ESCRS - P016 - TO EVALUATE A NOVEL DEEP LEARNING ALGORITHM-ARTIFICIAL INTELLIGENCE (DLA-AI) MODEL TO DISTINGUISH BETWEEN EYES THAT MAY/MAY NOT HAVE A GRAFT DETACHMENT BASED ON PRE-DMEK SURGERY OPTICAL COHERENCE TOMOGRAPHY (OCT) IMAGES.

TO EVALUATE A NOVEL DEEP LEARNING ALGORITHM-ARTIFICIAL INTELLIGENCE (DLA-AI) MODEL TO DISTINGUISH BETWEEN EYES THAT MAY/MAY NOT HAVE A GRAFT DETACHMENT BASED ON PRE-DMEK SURGERY OPTICAL COHERENCE TOMOGRAPHY (OCT) IMAGES.

Published 2023 - 27th ESCRS Winter Meeting

Reference: P016 | Type: Presented Poster & Poster Submission | DOI: 10.82333/jxw4-5j14

Authors: Alastair Simon Patefield* 1 , Yanda Meng 1 , Giulia Coco 2 , Sabrina Vaccaro 3 , Mohit Parekh 4 , Matteo Airaldi 5 , Stephen Kaye 1 , Yalin Zheng 6 , Vito Romano 7

1Department of Eye and Vision Sciences,Institute of Life Course and Medical Sciences, University of Liverpool,Liverpool,United Kingdom, 2Department of Clinical Science and Translational Medicine,University of Rome Tor Vergata,Rome,Italy, 3Department of Corneal Diseases,St. Paul's Eye Unit, Royal Liverpool University Hospital,Liverpool,United Kingdom, 4Schepens Eye Research Institute, Massachusetts Eye and Ear, Department of Ophthalmology,Harvard Medical School,Boston,United States, 5Department of Biomedical and Clinical Science Luigi Sacco,University of Milan,Milan,Italy, 6Liverpool Centre for Cardiovascular Science,University of Liverpool and Liverpool Heart and Chest Hospital,Liverpool,United Kingdom, 7Department of Eye and Vision Sciences ,Institute of Life Course and Medical Sciences, University of Liverpool,Liverpool,United Kingdom

Purpose

To evaluate a novel deep learning algorithm-artificial intelligence (DLA-AI) model to distinguish between eyes that may/may not have a graft detachment based on pre-DMEK surgery optical coherence tomography (OCT) images.

Setting

Department of Eye and Vision Sciences, Institute of Life Course and Medical Sciences, University of Liverpool, Liverpool, UK.

Department of Corneal Diseases, St. Paul's Eye Unit, Royal Liverpool University Hospital, Liverpool, UK.

 

Methods

Retrospective cohort study performed on images taken by an anterior segment OCT (AS-OCT) scanner. A multiple instance learning AI (MIL-AI) model using a Resnet-101 backbone was designed and applied to pre-DMEK surgery OCT images. The images were split into training and testing sets. The MIL-AI model was trained and validated on the training set and then its performance and heat maps were calculated from the testing set. F1 score, sensitivity, specificity and area under curve were used as performance metrics.

Results

9,466 images of 74 eyes (128 images per eye) were included in the study. Images from 50 eyes were used to train and validate the system, then the images of the remaining 24 eyes were used to determine the performance of the model whilst heat maps were generated for visualization. The best MIL model had a Resnet-101 model as the backbone.  The performance metrics on the test set (at a 95% confidence interval [CI]) were as follows: F1 score 0.77 (0.57 - 0.91); precision 0.67 (0.44 - 0.88); specificity 0.45 (0.15 - 0.75); sensitivity 0.92 (0.73 - 1.00) and area under the curve 0.63 (0.52 - 0.86). AI performance was more sensitive (92% vs 31%) but less specific (45% vs 64%) than an ophthalmologist’s performance.

Conclusion

The DLA-AI produces reasonable results on  predicting the eyes that may/may not have graft detachment post-DMEK surgery. A larger scale clinical trial to validate this high sensitivity model is warranted. Implications are opportunity for implementation into DMEK suitability screening.