Predictive value of an artificial intelligence algorithm to identify images with characteristics of exudative lesions.

Abstract

Purpose : Validation of AI (artificial intelligence) algorithms for evaluating OCT (Optical coherence tomography) in ocular diseases. Prove that AI has discriminatory power to separate patients with antiangiogenic therapy criteria and those who do not.

Methods : A retrospective study analysis of secondary sources where OCTs are collected from a database of an ophthalmological clinic. 136 Images were selected by two retina specialists who searched among OCT images between 2019 and 2022; The primary purpose was to determine which patients benefited from antiangiogenic therapy and those who did not. The model had to recognize variables that would identify eyes with fluid secondary to exudative pathologies.

Through transfer learning a deep learning model called convolutional neural networks was trained based on a state-of-the-art architecture. Deep learning algorithm was implemented to extract collected images' characteristics to predict patients that require immediate initiation of antiangiogenic therapy. A retention scheme was used for the model's training and validation stage, and selected image analysis model was pretrained. Using "Arkangel AI" (a tool that automatically performs training, validation, and deployment) tests were performed for different base architectures. For each best hyper-parameter configuration and base architecture, an iterative optimization was performed following binary cross-entropy loss function.

The final model was trained by letting each base architecture configuration and hyperparameters learn for a fixed number of iterations. Each network processed the entire data set four times to search for optimal hyperparameters. Best-performing configuration was then used to train ten times. Binary cross entropy loss function and RMSprop optimizer with a learning rate of 0.01 was used for training the proposed implementation.

Results : The data analysis yields accuracy of 97.9% and coverage of 90.9% . Time of image processing is two seconds

Conclusions : With AI algorithm, subjectivity of interpretation in OCT reading is eliminated, which improves efficiency in the opportunity to start treatment and decreases difficulties in decision-making.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

FIGURE 1 Proposed approach for obtaining a model from images through backbone learning