2308 - A Novel Mamba-Based 3D Dose Prediction Model with Channel-Aware Scan for Nasopharyngeal Carcinoma Radiotherapy
Presenter(s)
P. Zhou1,2, Q. Peng1, Y. Li3, and C. Li1,2; 1State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, Guangzhou, China, 2Sino-French Institute of Nuclear Engineering and Technology, Sun Yat-sen University, Zhuhai, China, 3Department of Radiation Oncology Physics, The First People’s Hospital of Foshan, Foshan, China
Purpose/Objective(s): This study introduces the 3D Channel-Aware Mamba Net (3D CAM Net), a novel Mamba-based model, to improve the accuracy of 3D dose distribution prediction for nasopharyngeal carcinoma (NPC) radiotherapy planning.
Materials/Methods: The 3D CAM Net integrates a 3D residual UNet backbone with the Mamba module, incorporating two key enhancements. First, a novel Channel-aware scan mechanism is introduced to improve the model's perception of multi-spatial and channel-specific information. Second, a hybrid loss function is designed to enhance model's sensitivity to dose gradient information. The model takes CT images and structure contours as input and predicts 3D dose distribution map as output. A dataset of 139 NPC patients undergoing intensity-modulated radiotherapy (IMRT) was used for training (95 plans), validation (20 plans), and testing (24 plans). Comparative evaluations were conducted against advanced models with diverse architectures: ResUNet++ (CNN architecture), Swin UNETR (Transformer architecture), and U-Mamba (Mamba architecture) to validate its effectiveness.
Results: 3D CAM Net achieves a mean voxel-based dose difference of 1.34 ± 0.22 Gy, which is significantly lower than that of all other evaluated models. It also demonstrates the smallest prediction error for 21 out of 27 evaluated anatomical structures and a significantly higher 3D gamma analysis (3%/3mm) pass rate of 82.83% ± 4.73%. Ablation studies further confirm the effectiveness of the proposed Channel-aware scan mechanism and hybrid loss function.
Conclusion: This study proposes 3D CAM Net, offering a novel and accurate solution for NPC dose prediction. The proposed framework demonstrates potential for enhancing the accuracy of 3D dose prediction.