CPGNet: Multimodal Graph Learning With Hierarchical Category Guidance for Multi-Label Whole Slide Image Classification

Abstract

The analysis of WSI categories in digital pathology serves as a critical basis for clinicians’ decision-making regarding the diagnosis, treatment, and prognosis of cancer patients. However, current automated methods for cancer type identification are predominantly formulated as single-label classification problems. These methods typically rely on datasets with relatively balanced and abundant samples, where each sample belongs to a single category. This approach does not fully align with real-world clinical scenarios, where cancer subtypes often exhibit multi-label characteristics and class imbalance, posing significant challenges. To address this issue, this paper proposes CPGNet, a category-prompted graph network designed as a multi-label whole slide image classifier that is better suited for clinical scenarios. CPGNet employs the MaskSLIC algorithm for superpixel segmentation of whole slide images, effectively capturing the nonlinear spatial distribution of cellular and tissue structures. The segmented superpixels are then encoded as graph nodes along with their corresponding features, while edges and edge features are constructed to abstractly model WSIs as graphs. Furthermore, the method introduces a Graph Local-Global Feature Interaction module, which aggregates features of neighboring nodes and edges via a Graph Neural Network to capture local information while leveraging a multi-head self-attention mechanism to model global dependencies, mimicking the diagnostic behavior of pathologists. Additionally, the Vision-Category Interaction module exploits semantic relationships between categories to guide visual feature classification, providing supplementary cues for accurate predictions. To enhance the model’s focus on hard-to-classify positive samples, we also implement a reweighting strategy. The proposed approach is evaluated on a private dataset (YNLUAD) and two public challenge datasets (BCNB and AGGC22). The experimental results demonstrate the superiority, generalizability, and robustness of CPGNet.