Computer-Aided Classification of Breast Lesions Based on US RF Time Series Using a Novel Machine Learning Approach.

Area of Science:

• Medical Imaging
• Machine Learning
• Biomedical Engineering

Background:

• Ultrasound (US) radio-frequency (RF) time series show promise for breast cancer screening.
• This method offers a non-invasive approach without requiring additional equipment.

Purpose of the Study:

• To develop a machine learning (ML) approach for automatic classification of breast lesions.
• Classify lesions into benign, probably benign, suspicious, and malignant categories using US RF time series features.

Main Methods:

• Analyzed 220 US RF time series datasets from 118 patients.
• Extracted 291 features from regions of interest (ROIs) in time, frequency, and time-frequency domains.
• Employed a reference classification method (RCM) and Lee filter for noise reduction.

Main Results:

• Achieved high classification accuracies: 98.59% (2-class), 98.13% (3-class), and 96.10% (4-class).
• Utilized Support Vector Machine (SVM) and K-nearest neighbor (KNN) classifiers with 5-fold cross-validation.
• Demonstrated the effectiveness of the Lee filter in reducing speckle noise.

Conclusions:

• The proposed CCRFML approach effectively differentiates breast lesions using in vivo RF time series and ML.
• The high classification accuracy supports its role in assisting medical professionals for noninvasive breast lesion differentiation.