MRI Radiomics Features for Prediction of Treatment Response in Colorectal Patients

Abstract

In this study, we assess the power of MRI radiomic features for prediction of locally advanced rectal cancer (LARC) patients' response to neoadjuvant chemoradiation. T2-Weighted MR images acquired 2 weeks before and 4 weeks after treatment of 50 patients were used. The tumor volume was delineated by an experienced radiologist on T2-weighted MR images followed by the extraction of radiomics features, including morphology, first-order, histogram, and texture from volumes of interest (VOI). First, univariate analysis was applied on features to identify predictive power of features. To build a predictive model, we used Random Forest (RF) algorithm along with Max-Relevance-Min-Redundancy (MRMR) feature selection algorithm for reducing complexity and improving generalization. Finally, the model was evaluated through the area under the receiver operator characteristic (ROC) curve (AVC), sensitivity, specificity and accuracy metrics. In univariate analysis, delta radiomics of LAE and LALGLE features from GLSZM had the highest predictive performance (AUC=0.67). In multivariate analysis, the highest predictive performance for response prediction in LARC patients was achieved using delta-radiomic features with AUC of 0.92 and 0.88 in training and validation datasets, respectively. The achieved results were promising to move towards personalized treatment for LARC patients.