A radiomics method to classify microcalcification clusters in digital breast tomosynthesis.

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Digital breast tomosynthesis (DBT) is increasingly used for breast cancer screening.
• Microcalcification clusters in DBT can span multiple slices, complicating radiologist assessment.
• Accurate classification of microcalcification clusters is crucial for diagnosis.

Purpose of the Study:

• To investigate a radiomics method for classifying microcalcification clusters in DBT.
• To assess the diagnostic performance of radiomics features extracted from semiautomatically segmented clusters.
• To differentiate between benign and malignant microcalcification clusters using machine learning.

Main Methods:

• Retrospective study of 550 DBT volumes from 275 patients (79 benign, 196 malignant).
• Semiautomatic segmentation of microcalcification clusters followed by radiomics feature extraction.
• Classification using a random forest (RF) classifier with selected features; evaluation using ROC curves and AUC.

Main Results:

• Twenty-six key radiomics features were selected from 170.
• Highest AUC achieved was 0.868 (case-based) and 0.834 (view-based) using 3D-domain features.
• 2D-domain radiomics features demonstrated statistically similar performance to 3D features.

Conclusions:

• Radiomics models show encouraging performance in classifying benign and malignant microcalcification clusters in DBT.
• Semiautomatic segmentation is a viable approach for feature extraction in DBT.
• This radiomics approach can potentially improve diagnostic accuracy for microcalcification clusters.