Evaluation of Trident Network against Atlas and 2D U-Net methods for Auto-Segmentation of Organs at Risk in Nasopharyngeal Carcinoma: A Comparative Study

Jinghui Pan ^1,2* Fengjiao Zhuo ³ Dan Li ⁴ Xiangpan Li ¹ Qibin Song ¹ Hao Jiang ² Jinsheng Xiao ² Changli Ruan ¹

• ¹ Department of Radiation Oncology, Renmin Hospital, Wuhan University, Wuhan, China
• ² School of Electronic Information, Wuhan University, Wuhan, China
• ³ Department of Radiation Oncology, Jiangling County People's Hospital, Jingzhou, China
• ⁴ Department of Radiation Oncology, Hubei Cancer Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China

This study evaluates the accuracy of automatic segmentation of Organs at Risk (OARs) in nasopharyngeal carcinoma (NPC) using three approaches: atlas-based, 2D U-Net, and a self-developed Trident network. Our aim was to develop, validate, and compare the performance of Trident network for precise delineation of NPC using our center's dataset, in comparison to conventional atlas-based and 2D U-Net techniques.We randomly selected 209 patients with NPC for this retrospective study, with OARs manually delineated by physicians. An atlas template library was generated using data from seventeen OARs obtained from a subset of these patients, while the remaining 29 cases constituted the test set. The performance of auto-delineation methods, including Atlas, 2D U-Net, and the Trident Network, was compared to manual delineations. Accuracy was evaluated using the Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD), alongside timing each method. Statistical analysis included one-way ANOVA for normally distributed data, Welch's test for data with uneven variance, and Post Hoc Comparison tests (Least Significant Difference [LSD] and Tamhane's T2) for pairwise comparisons. The Kruskal-Wallis H test was employed for non-normally distributed data.In the evaluation of segmentation results for all 23 OARs in NPC, the Trident Network achieved the highest DSC (0.87±0.07), significantly outperforming the Atlas (0.67±0.02) and 2D U-Net (0.71±0.02). Additionally, the mean HD values for all the three methods were below 5 mm. In particular, with the exception of the right eyeball, the Trident network demonstrated superior DSC for each organ compared to the other two methods. The Trident Network showed high morphological similarity from Atlas and 2D U-Net in most structures, performing significantly better than Atlas for multiple structures such as the Hypophysis, Optic chiasma, Esophagus, and others. Furthermore, it showed better performance than the 2D U-Net in several structures, including the optic chiasma, optic nerve, and larynx. Conversely, the 2D U-Net excelled over Atlas in structures like Brainstem and Lens.The Trident Network demonstrates superior morphological and geometric accuracy compared to Atlas and 2D U-Net in the delineation of OARs in nasopharyngeal carcinoma, significantly reducing the need for manual corrections and improving delineation efficiency.