Efficient AI model segments thyroid nodules on CT with 78.6% Dice score

A channel-attention high-resolution network segmented thyroid nodules on CT with a Dice coefficient of 78.6% overall, reaching 85.5% for benign nodules versus 66.2% for malignant ones, while reducing computational complexity by 73% compared to standard HRNet.

• In an internal test set of 100 patients, CA-HRNet + test-time augmentation achieved a Dice of 78.6% and an intersection over union of 70.0%.
• Segmentation performance was substantially lower for malignant nodules (Dice 66.2%) than for benign nodules (Dice 85.5%), attributed to irregular margins and less distinct boundaries of malignant lesions.
• The study was a single-center retrospective analysis without external or prospective validation.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.