{"id":"d67185fa-7920-4490-83f7-72b77bf2b701","arxiv_id":"2504.13717","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A thesis compiled from the author's published works reports that activation maximization cannot interpret medical image classifiers, while prototype-based and feature-co-occurrence methods improve explainable and robust medical classification.","lead":"This thesis compiles peer-reviewed work on explainable medical-image deep learning: activation maximization fails on chest X-rays, prototype models align with radiology, and co-occurrence-based 'causal' modules plus a biology-inspired network improve classification robustness. A generalist reader might care because it documents a concrete research path toward trustworthy clinical AI and organizes the XAI-causality literature into a scaffold.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"§6.1's 'causality map' is feature co-occurrence; with no causal identification or intervention test, co-occurrence is association. The thesis's core claim that weak causal signals are exploited, and CROCODILE/CoCoReco's causal grounding, are therefore unsupported.","rationale":"I read the manuscript as a thesis compiling published work, with the central claim defined in Chapter 1.2.2: feature co-occurrence in medical images yields weak causal signals that can be exploited without a priori knowledge, improving classification, explainability, and OOD robustness. The visible portions of Chapters 1–5 are methodologically careful: the ProtoPNet study includes patient-stratified cross-validation, hold-out test sets, and radiologist feedback, which count as real evidence for the explainability part. However, the causality chapters (6–8) are where the headline claim lives, and the full experimental bodies are truncated in the provided text, so the numbers could not be verified. The decisive weakness is conceptual rather than statistical: Section 6.1 defines the causality map as feature co-occurrence. Co-occurrence is a symmetric second-order association; it can reflect confounding, selection bias, or a shared cause. Without an identification argument (e.g., an assumed SCM, an environment-invariance condition, or a do-intervention on the image generation process), the map does not warrant the causal label. This is not an 'outside consensus' disagreement: the thesis itself adopts Pearl's ladder and warns against exactly this confusion in Sections 2.4.2–2.4.3. The concern is therefore internal consistency with the paper's own standard. If the modules are relabelled as attention-like feature enhancement using co-occurrence statistics, the accuracy and explainability findings may remain intact, but the abstract's key finding (iv) and CROCODILE's stated confounder-removal mechanism would need to be withdrawn or substantially weakened. This matches the reader's weakest assumption, so my agreement is 'agree.' The appropriate verdict is CONDITIONAL: the visible work is competent, but the central causal claim must be either supported by an identification argument and intervention-based validation or explicitly reframed as association-based feature enhancement. I would not reject outright, because the engineering contributions and the explainability chapters stand on their own; I would not accept, because the headline claim is currently unverified. Since the reader's verdict is already CONDITIONAL, I recommend no change to the verdict.","tokens_in":50041,"tokens_out":4587,"duration_ms":43218,"concrete_test":"Re-run the main and OOD experiments of Chapter 6 (and, if feasible, Chapters 7 and 8) with the 'causality map' component replaced by an identically computed co-occurrence map that carries no causal label, and with a standard channel-attention baseline (e.g., SE block) of matched parameter count, on the same datasets and splits. If the three conditions are statistically indistinguishable on ID and OOD metrics and on a localization metric such as pointing-game accuracy, the causal framing adds no measurable content and the central claim is vacuous. If the causality-labelled version wins, inspect the learned map to see whether it isolates known causal regions and test with a synthetic intervention that breaks a known spurious co-occurrence (e.g., removing a confounded region) to confirm the gain is not driven by the artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is the co-occurrence-as-causality identification in Chapter 6, Section 6.1 ('Causality Signals in Images'). The causality map is computed from feature co-occurrence statistics in CNN feature maps, and this map is then used as the causal signal in the causality-factors extraction module (Chapter 6.2.1), in CROCODILE's prior injection (Chapter 7.1.5), and in CoCoReco's Contextual Attention Block (Chapter 8.1.2). The thesis provides no structural causal model, no do-calculus derivation, and no controlled intervention on the image-generating process that would distinguish a cause from a mere statistical co-occurrence. This matters because the manuscript itself, in Sections 2.4.2 and 2.4.3, correctly emphasizes that association is not causation and that shortcut learning stems from spurious correlations. A scanner artifact, a demographic attribute, or a hospital-specific marker can co-occur with a disease label without causing it; a module that upweights such co-occurring features would amplify the shortcut rather than remove it. Consequently, the central claim in Section 1.2.2—that the modules 'automatically discover and exploit weak causal signals without a priori knowledge'—is not supported by the evidence presented, and the robustness story of CROCODILE (Chapter 7) inherits this gap. The accuracy numbers may survive, but the stated mechanism does not.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This thesis, a cumulative PhD dissertation posted on arXiv, argues for 'human-aligned deep learning' along three axes: explainability, causality, and biological inspiration. Two empirical chapters evaluate XAI methods for medical classifiers: activation maximization (AM), which the author finds to produce uninterpretable visualizations for chest X-ray models and motivates a set of novel image-prior losses; and ProtoPNet for breast mass classification, validated with patient-stratified cross-validation, a genuinely held-out test set, and expert radiologist feedback. A systematic review chapter proposes a scaffold that organizes the XAI-causality literature. The core methodological contribution is a family of 'causality-driven' modules (Chapters 6-8) built on the notion of a 'causality map,' operationalized as feature co-occurrence statistics in CNN feature maps, and deployed in the CROCODILE robustness framework and the CoCoReco biologically inspired network. The load-bearing claim, stated in Section 1.2.2, is that these modules automatically discover and exploit 'weak causal signals' without a priori knowledge, improving classification, interpretability, and out-of-distribution robustness.","tokens_in":50282,"tokens_out":11931,"duration_ms":98576,"significance":"The thesis is honestly reported and contains genuinely solid components. Chapter 4 stands out: its evaluation protocol (patient-stratified 5-fold CV plus a held-out test set, explicitly fixing the original ProtoPNet evaluation flaw) and its single-radiologist clinical assessment provide a valuable template for explainable-by-design medical imaging research, and code is shipped for this chapter. Chapter 3's negative results and Chapter 5's systematic scaffold are useful contributions. The significance of the central causality claim, however, is conditional on an identification step—linking feature co-occurrence to causation—that the manuscript never justifies. If the causal framing fails, the modules reduce to associative attention-like enhancements and the mechanistic robustness story is weakened, even if the reported accuracy numbers survive. As an engineering contribution the work is real but modest; as a demonstration that weak causal signals can be exploited without prior knowledge, the claim is, as written, unsupported.","major_comments":[{"comment":"The load-bearing claim of the thesis, stated in Section 1.2.2, is that the proposed modules 'automatically discover and exploit weak causal signals in medical images without requiring prior knowledge' and that this 'improves classification and produces more robust predictions by focusing on the relevant parts of the image.' The mechanism behind this claim is the 'causality map,' defined in Section 6.1 as feature co-occurrence statistics extracted from CNN feature maps, on which the causality-factors extraction module (§6.2.1), CROCODILE's prior injection (§7.1.5), and CoCoReco's Contextual Attention Block (§8.1.2) all build. The manuscript never provides the identification step that would connect co-occurrence to causation: there is no structural causal model, no do-calculus derivation, no intervention on the image-generating process, and no test distinguishing a genuinely causal feature from a merely associated one. This gap is consequential for the causal framing because the thesis itself, in Sections 2.4.2 and 2.4.3, correctly establishes that association is not causation and that shortcut learning arises from spurious correlations: a scanner artifact, a demographic attribute, or a hospital-specific marker can co-occur with a disease label without causing it, and a module that upweights such co-occurring features would amplify the shortcut rather than remove it. The robustness story of Chapter 7 and the abstract's domain-generalization claim inherit this premise. Two repairs are feasible within the manuscript's scope: (a) add an identification test, such as showing that the discovered maps track clinically known causal features, or that removing spurious co-occurrences leaves performance intact, or a capacity-matched comparison against a purely associative attention baseline; or (b) reframe the contribution as attention-like co-occurrence feature enhancement and remove the causal terminology accordingly. Without one of these, the central mechanistic claim is unsupported even if the reported accuracies hold.","section":"§6.1, §1.2.2, §7.1.5, §8.1.2"},{"comment":"Key finding (i), summarized as 'accuracy rises, interpretability falls,' is drawn from Section 3.3.4, where two training regimes (transfer learning vs. end-to-end fine-tuning) are compared on two binary tasks (finding vs. no-finding and male vs. female) from ChestX-ray14. The accuracies are reported as single values with no error bars or multiple seeds (e.g., 66.76% vs. 68.98%; 79.79% vs. 95.81%), and the interpretability judgment is made by the author's visual inspection of the AM images in Figure 3.6 rather than by any quantitative interpretability measure, user study, or external rater. Since the chapter concludes a general trade-off from this comparison, the evidence base is thin. Reporting variance across seeds, matching the training budgets of the two regimes more carefully, and adding a quantifiable proxy for AM interpretability (for example, a distance-based measure of deviation from a prototypical medical image, or ratings from multiple observers) would strengthen the claim; alternatively, the conclusion should be explicitly scoped to the two configurations studied.","section":"§3.3.4"}],"minor_comments":[{"comment":"'ViTss divide an image' is a typo for 'ViTs divide an image.'","section":"§2.1.2"},{"comment":"The weights λ1–λ4 of the proposed medical-image AM priors are not reported, and the text only states that tuning 'proved to be hard.' For a chapter whose principal result is negative, reporting the attempted ranges and final values would materially aid reproducibility.","section":"Eq. (3.6)"},{"comment":"The loss components are introduced in the order Classification, Symmetry, Histogram, Noise, Frequency, but appear in Eq. (3.6) as LC, LH, LN, LS, LF; aligning the order would prevent confusion.","section":"Eq. (3.6) and following list"},{"comment":"The causality-factors extraction module introduces free parameters such as co-occurrence thresholds and the number of factors; their values and an ablation over them should be reported so that the results are reproducible and the module's sensitivity is visible.","section":"§6.2.1"},{"comment":"The results and discussion supporting the abstract's domain-generalization claim are summarized very briefly in the thesis; reproducing the key quantitative tables would allow the reader to verify the in-distribution and OOD claims directly.","section":"§7.3"},{"comment":"The heading 'Exod-IA: the Forbidden One' is informal for a formal dissertation; a neutral heading (e.g., 'Exod-IA: toward a clinical decision support system') would be more appropriate.","section":"§9.2.1"}],"recommendation":"major_revision","confidential_remarks":"The self-citation pattern is expected for a cumulative dissertation and does not raise concerns by itself; each technical chapter is based on externally peer-reviewed publications and Chapter 4 ships code. The principal risk to the manuscript is the causal framing: if the authors cannot supply any identification test, the claims in §1.2.2 and finding (iv) of the abstract must be softened. The editor may also wish to consider whether the extensive background material and dissertation format fit the venue, and whether the 'causality' terminology in the titles of Chapters 6-8 fairly represents the content, given that the operational definition is feature co-occurrence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the thing to know: this is a PhD thesis posted to arXiv, composed almost entirely of the author's own peer-reviewed papers, chapters 4-8 being self-described compilations. There is no single new result here. For a thesis that is normal; for an arXiv research paper it means the novelty is in the framing.\n\nThat framing actually does useful work. Chapter 5's review of the XAI-causality literature is a genuine scaffold, and the updated search makes it a reasonable entry point. Chapter 4 is the strongest piece: patient-stratified five-fold CV, a held-out test set that explicitly fixes a flaw in the original ProtoPNet evaluation, and a radiologist assessment of the prototypes. That chapter deserves to be taken seriously. Chapter 3 is honest in the right way—it reports that the proposed medical-image priors for activation maximization did not work, and it says so without spin.\n\nThe soft spot is the one the stress-test note hits: Section 6.1 defines a 'causality map' as feature co-occurrence statistics computed from CNN feature maps, and this map is the causal signal for everything that follows—the causality-factors module, CROCODILE's prior injection, CoCoReco's attention block. The thesis itself, in Section 2.4.2, correctly explains that association is not causation and that shortcut learning is exactly the problem of spurious correlation. By that standard, no argument is given to show that the co-occurrence statistics capture causes rather than (say) hospital-specific markers or scanner artifacts. Upweighting such features could amplify shortcuts. So the load-bearing mechanistic claim—that the models exploit 'weak causal signals' and that robustness comes from causal grounding—is unsupported. The accuracy numbers may survive; the stated mechanism does not.\n\nThere are smaller issues. Chapter 3 reports accuracies without error bars and draws the 'accuracy rises, interpretability falls' conclusion from a qualitative reading. The experimental bodies of Chapters 6-8 are truncated in this arXiv version, so the OOD claims cannot be checked. And because the chapters are self-compilations of already-published work, the citation pattern is fine but the document adds little beyond the constituent papers.\n\nWho is this for? Anyone who wants a single-place overview of a coherent research arc in medical-image XAI, and a useful warning example of co-occurrence being branded as causality. It is not a paper that should be published as a new contribution without revision.\n\nFor peer review: if the venue accepts thesis-style or survey-style submissions, send it to referees, but require the author to either relabel 'causality-driven' as 'co-occurrence-driven' or supply a causal identification argument, and to add error bars and code for Chapters 6-8. If the venue expects a monolithic new result, desk reject.","headline":"A competent, honestly assembled thesis that re-presents already published work; its central 'causality-driven' claim equates co-occurrence with causation and needs a clear fix before the framework's robustness story can be taken seriously.","tokens_in":50965,"tokens_out":2701,"would_cite":false,"duration_ms":24908,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The thesis claims that CNN feature co-occurrence, packaged as a 'causality map,' can substitute for causal knowledge in medical image classification, improving accuracy, explainability, and out-of-distribution generalization.","keywords":["explainable AI","medical image classification","feature co-occurrence","causal deep learning","domain generalization","prototypical part learning","out-of-distribution robustness","biologically inspired vision"],"falsifier":"Train a causality-module classifier on chest X-rays into which a task-irrelevant but consistently co-occurring artifact has been embedded (for example, a synthetic watermark in one corner of every positive case), then evaluate on a test set from which the watermark is removed; a map that genuinely captures causal structure should ignore the watermark and keep accuracy, whereas an association map will depend on it and show a sharp drop.","tokens_in":49682,"feed_emoji":"🩻","tokens_out":8554,"duration_ms":72567,"temperature":0.7,"pith_summary":"This thesis argues that deep learning can be aligned with human reasoning by treating explainability, causality, and biological vision as one problem, and it stakes its central technical claim on a simple operational move: feature co-occurrence statistics extracted from CNN feature maps are 'weak causal signals' that a classifier can exploit without any causal knowledge about how the images were made. If that move holds, medical image classifiers can become more accurate, more explainable, and more reliable on out-of-distribution data than ordinary attention-based models, which matters because shortcut learning and domain shift are the main reasons medical AI fails in the clinic. The thesis supports the claim across three connected studies: prototypical-part learning for breast mass classification that a radiologist confirms aligns with clinical reasoning; causality-driven CNN modules that improve classification and explanation quality; and the CROCODILE framework that generalizes across medical domains. The biological-vision study, CoCoReco, extends the same co-occurrence idea into a context-aware attention mechanism, and the net contribution is a parameter-light recipe for causal-looking robustness that does not require intervention data.","feed_headline":"Feature co-occurrence stands in for causality in medical imaging","feed_subtitle":"Mining CNN feature co-occurrence boosts diagnostic accuracy and explainability, no causal labels required.","key_machinery":"The load-bearing object is the 'causality map': a two-dimensional score field derived from feature co-occurrence statistics computed over a CNN's feature maps, which the paper treats as a proxy for causal relevance in a medical image. Around it the thesis builds a 'causality-factors extraction' module that selects and re-weights feature maps, an attention-inspired enhancement scheme, and, in CROCODILE, a mechanism that injects the map as prior knowledge into a dual-branch architecture that disentangles disease features from domain features; in CoCoReco the same co-occurrence principle appears as a Contextual Attention Block that models visual context. The causality map is what carries the argument from association to the claimed causal robustness, so everything the thesis says about explainable, robust, generalizing classifiers depends on this one operationalization.","core_discovery":"On the paper's own terms, the discovery is that 'causality signals' need not be learned from interventions or expert graphs but can be read off from the statistical structure of a trained network: features that co-occur consistently in medical images carry weak causal information, and packaging that co-occurrence into a 'causality map' lets a CNN enhance the feature maps that matter, suppress those that do not, and thereby classify better while focusing on the relevant image regions. The thesis presents this as the thread linking its main results: activation maximization fails on medical images because it cannot separate class-relevant structure from anatomical regularity; prototypical part learning succeeds because it grounds decisions in radiologically meaningful parts; the causality-driven modules improve accuracy, explanation quality, and one-shot learning; CROCODILE shows that injecting the causality map as a task prior, alongside contrastive disentanglement of disease and domain branches, yields robust generalization across datasets and to out-of-distribution data; and CoCoReco shows that the same co-occurrence mechanism, dressed as biological context, improves recognition without added trainable parameters.","pith_inferences":["If co-occurrence is sufficient as a causal proxy, the thesis's recipe removes the need for causal discovery or intervention data in a whole class of medical imaging tasks; a natural test is to pit the causality-map module against a method using an explicit causal graph on the same out-of-distribution benchmark.","Because the causality map is a statistic of the training distribution, it will absorb any spurious correlation that is stable in training data, so the module's robustness may fail precisely on the shortcut it learned; an intervention experiment that removes a planted co-occurring artifact from the test set would separate causal from associational behavior.","The same co-occurrence machinery could plausibly transfer beyond radiology to histopathology, fundus photography, and dermoscopy, wherever lesions and tissue context co-occur, and it could be paired with counterfactual generation to give per-case explanations that distinguish causes from correlates."],"forward_implications":["A CNN equipped with the causality-factors module should classify more accurately and produce saliency that highlights disease-relevant image regions, improving clinician trust without retraining the backbone.","CROCODILE-style dual-branch training with a causality-map prior should transfer across medical domains and hold performance on out-of-distribution data better than standard fine-tuned CNNs.","Prototypical part learning is a viable ante-hoc explanation for breast mass classification, with prototypes that a radiologist judges clinically aligned, unlike activation maximization, which loses interpretability as accuracy rises.","The context-aware attention block of CoCoReco can be dropped into standard feed-forward architectures to improve recognition with no additional trainable parameters.","The thesis's XAI-causality scaffold gives future work a shared vocabulary, organizing the field into critics of XAI under the causality lens, XAI for causality, and causality for XAI."],"supporting_citations":[{"why":"Introduces ProtoPNet, the explainable-by-design prototypical part architecture whose applicability to breast mass classification the thesis validates and whose radiologist-aligned explanations anchor the explainability findings.","marker":"[76]"},{"why":"Supplies the CBIS-DDSM mammography dataset that defines the breast mass classification task in the ProtoPNet study.","marker":"[230]"},{"why":"Supplies the ChestX-ray14 dataset used for the activation maximization study and for the causality-module experiments.","marker":"[459]"},{"why":"Introduces the Ladder of Causation, the association-intervention-counterfactual hierarchy that frames the thesis's claim that co-occurrence counts as a weak causal signal.","marker":"[321]"},{"why":"Formalizes structural causal models and causal diagrams, the vocabulary the causality chapters adopt when describing disease and domain features.","marker":"[316]"},{"why":"Defines gray-level co-occurrence matrix texture statistics, the classical co-occurrence analysis that the thesis's causality map generalizes and operationalizes.","marker":"[153]"},{"why":"Introduces invariant risk minimization, the invariance-based domain generalization approach that CROCODILE's disentanglement strategy contrasts with and extends.","marker":"[18]"}],"fun_headline_variants":["Co-occurrence stands in for causality in medical imaging","Causality from co-occurrence: no causal labels needed","Causality maps from co-occurrence improve diagnosis","Feature co-occurrence as a causal proxy in medical AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole causality story rests on identifying feature co-occurrence with causal signal: the thesis never shows that the maps track the image-generating process rather than mere correlation, so if co-occurrence is only association, the claimed robustness and explanation improvements reduce to attention-style feature enhancement.","fun_headline_variants_meta":{"raw":{"variants":["Co-occurrence stands in for causality in medical imaging","Causality from co-occurrence: no causal labels needed","Causality maps from co-occurrence improve diagnosis","Feature co-occurrence as a causal proxy in medical AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000389,"raw_usage":{"total_tokens":2112,"prompt_tokens":1070,"completion_tokens":1042,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":686,"completion_tokens_details":{"reasoning_tokens":976}},"tokens_in":686,"tokens_out":1042,"duration_ms":9381,"temperature":1.0,"reasoning_tokens":976,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:02:51.785220+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a causality-module classifier on chest X-rays into which a task-irrelevant but consistently co-occurring artifact has been embedded (for example, a synthetic watermark in one corner of every positive case), then evaluate on a test set from which the watermark is removed; a map that genuinely captures causal structure should ignore the watermark and keep accuracy, whereas an association map will depend on it and show a sharp drop.","supporting_citations":[],"review_version":1}