Multimodal Retinal Imaging for the Detection of Early Alzheimer's Disease

The problem

Finding Alzheimer's early is hard, and it matters

In Alzheimer's disease, two proteins, amyloid-β and tau, slowly accumulate in the brain long before symptoms appear. Worldwide, an estimated 315 million people are thought to be in this silent "preclinical" phase: amyloid-positive, but cognitively normal.

Spotting the disease at this stage is becoming more important. New antibody treatments that target amyloid can moderately slow cognitive decline, clinical trials need to recruit people early, and lifestyle changes have the most room to help. But the reference tests, a PET scan of the brain or a spinal tap to analyse cerebrospinal fluid, are expensive or invasive, and not suited to screening large groups of people. Blood tests are a promising alternative. So is the eye.

Why the retina?

The retina grows out of the same embryonic tissue as the brain and is directly connected to it through the optic nerve. It is the only part of the central nervous system we can look at directly, with light, in a few seconds, and without any injection. And the cameras to do it are already in eye clinics. If Alzheimer's-related changes show up in the retina, an eye exam could one day help decide who should get a PET scan or be invited to a trial.

The approach

Three cameras, three different views of the same retina

Each imaging technique captures a different aspect of the retina. The underlying idea: Alzheimer's is not one process but several (protein build-up, loss of nerve cells, blood vessel problems), so a combination of views might see more than any single one.

Who took part

The 40 participants (aged 50–85) came from RetAD, a longitudinal study at UZ Leuven. Crucially, everyone had an amyloid PET scan, so we knew who had amyloid in their brain, rather than relying on symptoms alone.

For each modality we trained a simple, transparent model (logistic regression) to tell Aβ+ from Aβ− participants, and tested it with leave-one-out cross-validation: the model is trained on 39 people and asked to predict the 40th, repeated for every person. That way, nobody is ever scored by a model that has seen them.

A closer look

Reading texture, not just colour

Raw brightness in a hyperspectral image varies a lot from person to person: lens clarity, pupil size and camera flash all play a role. In our data, the normalised brightness per colour band overlapped heavily between the two groups. So instead of asking how bright the retina is, we asked how the brightness is arranged.

One classic texture measure is GLCM correlation (from the gray-level co-occurrence matrix). It asks: if I know the brightness of one pixel, how well can I predict the brightness of its neighbour? A smooth, uniform texture gives a value close to 1. A patchier, more heterogeneous texture gives a lower value. Try it:

What we found

Five findings, and how sure we are about each

Across all colour bands and regions, the strongest candidate signal was GLCM correlation at 466 nm, the shortest (bluest) wavelength we analysed. Amyloid-positive participants had lower values, meaning a more heterogeneous local texture. The difference was clearest in the lower (inferior) half of the macula, where texture alone separated the groups with an AUC of 0.72.

Why blue? Shorter wavelengths scatter more strongly off very small structures in tissue. Earlier studies, in Alzheimer's mouse models and in patients, also reported the largest retinal optical changes at short wavelengths, below about 500–565 nm.

We also checked how stable the measurement is. Repeating the scan with different flash settings on the same day gave very consistent values (ICC up to 0.88). Across visits months apart, repeatability was moderate (ICC 0.60 for the full macula).

On OCT, the overall pattern of GC-IPL thickness across the nine macular regions was associated with amyloid status (p = 0.009, adjusted for age and sex). The nerve fibre layer (RNFL) was not. Interestingly, no single region differed significantly on its own: the information lies in the spatial pattern, not in one spot.

A model using all nine GC-IPL regions reached an AUC of 0.73. And, like HSI, the inferior regions were far more informative than the superior ones (AUC 0.71 vs. 0.45). Two very different techniques pointing at the same part of the retina is intriguing. Since both were measured in the same people, though, it is a convergence worth following up, not an independent replication.

Of 24 vascular measures, only arteriole width (slightly wider in Aβ+) and the number of venous endpoints showed nominal differences, and neither survived correction for multiple comparisons. Arteriole and venule width together gave an AUC of 0.68, and adding them to the best model did not help (0.84 → 0.81).

That doesn't mean blood vessels are irrelevant in Alzheimer's. Static vessel width on a photo is a coarse measure, and vessel width is also strongly affected by blood pressure, diabetes and heart disease. Changes at the level of the tiniest capillaries (measurable with OCT angiography) or in how vessels react (dynamic vessel analysis, now being collected in RetAD) may be more sensitive.

The strongest model combined HSI texture with the nine OCT GC-IPL regions. How to read the numbers:

To look beyond the early stage, we added 39 participants from our earlier study (22 cognitively normal, 17 with Alzheimer's dementia), imaged with the same camera. Texture was lower in the early amyloid-positive group than in controls, but in people with dementia it was back up, not different from controls.

A similar "strongest in the middle" pattern was reported in an earlier hyperspectral study, whose authors suggested that small, soluble amyloid clusters early in the disease may affect light differently from the larger, fibrillar aggregates found later. That could explain our pattern too, but it remains a hypothesis. It also has a practical implication: a marker that goes down and then back up cannot simply be read as "lower is worse".

The bigger picture

Why combining views makes biological sense

Researchers classify Alzheimer's biomarkers by the process they reflect: Amyloid, Tau, Neurodegeneration, and Vascular damage as a common co-pathology. Our three retinal views loosely line up with that framework.

This offers a plausible explanation for why HSI and OCT might complement each other: they may sample different disease processes. But it is a way of framing hypotheses. The retinal measures do not directly quantify the corresponding processes in the brain, and future studies need to test these links against established brain and blood biomarkers.

Reading this responsibly

What this study does not show

• It is small. 40 participants, 17 amyloid-positive. That means wide confidence intervals and a real risk that the models are tuned to this particular group.
• It has not been externally validated. The key HSI feature was discovered in the same data it was tested on, and no independent cohort with PET-confirmed amyloid status was available.
• It doesn't test specificity. People with glaucoma or other neurological diseases were excluded, so we don't yet know whether these signals are specific to Alzheimer's.
• It is not a diagnostic test. The results are hypothesis-generating estimates of what might be possible, not measures of clinical performance.

What it does provide: preliminary evidence that the retina carries measurable signals linked to brain amyloid, a concrete candidate feature (466 nm texture), a candidate region (the inferior macula), and a reason to pursue multimodal retinal imaging in larger, longitudinal cohorts.

If you read nothing else

Five take-aways

1. The retina shows signals linked to brain amyloid, even before dementia.In preclinical and prodromal Alzheimer's, confirmed by PET.
2. Texture beats brightness.How light is arranged across the retina was more informative than how much light is reflected, with blue light (466 nm) standing out.
3. For OCT, the spatial pattern matters.No single region differed on its own, but the pattern of GC-IPL thickness across the macula did. Both HSI and OCT pointed to the lower half of the macula.
4. HSI and OCT look complementary.Together they reached AUC 0.84, the best of all combinations, though not yet significantly better than either alone.
5. Next: bigger, longer, prespecified.Larger longitudinal cohorts with the wavelength and region fixed in advance, plus capillary-level and functional vessel imaging, which is part of the ongoing RetAD study.

Cite

BibTeX

@article{ghesquiere2026multimodal,
  author  = {Ghesquiere, Michiel and Christinaki, Eirini and Lemmens, Sophie and
             Van Eijgen, Jan and Beeckmans, Lennert and Ghinis, Achilleas and
             Jacobs, Thomas and Van Keer, Karel and Wehbi, Zahi and De Groef, Lies and
             Dahdouh-Guebas, Yasmin and Charle, Wouter and Vande Casteele, Thomas and
             Vandenbulcke, Mathieu and Vanderlinden, Greet and Van Laere, Koen and
             Schaeverbeke, Jolien and Vandenberghe, Rik and Ceccarini, Jenny and
             De Vos, Maarten and Stalmans, Ingeborg},
  title   = {Multimodal Retinal Imaging for the Detection of Early {Alzheimer's} Disease:
             Combining Biochemical, Structural, and Vascular Biomarkers},
  journal = {Bioengineering},
  year    = {2026},
  volume  = {13},
  number  = {9},
  pages   = {1041},
  doi     = {10.3390/bioengineering13091041}
}