Local Tuning of an Existing Externally Developed Radiomic-Based Model for Predicting Patient Outcome in Locally Advanced Rectal Cancer

Abstract

Locally advanced rectum cancer (LARC) patients achieving a pathological complete response (pCR) may have a good oncological outcome and the appropriateness of a partial resection could be considered. We aim to further enhance a validated radiomics-based model to predict pCR after chemo-radiotherapy in LARC for use in clinical practice.A generalized linear model (GLM) to predict pCR in LARC patients previously trained in Europe and validated with external intercontinental cohort (59 patients), was first examined with further 88 intercontinental patient datasets to assess its reproducibility. Moreover, to further enhance the model pCR accurate prediction, new radiomics and clinical features, and validation methods were investigated. LARC patients included in this study underwent neo adjuvant chemo-radiotherapy (CRT) followed by total mesorectal excision. Response to CRT was determined by histopathological examination and classified using tumor regression grade (TRG) according to Mandard classification as pCR (TRG = 1), or non-responder (TRG > 1). Diagnostic MR images of all patients were acquired with the same acquisition protocol. The patients were divided homogenously into training group (75%) and testing group (25%). For features extraction, images were preprocessed with Laplacian of Gaussian convolution with kernel filter; sigma from 0.1 to 1.0 (step size 0.05) were applied using a statistical software package. Mann-Whitney test was used to find features significant for pCR. To avoid the multicollinearity among radiomics features, the optimal subset was selected by the least absolute shrinkage and selection operator (LASSO) binary logistic regression model through tuning the lambda parameter by 5-fold cross-validation while evaluating area under curve (AUC) of receiver operating curve (ROC). The performance of the new and original models was compared through AUC of the ROC.Patient's tumors characteristics and outcomes were reported in the table below. The value of AUC of the original model was 0.831 (95% CI, 0.701-0.961), and 0.828 (95% CI, 0.700 -0.956) in the original and new validation cohorts, respectively, showing a reproducibility in the applicability of the original model. Nine features (eight radiomic features and one clinical parameter) were found to be significant with LASSO model, and added in the new model. Rad-score between pCR and non pCR group occurred in both of training (P < .001) and testing cohorts (P < .001); the ROC of new model was 0.926 (95% CI, 0.859-0.993) for training and 0.926 (95% CI, 0.767-1.00) for testing groups, improving the performance of the original model.Previously GLM model show good reproducibility in predicting pCR to CRT in LARC; the enhanced LASSO model developed in this study has the potential to improve prediction accuracy.