A Preoperative CT-based Multiparameter Deep Learning and Radiomic Model with Extracellular Volume Parameter Images Can Predict the Tumor Budding Grade in Rectal Cancer Patients

RATIONALE AND OBJECTIVES: To investigate a computed tomography (CT)-based multiparameter deep learning-radiomic model (DLRM) for predicting the preoperative tumor budding (TB) grade in patients with rectal cancer. METHODS: Data from 135 patients with histologically confirmed rectal cancer (85 in the Bd1+2 group and 50 in the Bd3 group) were retrospectively included. Deep learning (DL) features and hand-crafted radiomic (HCR) features were separately extracted and selected from preoperative CT-based extracellular volume (ECV) parameter images and venous-phase images. Six predictive signatures were subsequently constructed from machine learning classification algorithms. Finally, a combined DL and HCR model, the DLRM, was established to predict the TB grade of rectal cancer patients by merging the DL and HCR features from the two image sets. RESULTS: In the training and test cohorts, the AUC values of the DLRM were 0.976 [95% CI: 0.942-0.997] and 0.976 [95% CI: 0.942-1.00], respectively. The DLRM had good output agreement and clinical applicability according to calibration curve analysis and DCA, respectively. The DLRM outperformed the individual DL and HCR signatures in terms of predicting the TB grade of rectal cancer patients (p < 0.05). CONCLUSION: The DLRM can be used to evaluate the TB grade of rectal cancer patients in a noninvasive manner before surgery, thereby providing support for clinical treatment decision-making for these patients.