An image-based instrument recognition system completed simulated postoperative counting tasks faster than manual counting, with identical results under controlled conditions. A randomised crossover evaluation at a regional hospital in northern Taiwan involved 34 operating theatre nurses. Published in Health Informatics Journal, the evaluation compared counting time, recognition results and participant questionnaire scores for an artificial intelligence-based instrument recognition and inventory system, known as AI-IRIS, and conventional manual counting. The scenarios used fixed lighting, non-overlapping orthopaedic instruments and no deliberate discrepancies, limiting conclusions about error detection or performance in routine operating theatre conditions.
Controlled Comparison Shows Matching Results
The image-based system combines a camera, an edge-computing platform and a dashboard that displays detected instruments against an expected inventory. Tray images are processed locally by a recognition model trained on 30 orthopaedic instruments, including haemostats, scalpels and needle holders. Matching results produce a confirmation, while missing or mismatched items generate a visual alert for clinical verification. The system does not require physical tags and operated offline during testing.
Each nurse completed two standardised postoperative counting scenarios using both the image-based system and manual counting. The order was randomised and separated by a five-minute washout period. Both methods were demonstrated before testing and participants practised the tasks. Hospital orthopaedic instrument sets were arranged in standard postoperative tray layouts. Camera position and lighting remained fixed, while manual counting used a standardised checklist.
Both methods produced the same recognition and quantity results. Each identified and counted 30 instruments in the first scenario and 27 in the second. The scenarios did not include missing, additional or incorrectly placed instruments. They also avoided overlap and other challenging visual conditions. The identical results therefore show that both methods completed the defined verification tasks correctly, but not that the image-based system detects discrepancies better.
The evaluation excluded the search procedures that would normally follow a count discrepancy, including checks of trays, drapes, linens, waste and suction containers. Its findings apply only to the initial counting tasks used in the simulation.
Counting Is Faster, but Wider Effects Remain Untested
The image-based system completed the first scenario in an average of 51.5 seconds, compared with 79.4 seconds for manual counting. In the second scenario, average times were 47.6 seconds and 74.0 seconds respectively. This represented a reduction of around one-third in the time required for the controlled tasks.
Participating nurses most often identified unfamiliarity among newer staff with instrument names, characteristics and quantities as a cause of counting errors. Other selected factors included the absence of detailed inventory lists and the use of lists that were unclear or not regularly updated. These responses indicate areas in which manual verification can depend on staff familiarity and supporting documentation.
Must Read: Voice AI Supports Pre-Procedural Cardiac Calls
The image-based system provides standardised visual identification and compares detected instruments with a predefined digital inventory. During the simulation, this reduced the time required to complete the specified counts. However, the evaluation did not measure workload, cognitive burden, operating theatre throughput, discrepancy resolution or retained surgical items. It also did not test whether the system reduces counting errors when instruments are missing, obscured or mixed across sets.
The reported time savings should therefore be interpreted narrowly. They reflect two short, standardised scenarios involving one group of orthopaedic instruments under stable conditions. No evidence was produced on cumulative time savings across complete surgical workflows or on the effect of the system during preoperative or intraoperative counting.
Questionnaire Scores Favour the Automated Method
Participants rated both methods through a 16-item questionnaire covering perceived advantage, compatibility, ease of use, trialability and visibility of benefits. The image-based system received a higher total questionnaire score than manual counting. Significant differences were reported for perceived work efficiency, time needed for inventory, simplicity of identification and willingness to try the method.
Ratings for workflow integration, ease of learning and alignment with safety policies also tended to favour the automated method, although not every item differed significantly. These responses reflect participant perceptions after brief demonstrations and simulated use. They do not establish long-term acceptance, sustained use or successful implementation in routine practice.
Several limitations restrict the findings. The evaluation involved a small, all-female sample from one hospital and only 30 orthopaedic instruments. The model’s performance remains unknown for larger or more complex sets, other surgical specialties and conditions involving overlap, occlusion, variable lighting or blood contamination. Participants were trained before testing and each session lasted around 15 to 20 minutes. The evaluated system is covered by a Taiwanese patent and was developed by three of the investigators involved in the evaluation. The work was funded by Taiwan Adventist Hospital through an industry–academia collaboration project.
In controlled simulations, the image-based system completed postoperative instrument verification faster than manual counting and produced the same counting results. Participants also gave it higher questionnaire scores on several measures. Because the scenarios contained no deliberate discrepancies and used fixed lighting and non-overlapping instruments, the evaluation does not demonstrate better discrepancy detection, improved patient safety or reduced workload. Further assessment is needed in routine operating theatres, using varied instrument sets and more challenging visual conditions, before conclusions can be drawn about clinical performance, implementation or broader operational effects.
Source: Health Informatics Journal
Image Credit: iStock
References:
Lin ZC, Hung HC, Koo M et al. (2026) Integrating artificial intelligence into surgical nursing workflows: Evaluation of an image-based Instrument verification system. Health Informatics Journal;32(3). https://doi.org/10.1177/14604582261468225
