Vertical Cup-to-Disc Ratio Estimation using Deep Learning
Care Process & Redesign
Technology
Ng Teng Fong Healthcare Innovation Programme
National Healthcare Group
National University Health System
31 December 2024
The purpose of this study is to develop an AI model for glaucoma screening through the measurement of vertical. We developed a deep learning system that estimates vCDR from standard, UWF, and smartphone-based images.
Year Submitted: 2024
Published Date: 31 December 2024
Tags: Care Process ＆ Redesign, Technology, Artificial Intelligence
About this Content
Aims
The purpose of this study is to develop an AI model for glaucoma screening through the measurement of vertical cup-disc-ratio. We then validate the effectiveness and generalizability of our deep learning system for vCDR estimation across standard, UWF, and smartphone-based images.
Background
Glaucoma, a leading cause of blindness, is projected to affect over 111.8 million people globally by 2040, with an increasing burden on eye care services due to aging populations. In Singapore, where the elderly population is expected to more than double by 2050, the demand for eye care services is rising rapidly. As the number of trained ophthalmologists struggles to keep pace, there is a growing need to leverage other healthcare professionals, such as optometrists, to manage stable chronic eye conditions.
Methods
In this study, 151 patients aged 27 to 87 from the Ang Mo Kio (AMK) Specialist Centre in Singapore participated. Their pupils were dilated before having fundus photos taken with the Zeiss CLARUS 500 and the oDocs Nun IR portable camera. A deep learning-based automated system was developed for estimating the vertical cup-to-disc ratio (vCDR), utilizing detection and segmentation CNNs based on the U-Net architecture with a ResNet-18 encoder. To enhance the accuracy of optic disc (OD) and optic cup (OC) delineations, test-time augmentation (TTA) was employed.
Results
For vCDR estimation, the automated system achieved MAEs of 0.040 (95% CI, 0.037–0.043), 0.068 (95% CI, 0.061–0.075), and 0.084 (95% CI, 0.075–0.092) in the test set of the REFUGE data set (n = 400), AMK UWF images (n = 296), and AMK smartphone-based images (n = 300), respectively. Without correcting for optic disc size, there was no correlation between OCT RNFL thickness and fundus CDR in all sectors. After correcting for disc area, a weak negative correlation (r = −0.4046, P < 0.05) was observed at the 90°/S2 sector.
Conclusion
We developed a deep learning system that estimates vCDR from standard, UWF, and smartphone-based images.
Lessons Learnt
At NHG, the development of a customized AI solution for NHG is a novel endeavour. NHG and TTSH have protocols for vendor-based solutions, and are building up infrastructure for in-house development. For NHG to truly see value in transforming care, the basic technological infrastructure, manpower and training will need to be developed.
Additional Information
Publication: https://tvst.arvojournals.org/article.aspx?articleid=2793526
Keywords
AI, deep learning, glaucoma screening, fundus photos, smartphone
Innovators' Details
Innovators' Details
Healthcare Cluster(s) | National Healthcare Group, National University Health System |
Organization(s) Involved | Tan Tock Seng Hospital, Nanyang Technological University |
Platform(s) | Ng Teng Fong Healthcare Innovation Programme |
Healthcare Professional Group(s) | Medical |
Applicable Specialty or Discipline | Ophthalmology |
Project Lead(s) | Kelvin Li Zhenghao |
Project Member(s) | Kelvin Li Zhenghao |
Connect with this contributor!
Dr Kelvin Li Zhenghao - Kelvin_li@ttsh.com.sg
Project Attachment
801_TTSH_NTFHIP 2024_Vertical Cup-to-Disc Ratio Estimation using Deep Learning.pdf