Content-based image retrieval applied to BI-RADS tissue classification in screening mammography

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Machine Learning

Background:

• Accurate breast density classification is crucial for mammography interpretation.
• Existing methods may lack efficiency in parameter adaptation for lesion detection.
• A robust content-based image retrieval (CBIR) system can enhance diagnostic workflows.

Purpose of the Study:

• To develop a CBIR system for breast tissue density classification.
• To integrate this system into a processing chain for adaptive lesion segmentation and classification.
• To improve the efficiency and accuracy of mammographic analysis.

Main Methods:

• Breast density characterization using singular value decomposition (SVD) for texture analysis and histograms.
• Pattern similarity computed via support vector machine (SVM) to categorize four BI-RADS tissue densities.
• Investigation of SVD parameter variations and SVM kernel types (linear, radial, polynomial) on a large mammogram database.

Main Results:

• A comprehensive reference database exceeding 10,000 mammograms from multiple sources (DDSM, MIAS, LLNL, RWTH) with verified ground truth.
• Achieved an average precision of 82.14% using 25 singular values (SVD) and a polynomial kernel with one-against-one SVM classification.
• Demonstrated the effectiveness of the chosen SVD and SVM parameters for accurate breast density classification.

Conclusions:

• Singular value decomposition (SVD) combined with support vector machines (SVM) effectively characterizes breast density from mammograms.
• The developed CBIR system offers a valuable tool to assist radiologists in their diagnostic process.
• This approach facilitates adaptive parameter tuning for subsequent lesion segmentation and classification tasks.