CPGNet: Multimodal Graph Learning with Hierarchical Category Guidance for Multi-Label Whole Slide Image...

Area of Science:

• Digital Pathology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Current automated cancer type identification in Whole Slide Images (WSIs) uses single-label classification, which is insufficient for complex clinical scenarios.
• Real-world digital pathology data often features multi-label characteristics and class imbalance, challenging existing automated methods.
• Accurate WSI analysis is crucial for cancer diagnosis, treatment planning, and prognosis.

Purpose of the Study:

• To develop an advanced multi-label Whole Slide Image (WSI) classifier, CPGNet, to better handle complex cancer subtypes and class imbalance in digital pathology.
• To mimic the diagnostic process of pathologists by integrating local and global feature extraction and leveraging semantic category relationships.
• To improve the accuracy and robustness of automated cancer subtyping in clinical settings.

Main Methods:

• CPGNet utilizes MaskSLIC for superpixel segmentation of WSIs, representing them as graphs with nodes and edges.
• A Graph Neural Network (GNN) with a multi-head self-attention mechanism (GLGFI module) captures local and global spatial dependencies.
• A Visual-Category Interaction (VCI) module leverages semantic relationships, and a reweighting strategy addresses class imbalance.

Main Results:

• CPGNet demonstrated superior performance, universality, and robustness across private (YNLUAD) and public (BCNB, AGGC22) datasets.
• The proposed multi-label classification approach effectively handles the complexities of real-world cancer subtype identification.
• The model successfully captures intricate spatial distributions and mimics expert pathologist diagnostic strategies.

Conclusions:

• CPGNet offers a significant advancement in automated multi-label WSI classification for digital pathology.
• The model's ability to handle class imbalance and complex spatial relationships enhances its clinical applicability.
• This approach provides a more realistic and effective tool for cancer diagnosis and prognosis support.