ESTRO 2025 - Abstract Book

S2444

Physics - Autosegmentation

ESTRO 2025

We retrospectively collected data from 126 NPC patients, including original planning CT images and CT images of ART. The dataset was randomly divided into a training set (84 cases), a validation set (22 cases), and a testing set (20 cases). Ground truth for GTVn (subdivided into GTVn_L in the left neck and GTVn_R in the right neck) and CTVs (including high-risk CTV1 and low-risk CTV2) were reviewed by two clinical experts. We introduced a channel attention module to build the ECA-UNet, using the target delineations from the original plan and CT images during ART as inputs. This approach utilizes prior information from the original plan to improve the delineation accuracy. The detailed algorithm flow diagram and network structure can be shown in Figure 1. The performance of the model was evaluated using the Dice similarity coefficient (DSC) and 95% Hausdorff distance (HD95). Additionally, ablation experiments were conducted to compare the impact of incorporating original plan target information on the model's effectiveness. Results: In the testing set, the average DSC values were 0.72 ± 0.06, 0.76± 0.03, 0.96 ± 0.07, and 0.93 ± 0.03 for GTVn_L, GTVn_R, CTV1, and CTV2, respectively. The average HD95 values were 2.52 ± 0.38 mm, 3.75 ± 1.64 mm, 1.25 ± 0.43 mm, and 1.86 ± 0.59 mm, respectively. Compared to not incorporating original plan target information, the introduction of this information improved the accuracy of automatic delineation for GTVn_L, GTVn_R, CTV1, and CTV2 by 19%, 17%, 11%, and 8%, respectively.

Conclusion: The proposed auto-delineation model leverages target information from the original plan to achieve high accuracy in delineating GTVn and CTVs, facilitating the rapid and effective implementation of online ART for NPC.

Keywords: Automatic Delineation, Nasopharyngeal Carcinoma