Radiomic prediction of early progression at 2 years post-treatment in patients with resectable rectal cancer based on rectal tumor and mesentery characteristics: a two-center study.

Background: Resectable rectal cancer carries a high risk of postoperative early progression (EP) within 2 years, which seriously affects patients' long-term survival outcomes. Current routine clinical evaluation strategies cannot accurately identify high-risk individuals, while radiomics and deep learning derived from preoperative magnetic resonance imaging (MRI) allow quantitative mining of hidden tumor microenvironmental features. This study aimed to explore the predictive value of preoperative MRI radiomics and deep learning for EP in patients with rectal cancer. Methods: A retrospective analysis included 255 patients with rectal cancer from The First Affiliated Hospital of Soochow University (Center 1) and 69 patients from the Suzhou Hospital Affiliated to Nanjing Medical University (Center 2). Patients from Center 1 were divided into a training set (n=204) and a test set (n=51) in an 8:2 ratio. Patients from Center 2 served as the external validation set (n=69). These patients were followed up, and the occurrence of EP (tumor recurrence or metastasis within 2 years of treatment) was recorded. Of the 324 patients with rectal cancer, 81 had EP and 243 did not. Features were extracted and filtered from MRI imaging data to construct radiomics and deep-radiomics models. Receiver operating characteristic (ROC) curves were plotted. The area under the ROC curve (AUC) was calculated to compare the models' predictive performance. Additionally, decision curves were used to assess the models' discriminatory performance. The SHapley Additive exPlanations (SHAP) method was used to visually demonstrate the feature contributions in the performance model and to illustrate the interpretability of the best model. Results: Nine radiomic features and ten deep radiomic features were selected. For the radiomics model, a multi-layer perceptron (MLP) was chosen as the classifier. For the deep radiomics model, extremely randomized trees (ExtraTrees) were chosen as the classifier. According to SHAP, the features contributing most markedly to the deep radiomics model were, in order: DL_resnet18_3, DL_resnet18_6, and T2surround_wavelet-HTL_glcm_Imc2. The deep-radiomics model [AUC =0.881; 95% confidence interval (CI): 0.821-0.940] significantly outperformed the radiomic model (AUC =0.777; 95% CI: 0.701-0.854). The Delong test revealed that the difference in predictive performance between the two models was extremely statistically significant (P<0.02). Conclusions: A deep-radiomics model based on pre-treatment MRI images demonstrates good performance in predicting post-treatment EP outcomes in patients with resectable rectal cancer. The SHAP algorithm provides interpretability at the individual level and supplies a basis for personalized post-treatment care in these patients.