Giap Binh Duong, Ballouz Dena, Srinivasan Karthik, Lustre Jefferson, Likosky Keely, Mahmoud Ossama, Mian Shahzad I, Tannen Bradford L, Nallasamy Nambi
Department of Ophthalmology & Visual Sciences, Kellogg Eye Center, University of Michigan, Ann Arbor, Michigan.
Department of Vitreo Retinal, Aravind Eye Hospital, Chennai, Tamil Nadu, India.
Ophthalmol Sci. 2025 Mar 14;5(4):100764. doi: 10.1016/j.xops.2025.100764. eCollection 2025 Jul-Aug.
To develop and validate a novel artificial intelligence (AI)-powered video analysis system to assess surgeon proficiency in maintaining (1) eye neutrality, (2) eye centration, and (3) adequate focus of the operating microscope in cataract surgery and evaluate differences in these metrics between attending cataract surgeons and ophthalmology residents.
A retrospective surgical video analysis.
Six hundred twenty complete surgical video recordings of 620 cataract surgeries performed by either attending surgeons or ophthalmology residents.
Performance of the proposed AI-powered video analysis system (CatSkill) for cataract surgery was evaluated at multiple stages. Anatomy and surgical landmark segmentation were reported as Dice coefficients. The proposed cataract surgery assessment metrics (CSAMs) were compared between attending and resident surgeons on a phase-wise basis. Surgery-level classification performance (attending vs. resident) of a machine learning (ML) algorithm trained on the CSAMs was assessed using area under the receiver operating characteristic curve (AUC).
An automated system involving video preprocessing, deep learning-based segmentation with limbus obstruction detection and compensation, and CSAM computation was designed to assess surgeon performance based on surgical videos. Three CSAMs were computed to analyze 430 cataract surgeries (254 attendings and 176 residents). An ML algorithm was developed to predict surgeon training level using only CSAMs.
The CatSkill system using FPN (VGG16) achieved a Dice coefficient of 94.03% for segmentation of palpebral fissure, limbus, and Purkinje image 1. The phase-wise mean CSAM scores were higher for attendings than residents across all surgical phases. Residents struggled with stability/centration during the Main Wound, Cortical Removal, Lens Insertion, and Wound Closure phases, and had difficulty maintaining adequate microscope focus during later phases of surgery. A random forest model using CSAMs achieved an AUC of 0.865 in predicting the skill level (attending or resident) of the surgeon.
The proposed AI-derived CSAMs provide a high level of reliability in assessing the ability of surgeons to maintain eye neutrality, centration, and focus level during cataract surgery. Furthermore, downstream analysis using an ML model for surgical-level classification indicates that the proposed CSAMs provide significant predictive value for assessing the overall training level of the surgeon.
Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
开发并验证一种新型的人工智能（AI）驱动的视频分析系统，以评估外科医生在白内障手术中维持（1）眼球中立性、（2）眼球中心定位以及（3）手术显微镜足够聚焦的熟练程度，并评估白内障主治外科医生和眼科住院医师在这些指标上的差异。
一项回顾性手术视频分析。
由主治外科医生或眼科住院医师进行的620例白内障手术的620份完整手术视频记录。
在多个阶段评估所提出的用于白内障手术的AI驱动视频分析系统（CatSkill）的性能。解剖结构和手术标志分割以Dice系数报告。在各个阶段对主治医生和住院医生之间所提出的白内障手术评估指标（CSAM）进行比较。使用在CSAM上训练的机器学习（ML）算法的手术水平分类性能（主治医生与住院医生）通过受试者操作特征曲线（AUC）下的面积进行评估。
设计了一个自动化系统，包括视频预处理、基于深度学习的带有角膜缘阻塞检测和补偿的分割以及CSAM计算，以根据手术视频评估外科医生的表现。计算三个CSAM以分析430例白内障手术（254例主治医生的手术和176例住院医生的手术）。开发了一种ML算法，仅使用CSAM来预测外科医生的培训水平。
使用FPN（VGG16）的CatSkill系统在睑裂、角膜缘和浦肯野图像1的分割中实现了94.03%的Dice系数。在所有手术阶段，主治医生的各阶段平均CSAM得分均高于住院医生。住院医生在主切口、皮质切除、晶状体植入和伤口闭合阶段在稳定性/中心定位方面存在困难，并且在手术后期难以维持手术显微镜的足够聚焦。使用CSAM的随机森林模型在预测外科医生的技能水平（主治医生或住院医生）方面的AUC为0.865。
所提出的源自AI的CSAM在评估外科医生在白内障手术中维持眼球中立性、中心定位和聚焦水平的能力方面具有高度可靠性。此外，使用ML模型进行手术水平分类的下游分析表明，所提出的CSAM在评估外科医生的整体培训水平方面具有显著的预测价值。
在本文末尾的脚注和披露中可能会找到专有或商业披露信息。
