An Adaptive Fusion Network for Breast Tumor Grading Based on Graph Structure Learning.

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Breast cancer poses a significant threat to women's health, necessitating accurate malignancy grading for improved survival rates.
• Current clinical diagnosis relies on multi-modal data, but effectively integrating this information for grading remains a challenge.
• Graph convolutional networks (GCNs) show promise for multi-modal data processing in tumor grading.

Purpose of the Study:

• To develop and evaluate a multi-graph adaptive fusion network for accurate breast tumor grading using multi-modal data.
• To address the limitations of single-graph models in utilizing comprehensive multi-modal information for diagnosis.

Main Methods:

• A multi-modal information extraction network, comprising a graph structure learning layer and a graph auto-encoder, was designed to extract modal-specific information and graph structures.
• Attention scores were used to weight-fuse modal features and graphs, creating aggregated representations.
• A neighborhood adaptive aggregation module was developed to optimize graph convolution by calculating neighborhood aggregation coefficients.

Main Results:

• The proposed multi-graph adaptive fusion network effectively processed multi-modal medical data.
• The graph learning-based diagnosis model demonstrated improved accuracy in multi-grading of breast tumors on both public and private datasets.

Conclusions:

• The developed multi-graph adaptive fusion network offers a robust approach for breast tumor grading.
• This graph learning-based method enhances the accuracy of multi-modal data utilization in clinical diagnosis, potentially improving patient outcomes.