Application of convolutional neural networks to define Fuchs endothelial dystrophy

Abstract

Purpose. To evaluate the application of convolutional neural networks for the automatic detection of Fuchs' dystrophy. Material and methods. The study included 700 biomicroscopic images of the corneal endothelium (Tomey EM-3000) randomly selected from the database of the Saint-Petersburg brunch of the S. Fyodorov Eye Microsurgery Federal State Institution. At the first stage, the images were divided into 2 groups. The first group included images with the presence of Fuchs' dystrophy, the second – another pathology or a healthy cornea. The corneal endothelial cell density images were divided into three categories: training, validation, and test datasets. In our study we tested various architectures of convolutional neural networks: ResNet18, ResNet50, VGG16, VGG19 and GoogleNet. Results. The approbation of the neural network on the test sample has given the following values of the F-measure: ResNet18: 0.985; ResNet50: 1,000; VGG16: 0.940; VGG19: 0.990; GoogleNet: 0.987. Pre-trained network ResNet50 performed best with frozen layers, Adam optimizer, cross-entropy as a loss function, and a training step of 0.000005. Conclusion. The use of convolutional neural networks for the automatic detection of Fuchs' dystrophy can be successfully implemented as part of a doctor's decision support system. ResNet50 showed the best results among all types of models and did not give a single error on the test sample, which indicates the high efficiency of using this network in the classification algorithm for corneal endothelial images. Keywords: artificial intelligence, Fuchs corneal dystrophy, convolutional neural networks