Predicting Neoadjuvant Therapy Response in Breast Cancer Patients via Radiomics Analysis of Dynamic Contrast-Enhanced MRI Imaging Features

This study introduces an accurate approach to forecast neoadjuvant therapy response in breast cancer patients leveraging dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) features. Pre-operative DCE-MRI data were retrieved from the Cancer Imaging Archive (TCIA), comprising axial breast MRI images acquired by 1.5T or 3T scanners, with sequences shared in DICOM format. Radiologist annotations facilitated accurate lesion location identification on DCE-MRI images, crucial for subsequent feature extraction. A comprehensive set of 529 computer-extracted DCE-MRI imaging descriptors, encompassing diverse tumor and tissue characteristics such as size, shape, texture, and enhancement, were derived using in-house software. To address class imbalance, oversampling techniques were implemented. Additionally, feature selection methods were employed to enhance model interpretability and efficiency. From the initial feature pool, the 40 most significant features were identified utilizing minimum redundancy maximum relevance (mRMR) analysis. Three distinct predictive models, namely Logistic Regression, AdaBoost, and XGBoost, were employed to forecast neoadjuvant therapy response. Evaluation of model performance revealed varying levels of accuracy, with XGBoost demonstrating superior performance metrics including accuracy (0.912), precision (0.917), recall (0.912), and area under the curve (AUC) value of 0.913, based on data from 416 patients, with a training-to-evaluation data split of 0.7:0.3. In conclusion, this study underscores the potential of DCE-MRI imaging features combined with comprehensive patient data for accurate neoadjuvant therapy response prediction in breast cancer patients. The findings highlight the promising role of imaging features in advancing personalized treatment strategies and ultimately improving patient outcomes.