Hyperspectral Imaging for the Classification and Analysis of Drusen in Age-Related Macular Degeneration

Abstract

Purpose

To characterize the spatial-spectral signatures of drusenoid deposits [soft, hard, and reticular pseudodrusen (RPD)] in eyes with non-exudative age-related macular degeneration (AMD).

Methods

In this observational, cross-sectional, prospective study, hyperspectral retinal images and corresponding Optical Coherence Tomography (OCT) images were collected from 152 eyes of 100 participants with non-exudative AMD. Drusenoid deposits were manually annotated in 148 eyes (97 participants) by retinal specialists using OCT as the reference standard. Automated drusen identification was then performed using spectral angle mapper (SAM) and random forest classifiers.

Results

Hyperspectral imaging revealed that RPD was most distinct at 470–500 nm, soft drusen at 525–575 nm, hard drusen at 620–675 nm, and pigmentary abnormalities at 650–725 nm. Using the SAM classifier, drusenoid deposits were identified with 97% sensitivity and 70% specificity. The random forest classifier achieved higher overall performance, with 88% sensitivity, 99% specificity, and an area under the receiver operating characteristic curve of 0.96. Pixel-wise SAM mapping delineated drusenoid deposits and healthy retina, extending beyond annotated regions.

Conclusions

Distinct spatial-spectral signatures for soft drusen, hard drusen, and RPD were identified using hyperspectral imaging. Both SAM and random forest classifiers demonstrated promising results for automated drusenoid deposit detection. These findings support the role of hyperspectral retinal imaging as a noninvasive tool for automated drusen mapping and patient risk stratification in nonexudative AMD.

Age-related macular degeneration (AMD) is a leading cause of moderate to severe vision impairment worldwide.^1,2 Clinically, AMD can be classified into early, intermediate, and advanced stages.^1,3 The early and intermediate stages are nonexudative and characterized by the presence of drusenoid deposits—extracellular accumulations that appear as yellowish-white, round lesions within the retina. The advanced stage is marked by the presence of either neovascularization or geographic atrophy (GA).^1–3 Intravitreal injections of anti-vascular endothelial growth factor (VEGF) drugs slow the exudative process, effectively reducing progression to blindness and, in some cases, restoring vision.^4,5 However, neovascular or exudative AMD account for only 10%–15% of all cases.⁶ Recent advancements in GA treatment targeting the complement pathway, mark significant progress in managing non-exudative AMD.^7–9 These developments bring earlier stages of the disease into focus, encouraging long-term interventions and innovations in imaging techniques, artificial intelligence applications, and efficacy assessment endpoints.

Drusenoid deposit can be classified into sub-RPE drusen (hard and soft drusen) and subretinal drusenoid deposits (reticular pseudodrusen [RPD]). Soft drusen have indistinct edges, whereas hard or small drusen present with well-defined boundaries. These types differ not only in morphology but also in biochemical composition and clinical relevance: hard drusen are commonly observed in aging eyes but generally do not confer significant risk for AMD progression.^10,11 In contrast, RPD, located above the RPE in the subretinal space, represent a distinct entity that is associated with increased risk of both GA and exudative AMD.¹²

Current in vivo imaging modalities for retinal drusen, such as spectral-domain optical coherence tomography (OCT),¹³ fundus autofluorescence,¹⁴ adaptive optics scanning light ophthalmoscopy,¹⁵ and color fundus photography,^13,15 have provided valuable insights into drusen morphology and distribution, yet each has inherent limitations. Fundus autofluorescence can yield ambiguous interpretations,¹⁶ and adaptive optics scanning light ophthalmoscopy, although capable of high-resolution imaging, remains technically complex and less accessible in clinical settings.¹⁷ OCT offers cross-sectional visualization of the retina, enabling precise localization and measurement of drusenoid deposits.¹³ However, its inherent anisotropic spatial resolution and limited field of view often restrict comprehensive en face mapping and quantification, especially for small or peripheral deposits. Moreover, OCT captures morphological features without revealing the biochemical composition of these lesions. These constraints highlight the need for complementary imaging approaches, such as hyperspectral imaging, to achieve a more complete characterization of drusen in vivo.

The Optina-4C hyperspectral retinal camera (Optina Diagnostics, Montreal, Canada) captures images across a continuous spectral range from visible to near-infrared light (450–905 nm) at 5 nm intervals, producing 92 narrowband images within approximately one second. Unlike conventional fundus imaging, which relies on three broad spectral bands (red, green, and blue), the Optina-4C acquires sequential images across discrete illumination wavelengths, providing fine spectral reflectance data. Similar to recent advances in wavelength-dependent retinal assessment, such as in electroretinography, which have demonstrated that specific spectral conditions yield clearer insights into retinal function,¹⁸ discretizing the illumination spectrum allows for the precise isolation and quantification of constituent molecular compounds, providing comprehensive insights that go beyond broad-spectrum methods. The spectral richness of this modality allows quantitative analysis of molecular content (e.g., hemoglobin, melanin), cellular organization (e.g., capillaries and nerve fiber layer structure), and tissue density or thickness (e.g., neurodegenerative changes). Consequently, hyperspectral imaging extends beyond anatomical visualization to reveal phenotypic signatures arising from both structural and metabolic changes reflected as functional alterations in retinal tissue.

Recent studies^19–21 have explored hyperspectral retinal images composed of 16 spectral bands (460–630 nm) combined with image feature extraction, such as texture analysis, to classify drusen from non-drusen areas, showing superiority over conventional RGB retinal images.²² Lee et al.²³ proposed combining hyperspectral imaging with a non-negative matrix factorization approach to characterize the constituent compounds of drusen and macular pigment. However, drusen types were not differentiated in these studies.

Using multimodal imaging and a Beer-Lambert law–based light absorption model, a hypothesis was proposed to explain the distinct visual appearance of different drusenoid deposit types.¹³ Hard and soft drusen, located beneath the RPE, strongly absorb short-wavelength (blue) light, producing a bright red channel and a dim blue channel in RGB fundus images. In contrast, RPD, situated above the RPE, exhibit minimal short-wavelength absorption and therefore appear more prominent under blue illumination.¹³ Because RPD are associated with progression to GA and neovascular AMD, accurate identification or detection can assist clinicians in providing personalized management, including more frequent follow-up. Thus we hypothesized that spatial-spectral features derived from hyperspectral imaging can discriminate among different drusenoid deposit types, offering new opportunities for improved diagnosis and monitoring.

Methods

Study Design

This observational, cross-sectional, prospective study included participants with nonexudative AMD in at least one eye presenting at the ophthalmology clinic Clinique Ophtalmologique 2121 (Montreal, Canada) between November 2022 and August 2023. The research adhered to the tenets of the Declaration of Helsinki and complied with the Health Insurance Portability and Accountability requirements and the Personal Information Protection and Electronic Act. Ethical approval was obtained from the Advarra Institutional Review Board (protocol no. 00065167), and all participants provided written informed consent before enrollment. Participants underwent hyperspectral retinal imaging and OCT, with OCT serving as the gold standard for identifying and classifying drusenoid deposits.

Participants

Inclusion criteria on the day of the examination includes (1) age ≥50 years old and (2) nonexudative AMD in at least one eye, with at least one type of the following retinal drusenoid deposits: hard drusen, soft drusen, RPD. The presence of multiple small drusen (<63 µm) or intermediate drusen (≥63 µm and ≥125 µm) is sufficient to qualify for this study.²⁴

Exclusion criteria includes any ophthalmologic condition that would prevent obtaining retinal imaging and/or could interfere with the analysis of the hyperspectral images and OCT images, including (1) pupil dilation contraindicated, (2) inadequate pupil dilatation (< 6mm diameter) preventing uniform illumination of the retina with the hyperspectral imaging camera, (3) nonexudative AMD presenting only with pigmentary changes or GA (no drusenoid deposits), (4) presence of GA in a cumulative area of >0.5 disc area, (5) presence of neovascular/exudative AMD, (6) signs of vascular occlusion or retinopathy (microaneurysm, exudate, hemorrhage, or edema) within a diameter of 10 mm from the mid-point between the optic nerve head and the macula (i.e. the area of interest for the hyperspectral imaging), (7) macular dystrophy, (8) nuclear sclerosis ≥3 on the Lens Opacities Classification System III²⁵ or presence of central cortical or central posterior subcapsular cataract, (9) deficient visual fixation (inability to fixate for at least 2 seconds), (10) refractive error outside the range of −15 D to 15 D, (11) corneal or media opacities affecting retinal imaging on a cumulative area > 1 disc area within a diameter of 10 mm from the mid-point between the optic nerve head and the macula, (12) scar, atrophy, nevus, tumor, epiretinal membrane or retinal pucker with a cumulative area > 1 disc area within a diameter of 10 mm from the mid-point between the optic nerve head and the macula, (13) papilledema. Besides, exclusion criteria also include inability to obtain an OCT image centered on the macular region of satisfactory quality for analysis of the drusenoid deposits (based on the ability of the eye specialist performing the image annotation of the drusenoid deposits), and the inability of obtaining at least three images of satisfactory quality with the hyperspectral imaging camera per the Optina Diagnostics quality index software.

Data Collection

All participants underwent an ophthalmic examination, including visual acuity and slit-lamp biomicroscopy, to determine their eligibility and document the condition of the fundus. Following pupil dilation, participants underwent spectral-domain optical coherence tomography volume scan of the macula (6 mm × 6 mm, 128 B-scans) followed by flash-based color fundus photography on a 45° field of view using the Topcon 3D OCT-1 Maestro 2 device (Topcon Corporation, Tokyo, Japan).

Hyperspectral retinal imaging was performed using the Optina-4C to acquire a hyperspectral data cube consisting of 92 images captured at sequential excitation wavelengths of visible and near-infrared light ranging from 450 to 905 nm in 5 nm increments. This acquisition strategy focuses on the excitation spectrum, enabling characterization of wavelength-dependent reflectance signatures of drusenoid deposits. Image acquisition was completed within approximately one second over a circular 31° field-of-view, covering the optic nerve head and the fovea (Fig. 1A). The Optina-4C hyperspectral retinal camera has a nominal pixel pitch of 8.6 µm on the fundus, calibrated based on the method described in the international standard ISO 10940:2009 Ophthalmic instruments—Fundus cameras assuming a standard emmetropic model eye. Individual variations in axial length or corneal curvature were not accounted for, because imaging was used primarily for spatial-spectral analysis rather than precise morphometric measurements. The nonexudative AMD status and stage, as well as information about eye diseases and conditions, were documented for both eyes. Other information that was collected from the participants of this study include whether the participant is on statins, age, gender, ethnicity, and the use of AREDS/AREDS 2 formulation.

Figure 1.

(A) Central view of the Optina-4C: circular 31° field of view including both the fovea and optic nerve disc. (B) Illustration of the pseudo-normalization process involving the division of an image by a gaussian-filtered version of itself.

Hyperspectral Images Preprocessing

The quality of the hyperspectral data cubes was assessed to identify at least one and up to three cubes with the highest quality, based on the absence of image artifact in the macular region. Image preprocessing was performed using proprietary software internally developed by Optina Diagnostics. The selected hyperspectral retinal images were first normalized to correct for spatial and spectral variations in retinal illumination using two reference hyperspectral image cubes—one captured from a model eye that has a “fundus” with a known reflectivity and the other with a light trap installed at the location of the eye. The spectral calibration step includes a correction for the native spectral gradient of 5nm between the central line and the top and bottom lines of the raw image (frame). Interpolation of the intensities between subsequent frames was used, and this resulted in a normalized hyperspectral cube with 91 images (frames) where all the lines of the image have the exact same wavelength. Additionally, images in different wavelengths were aligned to correct for any eye movement that may have occurred during the image acquisition. The alignment of the images was done using a non-rigid image registration. Therefore each pixel of the normalized and registered dataset contained a meaningful spectrum representative of its local reflectivity. Moreover, a denoising step in spatial-spectral direction were applied to reduce the noise level coming from the camera (Supplementary Figs. S1A, S1B).

Annotation of Drusenoid Deposits

Following preprocessing, hyperspectral data cubes were visually inspected alongside corresponding OCT images to identify and annotate drusenoid deposits (soft, hard, and RPD) based on their location and appearance on OCT B-scans. An internally developed segmentation tool was used to standardize annotations. This tool includes two visualization windows: one for selecting the wavelength that optimizes deposit visibility and another for segmentation using B-spline interpolation after placing reference points (Supplementary Figs. S1C, S1D).

Annotations were completed in two steps: (1) a retinal specialist segmented one drusenoid deposit of each type and a 125 µm-diameter background area (or equivalent disjoint areas) per the Early Treatment Diabetic Retinopathy Study grid region.²⁶ If fewer than 10 deposits were initially segmented, additional deposits were annotated until a total of 10 was reached or all available deposits were included; (2) a second retinal specialist reviewed the annotations to ensure accuracy and consistency. This method ensures a representative sample of drusenoid deposit types and sizes for each eye (Supplementary Fig. S1C). The eyes were classified into four groups (Rotterdam classification)²⁷ according to the amount of each type of deposit present in the area of acquisition of the hyperspectral data cube (10 mm diameter from central point between macula and optic nerve, Fig. 1A): (1) small drusen, (2) median distinct soft drusen, (3) large indistinct soft drusen, and (4) RPD (Table 1).

Table 1.

Classification Groups for the Study Eyes

AMD Classification	Stage (Rotterdam Classification)27
Group 1: Small drusen
Only drusen ≤ 63 µm	0b
Pigmentary changes	1b
Group 2: Median distinct soft drusen
Soft drusen (64∼124 µm)	1a
Soft drusen (64∼124 µm) with pigmentary changes	2b
Group 3: Large indistinct soft drusen
Soft drusen (≥125 µm)	2a
Soft drusen (≥125 µm) with pigmentary changes	3
Group 4: Reticular pseudodrusen
Reticular pseudodrusen	2a
Reticular pseudodrusen with pigmentary changes	3

Spatial-Spectral Feature Extraction

Internal Normalization

Hyperspectral retinal images were analyzed to extract spatial-spectral features of drusenoid deposits for classification. To isolate the drusenoid deposits’ spatial-spectral characteristics independently of the spectral characteristics from those of surrounding retinal tissue—where chromophore abundance (e.g., melanin, macular pigment, hemoglobin) and anterior eye structures (e.g., age-related lens changes, which increases absorption of short-wavelength light) may introduce variability—internal normalization based on a pseudo-normalization approach was performed (Supplementary Methods). Each frame of the hyperspectral data cube was normalized pixel-by-pixel using a blurred version of itself, generated via a spatial Gaussian filter with a mask size of 151 pixels (Fig. 1B). After normalization, spectra was expressed relative to the retinal background. The mean intensity values of segmented drusenoid deposits were compared to assess spectral differences.

Spectral Angle Mapper

The spectral angle mapper (SAM) classification method²⁸ was used to classify pixels based on their similarity to reference spectra. In this study, the similarity angle between the reference spectrum and the pixel spectrum was calculated as follows:

$$\begin{array}{c}\hfill \alpha ={cos}^{-1}\left(\frac{{\sum }_{i=0}^n{X}_i{Y}_i}{{\sum }_{i=0}^n{X}_i^2{\sum }_{i=0}^n{Y}_i^2}\right)\end{array}$$

where X is the reference spectrum and Y is the pixel spectrum, n is the wavelength band, and α represents the angle between them. It calculates the spectral angle between the pixel and reference spectra, where smaller angles indicate greater similarity. SAM was applied to segmented and non-segmented areas to classify drusenoid deposits versus the retinal background.

Random Forest Classifier

A random forest classifier²⁹ was also evaluated for the binary classification of drusenoid deposits versus the retinal background. The model was trained using pseudo-normalized hyperspectral data, with 91 spatial-spectral features derived from the average intensity of segmented regions across the spectral range. Despite the limited dataset for some drusenoid types, this evaluation may provide insight into its utility for classification.

Statistical Analysis

The demographic data for the study population and subgroups were summarized by descriptive statistics. For classification techniques, the dataset was split 80:20 by participant ID to ensure that images from the same individual did not appear in both training and testing sets. Although stratification by drusen type was not explicitly applied, the distribution of drusen subtypes was similar between the training and test datasets. The classification performances were reported by area under the receiver-operating curve (AUROC), sensitivity and specificity.

Results

Cohort and Demographic Characteristics

A total of 112 participants were enrolled in the study. After excluding 10 screen failures and two withdrawals, 100 participants with 152 eligible eyes were included in the final analysis. The participants had an average age of 75.5 ± 7.1 years, with 69% female and 93% identifying as white. Among the participants, 52 had both eyes imaged, 20 had only the right eye imaged, and 28 had only the left eye imaged. Additionally, 75 participants reported using AREDS or AREDS 2 formulations, while 42 reported taking statins. Demographic details and other collected information are summarized in Table 2.

Table 2.

Demographic Characteristics of the Study Participants (n = 100)

Age (years) mean ± SD	75.5 ± 7.1
Female	69 (69%)
Ethnicity
White	93 (93%)
Hispanic or Latino	2 (2%)
Asian	1 (1%)
Black or African American	1 (1%)
Other	3 (3%)
Study eye
Both eyes (OU)	52 (52%)
Only right eye (OD)	20 (20%)
Only left eye (OS)	28 (28%)
Medications
Use of statins	42 (42%)
Use of AREDS/AREDS 2	75 (75%)
Unknown use of statins or AREDS/AREDS 2	3 (3%)

AREDS, Age-Related Eye Disease Study.

During the annotation process, four eyes from three participants were excluded from the analysis because of insufficient OCT image quality, as determined by the retinal specialists. Consequently, the reported results are based on 148 eyes from 97 participants. Table 3 summarizes the distribution of segmented eyes according to drusenoid deposit classification and AMD stage based on the Rotterdam classification.²⁷ The majority of annotated eyes (95%) exhibited large soft drusen, although all drusenoid deposit types were represented, with RPD being the least prevalent, observed in 32% of eyes. Most eyes (86%) displayed more than one type of drusenoid deposit, whereas only 14% had a single type. The vast majority of eyes (97%) were classified as stage 2 or higher on the Rotterdam classification scale.

Table 3.

Number of Segmented Eyes Per Classification Group and Disease Stage

	n = 148
Classification group
Group 1: Small drusen	72 (49%)
Group 2: Median soft drusen	86 (58%)
Group 3: Large soft drusen	140 (95%)
Group 4: Reticular pseudodrusen	48 (32%)
Number of eyes classified
1 group	21 (14%)
2 groups	63 (43%)
3 groups	57 (39%)
4 groups	7 (5%)
Stage (Rotterdam classification)27
0b	3 (2%)
1a	1 (1%)
2a	82 (55%)
3	49 (33%)
4	13 (9%)

Spatial-Spectral Characterization of the Labeled Drusenoid Deposits

Visual Observations of Drusenoid Deposits in Hyperspectral Retinal Images

To identify spatial-spectral features characteristic of different drusenoid deposit types, side-by-side visualization of hyperspectral data cubes and corresponding OCT images was conducted by a retina specialist. This process aimed to locate and identify various drusenoid deposit types within both imaging modalities. In this exploratory analysis, images at different wavelengths were visually examined to determine the spectral ranges that provided the best contrast for drusenoid deposits and pigmentary changes. Figure 2 shows three representative case studies that highlight the key observations. Soft drusen showed the highest contrast in the spectral range close to 570 nm in the hyperspectral images, whereas hard drusen were most prominent around 620 nm. The appearance of soft drusen was significantly attenuated at this higher wavelength (Fig. 2A). As shown in Figure 2B, soft drusen were most visible at 570 nm in the hyperspectral images, while pigmentary abnormalities were best observed at 660 nm. Notably, these pigmentary changes were not as clearly discernible in the color fundus photograph. Figure 2C showed that soft drusen were most visible at 575 nm, hard drusen showed the greatest contrast at 620 nm (with soft drusen and pseudodrusen appearing much attenuated), and RPD were most distinct near 490 nm (where both soft and hard drusen were less visible). The corresponding OCT B-scans are provided in Supplementary Figures S2A–C.

Figure 2.

Hyperspectral imaging revealed distinct wavelength-dependent signatures for different types of drusenoid deposits and pigmentary changes. (A) This eye exhibited multiple soft and hard drusen. In the hyperspectral images, soft drusen showed the highest contrast at 575 nm, whereas hard drusen were most prominent at 620 nm. At this higher wavelength, the appearance of soft drusen was significantly attenuated. (B) This eye displayed soft drusen and pigmentary abnormalities. (a, b) Soft drusen were most visible at 570 nm, whereas (c) pigmentary abnormalities was best observed at 660 nm. Notably, these pigmentary changes were not clearly discernible in the color fundus photograph. (C) This eye exhibited soft drusen, hard drusen, and reticular pseudodrusen (RPD). (a) RPD was most distinct at 490 nm (where both soft and hard drusen were less visible). (b) Soft drusen were most visible at 575 nm, (c) hard drusen showed the greatest contrast at 620 nm, with soft drusen and pseudodrusen appearing much attenuated. Dark blue: large soft drusen; purple: medium soft drusen; dark green: pseudodrusen; red: hard drusen; light blue: large soft drusen with pigmentary abnormalities; yellow: background 1; orange: background 2; light green: background 3.

The visual inspection of annotated hyperspectral retinal images revealed the following key findings: (1) Soft drusen exhibited the highest visibility in the spectral range of 525–575 nm, with their contrast diminishing rapidly outside this range. (2) Hard drusen displayed relatively higher intensity compared to other drusenoid deposits in the spectral range of 620–675 nm. (3) RPD were most distinct in the spectral range of 470–500 nm, showing superior visibility relative to other drusenoid deposits. (4) Pigmentary abnormalities, which were often undetectable in color fundus photography, were readily discernible in the hyperspectral images within the spectral range of 650–725 nm. These findings underscore the potential of hyperspectral retinal imaging for quantitative analysis, enabling the extraction of characteristic spatial-spectral features specific to different drusenoid deposit types.

Relative Spatial-Spectral Signature from Internal Normalization

After pseudo-normalization of the hyperspectral images (Fig. 1B), we evaluated the mean spectrum for each segmented area corresponding to different types of drusenoid deposits across the 148 annotated eyes. As shown in Figure 3A, distinct spatial-spectral signatures were observed for each drusenoid deposit type. A strong correspondence between visual observations described above and the extracted spectral signatures was evident. Specifically, soft drusen were most prominent near the peak of their spectral signature in the 525–575 nm range, with their intensity decreasing rapidly outside this range. In contrast, hard drusen, which were less intense than soft drusen (and RPD) in the 525–575 nm range, maintained relatively higher intensity as the wavelength increased to 650 nm. Conversely, when the wavelength was decreased from 550 nm to 475 nm, the intensity of both soft and hard drusen diminished quickly, whereas the reflectance of RPD remained notably stronger. This observation is consistent with the anatomical location of RPD, which lie above the RPE where there is little melanin to attenuate light in the blue region of the spectrum (below 500 nm).

Figure 3.

(A) Mean spectrum of the different type of drusenoid deposits averaged over 148 eyes (left) and the same spectrum normalized between 0 and 1 (right). (B) Confusion matrix of the classification based on the 4 classes (soft drusen, hard drusen, reticular pseudodrusen and background) performed by SAM technique. (C) Confusion matrix of the binary classification (drusenoid deposit of any type and background) from the SAM technique, with a sensitivity of 97% and a specificity of 70%. (D) Confusion matrix of the binary classification (drusenoid deposit of any type and background) from the random forest model. (E) The random forest classifier achieved excellent performance with a sensitivity of 88%, specificity of 99%, and an AUROC of 0.96.

Drusen Classification Using SAM

We applied the SAM technique to classify pixels within segmented areas. The dataset was split into 80% for training and 20% for testing. To ensure independence of the testing dataset, the mean spectral curves for the training data (118 eyes) were used as reference spectra (Supplementary Fig. S1E). These curves closely matched those from the full dataset (Fig. 3A), confirming signal stability.

Each pixel was classified based on the smallest angle between its spectrum and the reference spectra. The label for each segmented area was determined by the most prevalent pixel label. A smaller weight (0.5) was assigned to background pixels, which were present along the boundaries of some drusenoid deposits. This process was repeated for all segmented areas.

To evaluate classifier performance, we visually assessed the test dataset (30 eyes), presenting five representative cases (Figs. 4A–E). Each case shows the annotated image (left), predicted labels for each pixel (middle), and the global label for each area (right).

Figure 4.

Comparison of the annotated labels by retinal specialists (left), the predicted labels by SAM on a pixel basis (middle), and the aggregated prediction based on label prevalence in the segmented areas (right, the global label). (A) Some pixels within the large soft drusen were misclassified as background in the SAM predicted labels, but the global labels correctly identified all soft drusen. (B) The annotated images presented nine RPD and one with pigmentary changes. The SAM predictions correctly identified seven out of nine RPD and the background labels. The larger RPD was misclassified as a soft druse. (C) The annotated image showed both soft drusen and RPD, the SAM predictions revealed some large soft drusen as background or RPD. However, global predictions correctly identified all three RPD and five of seven soft drusen. (D) The annotated image presented 10 large soft drusen, two of which had pigmentary changes. SAM predictions identified some soft drusen pixels as background or RPD. Global predictions correctly identified nine out of 10 large soft drusen, with one misclassified as RPD. (E) A case with nine large soft drusen in the annotated image, pixel-level predictions misclassified some soft drusen pixels as background or RPD. However, global predictions correctly identified eight out of nine large soft drusen, with one misclassified as background.

Figure 4A showed that some pixels within the large soft drusen were misclassified as background in the SAM predicted labels, but the global labels correctly identified all soft drusen. Figure 4B presented nine RPD and one with pigmentary changes in the annotated image. The SAM predictions correctly identified seven out of nine RPD and the background labels. The larger RPD was misclassified as a soft druse. Figure 4C showed a case with both soft drusen and RPD in the annotated image, the SAM predictions showed some large soft drusen as background or RPD. However, global predictions correctly identified all three RPD and five of seven soft drusen. Figure 4D presented 10 large soft drusen, two of which had pigmentary changes. SAM predictions on a pixel basis identified some soft drusen pixels as background or RPD. Global predictions correctly identified nine out of ten large soft drusen, with one misclassified as RPD. Figure 4E showed a case with nine large soft drusen, pixel-level predictions misclassified some soft drusen pixels as background or RPD. However, global predictions correctly identified eight out of nine large soft drusen, with one misclassified as background. The corresponding OCT B-scans are provided in Supplementary Figures S3A–E.

To evaluate the performance of SAM classification on segmented regions in the test dataset, we generated a 4 × 4 confusion matrix comparing true and predicted labels (Fig. 3B). We also assessed classification performance using binary classes: drusenoid deposits (any type) versus background (Fig. 3C). In this binary classification, the algorithm achieved a sensitivity of 97% and a specificity of 70%.

Drusen Classification With Random Forest Classifier

The random forest classifier, trained on 80% of the dataset for binary classification (drusenoid deposits vs. background), was tested on the remaining 20% within the segmented areas. The confusion matrix is shown in Figure 3D. The model correctly identified all but one of the 78 background areas, and most drusenoid deposits (262 correctly identified, 37 misclassified). The classifier achieved excellent performance with a sensitivity of 88% and specificity of 99% for the identification of drusen relative to background, and an AUROC of 0.96 (Fig. 3E).

Mapping Drusenoid Deposits Outside the Labels

To assess whether the trained classifiers could provide meaningful predictions for classifying pixels outside the segmented areas, SAM and random forest classifiers were performed.

Pixel Classification Using SAM

We evaluated the SAM technique to classify pixels outside the labeled regions, using four labels (three drusenoid types and background) and an additional label for pixels with near-unity values across wavelengths, indicating background. This analysis was performed with segmentations from 83 eyes, resulting in reference spectra similar to those used previously. Supplementary Figures S4A–D showed four representative cases.

In Supplementary Figure S4A, the algorithm correctly identified most large soft drusen, although their boundaries were classified as hard drusen. Additionally, unsegmented soft drusen were predicted correctly, and surrounding areas were labeled as background. In Supplementary Figure S4B, the three large soft drusen were correctly identified, but the boundary of drusen and regions without drusen were labeled as hard drusen. Further investigation is needed to explain this. Supplementary Figure S4C showed a large confluent soft drusen, predicted as a mix of soft drusen, RPD, hard drusen, and background. The prediction of RPD may relate to pigmentary abnormalities, affecting the spectrum at lower wavelengths. Supplementary Figure S4D displayed large confluent soft drusen, correctly predicted with some overrepresentation of hard drusen pixels.

Pixel Classification With the Random Forest Classifier

We also applied the random forest model, trained for binary classification of drusenoid deposits versus background, to predict regions outside the labeled areas. The mean spectrum of a 6 × 6 pixel window was used as input, with the binary output (drusenoid deposit or background) displayed on an image map for comparison with expert annotations. Supplementary Figure S4E showed promising performance results from the macular region of a subject. Given that only a subset of drusenoid deposits was annotated by retinal specialists in this case, the model successfully identified the annotated lesions and additionally detected other potential soft drusen. Further investigation is warranted to fully evaluate the extent of drusen presence.

Discussion

The biogenesis of drusen and RPD remains incompletely understood, but their role in AMD-related functional loss is well established. In wet AMD, hypoxic stress on RPE cells caused by separation from Bruch's membrane due to soft drusen induces VEGF secretion, promoting choroidal neovascularization.³⁰ Additionally, soft drusen recruit macrophages to the RPE, leading to TNF-α release and subsequent production of angiogenic factors like IL-8 and VEGF by the RPE.^5,31 In geographic atrophy, coalesced soft drusen detach the RPE from Bruch's membrane, resulting in RPE dysfunction, photoreceptor loss, and persistent basal laminar deposits, suggesting their pathogenic role.^32,33 Thickened basal lamina further slows lipoprotein transport, exposing the RPE to lipid stress.¹¹ Although hard drusen are generally associated with normal aging, their abundance correlates with an increased likelihood of soft drusen formation.³⁴

Although RPD can be observed across all stages of AMD and are not specific to the disease, their presence has been associated with an increased risk of progression to late-stage AMD, including both GA and neovascular forms.¹² However, their precise role in disease progression remains to be fully elucidated.³⁵ Studies have identified structural changes in areas of RPD regression, including reduced photoreceptor length and choroidal thinning, potentially caused by RPD extending toward the external limiting membrane and damaging photoreceptors before clearance by Müller glial cells.^36,37

The composition of drusen and RPD has been largely characterized through ex vivo histological studies. Soft drusen and basal linear deposits share a heterogeneous lipid-rich composition, including apolipoprotein A and esterified cholesterol, while RPD contain apolipoprotein E and unesterified cholesterol but lack esterified cholesterol.^37,38 In contrast, hard drusen are more homogeneous, comprising hyaline, eosinophilic material, and lipids.³⁹

A deeper understanding of the composition and formation of drusenoid deposits is essential for developing effective AMD therapies. Early intervention targeting the biochemical pathways underlying these deposits could slow or prevent progression to advanced nonexudative or exudative AMD. In vivo imaging is particularly valuable in this context, as it may also provide biomarkers for therapeutic efficacy. Currently, OCT is the most widely used method in clinical practice capable of distinguishing between different types of drusenoid deposits. By providing cross-sectional views of the retina, OCT reveals the location and size of deposits.¹³ Schlanitz et al.⁴⁰ generated drusen volume maps that revealed longitudinal changes in drusen volume. Both drusen type and area have been associated with the risk of progression to late-stage AMD.⁴¹ However, its limitations include incomplete en face mapping and anisotropic spatial resolution, with B-scan spacing potentially missing smaller deposits. Additionally, OCT's field of view is often restricted to the perifovea, and it offers morphological insights, lacking information about the biochemical composition of drusen.⁴⁰

The hyperspectral retinal imaging system (Optina Diagnostics) is an advanced tool designed for rapid and noninvasive retinal imaging. With its fast scanning speed (∼1 second) and multi-wavelength capabilities, hyperspectral retinal imaging captures detailed spectral data across a wide range of wavelengths and provides rich spectral reflectance data enabling biomolecular insight for a comprehensive assessment of retinal health for both clinical and research applications.^42,43

This exploratory study aims to evaluate the feasibility of identifying spatial-spectral features characteristic of drusenoid deposit types (soft, hard, and RPD) in nonexudative AMD participants. We developed a comprehensive database of hyperspectral retinal images, normalized, registered, segmented, and annotated based on drusenoid deposit types identified in OCT images. The dataset included 152 hyperspectral and corresponding OCT images from 100 participants, with annotations performed by retinal specialists on 148 eyes from 97 participants using OCT as the reference. Visual analysis of the hyperspectral images revealed that RPD were most distinct at 470–500 nm, consistent with previous observations,¹³ soft drusen exhibited peak visibility at 525–575 nm (with contrast decreasing outside this range), hard drusen showed higher intensity at 620–675 nm, and pigmentary abnormalities—often undetectable or difficult to confidently identify in color fundus photography—were clearly visible at 650–725 nm. The pseudo-normalization procedure enabled more reliable extraction of the spectral signatures characteristic of each drusenoid category. By minimizing variability attributable to uneven illumination, fundus pigmentation (including melanin and macular pigment), and ocular media influences such as age-related lens changes, this approach helped reduce potential confounding effects in the spectral analysis.

Classification was performed using SAM and random forest algorithms. Although SAM-based classification provided promising results for pixel-level mapping of drusenoid deposits in the macula, limitations included misclassification of small or confluent drusen, likely due to spectral noise or segmentation errors (Figs. 4A, 4C, 4D). Random forest achieved high sensitivity and specificity (AUROC = 0.96) in distinguishing drusen from background, although the sample size limited multiclass classification. Future studies with larger datasets and multimodal machine learning models could improve accuracy, particularly for distinguishing hard drusen, soft drusen, and RPD.

Hyperspectral imaging provides a noninvasive method for assessing drusen morphology, monitoring AMD progression, and identifying high-risk features such as RPD and confluent large soft drusen.^11,12 Its integration into clinical workflows, alongside OCT and fundus photography, may enhance diagnostic precision and enable personalized management of AMD with improved patient risk stratification. In future clinical applications, a one-second hyperspectral scan could automatically annotate different drusen subtypes through embedded analysis software, offering a spectral-based evaluation of AMD progression risk. The distinct spectral signatures identified in this study highlight the potential of this approach for precise phenotyping. While the current binary classification serves as a proof of concept, larger multicenter datasets will be necessary to achieve statistically powered multiclass classification and to validate its robustness for clinical deployment.

Several limitations should be considered when interpreting our findings. First, our segmentation was not exhaustive. The protocol was optimized for label fidelity—ensuring the retinal specialists correctly identified the type of drusenoid deposit (hard, soft, RPD) using OCT as the gold standard—rather than quantitative mapping of all visible drusen. This strategy was necessary to build a reliable ground truth for spatial-spectral analysis but means our classification performances results are based only on the segmented drusenoid deposits. Second, although the classification performance was robust in binary discrimination, the overall sample size, particularly for RPD, remains constrained. Future studies with larger cohorts are warranted to validate multiclass classification and ensure generalizability.

Conclusions

This exploratory study provides evidence that hyperspectral imaging can identify subtype-specific spectral signatures of drusenoid deposits and supports its potential role in non-exudative AMD diagnosis and monitoring. Future work should focus on expanding datasets, developing multimodal deep learning models, and evaluating performance in longitudinal settings. Combining hyperspectral retinal imaging with complementary structural and functional modalities, such as OCT, fundus imaging, localized or multifocal electroretinography,^44–46 and contrast sensitivity testing,^47–49 may further refine structure/function correlations and enhance early detection. As novel therapies for nonexudative AMD emerge, hyperspectral imaging could aid in detecting early pathologic changes and tailoring personalized follow-up and treatment strategies.