An Image Diagnosis Algorithm for Keratitis Based on Deep Learning

Clinical diagnosis of keratitis highly depends on the observation of medical images. Since there are many classifications of keratitis, and the pathogenic factors are different, ophthalmologists will be more demanding. In this paper, a multi-task recognition method is proposed for the automatic diagnosis of keratitis. The diagnosis basis of keratitis is obtained, and the image of the anterior segment is interpreted. Under the guidance of ophthalmologists, all anterior segment images are labeled from five signs, consisting of opacity area in the cornea (turbid and clear), boundary of the focus (distinct and vague), epithelium of the focus area (intact and incomplete), hyperemia (congestive and healthy), and neovascularization (yes and no), which are important in the diagnosis of keratitis. A multi-label image dataset is constructed, and the images are enhanced by horizontal flipping according to the image characteristics. In this paper, an improved multi-attribute network based on ResNet50 is proposed, including a feature extraction module and a classification module. The feature extraction module is to extract image features, and the classification module is a multi-output network in which each channel corresponds to each attribute. In order to improve the overall recognition accuracy of multi-task, the loss function is optimized. In the loss function, the loss weights of different tasks are determined based on the classification difficulty. A joint training approach is used to train the multi-attribute network which can simultaneously recognize the five attributes and obtain the specific symptoms of keratitis. The experimental results show that the average accuracy of these five attributes can be achieved 84.89% in the multi-attribute network, among which the highest accuracy can be achieved 89.51%.