Deep Learning-Based CAD System for Enhanced Breast Lesion Classification and Grading Using RFTSDP Approach.

Area of Science:

• Medical imaging
• Biophysics
• Artificial intelligence in medicine

Background:

• Accurate breast lesion classification is vital for treatment but limited by current diagnostic precision, often necessitating biopsies.
• Radio frequency time series dynamic processing (RFTSDP) was introduced to analyze tissue dynamics and scatterer displacement impacts on RF echoes during stimulation for enhanced diagnostics.

Purpose of the Study:

• To develop and evaluate a deep learning (DL)-based system for automated breast lesion classification and grading using RFTSDP.
• To compare the performance of a convolutional neural network (CNN)-based RFTSDP method against traditional machine learning techniques.

Main Methods:

• Developed a vibration-generating device for ultrafast ultrasound data acquisition from ex vivo breast tissues.
• Employed a CNN for automated feature extraction and classification of lesions into 2, 3, and 5 categories.
• Compared CNN-based RFTSDP performance with spectral and nonlinear feature extraction followed by support vector machine (SVM).

Main Results:

• The DL-based RFTSDP method achieved 99.53% accuracy in classifying and grading breast lesions under 65 Hz vibration.
• CNN consistently outperformed SVM, achieving 98.01% accuracy in 5-class classification compared to SVM's 95.64%.
• The CNN-based RFTSDP method demonstrated a 28.67% improvement in classification accuracy compared to non-stimulation conditions.

Conclusions:

• A DL-based computer-aided diagnosis (CAD) system was successfully developed for breast lesion classification and grading.
• The proposed RFTSDP system enhances classification accuracy, stability, and robustness over traditional methods.