Segmentation influence on the generalizability of the overall survival models in CT-based oropharyngeal carcinoma patients.

Objectives: To evaluate the impact of tumor volume segmentation variability on the repeatability of radiomic features (RFs) and to determine how RF repeatability influences the generalizability of radiomic models for predicting overall survival (OS) in patients with oropharyngeal carcinoma (OPC). Methods: We retrospectively analyzed CT images from 1017 patients with oropharyngeal carcinoma across three institutions. Perturbation methods were applied to simulate variations in gross tumor volume segmentation. RFs were extracted from both the original images and Laplacian of Gaussian-filtered images using different perturbation masks. RF repeatability was quantified using intra-class correlation coefficients (ICC). Repeatable RFs were progressively incorporated into the modeling process according to different ICC thresholds to assess the influence of feature repeatability on model generalizability. Results: Incorporation of RFs with ICC values between 0.7 and 0.8 improved the AUC index of the two-year and three-year OS models in external validation cohorts. Using an ICC threshold of 0.7, RFs were classified into high- and low-repeatability groups, and OS models were trained and validated using the training, internal testing, and external validation cohorts. Across all cohorts, the OS model trained with high-repeatability RFs demonstrated significantly superior performance compared to the model trained with low-repeatability RFs. Conclusion: The findings demonstrate that selecting RFs with ICC values greater than 0.7 substantially enhances both the generalizability and predictive performance of CT-based radiomic models for patients with OPC. This study further underscores the importance of considering RF repeatability, particularly in the presence of tumor volume segmentation variability, to improve the robustness and clinical reliability of radiomic models.