563 Automated Assessment of Microsurgical Skill Using Deep Learning and Procedural Video Analysis
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
INTRODUCTION: Technical skill assessment in neurosurgery remains dependent on expert observation, which is resource-intensive, subjective, and lacks scalability. An objective, automated method could provide consistent feedback to trainees while reducing faculty workload. The end-to-side vascular anastomosis, a standardized and reproducible task requiring practice and fine motor control, can serve as a useful proxy for microsurgical proficiency. Recent advances in AI and computer vision offer promising tools for automating such assessments using procedural video analysis. METHODS: Surgical videos were collected from 10 participants representing training levels ranging from medical students to attending neurosurgeons, each performing 5–12 end-to-side anastomosis procedures on artificial vessel models under standardized conditions. All videos were scored independently by three expert neurosurgeons using a 9-domain modified NOMAT scale (1–5 per category). If total scores across raters differed by more than 10 points, or if any individual category score differed by 3 or more points, the discrepancies were reviewed in group sessions and scores were realigned by consensus. The prototype model consists of two components: (1) a kinematic analysis pipeline using the YOLO object detection algorithm to extract motion-based features (velocity and acceleration), and (2) a temporal action recognition model using a custom video transformer trained to detect ten defined microsurgical actions. Model outputs were designed to generate total and subcategory NOMAT scores based on expert-labeled training data. RESULTS: The prototype system successfully detected microsurgical instruments using object detection and extracted motion-based features such as velocity and acceleration. The temporal action recognition model identified discrete surgical actions across predefined categories. Integration of both components enabled the system to process procedural videos and produce preliminary outputs aligned with the structure of NOMAT scoring. Work is ongoing to validate score predictions and assess concordance with expert ratings. CONCLUSIONS: This prototype demonstrates the feasibility of AI-driven, objective assessment of microsurgical skill. Ongoing work will focus on model validation, performance benchmarking, and exploring integration into neurosurgical training pathways.
Fields of Study & Subjects
Surgical Simulation and Training
Innovations in Medical Education
History of Medical Practice
Publication details
- Journal / source
- Neurosurgery
- Volume
- 72
- Issue
- Supplement_1
- Pages
- 169–170
- Publication date
- 26 March 2026
- Language
- EN