ESTRO 2024 - Abstract Book

S3119

Physics - Autosegmentation

ESTRO 2024

delineation task by harnessing the plain CT images from the training set. Specifically, the state-of-the-art nnUnet model was trained and evaluated to auto-delineate the CTV, and the resulting model performance is served as a baseline for assessment. To take advantage of the informative CECT images, we embarked on the creation of the Onet model (as shown in Figure 1.). This innovative network architecture consists of a dual Unet, a fusion module, and a fusion loss function which seamlessly amalgamated insights from both plain CT and CECT. This integrative scheme is designed to leverage the CECT information to improve automatic delineation accuracy. Ablation experiment where both plain CT and CECT served as multimodal input for training and prediction using the nnUnet (multimodal training strategy) was also performed and quantitatively compared with the Onet. The automated delineation of the CTV was conducted for a total of 11 patients, and the evaluation was based on multiple criteria, including Dice Coefficient, Average Surface Distance (ASD), and Hausdorff Distance (HD). The training and testing of the above three models were carried out on a GPU (RTX3090, NVIDIA) server.

Results:

1. DICE Scores: The DICE scores for the nnUnet (plain CT), nnUnet (plain CT + CECT), and Onet were 0.737 ± 0.054, 0.717 ± 0.07, and 0.817 ± 0.024, respectively. These scores indicated the Onet provides the best CTV overlap and similarity between the automated and manual delineations. 2. ASD Measurements were 1.138 ± 0.037 mm, 1.139 ± 0.056 mm, and 1.103 ± 0.044 mm for the nnUnet (plain CT), nnUnet (plain CT + CECT), and Onet, respectively, showing that the Onet provides the lowest average surface distance between the automated and manual CTV delineations 3. HD Values were 13.202 ± 3.132 mm, 13.258 ± 4.391 mm, and 11.219 ± 2.004 mm for the nnUnet (plain CT), nnUnet (plain CT + CECT), and Onet, respectively, showing the Onet provides the smallest Hausdorff distance between automated and manual CTV delineations