Purchase this article with an account.
Hong Qi, Davide Borroni, Rongjun Liu, Bryan Williams, Mike Beech, Yitian Zhao, Baikai Ma, Vito Romano, Uazman Alam, Stephen B Kaye, Yalin Zheng; Automated detection of corneal nerves using deep learning. Invest. Ophthalmol. Vis. Sci. 2018;59(9):5721.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
There has been increasing use of in-vivo confocal microscopy (IVCM) for the non-invasive examination of corneal nerves. This allows the study of nerve alterations in different ocular diseases, both before and after corneal surgery, and in systemic diseases such as diabetes. However, there has been no corresponding increase in the availability of software tools to support the automatic analysis of corneal nerves. We propose and evaluate a new computer aided detection model based on deep learning for the automatic detection of corneal nerves in IVCM images.
584 IVCM images were acquired using the Heidelberg Retina Tomograph 3/Rostock Cornea Module (Heidelberg Engineering, Heidelberg, Germany) from healthy corneas and corneas with various corneal conditions. The corneal nerves in each image were manually traced by a clinical ophthalmologist (DB) to provide a ground truth. 437 (76%) randomly chosen images were used to train the deep learning model while the remaining 147 (24%) were reserved for testing the model. A convolutional neural network (CNN) was adopted here for the purpose of segmenting the corneal nerves, comprising 10 convolution layers and 9 pooling layers. The Dice Similarity Coefficient (DSC) was used as the cost function to train the network. Drop-out (ratio 0.2) was used in order to reduce the problem of overfitting during training. In order to improve performance, 81 patches of size 128x128 pixels per image were produced and used as input for the model. The model was trained for 200 iterations and performance of the trained model was evaluated using DSC.
Excellent segmentation results were achieved. In terms of patches, the mean DSC values obtained were 0.916 for the validation set and 0.867 for the test set. Reconstructing the whole-image segmentations from the patches and comparing with expert annotations, we achieved a mean DSC of 0.856.
Our results have demonstrated that this deep learning based model can provide very encouraging localisation performance for the detection of corneal nerves. The developed model could be further refined and validated with large datasets in order to support the management of eye disease and systemic disease.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.
This PDF is available to Subscribers Only