Advancing Delineation of Gross Tumor Volume Based on Magnetic Resonance Imaging by Performing Source-Free Domain Adaptation in Nasopharyngeal Carcinoma

Hongqiu Wang, Shichen Zhang, Xiangde Luo, Wenjun Liao, Lei Zhu

Abstract: Nasopharyngeal carcinoma (NPC) is a common and significant malignancy that primarily affects the head and neck region. Accurate delineation of Gross Tumor Volume (GTV) is crucial for effective radiotherapy in NPC. Although Magnetic Resonance Imaging (MRI) allows precise visualization of tumor characteristics, challenges arise due to the complex anatomy of the nasopharyngeal region and GTV localization. Recently, deep learning based methods have been extensively explored in automatically segmenting various types of tumors, and promising progress has been made. However, the substantial reliance on particular medical images with adequate supervised annotations, coupled with the restricted access to clinical data in hospitals, presents significant obstacles in employing computer-aided segmentation for radiotherapy. Therefore, we propose a novel source-free domain adaptation framework that transfers knowledge of tumor segmentation learned in the source domain to the unlabeled target dataset without the access to the source dataset and annotate the target domain, for the NPC. Specifically, We enhances model performance by jointly optimizing entropy minimization and pseudo-labeling based on source rehearsal. We validated our approach on 406 patients data collected from three hospital centers and achieved an effective accuracy improvements compared to directly employing model trained on the source domain.