A Multimodal Graph Neural Network Based on MRI Radi-omics for Prognostic Prediction in Locally Advanced Cervical Squamous Cell Carcinoma.

IntroductionWe aimed to develop and validate a clinically interpretable multimodal deep learning framework integrating multi-sequence MRI and clinical variables for predicting overall survival (OS) in locally advanced cervical squamous cell carcinoma (LACSCC).MethodsThis retrospective cohort study included 247 patients with pathologically confirmed LACSCC. Pretreatment MRI sequences, including contrast-enhanced T1-weighted imaging (T1CE), T2-weighted imaging (T2WI), and diffusion-weighted imaging (DWI), were analyzed. Tumor regions of interest were manually delineated, and slice-level deep features were extracted using a fine-tuned ResNet-50 network. A graph neural network was then constructed to model spatial relationships among MRI slices, aggregate slice-level features into patient-level representations, and integrate imaging features with clinical variables to generate a Deep Learning Prognostic Score (DLPS). Model performance was assessed using time-dependent receiver operating characteristic (ROC) curves, Kaplan-Meier survival analysis, Cox regression analysis, and nomogram calibration.ResultsIn the validation cohort, the model achieved area under the ROC curve values of 0.747, 0.749, and 0.706 for predicting 1-year, 3-year, and 5-year OS, respectively. Kaplan-Meier survival analysis showed significantly better OS in the high-DLPS group than in the low-DLPS group. Multivariable Cox regression analysis identified DLPS as an independent predictor of OS. The nomogram integrating DLPS and significant clinical variables showed good calibration in both the training and validation cohorts.ConclusionWe developed and validated a multimodal MRI-based GNN framework for individualized OS prediction in LACSCC. By integrating spatially structured imaging features with clinical variables, the model provided effective prognostic stratification and may support personalized risk assessment and treatment decision-making.