MRI-based deep learning nomogram for prediction of early metachronous liver metastases in rectal cancer

Background: To investigate the potential of deep learning nomogram based on MRI for predicting early metachronous liver metastases (EMLM) in patients with rectal cancer. Methods A total of 170 patients with rectal cancer were enrolled and divided into training (2,080 images and later augmented to 12,480 images, 130 patients) and test (640 images, 40 patients) sets. A deep learning model was developed based on T2WI images using a 3D residual network (3D-ResNet). A clinical model based on independent predictors was then constructed via multivariate logistic regression. The nomogram was constructed by integrating the deep learning model with clinical model. Results The deep learning model exhibited good performance with AUCs of 0.856 and 0.812 in the training and test sets, respectively. Multivariate logistic regression revealed that mr-N stage (OR, 2.260; 95% CI: 1.322–3.865; p = 0.003) and mr-EMVI (OR, 2.682; 95% CI: 1.204–5.976; p = 0.016) were independent predictors of EMLM. The clinical model based on mr-N stage and mr-EMVI was developed with AUCs of 0.738 and 0.754 in the training and test sets, respectively. A nomogram was constructed based on three variables: deep learning signature, mr-N stage, and mr-EMVI. The AUCs of the nomogram were 0.906 and 0.895 in the training and test sets, respectively. The predictive performance of the nomogram was significantly superior to that of the clinical model and the deep learning model in both the training and test sets (p < 0.05). Conclusions The nomogram combining deep learning signature, mr-N stage and mr-EMVI may aid in the clinical prediction of EMLM in patients with rectal cancer.