Detection of Suspicious Lesions in Breast MRI: Radiomics Patch-based Granular Classification Approach.

Area of Science:

• Radiology
• Medical Imaging Analysis
• Machine Learning in Healthcare

Background:

• Breast magnetic resonance imaging (MRI) is valuable but limited by cost and reader availability.
• Existing research focuses on breast region segmentation in MRI.
• Automated detection of suspicious lesions is needed to enhance MRI's utility.

Purpose of the Study:

• To evaluate automated detection of suspicious breast lesions in MRI using radiomics analysis.
• To develop a processing pipeline for characterizing breast tissue patches.
• To classify tissue patches for differentiating lesions from normal tissue.

Main Methods:

• Breast MRI data processed via a grid division pipeline.
• Radiomics features extracted from image patches.
• Binary classification using Random Forest (RF) and XGBoost algorithms.
• Patch-wise classification to identify suspicious lesions.

Main Results:

• Achieved F1 scores of ≥ 0.92 for patch-wise classification.
• Demonstrated high balanced precision (0.94) and recall (0.95).
• Reassembled patch decisions enable lesion localization and patient-level decision support.

Conclusions:

• Automated radiomics analysis shows high accuracy for suspicious lesion detection in breast MRI.
• This method can support clinical decision-making and patient prioritization.
• The approach is particularly relevant for settings with limited radiologist availability.