Retinograd-AI: An Open-source Automated Fundus Autofluorescence Retinal Image Gradability Assessment for Inherited Retinal Dystrophies

Abstract

PurposeTo develop an automated system for assessing the quality of Fundus Autofluorescence (FAF) images in patients with inherited retinal diseases (IRD).MethodsWe annotated a dataset of 2445 FAF images from patients with Inherited Retinal Dystrophies which were assessed by three different expert graders. Graders marked images as either gradable (acceptable quality) or ungradable (poor quality), following a strict grading protocol. This dataset was used to train a Convolutional Neural Network (CNN) classification model to predict the gradability label of FAF images.ResultsRetinograd-AI achieves a performance of 91% accuracy on our held-out dataset of 133 images with an Area Under the Receiver Operator Characteristic (AUROC) of 0.94, indicating high performance in distinguishing between gradable and ungradable images. Applying Retinograd-AI to our full internal dataset, the highest proportion of gradable images was found in the 30-50 years age group, where 84.3% of images were rated as gradable, while the lowest was in 0-15 year olds, where only 45.2% of images were rated as gradable. 83.4% of images from male patients were rated as gradable, and 90.6% of images from female patients. By genotype, from the 30 most common genetic diagnoses, the highest proportion of gradable images was in patients with disease causing variants inPRPH2(93.9%), while the lowest wasRDH12(28.6%). Eye2Gene single-image gene classification top-5 accuracy on images rated by Retinograd-AI was 69.2%, while top-5 accuracy on images rated as ungradable was 39.0%. Retinograd-AI is open-sourced, and the source code and network weights are available under an MIT licence on GitHub athttps://github.com/Eye2Gene/retinograd-aiConclusionsRetinograd-AI is the first open-source AI model for automated retinal image quality assessment of FAF images in IRDs. Automated gradability assessment through Retinograd AI enables large scale analysis of retinal images, which is an essential part of developing good analysis pipelines, and real-time quality assessment, which is essential for deployment of AI algorithms, such as Eye2Gene, into clinical settings. Due to the diverse nature of IRD pathologies, Retinograd-AI may also be applicable to FAF imaging for other conditions, either in its current form or through transfer learning and fine-tuning.