Thesis title: Development of a Radiomic Model Using Machine Learning for Breast Lesion Classification and Breast Cancer Subtyping
This study investigates the potential of radiomics and machine learning in the non-invasive classification and subtyping of breast lesions, aiming to enhance diagnostic precision and personalized oncology. Using MRI data processed on the TRACE4Research™ platform, we developed four radiomic models targeting specific classification tasks: distinguishing benign from malignant lesions, differentiating luminal from non-luminal malignancies, and further subtyping luminal (Luminal A vs. Luminal B) and non-luminal (HER2+ vs. Triple-Negative) cancers. The models were trained on a balanced dataset and underwent internal and external validation to ensure robustness and generalizability.
The benign vs. malignant model demonstrated high accuracy, achieving a mean ROC-AUC of 88%, reflecting the model's reliability in distinguishing between benign and malignant lesions. The molecular subtyping models showed varied performance, with the luminal vs. non-luminal model achieving favorable sensitivity, while the Luminal A vs. Luminal B model exhibited moderate discriminatory power. The HER2+ vs. Triple-Negative model achieved high sensitivity, indicating its utility in assessing aggressive breast cancer subtypes.
The findings underscore radiomics’ potential to complement traditional diagnostic methods, particularly in cases requiring detailed molecular characterization. However, challenges such as variability in MRI protocols and sample size constraints for molecular subtypes highlight the need for standardization and larger datasets. Future research should explore multi-modal radiomics and the integration of molecular data to enhance model performance for clinical use.