3429 - Generalizability of an Automatic GTV Segmentation Model in the Reirradiation Setting for Oropharyngeal Cancer
Presenter(s)
A. Beddok1,2, Y. Dong1,3, K. Grogg1,3, E. Lissavalid1,3, M. Galal1,3, M. Moteabbed4, J. Woo5, H. A. Shih4, C. Ma1,3, G. El Fakhri1,3, and T. Marin1,3; 1Yale Biomedical Imaging Institute, Yale University School of Medicine, New Haven, CT, 2Department of Radiation Oncology, Institut Godinot, Reims, France, 3Department of Radiology and Biomedical Imaging, Yale University School of Medicine, New Haven, CT, 4Department of Radiation Oncology, Mass General Brigham / Massachusetts General Hospital, Boston, MA, 5Gordon Center for Medical Imaging, Harvard Medical School, Boston, MA
Purpose/Objective(s): Deep learning–based tumor segmentation models perform well in primary oropharyngeal squamous cell carcinoma (OSCC), but their robustness in patients requiring reirradiation (reRT) is unclear. This study evaluated whether a model trained on primary OSCC generalizes to the reRT setting and whether performance differences are independent of tumor volume.
Materials/Methods: A nnU-Net–based automatic gross tumor volume (GTV) segmentation model was trained on 148 patients with primary OSCC from three publicly available multi-institutional datasets. All patients underwent pre-treatment PET-CT imaging. External validation was performed on an independent primary cohort (n=55). The trained model was then applied to 15 patients with recurrent OSCC treated with reRT at an independent institution. To distinguish institutional from anatomical effects, performance was also assessed in 24 primary OSCC patients from the same institution. Institutional GTVs were delineated on PET-CT by an expert radiation oncologist. Segmentation accuracy was evaluated using the Dice similarity coefficient (DSC) and 95th percentile Hausdorff distance (HD95). Group differences were assessed using Kruskal–Wallis and pairwise Wilcoxon tests. Multivariable linear regression evaluated the independent association between reRT status and DSC after adjustment for tumor volume.
Results: In the external primary cohort (n=55), median DSC was 0.74 (IQR 0.13), with a mean of 0.68 ± 0.18. In institutional primary cases (n=24), median DSC was 0.60 (IQR 0.14), mean 0.57 ± 0.18, indicating a modest decrease in performance across primary cohorts. In contrast, performance markedly declined in the reRT cohort (n=15), with median DSC 0.00 (IQR 0.33) and mean 0.19 ± 0.28. Median HD95 was comparable between primary cohorts (6.67 mm and 6.78 mm), despite differences in tumor volume, but increased to 15.40 mm in the reRT cohort. Differences in DSC were significant across groups (Kruskal–Wallis p = 2.5×10?7). Pairwise analyses showed a modest but significant difference between primary cohorts (p = 0.002), whereas performance in the reRT cohort was substantially lower than in both primary cohorts (p = 0.001). Tumor volumes differed across groups (p = 4.5×10?6), with smaller lesions observed in both institutional cohorts, particularly in reRT patients. After adjustment for tumor volume, institutional differences were no longer significant, whereas reRT status remained independently associated with decreased DSC in multivariable analysis (ß = -0.41, p = 1.0×10??).
Conclusion: A segmentation model trained on primary OSCC does not generalize to recurrent tumors requiring reRT. Performance degradation persists after adjustment for tumor volume, indicating a domain shift related to prior irradiation. reRT-specific training or domain adaptation may be necessary for safe clinical implementation.