A Practical Computer Aided Diagnosis System for Breast Ultrasound Classifying Lesions into the ACR BI-RADS Assessment

Abstract

Purpose: In this paper, we propose an open-source deep learning-based computer-aided diagnosis system for breast ultrasound images based on the Breast Imaging Reporting and Data System (BI-RADS). Methods: Our dataset with 8,026 region-of-interest images preprocessed with ten times data augmentation. We compared the classification performance of VGG-16, ResNet-50, and DenseNet-121 and two ensemble methods integrated the single models. Results: The ensemble model achieved the best performance, with 81.8% accuracy. Our results show that our model is performant enough to classify Category 2 and Category 4/5 lesions, and data augmentation can improve the classification performance of Category 3. Conclusion: Our main contribution is to classify breast ultrasound lesions into BI-RADS assessment classes that place more emphasis on adhering to the BI-RADS medical suggestions including recommending routine follow-up tracing (Category 2), short-term follow-up tracing (Category 3) and biopsies (Category 4/5).