Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ReaMIL: Reasoning- and Evidence-Aware Multiple Instance Learning for Whole-Slide Histopathology

Hyun Do Jung, Jungwon Choi, Hwiyoung Kim

TL;DR

ReaMIL addresses the interpretability gap in MIL for whole-slide histopathology by learning a compact, evidence-based set of tiles that suffices for slide-level predictions. It augments a strong MIL backbone with a lightweight evidence head that produces soft tile selections via a Concrete gate, creating three slide views (full, keep, drop) and training with a budgeted-sufficiency objective that enforces sufficiency, exclusion, contiguity, and sparsity. On TCGA-NSCLC, TCGA-BRCA, and PANDA, ReaMIL maintains or slightly improves AUC while dramatically reducing the number of tiles needed for high-confidence decisions (MSK $\ o$ approximately 8.2 tiles at $\tau=0.90$, AUKC $\to$ roughly 0.864), demonstrating sharp confidence accumulation with small, spatially coherent evidence sets. The approach requires only slide-level labels, integrates with standard MIL pipelines, and yields slide-level overlays, offering practical interpretability for clinical deployment without additional supervision.

Abstract

We introduce ReaMIL (Reasoning- and Evidence-Aware MIL), a multiple instance learning approach for whole-slide histopathology that adds a light selection head to a strong MIL backbone. The head produces soft per-tile gates and is trained with a budgeted-sufficiency objective: a hinge loss that enforces the true-class probability to be $\geq τ$ using only the kept evidence, under a sparsity budget on the number of selected tiles. The budgeted-sufficiency objective yields small, spatially compact evidence sets without sacrificing baseline performance. Across TCGA-NSCLC (LUAD vs. LUSC), TCGA-BRCA (IDC vs. Others), and PANDA, ReaMIL matches or slightly improves baseline AUC and provides quantitative evidence-efficiency diagnostics. On NSCLC, it attains AUC 0.983 with a mean minimal sufficient K (MSK) $\approx 8.2$ tiles at $τ= 0.90$ and AUKC $\approx 0.864$, showing that class confidence rises sharply and stabilizes once a small set of tiles is kept. The method requires no extra supervision, integrates seamlessly with standard MIL training, and naturally yields slide-level overlays. We report accuracy alongside MSK, AUKC, and contiguity for rigorous evaluation of model behavior on WSIs.

ReaMIL: Reasoning- and Evidence-Aware Multiple Instance Learning for Whole-Slide Histopathology

TL;DR

ReaMIL addresses the interpretability gap in MIL for whole-slide histopathology by learning a compact, evidence-based set of tiles that suffices for slide-level predictions. It augments a strong MIL backbone with a lightweight evidence head that produces soft tile selections via a Concrete gate, creating three slide views (full, keep, drop) and training with a budgeted-sufficiency objective that enforces sufficiency, exclusion, contiguity, and sparsity. On TCGA-NSCLC, TCGA-BRCA, and PANDA, ReaMIL maintains or slightly improves AUC while dramatically reducing the number of tiles needed for high-confidence decisions (MSK approximately 8.2 tiles at , AUKC roughly 0.864), demonstrating sharp confidence accumulation with small, spatially coherent evidence sets. The approach requires only slide-level labels, integrates with standard MIL pipelines, and yields slide-level overlays, offering practical interpretability for clinical deployment without additional supervision.

Abstract

We introduce ReaMIL (Reasoning- and Evidence-Aware MIL), a multiple instance learning approach for whole-slide histopathology that adds a light selection head to a strong MIL backbone. The head produces soft per-tile gates and is trained with a budgeted-sufficiency objective: a hinge loss that enforces the true-class probability to be using only the kept evidence, under a sparsity budget on the number of selected tiles. The budgeted-sufficiency objective yields small, spatially compact evidence sets without sacrificing baseline performance. Across TCGA-NSCLC (LUAD vs. LUSC), TCGA-BRCA (IDC vs. Others), and PANDA, ReaMIL matches or slightly improves baseline AUC and provides quantitative evidence-efficiency diagnostics. On NSCLC, it attains AUC 0.983 with a mean minimal sufficient K (MSK) tiles at and AUKC , showing that class confidence rises sharply and stabilizes once a small set of tiles is kept. The method requires no extra supervision, integrates seamlessly with standard MIL training, and naturally yields slide-level overlays. We report accuracy alongside MSK, AUKC, and contiguity for rigorous evaluation of model behavior on WSIs.
Paper Structure (19 sections, 5 equations, 3 figures, 3 tables)

This paper contains 19 sections, 5 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Multiple instance learning for whole-slide histopathology
  4. Interpretability of attention in MIL
  5. Budgeted evidence and selective prediction
  6. Methodology
  7. Problem setup and backbone
  8. Evidence selection head
  9. Evidence-aware training objectives
  10. Evidence-efficiency metrics
  11. Minimal Sufficient K (MSK).
  12. Area Under K-Curve (AUKC).
  13. Experiments
  14. Datasets and setup
  15. Slide-level performance
  16. ...and 4 more sections

Figures (3)

  • Figure 1: Overview of ReaMIL. Frozen UNI2-h features and patch coordinates are extracted from each WSI and mapped to tokens with positional embeddings. An evidence head produces soft selection scores $z \in (0,1)^N$ via a Concrete (Gumbel--sigmoid) gate, and defines three bags: the full bag $x$, a keep bag $z \cdot x$, and a drop bag $(1-z) \cdot x$. All three bags are processed by a shared TransMIL encoder and slide head. Losses encourage (i) correct predictions on the full and keep bags (cross-entropy on $\ell_{\text{full}}$ and $\ell_{\text{keep}}$ plus a sufficiency hinge at confidence $\tau$), (ii) low true-class probability on the drop bag (exclusion), (iii) spatially compact selections (contiguity on coordinates), and (iv) a small evidence budget via an $\ell_1$ penalty on $z$. At test time, the model outputs both slide predictions and ranked evidence coordinates. Reasoning metrics are computed by probing the top-K curve of true-class probability $p_y(K)$: AUKC summarizes the area under this curve, and $\mathrm{MSK}@\tau$ measures the minimal number of tiles required to reach confidence $\tau$.
  • Figure 2: K-curve on NSCLC (test set). True-class probability $p_y(K)$ as top-$K$ tiles (ranked by selection score) are revealed. Solid line: mean across slides; shaded region: $\pm$1 std. Vertical dashed line: mean MSK@$\tau=0.9$. MSK is computed per-slide before averaging, so individual slides may cross $\tau$ even when the mean curve does not.
  • Figure 3: Evidence visualization on TCGA-NSCLC. Left: LUSC (squamous cell carcinoma) case with relatively compact evidence clusters over squamous tumor nests. Right: LUAD (adenocarcinoma) case with more diffuse selection over gland-forming tumor regions. Each panel shows selected tile locations (green boxes) and the corresponding top-$K$ zoomed patches. For visualization, we show zoomed-in regions (left: $8192 \times 8192$; right: $16384 \times 16384$ pixels), where the selected tiles (size $256 \times 256$) are outlined in green.