Validation of MRI-based nnU-Net model for automated segmentation of neck lymph nodes in head and neck squamous cell carcinoma: a multicenter study.
Authors
Affiliations (5)
Affiliations (5)
- Beijing Tongren Hospital, Capital Medical University, Beijing, China.
- Beijing Friendship Hospital, Capital Medical University, Beijing, China.
- Beijing Tongren Hospital, Capital Medical University, Beijing, China. [email protected].
- Beijing Friendship Hospital, Capital Medical University, Beijing, China. [email protected].
- Beijing Tongren Hospital, Capital Medical University, Beijing, China. [email protected].
Abstract
To develop and externally validate an MRI-based deep learning framework for automated 3D segmentation of neck lymph nodes (LNs) in head and neck squamous cell carcinoma (HNSCC), and to assess segmentation accuracy, volumetric agreement, and time efficiency. Axial head and neck MRI data were retrospectively collected from two centers. Each LN was manually delineated following two strategies: margin-excluding approach (excluding a 1-2 mm perinodal margin) and boundary-adherent approach (following the LN contour). For each strategy, nnU-Net (a DL framework) was applied for training three single-sequence models using contrast-enhanced T1-weighted (CE-T1w), T1-weighted (T1w), and T2-weighted (T2w) imaging. The model performance was evaluated in the internal validation cohort (n = 99) and the external test cohort (n = 150). Efficiency was assessed by comparing manual vs. DL-assisted contouring times in a separate cohort (n = 851). In the internal validation cohort, the model trained on CE-T1w images with boundary-adherent annotations achieved the highest overlap (median Dice similarity coefficient of 0.795). CE-T1w model showed high volumetric agreement with manual segmentation (concordance correlation coefficient of 0.939 for margin-excluding and 0.899 for boundary-adherent strategies), and its performance was confirmed in the external test cohort. DL assistance reduced contouring time by 33.14% (n = 851, mean 100.22 s vs. 149.89 s, p < 0.001). The two annotation policies yielded similar LN counts but significantly different volume endpoints (p < 0.001). We developed and validated a deep learning framework capable of accurate and efficient 3D LN segmentation using single-sequence model inputs after CE-T1w-referenced preprocessing. The model shows strong generalization across multiple centers. This tool can standardize volumetric assessments and improve clinical workflow efficiency in HNSCC.