Objectives: Our aim was to develop an optimal model for predicting distant metastasis (DM) in rectal cancer (RC) via integrating tumor lymphocytic reaction (LR) with deep learning radiomic (DLR) features. Methods: A total of 190 patients with RC were divided into a training cohort (n = 133) and a validation cohort (n = 57) in a 7:3 ratio. Preoperative radiomics (Rad), deep transfer learning (DTL), and DLR features were extracted from T2WI scans, and clinicopathological variables were collected. All models were constructed using the support vector machine (SVM) algorithm, and their performance was evaluated using the area under the receiver operating characteristic curve (AUC). A 3-year follow-up was conducted to analyze 3-year distant-metastasis-free survival (DMFS) outcomes and DM risk. Results: Multivariate analysis confirmed LR as an independent predictor of 3-year DMFS (p < 0.05). The fusion model integrating LR and DLR exhibited preliminary predictive performance, with AUC values of 0.911 and 0.884 in the training and validation cohorts, respectively. Subgroup observational analyses showed apparent DMFS differences associated with adjuvant chemotherapy exposure among low-risk patients, while such survival patterns were not evident in the high-risk subgroup defined by the nomogram cut-off of 0.3. A moderate negative correlation was detected between DLR signatures and LR (r = −0.44, p < 0.001), providing preliminary immune-related clues for interpreting deep learning radiomic biomarkers. Conclusions: The multimodal fusion nomogram combining pathological LR and DLR signatures shows encouraging preliminary predictive performance for 3-year DMFS risk in patients with RC. This postoperative multimodal tool may provide a preliminary reference for clinicians to implement individualized postoperative surveillance for patients with RC, and the inverse association between DLR and LR helps reveal the immune-related background of imaging signatures.
Objectives: To develop and validate a hybrid deep learning–machine learning (DL-ML) fusion model for noninvasive preoperative differentiation of solitary pulmonary mucinous adenocarcinoma (SPMA). Methods: A total of 200 patients with pathologically confirmed lung adenocarcinoma, including 37 SPMA cases, treated at Tianjin Union Medical Center between 2018 and 2025 were retrospectively enrolled. Patients were randomly assigned to a training cohort (n = 140) and a test cohort (n = 60). Clinical characteristics, radiomic features, and deep learning features extracted via ResNet50 were integrated. Least absolute shrinkage and selection operator (LASSO) regression was used for feature selection, and multiple machine learning classifiers were compared. Model performance was assessed using the receiver operating characteristic (ROC) curve, decision curve analysis (DCA), and calibration curve analysis (CCA). Results: LASSO regression identified 10 optimal features, comprising 8 radiomic and 2 deep learning features. The random forest classifier yielded the best performance. The hybrid DL-ML fusion model yielded the highest AUC of 0.982 in the training cohort and 0.878 in the test cohort, significantly outperforming the clinical model (Clinic), radiomic model (Rad), deep transfer learning (DTL) model, deep learning–radiomics (DLR) model. In the test cohort, the hybrid DL-ML fusion model achieved an AUC of 0.878, which was significantly higher than that of the clinical model (0.755; p < 0.05). DCA and CCA confirmed favorable clinical utility and calibration. Conclusions: The hybrid DL-ML fusion model enables accurate, noninvasive preoperative differentiation of SPMA. It outperforms conventional clinical assessment and single-modality imaging models, with promising potential for noninvasive preoperative differential diagnosis.