Interpretable Radiomic Signature for Breast Microcalcification Detection and Classification.

Area of Science:

• Radiology
• Medical Imaging
• Computational Pathology

Background:

• Breast microcalcifications are common in mammograms, with some indicating invasive tumors.
• Accurate diagnosis is challenging due to variations in size, shape, and subtle differences.
• Radiomics offers a quantitative approach to extract imaging features for improved diagnostic accuracy.

Purpose of the Study:

• To develop and validate a radiomic signature for differentiating healthy tissue, benign, and malignant breast microcalcifications.
• To assess the performance of machine learning models in classifying microcalcification types.
• To enable clinical validation through interpretable radiomic features.

Main Methods:

• Extraction of radiomic features from a dataset of healthy, benign, and malignant microcalcification regions of interest (ROIs).
• Selection of distinct radiomic signatures for detection (healthy vs. microcalcifications) and classification (benign vs. malignant).
• Training and evaluation of machine learning models (SVM, Random Forest, XGBoost) for multi-class classification.

Main Results:

• A shared radiomic signature was identified for both detection and classification tasks.
• XGBoost model achieved high performance: AUC-ROC of 0.830 (healthy), 0.856 (benign), and 0.876 (malignant).
• Key features like GLCM Contrast, FO Minimum, and FO Entropy were identified as important and clinically relevant.

Conclusions:

• The proposed radiomic signature effectively differentiates between healthy tissue, benign, and malignant breast microcalcifications.
• Machine learning models, particularly XGBoost, demonstrate significant potential for accurate microcalcification classification.
• Radiomic feature interpretability facilitates clinical validation and understanding of diagnostic markers.