Repurposing a commercial artificial intelligence vision platform for automated surgical instrument identification: a proof-of-concept study.

Abstract

Background: Retained surgical items are classified as never events but continue to occur, and the count meant to prevent them is often recorded as correct even when an item is retained. Computer vision could provide an independent check on what is on the instrument table, but purpose-built surgical systems are scarce. Industrial vision platforms offer a faster route, although their surgical performance is unknown.

Methods: In a prospective experimental study, a commercial artificial intelligence vision platform developed for industrial inspection was applied to 19 surgical instruments and 3 suture items, each presented once in 100 runs (2,200 experiments). No software redesign was required, but the classifier had been trained on this item set, making this a supervised rather than a zero-shot setting. Two outcomes were recorded: whether the platform assigned the exact item identity, and whether it registered an item as present at the packing step irrespective of type, termed count-presence detection. Because experiments are clustered within items and runs, the primary estimate is the mean per-item accuracy with cluster-adjusted intervals.

Results: The correct instruments were identified in 2,036 of 2,200 experiments (92.5%). The mean per-item accuracy was 92.5% (95% confidence interval 88.4 to 96.7; cluster-adjusted 87.7 to 95.6). Count-presence detection was 99.1% (98.7 to 99.4). Performance was reproducible across runs (intraclass correlation 0.02), whereas variance clustered by item (0.12). Accuracy ranged from 64% (suture needle 6-0) to 100% (six items). Errors were directional and almost always within visually near-identical sub-types.

Conclusions: An industrial vision platform detected item presence with high reliability and identified item type with high accuracy, however, with a residual error rate of 7.5%. The errors were predictable and confined to visually similar classes, pointing to correctable mitigations rather than a fundamental limit. These figures represent an upper bound under highly controlled conditions rather than expected intraoperative performance. Such a system is intended to support rather than replace the perioperative team. This study establishes feasibility rather than clinical benefit; a prospective intraoperative study is the necessary next step.