Graph Neural Networks for Gleason Grading in Prostate Histopathology Images

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Digital pathology

Background:

• Prostate cancer is a significant cause of cancer mortality.
• Accurate Gleason grading is essential for determining treatment strategies.
• Current grading methods can be subjective and time-consuming.

Purpose of the Study:

• To develop and validate a Graph Neural Network (GNN) model for automated Gleason grading.
• To assess the efficacy of GNNs in classifying prostate cancer aggressiveness from histopathology images.
• To improve the detection of aggressive cancer grades, particularly Gleason Grade 5.

Main Methods:

• Utilized the Automated Gleason Grading Challenge 2022 dataset.
• Constructed patch-level graphs from Hematoxylin and Eosin-stained Whole-Slide Images.
• Employed Graph Attention Networks (GAT) and Graph Convolutional Networks (GCN) for classification.
• Integrated Focal Loss to address class imbalance, especially for Gleason Grade 5.

Main Results:

• GNNs effectively distinguished between Gleason grades, demonstrating robustness to class imbalance.
• Focal Loss significantly improved the classification accuracy of the minority Gleason Grade 5.
• The proposed GNN models achieved superior performance compared to state-of-the-art methods.
• High F1-scores were obtained without relying on scanner generalization techniques.

Conclusions:

• Graph Neural Networks offer a powerful and accurate approach for automated prostate cancer Gleason grading.
• The developed models show potential for improving diagnostic efficiency and patient stratification.
• This AI-driven method can aid pathologists in identifying aggressive prostate cancer more effectively.