BI-RADS-NET: AN EXPLAINABLE MULTITASK LEARNING APPROACH FOR CANCER DIAGNOSIS IN BREAST ULTRASOUND IMAGES.
Boyu Zhang1, Aleksandar Vakanski2, Min Xian3
1Institute for Modeling Collaboration and Innovation, University of Idaho, Moscow, USA.
IEEE International Workshop on Machine Learning for Signal Processing : [Proceedings]. IEEE International Workshop on Machine Learning for Signal Processing
|May 5, 2022
Summary
This study presents BI-RADS-Net, an explainable AI for breast cancer detection in ultrasounds. It provides clinical explanations for diagnoses, improving trust and accuracy in medical imaging.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Healthcare
- Oncology
Background:
- Explainability is crucial for clinician trust in machine learning (ML) for healthcare decisions.
- Current ML models for breast cancer detection often lack transparent decision-making processes.
Purpose of the Study:
- Introduce BI-RADS-Net, a novel explainable deep learning model for breast cancer detection in ultrasound images.
- To provide clinically relevant explanations for model predictions using the BI-RADS lexicon.
Main Methods:
- Developed a deep learning approach (BI-RADS-Net) integrating classification and explanation tasks.
- Learned feature representations aligned with clinical diagnosis, focusing on BI-RADS morphological descriptors (shape, orientation, margin, echo, posterior features).
- Predicted malignancy likelihood corresponding to BI-RADS assessment categories.
Main Results:
- Experimental validation on 1,192 images demonstrated improved model accuracy.
- The model provided explanations in terms of established BI-RADS features used in clinical practice.
- Achieved accurate prediction of malignancy likelihood aligned with clinical assessment categories.
Conclusions:
- BI-RADS-Net enhances trustworthiness in AI for breast cancer detection through explainable predictions.
- The approach successfully translates complex model decisions into clinically interpretable BI-RADS features.
- This work bridges the gap between AI capabilities and clinical utility in breast ultrasound diagnosis.
