{"id":"dc173966-b449-473b-bc57-67d625a47543","arxiv_id":"2502.12181","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A causality-based explainability tool for 3D medical image classifiers, demonstrated on stroke detection, produces voxel-level responsibility maps without accessing the model's internals.","lead":"This paper presents 3D ReX, a tool that explains why a 3D medical image classifier makes its decision by testing which brain regions, when masked, change the prediction. It is the first claimed causality-based explanation method for 3D models, and could give clinicians a way to see what a stroke detection network actually relies on.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The two-axis random split in Algorithm 2 has no coverage guarantee and can silently miss the causal lesion; the single qualitative example and post hoc occlusion choice do not rule this out.","rationale":"The reader's weakest-assumption analysis identifies the Algorithm 2 random two-axis sampling heuristic as the point where the argument is least secure, and my reading concurs. The paper's evidence is a single qualitative patient example, no quantitative metrics, no random-seed sensitivity analysis, and an occlusion value selected after comparing several alternatives, all of which leaves the central empirical claim underdetermined. The failure mode I describe is concrete: a lesion whose causal mass is split by the random partition can produce no passing child, causing the recursion to halt before the lesion is ever refined. This would make the responsibility map highlight nothing or an unrelated region even for a perfectly trained model. The proposed synthetic test directly settles whether this happens. I would not move the verdict to REJECT because the method may still work in practice and the paper explicitly disclaims a comprehensive evaluation; the appropriate verdict remains CONDITIONAL pending the quantitative and sensitivity checks. The reader's judgement is therefore unchanged.","tokens_in":7061,"tokens_out":5915,"duration_ms":70666,"concrete_test":"Construct a synthetic 3D volume with a single contiguous lesion-like blob and a classifier whose decision is known to depend only on that blob, e.g., a logistic model on the blob's mean intensity with a fixed threshold. Run 3D ReX with a fixed 0-occlusion value for at least 20 random seeds, then compute Dice/Jaccard overlap between the responsibility map thresholded at the reported explanation level and the true blob. Repeat for multiple blob sizes, positions, and especially blobs that straddle likely partition boundaries. If overlap is low or highly variable across seeds, or if any non-trivial blob is systematically missed, the Algorithm 2 sampling heuristic is the bottleneck. A complementary check on one seed is to compare against exhaustive eight-way splitting to quantify the information lost by dropping the third axis.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that 3D ReX's responsibility map highlights the voxels most causal to a 3D classifier's decision, with the supporting observation being that 0-value occlusion produces explanations overlapping stroke lesions. This inference rests on the sampling scheme in Algorithm 2. At each recursion, the current input box is split along only two randomly chosen axes, so the third axis remains unsplit for that child. If the causal evidence for the lesion is distributed so that no single resulting quadrant preserves the original prediction, no child passes, no responsibility is assigned, and recursion terminates without ever isolating the lesion. No convergence argument, coverage guarantee, or sensitivity analysis over random seeds is provided, and the search limits (dmax, lmax) are not reported. Because the only reported result is a single patient's visual comparison and the 0-occlusion value was chosen after trying several occlusion strategies, the current evidence cannot distinguish a genuine causal localization from an artifact of the random two-axis partitioning and post hoc occlusion selection. This is not a claim that the method is wrong, but that the central empirical assertion is unsupported until this failure mode is ruled out.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents 3D ReX, a model-agnostic, causality-based post-hoc explainability method for 3D medical images. The method extends the REX framework by partitioning an input volume into supervoxels, occluding supervoxels with a chosen value, querying the classifier, and assigning responsibility to supervoxels whose occlusion preserves the original hard-label prediction. The authors evaluate 3D ReX on a 3D ResNet18 stroke classifier trained on ATLAS and IXI MRI data, reporting that a 0-value occlusion produces explanations aligned with the annotated lesion for a single patient, while a healthy-MRI occlusion highlights different regions. The paper explicitly states that a comprehensive numerical evaluation remains future work.","tokens_in":7367,"tokens_out":3926,"duration_ms":40796,"significance":"If the central claim holds, 3D ReX would provide a useful model-agnostic alternative to gradient-based explainers for 3D neuroimaging, built on a formal theory of actual causality rather than on ad-hoc saliency. The algorithm is clearly specified in pseudocode, the method does not require access to model weights, and the authors are transparent about the limitations of their evaluation. However, the empirical support for the headline claim is currently limited to one qualitative example with a post hoc chosen occlusion value, so the contribution is best viewed as a plausible proof-of-concept rather than a validated tool. The theoretical foundation is sound, but the experimental evidence is not yet commensurate with the abstract's claim that 3D ReX highlights the regions most crucial to the model's decision.","major_comments":[{"comment":"The two-axis random split has no coverage guarantee. If the causal evidence for a lesion is distributed so that no single quadrant preserves the original prediction, then no child passes, no responsibility is assigned, and recursion terminates without isolating the lesion even for a perfectly trained model. The paper asserts that splitting on two randomly chosen axes can 'adequately capture (or break) spatial dependencies' but provides no convergence argument, no failure-probability bound, and no sensitivity analysis over random seeds. The search limits dmax and lmax are not reported either. Since the central empirical observation depends on this sampling heuristic, the authors should provide a coverage analysis or empirically demonstrate stability across multiple random runs and parameter settings.","section":"Section 3, Algorithm 2"},{"comment":"The claim that 0-value occlusion produces explanations 'closely aligned/overlapping with lesion location' is supported only by a visual comparison on one patient (Patient A). No quantitative metric such as Dice score, overlap fraction, centroid distance, or voxel-level precision/recall is reported, and there is no comparison against a baseline explainer. The Discussion candidly states that numerical evaluation of a large-scale dataset remains to be done, but this means the abstract's general claim is not yet supported by the presented evidence. Either the abstract and introduction should be tempered to reflect a proof-of-concept, or the experiments should include quantitative evaluation over a larger set of test patients.","section":"Section 4, Examples of Results"},{"comment":"The occlusion value selection is circular as presented. The authors state that they compared mean intensity, healthy-MRI patches, and a 0 value, and then report that the 0-value occlusion produced explanations aligned with the lesion. Because the same patient and the same test-set example are used both to select the occlusion value and to demonstrate the alignment, the reported result is partly a product of post hoc selection. The paper should either prespecify the occlusion-selection criterion, validate the chosen occlusion on an independent set of patients, or explicitly frame the comparison as exploratory rather than as evidence for the method's correctness.","section":"Section 4, Settings for 3D-ReX Explanations"},{"comment":"The pseudocode is ambiguous about whether 'target' is the hard class label or the raw model output vector. If line 6 'target <- m(d)' stores a probability vector, then the exact equality test 'prediction = target' in line 16 will essentially never hold for real-valued outputs, and the algorithm will produce an empty responsibility map. If the test is intended to compare hard class labels, the pseudocode should say so explicitly (for example, 'target <- argmax m(d)' or 'prediction = target' with defined hard-label equality). The distinction is load-bearing for the correctness of the algorithm and should be clarified.","section":"Algorithm 1, lines 6 and 16"}],"minor_comments":[{"comment":"The name of the tool is rendered inconsistently as '3D-R EX', '3D-REX', '3D-ReX', and '3D-R EX' in different places; please use a single consistent form.","section":"Throughout"},{"comment":"The text says the explanation is 'positioned perpendicularly in comparison to the lesion in 3D space' while also claiming close alignment/overlap with the lesion; these statements are in tension and should be reconciled or quantified.","section":"Section 4, Examples of Results"},{"comment":"The statement that mean-value occlusion 'failed to yield any interpretable results' is not supported by a definition of 'interpretable' or by any figure or metric; please clarify the criterion used.","section":"Section 4, Settings for 3D-ReX Explanations"},{"comment":"The phrase 'resized to 96×96×96mm' appears to mix voxel dimensions with physical units; if this refers to voxel counts, please correct the notation.","section":"Section 4, Data"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a medical-imaging XAI venue, and the authors make a strong novelty claim ('the first causality-based post-hoc explainability tool for 3D models'). I would encourage the editor to ask the authors to broaden the related-work search beyond the self-cited REX line to ensure the novelty claim is accurate, and to require code and data availability if the paper is accepted, since the reproducibility of the single reported example would otherwise be difficult to assess."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, what's actually new: this is a clean port of the REX responsibility-map algorithm to 3D medical volumes, using random two-axis quadrant splits to keep the search tractable. That is a real extension of the prior REX line, not a new conceptual framework. The paper is clearly written and the algorithm is specified precisely enough to reimplement. The occlusion-value comparison (zero, mean intensity, healthy patch) is a useful practical experiment, and the authors deserve credit for stating outright in the limitations that this is not a comprehensive evaluation.\n\nThe soft spots are the ones the reader flagged. The only demonstration is a single stroke patient with visual overlap against a lesion annotation. There are no numeric metrics, no baseline comparison, and no indication of how representative that patient is. The occlusion value that produced the nice overlap was chosen after comparing several options, which is post hoc selection; the paper doesn't state whether that choice was prespecified. Since the explanation changes qualitatively with the occlusion value, this weakens the central empirical claim.\n\nThe Algorithm 2 concern from the stress-test is legitimate but not fatal. Randomly splitting on two of three axes has no coverage guarantee; if the causal region is split so that no quadrant preserves the prediction, the recursion terminates without localizing it. The authors explicitly trade exhaustive coverage for computational cost, but they don't analyze sensitivity to random seeds or report dmax and lmax. That's a missing robustness analysis, not a contradiction in the method. For a methods paper, this is acceptable if presented as an open question; the paper partially does this under 'computational complexity.'\n\nCitation pattern is fine. Self-cites to REX are appropriate because the 3D adaptation builds directly on that work.\n\nBottom line: this is a workshop-quality contribution that deserves a serious referee. The algorithm is well-defined and the underlying causal theory is sound. The claims, however, need to be scaled back or the evaluation expanded with multiple patients, quantitative overlap metrics, and a comparison to existing 3D saliency methods before I would cite it as a validated tool. I'd send it to peer review with the expectation of major revision.","headline":"Honest, clearly specified 3D port of REX for stroke MRI; algorithm is sound, but evidence is one patient and a post hoc occlusion choice, so the empirical claim is unproven.","tokens_in":7785,"tokens_out":2616,"would_cite":false,"duration_ms":26265,"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":"3D ReX, a causality-based explanation tool for 3D medical classifiers, generates responsibility maps that highlight the brain regions driving a stroke-detection model's decision.","keywords":["3D ReX","causal explanation","responsibility map","actual causality","stroke detection","explainable AI","neuroimaging","model-agnostic"],"falsifier":"Run 3D ReX multiple times with different random seeds on the same high-confidence stroke MRI and measure the overlap of the resulting responsibility maps with the annotated lesion; if the maps vary widely from seed to seed or frequently miss the lesion for a correctly classified stroke, the sampling heuristic does not reliably recover the model's causal regions.","tokens_in":6821,"feed_emoji":"🧠","tokens_out":3076,"duration_ms":30283,"temperature":0.7,"pith_summary":"This paper introduces 3D ReX, a post-hoc explainability tool for 3D medical image classifiers that does not need access to a model's internal weights. It adapts a 2D causal-explanation framework so that responsibility maps are built by repeatedly occluding supervoxels and keeping only those occlusions that preserve the model's original classification. The central claim is that these maps reveal the voxel regions most crucial to a 3D model's decision, and the authors test this on a stroke-detection MRI classifier. With a zero-valued occlusion, the generated explanations overlapped closely with annotated lesion locations, suggesting the model relied on clinically relevant regions. This matters because 3D medical imaging has few model-agnostic explanation tools, and stroke data provides a rare ground truth for judging whether an explanation points to the right anatomy.","feed_headline":"Causal 3D explainer finds stroke lesions from MRI","feed_subtitle":"A model-agnostic tool builds responsibility maps that overlap with annotated lesions in stroke classification.","key_machinery":"The central mechanism is causal responsibility, as formalized in the structural-model approach to actual causality, applied through an iterative occlusion algorithm. The algorithm initializes a zero responsibility map, obtains the model's prediction on the unmodified input as a target, and then repeatedly partitions the input space. To limit computational cost, each region is split along two randomly chosen axes out of x, y, and z, producing four subregions; each subregion is masked and tested against the target, and only those mutants that preserve the classification are retained and recursively refined. Responsibility is accumulated over passing supervoxels and non-contributing supervoxels are discarded, producing a final map whose peaks mark the voxels most responsible for the decision.","core_discovery":"The paper claims 3D ReX is the first causality-based post-hoc explainability tool for 3D models. Using the theory of actual causality, it quantifies each voxel's responsibility for a classification by masking supervoxels and checking whether the model's prediction survives. The resulting responsibility map accumulates peaks in regions that are causally necessary for the output, and an approximately minimal sufficient subset of voxels can be extracted as a causal explanation. In the stroke-detection experiments, the 0-value occlusion condition produced explanations that aligned or overlapped with the annotated lesion location, indicating that the model's focus corresponded to regions typically associated with stroke pathology. The paper also reports that a healthy-MRI occlusion highlighted different brain regions, revealing that the model may draw on a more complex set of areas, and that a mean-intensity occlusion failed to yield interpretable results.","pith_inferences":["Inference: If 0-value occlusions reliably recover lesion regions across many patients, 3D ReX could act as a lesion-localization validator for classifiers trained without segmentation labels, giving a causal grounding to saliency-based claims.","Inference: The random two-axis splitting is a sampling heuristic whose coverage of the unsplit axis is untested; running the method with multiple random seeds and measuring the stability of the responsibility map would reveal whether the heuristic is reliable for contiguous lesions.","Inference: The observed divergence between 0-value and healthy-MRI explanations suggests that different occlusion values probe different causal pathways; a systematic comparison against lesion segmentation metrics could turn this observation into a quantitative evaluation protocol for 3D explainability tools."],"forward_implications":["Zero-valued occlusion can produce 3D explanations that align with annotated stroke lesion locations, offering a way to verify whether a classifier uses clinically meaningful anatomy.","Different occlusion values yield different responsibility maps, showing that the choice of occlusion value is itself a meaningful axis of analysis for understanding a model's decision.","Because 3D ReX only queries model outputs, it can be applied to any 3D classifier without needing gradients or internal activations.","The method's explanatory output can be rendered both as 2D slice overlays and as 3D volumetric explanations, supporting clinical review across anatomical planes.","The paper's approach could extend beyond stroke to other 3D medical imaging tasks with localized pathologies, such as tumor or hemorrhage detection."],"supporting_citations":[{"why":"Supplies the base REX framework that 3D ReX extends from a 2D image classifier to 3D inputs.","marker":"Chockler et al. 2024"},{"why":"Provides the theory of actual causality that underpins the notion of causal explanations used by 3D ReX.","marker":"Halpern and Pearl 2005"},{"why":"Defines causal responsibility, the quantity that the responsibility map visualizes for each voxel.","marker":"Chockler and Halpern 2004"},{"why":"Demonstrates that a 0-value occlusion with REX on MRI slices produces explanations coinciding with human segmentations, motivating the same occlusion value here.","marker":"Blake et al. 2024"},{"why":"Provides the ATLAS stroke neuroimaging dataset with lesion annotations, used as the ground truth for comparing explanations against lesion locations.","marker":"Liew et al. 2022"},{"why":"Establishes the meaningful-perturbation framework that motivates the selection and evaluation of occlusion values in perturbation-based explainability.","marker":"Fong and Vedaldi 2017"}],"fun_headline_variants":["First causal 3D explainer maps stroke lesions","Causal voxel maps explain stroke AI decisions","3D ReX: causal tool reveals stroke lesion locations","Causal AI explainer pinpoints stroke lesions in 3D"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's correctness rests on the assumption that randomly splitting an input region along two of three axes, and recursively refining only the supervoxels that preserve the classification, adequately captures the 3D regions a model actually uses.","fun_headline_variants_meta":{"raw":{"variants":["First causal 3D explainer maps stroke lesions","Causal voxel maps explain stroke AI decisions","3D ReX: causal tool reveals stroke lesion locations","Causal AI explainer pinpoints stroke lesions in 3D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000987,"raw_usage":{"total_tokens":4105,"prompt_tokens":782,"completion_tokens":3323,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":398,"completion_tokens_details":{"reasoning_tokens":3257}},"tokens_in":398,"tokens_out":3323,"duration_ms":21003,"temperature":1.0,"reasoning_tokens":3257,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T19:21:27.954884+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run 3D ReX multiple times with different random seeds on the same high-confidence stroke MRI and measure the overlap of the resulting responsibility maps with the annotated lesion; if the maps vary widely from seed to seed or frequently miss the lesion for a correctly classified stroke, the sampling heuristic does not reliably recover the model's causal regions.","supporting_citations":[{"cited_title":"Y.; and Pearl, J","cited_arxiv_id":null,"evidence_quote":"Provides the theory of actual causality that underpins the notion of causal explanations used by 3D ReX."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines causal responsibility, the quantity that the responsibility map visualizes for each voxel."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates that a 0-value occlusion with REX on MRI slices produces explanations coinciding with human segmentations, motivating the same occlusion value here."},{"cited_title":"P.; Donnelly, M","cited_arxiv_id":null,"evidence_quote":"Provides the ATLAS stroke neuroimaging dataset with lesion annotations, used as the ground truth for comparing explanations against lesion locations."},{"cited_title":"C.; and Vedaldi, A","cited_arxiv_id":null,"evidence_quote":"Establishes the meaningful-perturbation framework that motivates the selection and evaluation of occlusion values in perturbation-based explainability."}],"review_version":1}