Purchase this article with an account.
Ling-Ping Cen, Jie Ji, Jianwei Lin, Calvin C P Pang, Mingzhi Zhang; Development and Validation of a Cloud-based Deep Learning Platform for Detection of 37 Fundus Diseases in Retinal Photographs. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1457.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop and apply a cloud-based deep learning platform (DLP) for automated detection of 37 types of ocular fundus diseases and conditions in retinal photographs.
A DLP composed of various convolutional neural networks (CNNs) was constructed. After being trained with 96 758 fundus images, the diagnostic performance of DLP was validated with three datasets using 110 738 images for detection of 37 types of fundus diseases and conditions. Confusion matrix, overall accuracy (OA), Cohen’s Kappa and relative classifier information (RCI) were applied to evaluate the DLP performance of multi-category classification. Sensitivity, specificity and area under the receiver operating characteristic curve (AUC) were applied for binary classification. All these results of the DLP were generated based on the reference standard of licensed ophthalmologists.
The primary validation dataset consisted of 41 468 images from 32 614 subjects (mean age, 49.9 years, 51.2% men). The multihospital test sets had 58 251 images from 25 515 subjects (mean age, 55.0 years, 49.3% men). The screening categorized dataset had 11 019 images from 2594 subjects (mean age, 61.6 years, 27.8% men). For Classification of 29 categories, the DLP had an OA of 97.9%, a Kappa of 0.968 and an RCI of 0.939 for primary validation. For multihospital tests, OA range was 95.6% to 97.5%. For screening categorized, OA was 97.0%. For detecting referable in the “nonreferable-vs-referable” classification, AUC range was 0.999 (95%CI, 0.999-1.000), sensitivity was 99.4% (95% CI, 99.3%-99.5%) and specificity was 98.6% (95% CI, 98.5%-98.8%) for primary validation. For multihospital tests, AUC range was 0.996 to 0.999. For screening categorized, AUC was 0.995 (95%CI, 0.994-0.996). For classification of subcategories, accuracy range was 85.5% to 97.8%, AUC range 0.868 to 0.996. The DLP has been applied to community screening of retinal diseases and opened online for public use.
We have developed a cloud-based DLP for automated detection of 37 types of ocular fundus diseases and conditions with high OA, Kappa, RCI as well as high sensitivity, specificity and AUC.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
Example of Original Images, Preprocessed Images, Heatmaps, (Cropped optic disc and Optic disc mask) of Selected Categories.
Examples of Online Application of the DLP for Clients.
This PDF is available to Subscribers Only