Discover the SciOpen Platform and Achieve Your Research Goals with Ease.
Search articles, authors, keywords, DOI, etc.
To address the clinical challenges of poor esophageal tissue contrast, inherent segmentation difficulty, and suboptimal performance of generic segmentation models on non-contrast computed tomography (NCCT), this study developed a dual-layer computed tomography (DLCT) annotation-guided 3D no new U-Net (3D nnU-Net) model to improve the accuracy and generalizability of NCCT-based esophageal segmentation, and compared its performance with the widely used generic model Total Segmentator.
A total of 300 patients who underwent chest DLCT examination at the Fourth Affiliated Hospital of Nanjing Medical University were retrospectively enrolled. The manual esophageal segmentation gold standard was established by fusing virtual non-contrast (VNC) and contrast-enhanced CT images. The 3D nnU-Net model was trained on VNC images, with true non-contrast (TNC) images serving as the internal validation set. Two independent external test sets were included: 100 NCCT cases from Zhongda Hospital Affiliated to Southeast University, and 48 NCCT cases from the National Center for Bioinformation. Segmentation performance of the two models was compared via quantitative metrics (Dice similarity coefficient, DSC; 95% Hausdorff distance, HD95; intraclass correlation coefficient, ICC) and qualitative 5-point Likert subjective scoring.
Across the internal validation set and two external test sets, the proposed 3D nnU-Net model achieved excellent core segmentation performance (DSC: 0.756-0.798, HD95: 4.00-4.10 mm), and significantly outperformed the Total Segmentator model (DSC: 0.687-0.697, HD95: 5.00~5.00 mm; Z=3.875-13.045, all P < 0.001). The proposed model exhibited stronger agreement with the manual gold standard (Spearman rs=0.821-0.887, ICC=0.84-0.91), markedly higher than the control model (Spearman rs=0.626-0.716, ICC=0.64-0.70). Bland-Altman analysis confirmed the 3D nnU-Net model had narrower 95% limits of agreement and smaller measurement bias, with volumetric deviation within the clinically acceptable range. Meanwhile, the proposed model yielded a stable median subjective segmentation quality score of 5.00, significantly higher than that of the control model (P < 0.001).
The DLCT annotation-guided 3D nnU-Net model achieves superior esophageal segmentation performance and generalization capability, providing a novel technical strategy for automated NCCT-based esophageal segmentation.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Comments on this article