Development of a hybrid 2.5D deep learning model for glioma survival prediction using T1-weighted MRI from the CGGA database.

Background: Current glioma survival prediction relies on invasive molecular profiling. To overcome this, a non-invasive deep learning framework using T1-weighted contrast-enhanced MRI (T1CE) was developed to predict overall survival. This framework addresses computational limitations associated with the volumetric analysis while preserving important spatial information. Methods: We designed a hybrid 2.5D convolutional neural network to process multi-slice inputs, including the center slice and its adjacent slices, from 217 patients in the CGGA database. Transfer learning using ResNet and DenseNet architectures were employed to initialize the models. These models were subsequently fine-tuned with the Cox proportional hazards loss function. After the fine-tuning process was completed, the imaging signature was combined with clinical and molecular variables, including IDH and 1p19q status, to build an integrated model. Performance was evaluated via C-index, time-dependent AUC, and Kaplan-Meier analysis in independent training (70 %) and testing (30 %) cohorts. Results: The Combined model achieved superior discrimination, with a training C-index of 0.819 (95 % CI: 0.758-0.880) and a testing C-index of 0.804 (95 % CI: 0.708-0.900). It significantly outperformed the isolated Radiomic, deep learning (2D and 2.5D), and Clinical models (all p < 0.05). Moreover, time-dependent ROC analysis demonstrated consistent model performance over 1-5 years, with AUC values ranging from 0.851 to 0.906. The stratified survival curves clearly revealed distinct prognostic groups (log-rank p < 0.001). Conclusions: The 2.5D multi-source framework provides a clinically feasible, non-invasive tool for preoperative survival prediction, enabling personalized therapeutic strategies for glioma patients.