2765 - Expert-Constrained Deep Learning Framework for Nasopharyngeal Carcinoma Automated Target Delineation
Presenter(s)
S. Zhao1, Q. Liu1, L. Jia2, Z. Wei2, Z. Zhang1, K. Yang1, J. Han3, and J. Huang1; 1Cancer Center, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China, 2Cooperation Innovation Department, Shanghai United Imaging Healthcare Co., Ltd, Shanghai, China, 3Institute of Radiation Oncology, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China
Purpose/Objective(s): Radiotherapy is the primary treatment for non-metastatic nasopharyngeal carcinoma (NPC), where both therapeutic efficacy and toxicity depend on precise target delineation. However, manual contouring for NPC remains time-consuming and labor-intensive. This study proposes an expert-constrained auto-contouring algorithm to improve efficiency, while validating its accuracy and clinical acceptability.
Materials/Methods: Between October 2020 and October 2023, 84 NPC patients treated with intensity-modulated radiotherapy (IMRT) were retrospectively enrolled and randomly divided into a training set (60 cases), a validation set (9 cases), and a test set (15 cases) for model development, validation, and evaluation. Two senior radiation oncologists independently reviewed the ground truth delineations for the primary gross tumor volume (GTVp), nodal gross tumor volume (GTVn), high-risk clinical target volume (CTV1), and low-risk clinical target volume (CTV2). We developed ECAUNet, a novel deep learning framework integrating spatial constraints from expert-modified GTVp delineations to guide automated contouring of GTVn, CTV1, and CTV2. An additional prospective cohort of 15 NPC patients (January to October 2024) was analyzed and compared to a baseline nnUnet model trained on identical data. Model performance was quantified using the Dice similarity coefficient (DSC), average surface distance (ASD), and the time required for contour modification.
Results: In the 15 test patients, using ground truth delineations as references, our model achieved an average DSC of 0.70 ± 0.08, 0.92 ± 0.05, and 0.89 ± 0.03 for GTVn, CTV1, and CTV2, respectively, with an average ASD of 1.12 ± 0.07, 1.13 ± 0.07, and 0.52 ± 0.19. Compared to a model trained without expert-constrained GTVp input (mean DSC: 0.61 ± 0.11 for GTVn, 0.83 ± 0.03 for CTV1, and 0.87 ± 0.03 for CTV2), our expert-constrained model improved the DSC by 16%, 11%, and 2% for GTVn, CTV1 and CTV2 (P<0.01, P<0.01 and P=0.09, respectively). In the prospective cohort of 15 patients, the average DSC between physician-modified contours and ECAUNet-generated contours was 0.64 ± 0.10 for GTVn, 0.94 ± 0.01 for CTV1, and 0.89 ± 0.01 for CTV2, compared to 0.63 ± 0.11, 0.93 ± 0.01, and 0.89 ± 0.01 from the baseline nnUNet model. Additionally, physicians required an average of 29.7 minutes to review and refine ECAUNet-generated contours, demonstrating a 41% reduction in time compared to full manual contouring (50.3 minutes).
Conclusion: The proposed model effectively integrates expert knowledge into automated contouring, demonstrating improved accuracy and clinical applicability for NPC target auto-contouring. This approach has the potential to standardize radiotherapy workflows while reducing time costs in clinical practice.