Knowledge-Powered Deep Breast Tumor Classification With Multiple Medical Reports.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Computational Pathology

Background:

• Accurate breast tumor classification using diverse medical reports (B-ultrasound, Mammography, MRI) is vital for intelligent cancer diagnosis.
• Medical reports possess complex hierarchical syntactic structures and rich semantic information, posing challenges for traditional classification methods.

Purpose of the Study:

• To propose a Knowledge-powered Deep Breast Tumor Classification (KDBTC) model that leverages semantic information as prior knowledge for improved accuracy.
• To enhance the extraction of syntax-aware and semantic representations from medical reports for breast cancer classification.

Main Methods:

• Developed a Knowledge-powered Deep Breast Tumor Classification (KDBTC) model integrating semantic information into deep neural networks.
• Employed Hierarchical Attention Bidirectional Recurrent Neural Networks (HA-BiRNNs) for hierarchical encoding of syntax-aware representations.
• Utilized Tree Structured Recurrent Neural Networks with gated recursive units (Tree-GRUs) to encode semantic representations from a clinical domain semantic tree.

Main Results:

• The KDBTC model effectively combines syntax and semantic representations for breast tumor classification.
• Evaluated on real-world breast cancer medical reports, the proposed method demonstrated superior performance compared to existing approaches.

Conclusions:

• The KDBTC model offers a promising approach for accurate breast tumor classification by effectively utilizing the rich information within medical reports.
• Integrating semantic prior knowledge into deep learning models significantly enhances the performance of intelligent cancer diagnosis systems.