Research map / Label-efficient and robust learning / Recognition and long-tail

Symb-xMIL: Symbolic Explanations for Multiple Instance Learning in Digital Pathology

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Prototype and concept dictionary models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Symb-xMIL, a post-hoc framework that scores how well a MIL model's sub-bag predictions align with logical rules over semantic patch labels, yielding a symbolic representation space for local and cohort-level explanation.

Problem

Heatmap-based MIL explanations show which regions matter but not how evidence from different tissue regions is combined logically, limiting interpretability of higher-order interactions in digital pathology.

Approach

Replace Harsanyi-dividend-based relevance in Symbolic XAI with direct correlation between a logical query's indicator function and the MIL model's predictions on semantic sub-bags; stack alignment scores over a query space into a symbolic representation for clustering samples.

Key results

• On synthetic MIL data, the top-aligned query matched the ground-truth rule in all three datasets (e.g., 0.951 for 4 AND 7, 0.967 for 4 to 7), outperforming Symbolic XAI on AND and OR rules.
• On Camelyon16 (AUROC 0.995 ± 0.010), clustering found a C3 cluster of 25 slides where counterfactual normal-patch transfer flipped 21/25 (84%) to tumor-positive, suggesting a Clever Hans signal.
• On TCGA-HNSCC, a symbolic HPV-like/unlike partition gave CoxPH z = 2.04 (one-sided p = 0.021, logrank p = 0.04), slightly stronger than clinical HPV status (logrank p = 0.06).

Surprising findings

• Clever Hans-like normal patches in Camelyon16 flipped 84% of counterfactual slides to tumor-positive.
• The symbolic HPV-like partition stratified survival slightly better than clinical HPV status, though the separation is exploratory.

Limitations

• Depends on quality of semantic maps and patch labels (e.g., missed tumor region in tumor_018).
• HNSCC survival findings are exploratory and confounded by other prognostic factors; not clinically actionable.
• Small HPV+ sample size (43 HPV+ slides) and moderate significance.

Experimental setup

Datasets: MNIST Camelyon16 TCGA-HNSCC

Benchmarks: Camelyon16 TCGA-HNSCC MNIST-based synthetic MIL

Metrics: AUROC alignment score (correlation) silhouette score CoxPH z-score logrank p-value permutation test p-value

Baselines: Symbolic XAI xMIL

Builds on: Symbolic XAI xMIL TransMIL LRP

Closest papers by meaning

• xMIL: Insightful Explanations for Multiple Instance Learning in Histopathology (NeurIPS 2024 · similarity 0.94)
• SI-MIL: Taming Deep MIL for Self-Interpretability in Gigapixel Histopathology (CVPR 2024 · similarity 0.89)
• Do Multiple Instance Learning Models Transfer? (ICML 2025 · similarity 0.87)
• Revisiting End-to-End Learning with Slide-level Supervision in Computational Pathology (NeurIPS 2025 · similarity 0.86)
• Patho-R1: A Multimodal Reinforcement Learning-Based Pathology Expert Reasoner (AAAI 2026 · similarity 0.86)
• Feature Re-Embedding: Towards Foundation Model-Level Performance in Computational Pathology (CVPR 2024 · similarity 0.85)