Generalizable Magnetic Resonance Imaging-based Nasopharyngeal Carcinoma Delineation: Bridging Gaps Across Multiple Centers and Raters With Active Learning

• DOI: 10.1016/j.ijrobp.2024.11.064
• Publication Date: 2024-11-16T16:16:39Z
• Authors: Xiangde Luo; Hongqiu Wang; Jinfeng Xu; Lu Li; Yue Zhao; Yuan He; Hui Huang; Jianghong Xiao; Tao Song; Shichuan Zhang; Shaoting Zhang; Guotai Wang; Wenjun Liao

To develop a deep learning method exploiting active and source-free domain adaptation for gross tumor volume delineation in nasopharyngeal carcinoma (NPC), addressing the variability inaccuracy when deploying segmentation models multicenter multirater settings. One thousand fifty-seven magnetic resonance imaging scans of patients with NPC from 5 hospitals were retrospectively collected annotated by experts same medical group consensus evaluation. data set was used model development (source domain), remaining 4 testing (target domains). Meanwhile, another 170 NPC, annotations delineated independent experts, created We evaluated pretrained model's migration ability to target domains. Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), other metrics quantitative evaluations. In dataset5 sets, our only requires limited labeled samples (only 20%) achieve median DSC ranging 0.70 0.86 HD95 3.16 7.21 mm centers, 0.78 0.85 3.64 6.00 sets. For DSC, results 3 sets all showed no statistical difference compared fully supervised U-Net (P values > 0.05) significantly surpassed comparison < 0.05). Clinical assessment that method-generated delineations can be both scenarios after minor refinement (revision ratio <10% time <2 minutes). The proposed effectively minimizes gaps delivers encouraging performance training samples, offering promising practical solution accurate generalizable NPC.