Radiomics of rectal cancer for predicting distant metastasis and overall survival

Abstract

BACKGROUND: Rectal cancer (RC) patient stratification by different factors may yield variable results. Therefore, more efficient prognostic biomarkers are needed for improved risk stratification, personalized treatment, and prognostication of RC patients. AIM To build a novel model for predicting the presence of distant metastases and 3-year overall survival (OS) in RC patients. METHODS This was a retrospective analysis of 148 patients (76 males and 72 females) with RC treated with curative resection, without neoadjuvant or postoperative chemoradiotherapy, between October 2012 and December 2015. These patients were allocated to a training or validation set, with a ratio of 7:3. Radiomic features were extracted from portal venous phase computed tomography (CT) images of RC. The least absolute shrinkage and selection operator regression analysis was used for feature selection. Multivariate logistic regression analysis was used to develop the radiomics signature (Rad-score) and the clinicoradiologic risk model (the combined model). Receiver operating characteristic curves were constructed to evaluate the diagnostic performance of the models for predicting distant metastasis of RC. The association of the combined model with 3-year OS was investigated by Kaplan-Meier survival analysis. RESULTS A total of 51 (34.5%) patients had distant metastases, while 26 (17.6%) patients died, and 122 (82.4%) patients lived at least 3 years post-surgery. The values of both the Rad-score (consisted of three selected features) and the combined model were significantly different between the distant metastasis group and the non-metastasis group (0.46 ± 0.21 vs 0.32 ± 0.24 for the Rad-score, and 0.60 ± 0.23 vs 0.28 ± 0.26 for the combined model; P < 0.001 for both models). Predictors contained in the combined model included the Rad-score, pathological N-stage, and T-stage. The addition of histologic grade to the model failed to show incremental prognostic value. The combined model showed good discrimination, with areas under the curve of 0.842 and 0.802 for the training set and validation set, respectively. For the survival analysis, the combined model was associated with an improved OS in the whole cohort and the respective subgroups. CONCLUSION This study presents a clinicoradiologic risk model, visualized in a nomogram, that can be used to facilitate individualized prediction of distant metastasis and 3-year OS in patients with RC.