Multimodal Retinal Imaging for the Detection of Early Alzheimer’s Disease: Combining Biochemical, Structural, and Vascular Biomarkers

1. Introduction

Dementia is a major public health concern, with more than 150 million people expected to be affected by 2050 due to demographic aging [1]. Alzheimer’s disease (AD), the leading cause of dementia, progresses along a biological continuum that begins decades before clinical symptoms emerge, marked by the silent accumulation of amyloid-β (Aβ) and tau [2]. As the disease progresses, these protein aggregates trigger synaptic dysfunction and neuronal loss, alongside cerebral amyloid angiopathy, neuroinflammation, and gliosis, reflecting the multifactorial nature of the disease [3]. Current global estimates suggest that approximately 315 million individuals are in the preclinical phase (Aβ-positive but cognitively normal, CN), 69 million in the prodromal stage (mild cognitive impairment, MCI due to AD), and over 32 million live with AD dementia (ADD) [4]. Given this large at-risk population, there is an urgent need for effective early-detection strategies. The recent arrival of monoclonal antibodies as a promising therapeutic approach, showing moderate effect in decelerating cognitive decline, further underscores the importance of timely identification and intervention [5]. Moreover, early biomarker-based identification not only facilitates timely lifestyle interventions but also enables clinical trial recruitment, which is crucial for advancing the development of disease-modifying treatments [6,7]. Furthermore, understanding biomarkers trajectories across the disease continuum is crucial for monitoring disease progression [8]. However, current gold-standard biomarkers, such as cerebrospinal fluid analysis (CSF) assays and positron emission tomography (PET), are invasive, expensive, and unsuitable for population-level screening [9]. This has driven efforts towards non-invasive, scalable, and cost-effective alternatives [10], including promising blood-based biomarkers (BBMs) such as plasma phosphorylated tau (p-tau), which demonstrate high diagnostic accuracy [11].

Alternatively, due to its common embryological origin and direct anatomical continuity with the brain, the retina has gained prominence as a non-invasive biomarker site for detecting early AD-related changes [12,13,14]. Among the available retinal imaging modalities, hyperspectral imaging (HSI) offers a unique advantage by capturing both spatial and spectral information (Figure 1a), enabling detection of subtle biochemical alterations that may precede any discernible structural damage [15,16]. Furthermore, both postmortem and in vivo studies have demonstrated that HSI can identify spectral signatures associated with retinal Aβ deposition [15,16,17]. To mitigate the high inter-subject variability of raw spectral reflectance data [18,19], recent studies have incorporated texture features derived from the gray-level co-occurrence matrix (GLCM), which capture spatial relationships between pixel intensities and improve discrimination robustness [19]. Nonetheless, the potential of HSI remains underexplored in AD research. The macula has been suggested as a particularly susceptible site in preclinical AD, with several studies reporting subtle thinning of macular layers in individuals with MCI [20,21,22]. These findings indicate that the macula may exhibit early pathological changes, potentially before widespread neurodegeneration is detectable, which could be more sensitively captured through GLCM-based texture analysis [23]. Yet, no HSI study to date has systematically examined how texture alterations vary across the macula, representing a critical step toward identifying regions of early retinal vulnerability in AD. Moreover, establishing the measurement reliability of such novel features is critical for their validation as clinical biomarkers.

Adding to the biochemical insights from HSI, other complementary retinal imaging biomarkers can capture different aspects of AD pathology, such as structural and vascular changes. Optical coherence tomography (OCT) provides high-resolution cross-sectional imaging of the retina, enabling quantification of structural neurodegeneration through layers such as the retinal nerve fiber layer (RNFL) consisting of ganglion cell axons, the ganglion cell layer (GC) containing their cell bodies, and the inner plexiform layer (IPL) that includes their dendrites and synapses (Figure 1b) [22]. Standard OCT protocols often quantify the latter two as a single unit, the ganglion cell-inner plexiform layer (GC-IPL), which has emerged as a sensitive biomarker of neurodegeneration in MCI and AD [24]. In parallel, vascular biomarkers (VBMs) can shed light on the microvascular dysfunction known to contribute to AD progression [25].

Advances in machine learning (ML) and deep learning (DL) techniques are increasingly used in ophthalmic imaging for automated extraction of quantitative biomarkers and disease classification [26]. Using DL based segmentation of color fundus photography (CFP), VBMs such as the Central Retinal Arteriolar and Venular Equivalents (CRAE/CRVE) can be automatically extracted for large-scale analysis (Figure 1c) [27,28].

Prior studies have demonstrated the added value of combining imaging modalities. For instance, Sharafi et al. [19] used a support vector machine (SVM) to show that supplementing vascular measures with HSI-based texture features improved detection accuracy from 76% to 85%. Similarly, in our previous work, we demonstrated that a linear discriminant analysis (LDA) model combining HSI with OCT features outperformed HSI alone, enabling more accurate classification of individuals with established ADD [29]. Despite these insights, it remains unknown whether this diagnostic ability and synergy extend to the more subtle pathological changes present in early, Aβ PET-confirmed stages of AD. A comprehensive model integrating biochemical (HSI), structural (OCT), and vascular (CFP-based) biomarkers for early AD detection has yet to be systematically developed and validated.

To address these knowledge gaps, we conducted a proof-of-concept study focused on both early AD detection and biomarker trajectories. Our primary classification analyses were performed on a well-characterized cohort of CN and MCI participants with PET-confirmed Aβ status. For an exploratory analysis of how biomarkers behave across the disease continuum, this cohort was supplemented with a group of patients with ADD.

Our aims were threefold: (1) to evaluate HIS-derived texture features for classifying Aβ status and to identify topographic patterns of retinal vulnerability; (2) to explore the trajectory of the most salient HSI biomarker across the AD continuum using the augmented cohort; and (3) to systematically compare the diagnostic performance of biochemical (HSI), structural (OCT), and vascular (VBMs) biomarkers, both individually and in combination.

2. Materials and Methods

2.1. Study Participants and Cohort Definitions

To address the study aims, two distinct participant cohorts were included. The primary CN/MCI cohort was used for unimodal and multimodal classification analyses, while an additional CN/ADD cohort was combined with the primary cohort for exploratory analyses of biomarker trajectories across the AD continuum.

The primary CN/MCI cohort consisted of 40 participants recruited for the prospective, longitudinal biomarker study Multimodal Retinal Imaging in the Detection and Follow-up of Alzheimer’s Disease (RetAD) (ClinicalTrials.gov: NCT03466177). We included participants aged 50–85 years with complete multimodal retinal imaging and Aβ PET data. For each participant, a single study eye was selected and used consistently for all imaging modalities and follow-up visits. Exclusion criteria included major neurological disorders other than AD, large-vessel stroke, focal brain lesions on MRI, or ocular diseases other than refractive error or cataract (e.g. glaucoma). Participants underwent a standardized neuropsychological assessment to evaluate global cognition (Mini-Mental State Examination [MMSE]) and specific domains including memory (Auditory Verbal Learning Test [AVLT]), language (Boston Naming Test [BNT]), and verbal fluency (Animal Verbal Fluency [AVF]). Amyloid burden was assessed via static ¹⁸F-Flutemetamol PET scans acquired 90–110 min post-injection [30]. Standardized uptake value ratios (SUVRs) were converted to Centiloid (CL) values, with participants classified as Aβ positive (Aβ+) if CL values > 23.5 and Aβ negative (Aβ−) if CL ≤23.5 [31]. Study protocols were approved by the Ethics Committee of UZ Leuven and were conducted with the EU Directive on Clinical Trials (2001/20/EC) and the Declaration of Helsinki.

For the exploratory trajectory analysis, the primary cohort was supplemented with an additional CN/ADD cohort from our previously published study [29], comprising 22 CN controls and 17 patients with clinically or biomarker-confirmed ADD, with no overlap in participants between the cohorts.

2.2. Image Acquisition and Feature Extraction

2.2.1. Hyperspectral Retinal Imaging (HSI)

Acquisition and pre-processing

Macula-centered HSI was acquired using a MQ022HG-IM-SM4X4-VIS2 snapshot camera (Ximea, Münster, Germany) connected to a Topcon TRC-50DX fundus camera (Topcon Corporation, Japan), capturing 16 spectral bands in a single 0.2 ms exposure (50° field of view (FOV), 460-620 nm, bit depth 10, analog gain 3.2x) [32,33]. Following the acquisition protocol described by Lemmens et al. [29], raw images were pre-processed by computing relative reflectance using dark and white reference frames, followed by application of a spectral correction matrix [34]. This correction applied the sensor’s spectral response model to account for cross-talk between neighboring spectral bands, converting raw digital readings into effective spectral reflectance values and yielding 14 virtual bands per image for analysis.

All participants from both the CN/MCI and CN/ADD cohorts underwent HSI using the same imaging setup and protocol, ensuring full direct comparability of extracted HSI texture features across cohorts.

Quality control and regions of interest (ROIs)

Each pre-processed HSI underwent rigorous visual inspection, including evaluation of all 14 spectral bands, a pseudo-RGB reconstruction approximating a color fundus image, and the corresponding vessel segmentation mask generated by a DL model [27]. Images with poor vessel delineation [35], significant motion blur, or illumination artifacts were excluded from further analysis. For analysis of these high-quality images, a 6 mm circular macular grid was applied, corresponding to the diameter of the Early Treatment Diabetic Retinopathy Study (ETDRS) grid [36]. This grid was precisely centered on the fovea, which was manually annotated by an ophthalmologist on a pseudo-RGB HSI reconstruction. The optic disc (OD) center was automatically detected using a DL model [28]. This macular region was then divided into superior and inferior hemispheres using the line connecting the OD center and fovea as the boundary. To ensure consistent ROI sizing despite slight variations in eye position and field of view, the physical distance between the OD center and fovea (~4.5 mm) was mapped to pixel coordinates. These ROIs were then applied uniformly across all 14 spectral channels, after masking out retinal vessels [27].

Feature extraction

From each retinal vessel-masked ROI, two distinct feature sets were extracted for each of the 14 spectral bands:

• Normalized spectral reflectance: To minimize inter-subject variability arising from imaging factors (e.g. illumination differences), relative reflectance values within each ROI were standardized using the mean and standard deviation of the entire image, as described by Lemmens et al. [29]. The mean standardized reflectance value within each ROI was then calculated, yielding a 14-dimensional spectral feature vector.
• GLCM texture features: To quantify retinal tissue organization, five second-order statistical descriptors were derived from the Gray Level Co-occurrence Matrix (GLCM): contrast (CON), dissimilarity (DIS), homogeneity (HOM), energy (ENE), and correlation (COR) [23,37]. Using the “graycomatrix” function from the scikit-image Python package (version 0.23.2), GLCM features were calculated with a pixel-pair distance of 1 (capturing fine-grained local texture), averaged across four orientations (0°, 45°, 90°, and 135°) to ensure rotational invariance, and normalized to allow comparability across ROIs and imaging conditions.

2.2.2. Optical Coherence Tomography (OCT)

Spectral-domain OCT (SPECTRALIS, Heidelberg Engineering) was performed on a single study eye centered on the macula. The Heidelberg Eye Explorer software automatically extracted RNFL and GC-IPL thickness values across the nine-region ETDRS grid. Subsequently, GC-IPL thickness values were used as features, either from all nine regions combined (for the full macular grid) or from subsets corresponding to the HSI grids (e.g., Inner and Outer Inferior regions for the inferior hemisphere analysis).

2.2.3. Color Fundus Photography (CFP)

Disc-centered 30° field-of-view CFP were acquired using the Visucam PRO NM camera (Carl Zeiss Meditec, Germany). In case of multiple CFPs captured on the same day, the highest-quality image was manually selected for the analysis. Following the same protocol as in Fhima et al. [38], the arterioles and venules were segmented from the CFPs using a DL model, after which VBMs were extracted from the segmentation masks using the Python Vasculature BioMarker (PVBM) toolbox [28]. The CRAE and CRVE were calculated using the Knudtson method [39]. For one subject with discontinuous vessel segmentation, VBMs were imputed using the dataset mean.

2.3. Statistical Analysis

Group Comparisons

Differences in retinal features between Aβ− and Aβ+ participants were evaluated using two-sided Mann–Whitney U tests. As this study represents an exploratory biomarker screen aimed at identifying candidate features for classification, group differences were evaluated using unadjusted p-values without correction for multiple comparisons [40]. We recognize that this increases the likelihood of false positive findings (Type I error), and accordingly, all reported p-values should be interpreted as hypothesis-generating rather than confirmatory. Demographic variables (age and sex) were compared between groups using appropriate statistical tests following normality evaluation. For continuous variables (age), independent sample t-tests were used when data followed a normal distribution, otherwise Mann-Whitney U tests were applied. For multiple group comparisons, the Kruskal-Wallis test was used for continuous variables. For categorical variables, chi-square tests were performed for both pairwise and multiple group comparisons. To assess the overall effect of group differences on structural OCT measures (RNFL and GC-IPL), a multivariate analysis of covariance (MANCOVA) was conducted, with age and sex included as covariates, using Wilks' Lambda to test for significance (p < 0.05).

2.4. Repeatability and Reproducibility

To assess the robustness of the his-derived texture features (e.g. GLCM COR at 466 nm) across the defined macular ROIs, intra-visit repeatability and inter-visit reproducibility analyses were performed. For this, we utilized data from participants within the RetAD cohort who had the necessary repeated image acquisitions. Reliability was quantified using the Intraclass Correlation Coefficient (ICC) based on a two-way random-effects model for absolute agreement (ICC(2,1), pingouin Python package). Repeatability, defined as test-retest consistency across longitudinal timepoints (e.g., baseline vs. 6-month follow-up) with identical flash settings, was assessed using 20 unique subject-flash pairs from 14 subjects. Reproducibility, defined as consistency across different flash intensities within a single visit, was assessed using 31 distinct image pairs from 22 subjects. The number of image pairs exceeded the number of individual participants because some individuals contributed multiple eligible data points, for instance at different follow-up visits or with various acquisition settings.

2.5. Biomarker Trajectory Across AD Stages

For exploratory trajectory analysis, the combined cross-sectional cohort was organized into three distinct groups: (1) CN controls, comprising all Aβ− CN individuals from the primary cohort and the CN controls from the additional cohort; (2) early Aβ+ individuals, representing the preclinical (CN) or prodromal (MCI) group from the primary cohort; and (3) AD patients with clinical or biomarker-confirmed ADD from the additional cohort [29]. Comparison of the GLCM COR at 466 nm (full macular grid) were performed across these three groups. Group differences in age and sex were also assessed to evaluate potential confounding effects.

2.6. Classification and Model Evaluation

Logistic regression classifiers were trained to predict Aβ status (Aβ− vs. Aβ+) using various sets of input features, including GLCM COR (466 nm), VBMs, and GC-IPL thickness measures, both individually (unimodal) and in combination (multimodal). To obtain an unbiased performance estimate, model performance was evaluated using a leave-one-subject-out cross-validation (LOOCV) framework. Within each fold, feature scaling (z-score normalization) was applied using only the training set statistics to prevent data leakage. The logistic regression models were trained with L2 regularization, the liblinear solver, a maximum of 1000 iterations, and balanced class weights. Performance metrics included the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, precision, and the F1-score. Uncertainty was quantified using 95% confidence intervals for the AUC, estimated via 1000 bootstrap resamples.

3. Results

3.1. Participant Characteristics

The primary CN/MCI cohort consisted of 40 adults with PET-confirmed Aβ status. Seventeen participants were Aβ+, including 10 CN and 7 MCI, representing preclinical and prodromal AD stages [2]. The remaining 23 participants were Aβ− CN controls. There were no significant differences in age or sex (Table 1). Cognitive performance differed primarily in the Aβ+ MCI subgroup, which showed lower MMSE scores and worse performance on memory and language measures (auditory verbal learning test, Boston Naming Test, and animal verbal fluency) compared with CN participants. APOE genotype distributions showed a trend toward differences across subgroups. Additional subject-level demographics and cognitive test scores are provided in Supplementary File 2. Demographic characteristics for the additional ADD cohort have been described previously [29].

3.2. Hyperspectral Imaging Texture by Macular Region and Amyloid Status

The investigation of retinal biochemical alterations linked to amyloid pathology showed that normalized reflectance spectra demonstrated substantial overlap between Aβ− and Aβ+ participants in all regions, providing little discriminatory value. In contrast, texture features revealed a consistently different signal (Supplementary Figure 1). Among the five GLCM descriptors, COR at 466 nm emerged as the most discriminative feature. Aβ+ individuals showed significantly lower COR values, particularly in the inferior macula (p = 0.010) and full macular grid (p = 0.031), while the superior hemisphere did not differ significantly (p = 0.106) (Figure 2). These lower COR values indicate reduced similarity between neighboring pixel intensities and thus more heterogeneous local patterns.

To quantify the diagnostic utility of HSI, logistic regression classifiers were trained using GLCM COR at 466 nm as the input feature for each macular region. The inferior hemisphere model achieved the highest performance (AUC = 0.72), followed by the full macular grid (AUC = 0.69) and the superior hemisphere (AUC = 0.62). Reliability analysis indicated that the full macular grid exhibited the highest intra-visit repeatability (ICC = 0.60), while both the full macular grid and inferior hemisphere had excellent inter-visit reproducibility (ICC = 0.88) (Table 2).

3.3. Hyperspectral Imaging GLCM COR Across the AD Continuum

To examine retinal texture changes with disease progression, we assessed GLCM COR at 466 nm across CN controls (n = 45; 23 from the primary and 22 from the supplemented cohort), early Aβ+ individuals (n = 17), and ADD patients (n = 17). A clear stage-dependent trajectory was observed. Compared with CN controls, individuals in the preclinical/prodromal stage (CN/MCI Aβ+) showed significantly lower COR values (p = 0.015), reflecting increased local tissue heterogeneity. In contrast, COR values in the ADD group rose again, approaching levels observed in controls (CN) and showing no significant differences (p = 0.218) (Figure 3). This non-monotonic pattern suggests that the discriminative value of this feature may be strongest during the earliest stages of AD. Demographic comparisons confirmed that these group differences were unlikely to be explained by age (p = 0.644) or sex (p = 0.428) (Supplementary Table 1).

3.4. Optical Coherence Tomography-Based Thickness Analysis

To investigate whether the spatial pattern of macular neurodegeneration was associated with Aβ status, we performed separate MANCOVA analyses for GC-IPL and RNFL thickness, treating the values from all nine ETDRS regions as distinct outcomes (Table 3). This approach revealed a significant association between the overall GC-IPL thickness profile and Aβ status (p = 0.009), with sex also emerging as a significant factor (p = 0.041). In contrast, the RNFL thickness profile showed no association with Aβ status (p = 0.694), though it was significantly influenced by age (p = 0.024). Notable, when comparing GC-IPL thickness within individual ETDRS regions, no single region yielded statistically significant differences (Supplementary Fig. 2).

To assess the diagnostic value of these OCT structural measures, logistic regression models were constructed using GC-IPL thickness as input features. Model performance varied across retinal regions: inferior macular regions (II, IO) showed markedly higher discriminative ability (AUC = 0.71) compared with superior regions (AUC = 0.45). The model incorporating all nine macular regions achieved the highest overall classification performance overall (AUC = 0.73) (Table 4).

3.5. Color Fundus Photography-Based Vascular Biomarker Analysis

Finally, we assessed whether retinal vascular alterations were associated with amyloid status by extracting a comprehensive set of VBMs (Supplementary Table 2). Of the 24 vascular parameters evaluated, only two showed significant differences between the Aβ+ and Aβ− groups: the Aβ+ group had significantly wider arteriolar diameters (CRAE; p = 0.031) and a greater number of venous endpoints (Endp; p = 0.043). The remaining metrics, including those that measure tortuosity and fractal dimension, were comparable between groups (Supplementary Table 2).

To evaluate the diagnostic utility of these vascular changes, we built logistic regression models using the well-established vessel caliber parameters, CRAE and CRVE respectively (Figure 4) [41]. The significant arteriolar widening observed in the Aβ+ group translated into modest discriminatory power (AUC = 0.66). In contrast, the venular diameter (CRVE), which showed no group difference (p = 0.662), offered no classification value (AUC = 0.48). A model combining both metrics achieved the highest VBM-based performance with an AUC of 0.68 (Table 5).

3.6. Unimodal and Multimodal Retinal Imaging Classification Performance

To assess the synergistic added value of integrating multiple retinal imaging biomarkers, we compared the performance of unimodal and multimodal logistic regression models (Figure 5). The best overall classification performance was achieved by combining HSI-derived texture features (GLCM COR at 466, full macular grid) with GC-IPL thickness across the 9 ETDRS regions, yielding an AUC of 0.84, sensitivity of 0.82, and specificity of 0.78. Adding VBMs to this model did not yield further improvement (AUC = 0.81). Detailed classification results for all feature combinations across macular regions (full macular grid, superior, and inferior) are provided in Supplementary Table 2.

4. Discussion

4.1. Hyperspectral Imaging Texture Reflects Early Amyloid-Related Biochemical Changes

This study provides proof-of-concept evidence that multimodal retinal imaging, integrating HSI texture and OCT-derived GC-IPL thickness, can predict PET-confirmed amyloid status in the preclinical and prodromal stages of AD. This synergy suggests that different imaging modalities capture distinct, complementary aspects of early retinal pathology.

Notably, texture-based HSI features outperformed normalized spectral reflectance in discriminating Aβ status. The increased heterogeneity (lower GLCM COR) observed in Aβ+ individuals, particularly in the inferior macula, points to localized retinal disorganization. Although the full macular grid yielded slightly lower classification performance, it demonstrated higher repeatability over 6-months’ time intervals, indicating a potential trade-off between localized sensitivity and biomarker robustness. While Lemmens et al. [29] found spectral features discriminative in patients with ADD, Hadoux et al. reported that spectral data alone could not distinguish amyloid status, prompting them to apply dimensionality reduction to correct for non-AD related spectral variability [18]. Beyond spectral information, other studies have shown the added value of spatial patterns. The diagnostic potential of our texture-based approach is supported by the work of Sharafi et al. [19], who also found that GLCM texture features could discriminate Aβ status, albeit from perivascular regions. Collectively, these findings reinforce that robust spatial-spectral methods are needed to overcome low interclass discrimination and high intraclass variance inherent to pixel-wise spectral analysis in HSI [42,43].

While some histopathological studies report denser Aβ deposits in peripheral regions such as the superior-temporal and inferior-temporal retina [14,44,45], others have identified Aβ deposits in the central retina, particularly in the perifoveal ganglion cell layer [46]. In practice, however, taking peripheral retinal pictures in a consistent and reproducible way is more challenging than central retinal pictures, particularly in elderly individuals and patients with cognitive impairment, making central pictures a more attractive and feasible option. Aβ depositions has also been observed along retinal vasculature [47,48], raising the possibility that perivascular pathology may also contribute to our texture signal. GLCM COR quantifies the correlation between the gray levels of neighboring pixels in the co-occurrence matrix, with higher values indicating more similarity and lower values reflecting greater variability. Lower GLCM COR at 466 nm in Aβ+ participants indicates greater tissue heterogeneity and local disorganization. This finding is consistent with ex vivo work by Zaletel et al. [49], who reported that dense-core amyloid plaques in AD brain tissue exhibit lower GLCM COR values than diffuse plaques in non-AD tissue.

The most discriminative HSI-derived texture feature occurred at 466 nm in both the full macular grid and the inferior hemisphere, which corresponds to the shortest wavelength captured by our snapshot camera (Supplementary Figure S1). This finding aligns with prior literature linking Aβ accumulation to increased light scattering and altered reflectance in the blue-light spectrum. Vince and More reported reduced short-wavelength intensities (480–560 nm) in transgenic AD mice months before plaque formation, attributing this to Rayleigh scattering by sub-wavelength soluble Aβ [15]. Rayleigh scattering is more pronounced at shorter wavelengths, making this region particularly sensitive to small-scale tissue alterations, such as Aβ deposition, which can lead to localized changes in reflectance. Subsequent in vivo and clinical studies similarly identified strongest spectral deviations below 500 nm [16,17,50,51], while Hadoux et al. found the most discriminative bands below 565 nm [18]. Sharafi et al. extended these insights to spatial texture features, finding that the strongest group differences in retinal GLCM metrics emerged in the 450–550 nm range [19], further supporting 466 nm as a particularly informative wavelength for Aβ-related tissue disorganization.

Other pathological processes including deposition of phosphorylated tau, early neurodegeneration, or inflammatory responses may underlie the observed texture changes [44,52,53]. Phosphorylated tau has been detected in post-mortem retinas of AD patients [54,55], with recent evidence showing distinct light scattering signatures for both Aβ and tau [56]. Importantly, the presence of retinal Aβ remains contested: while many studies confirm its presence, others report negative findings [57,58,59]. Future research should validate these texture features against histopathology to determine their pathological specificity, clarify regional differences, and support the development of anatomically targeted imaging strategies.

4.1.1. Non-Monotonic Biomarker Dynamics Across the AD Continuum

Our exploratory analysis suggested that HSI-based retinal texture features may follow a non-monotonic, U-shaped trajectory across the AD continuum. This pattern is consistent with findings from More et al., who reported the largest spectral deviation from control subjects at the MCI stage (MMSE ⩾22), with the discriminative signal progressively weakening in more advanced dementia [16]. They attributed this peak sensitivity in early disease to small, soluble Aβ oligomers that are potent sources of Rayleigh scattering, while later stages feature larger and fibrillar Aβ. The initial decrease in GLCM COR in our cohort might thus reflect the increased local texture variation from this peak scattering effect, while we speculate the subsequent rise in the ADD group represents a more optically uniform texture resulting from the shift to fibrillar Aβ.

While this hypothesis provides a compelling biological rationale, our findings should be considered mainly exploratory. The analysis combined two cross-sectional cohorts with modest sample sizes, highlighting the need for validation in larger, longitudinal studies. Furthermore, while the imaging protocol was consistent, the potential influence of subtle confounders, such as variations in acquisition settings or noise on texture metrics, cannot be fully excluded. Our results contrast with those of Jin et al., who analyzed texture features within retinal layers (GC-IPL, OPL, OS, RPE) using cross-sectional (B-scan) OCT and reported a monotonic decline in GLCM COR features from CN to MCI to AD [23]. This divergence may reflect the differences in what both different imaging modalities and different imaging planes (en face and cross-sectional) capture. Our en face HSI provides a planar projection map that appears sensitive to the biochemical Aβ changes. In contrast, cross-sectional (B-scan) OCT yields high-resolution, depth-resolved measurements within isolated layers, which may probe a different facet of pathology and therefore produce a different trend.

These stage-specific patterns across the AD continuum have important implications. Models trained at one disease stage may not generalize to others. Future work should therefore develop stage-specific algorithms and prioritize longitudinal cohorts to map the full evolution of retinal biomarkers across the AD continuum.

4.2. Macular GC-IPL Measures Suggest Retinal Neurodegenerative Changes

Our structural analysis with OCT complements our biochemical findings, identifying the macular GC-IPL, rather than the macular RNFL, as the primary site of amyloid-related thickness changes. This reinforces previous findings that the macular GC-IPL is a more sensitive marker for AD-related pathology, likely due to its high ganglion cell density and lower inter-individual variability [22]. Consistent with this, macular thinning has been reported in cognitively unimpaired individuals with risk factors for AD (e.g. APOE ε4, family history) [60]. Moreover, GC-IPL thinning (but not RNFL) has been associated with higher pTau burden [61] and thinner GC-IPL at baseline carries an increased risk in cognitive decline [62].

At the same time, evidence from Jali et al. highlights a different challenge: although they observed progressive GC-IPL (p < 0.001) and RNFL (p = 0.030) thinning across the continuum from normal cognition to dementia, average GC-IPL (AUC = 0.60) and RNFL (AUC = 0.56) measures showed limited discriminatory utility for distinguishing normal or mildly cognitively impaired individuals from those with more advanced disease stages [63]. This limited utility may stem from non-uniform retinal thickness changes in AD, as studies have shown dynamic patterns with GC-IPL and RNFL thickening occurring alongside thinning, possibly reflecting glial hyperplasia in early disease stages [64,65]. This variability underscores the limitations of relying solely on average thickness measures for early AD screening, as subtle yet clinically relevant changes might be missed. In our cohort, using all nine GC-IPL regions as input features achieved the strongest OCT-based classification performance (AUC = 0.73). By contrast, a model using a single global GC-IPL thickness feature, using the mean of those same nine sectors, resulted in classification performance around chance level (AUC = 0.45), suggesting that spatial heterogeneity provides discriminative signal that is lost when averaging across larger areas.

Interestingly, we observed regional asymmetry in discriminative power. The inferior ETDRS regions (II and IO) yielded an AUC of 0.71, while the superior regions (SI and SO) performed poorly (AUC = 0.45). This spatial bias mirrored our HSI findings, with both modalities converging on an “inferior vulnerability zone”, where inferior regions consistently showed stronger group differences. These patterns resemble those seen in glaucoma, suggesting a potential shared regional susceptibility across glaucoma and AD [66,67].

In our cohort, age was significantly associated with macular RNFL but not GC-IPL thickness, while sex was significantly associated with GC-IPL but not RNFL. This pattern is somewhat unexpected, as previous studies generally report both age- and sex-related differences across several macular layers [68,69]. Overall, the literature indicates that macular thickness measures are influenced by demographic factors, underscoring the importance of adjusting for these confounders. The divergent associations we observed may therefore reflect cohort-specific variations, such as a narrow age range and our modest sample size, rather than a true biological absence of effects.

Beyond thickness, OCT images can also be analyzed using texture metrics. In a recent study, OCT-based GLCM features outperformed average thickness measures in distinguishing AD and MCI from controls (AUC = 0.935 CN/AD; AUC = 0.830 CN/MCI vs. 0.795 and 0.705, respectively) [23]. Furthermore, emerging DL models that process full GC-IPL or RNFL thickness maps directly, rather than relying on handcrafted features, have shown promising results in AD and MCI classification [70]. Despite promising performance, the lack of large, publicly available OCT datasets remains a major limitation for DL, with most studies based on small, private cohorts collected across heterogeneous clinical settings [71].

4.3. Wider Arteriolar Diameters Indicate Early Vascular Involvement

Our comprehensive analysis of 24 VBMs revealed that most were not associated with amyloid status, with the notable exception of wider retinal arterioles (CRAE) in the Aβ+ group. When used as a combined input, CRAE and CRVE yielded modest classification performance. Our results align with Sharafi et al. [19], who also reported wider arteriolar diameters in Aβ+ subjects, suggesting early vascular changes may accompany biochemical and structural alteration. Conversely, a meta-analysis pooling results from four studies reported no significant association between CRAE or CRVE and cognitive impairment [72]. The analysis revealed substantial heterogeneity between studies (I² = 93%), pointing to possible differences in study design, disease stage, imaging protocols, and vascular measurement techniques as contributing factors. Further nuance comes from research investigating both OCT angiography (OCTA) and dynamic vessel analysis (DVA) in MCI and ADD subjects. Whereas no significant structural vessel changes were observed, impaired vascular reactivity to flicker stimulation was detected in individuals with MCI, suggesting that functional alterations may precede morphological changes in retinal vessels [73]. Functional vascular imaging, such as DVA, may offer greater sensitivity to early AD-related changes and should be explored further. In this context, we are currently collecting DVA data in the ongoing RetAD trial, which will be analyzed in future work to further elucidate early functional vascular alterations in AD.

4.4. Multimodal Retinal Imaging Maps onto the ATV(N) Framework

The AT(N) framework classifies AD biomarkers by the pathological processes they measure: Aβ deposition (A), pathologic tau (T), and neurodegeneration (N) [74]. Our results indicate that HSI may reflect biochemical changes related to amyloid (A), while others have found HSI scattering signatures specific to tau (T) [56]. OCT thickness measures, in turn, are indicative of neurodegeneration (N), highlighting how each retinal imaging modality targets a distinct facet of AD pathology.

Although adding VBMs to the multimodal model did not further improve classification performance in this cohort, these findings are hypothesis-generating and suggest that retinal vascular alterations may still warrant further investigation in larger studies. Notably, the 2024 revised criteria for AD diagnosis and staging expanded the classic AT(N) framework to include vascular brain injury (V) as a marker of non-AD co-pathology [2]. Supporting this, Chun et al. showed that vascular burden, defined by severe white matter hyperintensities, significantly accelerates cognitive decline, particularly in individuals with early Alzheimer’s pathological changes (A+T−) [75]. Such evidence underscores the value of VBMs as potential early indicators of risk [76]. Within this context, our multimodal approach provides a strong biological rationale for integrating complementary retinal imaging modalities, offering a more comprehensive view of retinal involvement in AD and improving the specificity of screening strategies [74,75].

4.5. Strengths and Limitations

A key strength is the use of a well-characterized cohort in the earliest preclinical and prodromal stages of AD, where scalable screening tools are most needed. Critically, amyloid PET imaging served as the gold-standard reference for AD pathology, enhancing the diagnostic validity of the outcome labels. Moreover, the use of a snapshot HSI camera allows for rapid imaging acquisition, reducing potential motion artifacts and screening time [33].

The primary limitation is the modest sample size, which increases the risk of overfitting and is reflected in the wide AUC confidence intervals. Therefore, our findings require external validation in larger cohorts to confirm their generalizability. Furthermore, as our study excluded individuals with other neurodegenerative or common retinal diseases (e.g. Parkinson’s disease or glaucoma), the specificity of these biomarkers remains untested. Future research should test whether the observed retinal changes are specific to AD or also present in other conditions, which is critical for clinical translation.

5. Conclusions

Multimodal retinal imaging, combining HSI-derived texture features and OCT-based structural measurements, shows promise as a non-invasive screening tool for early-stage AD. Integrating complementary biomarkers improved diagnostic accuracy, potentially capturing distinct aspects of AD pathophysiology. However, confirming the robustness and specificity of these findings will require validation in larger, longitudinal cohorts that include other neurodegenerative and ocular diseases. Upon such validation, this approach could become a scalable and cost-effective tool for population screening and clinical trial enrichment, leveraging existing ophthalmic infrastructure.