Related Experiment Video
Updated: Jun 14, 2025
13:47
Laser Capture Microdissection of Highly Pure Trabecular Meshwork from Mouse Eyes for Gene Expression Analysis
Published on: June 3, 2018
9.3K
Beyond PhacoTrainer: Deep Learning for Enhanced Trabecular Meshwork Detection in MIGS Videos
Su Kara1, Michael Yang1, Hsu-Hang Yeh2
1Department of Ophthalmology, Stanford University, Palo Alto, CA, USA.
Translational Vision Science & Technology
|September 3, 2024
Summary
Deep learning models accurately identify minimally invasive glaucoma surgery (MIGS) and locate the trabecular meshwork (TM) in surgical videos. Transfer learning enhances surgical video analysis for improved training and patient care.
Area of Science:
- Ophthalmology
- Computer Vision
- Medical Imaging
Background:
- Surgical video analysis is crucial for training and quality improvement.
- Existing computer vision models primarily focus on standard cataract surgery.
Purpose of the Study:
- Develop deep learning models for surgical video analysis.
- Identify minimally invasive glaucoma surgery (MIGS) procedures.
- Precisely locate the trabecular meshwork (TM) within surgical videos.
Main Methods:
- Utilized transfer learning to adapt a pre-trained cataract surgery model for MIGS identification.
- Developed and compared U-Net, Y-Net, and Cascaded models for TM localization.
- Employed pixel error and Intersection over Union (IoU) for segmentation accuracy assessment.
Main Results:
- Achieved 87% accuracy and 0.99 AUROC for MIGS frame classification.
- Maintained 79% accuracy for identifying 14 standard cataract surgery steps.
- The U-Net model demonstrated superior TM segmentation with 0.9988 IoU and 1.47 average pixel error.
Conclusions:
- Developed high-performing computer vision models for MIGS recognition and TM localization.
- Transfer learning enables efficient extension of models to new surgical procedures.
- Demonstrated the potential for automated feedback systems in surgical training.
