OP-3 The effectiveness of artificial intelligence in annotating and measuring corneal pathology on OCT

Le résumé fourni par la source

*Correspondence, Colby Hart: colbythomashart@gmail.com Objective To determine if corneal OCT images can be characterised and measured using artificial intelligence (AI) and how this compares to manual subjective assessment. Methods Phase one. Casia OCT images of patients with corneal disease were included from Birmingham and Liverpool. Individual images were annotated by expert clinicians after concordance training sessions. Two annotations were made: high and low confidence lesion borders. Images were split into training and testing sets. Training data were used to train a DeepLabV3 deep learning model. Testing sets were used to evaluate performance. Lesions were independently evaluated by three masked experts. Phase two. OCT images from patients with microbial keratitis (MK) on days 0, 3, 7 and 28 were annotated by AI after training on normal corneal OCTs. Nonparametric analysis was undertaken using SPSS v25. Results Phase one. 456 images from patients with primary cornea disease were used to train the AI model and 43 were used for testing the model. Comparing manual and automated annotation, there was a significant difference between expert clinicians (p=0.03, p=0.001) in deciding whether the AI or subjective annotation was a better representation. This may reflect the variety of lesions included. Phase two. Images of 102 patients with MK were selected from days 0, 3, 7 and 28 and subjected to automated annotation. Data analysis on AI annotation of improvement in MK is due March 2022. Conclusion The usefulness of AI for annotating corneal OCT lesions depends on the homogeneity and quality of the image. OCT systems which provide higher resolution images enable better automated annotation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.