Computer-Aided Diagnosis for Breast Ultrasound Using Computerized BI-RADS Features and Machine Learning Methods

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast ultrasound is crucial for lesion characterization.
• Accurate differentiation of benign and malignant breast lesions remains a challenge.
• The Breast Imaging Reporting and Data System (BI-RADS) provides standardized descriptors for breast imaging findings.

Purpose of the Study:

• To identify effective computable features from BI-RADS for breast ultrasound.
• To develop and evaluate a computer-aided diagnosis (CAD) system for breast ultrasound.
• To assess the performance of different machine learning methods in classifying breast lesions.

Main Methods:

• A database of 283 pathology-proven breast lesions was utilized.
• Computable features based on ultrasound BI-RADS categories were designed.
• Machine learning algorithms including decision tree, artificial neural network, random forest, and support vector machine were employed.
• A "bottom-up" feature selection approach and 10-fold cross-validation were used.

Main Results:

• The highest area under the receiver operating characteristic (ROC) curve (AUC) achieved was 0.84 with a support vector machine (SVM), yielding 77.7% overall accuracy.
• The highest overall accuracy of 78.5% was obtained using a random forest classifier, with an AUC of 0.83.
• Lesion margin and orientation were identified as optimal features, consistently effective across different machine learning methods.

Conclusions:

• The developed CAD system effectively utilizes BI-RADS features for breast ultrasound analysis.
• Machine learning models, particularly SVM and random forest, demonstrate strong performance in differentiating benign and malignant breast lesions.
• Identified optimal features like lesion margin and orientation can enhance CAD system capabilities for clinical decision support.