Related Experiment Video
Updated: May 12, 2026
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Deep learning-based automatic segmentation and CT elastography radiomics for preoperative prediction of right
Chao Ji1, Qingqing Li2, Shumin Jiang1
1Department of Thoracic Surgery, Zhongshan Hospital Xiamen University, School of Medicine, Xiamen University, Xiamen, China.
Objective:
Accurate preoperative prediction of lymph node metastasis adjacent to the right recurrent laryngeal nerve (RLN) in esophageal squamous cell carcinoma (ESCC) is crucial for treatment planning. This study aimed to develop and validate an imaging approach that integrates deep learning-based automatic segmentation (nnU-Net) with computed tomography (CT)-derived differential elasticity map (DEM) to predict RLN lymph node metastasis in ESCC.
Methods:
This retrospective study included 415 patients diagnosed with ESCC. An automatic segmentation model was trained using the nnU-Net framework to delineate lymph nodes near the right RLN. Three-dimensional CT elasticity images were generated from segmented CT voxels, from which radiomic features, including first-order entropy and multi-scale fractal dimensions, were extracted. Statistically significant features were selected using statistical tests and area under the curve (AUC) analyses, and their diagnostic efficacy, probability calibration, and clinical decision-making value were further evaluated in a validation cohort.
Results:
The automatic segmentation model achieved a Dice coefficient of 0.898 ± 0.024. Five DEM-derived radiomic features were ultimately selected: one first-order entropy feature (E_original_firstorder_Entropy) and four fractal dimension-related features. The entropy feature exhibited the highest diagnostic performance (AUC = 0.814, sensitivity = 0.895, specificity = 0.709), and fractal dimension-related features were significantly elevated (all P < 0.001) in the metastatic group, indicating increased textural complexity and multi-scale irregularity. Calibration curves demonstrated the robustness of entropy-based probability estimation. Decision curve analysis confirmed the clinical utility of these features, showing positive net benefits across a wide range of threshold probabilities (10%-70%).
Conclusion:
The proposed automated workflow, combining nnU-Net segmentation and DEM-based radiomics, enables accurate and non-invasive prediction of right RLN lymph node metastasis in ESCC. First-order entropy and fractal dimension features offer valuable complementary information beyond conventional radiomics, providing a promising tool for preoperative decision support and personalized treatment planning.
