Histomorphology-Guided Prototypical Multi-Instance Learning for Breast Cancer WSI Classification

TL;DR

This work tackles the challenge of embedding histomorphology into MIL-based WSI classification for breast cancer. It introduces HGPMIL, a histomorphology-guided prototypical MIL framework with three components: (i) a histomorphology-centric importance estimation network combining a Cellularity Prediction Network and Architecture Grading Network, (ii) histomorphology-prototypical clustering that creates prototypes for biologically meaningful regions, and (iii) histomorphology-guided prototypical aggregation that informs robust WSI representations. The approach yields consistent improvements in molecular subtyping, cancer subtyping, and survival analysis across multiple datasets and MIL baselines, with ablations showing contributions from both tumor cellularity and architecture prototypes. The method demonstrates strong generalizability and interpretability, and the authors provide code to facilitate adoption and further research in computational pathology.

Abstract

Histomorphology is crucial in cancer diagnosis. However, existing whole slide image (WSI) classification methods struggle to effectively incorporate histomorphology information, limiting their ability to capture key pathological features. Particularly when the number of instances within a bag is large and their features are complex, it becomes challenging to accurately identify instances decisive for the bag label, making these methods prone to interference from ambiguous instances. To address this limitation, we propose a novel Histomorphology-Guided Prototypical Multi-Instance Learning (HGPMIL) framework that explicitly learns histomorphology-guided prototypical representations by incorporating tumor cellularity, cellular morphology, and tissue architecture. Specifically, our approach consists of three key components: (1) estimating the importance of tumor-related histomorphology information at patch-level based on medical prior knowledge; (2) generating representative prototypes through histomorphology-prototypical clustering; and (3) enabling WSI classification through histomorphology-guided prototypical aggregation. HGPMIL adjusts the decision boundary by incorporating histomorphological importance to reduce instance label uncertainty, thereby reversely optimizing the bag-level boundary. Experimental results demonstrate its effectiveness, achieving high diagnostic accuracy for molecular subtyping, cancer subtyping and survival analysis. The code will be made available at https://github.com/Badgewho/HMDMIL.

Figures (8)

• Figure 1: (a) Decision boundaries of MIL and HGPMIL. Left: MIL determines the decision boundary based on all patches equally. Right: HGPMIL adjusts the decision boundary by incorporating histomorphological importance to reduce instance label uncertainty, thereby reversely optimizing the bag-level boundary. (b) Examples illustrating histomorphology features characterized by tumor cellularity and the Nottingham Grading System.
• Figure 2: Overview of HGPMIL. Given a set of patches cropped from a slide, we sequentially utilize Patch Encoder, Cellularity Prediction Network, Architecture Grading Network, Histomorphology-Guided Prototypical Clustering, Histomorphology-Guided Prototypical Aggregation, and Multi-Instance Learning for WSI analysis of different downstream tasks. The shade of color within the rectangles reflects the importance of sorted tumor-related histomorphological information: the darker the color, the higher the importance, while gray ones represent middle-ranked uncertain instances.
• Figure 3: Survival prediction results on the TCGA-BRCA dataset.
• Figure 4: Ablation experiment results on AUC change.
• Figure 5: Ablation experiment results on ACC change.