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

Baizhi Wang, Rui Yan, Wenxin Ma, Xu Zhang, Yuhao Wang, Xiaolong Li, Yunjie Gu, Zihang Jiang, S. Kevin Zhou · 2025

Read Histomorphology-Guided Prototypical Multi-Instance Learning for Breast Cancer WSI Classification by Baizhi Wang, Rui Yan, Wenxin Ma, Xu Zhang, Yuhao Wang, Xiaolong Li, Yunjie Gu, Zihang Jiang, S. Kevin Zhou (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-17983.txt.

Introduction

Baizhi Wang and co-authors build a method that sorts breast cancer slides by what the tissue looks like under a scope, in a paper from 2025. Whole slide scans hold too many patches to label one by one, so methods learn from the slide tag alone. That leaves the key patches lost in a sea of bland ones, and calls grow shaky.

The new frame scores each patch by how much tumor form it holds, by cell bulk, cell shape, and tissue lay. Like patches then clump into type marks that stand for the slide. Slide calls are made by pool of those marks, with the form scores set to tune the call edge. Tests show strong scores on subtype, broad cancer sort, and life span tasks.

Catalog record

CREATORBaizhi Wang, Rui Yan, Wenxin Ma, Xu Zhang, Yuhao Wang, Xiaolong Li, Yunjie Gu, Zihang Jiang, S. Kevin Zhou
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE31 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-17983

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Histomorphology-Guided Prototypical Multi-Instance Learning for Breast Cancer WSI Classification Baizhi Wang, Rui Yan, Wenxin Ma, Xu Zhang, Yuhao Wang, Xiaolong Li, Yunjie Gu, Zihang Jiang, S. Kevin Zhou 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 available. 1 Introduction 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. Breast cancer is the most prevalent and lethal malignancy among women [ 17 , 1 ] . Histomorphology features are critical for breast cancer [ 27 ] diagnosis, grading, and prognosis. Histomorphology is typically evaluated based on cellular morphology and tissue architecture, as formalized in the Nottingham Grading System (NGS) [ 8 ] , which scores nuclear atypia, mitotic count, and tubule formation, with these factors collectively determining the tumor grade (I–III); higher grades indicate poorer differentiation and more aggressive behavior, as illustrated in Fig. 1 (b). Meanwhile, tumor cellularity, defined as the percentage of malignant cells within the tumor bed, is another key indicator of tumor burden and is used to assess treatment response, especially after neoadjuvant therapy [ 20 ] . A substantial reduction in cellularity correlates with favorable outcomes [ 10 ] . Computational pathology has significantly advanced whole slide image (WSI) analysis through deep learning [ 3 , 5 , 16 ] , with most approaches relying on the weakly supervised learning paradigm known as multi-instance learning (MIL) [ 12 , 6 ] . However, existing computational models often fail to explicitly incorporate tumor tissue architecture and cellularity into decision making process, limiting their ability to capture key pathological features. Particularly when the number of instances (i.e., patches) within a bag (i.e., WSI) 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 overcome these limitations, prototype learning has emerged as an effective approach for improving model performance and interpretability in computational pathology [ 11 , 22 , 37 , 7 ] . Prototypes, as learned representative feature centroids, allow the model to focus on biologically meaningful tumor regions, such as those with high cellularity or architecture [ 19 , 36 , 39 ] . By learning prototypes that represent the idealized features of tumor regions, the model can better differentiate between diagnostically relevant and irrelevant areas [ 26 , 30 , 39 ] . This enhances the decision boundary by reducing overlap between classes and ensuring that the model pays more attention to regions that are more informative and unambiguous for classification [ 25 ] . In this paper, we propose H istomorphology- G uided P rototypical M ulti- I nstance L earning (HGPMIL), a simple yet effective framework that learns histomorphological prototypes to guide instance aggregation, enabling more robust WSI classification. As shown in Fig. 1 (a), HGPMIL achieves a more discriminative decision boundary than traditional MIL by leveraging histomorphological knowledge, leading to improved separation between positive and negative bags. Extensive experimental results demonstrate a strong correlation between utilizing tumor histomorphology and improving the accuracy of MIL-based WSI classification. The contributions of this paper can be summarized as follows: • We explicitly integrate histomorphological knowledge into WSI classification by learning prototypical representations guided by medical priors. This enables the model to capture critical pathological patterns while filtering out ambiguous instances, which are essential for MIL-based classification. • We propose HGPMIL, a novel and adaptable histomorphology-guided prototypical multi-instance learning framework, designed to enhance prototypical aggregation through three key components: histomorphology importance estimation network, histomorphology-prototypical clustering, and histomorphology-guided prototypical aggregation. • Extensive experiments demonstrate that HGPMIL is highly flexible and broadly applicable across various MIL models and downstream tasks, consistently leading to significant performance improvements. 2 Related Work 2.1 Attention-based and Knowledge-guided MIL MIL has become a standard weakly supervised learning paradigm in computational pathology, especially when detailed cell- or region-level annotations are unavailable [ 24 ] . In MIL, a WSI is treated as a bag of instances (typically patches), and the goal is to infer a slide-level label by aggregating instance-level information [ 13 , 28 ] . Attention-based methods, such as ABMIL [ 12 ] , CLAM [ 14 ] , and TransMIL [ 23 ] , have shown strong performance by learning attention weights that emphasize informative patches. These models improve both accuracy and interpretability by enabling instance-level localization of diagnostically relevant regions. More recently, HDMIL [ 34 ] extends this line of work with a hierarchical attention mechanism to capture both local and global context. However, while effective, most of these models rely solely on data-driven attention and do not explicitly incorporate medical priors. 2.2 Efficient and Long-sequence MIL Processing WSIs with thousands of patches introduces computational and modeling challenges [ 4 ] . Transformer-based MIL methods like TransMIL and DTFD-MIL [ 40 ] address this by modeling long-range dependencies. Meanwhile, state-space sequence models, such as S4MIL [ 9 ] and MambaMIL [ 38 ] , provide more scalable alternatives with linear-time complexity and strong capacity for long-context modeling. These methods achieve competitive performance while reducing computational costs. However, they typically treat all instances equally and overlook histopathological relevance at the patch level. Despite these advances, existing MIL frameworks [ 31 ] often process large numbers of redundant patches and lack mechanisms to incorporate domain knowledge into patch evaluation. This limits efficiency and clinical interpretability. To address these challenges, we propose HGPMIL, a plug-and-play framework that integrates tumor tissue architecture and cellularity into MIL process. By guiding patch importance estimation and aggregation with histomorphological priors, HGPMIL enhances both diagnostic accuracy and biological relevance in WSI-level classification. 3 Methods 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

Read the complete text