall.\n\nWhat impressed me: the negative result on atelectasis/nodule is reported honestly, and the stability check against random vectors is a reasonable sanity check. The averaged cross-dataset cosine similarity of 0.406 is modest in absolute terms but clearly above random, so the claim that CAVs are not dataset-specific is plausible.\n\nThe soft spots are real but mostly fixable. First, the abstract says 'without explicit label training,' but Section 3.1 uses class labels as proxies for concepts. That overstates the method. Second, the averaging count is inconsistent: Section 4.2 says '4 sets of concept vectors' while Table 5 says 10; that needs cleaning up. Third, the IoU numbers for the averaged vectors have no uncertainty, so we can't tell if the NIH-mean cardiomegaly 0.363 vs baseline 0.286 is robust.\n\nThe bigger issue, and the one that matters most for the central claim, is the missing control for random-vector traversals. The attribution map is computed as the absolute difference between the original and the traversed reconstruction. Any latent direction that shifts global intensity or contrast will overlap with a large central bounding box like cardiomegaly, producing a high IoU without encoding the clinical concept. The paper shows CAVs are far from random in cosine similarity, but it never shows that traversing along a random vector produces substantially worse attribution maps under the same step size and reconstruction protocol. Without that null control, the observed IoU gain does not demonstrate concept-specific explanation quality.\n\nMy overall read: this is an honest, preliminary empirical report, not an overclaiming mess. The missing random-direction control is a genuine gap but a curable one. The paper deserves a serious referee, and I would engage with it in a reading group, but I wouldn't cite the explanation-quality result until the control is added and the consistency issues are fixed.","headline":"Honest preliminary CAV+latent shift study for chest X-rays; the stability claim holds up, but the explanation-quality claim needs a random-direction control before it convinces.","tokens_in":7667,"tokens_out":2066,"would_cite":false,"duration_ms":19947,"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":"The paper claims that user-defined clinical concepts can be recovered as stable directions in a chest X-ray autoencoder's latent space, and that traversing along those directions produces visual counterfactual explanations that highlight…","keywords":["Interpretability","Concept Activation Vectors","Autoencoders","Counterfactual Explanations","Visual Explanations","Explainable AI","Feature Attribution","Chest X-rays"],"falsifier":"Recompute the cross-dataset cosine similarities after permuting the positive and negative labels within each dataset, or after training on NIH positives versus CheXpert negatives; if the similarity does not drop toward the random-vector baseline near zero, the stability in Table 1 is not evidence of concept capture. A complementary check is to measure reconstruction error inside the annotated bounding boxes for small pathologies, since the paper itself attributes the atelectasis failures to reconstruction fidelity.","tokens_in":6702,"feed_emoji":"🩻","tokens_out":10316,"duration_ms":90614,"temperature":0.7,"pith_summary":"This paper tries to establish that clinical concepts, such as an enlarged heart, have recognizable directions in the mathematical space learned by a simple image-reconstruction model, and that moving an X-ray's representation along those directions makes the feature appear or disappear. If true, this matters because explanations would be stated in clinician-defined terms rather than in model gradients, and because the directions transfer across datasets, the same concept vectors could be used on images from other hospitals. The authors' evidence is a set of stability measurements, with an averaged cross-dataset cosine similarity of 0.406 versus near zero for random vectors, and localization measurements, with a cardiomegaly IoU of 0.363 for the averaged NIH vector versus 0.286 for the Latent Shift baseline. They are explicit that the method does not beat the baseline overall and that small pathologies remain challenging due to reconstruction limits. The contribution is therefore a demonstration of a feasible, label-free route to concept-based counterfactual explanations, with limitations clearly stated.","feed_headline":"X-ray concept directions survive dataset changes","feed_subtitle":"Averaged concept vectors score 0.363 IoU on cardiomegaly vs. the 0.286 baseline.","key_machinery":"The central object is the Concept Activation Vector (CAV): the vector orthogonal to the decision boundary of a linear classifier trained in the autoencoder's flattened latent space to separate positive concept samples from negative samples. The paper uses this vector as a movable direction in latent space, adding it to an image's embedding to exaggerate the concept and subtracting it to curtail the concept; decoding both modified embeddings and taking the absolute pixelwise difference produces the 2D attribution map used as the visual explanation. The same traversal protocol and IoU evaluation are inherited from the Latent Shift baseline, with the step size fixed at 10 and concept vectors averaged over multiple batches.","core_discovery":"The paper's central claim is that Concept Activation Vectors computed in the latent space of a reconstruction autoencoder recover user-defined clinical concepts, even though the autoencoder is trained only to reconstruct images and never sees pathology labels. The vectors are obtained by training a linear classifier to separate latent codes of images labeled with the target concept from latent codes of images without it, and the vector orthogonal to the decision boundary is taken as the concept direction. Averaging several such vectors yields a stable direction: the mean NIH and mean CheXpert vectors for the overlapping pathologies have a cosine similarity of 0.406, versus near zero for random vectors. Traversing an image's embedding along this direction and decoding produces exaggerated and curtailed reconstructions, and the pixelwise difference between them is used as a visual explanation. The paper's best result is for cardiomegaly, where the NIH averaged vector reaches an Intersection over Union (IoU) of 0.363 against 0.286 for the Latent Shift baseline; for smaller and more variably localized pathologies such as atelectasis the IoU is much lower, and the method does not outperform the baseline across the board.","pith_inferences":["The authors do not run this experiment, but the same concept vectors could be used as probes on other chest X-ray models: measuring whether a classifier's decision boundary aligns with the NIH-mean cardiomegaly vector would test whether the model and the concept share a common geometry.","IoU rewards tight pixel overlap with bounding boxes, so it may undervalue explanations that point a clinician to the right region without exact alignment; a small reader study asking clinicians to localize the concept from the attribution map would test the interpretability claim more directly.","If stability across datasets is genuine, a practical byproduct is a dataset-comparability check: a new, unlabeled X-ray collection whose concept directions lie close to the NIH-mean vectors is probably clinically similar to the training distribution.","The linear-separator premise could be stress-tested by fitting a low-dimensional concept subspace, for example the principal components of the positive-set latent codes, and traversing along several components; this would reveal whether the concept is better described by a cone or a manifold than by a single direction."],"forward_implications":["Averaging several concept vectors noticeably improves their stability and cross-dataset alignment, so a mean vector rather than a single-batch vector is the more reliable unit of explanation.","Because concept vectors transfer between datasets, the same directions could in principle be used to explain images from another scanner or hospital without retraining the generative model.","The generative model itself needs no label training: clinical knowledge enters only through the concept vectors, so one reconstruction model can be steered toward many different concepts.","For large, consistently located pathologies such as cardiomegaly the traversal localizes the feature better than the baseline, while small, variably located pathologies need better reconstruction fidelity before the same mechanism can work.","Modeling a concept as a subspace or a distribution, rather than a single vector, is the paper's own next step for handling concept variability."],"supporting_citations":[{"why":"Introduces Concept Activation Vectors and the linear-classifier procedure the paper uses to build concept directions.","marker":"(Kim et al., 2018)"},{"why":"Supplies the Latent Shift traversal method, the IoU evaluation protocol, and the quantitative baseline the paper compares against.","marker":"(Cohen et al., 2021)"},{"why":"Provides the pre-trained autoencoder models whose latent spaces are used for concept vectors and traversal.","marker":"(Cohen et al., 2022)"},{"why":"Supplies the NIH chest X-ray dataset, its disease labels, and the bounding-box annotations used to compute IoU for explanations.","marker":"(Wang et al., 2017)"},{"why":"Supplies the CheXpert dataset and labels used for the cross-dataset concept-vector stability experiments.","marker":"(Irvin et al., 2019)"},{"why":"Supports the paper's stated limitation that explainability benchmarks are lacking, so IoU is only a proxy.","marker":"(Holmberg, 2022)"},{"why":"Supports the claim that IoU is a proxy that needs to be complemented by human evaluation of explanations.","marker":"(Naveed et al., 2024)"}],"fun_headline_variants":["Stable concept vectors pinpoint cardiomegaly on X-rays","Concept directions survive dataset shifts in chest X-rays","Latent space concepts beat baseline on cardiomegaly","Averaged CAVs stabilize chest X-ray explanations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the premise that a straight boundary learned to separate images that carry a disease label from images that do not points at the disease itself rather than at incidental differences between the two image groups; if that premise fails, the reported stability and overlap numbers would not show that the explanations are about clinical concepts.","fun_headline_variants_meta":{"raw":{"variants":["Stable concept vectors pinpoint cardiomegaly on X-rays","Concept directions survive dataset shifts in chest X-rays","Latent space concepts beat baseline on cardiomegaly","Averaged CAVs stabilize chest X-ray explanations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1337,"prompt_tokens":903,"completion_tokens":434,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":371}},"tokens_in":519,"tokens_out":434,"duration_ms":4657,"temperature":1.0,"reasoning_tokens":371,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:47:48.887918+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the cross-dataset cosine similarities after permuting the positive and negative labels within each dataset, or after training on NIH positives versus CheXpert negatives; if the similarity does not drop toward the random-vector baseline near zero, the stability in Table 1 is not evidence of concept capture. A complementary check is to measure reconstruction error inside the annotated bounding boxes for small pathologies, since the paper itself attributes the atelectasis failures to reconstruction fidelity.","supporting_citations":[{"cited_title":"Gifsplanation via latent shift: a simple autoencoder approach to counterfactual generation for chest x-rays","cited_arxiv_id":null,"evidence_quote":"Supplies the Latent Shift traversal method, the IoU evaluation protocol, and the quantitative baseline the paper compares against."},{"cited_title":"Viviano, Paul Bertin, Paul Morrison, Parsa Torabian, Matteo Guarrera, Matthew P Lungren, Akshay Chaudhari, Rupert Brooks, Mohammad Hashir, and Hadrien Bertrand","cited_arxiv_id":null,"evidence_quote":"Provides the pre-trained autoencoder models whose latent spaces are used for concept vectors and traversal."},{"cited_title":"Towards Benchmarking Explainable Artificial Intelligence Methods","cited_arxiv_id":"2208.12120","evidence_quote":"Supports the paper's stated limitation that explainability benchmarks are lacking, so IoU is only a proxy."}],"review_version":1}