MRI-based texture analysis for breast cancer subtype classification in a multi-ethnic population

Area of Science:

• Radiology and Medical Imaging
• Oncology
• Artificial Intelligence in Medicine

Background:

• Breast cancer subtyping is crucial for treatment and prognosis, but traditional methods are invasive.
• Magnetic Resonance Imaging (MRI) radiomics offers a non-invasive approach to extract quantitative imaging features.
• Limited research exists on MRI radiomics for breast cancer subtyping in diverse populations.

Purpose of the Study:

• To identify predictive radiomic features from multiple MRI sequences for breast cancer subtype classification.
• To compare the performance of four MRI sequences in breast cancer subtyping.
• To determine the optimal machine learning (ML) model for non-invasive breast cancer subtyping.

Main Methods:

• Retrospective analysis of 162 breast cancer MRI cases with semi-automated segmentation.
• Extraction of 256 radiomic features from multiple MRI sequences.
• Development of a multimodal ML framework using random forest and recursive feature elimination for feature selection based on AUROC.

Main Results:

• Key predictive features included patient age, tumor size, margin characteristics, and intra-tumoral intensity patterns.
• Inversion recovery and T1 post-contrast MRI sequences demonstrated superior performance for subtyping.
• Texture-based ML models achieved AUROC values of 0.735 (luminal), 0.630 (HER2-enriched), and 0.747 (triple-negative), emulating visual assessment.

Conclusions:

• MRI-based texture features and advanced ML show significant potential for enhancing breast cancer diagnosis.
• Radiomics provides a non-invasive tool for personalized treatment planning in breast cancer.
• This approach can complement existing clinical workflows for breast cancer management.