Thyroid gland delineation in noncontrast-enhanced CTs using deep convolutional neural networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate thyroid gland delineation is crucial for diagnosing and managing thyroid carcinoma.
• Current methods may lack efficiency and robustness, especially in noncontrast-enhanced CT scans.
• Deep learning offers potential for automated and precise medical image segmentation.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) method for accurate, efficient, and robust thyroid gland segmentation in noncontrast-enhanced head and neck CTs.
• To assess the DL method's performance against established segmentation metrics and compare it with other DL models.
• To validate the DL method's accuracy in a large cohort of patients with suspected thyroid carcinoma.

Main Methods:

• Retrospective analysis of 1977 noncontrast-enhanced head and neck CT scans from patients with suspected thyroid carcinoma.
• Development of a DL-based segmentation method for the thyroid gland.
• Evaluation using metrics such as Dice Similarity Coefficient (DSC), Jaccard Index (JAC), sensitivity, specificity, Hausdorff Distance (HD), Mean Surface Distance (MSD), Residual Mean Square Distance (RMSD), and Center of Mass Distance (CMD).
• Cross-validation, hold-out experiments, cancer vs. benign prediction accuracy, and cross-gender analysis were performed.

Main Results:

• The DL method achieved high accuracy, with median DSC > 0.913, JAC > 0.839, sensitivity > 0.856, and specificity > 0.979.
• Excellent agreement with clinical contours was observed, with median MSD < 0.31 mm, RMSD < 0.48 mm, HD < 2.06 mm, and CMD < 0.50 mm.
• The proposed method demonstrated significantly improved performance compared to 3D U-Net and V-Net, with robust results across six cross-sectional tests.

Conclusions:

• The developed DL method provides accurate, efficient, and robust thyroid gland delineation in noncontrast-enhanced head and neck CTs.
• This AI-driven approach shows significant potential for improving diagnostic accuracy and workflow efficiency in thyroid cancer evaluation.
• The method's strong performance across multiple validation tests supports its clinical applicability.