Fudan University Shanghai Cancer Center Shanghai, Shanghai
X. Ou1,2, J. Wang1,2, W. Yan1,2, Z. Wei3, L. C. Jia4, Y. Wang5, and C. Hu1,2; 1Department of Oncology, Fudan University Shanghai Medical College, Shanghai, China, 2Department of Radiation Oncology, Fudan University Shanghai Cancer Center, Shanghai, China, 3Real Time Laboratory, United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China, 4Shenzhen United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China, 5Central Research Institute, United Imaging Healthcare Group, Shanghai, China
Purpose/Objective(s): The precise delineation of tumor targets in the radiotherapy of nasopharyngeal carcinoma (NPC) is critically linked to the treatment outcomes. The application of Artificial Intelligence (AI) can significantly enhance the accuracy and efficiency of clinical target volume (CTV) delineation. However, there is a paucity of research investigating the ability of AI-delineated nasopharyngeal carcinoma CTV to encompass areas of recurrence. To address this, we proposed an AI-driven approach for the automated delineation of nasopharyngeal carcinoma CTV that encompasses recurrent regions on CT images. Materials/
Methods: We retrospectively collected planning CT scan data from 51 NPC patients, who were diagnosed as local recurrent disease after definitive radiotherapy using intensity-modulated radiation therapy. This cohort was randomly divided into three datasets for training (35 cases), validation (8 cases), and testing (6 cases). The gold standard delineations of the primary gross tumor volume (GTVp), clinical target volume (CTV), and recurrent regions (GTV-re) were manually outlined by two experienced radiation oncologists according to international guidelines of CTV delineation of NPC. We introduced a three-dimensional UNet with an effective channel attention mechanism (ECAUNet), utilizing GTVp as an empirical constraint input to guide the automatic delineation of CTV. The automatically delineated recurrent region was incorporated into the model parameter optimization solely as part of the loss function. The performance of the automatic delineation of CTV was evaluated using the mean Dice Similarity Coefficient (DSC) and the average Surface Distance (ASD). The ability of the AI-delineated CTV to cover recurrent areas was assessed using Sensitivity, defined as the proportion of the recurrent area within the AI-delineated CTV relative to the entire recurrent area. Results: The method we proposed showed superior segmentation accuracy in the CTV target region delineation compared to the gold standard manually outlined by medical professionals. This was evident across a test set comprising 6 cases, where the average Dice Similarity Coefficient (DSC) for CTV reached 0.95±0.01, accompanied by an average Surface Distance (ASD) of 1.76±2.91. Additionally, the Sensitivity of CTV was 0.98±0.03, indicating a high degree of coverage of the recurrent regions by the automatically delineated. Conclusion: Our AI-driven auto-delineation algorithm exhibits high accuracy and extensive coverage of recurrent areas in the delineation of CTV for NPC on non-contrast CT scans. By integrating expert knowledge of the GTVp as an a priori constraint, we have effectively improved the accuracy of CTV delineation. Furthermore, the inclusion of recurrent region prediction in the model parameter optimization process helps ensure that the automated CTV delineations more closely meet clinical standards.
