Dementia Detection via Retinal Hyperspectral Imaging and Deep Learning: Clinical Dataset Analysis and Comparative Evaluation of Multiple Architectures

Abstract

This study aimed to detect dementia using intelligent hyperspectral imaging (HSI), which enables the extraction of detailed spectral information from retinal tissues. A total of 3256 ophthalmoscopic images collected from 137 participants were analyzed. The spectral signatures of selected retinal regions were reconstructed using hyperspectral conversion techniques to examine wavelength-dependent variations associated with dementia. To assess the diagnostic capability of deep learning models, four convolutional neural network (CNN) architectures-ResNet50, Inception_v3, GoogLeNet, and EfficientNet-were implemented and benchmarked on two datasets: original ophthalmoscopic images (ORIs) and hyperspectral images (HSIs). The HSI-based models consistently demonstrated superior accuracy, achieving 84% with ResNet50, 83% with GoogLeNet, and 82% with EfficientNet, compared with 80-81% obtained from ORIs. Inception_v3 maintained an accuracy of 80% across both datasets. These results confirm that integrating spectral information enhances model sensitivity to dementia-related retinal changes, highlighting the potential of HSI for early and noninvasive detection.

Keywords: artificial intelligence; dementia detection; hyperspectral imaging.

Figure 1

Overall experimental workflow for dementia detection using hyperspectral imaging. The process involves: (1) Acquiring fundus images and extracting spectral features from five Regions of Interest (ROIs); (2) Reconstructing hyperspectral (HS) cubes (512 × 512 × 401) from RGB images; (3) Preprocessing HS images into three-channel inputs (512 × 512 × 3) suitable for Deep Learning; and (4) Training and evaluating four CNN architectures (ResNet50, Inception_v3, GoogLeNet, EfficientNet) to classify cognitive status (Normal, MCI, Dementia) based on Mini-Mental State Examination (MMSE) scores.

Figure 2

Representative Hyperspectral (HS) montages compared with original ophthalmoscopic images. The figure displays color-reconstructed HS images at specific spectral bands (550, 600, 650, 700, and 780 nm). These wavelength-dependent views reveal structural and vascular variations—such as vessel contrast and pigmentation—that are less discernible in standard RGB imaging. The range 550–780 nm was specifically selected to highlight these features.

Figure 3

Schematic representation of the five Regions of Interest (ROIs) selected for spectral analysis. The standardized sampling areas (240 × 240 pixels) are located at: the fovea (F), the superior temporal arcade (S1, S2), and the inferior temporal arcade (I1, I2). These regions were chosen to ensure consistent spatial sampling of vascular and neural tissues across all participants.

Figure 4

Age-dependent variations in retinal spectral reflectance among cognitively normal participants. The panels display mean spectral reflectance curves for three age groups (60 s, 70 s, and 80 s) across five retinal regions: (a) Fovea (F), (b) Inferior 1 (I1), (c) Inferior 2 (I2), (d) Superior 1 (S1), and (e) Superior 2 (S2). Shaded bands represent the spectral variability (range). A progressive increase in reflectance intensity is observed in the longer wavelength range (>600 nm) for the oldest age group (80 s), particularly in the S1 region.

Figure 5

Comparative spectral reflectance profiles distinguishing Normal, Mild Cognitive Impairment (MCI), and Dementia groups. The graphs illustrate the mean spectral intensity across five retinal regions: (a) F, (b) I1, (c) I2, (d) S1, and (e) S2. Solid lines represent the mean reflectance, while shaded areas indicate the standard deviation. A statistically significant increase in reflectance is evident in the long-wavelength range (600–780 nm) for the Dementia group compared to the Normal and MCI groups, particularly in regions S1 and I2.

Figure 6

Spectral reflectance profiles of female participants stratified by cognitive status. Comparison of Normal, MCI, and Dementia groups across five retinal regions: (a) F, (b) I1, (c) I2, (d) S1, and (e) S2. Similarly to the overall population, female participants with dementia exhibit higher spectral reflectance in the longer wavelengths (>650 nm), with the most distinct separation observed in the inferior (I2) and superior (S2) temporal arcades.

Figure 7

Comparison of retinal spectral variations across Diabetic Retinopathy (DR) severity stages. The curves show spectral reflectance for Normal controls, Background DR (BDR), Pre-proliferative DR (PPDR), and Proliferative DR (PDR) across regions (a) F, (b) I1, (c) I2, (d) S1, and (e) S2. A progressive elevation in long-wavelength reflectance correlates with increasing disease severity, serving as a comparative reference for spectral changes driven by retinal pathology.

Figure 8

Classification accuracy comparison of four Deep Learning models on Original Retinal Images (ORI) versus Hyperspectral Images (HSI). The bar chart displays the overall accuracy achieved by ResNet50, GoogLeNet, Inception_v3, and EfficientNet. HSI-based models (orange bars) consistently outperform ORI-based models (blue bars), with ResNet50 achieving the highest accuracy of 84% using hyperspectral data.