Input of pre-operative radiomics to select patients with locally advanced rectal cancers

With the advent of total neoadjuvant therapy (TNT) for locally advanced rectal adenocarcinomas (LARC), high rates of pathologic complete response (pCR) are achieved, paving the way for organ preservation strategies. Identifying complete responders is a challenge. In this ancillary study of the GRECCAR 12 prospective multicentric randomized trial, we aimed to assess the input of baseline T2-weighted MRI radiomics associated with standard clinical and radiological data to predict pCR after neoadjuvant therapy. We retrospectively considered every patient included in the GRECCAR 12 trial. After pre-sampling of T2-weighted pretherapeutic MRI sequences, and a posteriori ComBat harmonization, 4 sets of 100 radiomic features were extracted from manually delineated tumors. Following random division between training and testing sets on a 3:1 ratio, univariate selection with stepwise regression and redundancy reduction, two models (with and without ComBat) using radiomic and non-radiomic variables were built. The prediction of pCR was assessed using logistic regression models, with ROC-AUC and accuracy of prediction as performance gauges. One hundred and sixty-one patients were included in the final radiomic analysis, 82 in the chemo-radiotherapy arm and 79 in the TNT arm. The rate of pCR was 46%. The best performing models including 2 non-radiomic features, ymrTRG and radiomic features (3 for No-ComBat and 4 for ComBat) obtained AUCs of 0.65 for No-ComBat and 0.74 for ComBat on the testing set, beating clinical variables (AUC of 0.56) or ymrTRG (AUC of 0.58) alone. Radiomics has the potential to better select patients for organ preservation, increasing pCR prediction accuracy in LARC patients after neoadjuvant therapy in association with clinical data and ymrTRG. Prospective trials and intergroup workforce are warranted, as well as evaluating the addition of other biomarkers (MSI status, circulating DNA...). (author)