Multi-institutional deep learning for GTV segmentation and survival prediction in nasopharyngeal carcinoma

To develop an automated workflow for gross tumor volume (GTV) segmentation in radiotherapy planning CT images of nasopharyngeal carcinoma (NPC) patients and to evaluate the 5-year Disease-Free Survival (DFS) predictive performance of radiomics, clinical and combined features. Contrast-enhanced planning CTs of 75 NPC patients were collected. SwinUNETR, UNETR and nnU-Net models were trained with five-fold cross-validation; performance was quantified by Dice similarity coefficient (DSC). Additional 120 scans from SeGrap2023 were incorporated to assess the model performances on diverse cohorts. For DFS prediction, 1059 slice-wise radiomic features were extracted. Feature selection used univariate filtering, correlation thresholding and LASSO, followed by machine-learning modelling with radiomic, clinical and combined inputs. The highest 5-fold cross-validation DSC performance was achieved by nnU-Net (DSC = 0.79) when trained and internally validated only on the SeGrap2023 dataset. However, DSC dropped to 0.36 when Acibadem cohort data were utilized for external test set, suggesting a significant effect of the domain shift. When the Acibadem and SeGrap2023 datasets were combined, nnU-Net achieved an average DSC of 0.73 in 5-fold cross-validation and 0.69 on subset of Acibadem internal test cases. In predicting 5-year DFS, a Logistic Regression model using combined radiomic and clinical features provided the highest AUC score (0.79), outperforming clinical-only (AUC = 0.59) and radiomics-only (AUC = 0.63) feature sets. Multi-institutional training mitigates domain shift and boosts segmentation robustness. In addition, integrating radiomics with clinical data enhances DFS prediction in NPC. Advanced deep-learning and machine-learning pipelines can refine radiotherapy planning and prognostication, supporting personalized management and improved outcomes.