Bi-DenseNet: Automatic Recognition of Ocular Surface Disease Using Smartphone Imaging

Xiongbiao Luo,Xiang Lin,Weijie Ouyang,Song Zheng,Jianhui Chen,Zuguo Liu

DOI: https://doi.org/10.1016/j.bspc.2024.106622

IF: 5.1

2024-01-01

Biomedical Signal Processing and Control

Abstract:Ocular surface disease is most common and prevalent eye disorder that is difficult to precisely diagnose due to its complexity more than ophthalmologists realize. Biomicroscopic diagnosis of this disease at the hospital is large economic and humanistic burden such as cost and visit duration. This work presents the first study on automatic recognition of ocular surface disease by smartphone imaging and bilateral densely connected convolutional networks. Our new solution is to use cheap smartphones as a replacement of expensive biomicroscopes to image the ocular surface. To recognize ocular surface disease on images, we propose new bilateral densely connected convolutional networks that introduce an additional cheap hybrid operator to extract more diverse features and concatenate double outputs of each layer into the feature maps as inputs of the next layer, leading to a significant increase of efficient feature reuse of dense connections while critically reducing the amount of parameters. Three ophthalmologists used different smartphones to collect a large amount of ophthalmic images with various ocular surface diseases to evaluate our method, with the validation results showing that our method provides a promising and effective strategy to accurately diagnose ocular surface disorders. In particular, our smartphone image deep learning based automatic recognition method attains much higher average recognition accuracy of 91.3% than 87.1% of densely connected networks, while it only needs 0.25M model parameters. Additionally, our method can be conveniently integrated into smartphones for subjects to monitor eye surface disorders without going to hospital. The pre-trained Bi-DenseNet model with its code will be available at GitHub soon.