Résumé
To compare different radiomics approaches to predict patients outcome in locally advanced rectal cancer (LARC) using MRI radiomics engineered features or deep learning models at baseline and after neoadjuvant chemoradiotherapy (CRT).Methods: 98 patients from a phase II, prospective, multicenter, randomized study (GRECCAR4-NCT01333709) were included in this study. T2-weighted sequences in conjunction with diffusion weighted images (DWI) from baseline and post CRT images were used for the analysis of 3 radiomics models at baseline and after CRT using conventional machine learning (CML) techniques. For both time point, features were extracted from 2D manual segmentation (MS), 3D MS and from bounding boxes (BBs). Recursive feature elimination and a random forest classifier were used to select features and build each model. Additionally, a deep learning model (DL), using a 16-layer convolutional neural network (CNN) was evaluated as well.Results: all 6 models of CML were able to predict patient outcome with AUCs ranging from 0.69 to 0.92. Comparison of performances among the 3 baseline models and after CRT did not show significant differences. In contrast, deep learning model demonstrated hazardous performance with an average accuracy of 0.51+/-0.3. BB Model had the highest stability across the 5 iterations of the cross-validation, with an AUC of 0.8+/-0.12 at baseline and 0,83+/-0.04 after CRT to predict patient outcome.Conclusion: among the 3 radiomics approaches to predict patient outcome using CML techniques, BBs were the easiest to obtain and be translated to a busy routine practice.