On this episode of Evidence Based Retina, Rishi P. Singh, MD, FASRS, is joined by Sharon Fekrat, MD, FACS, FASRS, to discuss using multimodal retinal imaging and AI to identify neurocognitive disease.
Dr. Fekrat is Vice Chair of Faculty Affairs and Professor of Ophthalmology, Neurology, and Surgery at Duke University School of Medicine, and Founder and Director of the iMIND Study Group.
iMIND study fast facts
- Study group: The iMIND (Eye Multimodal Imaging in Neurodegenerative Disease) Study Group, based in the Duke Neurology clinic since 2017, is investigating whether multimodal retinal and choroidal imaging, analyzed with AI, can detect neurodegenerative conditions before symptoms appear.
- Key findings: Two consistent findings have emerged across neurocognitive conditions in the iMIND cohort:1,2
- Thinning of the ganglion cell-inner plexiform layer (GC-IPL) on OCT
- Decreased vessel and perfusion density in the retinal microvasculature on OCT angiography (OCTA)
- AI application: Convolutional neural network (CNN) models trained using multimodal retinal imaging have shown promise for identifying symptomatic Alzheimer's disease and GC-IPL maps were the most significant contributor to the model.3
- Standardization: Challenges with image standardization remain before this approach can enter routine clinical care. These include harmonized acquisition protocols, segmentation rules, image quality thresholds, and reporting standards.
Deeper dive into the iMIND study
The retina shares embryonic origins with the brain and, as neurosensory tissue, may reflect changes occurring in the central nervous system.4 Alzheimer's disease has a preclinical phase, a nearly 20-year asymptomatic period, in which pathophysiological changes can precede symptom onset by decades.5
Identifying patients during that asymptomatic window could open opportunities for clinical trial enrollment, risk factor modification, and earlier therapeutic intervention. The iMIND Study Group is investigating whether multimodal retinal and choroidal imaging analyzed with AI can identify neurodegenerative disease during this window, across a large participant cohort.
To date, over 2,000 participants between the ages of 18 and 90 have been enrolled into iMIND, mostly from Duke Neurology clinics, including individuals with:1,2
- Alzheimer's disease
- Mild cognitive impairment
- Parkinson's disease
- Frontotemporal dementia
- Dementia with Lewy bodies
- Multiple sclerosis
- ALS
- Huntington's disease
- Traumatic brain injury
- PTSD
Controls include cognitively normal adults as well as individuals who know their APOE genetic status but have not yet developed symptoms, a group Dr. Fekrat described as particularly valuable.
Multimodal imaging approach
The group uses several imaging modalities; the ZEISS Cirrus HD-5000 OCT captures undilated macular cube and optic disc cube data across a range of structural metrics and the ZEISS AngioPlex OCTA measures vessel and perfusion densities in the superficial capillary plexus as well as peripapillary capillary flux index and perfusion density.
Ultra-widefield imaging with the OPTOS California device captures red-green color images and fundus autofluorescence, though lid and lash artifacts in these undilated images remain a challenge for reliably capturing the peripheral retina.
Common themes across conditions
GC-IPL thinning on OCT has been documented in Alzheimer's disease and mild cognitive impairment.1 Decreased vessel and perfusion densities on OCTA have been observed across the full range of neurocognitive conditions studied.1,2 Both findings have held across the iMIND cohort since imaging began in 2017, Dr. Fekrat noted.
Since microvascular loss and GC-IPL thinning have many potential causes, the iMIND study excludes participants with diabetes, glaucoma, or other conditions that could independently produce similar findings. This supports the interpretation that the retinal changes observed may indeed be specific to the neurocognitive and neurodegenerative conditions under study.
AI and oculomics
Traditional cross-sectional statistical analyses have provided useful data, but AI may detect patterns that those methods cannot. The iMIND group has developed CNN models using multimodal retinal imaging that have successfully differentiated symptomatic Alzheimer's disease and mild cognitive impairment from normal cognition.3,6
GC-IPL thickness maps had the greatest impact on model performance, with quantitative patient data and OCTA images also contributing.3,6 The iMIND group has also developed a model that can differentiate Parkinson’s disease from controls.7 Dr. Fekrat also noted interest in incorporating whole-fundus OCTA, which could allow imaging of the entire fundus, including the peripheral retina that current devices have not fully captured.
Barriers to clinical use
Several hurdles remain before this can be applied in clinical care. Broader and more diverse patient enrollment is needed, along with harmonized acquisition protocols, rules around segmentation, image quality thresholds, and reporting standards, similar to what the DRCR Retina Network has done for diabetic retinal imaging research.
When these are addressed, Dr. Fekrat noted potential roles could include population screening, primary care triage, adjunctive use in memory clinics, and pre-selection for clinical trial entry.
This article was written by Sonia Kelley, OD, MS, based on the recorded video from Drs. Singh and Fekrat.
