Generalizable MRI-based Nasopharyngeal Carcinoma Delineation: Bridging Gaps across Multiple Centers and Raters with Active Learning.

Purpose: To develop a deep learning (DL) method exploiting active learning and source-free domain adaptation for gross tumor volume (GTV) delineation in nasopharyngeal carcinoma (NPC), addressing the variability and inaccuracy when deploying segmentation models in multi-center and multi-rater settings.

Materials and methods: 1057 MRI scans of NPC patients from five hospitals were retrospectively collected and annotated by experts from the same medical group with consensus for multi-center adaptation evaluation. One dataset was used for model development (source domain), with the remaining four for adaptation testing (target domains). Meanwhile, another 170 NPC patients with annotations delineated by four independent experts were built for multi-rater adaptation evaluation. We evaluated the pre-trained model's migration ability to the four multi-center and four multi-rater target domains. Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95) and other metrics were used for quantitative evaluations.

Results: In the adaptation of dataset5 to other datasets, our source-free active learning adaptation method only requires limited labeled target samples (only 20%) to achieve a median DSC ranging from 0.70 to 0.86 and a median HD95 ranging from 3.16mm to 7.21mm for four target centers, and 0.78 to 0.85 and 3.64mm to 6.00mm for four multi-rater datasets. For DSC, our results for three of four multi-center datasets and all multi-rater datasets showed no statistical difference compared to the fully supervised U-Net model (P-values > 0.05) and significantly surpassed comparison models for three multi-center datasets and all multi-rater datasets (P-values < 0.05). Clinical assessment showed that our method-generated delineations can be used both in multi-center and multi-rater scenarios after minor refinement (revision ratio < 10% and median time < 2 minutes).

Conclusion: The proposed method effectively minimizes domain gaps and delivers encouraging performance compared with fully supervised learning models with limited labeled training samples, offering a promising and practical solution for accurate and generalizable GTV segmentation in NPC.