SEW: Self-calibration Enhanced Whole Slide Pathology Image Analysis

TL;DR

SEW addresses the challenge of efficiently and accurately analyzing gigapixel whole-slide pathology images by integrating global structural features from thumbnail-level superpixel graphs with targeted local features from focused lesion regions. It introduces a three-branch architecture (global, focus predictor, local) coupled with a global–local consistency constraint and a pathological prototype vocabulary to enforce cross-sample feature consistency and enable tumor-marker mining. The method achieves state-of-the-art accuracy and speed across multiple cancer datasets, and its focused features reveal novel prognostic markers validated by pathologists, with the reconstructed tissue distributions offering insight into marker localization. Overall, SEW provides a scalable, explainable framework that couples fast coarse classification with precise local analysis and biomarker discovery.

Abstract

Pathology images are considered the ``gold standard" for cancer diagnosis and treatment, with gigapixel images providing extensive tissue and cellular information. Existing methods fail to simultaneously extract global structural and local detail features for comprehensive pathology image analysis efficiently. To address these limitations, we propose a self-calibration enhanced framework for whole slide pathology image analysis, comprising three components: a global branch, a focus predictor, and a detailed branch. The global branch initially classifies using the pathological thumbnail, while the focus predictor identifies relevant regions for classification based on the last layer features of the global branch. The detailed extraction branch then assesses whether the magnified regions correspond to the lesion area. Finally, a feature consistency constraint between the global and detail branches ensures that the global branch focuses on the appropriate region and extracts sufficient discriminative features for final identification. These focused discriminative features prove invaluable for uncovering novel prognostic tumor markers from the perspective of feature cluster uniqueness and tissue spatial distribution. Extensive experiment results demonstrate that the proposed framework can rapidly deliver accurate and explainable results for pathological grading and prognosis tasks.

Figures (3)

• Figure 1: The SEW framework comprises a global branch, a focus predictor, and a detailed extraction branch. The global branch initially classifies the pathological thumbnail using loss function $\mathcal{L}^{global}_{\text{CLS}}$, while the focus predictor identifies relevant regions for classification based on the global branch's last layer features, guided by $\mathcal{L}^{focus}$. The detailed extraction branch then evaluates whether the magnified regions correspond to the lesion area using $\mathcal{L}^{local}_{\text{CLS}}$. Additionally, the feature similarity constraint $\mathcal{L}^{cst}$ between the global token and its corresponding local class token enhances the global branch's ability to extract discriminative features. With the aggregated graph features, pathological prototypes are clustered to reinforce feature consistency across diverse WSI samples, a crucial step for tumor marker discovery.
• Figure 2: Visualization of mined tumor markers in colorectal cancer samples: a) The SEW model is employed to extract pathological tissue-scale features from focused areas of colorectal cancer samples and perform clustering analysis, with particular emphasis on two feature clusters (with only red points) linked to poor prognosis. b) and c) showcase two novel tumor markers (verified by the pathologist) identified in the WSIs, along with their corresponding locations. d) presents the reconstructed WSI with pathological prototypes, where the spatial distribution of cancerous tissue (denoted in red) reveals the third tumor marker: the degree of tumor infiltration.
• Figure 3: The accuracy curve for various magnification rates (8x, 16x, and 32x) with different numbers of focus areas.