A novel automated parathyroid glands detection and segmentation method in thyroidectomy

Area of Science:

• * Medical imaging and artificial intelligence.
• * Surgical technology advancements.

Background:

• * Preserving parathyroid glands (PGs) during thyroidectomy is crucial but challenging.
• * Deep learning (DL) shows promise for enhancing intraoperative decision-making.

Purpose of the Study:

• * To introduce PG-AI, a novel DL-based system for intraoperative parathyroid gland identification.
• * To evaluate the performance of PG-AI compared to existing methods and surgical expertise.

Main Methods:

• * Development of PG-AI, integrating a YOLOX localization subnet and a Trans-U-HRNet segmentation model.
• * Training and validation on 976 intraoperative images from 121 patients undergoing open thyroidectomy.
• * Comparative analysis with near-infrared autofluorescence (NIRAF) imaging and surgeon assessments.

Main Results:

• * PG-AI achieved 91.1% accuracy and 86.5% recall on internal validation.
• * Visualization recognition rates were 88.7% (internal) and 85.0% (external).
• * PG-AI demonstrated superior performance compared to junior surgeons and showed 72.1% agreement with NIRAF.

Conclusions:

• * PG-AI provides accurate, real-time segmentation masks for PGs during surgery.
• * The system offers reliable intraoperative visual guidance, aiding surgeons in PG identification and preservation.