Ultrasound Radio Frequency Time Series for Tissue Typing: Experiments on In-Vivo Breast Samples Using...

Area of Science:

• Medical Imaging
• Machine Learning
• Oncology

Background:

• Ultrasound (US) radio-frequency (RF) time series show promise for breast cancer (BC) screening.
• US-based methods offer advantages as they do not require supplementary equipment.

Purpose of the Study:

• To develop a machine learning (ML) method for automated classification of breast lesions using US RF time series data.
• To categorize lesions into benign, probably benign, suspicious, or malignant types.

Main Methods:

• Analysis of 220 data points from 118 patients using the RFTSBU dataset.
• Extraction of 283 features from manually selected regions of interest (ROIs) using textural analysis (Gabor filter, GLCM, GLRLM, GLSZM, GLDM).
• Feature selection via particle swarm optimization (PSO) to 131 features, followed by classification using a multi-origin method classification (MOMC) approach.

Main Results:

• Achieved high accuracy rates with 5-fold cross-validation.
• 2-class classification accuracy: 98.57% ± 1.09% (MOMC-SVM and MOMC-ensemble).
• 3-class classification accuracy: 91.53% ± 0.89%.
• 4-class classification accuracy: 83.71% ± 1.30%.

Conclusions:

• An innovative ML-based approach effectively differentiates breast lesion types using in vivo US RF time series data.
• The method demonstrates high classification accuracy, advancing computer-aided diagnosis (CAD) for BC screening.