ABS 023: AI-powered Surgical Phase Classification and Segmentation System for DMEK Surgical Analysis

Christian Reinhardt ¹ , Joshua Ong MD ¹ , Jefferson Lustre ¹ , Chanon Thanitcul MD ¹ , Binh D. Giap PhD ¹ , Hamza Khan ² , Keely Likosky ¹ , Ossama Mahmoud ² , Shahzad I. Mian MD ¹ , Bradford Tannen MD, JD, MBA ¹ , Nambi Nallasamy MD ¹ ³

¹ Department of Ophthalmology and Visual Sciences, Kellogg Eye Center, University of Michigan, Ann Arbor, Michigan
² School of Medicine, Wayne State University, Detroit, Michigan
³ Department of Computational Medicine and Bioinformatics, University of Michigan, Ann Arbor, Michigan

Van Wickle (2025) Volume 1, ABS 023

Introduction: Introduction: Descemet Membrane Endothelial Keratoplasty (DMEK) is a specialized form of partial-thickness corneal transplant surgery for corneal endothelial dysfunction. This technique offers the potential to improve vision for patients with corneal edema. As a relatively new technique that is becoming essential to know for cornea specialists, training surgeons is of utmost importance. However, there exist few objective measures of surgeon performance for DMEK. In this study, we developed and validated machine learning-based surgical phase segmentation model to analyze DMEK videos to automate the segmentation of DMEK surgeries into their component phases.
Methods: We developed a dataset consisting of 77 DMEK videos. The dataset was split into training (54 videos), validation (12 videos), and testing (11 videos) subsets. Each frame of each video was labeled as 1 of 18 active surgical phases or “No Activity” wherein no active surgical maneuvers were performed, resulting in more than 2,808,000 annotated video frames . A 2D Densenet model was trained using the training set and refined using the validation set. Performance metrics were computed on the testing set to assess frame-level phase classification and phase-level boundary identification accuracy.
Results: The model performed at 0.97 Classification Accuracy, 0.62 Precision, 0.53 Recall, 0.76 F1-score, and 0.72 Average ROC-AUC.
Conclusion: DMEK offers the potential to improve vision for patients with corneal endothelial dysfunction. Automated segmentation of DMEK surgical video recordings, as demonstrated here, can aid in the analysis of surgeon performance and in providing focused feedback to trainee surgeons. Improving trainee surgeon comfort with DMEK can help further expand access to this important surgical technique.
Methods: We developed a dataset consisting of 77 DMEK videos. The dataset was split into training (54 videos), validation (12 videos), and testing (11 videos) subsets. Each frame of each video was labeled as 1 of 18 active surgical phases or “No Activity” wherein no active surgical maneuvers were performed, resulting in more than 2,808,000 annotated video frames . A 2D Densenet model was trained using the training set and refined using the validation set. Performance metrics were computed on the testing set to assess frame-level phase classification and phase-level boundary identification accuracy.

Results: The model performed at 0.97 Classification Accuracy, 0.62 Precision, 0.53 Recall, 0.76 F1-score, and 0.72 Average ROC-AUC.

Discussion: DMEK offers the potential to improve vision for patients with corneal endothelial dysfunction. Automated segmentation of DMEK surgical video recordings, as demonstrated here, can aid in the analysis of surgeon performance and in providing focused feedback to trainee surgeons. Improving trainee surgeon comfort with DMEK can help further expand access to this important surgical technique.