MorphoSHAP: Rethinking the Unit of Attribution in Explanation for Deep Visual Models
Anirudh Prabhakaran, Alexandre Rocchi, Gianni Franchi
cs.CV · cs.AI
Quick verdict
MorphoSHAP makes morphological shapes the Shapley players for image attribution; strong insertion/deletion results, but faithfulness metrics need scrutiny.
Summary
MorphoSHAP proposes a post-hoc attribution method that replaces pixels, superpixels, or patches with morphological shapes extracted from the Tree of Shapes as players in a KernelSHAP game. Each shape carries a scale label (eight relative-area bins), a geometric label (six-class vocabulary assigned by deterministic rules or an MLP trained on 5,000 synthetic shapes), and a signed Shapley contribution, enabling heatmaps, waterfall plots, template-based textual explanations, and dataset-level class-shape statistics from one attribution. Across five datasets (ImageNet, Galaxy10, EuroSAT, Waterbirds, TissueMNIST) and three architectures, the method reports the best Insertion AUC on four datasets (e.g., EuroSAT 0.885) and lowest Deletion AUC on three, at 0.13 s/image on ImageNet (about 6x faster than SLIC-KernelSHAP). A user study (43 participants, 976 trials) shows 60.1% preference and 89.7% guess-the-class accuracy, though Grad-CAM scores higher on accuracy (91.7%) and the claimed decision-time advantage is not statistically significant by the authors' own ANOVA. Key limitations: weak on texture-driven predictions; geometry tagging only 74.4% accurate on GTSRB; shape-level insertion/deletion may not be comparable to pixel-level baselines; and no code release is stated.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The choice of elementary attribution unit is a central design decision in visual XAI, and morphological shapes from the Tree of Shapes are better players than pixels, regular patches, or superpixels.
- MorphoSHAP is the first SHAP-based image attribution framework to combine spatial, geometric, textual, and dataset-level (global) explanations from a single shared vocabulary.
- Across five datasets and three architectures it achieves the best Insertion AUC on ImageNet, Galaxy10, EuroSAT, and TissueMNIST and the lowest Deletion AUC on ImageNet, EuroSAT, and Waterbirds, at about 0.13 s per ImageNet image.
- A 43-participant user study shows its explanations are preferred (60.1% of trials) and support high guess-the-class accuracy (89.7%) with the fastest median decision time.
Strengths
Genuinely novel attribution unit with a principled hierarchy
Defining the Shapley game over Tree-of-Shapes morphological shapes (Sec. 3.2) gives nested, self-dual, multi-scale players without external segmentation (contrast SAM-based Explain Any Concept) and with only black-box logit access (Eq. 19); Sec. 2.3 credibly positions this as the first morphology-to-XAI link used as attribution units.
Unusually thorough ablations and component validation
Appendix C.3 ablates sampling budget (S9: plateau at N=1024), overlap threshold (S10), scale granularity (S11), and background baselines (S12, honestly showing pure-black masking collapses EuroSAT Insertion to 0.786); Tables S13-S14 validate the tagger itself (pipeline accuracy 77.86% MLP vs 59.83% rules; GTSRB ground-truth geometry recovery 74.44% vs 57.00%).
Strong speed with a clear structural reason
Honest scoping of wins and losses
Deterministic textual generation avoids hallucination
Weaknesses
Insertion/deletion comparison is unit-dependent
User-study statistics contradict the prose
Internal inconsistencies in numbers
Tagger choice contradicts its own justification
Main configuration is dominated in its own ablation; no code
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
MorphoSHAP (arXiv 2609.25815, cs.CV/cs.AI) is a method paper that repositions the *unit* of visual attribution. Instead of pixels, superpixels, or regular patches, the players of the Shapley game are morphological shapes extracted with the Tree of Shapes (ToS) — a self-dual hierarchical decomposition (Sec. 3.2). Each shape is tagged with a scale level (eight relative-area bins, App. A.1) and a geometric category (six-class vocabulary via deterministic rules or a learned MLP, Sec. 3.4), then KernelSHAP (N=1024 coalitions) estimates signed contributions (Sec. 3.5). The paper claims: (i) unit choice is a first-class design decision; (ii) the shared shape vocabulary unlocks spatial, textual, and dataset-level explanations from one attribution; (iii) state-of-the-art faithfulness on most of five datasets x three architectures; (iv) human preference in a 43-participant study.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is MorphoSHAP and how does it differ from regular SHAP?
MorphoSHAP is a post-hoc, model-agnostic image attribution method that runs the Shapley game over morphological shapes extracted from the Tree of Shapes decomposition, rather than over pixels, superpixels, or patches. Each shape is tagged with a scale level (eight relative-area bins) and a geometric category (circle, rectangle, triangle, elongated, polygon, complex), so one attribution yields a heatmap, a SHAP-style waterfall, a template-based text summary, and dataset-level shape statistics without any language model.
Does MorphoSHAP outperform Grad-CAM and KernelSHAP?
In the paper's Table 1 it reports the best Insertion AUC on ImageNet (0.876), Galaxy10 (0.876), EuroSAT (0.885) and TissueMNIST (0.820), and the lowest Deletion AUC on ImageNet, EuroSAT and Waterbirds; but SHAP-BPT wins some entries (e.g., Waterbirds insertion 0.907, Galaxy10 deletion 0.074). Note the insertion/deletion values are computed in each method's own explanatory unit space, so cross-method numbers should be read with caution.
Is code or data available for MorphoSHAP?
The paper text reviewed here does not mention a released code repository. The method is described in enough detail to reimplement (scale thresholds, overlap rule, KernelSHAP budget N=1024, MLP architecture trained on 5,000 synthetic shapes), and all five evaluation datasets plus the three model architectures are public, but the Tree of Shapes decomposition and the synthetic tagging pipeline make replication nontrivial.
What are the main limitations of MorphoSHAP?
The authors acknowledge it may fail when predictions rely on fine textures rather than shapes. Independently, its geometric naming is only 74.4% accurate on GTSRB traffic signs (rules: 57.0%), so textual and dataset-level explanations can misname the governing structure; shape-level insertion/deletion metrics may not be directly comparable to pixel-level baselines; and the user-study decision-time advantage is not statistically significant by the paper's own ANOVA.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-4 and S5-S9 are available only as captions in the extracted text; visual quality of the heatmaps, waterfalls, and study interface could not be assessed.
- Table S4 extraction may contain formatting noise, but the mismatches flagged between the A.5 narrative and the table involve large, systematic discrepancies, not single-digit rounding.
- Appendix A.4's sentence justifying the learned MLP with 'maximum transparency, domain-independence' is reproduced as written and appears self-contradictory; it could be a drafting error.
- References could not be externally verified (e.g., ShapBPT AAAI 2026, AA-Shap 2026); novelty assessment relies on the paper's related-work section.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.