Automatic screening of pathological myopia using deep learning

Pathological Myopia (PM) is one of main causes of visual impairment and irreversible blindness worldwide. Its social and economic burden has been demonstrated by epidemiological studies. Early detection and intervention of PM are crucial to slow down the degree of retinal damage and decrease the risk of blindness. However, manually scrutinizing is tedious and subject to uncertainties. In the paper, therefore, we propose an AI-driven fundus screening system for PM, which can identify characteristic of PM directly from fundus imaging without expert intervention. The automatic PM screening system consists of a set of advanced Convolutional Neural Networks (CNNs) based on 3,796 color fundus photographs, via Transfer Learning and Ensemble Learning technology. Furthermore, we empirically investigate the performances of both the lightweight and larger-scale networks on this work to construct the optimal model. Finally, we evaluate the validity and reliability of the PM screening system through six metrics. Experimental results show that the screening system for PM performs excellently, with an accuracy of 99.7%, a sensitivity of 99.5% and a specificity of 100%. This work has the potential to provide insights and reduce workload in PM screening for ophthalmologists.