Abstract
<title>Abstract</title> <p>Objective To develop a nomogram used to non-invasive prediction of nCRT outcome in patients with rectal cancer. Materials and Methods A total of 292 rectal cancer patients were retrospectively collected and randomly divided into a training (n = 230) and a validation dataset (n = 62). Both intra-tumoral and peri-tumoral regions were manually labeled on the T2-weight MR images. Deep features and radiomics features were extracted from the intra-tumor regions only and the combined intra- and peri-tumoral regions, respectively, and used for the development of deep learning models and radiomics model. Clinical risk factors were selected through univariate and multivariate regression analysis. Furthermore, a nomogram was constructed by integrating DL features, radiomics features and clinical risk factors. The performance of these models was evaluated by receiver operating characteristics curve (ROC) analysis and decision curve analysis (DCA). Results The AUCs of the DL model and radiomics model based on intra-tumor regions were 0.666 (95% CI, 0.534–0.780) and 0.817 (95% CI, 0.698–0.903) in the validation dataset, respectively. The performance of DL model and radiomics model was improved after integrating peri-tumoral regions, with AUCs increasing to 0.739 (95% CI, 0.612–0.842) and 0.845 (95% CI, 0.731–0.924) in the validation dataset, respectively. The nomogram showed the best predictive capability with AUCs achieving 0.921 (95% CI, 0.824–0.874) in the validation dataset, which was significantly higher than other models (all p-values < 0.05). All models showed good calibration, and the DCA demonstrated that the net benefit of the nomogram was higher than other models across majority range of threshold probabilities. Conclusion The nomogram incorporating deep learning features, radiomics signatures and clinical variables showed good performance and can be served as a noninvasive tool for predicting postoperative outcomes in patients with rectal cancer.</p>