Pith. sign in

REVIEW 4 major objections 4 minor 25 references

3D ReX: Causal Explanations in 3D Neuroimaging Classification

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 2502.12181 v3 pith:QX67M2LK submitted 2025-02-14 eess.IV cs.AIcs.CVcs.LG

Melane Navaratnarajah , Sophie A. Martin , David A. Kelly , Nathan Blake , Hana Chockler This is my paper · ORCID
classification eess.IVcs.AIcs.CVcs.LG
keywords 3DReXcausalexplanationresponsibilitymapactualcausalitystrokedetectionexplainableAIneuroimagingmodel-agnostic
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

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.

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 (4)
  1. [Section 3, Algorithm 2] 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.
  2. [Section 4, Examples of Results] 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.
  3. [Section 4, Settings for 3D-ReX Explanations] 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.
  4. [Algorithm 1, lines 6 and 16] 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.
minor comments (4)
  1. [Throughout] 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.
  2. [Section 4, Examples of Results] 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.
  3. [Section 4, Settings for 3D-ReX Explanations] 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.
  4. [Section 4, Data] 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.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity in the derivation chain; minor non-load-bearing self-citations and a post hoc occlusion-value choice weaken the evaluation but do not make the method's output reduce to its inputs.

full rationale

The core derivation of 3D ReX is self-contained. The responsibility map is constructed by occluding supervoxels with the trained model, querying the model on each mutant, and accumulating responsibility for partitions that preserve the original classification. Nowhere does the algorithm use the lesion annotation as an input; the lesion is used only as an external reference for visual comparison. The theoretical foundation in actual causality comes from Halpern and Pearl (2005) and Chockler and Halpern (2004), which are prior formal works rather than assumptions of the present result. The 3D adaptation in Algorithm 2 is an engineering modification that is honestly described as a computational compromise, and the paper explicitly disclaims comprehensive evaluation: 'We do not claim a comprehensive evaluation of 3D-REX. Numerical evaluation of a large-scale dataset remains to be done.' Two concerns are present but do not rise to derivation circularity. First, the paper self-cites prior REX and Blake et al. work to motivate the framework and the zero occlusion value; these self-citations are not load-bearing because the current method still derives its output from fresh model queries, and the paper performs its own comparison of occlusion strategies. Second, the choice of the 0-value occlusion was made after comparing several options on the same test example, so the subsequent observation that 0-value occlusion overlaps the lesion is weakened by selection bias; however, this is an evaluation weakness, not a case where a fitted parameter is presented as a prediction that is forced by construction. The Algorithm 2 two-axis random split has no coverage guarantee, which is a genuine robustness and correctness risk, but that is a sampling limitation rather than a circular definition. Overall, no equation or algorithmic step in the derivation is equivalent to its own input by construction, and no central claim rests solely on a self-citation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims rest on the established causal responsibility formalism plus three practical assumptions about pass/fail testing, random splitting, and lesion annotations as ground truth. The only fitted value is the occlusion value, selected post hoc. No new physical or conceptual entities are postulated.

free parameters (3)
  • occlusion value = 0
    Chosen after comparing mean intensity, healthy MRI patch, and zero; the zero value was selected because it produced explanations overlapping the lesion location (Section 4). This is a post hoc selection of a hyperparameter based on the desired result.
  • dmax (max depth)
    Search depth limit in Algorithm 1; no value or tuning procedure given in the paper.
  • lmax (search limit)
    Search step limit in Algorithm 1; no value or tuning procedure given in the paper.
assumptions (4)
  • domain assumption Halpern-Pearl actual causality and responsibility theory correctly capture 'cause' for classifier decisions.
    The paper inherits the causal responsibility framework from Chockler and Halpern (2004, 2024) without justifying its applicability to deep network classification (Sections 2-3).
  • ad hoc to paper Pass/fail is determined by exact equality of the hard class prediction (prediction = target), not by probability changes.
    Algorithm 1 line 16 compares predictions for equality; this ignores confidence shifts and may make responsibility maps sensitive to decision-boundary artifacts.
  • ad hoc to paper Random two-axis quadrant splits sample the 3D occlusion space sufficiently.
    Algorithm 2 chooses two random axes and random split coordinates; there is no proof or sensitivity study that this captures 3D lesion structure.
  • domain assumption ATLAS lesion annotations are an appropriate ground truth for judging explanation quality.
    The evaluation compares explanations to annotated lesion locations (Section 4, Figures 1-2), assuming these annotations are accurate and relevant to the model's learned features.

how reviews work

0 comments
Cite this review

Pith. "Pith review of 3D ReX: Causal Explanations in 3D Neuroimaging Classification." pith.science (2026). https://pith.science/paper/QX67M2LK

@misc{pith2026250212181,
  author       = {Pith},
  title        = {Pith review of: 3D ReX: Causal Explanations in 3D Neuroimaging Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QX67M2LK}},
  note         = {Machine review of arXiv:2502.12181}
}
read the original abstract

Explainability remains a significant problem for AI models in medical imaging, making it challenging for clinicians to trust AI-driven predictions. We introduce 3D ReX, the first causality-based post-hoc explainability tool for 3D models. 3D ReX uses the theory of actual causality to generate responsibility maps which highlight the regions most crucial to the model's decision. We test 3D ReX on a stroke detection model, providing insight into the spatial distribution of features relevant to stroke.

Figures

Figures reproduced from arXiv: 2502.12181 by the authors.

Figure 1
Figure 1. Visualization of the model’s explanation for patient A, predicted as stroke with a confidence of [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. 3D rendering of patient A’s results showing the pa [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 21 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Blake, N.; Kelly, D.; Pe \ n a, S.; Chanchal, A.; and Chockler, H. 2024. Explainable AI for the classification of brain MRIs. Nature Portfolio

    work page 2024

  4. [4]

    Chockler, H.; and Halpern, J. Y. 2004. Responsibility and Blame: A Structural-Model Approach. J. Artif. Intell. Res., 22: 93--115

    work page 2004

  5. [5]

    Chockler, H.; and Halpern, J. Y. 2024. Explaining Image Classifiers. arXiv preprint arXiv:2401.13752

    arXiv 2024

  6. [6]

    A.; Kroening, D.; and Sun, Y

    Chockler, H.; Kelly, D. A.; Kroening, D.; and Sun, Y. 2024. Causal Explanations for Image Classifiers. arXiv preprint arXiv:2411.08875

    arXiv 2024

  7. [7]

    Esteva, A.; Robicquet, A.; Ramsundar, B.; Kuleshov, V.; DePristo, M.; Chou, K.; Cui, C.; Corrado, G.; Thrun, S.; and Dean, J. 2019. A guide to deep learning in healthcare. Nature medicine, 25(1): 24--29

    work page 2019

  8. [8]

    C.; and Vedaldi, A

    Fong, R. C.; and Vedaldi, A. 2017. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, 3429--3437

    work page 2017

Show all 25 references
  1. [9]

    K.; and Jimma, W

    Gurmessa, D. K.; and Jimma, W. 2023. A comprehensive evaluation of explainable Artificial Intelligence techniques in stroke diagnosis: A systematic review. Cogent Engineering, 10(2): 2273088

    2023

  2. [10]

    Halpern, J. Y. 2016. Actual Causality. The MIT Press. ISBN 9780262336611

    2016

  3. [11]

    Y.; and Pearl, J

    Halpern, J. Y.; and Pearl, J. 2005. Causes and Explanations: A Structural-model Approach. Part II : Explanations. British Journal for the Philosophy of Science, 56(4)

    2005

  4. [12]

    Imperial College London . 2015. IXI Dataset – Brain Development

    2015

  5. [13]

    Jin, W.; Li, X.; Fatehi, M.; and Hamarneh, G. 2023. Guidelines and evaluation of clinical explainable AI in medical image analysis. Medical Image Analysis, 84: 102684

    2023

  6. [14]

    Lenis, D.; Major, D.; Wimmer, M.; Berg, A.; Sluiter, G.; and B \"u hler, K. 2020. Domain aware medical image classifier interpretation by counterfactual impact analysis. In Medical Image Computing and Computer Assisted Intervention--MICCAI 2020: 23rd International Conference, ...

    2020

  7. [15]

    P.; Donnelly, M

    Liew, S.-L.; Lo, B. P.; Donnelly, M. R.; Zavaliangos-Petropulu, A.; Jeong, J. N.; Barisano, G.; Hutton, A.; Simon, J. P.; Juliano, J. M.; Suri, A.; Wang, Z.; Abdullah, A.; Kim, J.; Ard, T.; Banaj, N.; Borich, M. R.; Boyd, L. A.; Brodtmann, A.; Buetefisch, C. M.; Cao, L.; Cassi...

    2022

  8. [16]

    M.; and Lee, S.-I

    Lundberg, S. M.; and Lee, S.-I. 2017. A Unified Approach to Interpreting Model Predictions. In Guyon, I.; Luxburg, U. V.; Bengio, S.; Wallach, H.; Fergus, R.; Vishwanathan, S.; and Garnett, R., eds., NIPS, 4765--4774. Curran Associates, Inc

    2017

  9. [17]

    Monai, P. 2020. The MONAI Consortium . Type: Computer Program

    2020

  10. [18]

    Petsiuk, V.; Das, A.; and Saenko, K. 2018. RISE: Randomized Input Sampling for Explanation of Black-box Models. In British Machine Vision Conference ( BMVC ) . BMVA Press

    2018

  11. [19]

    Why Should I Trust You?

    Ribeiro, M. T.; Singh, S.; and Guestrin, C. 2016. "Why Should I Trust You?": Explaining the Predictions of Any Classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '16, 1135–1144. New York, NY, USA: Association ...

    2016

  12. [20]

    Saeed, W.; and Omlin, C. 2023. Explainable AI (XAI): A systematic meta-survey of current challenges and future opportunities. Knowledge-Based Systems, 263: 110273

    2023

  13. [21]

    R.; Cogswell, M.; Das, A.; Vedantam, R.; Parikh, D.; and Batra, D

    Selvaraju, R. R.; Cogswell, M.; Das, A.; Vedantam, R.; Parikh, D.; and Batra, D. 2017. Grad- CAM : Visual explanations from deep networks via gradient-based localization. In International Conference on Computer Vision (ICCV), 618--626. IEEE

    2017

  14. [22]

    Sheu, R.-K.; and Pardeshi, M. S. 2022. A Survey on Medical Explainable AI (XAI): Recent Progress, Explainability Approach, Human Interaction and Scoring System. Sensors, 22(20)

    2022

  15. [23]

    Uzunova, H.; Ehrhardt, J.; Kepp, T.; and Handels, H. 2019. Interpretable explanations of black box classifiers applied on medical images by meaningful perturbations using variational autoencoders. In Medical Imaging 2019: Image Processing, volume 10949, 264--271. SPIE

    2019

  16. [24]

    H.; Kuijf, H

    Van der Velden, B. H.; Kuijf, H. J.; Gilhuijs, K. G.; and Viergever, M. A. 2022. Explainable artificial intelligence (XAI) in deep learning-based medical image analysis. Medical Image Analysis, 79: 102470

    2022

  17. [25]

    A.; Kafiabadi, S.; Al Busaidi, A.; Guilhem, E.; Montvila, A.; Lynch, J.; Townend, M.; Agarwal, S.; Mazumder, A.; Barker, G

    Wood, D. A.; Kafiabadi, S.; Al Busaidi, A.; Guilhem, E.; Montvila, A.; Lynch, J.; Townend, M.; Agarwal, S.; Mazumder, A.; Barker, G. J.; et al. 2022. Deep learning models for triaging hospital head MRI examinations. Medical Image Analysis, 78: 102391

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.