This study retrospectively utilized 521 colorectal cancer cases and actively took cases from 381 patients. Preoperative CT and postoperative hematoxylin-eosin stained slides were split into training and internal validation sets (7:3 ratio) for a deep learning-based multiomics model. This model was then used to predict postoperative distant metastasis and evaluate survival prognosis for colorectal cancer patients. When patients took part in a follow-up, their distant metastasis status was used to assert whether the systems had accurately predicted disease-free survival. With the information from these patients, two models were developed: a deep-learning radionics (DLRS) and deep-learning pathomics (DLPS) model. Within the training set, area under the curve (AUC) for radiological, pathological, DLRS, DLPS, Nomogram 1, and Nomogram 2 models were 0.657, 0.687, 0.931, 0.914, 0.938, and 0.930. These models outperformed conventional methods with statistical significance (P < 0.05). These models enable reliable disease-free survival risk stratification.