Predicting Response to Neoadjuvant Chemotherapy using Multi-Modal MRI Radiomics and Machine Learning Approaches

A precise computer-aided diagnosis (CAD) system is introduced to predict two different tumor responses to neoadjuvant chemotherapy (NAC) by studying the correlation between radiological and clinical markers, and treatment response: complete response and stable disease. Predicting the NAC response assists physicians in managing the treatment plan effectively and reduces unnecessary use of NAC. First, the breast tumor is delineated from MRI modalities. Then, the proposed system extracts functional features, including the apparent diffusion coefficient (ADC), from diffusion-weighted imaging (DWI), and 2^nd-order texture features, such as gray-level co-occurrence matrix (GLCM) and gray-level run length matrix (GLRLM), from various MRI modalities. Afterward, statistical features are extracted from these matrices. Moreover, ADC features are represented statistically using percentiles with 2% intervals. Subsequently, a feature selection approach, namely the genetic algorithm (GA), is employed for each fused features individually (i.e., ADC, GLCM, and GLRLM) to extract the most relevant features. The resulting feature set undergoes a multi-stage feature generation system, introducing new features (i.e., probabilities), which are fused with its input and clinical markers and are then fed into a lightGBM classifier for final prediction. Our system was evaluated on 54 patients using leave-one-subject-out (LOSO) cross-validation, achieving a sensitivity of 89.89%, a specificity of 89.29%, and an accuracy of 89.09%, emphasizing the promise of our system in predicting breast tumor response to NAC.