Research map / Label-efficient and robust learning / Recognition and long-tail
xMIL: Insightful Explanations for Multiple Instance Learning in Histopathology
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.
Introduces xMIL, a general MIL framework built on evidence scores, and xMIL-LRP, an LRP adaptation for attention and transformer MIL models that yields more faithful instance-level explanations on toy and histopathology tasks.
Problem
MIL explanation methods in histopathology are limited to small bags or ignore instance interactions, and common attention heatmaps do not distinguish positive from negative evidence or reliably reflect predictions.
Approach
Formalize MIL explanations via context-aware evidence scores (allowing instance interactions and signed evidence) and adapt layer-wise relevance propagation to attention MIL (AttnMIL, TransMIL) using the AH-rule and LN-rule, producing signed instance relevances that sum to the bag output and scale to large bags.
Key results
- On toy tasks, xMIL-LRP achieved the highest AUPRC-2 in most settings (e.g., 0.91 on 4-Bags AttnMIL and 0.90 on 4-Bags TransMIL vs. 0.88 for IG and 0.87 for single).
- In histopathology faithfulness (AUPC, lower is better), xMIL-LRP obtained the lowest AUPC in all eight model-dataset settings, e.g., 0.24 vs. 0.73 for single on TransMIL LUAD TP53 and 0.29 vs. 0.88 for IG on TransMIL CAMELYON16.
- Attention scores were not faithful outside CAMELYON16, with AUPC 0.94 on AttnMIL HNSC HPV, close to random (0.97).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use xMIL-LRP instead of the strongest baseline IG | AttnMIL explanations · 4-Bags | AUPRC-2: 0.88 ± 0.01 -> 0.91 ± 0.01 | Helped |
| Use xMIL-LRP instead of the strongest baseline Single | TransMIL explanations · 4-Bags | AUPRC-2: 0.85 ± 0.07 -> 0.90 ± 0.02 | Helped |
| Use xMIL-LRP instead of IG | AttnMIL explanations · Pos-Neg | AUPRC-2: 0.93 ± 0.00 -> 0.91 ± 0.01 | Hurt |
| Use xMIL-LRP instead of the strongest baseline Single | TransMIL explanations · Pos-Neg | AUPRC-2: 0.91 ± 0.02 -> 0.92 ± 0.01 | Helped |
| Use xMIL-LRP instead of the strongest baseline IG | AttnMIL explanations · Adjacent Pairs | AUPRC-2: 0.75 ± 0.03 -> 0.77 ± 0.04 | Helped |
| Use xMIL-LRP instead of the strongest baseline Single | TransMIL explanations · Adjacent Pairs | AUPRC-2: 0.77 ± 0.06 -> 0.81 ± 0.04 | Helped |
| Use xMIL-LRP instead of the best baseline | AttnMIL explanations · CAMELYON16 | AUPC: 0.61 ± 0.43 -> 0.51 ± 0.38 | Helped |
| Use xMIL-LRP instead of the best baseline | AttnMIL explanations · NSCLC | AUPC: 0.42 ± 0.26 -> 0.25 ± 0.22 | Helped |
| Use xMIL-LRP instead of the best baseline | AttnMIL explanations · HNSC HPV | AUPC: 0.78 ± 0.23 -> 0.71 ± 0.24 | Helped |
| Use xMIL-LRP instead of the best baseline | AttnMIL explanations · LUAD TP53 | AUPC: 0.38 ± 0.20 -> 0.31 ± 0.16 | Helped |
| Use xMIL-LRP instead of the best baseline | TransMIL explanations · CAMELYON16 | AUPC: 0.42 ± 0.35 -> 0.29 ± 0.30 | Helped |
| Use xMIL-LRP instead of the best baseline | TransMIL explanations · NSCLC | AUPC: 0.53 ± 0.26 -> 0.45 ± 0.26 | Helped |
Surprising findings
- Raw attention scores performed near random in faithfulness for HNSC HPV and were poor at separating positive from negative evidence in toy experiments.
- Single-instance perturbation gave nearly identical relevance (0.95-0.97) across instances despite a near-0 prediction score, suggesting context-free single-instance bags are uninformative for HPV prediction.
Limitations
- Evaluation relies on faithfulness metrics because true evidence functions are unknown for real histopathology datasets.
- Only AttnMIL and TransMIL (and AddMIL in toy settings) were evaluated; AddMIL was excluded from real-world experiments.
- Histopathology experiments use a single foundation feature extractor (CTransPath) and TransMIL excluded slides with more than 24,000 patches.
Experimental setup
Datasets: MNIST CAMELYON16 TCGA NSCLC TCGA HNSC HPV TCGA LUAD TP53
Benchmarks: 4-Bags Pos-Neg Adjacent Pairs CAMELYON16 TCGA NSCLC TCGA HNSC HPV TCGA LUAD TP53
Metrics: AUPRC-2 test AUROC AUPC (patch-dropping faithfulness)
Baselines: Attention single Gradient x Input Integrated Gradients Additive MIL attention rollout random
Builds on: layer-wise relevance propagation Attention MIL TransMIL MILLI Additive MIL
Closest papers by meaning
- Symb-xMIL: Symbolic Explanations for Multiple Instance Learning in Digital Pathology (NeurIPS 2026 · similarity 0.94)
- SI-MIL: Taming Deep MIL for Self-Interpretability in Gigapixel Histopathology (CVPR 2024 · similarity 0.89)
- Revisiting End-to-End Learning with Slide-level Supervision in Computational Pathology (NeurIPS 2025 · similarity 0.88)
- Do Multiple Instance Learning Models Transfer? (ICML 2025 · similarity 0.88)
- Feature Re-Embedding: Towards Foundation Model-Level Performance in Computational Pathology (CVPR 2024 · similarity 0.86)
- Memory-Efficient Prompt Tuning for Incremental Histopathology Classification (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.