“MRI-based radiomics machine learning model for tumour response prediction to neoadjuvant chemoradiotherapy (nCRT) in locally advanced rectal cancer (LARC): a retrospective study

Abstract

Rectal cancer (RC) is one of the most commonly diagnosed malignant tumours worldwide. Its most aggressive form, Locally Advanced Rectal Cancer (LARC), is treated with neoadjuvant chemoradiotherapy (nCRT) to downstage the tumour and improve results of the total mesorectal excision (TME). Approximately 20% to 25% achieve complete response and may be eligible for a more conservative watch and wait strategy. Early prediction of response to nCRT using information from the staging MRI could help adjust neoadjuvant treatment andimprove response, potentially avoiding surgeries. The aim of this project is to develop a machine learning (ML) model capable of predicting response to nCRT using clinical data and radiomics characteristics extracted from the pre-treatment MRI. The radiomics are extracted from manually delineated tumour masks (Core Tumour Radiomics) and from border masks computed from the manual segmentations (Border Radiomics). Nine ML models have been optimized and tested across the seven feature set combinations. The models with best performance were Random Forest for the core tumour radiomics dataset (Accuracy = 0.77, AUC = 0.70, Sensitivity = 0.67, Specificity = 0.86) and Multilayer Perceptron for the dataset with all features (Accuracy = 0.77, AUC = 0.70, Sensitivity = 0.83, Specificity = 0.71). However, the predictive capability of tumour borders could not be confirmed as these models yielded worse performance. Furthermore, via a feature importance analysis, it has been concluded that both shape and texture related radiomic features are predictors of treatment response although no specific marker has been identified.