External Validation of a Radiomics Model for the Prediction of Complete Response to Neoadjuvant Chemoradiotherapy in Rectal Cancer

Simple Summary In locally advanced rectal cancer (LARC), a minority of patients presents a pathological complete response (pCR) after neoadjuvant chemoradiotherapy (CRT). In this sub-population, organ preservation could be proposed without compromising overall survival. Using a robust neural network based statistical approach, correction of imbalanced data and inter-center variability, a radiomics-based model was externally validated with a balanced accuracy of 85.5%. This model efficiently predicted the patients with a pCR in an external cohort and could be used to select the patients eligible for organ preservation. Objective: Our objective was to develop a radiomics model based on magnetic resonance imaging (MRI) and contrast-enhanced computed tomography (CE-CT) to predict pathological complete response (pCR) to neoadjuvant treatment in locally advanced rectal cancer (LARC). Material: All patients treated for a LARC with neoadjuvant CRT and subsequent surgery in two separate institutions between 2012 and 2019 were considered. Both pre-CRT pelvic MRI and CE-CT were mandatory for inclusion. The tumor was manually segmented on the T2-weighted and diffusion axial MRI sequences and on CE-CT. In total, 88 radiomic parameters were extracted from each sequence using the Miras (c) software, with a total of 822 features by patient. The cohort was split into training (Institution 1) and testing (Institution 2) sets. The ComBat and Synthetic Minority Over-sampling Technique (SMOTE) approaches were used to account for inter-institution heterogeneity and imbalanced data, respectively. We selected the most predictive characteristics using Spearman's rank correlation and the Area Under the ROC Curve (AUC). Five pCR prediction models (clinical, radiomics before and after ComBat, and combined before and after ComBat) were then developed on the training set with a neural network approach and a bootstrap internal validation (n = 1000 replications). A cut-off maximizing the model's performance was defined on the training set. Each model was then evaluated on the testing set using sensitivity, specificity, balanced accuracy (Bacc) with the predefined cut-off. Results: Out of the 124 included patients, 14 had pCR (11.3%). After ComBat harmonization, the radiomic and the combined models obtained a Bacc of 68.2% and 85.5%, respectively, while the clinical model and the pre-ComBat combined achieved respective Baccs of 60.0% and 75.5%. Conclusions: After correction of inter-site variability and imbalanced data, addition of radiomic features enhances the prediction of pCR after neoadjuvant CRT in LARC.

Automated summary

Developing a radiomics model based on MRI and CE-CT using a neural network approach, this study aimed to predict pCR in LARC patients undergoing neoadjuvant treatment. By identifying the most predictive features and validating on datasets from two separate institutions, it was found that adding radiomic features improves the prediction accuracy for pCR.