Research Article
Automated Cataracts Screening from Slit-Lamp Images Employing Deep Learning
@INPROCEEDINGS{10.1007/978-3-030-67514-1_23, author={Zhipeng Zhang and Wenhui Shou and Dongjia Xing and Wenting Ma and Qingqing Xu and Wei Wang and Li-Qun Xu and Ziming Liu and Ling Xu}, title={Automated Cataracts Screening from Slit-Lamp Images Employing Deep Learning}, proceedings={IoT as a Service. 6th EAI International Conference, IoTaaS 2020, Xi’an, China, November 19--20, 2020, Proceedings}, proceedings_a={IOTAAS}, year={2021}, month={1}, keywords={Deep learning Mobile healthcare application Automated cataracts screening}, doi={10.1007/978-3-030-67514-1_23} }- Zhipeng Zhang
Wenhui Shou
Dongjia Xing
Wenting Ma
Qingqing Xu
Wei Wang
Li-Qun Xu
Ziming Liu
Ling Xu
Year: 2021
Automated Cataracts Screening from Slit-Lamp Images Employing Deep Learning
IOTAAS
Springer
DOI: 10.1007/978-3-030-67514-1_23
Abstract
To assess the feasibility and performance using deep learning networks to automatically detect cataracts from slit-lamp images in large-scale eye diseases screening scenarios. Two datasets were collected using, respectively, the professional Slit-Lamp Microscopes (SLM) and the portable Slit-Lamp Devices (SLD) clipped on a Smartphone, during routine eye disease screening programs in China. The former Dataset-M comprised 4891 images from 1670 subjects and the latter Dataset-D comprised 2516 images from 802 subjects. Each image was then labelled by three ophthalmologists as one of the three classes: 1) un-gradable image, 2) cataract, and 3) normal. For each dataset, two deep learning models were created: one for image quality assessment, and the other for cataracts detection, and the performance of which was assessed by the Area Under a ROC Curve (AUC) and kappa agreement. For the quality assessment models, on Dataset-M (Dataset-D), the corresponding AUC achieved were 0.929 (0.881), with kappa agreements of 0.628 (0.590) and p < 0.001, respectively. For the cataract detection models, the corresponding AUC were 0.997 (0.987), with kappa agreements of 0.912 (0.893) and p < 0.001, respectively. Furthermore, based on these models we built a practical cloud application that has been trialled in 25 real-world screening settings in China, receiving favourable feedbacks from clinicians, primary care physicians and patients alike.
