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

Accurate automatic recognition of surgical instruments is important for enhancing the safety and efficiency of transcanal endoscopic ear surgery (TEES). Segment Anything Model 2 (SAM 2) is a state-of-the-art artificial intelligence (AI) model featuring "zero-shot segmentation", which allows it to identify objects without extensive training data. The purpose of this study was to be the first to apply SAM 2 to TEES videos to evaluate its performance in segmenting and classifying surgical instruments.

From videos of three clinical TEES cases (tympanoplasty), we created a 684-frame evaluation video that simulated the sequential use of four types of surgical instruments: cupped forceps, pick, alligator forceps, and circular knife. SAM 2 was instructed to identify the instruments using only 12 reference images. Performance was quantitatively evaluated using four metrics-Precision, Recall, dice similarity coefficient (DSC), and intersection over union (IoU)-as well as a confusion matrix to assess frame-level classification accuracy by comparing the model's output to manually created ground-truth annotations.

For segmentation without classification, SAM 2 demonstrated excellent performance, with a mean Precision of 0.98, a mean Recall of 0.89, a mean DSC of 0.93, and a mean IoU of 0.87. When classifying the four specific instrument types, performance varied, with mean Precision ranging from 0.68 to 0.98, mean Recall from 0.79 to 0.86, mean DSC from 0.62 to 0.89, and mean IoU from 0.58 to 0.84. Notably, false-positive detections in non-instrument regions were extremely rare. A trend of lower classification accuracy was observed for articulating instruments with moving parts (cupped forceps and alligator forceps).

This study demonstrated that SAM 2 can accurately identify surgical instruments in TEES videos without the need to prepare large, costly training datasets. This zero-shot segmentation characteristic significantly lowers the barrier to clinical implementation of AI, suggesting it is a promising foundational technology for future surgical navigation and robotic automation. Key future challenges include improving classification accuracy for articulating instruments and accelerating processing speed for practical, real-time use.