Deep learning-based independent lymph node segmentation in esophageal cancer: a precise and efficient approach for radiotherapy planning.

Background: Esophageal cancer (EC) is among the most commonly diagnosed malignancies and ranks as the seventh leading cause of cancer-related mortality globally. Radiotherapy (RT) plays a key role in the management of EC by controlling local recurrence and improving survival outcomes. However, in current clinical practice, the delineation of lymph node (LN) clinical target volumes (CTVs) for RT remains time-consuming and labour-intensive. Deep learning-based auto-segmentation (DLBAS), as the fourth generation of this technology, offers superior results in medical image segmentation. The purpose of this study was to train and validate DLBAS models for the segmentation of distinct LN CTVs in EC, with the aim of supporting and streamlining the delineation process. Methods: Contrast-enhanced computed tomography (CT) scans were acquired from EC patients between October 2013 and December 2018, within 2 weeks prior to surgery. Two experienced radiation oncologists manually delineated LN CTVs, including stations 101, 104, 105, 106, 107, 108 and 109. We developed DLBAS models based on the no-new-U-Net (nnU-Net) and quantitatively evaluated their segmentation performance using metrics such as the dice similarity coefficient (DSC) and the 95% Hausdorff distance (95HD). To assess clinical utility, another two radiation oncologists scored the automated segmentations in the test sets. Additionally, the time required for manual versus automated segmentation was recorded. Results: A total of 364 patients with 1,131 LN CTVs were included in this study. The DLBAS models achieved a mean DSC ranging from 0.71 to 0.79 for all LN CTV stations. The majority of automated segmentations received a score of ≥2 (indicating minimal or no correction required), accounting for 89.8% of the total cases, and none of the structures received a score of 0 (unusable). Furthermore, the DLBAS model significantly reduced the contouring time, decreasing the average time per LN CTV from 257 seconds to 79 seconds compared to manual segmentation. Conclusions: The proposed DLBAS model effectively improves both the efficiency and accuracy of LN CTV segmentation. These findings demonstrate its suitability for clinical application and highlight its promise for facilitating CT-guided adaptive radiation therapy in EC.