Surgical instrument segmentation and classification in transcanal endoscopic ear surgery video using Segment Anything...

Area of Science:

• Artificial Intelligence in Medicine
• Surgical Technology
• Medical Imaging Analysis

Background:

• Accurate recognition of surgical instruments is crucial for enhancing safety and efficiency in transcanal endoscopic ear surgery (TEES).
• The Segment Anything Model 2 (SAM 2) offers advanced zero-shot segmentation capabilities, enabling object identification without extensive training data.
• Evaluating AI models like SAM 2 for surgical instrument recognition in TEES is essential for advancing clinical applications.

Purpose of the Study:

• To be the first to apply SAM 2 to TEES videos.
• To evaluate SAM 2's performance in segmenting and classifying surgical instruments within TEES procedures.
• To assess the feasibility of using zero-shot segmentation for surgical instrument identification.

Main Methods:

• A 684-frame evaluation video was created from three clinical TEES (tympanoplasty) cases.
• Four surgical instruments (cupped forceps, pick, alligator forceps, circular knife) were simulated.
• SAM 2 was tasked with instrument identification using only 12 reference images, with performance measured by Precision, Recall, DSC, and IoU.

Main Results:

• SAM 2 achieved excellent segmentation performance (mean DSC: 0.93, mean IoU: 0.87) without classification.
• Instrument classification performance varied (mean DSC: 0.62-0.89, mean IoU: 0.58-0.84), with lower accuracy for articulating instruments.
• False-positive detections were minimal, indicating high specificity.

Conclusions:

• SAM 2 accurately identifies surgical instruments in TEES videos without large training datasets.
• Zero-shot segmentation significantly lowers the barrier for AI clinical implementation in surgery.
• Future work should focus on improving classification accuracy for articulating instruments and enhancing processing speed for real-time application.