PURPOSE: With recent technical advancements, the concept of robotic-assisted microsurgery is gaining clinical relevance, holding great potential to revolutionize microsurgery standards. Its successful clinical integration, however, will depend on specific user training and evaluation. This study aimed to develop and verify automated tools for manual and robotic microsurgical performance assessment.
METHODS: Two deep convolutional neural network-based computer algorithms were developed to facilitate computer-assisted 1) microsurgical instrument tracking and 2) anastomosis quality evaluation. First, supervised and semi-supervised learning was used to train models to track both instruments in anastomoses videos. Instrument trajectory was plotted and tremor-movements were measured as absolute deviations from the smooth trajectory. Second, images of everted anastomoses were used to train an algorithm to detect stitches, stitch metrics, and tears while filtering out common nuisance variabilities. Both algorithms were trained and verified using 90 manual and 90 robot-assisted anastomoses, collected in a pre-clinical two-center trial. Robot-assisted anastomoses were performed using the Symani Surgical System with the RoboticScope.
RESULTS: We successfully developed and verified two computer algorithms for manual and robotic microsurgical self-assessment that work without tracking sensors or markers. The instrument tracking algorithm revealed a markedly higher degree of tremor in manually performed anastomoses. The anastomosis quality assessment algorithm reliably evaluates the anastomosis line, inter-stitch distance, stitch angles, and tears.
CONCLUSION:
We established two completely automated algorithms for reliable skill self-assessment applicable to both conventional and robotic microsurgery. These algorithms can increase microsurgical assessment accessibility and accelerate robotic training and evaluation processes, ultimately promoting their clinical integration.
