Artificial Intelligence for Classification of the Cause of Retinal Hemorrhages in Fundus Photos

Abstract

Purpose : To implement deep learning models for classifying the underlying cause of retinal hemorrhages in fundus photos.

Methods : This study included 597 standard fundus photos with retinal hemorrhages (RH) from children and adults and their diagnoses from multiple institutions worldwide. We separated the diagnosis into two groups: medical cases with a diagnosis of diabetic retinopathy, anemic retinopathy, coagulopathy, leukemia, papilledema, and retinal vein occlusions, and the trauma cases comprised of vaginal births, accidental trauma, and abusive head trauma. The dataset was randomly divided at the patient level in training (80%) and test (20%) sets. We developed two deep learning models, a simple convolutional neural network (CNN) and a transfer learning model using ResNet50, to predict if RH in the fundus photos were from a medical or a traumatic case. The models' performance was evaluated using the area under the receiver operating characteristic curve (AUC) on the test dataset. We used gradient-weighted Class Activation Mapping (Grad-CAM) for the models' visual explanation.

Results : There were 298 (49.9%) fundus photos with RH due to a medical cause and 299 (50.1%) with RH due to trauma. The CNN model achieved an AUC of 0.89, while the Resnet50 model with transfer learning achieved an AUC of 0.93, meaning that there is a 93% chance the Resnet50 model would be able to segregate medical and traumatic RH in this limited dataset. We were able to visualize the features used for classification using Grad-CAM (Figure 1).

Conclusions : This preliminary study shows that deep learning models have potential to classify the underlying cause of RH using fundus photos into either a traumatic or medical category. Grad-CAM allowed for the evaluation of visualization of the features used by the models for classification, providing insight into the decision-making process of artificial intelligence. While these findings suggest that artificial intelligence may eventually be a helpful tool for inferring the cause of RH in young children, a detailed systemic, multidisciplinary evaluation still will likely be required to determine the underlying cause of RH.

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

 

Figure 1. The fundus photos with RH A) anemic retinopathy (Medical) and B) vaginal delivery baby (Trauma), followed by the Grad-CAM output highlighting the features used by the ResNet model with the area of most and least attention in blue and yellow, respectively.