Artificial Intelligence for Segmentation and Classification of Retinal Hemorrhages in Fundus Photos

Abstract

Purpose : To develop a deep learning model using artificial intelligence (AI) for accurately identifying and segmenting retinal hemorrhages in fundus images. Methods : A dataset of 365 standard fundus photos from children and adults with retinal hemorrhages (RH) was collected from multiple institutions worldwide. Four medical students were trained to evaluate RH on fundus photos and then labeled 345 of the images selected at random. For evaluation, four ophthalmologists labeled the remaining 20 images. The labels were created using ImageJ software and included three mask layers for preretinal, intraretinal, and subretinal hemorrhages. The machine learning model UNet, a convolutional network model for image segmentation and classification, was trained on the 345 images labeled by medical students. The model performance was evaluated on the 20 physician-labeled images using the dice score coefficient, which measures the similarity between two images by counting the number of matching pixels and dividing by the total number of pixels. Results : The AI achieved an average dice score of 79.4% for detecting the location of the RH compared to the physician labels, with the physicians having an average dice score of 78.5% when compared pairwise. The physicians had an average dice score of 72.5%, and the AI score was 56.4% for preretinal hemorrhages. For intraretinal hemorrhages, the average dice scores were 51.7% and 12.5% for the physicians and the AI, respectively. For subretinal hemorrhages, the physicians had a score of 4.4%, and the AI had a score of 4.8%. Conclusions : Deep learning models can predict the location of RH on fundus photos with good agreeability with physician experts; however, more work is needed to predict the correct hemorrhage classification and improve inter-rater agreement. A larger, highquality dataset is needed for training and testing with an increase in physician numbers to help address limitations in the classification and inter-rater variability of intraretinal and subretinal hemorrhages. Future work will focus on improving the model's accuracy and addressing these limitations to enhance its utility in clinical settings.