DAG KITTLAUS
December 22, 2023Article of the Month – December 2023
December 29, 20232018
In ophthalmology, DL has demonstrated clinically acceptable diagnostic performance in identifying a variety of retinal disorders, especially DR and AMD...
DL Applications in Ophthalmology :
Modern machine learning techniques known as “deep learning” (DL) have attracted much attention from around the world in recent years. By 2040, a third of the 600 million people with diabetes worldwide will also have Diabetic Retinopathy (DR). Programs for DR screening face difficulties with execution, a lack of human assessors, and long-term financial viability.(1)
The diagnostic performance in identifying DR has been revolutionized over the past few years by DL. As an example, Abràmoff et al. demonstrated that a DL system was able to achieve an area under the receiver operating characteristic curve (AUC) of 0.980, with sensitivity and specificity of 96.8% and 87.0%, respectively, in the detection of referable DR on Messidor-2 data set. Many groups have demonstrated excellent diagnostic performance using this technique.(2)
More recently, Gulshan and colleagues from Google AI Healthcare reported DL systems with good diagnostic performance. A panel of 54 US-licensed ophthalmologists and ophthalmology residents graded 128,175 retinal pictures for DR and DMO between May and December 2015 and used the results to construct the DL system. The test set included roughly 10,000 photos taken from two publicly accessible databases (EyePACS-1 and Messidor-2), and they were assessed with high intragrader consistency by at least seven US board-certified ophthalmologists. For EyePACS-1 and Messidor-2, the AUC was 0.991 and 0.990, respectively.(3)
Age-related Macular Degeneration and Retinopathy of Prematurity:
Globally, AMD is a significant contributor to vision impairment in older adults. The stages of AMD were divided into none, early, middle, and late AMD by the Age-Related Eye Disease Study (AREDS). Due to the aging population, a strong DL system is urgently needed to screen these patients for subsequent assessment in tertiary eye care facilities. Ting et al. reported a DL system diagnostic performance in diagnosing referable AMD that was clinically acceptable.
Specifically, 108,558 retinal pictures from 38,189 patients were used to train and evaluate the DL system. In this investigation, fovea-centered images without macula segmentation were employed. As this cohort was being screened for DR, few patients had referrable AMD.
Age-related Macular Degeneration and Retinopathy of Prematurity:
Globally, AMD is a significant contributor to vision impairment in older adults. The stages of AMD were divided into none, early, middle, and late AMD by the Age-Related Eye Disease Study (AREDS). Due to the aging population, a strong DL system is urgently needed to screen these patients for subsequent assessment in tertiary eye care facilities. Ting et al. reported a DL system diagnostic performance in diagnosing referable AMD that was clinically acceptable.
Specifically, 108,558 retinal pictures from 38,189 patients were used to train and evaluate the DL system. In this investigation, fovea-centered images without macula segmentation were employed. As this cohort was being screened for DR, few patients had referrable AMD.
With a large number of referable AMD, DL systems were created for the other two studies,(17,18) using the AREDS data set (intermediate AMD or worse). Burlina et al. found a diagnosis accuracy between 88.4% and 91.6% using fivefold cross-validation, with an AUC of between 0.94 and 0.96.(4)
ROP is the most common cause of pediatric blindness in the world, accounting for 32,000 new cases of blindness each year. Early attempts have been made to apply DL for automated ROP diagnosis, which can overcome both implementation obstacles for ROP screening. Brown et al. most recently published the findings of a wholly automated DL system that could diagnose plus disease, the most significant aspect of severe ROP, with an AUC of 0.98 compared to a consensus reference standard diagnosis combining image-based diagnosis and ophthalmoscopy. When compared directly, the i-ROP DL method agreed with the consensus diagnosis more frequently than six out of eight international experts in ROP diagnosis.(4)