Automated Lymph Node Localization and Segmentation in Patients with Head and Neck Cancer: Opportunities and...

Background/Objectives: Accurate assessment of lymph nodes is of paramount importance for correct cN staging in head and neck cancer; however, it is very time-consuming for radiologists, and lymph node metastases of head and neck cancers may show distinct characteristics, such as central necrosis or very large size. Here, we evaluate the performance of a previously developed generic cervical lymph node segmentation model in a cohort of patients with head and neck cancer. Methods: In our retrospective single-center, multi-vendor study, we included 125 patients with head and neck cancer with at least one untreated lymph node metastasis. On the respective cervical CT scan, an experienced radiologist segmented lymph nodes semi-automatically. All 3D segmentations were confirmed by a second reader. These manual segmentations were compared to segmentations generated by an AI model previously trained on a different dataset of varying cancers. Results: In cervical CT scans from 125 patients (61.9 years ± 10.6, 100 men), 3656 lymph nodes were segmented as ground-truth, including 544 clinical metastases. The AI achieved an average recall of 0.70 with 6.5 false positives per CT scan. The average global Dice accounts for 0.73 per scan, with an average Hausdorff distance of 0.88 mm. When analyzing the individual nodes, segmentation accuracy was similar for non-metastatic and metastatic lymph nodes, with a sensitivity of 0.89 and 0.85. Localization performance was lower for metastatic than for non-metastatic lymph nodes, with a recall of 0.65 and 0.74, respectively. Model performance was worse for enlarged nodes (short-axis diameter ≥ 15 mm), with a recall of 0.36 and a sensitivity of 0.67. Conclusions: The AI model for generic cervical lymph node segmentation shows good performance for smaller nodes (SAD ≤ 15 mm) with respect to localization and segmentation accuracy. However, for clearly enlarged and necrotic nodes, a retraining of the generic AI algorithm seems to be required for accurate cN staging.