How computer vision is learning surgical instrument counting
A computer vision model achieved a 72.7% success rate when detecting and counting surgical tools in simulated operation videos. While promising for patient safety, it struggled to spot tiny items like needles.
Reading level
The full story with proper science words explained.
A missing pair of forceps
Nobody wants a surgical tool left behind inside a patient. Teams manually tally every item before and after an operation, but human eyes get tired. Could an algorithm help?
To explore this, scientists turned to computer vision for automatic surgical instrument counting. They wanted to see if software could watch video footage and reliably keep score as tools moved in and out of view.
Teaching an algorithm to spot steel
The team trained a convolutional neural network known as YOLO (You Only Look Once). This fast system scans whole video frames in a single glance to locate and classify objects. To teach it, researchers used expert annotation, where specialists manually marked tools across hundreds of images.
Their training library included 500 pictures of six tool types, 500 images pulled from online operations, and 100 minutes of simulated surgery showing 600 tool transfers. Eighty per cent of this footage trained the software, leaving the remaining fifth to test it.
Instead of watching the whole room, the model focused on a region of interest: a set zone on screen where instruments change hands. Across testing, it achieved a 72.7% success rate at identifying and counting items crossing that zone.
What we still do not know
The algorithm was far from perfect. It easily tracked bulkier instruments, but performed far less effectively on slender items such as needles and small hemostats, which are clamps used to control bleeding. In the video, tiny tools simply faded into the background.
Most importantly, this setup was tested only on staged footage in a simulation lab. We do not yet know how the system will cope with the unpredictable lighting, rapid movements, and messy workflows of an actual hospital operating theatre. Researchers still need new ways to catch those elusive tiny tools before machine eyes can assist real operations.
Science words
- Convolutional neural network
- A type of artificial intelligence specifically designed to analyse visual images.
- YOLO
- You Only Look Once; a rapid algorithm that locates and identifies objects in an image in a single pass.
- Annotation
- The process where human experts label images with bounding boxes and names to train an AI.
- Region of interest
- A chosen sub-area within a camera frame selected for specific analysis.
- Hemostat
- A surgical clamp used to compress blood vessels and control bleeding.
Check it yourself
This story is based on a real research paper in Surgical Innovation by Zhu, Wu, Lai et al. We write with AI help and check it against the paper, but the original is the final word.