Automated Deep Learning and CT Elastography Radiomics Predict RLN Lymph Node Metastasis in ESCC

Overview

This study developed an automated workflow combining nnU-Net deep learning segmentation with CT elastography-derived radiomic features to predict lymph node metastasis near the right recurrent laryngeal nerve in esophageal squamous cell carcinoma. The model demonstrated high accuracy, with entropy and fractal dimension features showing strong diagnostic performance and clinical utility.

Background

Esophageal squamous cell carcinoma (ESCC) frequently metastasizes to lymph nodes adjacent to the right recurrent laryngeal nerve (RLN), significantly impacting prognosis and treatment decisions. Conventional CT imaging relies on size and morphology criteria, which often miss micrometastases, limiting sensitivity and specificity. Radiomics offers quantitative imaging biomarkers but is hindered by manual segmentation variability and neglect of tissue biomechanical properties. Integrating automated segmentation with CT elastography-derived features may improve preoperative metastasis prediction and personalized treatment planning.

Data Highlights

Key Findings

• The nnU-Net model achieved high segmentation accuracy with a Dice coefficient of 0.898 ± 0.024 for lymph nodes near the right RLN.
• Five DEM-derived radiomic features were selected, including one first-order entropy and four fractal dimension-related features.
• The entropy feature showed the highest diagnostic performance with an AUC of 0.814, sensitivity of 0.895, and specificity of 0.709.
• Fractal dimension features were significantly elevated in metastatic lymph nodes (P < 0.001), indicating increased textural complexity.
• Calibration curves confirmed robust probability estimation based on entropy features.
• Decision curve analysis demonstrated positive net clinical benefits across a wide range of threshold probabilities (10%–70%).

Clinical Implications

The automated segmentation combined with CT elastography radiomics provides a non-invasive, reproducible method to accurately predict right RLN lymph node metastasis preoperatively in ESCC patients. This approach may enhance clinical decision-making by identifying patients at higher metastatic risk, facilitating personalized treatment strategies and potentially improving outcomes. Adoption of such automated workflows could reduce operator dependency and improve diagnostic consistency.

Conclusion

Integrating deep learning-based automatic segmentation with CT elastography-derived radiomic features enables accurate and clinically useful prediction of right RLN lymph node metastasis in esophageal cancer. This method offers a promising tool for personalized preoperative assessment and treatment planning.

References

1. Weng et al. 2021 -- Combining YOLO Models with Hyperspectral Imaging for EC Detection
2. Chang et al. 2020 -- Spectral Imaging Technology to Improve Early EC Detection
3. Chen et al. 2019 -- Optimizing Band Selection for Spectrum-Aided Visual Enhancer