H. Sun1, and L. N. Zhao2; 1Department of Radiation Oncology, Xijing Hospital, Fourth Military Medical University, xian, Shaanxi, China, 2Department of Radiation Oncology, Xijing Hospital, Air Force Medical University, Xian, China
Purpose/Objective(s): The gross tumor volume (GTV) for nasopharyngeal carcinoma (NPC) cancer exhibits a highly complex distribution characterized by uncertainty regarding location, size, and shape. A novel adaptive segmentation model is proposed to improve the accuracy of GTV delineation for NPC. Materials/
Methods: A total of 496 nasopharyngeal carcinoma patients from four centers were collected. The imaging data included CT, T1, T1-enhanced, and T2 MRI scans. A dynamic information fusion mechanism is introduced into the adaptive segmentation model. On the one hand, feature information among multimodal medical images is horizontally filtered based on the frequency extracted during a single training session. On the other hand, feature information within each modality image is vertically filtered based on the frequency extracted during multiple training sessions. According to the two types of screening methods, weight coefficients are dynamically assigned. Through dynamically assigning weighting coefficients based on two filtering methods, the crucial feature information from multimodal medical is adaptive fused, thereby ensuring the effectiveness of the fused feature information for delineating GTV(including GTVnx and GTVnd) across multiple centers. Based on the above-mentioned fusion features, the Markov chain learning model of the conditional diffusion probability model is used to achieve quantitative segmentation of GTV. The segmentation accuracy of the model was evaluated using dice similarity coefficient (DSC) (%), Hausdorff distance of 95% (HD95%) (mm), mean surface distance (MSD) (mm). Results: For the test set, the results of the three measurement indicators of DSC, HD95% and MSD based on the adaptive segmentation model are expressed as means (standard deviations), which are 81.05±2.16, 7.53±4.98, and 1.96±1.04 respectively. From the results of five-fold cross-validation, it can be seen that the average DSC accuracy in the validation set fluctuates within 3%, and the model shows good stability. In addition, through ablation comparison experiments, models with dynamic information fusion mechanisms can improve DSC delineation accuracy by an average of 9.8%. Conclusion: The new model can achieve more accurate and stable GTV segmentation of multi-center NPC cases. The adaptive segmentation scheme provides new ideas for improving the efficacy of adaptive radiotherapy.
