3466 - Delineation of Tumor Metabolic Volume in Nasopharyngeal Carcinoma on Computed Tomography Images Using a 3D U-Net Model

G. Zhou¹, Z. Mo², Y. X. Yang¹, Y. Xiong¹, H. Li², L. Jia², Y. He³, G. Y. Wang¹, X. Jiang¹, F. Chi¹, and Y. Sun⁴; ¹Sun Yat-sen University Cancer Center; State Key Laboratory of Oncology in South China; Collaborative Innovation Center for Cancer Medicine; Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangzhou, China, ²Shenzhen United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China, ³Shanghai United Imaging Healthcare Co., Ltd., Shanghai, China, ⁴Department of Radiation Oncology, Sun Yat-sen University Cancer Center, State Key Laboratory of Oncology in South China, Collaborative Innovation Center for Cancer Medicine, Guangzhou, China

Purpose/Objective(s): Monitoring metabolic activity within the tumor region during radiotherapy is crucial for patients with Nasopharyngeal Carcinoma (NPC). This practice enables the precise evaluation of treatment efficacy and may facilitate the application of adaptive radiotherapy (ART) strategies, leading to the optimization of irradiation regions and dosing. Materials/

Methods: This retrospective study was conducted at a single center in 18 NPC patients. Based on the deformation registration algorithm, planning Computed Tomography (CT) images for radiotherapy and pretreatment PET/CT images were registered to obtain the Gross Tumor Volume of the primary tumor (GTVp) delineation on PET and PET/CT. Subsequently, the delineation of Tumor Metabolic Volume (TMV) on PET images was accomplished by identifying areas with Standardized Uptake Values exceeding 2.5 within GTVp. Finally, the cavity and bone parts were removed from the delineation of TMV on PET/CT. For accurate segmentation of TMV on PET/CT, a seven-layer 3D U-Net model was established. To investigate the significance of localization information provided by the GTVp, we implemented diverse data preprocessing techniques for control experiments. One method involved using the center of the GTVp as the focal point for preprocessing cropping, while another entailed setting all voxel values outside the GTVp on PET/CT to -1000, akin to presenting the model with either ambiguous or precise localization cues. The consistency between label and prediction was evaluated using average dice similarity coefficient (DSC) and average symmetric surface distance (ASSD) with five-fold cross-validation. Furthermore, to assess the robustness of the model, we conducted volumetric analysis and shape evaluations of model predictions on 99 sets of daily Fan-beam Computed Tomography (FBCT) for 33 fractions in 3 patients.
Results: In the 5-fold cross-validation of 18 patients, the TMV model with ambiguous localization information achieved an average DSC of 0.69±0.11 and an ASSD of 1.8±0.7. In contrast, the TMV model with precise localization information attained an average DSC of 0.72±0.10 and an ASSD of 1.6±0.6. The TMV predicted by the model on daily FBCT showed a consistent decreasing trend, indicating progressive tumor shrinkage during treatment, aligning with the expected therapeutic effect.
Conclusion: We harnessed a seven-layer 3D U-Net architecture to delineate TMV on CT images with high precision, notwithstanding the use of a restricted dataset. The implementation of the TMV model aids in the delineation of biological target volumes for ART.