Customised artificial intelligence toolbox for detecting diabetic retinopathy with confocal truecolor fundus images using object detection methods

Abstract

Purpose: A novel convolutional neural network approach in detecting diabetic retinopathy (DR) was employed to overcome the black box dilemma in artificial intelligence (AI). In addition to identification and classification, this tool is intended to identify signs such as microaneurysms, hard exudates, dot-blot haemorrhages and flame-shaped haemorrhages, and neovascularisation with the help of customised human annotations. Design: This is a prospective cross-sectional study. Subjects: Eight thousand confocal high-resolution fundus images of 5,174 patients were included in this study. Methods: Dataset involved 8,000 fundus images of DR with 5,200 images for training, 1,400 images for validation and 1,400 images for the held-out test. The 1,400 images used for the held-out test were non-annotated fundus images. You Only Look Once (YOLO) 5 algorithms were used for detection. Main Outcome Measures: The AI tool was evaluated with mean average precision, objectness loss, classification loss, precision and recall. The number of images in which all the clinical signs of DR were correctly predicted, wrongly predicted and missed were also calculated. Results: Tests showed consistent increments from 79.5% to 91% accuracy in predicting the diagnosis, severity, and clinical fundus signs pertaining to DR. The overall sensitivity was 81.6% and the specificity was 100%. Conclusion: To our knowledge, this is the first paper to train fundus images with high-resolution confocal images and annotate every clinical sign of the DR fundus along with diagnosis and severity for accurate predictions with their various fundus signs, thus overcoming the black box dilemma. With constant training via a feedback mechanism, there was a continuous upsurge in prediction accuracy.