[Paper Review] An AI Framework for Microanastomosis Motion Assessment

Yan Meng, Eduardo Torres-Rodríguez+6 more|arXiv (Cornell University)|Jan 28, 2026

Surgical Simulation and Training0 citations

TL;DR

The paper presents an end-to-end AI framework for automated assessment of microanastomosis instrument handling using YOLO-based detection, DeepSORT tracking, instrument tip localization, and NOMAT-based skill classification, achieving high detection accuracy and solid agreement with expert ratings.

ABSTRACT

Proficiency in microanastomosis is a fundamental competency across multiple microsurgical disciplines. These procedures demand exceptional precision and refined technical skills, making effective, standardized assessment methods essential. Traditionally, the evaluation of microsurgical techniques has relied heavily on the subjective judgment of expert raters. They are inherently constrained by limitations such as inter-rater variability, lack of standardized evaluation criteria, susceptibility to cognitive bias, and the time-intensive nature of manual review. These shortcomings underscore the urgent need for an objective, reliable, and automated system capable of assessing microsurgical performance with consistency and scalability. To bridge this gap, we propose a novel AI framework for the automated assessment of microanastomosis instrument handling skills. The system integrates four core components: (1) an instrument detection module based on the You Only Look Once (YOLO) architecture; (2) an instrument tracking module developed from Deep Simple Online and Realtime Tracking (DeepSORT); (3) an instrument tip localization module employing shape descriptors; and (4) a supervised classification module trained on expert-labeled data to evaluate instrument handling proficiency. Experimental results demonstrate the effectiveness of the framework, achieving an instrument detection precision of 97%, with a mean Average Precision (mAP) of 96%, measured by Intersection over Union (IoU) thresholds ranging from 50% to 95% (mAP50-95).

Motivation & Objective

• Address the need for objective, scalable microsurgical skill assessment in neurosurgical training.
• Develop an end-to-end AI system for automated instrument handling evaluation in microanastomosis.
• Integrate robust detection, tracking, tip localization, and expert-labeled skill classification.
• Provide real-time, hardware-efficient assessment suitable for training environments.

Proposed method

• Instrument detection using a customized YOLOv11 model with a two-instrument-per-frame and different-types constraint.
• Hybrid instrument tracking that prioritizes YOLO detections over DeepSORT and preserves consistent IDs to improve temporal robustness.
• Instrument tip localization via multi-point shape descriptors matched to a reference descriptor with cosine similarity.
• A supervised classification module trained on expert-labeled NOMAT-based assessments to grade instrument handling proficiency.
• Motion feature extraction from instrument tip trajectories (velocity, acceleration, jerk) with TSFresh-based scalable feature representation and gradient boosting for skill classification.

Experimental results

Research questions

• RQ1 How accurately can instrument detection be performed in microanastomosis videos using a YOLO-based detector augmented for microsurgical instrument variations?
• RQ2 Can integrated tracking and tip localization provide reliable, continuous kinematic data for objective skill assessment?
• RQ3 How well does a supervised model trained on NOMAT labels reproduce expert assessments across skill categories?
• RQ4 What is the real-time computational performance of the proposed framework in training environments?

Key findings

• Instrument detection precision overall: 0.969 with mAP50 0.987 and mAP50–95 0.906.
• Average real-time processing speed: 29.7 frames per second.
• Overall skill-classification accuracy: 0.87 across Poor, Moderate, and Good categories.
• Class-wise skill metrics indicate better performance for Moderate and Good classes than Poor, pointing to annotation variability and dataset size issues.
• Tracking improvements yielded high recovery (98.7%) and correction (90.6%) rates for YOLO misclassifications or misses.
• Instrument tip localization maintained high spatial accuracy across frames, enabling reliable downstream motion analysis.