# Applying machine learning to safe vascular anastomosis

## TL;DR

This paper explores using machine learning on exoscope images to detect signs of unsafe blood vessel connections during surgery, aiming to improve surgical safety and training.

## Contribution

The study demonstrates the feasibility of using exoscope images and YOLO for detecting thrombus-predicting signs in microsurgery.

## Key findings

- Exoscope images were successfully used to train a YOLO model to detect thrombus-predicting signs.
- The trained algorithm detected four objects in real time but had high false-positive and false-negative rates.
- The study highlights how accessible tools like Python and Google Colaboratory enable non-experts to develop machine-learning models for surgery.

## Abstract

Machine-learning technology is currently being introduced into the medical field and has been shown to aid diagnostic imaging, patient examinations, patient-data analysis, various surgical aspects, and medical education. Recent advances in exoscopes and monitors are prompting a shift from optical microscope-based microsurgery to heads-up microsurgery. The high-definition exoscope images are highly suitable for machine learning. Since an algorithm that detects predictive signs of thrombus formation would aid microsurgery and help train surgeons to identify vessels at risk of unsafe microvascular anastomosis, we here asked whether we could use exoscope images to train such a machine-learning algorithm.

Arterial clots, intimal-wall damage, debris, and stumps in 9150 ORBEYE™ exoscope images of arterial anastomosis obtained in 2023–2024 were annotated with RectLabel pro™. These images were used to train the You Only Look Once (YOLO) model (Ultralytics) to detect the thrombus-predicting signs. The YOLO code was executed within Google Colaboratory™.

After algorithm training for 100 epochs, the four objects were detected in real time, albeit with high levels of false-positive and false-negative detections.

Our study shows the potential of machine learning on exoscope images to generate algorithms that promote safe microsurgical anastomosis. It also shows how the recent emergence of Python code, Google Colaboratory™, and machine-learning models such as YOLO has made it possible for even programming amateurs to develop effective machine-learning algorithms. Further development of new central and graphics processing units and computational processing methods will likely lead to machine-learning applications that improve surgery and facilitate medical training.