Glaucoma is a progressive optic neuropathy and the leading cause of irreversible blindness in the world. As the disease remains largely asymptomatic until late stages, there is a pressing need to develop affordable approaches for screening before visual impairment occurs. Although sophisticated imaging technologies such as Spectral domain-optical coherence tomography (SDOCT) can provide highly reproducible and accurate quantitative assessment of glaucomatous damage, their application in widespread screening or non-specialized settings is unfeasible, given the high cost and operator requirements. Fundus photography is a low-cost alternative that has been used successfully in teleophthalmology programs. However, subjective human grading of fundus photos for glaucoma is poorly reproducible and highly inaccurate, as gradings tend to largely over- or underestimate damage. We propose a new paradigm for assessing glaucomatous damage by training a deep learning (DL) convolutional neural network to provide quantitative estimates of the amount of neural damage from fundus photographs. In our Machine-to-Machine (M2M) approach, we trained a DL network to analyze fundus photos and predict quantitative measurements of glaucomatous damage provided by SDOCT, such as retinal nerve fiber layer (RNFL) thickness and neuroretinal rim measurements. Our preliminary results showed that the M2M predictions have very high correlation and agreement with the original SDOCT observations. This provides an objective method to quantify neural damage in fundus photos without requiring human graders, which could potentially be used for screening, diagnoses and monitoring in teleophthalmology and non- specialized point-of-care settings. In this proposal, we aim at refining and validating the M2M model in suitable, large datasets from population-based studies, electronic medical records, and clinical trial data. Our central hypothesis is that the M2M approach will be more accurate than subjective human gradings in screening, diagnosing, predicting and detecting longitudinal damage over time.
In Aim 1, we will investigate the performance of the M2M model to screen for glaucomatous damage using large datasets from 6 population-based studies: Blue Mountains Eye Study, Los Angeles Latino Eye Study, Tema Eye Survey, Beijing Eye Study, Central India Eye and Medical Study and the Ural Eye and Medical Study, which will provide data on over 25,000 subjects of diverse racial groups.
In Aim 2, we will investigate the ability of the M2M model to predict future development of glaucoma in eyes of suspects using the data from the Ocular Hypertension Treatment Study (OHTS).
In Aim 3, we will investigate the ability of the M2M model in detecting glaucomatous progression over time using data from the Duke Glaucoma Registry, a large database of longitudinal structure and function data in glaucoma with over 25,000 patients followed over time. If successful, this proposal will lead to a validated, inexpensive, and widely applicable tool for screening, early diagnosis and monitoring of glaucoma, that could be applied under population-based settings and also at non-specialized point-of-care settings.
Glaucoma is a leading cause of irreversible visual impairment in the world. This proposal will employ a novel artificial intelligence paradigm for quantifying neural damage on ocular fundus photographs for the purpose of screening, diagnosing and monitoring glaucoma damage. The approach will be validated on large datasets from population-based studies, electronic medical records and clinical trial data.
