Pith. sign in

REVIEW 3 major objections 5 minor 26 references

COGENT: Counterfactual Gaussian Explanations for Volumetric Medical Images

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Counterfactual edits in Gaussian space find Sybil's tumour evidence

desk verdict Clever idea, but the central localization experiment is circular: the counterfactual is only allowed to move Gaussians inside the tumour mask and is then scored against that same mask. read the letter →

arxiv 2608.11422 v1 pith:REKLGJMP submitted 2026-08-11 cs.CV

Dorian Rząsa , Bartosz Zabdyr , Krzysztof Piekarz , Jakub Grzywaczewski , Bartlomiej Sobieski , Przemyslaw Biecek , Żaneta Świderska-Chadaj , Olga Śliwicka
show 2 more authors
Przemys{l}aw Spurek Joanna Świebocka-Więk
This is my paper · ORCID
classification cs.CV
keywords counterfactualexplanations3DGaussiansplattingvolumetricmedicalimaginglungcancerriskpredictionSybilMedGSexplainableAIcomputedtomography
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

COGENT asks a different question from standard saliency maps: instead of highlighting voxels, it treats the CT scan as a 3D scene built from Gaussian primitives and searches for the smallest parameter changes to those primitives that flip Sybil's predicted lung-cancer risk. The paper claims that the Gaussians that change most mark the evidence driving the model, and that this localization lands on the ground-truth tumour: COGENT scores 0.2837 on Relative Random Area, far above pixel-space baselines like GradCAM (0.0805) and Saliency (0.0702). If the claim holds, representation-space counterfactual optimization gives clinicians a way to see which part of a scan the model relies on, in a form that stays anatomically coherent.

What carries the argument

The load-bearing object is the Gaussian primitive set Θ = {θ_i} with θ_i = (position, color, opacity, scale, rotation), rendered by MedGS into 2D slices that are assembled into V(Θ), the volume fed to Sybil. Because the entire chain Θ → R(Θ, c_j) → V(Θ) → f_m(V(Θ)) → L_cf is differentiable, gradients from Sybil's raw logits flow back to individual Gaussians. COGENT selects a subset Ω of Gaussians via anatomical-mask filtering, applies random-start projected gradient descent with an L∞ budget ε around the initial parameters, and ranks Gaussians by the magnitude of their parameter changes; the top 10% are highlighted as the explanation. Raw logits instead of calibrated probabilities are used to avoid gradient saturation.

What would settle it

Run COGENT with the mask-filtering step disabled, allowing all Gaussians to be optimized, and compute RRA against the ground-truth mask; if the top-changed Gaussians fall largely outside the lesion while risk still flips, the localization claim is refuted. Separately, feed Sybil both the original CT volume and the MedGS-rendered volume of the same scan and compare risk predictions; material divergence would mean the counterfactual is explaining the rendering, not the scan.

Watch

Extended reading notes

Core claim

The paper's central claim is that counterfactual optimization performed directly in the parameter space of a 3D Gaussian scene representation—not in voxel space—identifies the radiological features that drive Sybil's risk prediction. The pipeline reconstructs each LDCT volume with MedGS, renders slices that Sybil scores, and runs projected gradient descent on the Gaussian parameters (position, opacity, scale, color) under an L∞ bound, using raw risk logits as the objective so gradients stay strong. The most-changed Gaussians then form the explanation. COGENT reports the best localization among compared methods (RRA and RRA_abs = 0.2837, versus 0.0805 for GradCAM, 0.0702 for Saliency, 0.0254 for Integrated Gradients, 0.0218 for Input×Gradient, 0.0134 for Kernel SHAP) while staying competitive on sparsity and perturbation metrics, and expert review rated 40% of counterfactual edits as attenuating malignancy markers, 40% neutral, and 20% as exacerbating them.

Load-bearing premise

The localization score is measured against a tumour mask that was already used to decide which Gaussians COGENT may modify, so high RRA may be a consequence of that pre-selection rather than proof that Sybil's evidence is the visible lesion.

Editorial extensions

If this is right

  • If COGENT is correct, explainability for volumetric medical AI no longer needs to operate on pixels or voxels; any differentiable scene representation can host counterfactual explanations.
  • Clinicians could audit Sybil-type risk models by asking which localized 3D region, when altered, moves the prediction, rather than by inspecting 2D heatmaps.
  • Because the counterfactual volume stays close to the original under an L∞ bound, the explanation corresponds to a minimal, anatomically plausible edit, which is closer to what a radiologist would consider a cause.
  • The framework is presented as a general paradigm: it should transfer to other volumetric predictors as long as the rendering pipeline is differentiable.

Reading between the lines

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

  • The reported localization advantage is partly by construction: Figure 2 shows that mask filtering restricts optimization to Gaussians inside the tumour, so RRA measures how well COGENT uses an already-tumour-confined set rather than whether the model's evidence is the lesion. Running COGENT without the mask filter would test the stronger claim.
  • The method implicitly assumes that Sybil's risk judgment on MedGS-rendered slices equals its judgment on the original CT; if the rendering pipeline distorts low-dose CT appearance, the counterfactual may explain the rendering rather than the scan.
  • A natural testable extension is to use the same Gaussian-space optimization to locate evidence for other volumetric biomarkers (e.g., emphysema or fibrosis) where the relevant mask is not predefined, letting the optimizer discover regions of interest without annotation.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces COGENT, a counterfactual explanation method for volumetric medical images that operates directly in the parameter space of 3D Gaussian splatting representations (MedGS). COGENT first selects Gaussian primitives via mask filtering, then optimizes their parameters with projected gradient descent to shift the logits of the Sybil lung cancer risk prediction model, and finally uses the most-changed Gaussians as the explanation. The authors evaluate localization by comparing the resulting attribution map to ground-truth tumour masks (reporting RRA/RRA_abs) and supplement this with expert ratings. The central claim is that Gaussian-space counterfactual explanations localize the evidence driving Sybil's prediction to the tumour region.

Significance. If the localization claim were valid, COGENT would offer a novel and potentially useful representation-space explainability paradigm for volumetric medical imaging, with publicly released code and a fresh integration of MedGS with a clinical risk model. However, the evaluation design contains a fundamental circularity: the method's candidate Gaussians are preselected using the ground-truth tumour mask, and the localization score is computed against that same mask. This makes the central claim unfalsifiable in its current form. The expert evaluation, which shows only 40% positive ratings, does not rescue the claim. The underlying idea is interesting, but the manuscript's evidence is not.

major comments (3)
  1. [Figure 2 and Method description] The optimization is restricted to Gaussians preselected via mask filtering; Figure 2's caption states that 'mask filtering keeps only the Gaussians inside the tumour,' and the Method section describes identifying 'the Gaussians intersecting regions defined by anatomical masks.' Table 1 then scores the resulting explanation against the same ground-truth lesion mask (RRA/RRA_abs). Because the candidate Gaussians are confined to the tumour mask, the reported RRA = 0.2837 is largely determined by the preselection step rather than by whether Sybil's evidence actually lies in the lesion. The pixel-space baselines are not given this mask-based preselection, so the comparison in Table 1 is not apples-to-apples. This circularity invalidates the central claim that COGENT localizes Sybil's evidence; the experiment cannot falsify the possibility that Sybil relies on features outside the visible lesion, because outside-mask Gaussians are never eligible for selection.
  2. [Preliminary studies / MedGS and Method description] The counterfactual objective is computed on V(Theta), the volume rendered from MedGS slices, but the explanation is claimed to explain Sybil's prediction on the original CT volume. The paper does not report whether Sybil's raw hazard logits on V(Theta) are close to those on the original CT volume. Without such a fidelity check, the optimized gradients may be explaining the reconstructed volume rather than the original scan, and the counterfactual may succeed only because of rendering artifacts. The authors should report a concordance measure (e.g., logit correlation or prediction agreement) between original and reconstructed volumes for the evaluation dataset, and restrict or qualify the claims accordingly.
  3. [Expert Qualitative Evaluation and Results] The paper reports that only 40% of the evaluated cases received positive expert ratings (+1/+2), with 40% neutral and 20% negative. The statement that 'these findings demonstrate that Gaussian-space counterfactual explanations effectively localize radiologically relevant disease features' is not supported by these numbers; a 40% positive rate is inconclusive, and the conclusion is overstated. The expert evaluation should be presented as a preliminary qualitative observation, not as confirmatory evidence for the localization claim.
minor comments (5)
  1. [Table 1] No error bars, confidence intervals, or statistical significance tests are reported, and the number of scans in the evaluation is not stated.
  2. [Experiments] RRA and RRA_abs are not formally defined in the main text; a precise definition with the exact formula is needed to interpret the reported values.
  3. [Figure 3] The pixel-space baseline is not identified; the text only refers to 'a pixel-space counterfactual baseline' without specifying which method produced it.
  4. [Method description] The hyperparameters used for the main experiments (number of PGD steps, learning rate, epsilon, number of evaluated scans) are not given in the main text or the appendix; only a limited epsilon sensitivity analysis appears in the appendix.
  5. [General] The manuscript contains numerous typographical and formatting issues (e.g., missing spaces in the abstract, inconsistent capitalization of 'Gaussian Splatting'), which should be corrected in revision.

Circularity Check

1 steps flagged · score 8.0 of 10

Tumour-mask prefiltering makes the RRA localization result circular: the explanation is selected from Gaussians inside the lesion and then scored against the same lesion mask.

  1. self definitional [Figure 2 caption; Method description (subset selection); Table 1 / Results (localization score)]
    "mask filtering keeps only the Gaussians inside the tumour. Their parameters are optimized with PGD using gradients from the frozen Sybil classifier"

    The set of optimizable Gaussians is defined by the tumour mask (Figure 2: 'mask filtering keeps only the Gaussians inside the tumour'; Method: 'identifying the Gaussians intersecting regions defined by anatomical masks'). Since the explanation is formed by ranking members of that set by parameter-change magnitude, every retained Gaussian lies inside the lesion mask by construction. Table 1 then scores the explanation by 'the overlap between the attribution map and the ground-truth lesion mask' — the same mask used for filtering — and reports RRA = 0.2837. This overlap is a partial restatement of the selection constraint, not an independent measurement of whether Sybil's evidence is the visible lesion. Baselines do not receive the mask prefilter, so the comparison is biased.

full rationale

The paper's central quantitative claim is that COGENT localizes Sybil's evidence to the ground-truth tumour, supported by Table 1 (RRA = RRA_abs = 0.2837, well above all baselines). That claim is circular. Figure 2 states that 'mask filtering keeps only the Gaussians inside the tumour' and that only those parameters are optimized; the Method section likewise selects 'the Gaussians intersecting regions defined by anatomical masks.' The explanation is then the top-ranked subset of those same in-mask Gaussians, so by construction every explained primitive lies inside the lesion mask. Table 1 then measures 'the overlap between the attribution map and the ground-truth lesion mask' — the same mask used for filtering. The reported localization is therefore a restatement of the selection constraint, not an independent probe of where Sybil's evidence lies, and the comparison to pixel-space baselines (which are not mask-filtered) is not apples-to-apples. The qualitative expert evaluation is not itself circular, and the use of MedGS/Sybil as external components is legitimate; but the headline localization result, as quantified, reduces to the input mask by construction.

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

No new particles, forces, or physical entities are introduced; the only new artifact is the COGENT framework itself, which is a method rather than an entity. The main burden is the use of ground-truth tumour masks both to constrain the counterfactual and to score localization, plus the unvalidated fidelity of MedGS rendering for Sybil.

free parameters (3)
  • epsilon (L-infinity perturbation bound) = swept 0.1, 0.5, 1.0, 2.0, 4.0; final value not fixed
    Chosen by hand; controls how much Gaussians can move and thus the magnitude of the counterfactual edit.
  • top-k threshold for most changed Gaussians = 10%
    The attribution map is built from the 10% most-changed Gaussians; RRA and sparseness depend on this choice.
  • number of PGD steps = 10 (appendix), N in main text
    Optimization length is a free hyperparameter; sensitivity to N is not reported.
assumptions (3)
  • domain assumption MedGS-rendered volume V(Theta) is a faithful proxy for the original CT as far as Sybil is concerned.
    The gradient chain Theta -> R -> V -> Sybil is meaningful only if Sybil's prediction on the render matches its prediction on the original scan; not validated in the paper. See MedGS paragraph and Eq. V(Theta)=P(...).
  • domain assumption Ground-truth tumour masks are available, accurate, and used to restrict optimization.
    Figure 2 caption states mask filtering keeps only Gaussians inside the tumour; the accuracy of these masks is not assessed.
  • standard math PGD with random start finds a meaningful counterfactual in the constrained Gaussian parameter space.
    PGD is a standard adversarial tool; there is no guarantee that the found local optimum corresponds to a clinically meaningful counterfactual.

how reviews work

0 comments
Cite this review

Pith. "Pith review of COGENT: Counterfactual Gaussian Explanations for Volumetric Medical Images." pith.science (2026). https://pith.science/paper/REKLGJMP

@misc{pith2026260811422,
  author       = {Pith},
  title        = {Pith review of: COGENT: Counterfactual Gaussian Explanations for Volumetric Medical Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/REKLGJMP}},
  note         = {Machine review of arXiv:2608.11422}
}
read the original abstract

Explainability is essential for deploying deep learning models in high-stakes medical applications. Existing explainability methods for volumetric imaging predominantly operate in voxel space, overlooking the structured representations introduced by recent advances in 3D scene modeling. We present COGENT (Counterfactual Gaussian Explanations), a framework that generates counterfactual explanations directly in the parameter space of Gaussian-based volumetric representations. Built upon MedGS and the Sybil lung cancer risk prediction model, COGENT optimizes selected Gaussian primitives through a differentiable rendering pipeline, enabling gradients from the downstream predictor to identify representation components that most influence model decisions. Unlike conventional pixel- or voxel-level attribution methods, our approach formulates explainability as a counterfactual optimization problem over an explicit 3D scene representation, producing sparse and spatially localized explanations while preserving anatomical consistency. We evaluate COGENT on lung CT scans using quantitative comparisons with existing explainability methods together with qualitative analysis by medical experts. The results demonstrate that representation-space counterfactual optimization provides clinically meaningful explanations while offering a new perspective on interpreting volumetric deep learning models.

Figures

Figures reproduced from arXiv: 2608.11422 by the authors.

Figure 1
Figure 1. Overview of the COGENT framework. Given a scan that Sybil flags as high risk, COGENT asks what would make it healthy, removing the suspicious nodule with a localized Gaussian perturba￾tion while leaving the surrounding anatomy untouched (before/after insets). Re-scoring the edited volume with the same model flips the prediction to low risk, revealing the evidence behind the original decision. prediction, and disease… view at source ↗
Figure 2
Figure 2. Gaussian-space counterfactual explanation generated for Sybil. An LDCT volume and its tumour mask are encoded [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison between the original CT [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Tumour localisation in the 3D lung model and the corresponding local CT appearance. The highlighted Gaussian [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Representative counterfactual explanations generated by COGENT. Expert ratings are shown above each example. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 7
Figure 7. Figure 7: Original axial CT slices (top) and their correspond [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 6
Figure 6. Figure 6: Average change in probability per step as a function [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 18 canonical work pages

  1. [1]

    Journal of Clinical Oncology , volume=

    Sybil: A Validated Deep Learning Model to Predict Future Lung Cancer Risk from a Single Low-Dose Chest Computed Tomography Scan , author=. Journal of Clinical Oncology , volume=

    work page

  2. [2]

    ACM SIGGRAPH Conference Proceedings , year=

    3D Gaussian Splatting for Real-Time Radiance Field Rendering , author=. ACM SIGGRAPH Conference Proceedings , year=

    work page

  3. [3]

    arXiv preprint arXiv:2509.16806 , year=

    MedGS: Gaussian Splatting for Multi-Modal 3D Medical Imaging , author=. arXiv preprint arXiv:2509.16806 , year=

    arXiv

  4. [4]

    arXiv preprint arXiv:2602.02560 , year=

    Auditing Sybil: Explaining Deep Lung Cancer Risk Prediction Through Generative Interventional Attributions , author=. arXiv preprint arXiv:2602.02560 , year=

    arXiv

  5. [5]

    International Conference on Learning Representations , year=

    Explaining and Harnessing Adversarial Examples , author=. International Conference on Learning Representations , year=

  6. [6]

    2025 , issn =

    Interpretable machine learning for time-to-event prediction in medicine and healthcare , journal =. 2025 , issn =. doi:https://doi.org/10.1016/j.artmed.2024.103026 , url =

    work page arXiv 2025

  7. [7]

    Explainable AI Methods - A Brief Overview

    Holzinger, Andreas and Saranti, Anna and Molnar, Christoph and Biecek, Przemyslaw and Samek, Wojciech. Explainable AI Methods - A Brief Overview. xxAI - Beyond Explainable AI: International Workshop, Held in Conjunction with ICML 2020, July 18, 2020, Vienna, Austria, Revised and Extended Papers. 2022. doi:10.1007/978-3-031-04083-2_2

    doi:10.1007/978-3-031-04083-2_2 2020

  8. [8]

    International Conference on Learning Representations , year=

    Towards Deep Learning Models Resistant to Adversarial Attacks , author=. International Conference on Learning Representations , year=

Show all 26 references
  1. [9]

    Advances in Neural Information Processing Systems , year=

    A Unified Approach to Interpreting Model Predictions , author=. Advances in Neural Information Processing Systems , year=

  2. [10]

    IEEE International Conference on Computer Vision , pages=

    Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization , author=. IEEE International Conference on Computer Vision , pages=

  3. [11]

    International Conference on Machine Learning , year=

    Axiomatic Attribution for Deep Networks , author=. International Conference on Machine Learning , year=

  4. [12]

    PLoS ONE , volume=

    On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation , author=. PLoS ONE , volume=

  5. [13]

    Harvard Journal of Law and Technology , volume=

    Counterfactual Explanations without Opening the Black Box: Automated Decisions and the GDPR , author=. Harvard Journal of Law and Technology , volume=

  6. [14]

    2024 , publisher=

    Interpretable Machine Learning , author=. 2024 , publisher=

    2024

  7. [15]

    Explanatory Model Analysis , author=

  8. [16]

    International Conference on Machine Learning , year=

    Position: Explain to Question, Not to Justify , author=. International Conference on Machine Learning , year=

  9. [17]

    European Conference on Artificial Intelligence , year=

    Model Science: Getting Serious About Verification, Explanation and Control of AI Systems , author=. European Conference on Artificial Intelligence , year=

  10. [18]

    International Conference on Artificial Intelligence and Statistics , year=

    From Shapley Values to Generalized Additive Models and Back , author=. International Conference on Artificial Intelligence and Statistics , year=

  11. [19]

    International Conference on Machine Learning , year=

    The Shapley Taylor Interaction Index , author=. International Conference on Machine Learning , year=

  12. [20]

    Explainable AI: Interpreting, Explaining and Visualizing Deep Learning , author=

  13. [21]

    Natural Language Processing and Chinese Computing , pages=

    Explainable AI: A brief survey on history, research areas, approaches and challenges , author=. Natural Language Processing and Chinese Computing , pages=. 2019 , organization=

    2019

  14. [22]

    BMC Medical Imaging , year=

    Explainable artificial intelligence (XAI) in medical imaging: a systematic review of techniques, applications, and challenges , author=. BMC Medical Imaging , year=

  15. [23]

    Nature machine intelligence , volume=

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead , author=. Nature machine intelligence , volume=. 2019 , publisher=

    2019

  16. [24]

    Radiology: Artificial Intelligence , volume=

    TotalSegmentator: robust segmentation of 104 anatomic structures in CT images , author=. Radiology: Artificial Intelligence , volume=. 2023 , publisher=

    2023

  17. [25]

    2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=

    Aggregated attributions for explanatory analysis of 3D segmentation models , author=. 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=. 2025 , organization=

    2025

  18. [26]

    International Conference on Machine Learning , pages=

    Manifold Integrated Gradients: Riemannian Geometry for Feature Attribution , author=. International Conference on Machine Learning , pages=. 2024 , organization=

    2024

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.