Indocyanine green angiography and machine learning analysis determine topographic distribution of peripheral hard drusen in a Chinese cohort

Abstract

Background

Hard drusen appear as hyperfluorescent dots on indocyanine green angiography (ICGA) due to their high phospholipid content. This study aims to explore the distribution of peripheral hard drusen across age groups and retinal regions, and their correlation with macular pathology including hard drusen, soft drusen, and progression to neovascular age-related macular degeneration (nAMD).

Methods

A hospital-based retrospective study was conducted at Zhongshan Ophthalmic Center. Patients who underwent ICGA over a five-year period were included and categorised into seven age groups. Hard drusen were quantified across nine retinal quadrants using the Trainable Weka Segmentation machine learning plugin. A subgroup of patients underwent imaging follow-up to assess progression to nAMD.

Results

Here we show, across 1562 normal eyes (14,058 ICGA images) quantified using a validated machine learning model (Dice coefficient: 0.697; sensitivity: 82.93%; ICC: 0.944), that age 50 represents a critical threshold for drusen accumulation, with significant increases observed across all nine quadrants. The nasal quadrants demonstrate earlier and more prominent drusen accumulation compared to the temporal quadrants (P < 0.001 across ages 20–69 years). Notably, the central superior quadrant shows consistent associations with macular hard drusen counts (Stdβ = 0.104, P = 0.010), soft drusen volume (ρ = 0.657, P = 0.020), and progression to nAMD (OR = 1.324, P = 0.040).

Conclusions

Peripheral hard drusen accumulate preferentially in the nasal quadrants and increase markedly after age 50. Comprehensive peripheral retinal evaluation during ICGA examination may hold potential clinical value for AMD monitoring, pending further prospective validation.

Plain language summary

Age-related macular degeneration (AMD) is a leading cause of vision loss in older adults. Small deposits called drusen may be an early sign that AMD might develop later in life. Understanding where these deposits appear and how they relate to progression of AMD could help doctors identify at-risk patients earlier. We analysed 14,058 images of the retina from 1562 eyes and used a computer-assisted tool to map hard drusen across different regions of the retina. We found that drusen accumulate faster in the nasal region of the retina and that deposits in the upper-central region are linked to more advanced degeneration of the macula, which is the central part of the eye. These findings suggest that carefully examining particular regions of the peripheral retina may help identify people at higher risk of developing serious AMD.

Pu et al. use machine learning to quantify peripheral hard drusen across nine retinal quadrants in 1,562 normal eyes using indocyanine green angiography. Nasal quadrants show earlier and greater drusen accumulation, and central superior drusen count independently predicts progression to neovascular age-related macular degeneration.

Introduction

Drusen refers to localized accumulation of extracellular material located between retinal pigment epithelium (RPE) basal lamina and the inner collagenous layer of Bruch’s membrane, which is commonly observed in both normally aging eyes and in cases associated with retinal disease^1,2. Lipids, primarily esterified and unesterified cholesterol, along with phosphatidylcholine, comprise the main component of drusen^3,4. Biochemical process linked to aging, influenced or altered by genetic and environmental variables, may lead to the formation of drusen and disease progression⁵. Over the past decade, considerable progress has been achieved in comprehending drusen, particularly through the emergence of high-resolution in vivo fundus imaging^5,6.

Small hard drusen are uniformly bright white in color with a sharp outline and, by conventional classification, are smaller than 63 µm in diameter⁵. They often appear as common incidental findings of unclear significance during fundus examinations, with their presence visible in fundus photographs for over 90% of adults aged 20 to 46 years⁷. Previous research has shown that the presence of numerous small hard drusen in adults is associated with the later development of soft drusen, indicative of early stages of age-related macular degeneration (AMD)^8–10. Most research has focused on hard drusen in the macular area, yet these lesions also occur extensively in the peripheral retina. Recent studies have begun investigating peripheral hard drusen using ultra-widefield (UWF) imaging, but results are often limited by the subjective nature of drusen assessment^11–13. This is due to the low contrast between drusen and surrounding retinal tissue in pseudocolor and autofluorescence assessments, as well as poor visualization of drusen in the superior and inferior quadrants caused by eyelash obstruction.

Hard drusen may exhibit mild to intense hyperfluorescence on fundus fluorescein angiography (FFA) and indocyanine green angiography (ICGA), with intensity determined by their biochemical composition—high phospholipid content binding more fluorescein while neutral lipid-rich deposits show less binding^5,14. This fluorescence binding property creates higher contrast with surrounding tissue¹⁵, particularly in late-phase ICGA, which provides an ideal foundation for machine learning-based automated quantification. Here, leveraging this advantage, we conducted a hospital-based retrospective study of 1562 normal eyes using nine-quadrant late-phase ICGA and automated machine learning segmentation to characterise the topographic distribution of peripheral hard drusen across age groups and its associations with macular pathology. We show that nasal quadrants accumulate drusen earlier than temporal quadrants, with age 50 as a critical threshold across all quadrants, and that the central superior quadrant is independently associated with macular hard drusen burden, soft drusen volume, and progression to neovascular AMD (nAMD).

Methods

Study participants

This is a hospital-based retrospective study conducted in Zhongshan Ophthalmic Center. This study reviewed consecutive patients who underwent ICGA and were diagnosed with monocular fundus disease or angiographically normal eyes over a five-year period (May 2019 to May 2024). Follow-up data were obtained through systematic review of existing records and active patient recall (April-October 2025). The research adhered to the Declaration of Helsinki and received approval from the Institutional Review Board (IRB) at the Zhongshan Ophthalmic Center (2022KYPJ117 and 2024KYPJ099). Given the retrospective design, the requirement for informed consent for research participation was waived by the IRB. Written informed consent obtained from patients pertained solely to the clinical ICGA procedure, covering associated procedural risks, and was not specific to inclusion in this research study.

Each patient underwent multimodal fundus imaging as assessed by the initial consulting ophthalmologist to confirm the diagnosis. Only late-phase ICGA images (Spectralis; Heidelberg Engineering, Heidelberg, Germany), taken more than 25 minutes after dye injection and following full mydriasis induced by tropicamide, were selected for hard drusen analysis. To obtain a thorough assessment of the fundus, ICGA images of nine retinal quadrants were captured by having patients fixate on nine directional lights within the imaging device. These quadrants include superior temporal, central superior, superior nasal, central temporal, central, central nasal, inferior temporal, central inferior, and inferior nasal (Supplementary Figs. 1-2). A circle centered on the fovea with a radius stretching to the temporal rim of the optic disc was used to define the macular area for further macular drusen identification^16,17.

To avoid interference from other lesions, only fellow eyes with normal ICGA appearances were included in the study (study eye), including eyes exhibiting solely hard drusen or age-related scattered hypofluorescent spots on late-phase ICGA (ASHS-LIA)¹⁸. If both eyes were found to be angiographically normal, the eye with the higher imaging quality was included. Eyes with poor cooperation during the examination, insufficient ocular movement for peripheral image acquisition, or severe refractive media opacities were excluded from the study. Additionally, considering the potential differences in formation mechanisms of pachydrusen, optic disk drusen (ODD), and subretinal drusenoid deposits (SDD) compared to conventional hard drusen^19,20, eyes with these drusen subtypes were excluded. Drusen located over choroidal nevi were omitted from analysis. The data selection workflow is presented in Fig. 1.

Fig. 1. Flowchart of participant screening and enrollment for study.

ICGA, indocyanine green angiography.

Eligible patients in the study were divided into seven age groups based on their actual age at the time of examination: under 20 years, 20−29 years, 30−39 years, 40−49 years, 50−59 years, 60−69 years, and over 70 years. For ease of reference in subsequent description, these groups will be referred to as ‘under 20 s,’ ‘20 s,’ ‘30 s,’ ‘40 s,’ ‘50 s,’ ‘60 s,’ and ‘70 s and over’ respectively. The distribution of drusen across different quadrants in each age group was studied further.

Imaging processing

Hard drusen are defined as yellow, round deposits with a diameter smaller than 63 µm at the sub-RPE level in color photography, and as dotted highly fluorescent in the late-phase ICGA, and sometimes in FFA. Drusen diameter was assessed by visual grading, and those with a diameter less than half the width of a large arcade vein near the optic disk on ICGA were identified²¹. During image selection, we observed that hard drusen in some elderly participants merged into confluent patches appearing as intense hyperfluorescence on late-phase ICGA (Supplementary Fig 2). This confluence occurred in peripheral, peripapillary, and macular regions, obscuring individual drusen contours. To ensure accurate automated counting, eyes with confluent hard drusen were analyzed separately. Cases were classified as either having confluent hard drusen or only scattered hard drusen by two well-trained retinal specialists (J. P. and Y. J.). In instances of disagreement, the case was publicly discussed, and a third retinal specialist (F. W.) made the final determination. Cases with only scattered hard drusen were included in the subsequent machine training for image segmentation and counting.

The Fiji-ImageJ plugin: trainable weka segmentation (TWS) was used to analyse ICGA images of the hard drusen sections (ImageJ, version 1.8.0_322; National Institutes of Health, USA, available at https://imagej.nih.gov/ij). To achieve optimal recognition and segmentation of regions of interest (ROI), we prepared four training sets (A-D), each containing 20 late-phase ICGA images covering all nine quadrants. The images in these training sets displayed an increasing gradient of drusen count: training set A contained images with fewer than 5 drusen, training set B with 6-10 drusen, training set C with 11-20 drusen, and training set D with more than 20 drusen. The image stacks from A to D were sequentially trained on the TWS interface until all drusen in the training sets were clearly identifiable.

The training steps were as follows: (i) Stack Creation: The Stacking tool in ImageJ was utilized to create four training stacks, each containing 20 images. (ii) Opening TWS: The TWS plugin was launched, and training stack A was initially loaded. (iii) Algorithm Selection: The classifier was configured to use the following algorithms: Gaussian blur, Sobel filter, Hessian, difference of Gaussians, and membrane projections. Membrane Thickness was set to 1, Membrane Patch Size to 19, and Minimum and Maximum Sigma values were set to 1.0 and 16.0, respectively. (iv) ROI Segmentation: Selection tools, including circular, rectangular, and freehand line tools, were employed to delineate the drusen area in the 20 ICGA images. A line tool was used to separate overlapping drusen. The samples were manually classified into two classes: class one (drusen) and class two (background). (v) Classifier Training: The classifier was trained with the selected samples until it could recognize all drusen particles in the training stack A. Subsequently, stacks B through D were sequentially loaded and trained further on the existing model. Finally, the classifier was saved, and a probability map was generated. (vi) Particle Counting: The probability map was converted from a 32-bit to an 8-bit image, and the threshold was adjusted to above 90% to obtain a binary image, where drusen appeared as white particles against a black background. The ‘Analyze Particles’ tool in ImageJ was employed for counting the drusen, with a size threshold set from 30 to infinity and a circularity threshold from 0.2 to infinity. (vii) Workflow Standardization: The segmentation and quantification steps were recorded and compiled into a batch of commands using macros in ImageJ, facilitating easy execution in subsequent large-sample analyses.

A separate validation set comprising 80 ICGA images was obtained from an independent retrospective cohort examining ICGA-based choroidal features in pachychoroid spectrum diseases (IRB approval: 2023KYPJ114, Zhongshan Ophthalmic Center). As with the main study, informed consent for research use of images was waived by the IRB due to the retrospective nature of the study. Drusen counts in these 80 ICGA images were automatically quantified by the trained machine learning program, as well as manually by a trained observer (J. P.). The consistency between the model’s output and manual measurements was evaluated to assess the model’s performance. Representative hard drusen appearance on late-phase ICGA is shown in Fig. 2, the TWS operating interface is illustrated in Supplementary Fig. 3, and the overall machine learning workflow is presented in Supplementary Fig. 4.

Fig. 2. Identification and Segmentation of Small Hard Drusen in Late-Phase Indocyanine Green Angiography (ICGA).

A1, B1, C1, D1 These original late-phase ICGA images were captured with patients fixating on central, superior nasal, superior temporal, and inferior temporal directional lights within the imaging device. A2, B2, C2, D2 Hard drusen are highlighted by white dashed boxes and indicated with yellow or orange arrows in the corresponding original images. Hard drusen are defined as round or quasi-round highly fluorescent spots in the late phases of ICGA, with diameters less than half the width of a large arcade vein near the optic disk. A3, B3, C3, D3 Magnified views within the white dashed boxes are provided to better visualize the drusen, where hard drusen are outlined by dotted yellow circles. The number of drusen increases from top to bottom, with respective counts of 1, 2, 6, and 25 drusen. ICGA indocyanine green angiography.

Follow-up data collection

Systematic follow-up was conducted for participants aged ≥50 years through review of existing medical records and active patient recall (April–October 2025). Follow-up examinations included optical coherence tomography (OCT), with OCT angiography (OCTA) or FFA/ICGA performed when clinically indicated. To investigate the relationship between peripheral hard drusen and macular soft drusen (or drusenoid pigment epithelial detachment [PED]), eyes with soft drusen (or drusenoid PED) identified at baseline or developed during follow-up were combined for analysis. All eyes with soft drusen (or drusenoid PED) had volumetric OCT imaging available. Soft drusen volumes were quantified using manual three-dimensional segmentation on OCT as we previously described^22–24. Representative multimodal imaging and three-dimensional volumetric reconstruction of soft drusen are shown in Fig. 3.

Fig. 3. Multimodal imaging and three-dimensional volumetric quantification of soft drusen and drusenoid PED.

A−C Early-phase (37 s), mid-phase (12 min 05 s), and late-phase (37 min 56 s) ICGA images. Soft drusen and drusenoid PED areas consistently appear hypofluorescent throughout all phases. Scattered areas of ASHS-LIA are visible in the posterior pole. D Late-phase ICGA with mask soft drusen and drusenoid PED overlay. E, F Two representative B-scans from the volumetric OCT through the soft drusen region, showing dome-shaped RPE elevations with medium-to-high reflectivity material between the RPE and Bruch’s membrane. G Stack of B-scans used for three-dimensional volumetric measurement and reconstruction. H Three-dimensional rendering of the retinal and choroidal structure from volumetric OCT, demonstrating parafoveal annular elevation. I Three-dimensional visualization of the segmented soft drusen volume (orange), with retinal and choroidal layers rendered semi-transparent for enhanced drusen visualization. ASHS-LIA age-related scattered hypofluorescent spots on late-phase ICGA, ICGA indocyanine green angiography, OCT optical coherence tomography, PED pigment epithelial detachment, RPE retinal pigment epithelium.

Additionally, we evaluated the development of nAMD in study eyes that were angiographically normal at baseline, diagnosed based on clinical symptoms and the presence of macular neovascularization (MNV) on multimodal imaging according to established diagnostic criteria²⁵. Two representative cases of nAMD at follow-up are illustrated in Fig. 4. For study eyes that progressed to nAMD, baseline information was recorded including age, sex, fellow eye nAMD status at baseline, and follow-up duration (defined as the time interval from study enrollment to nAMD diagnosis, measured in months).

Fig. 4. Representative cases of nAMD. (A-F) Multimodal imaging of an eye with Type 2 macular neovascularization.

A IR image and corresponding OCT B-scan. The horizontal green arrow in the left panel indicates the B-scan location. OCT reveals retinal and sub-RPE hyperreflective material with extensive intraretinal fluid. Clustered hyperreflective intraretinal foci are visible on the right side of the PED. B OCTA demonstrates a neovascular network in the macular region (yellow arrowhead), partially obscured by exudation. C Early-phase FFA (29 s) shows hyperfluorescence from the inferior parafoveal neovascular network. D Late-phase FFA (12 min 24 s) demonstrates dye pooling in the macular region. E Early-phase ICGA (39 s) reveals the macular neovascular network and peripheral hypofluorescent areas due to exudative obscuration. F Mid-phase ICGA (12 min 24 s) shows staining of the neovascular network. Although CVH superior and inferior to the optic disc suggests pachychoroid neovasculopathy, the presence of Type 2 MNV without significant pachyvessels or marked choroidal thickening, along with hard exudates, led the expert panel to establish nAMD as the primary diagnosis. G−L Multimodal imaging of an eye with PCV. G IR image and corresponding OCT B-scan. The horizontal green arrow in the left panel indicates the B-scan location. OCT demonstrates the double-layer sign, a hyporeflective ring within the PED (yellow arrowhead), extensive subretinal hyperreflective material, and minimal subretinal fluid. H OCTA reveals a partial macular neovascular network (yellow arrowhead), partially obscured by exudation. I Early-phase FFA (59 s) shows extensive atrophic region in posterior pole with visible underlying choroidal vessels and focal hyperfluorescence in the macula (likely polypoidal lesions). J Late-phase FFA (19 min 18 s) demonstrates macular dye pooling. K, L Early-phase (1 min 34 s) and late-phase (19 min 18 s) ICGA reveal extensive atrophic region in posterior pole. Late-phase imaging shows staining of neovascular tissue in the macular region (likely polypoidal lesions), which were not prominent in the early phase. CVH choroidal vascular hyperpermeability; FFA fluorescein angiography; ICGA indocyanine green angiography; IR infrared reflectance; MNV macular neovascularization; nAMD neovascular age-related macular degeneration; OCT optical coherence tomography; OCTA optical coherence tomography angiography; PCV polypoidal choroidal vasculopathy; PED pigment epithelial detachment; RPE retinal pigment epithelium.

Statistics and reproducibility

Model performance was evaluated using pixel-level metrics (sensitivity, specificity, accuracy, precision, Dice coefficient, and intersection over union [IoU/Jaccard Index]) and object-level metrics (mean absolute error [MAE] for drusen counting). Inter-observer reliability for manual drusen counting in the validation set was assessed using intraclass correlation coefficient (ICC) and Pearson correlation analysis. Bland-Altman analysis was performed to evaluate agreement between automated and manual drusen counts. For clinical data analysis, descriptive statistics were reported as means ± standard deviations (SD) or medians with interquartile ranges (IQR) for continuous variables, and frequencies and percentages for categorical variables. Chi-square or Fisher’s exact tests were used to compare categorical variables (presence of scattered and confluent hard drusen across age groups). Mann-Whitney U tests were used to compare continuous variables between groups (drusen counts between age groups, drusen counts between progressors and non-progressors). Wilcoxon signed-rank tests were used to evaluate differences in drusen counts across different retinal quadrants. Linear regression models assessed the relationship of peripheral and macular hard drusen counts. Spearman correlation analysis was performed to assess the relationship between peripheral hard drusen counts and macular soft drusen volume. Logistic regression analysis was conducted to identify factors associated with progression to neovascular AMD, with age, sex, follow-up duration, and fellow eye nAMD status as candidate variables. All statistical analyses were conducted using IBM SPSS software (version 26.0, IBM Corp., Armonk, NY, USA). Significance was set at P < 0.05 (two-sided).

Results

Demographic data for study population

ICGA images from a total of 6175 patients were collected and screened. After applying stringent exclusion criteria, images from 1562 normal eyes—totaling 14,058 late-phase ICGA images across nine quadrants in 1562 patients—were included in the study, as depicted in the flowchart of Fig. 1. In this cohort, males accounted for 64.21% of the subjects, with an almost equal distribution between left and right eyes (50.19% for right eyes). The age distribution across the seven groups was as follows: 132 eyes in the under 20 s group, 132 in the 20 s group, 237 in the 30 s group, 302 in the 40 s group, 388 in the 50 s group, 299 in the 60 s group, and 72 in the 70 s and over group. Regarding the diagnosis of the contralateral symptomatic eyes, the electronic medical records system was queried, with uveitis (379 cases, 24.26%) being the most common diagnosis, including infectious and non-infectious types, followed by nAMD (244 cases, 15.62%), central serous chorioretinopathy (CSC) (222 cases, 14.21%), and angiographically normal eyes (170 cases, 10.88%). Detailed demographic and diagnostic information are presented in Table 1.

Table 1.

Demographic and Diagnostic Information of the Included Patients

Demographic and diagnostic characteristic	Value
Demographic information
Patients, n	1562
Eyes, n	1562
Sex, Males (%)	1003 (64.21%)
Laterality, Right (%)	784 (50.19%)
Age Groups, n (No. of Males, Male%)
< 20 yrs	132 (76, 57.6%)
20−29 yrs	132 (69, 52.27%)
30−39 yrs	237 (143, 60.34%)
40−49 yrs	302 (213, 70.53%)
50−59 yrs	388 (248, 63.92%)
60−69 yrs	299 (207, 69.23%)
> 70 yrs	72 (47, 65.28%)
Diagnostic informationa,n (%)
Uveitis	379 (24.26%)
Neovascular age-related macular degeneration	244 (15.62%)
Central serous chorioretinopathy	222(14.21%)
Angiographically normal eyes	170 (10.88%)
Pachychoroid pigment epitheliopathy	90 (5.71%)
Vitreous floater	67 (4.29%)
Choroiretinal tumor	66 (4.22%)
Optic neuropathy	57 (3.65%)
Choroidal neovascularization secondary to uveitis	48 (3.07%)
Chorioretinal atrophic lesion	41 (2.62%)
Multiple evanescent white dot syndrome	18 (1.15%)
Retinal vein occlusion	17 (1.09%)
Age-related retentional avascular pigment epithelial detachment	16 (1.02%)
Traumatic retinopathy	16 (1.02%)
Punctate inner choroidopathy	15 (0.96%)
Epiretinal membranes	14 (0.90%)
Pachychoroid neovasculopathy	13 (0.83%)
Others	69 (4.42%)

^aOnly diagnoses with more than 10 cases are presented. Other diseases included hereditary chorioretinopathy (8 cases), congenital hypertrophy of the retinal pigment epithelium, retinal macroaneurysm, macular hole (6 cases each), Coats' disease, laser-induced maculopathy, idiopathic choroidal neovascularization, drusenoid pigment epithelial detachment (5 cases each), choroidal nevus (4 cases), rhegmatogenous retinal detachment, papilledema secondary to idiopathic intracranial hypertension (3 cases each), acute macular neuroretinopathy, type 1 macular telangiectasia, congenital choroidal coloboma, Valsalva retinopathy, glaucoma (2 cases each), and focal choroidal excavation, persistent placoid maculopathy, central retinal artery occlusion (1 case each).

Hard drusen presence across different age groups

Initially, we focused on the presence of hard drusen across different age groups, as demonstrated in Table 2 and visualized in Fig. 5. As age increases, the proportion of eyes with hard drusen rises, with the percentages in ascending age groups being 28.79%, 53.03%, 54.21%, 72.52%, 90.47%, 99.67%, and 100%, respectively. Significant differences were observed between most adjacent age groups (P < 0.001 for all except for the 20 s vs. 30 s, and 60 s vs. 70 s and over, with P-values of 0.857 and 0.623, respectively). When focusing solely on the confluent type of hard drusen, we found only 2 cases in the 40 s group. With each advancing decade, the proportion of such cases significantly increased (P < 0.001 for all). Age group was found to be significantly correlated with the presence of drusen in regression models adjusted for sex (P < 0.001 for both drusen and the confluent subtype). However, sex was not recognized as a significant factor for drusen presence when adjusted for age group (P = 0.569 and 0.517 for drusen and the confluent subtype, respectively).

Table 2.

Proportion of Hard Drusen Among Different Age Groups

Age Groups	Hard Drusen		Confluent hard drusen only
	Presence, n (%)	Pa	Presence, n (%)	Pa
<20 yrs	38/132 (28.79%)	-	0 (0%)	-
20−29 yrs	70/132 (53.03%)	<0.001	0 (0%)	-
30−39 yrs	128/237 (54.21%)	0.857	0 (0%)	-
40−49 yrs	219/302 (72.52%)	<0.001	2/302 (0.66%)	-
50−59 yrs	351/388 (90.47%)	<0.001	46/388 (11.86%)	<0.001
60−69 yrs	298/299 (99.67%)	<0.001	119/299 (39.80%)	<0.001
> 70 yrs	72/72 (100%)	0.623	51/72 (70.83%)	<0.001
Total	1176/1562 (75.29%)		218/1562 (13.96%)
Sex adjusted for age group,
OR (95% CI)	1.085 (0.819, 1.437)	0.569	0.883 (0.607, 1.286)	0.517
Age group adjusted for sex,
OR (95% CI)	2.326 (2.113, 2.559)	<0.001	5.521 (4.369, 6.976)	<0.001

^aChi-square tests or Fisher’s exact tests were used to compare the differences in drusen proportions between corresponding age groups and the preceding age groups. Logistic regression models were employed to investigate the impact of sex and age on drusen presence by treating age as a continuous ordered factor. All statistical tests were two-sided, with significant P-values highlighted in bold. No adjustments were made for multiple comparisons.

Fig. 5. Distribution of drusen across nine quadrants and different age groups.

A Stacked Bar Chart of Drusen Presence Across Different Age Groups. The proportion of drusen presence shows a continuous increase with age, with about half of the participants exhibiting drusen by age 30, and nearly every individual displaying drusen presence by age 60. B Stacked Bar Chart of Confluent Drusen Presence Across Different Age Groups. For eyes with only confluent hard drusen, their appearance is generally later than that of scattered hard drusen, first observed after age 40, followed by a consistent increase as age increases by decades. By age 70 and beyond, over 70% of participants exhibit confluent hard drusen. C Radar Chart of Scattered Hard Drusen Presence in Nine Quadrants Across Different Age Groups. This chart shows a higher proportion of drusen in the central nasal quadrant at earlier ages. D Radar Chart of Confluent Hard Counts in Nine Quadrants Across the 50 s, 60 s, and 70 s Age Groups. It significantly highlights that the three nasal quadrants (central nasal, superior nasal, and inferior nasal) have the highest proportion of confluent hard counts across all age groups, while the three temporal quadrants (central temporal, superior temporal, and inferior temporal) have the lowest, regardless of age. E Heatmap of Specific Drusen Counts in Nine Quadrants Across Different Age Groups, where deeper red indicates a higher number of drusen. The cells display the median (first and third quartiles), and an asterisk (*) above a number indicates a significant increase in drusen count compared to the previous age group. ST superior temporal, CS central superior, SN superior nasal, CT central temporal, CN central nasal, IT inferior temporal, CI central inferior, IN inferior nasal.

Model performance and validation

The Trainable Weka Segmentation model demonstrated good performance for automated drusen quantification on the 80-image validation set. Pixel-level metrics revealed a sensitivity of 82.93% ± 12.28%, specificity of 99.95% ± 0.07%, overall accuracy of 99.93% ± 0.09%, and precision of 61.83% ± 10.77%. The Dice coefficient was 0.697 ± 0.069, and intersection over union (IoU) was 0.539 ± 0.087. Object-level performance showed a mean absolute error of 2.19 ± 2.98 drusen per image, with 73.1% of images showing errors within ±2 drusen, indicating good spatial agreement between automated and manual segmentations. The high specificity and accuracy primarily reflect the extreme class imbalance, where drusen pixels constitute <0.1% of total image area. Bland-Altman analysis demonstrated excellent agreement between automated and manual quantification, with a mean bias of 0.47 drusen (95% limits of agreement: −8.32 to 7.37 drusen), and 97.5% (78/80) of measurements fell within the limits of agreement (Supplementary Fig. 5A). Inter-observer reliability for manual drusen counting was excellent, with an ICC of 0.944 (95% CI: 0.917−0.963) and Pearson correlation of r = 0.946 (P < 0.0001) between two independent raters (Supplementary Fig 5B), confirming high reproducibility of manual drusen quantification in late-phase ICGA images.

Topographic distribution differences of hard drusen

By applying such TWS classifier, we obtained drusen counts in nine quadrants across different age groups, the results are presented in Table 3 and visualized in a heatmap in Fig. 5. A significant increase in drusen counts was observed at ages of 50 s and 60 s in each quadrant (P < 0.05 for all). Interestingly, in the nasal and inferior quadrants, the onset of significant increases in drusen counts occurred earlier, with the first significant increase observed at the age of 40 s in the three inferior quadrants, including the inferior temporal, central inferior, and inferior nasal quadrants (P < 0.05 for all). The superior nasal quadrant showed its first significant increase at the age of 30 s (P = 0.013). Additionally, there was a trend of increasing drusen in the central nasal quadrant at the age of 40 s, although the P-value approached but did not reach statistical significance (P = 0.092).

Table 3.

Counts of Hard Drusen in Nine Quadrants Among Different Age Groups

Age Groups	ST		CS		SN
	Counts	Pa	Counts	Pa	Counts	Pa
<20 yrs	0 (0, 0)	-	0 (0, 0)	-	0 (0, 0)	-
20−29 yrs	0 (0, 1)	0.332	0 (0, 1)	0.141	0 (0, 1)	0.272
30−39 yrs	0 (0, 2)	0.156	0 (0, 2)	0.123	1 (0, 2)	0.013
40−49 yrs	0 (0, 2)	0.053	1 (0, 3)	0.082	1 (0, 3)	0.154
50−59 yrs	2 (0, 6)	<0.001	3 (1, 7)	<0.001	3 (1, 8)	<0.001
60−69 yrs	5 (2, 8)	<0.001	4 (2, 8)	<0.001	5 (3, 9)	<0.001
> 70 yrs	5 (3, 8)	0.933	6 (4, 10)	0.187	5 (4, 7)	0.719
	CT		Central		CN
	Counts	Pa	Counts	Pa	Counts	Pa
<20 yrs	0 (0, 0)	-	0 (0, 0)	-	0 (0, 1)	-
20−29 yrs	0 (0, 1)	0.145	0 (0, 1)	0.066	1 (0, 3)	0.167
30−39 yrs	0 (0, 1)	0.059	0 (0, 2)	0.542	1 (0, 3)	0.716
40−49 yrs	0 (0, 2)	0.948	0 (0, 1)	0.845	1 (0, 4)	0.092
50−59 yrs	2 (0, 4)	<0.001	2 (0, 5)	<0.001	4 (1, 10)	<0.001
60−69 yrs	4 (2, 7)	<0.001	5 (3, 9)	<0.001	5 (3, 10)	0.017
> 70 yrs	4 (2, 6)	0.486	5 (3, 8)	0.965	4 (3, 9)	0.600
	IT		CI		IN
	Counts	Pa	Counts	Pa	Counts	Pa
<20 yrs	0 (0, 0)	-	0 (0, 0)	-	0 (0, 0)	-
20−29 yrs	0 (0, 1)	0.301	0 (0, 0)	0.954	0 (0, 2)	0.089
30−39 yrs	0 (0, 1)	0.760	0 (0, 1)	0.149	0 (0, 2)	0.120
40−49 yrs	0 (0, 2)	<0.001	0 (0, 3)	<0.001	1 (0, 4)	0.025
50−59 yrs	2 (0, 6)	<0.001	2 (0, 6)	<0.001	3 (1, 8)	<0.001
60−69 yrs	4 (2, 7)	<0.001	3 (2, 8)	<0.001	5 (2, 8)	0.002
> 70 yrs	4 (2, 6)	0.278	2 (1, 8)	0.197	5 (2, 10)	0.720

Data are presented as median (Interquartile range, IQR) for counts of hard drusen.

^aMann-Whitney U tests were used to compare differences in hard drusen counts between corresponding age group and its preceding age group. All statistical tests were two-sided, with significant P-values highlighted in bold. No adjustments were made for multiple comparisons.

ST superior temporal, CS central superior, SN superior nasal, CT central temporal, CN central nasal, IT inferior temporal, CI central inferior, IN inferior nasal.

To further explore variations in drusen distribution in each age group, we compared drusen counts between the superior and inferior quadrants, as well as between the temporal and nasal quadrants, the results are shown in Table 4. We found that in almost all age groups, the nasal quadrants had significantly higher drusen counts than the temporal quadrants (P < 0.001 for all age groups except for the youngest and oldest). Meanwhile, the differences between the superior and inferior quadrants were not significant, with the exception of the 30 s group (P = 0.031). Typical ICGA images demonstrating predominantly nasal scattered drusen distribution are shown in Supplementary Fig. 1. Confluent hard drusen exhibited a similar topographic distribution pattern. Only two eyes in the 40 s age group had confluent drusen, both located in the central nasal quadrant (100%). In older age groups (50−70s), the central nasal quadrant consistently had the highest prevalence of confluent drusen (94.12%, 95.00%, and 97.44%, respectively), while the central temporal quadrant had the lowest (17.65%, 15.00%, and 15.38%, respectively). Typical ICGA images demonstrating predominantly nasal confluent drusen distribution are shown in Supplementary Fig. 2.

Table 4.

Comparison of Hard Drusen Counts Across Four Quadrants

Age Groups	Superior vs. Inferior			Temporal vs. Nasal
	Superior Quadrantsb	Inferior Quadrantsc	Pa	Temporal Quadrantsd	Nasal Quadrantse	Pa
<20 yrs	0 (0, 0)	0 (0, 0)	0.685	0 (0, 0)	0 (0, 0)	0.091
20−29 yrs	0 (0, 0)	0 (0, 0)	0.993	0 (0, 0)	0 (0, 1)	<0.001
30−39 yrs	0 (0, 1)	0 (0, 1)	0.031	0 (0, 0)	0 (0, 1)	<0.001
40−49 yrs	0 (0, 2)	0 (0, 2)	0.792	0 (0, 2)	1 (0, 3)	<0.001
50−59 yrs	3 (1, 5)	0 (0, 0)	0.197	2 (0, 5)	3 (1, 7)	<0.001
60−69 yrs	5 (3, 7)	2 (0, 6)	0.172	5 (3, 7)	5 (3, 10)	<0.001
> 70 yrs	6 (4, 7)	4 (3, 7)	0.277	4 (2, 6)	5 (4, 8)	0.199

Median (interquartile range, IQR) was presented for counts of hard drusen.

^aWilcoxon signed-rank tests were used to compare differences in hard drusen counts across various quadrants. All statistical tests were two-sided. Statistically significant P-values are highlighted in bold. No adjustments were made for multiple comparisons.

^bThe count of hard drusen in superior quadrant was defined by averaging the counts in the superior temporal, central superior, and superior nasal quadrants.

^cThe count of hard drusen in inferior quadrant was defined by averaging the counts in the inferior temporal, central inferior, and inferior nasal quadrants.

^dThe count of hard drusen in temporal quadrants was defined by averaging the counts in the superior temporal, central temporal, and inferior temporal.

^eThe count of hard drusen in nasal quadrants was defined by averaging the counts in the superior nasal, central nasal, and inferior nasal.

Correlation between Periphery Drusen Counts and Macular Drusen Burden

We explored the association between peripheral drusen counts and macular drusen burden, including both macular hard drusen counts in the baseline study cohort and macular soft drusen volume in eyes with soft drusen present either at baseline or during follow-up. For macular hard drusen counts (Table 5), univariate regression analysis revealed that drusen counts in all eight peripheral quadrants were significantly associated with macular drusen counts (P < 0.001 for all quadrants). However, in multivariable regression analysis, only three peripheral quadrants remained independent associations: central superior (standardized β = 0.104, P = 0.010), central temporal (standardized β = 0.163, P < 0.001), and inferior temporal (standardized β = 0.091, P = 0.040).

Table 5.

Regression Analysis for Correlation between Peripheral and Macular Drusen

Variables	Counts of macular hard drusen
	Univariate		Multivariate
	Stdβ	Pa	Stdβ	Pa
Basic information
Age	0.334	<0.001	0.168	<0.001
Sex	−0.082	0.003	−0.042	0.142
Laterality	−0.003	0.911
Drusen counts
ST	0.322	<0.001	−0.028	0.557
CS	0.334	<0.001	0.104	0.010
SN	0.311	<0.001	0.021	0.628
CT	0.356	<0.001	0.163	<0.001
CN	0.231	<0.001	−0.040	0.295
IT	0.307	<0.001	0.091	0.040
CI	0.331	<0.001	−0.005	0.904
IN	0.314	<0.001	0.068	0.091

^aLinear regression analyses. Variables demonstrating a P-value < 0.05 in univariate models were included in the multivariate model. All statistical tests were two-sided, with significant P-values highlighted in bold. No adjustments were made for multiple comparisons.

Extreme values have been removed from regression analyses, which are defined as values less than Q1−1.5 * IQR or greater than Q3 + 1.5 * IQR, where Q1 is the first quartile and Q3 is the third quartile of the data distribution.

ST superior temporal, CS central superior, SN superior nasal, CT central temporal, CN central nasal, IT inferior temporal, CI central inferior, IN inferior nasal.

Combining baseline and follow-up data, we identified 52 eyes with macular soft drusen that underwent nine-quadrant ICGA imaging (5 eyes at baseline from fellow eyes, 47 eyes with newly developed soft drusen during follow-up). Of these, 40 eyes (76.92%) exhibited confluent peripheral hard drusen (representative case shown in Supplementary Fig. 6), which precluded quantitative drusen counting and correlation analysis. In the remaining 12 eyes with both macular soft drusen and countable peripheral hard drusen, three-dimensional volumetric measurements of soft drusen were performed (Fig. 3). Spearman correlation analysis revealed that only central superior hard drusen count showed significant correlation with macular soft drusen volume (R_s = 0.657, P = 0.020), while all other quadrants showed no significant associations (Supplementary Table 1 and Fig. 6).

Fig. 6. Correlation between peripheral hard drusen counts and macular soft drusen volume.

Spearman correlation between hard drusen counts across retinal quadrants and macular soft drusen volume (n = 12). Scatter plots show individual data points with regression lines and 95% CI. All statistical tests were two-sided. No adjustments were made for multiple comparisons. Only central superior hard drusen count showed significant correlation (R_s = 0.657, P = 0.020). Composite quadrant counts represent averaged values from their respective subfields (see Note). Note: Superior = mean (ST, CS, SN); Inferior = mean (IT, CI, IN); Temporal = mean (ST, CT, IT); Nasal = mean (SN, CN, IN). CI, central inferior; CN, central nasal; CS, central superior; CT, central temporal; IN, inferior nasal; IT, inferior temporal; Rs, Spearman correlation coefficient; SN, superior nasal; ST, superior temporal.

Correlation between baseline periphery drusen counts and nAMD progression

Among the 246 participants aged ≥50 years who completed imaging follow-up (mean follow-up duration: 26 ± 18 months), 49 eyes (19.9%) developed newly diagnosed nAMD. Comparison between progressors (n = 49) and non-progressors (n = 197) revealed significantly greater hard drusen counts in the central superior (P = 0.001) and superior nasal (P = 0.001), as shown in Table 6. Moreover, confluent hard drusen in the central superior quadrant were more frequent in progressors than in non‑progressors (22/54 vs 1/13, P = 0.024). Logistic regression analysis (Table 7) demonstrated that central superior hard drusen count was the only factor independently associated with progression to nAMD (OR: 1.324 per drusen, 95% CI: 1.013−1.731, P = 0.040), whereas age, sex, follow-up duration, and fellow-eye nAMD status at baseline showed no significant associations in this cohort. A sensitivity analysis was conducted in the subgroup of 198 eyes with complete data on clinical covariates, including hypertension, diabetes mellitus, hyperlipidemia, smoking history, and alcohol consumption (Supplementary Table 2). Among these 198 eyes, 40 (20.2%) progressed to nAMD. After forcing all available clinical covariates into the logistic regression model, both central superior and superior nasal hard drusen counts remained independently associated with nAMD progression (CS: OR = 1.455, 95% CI: 1.031–2.052, P = 0.033; SN: OR = 1.398, 95% CI: 1.043–1.874, P = 0.025).

Table 6.

Comparison of Drusen Counts and Proportion of Confluent Hard Drusen Between Progressors and Non-progressors

Age Groups	Hard Drusen			Confluent hard drusen only
	Progressors	Non-progressors	Pa	Progressors	Non-progressors	Pb
	(n = 49)			(n = 49)
(n = 197)
(n = 197)
ST	2 (0, 3)	2 (0, 3)	0.116	11/54	1/13	0.285
CS	3 (2, 5)	2 (0, 4)	0.001	22/54	1/13	0.024
SN	3 (2, 5)	2 (0, 4)	0.007	31/54	5/13	0.219
CT	0 (0, 2)	1 (0, 3)	0.717	4/54	2/13	0.366
CN	3 (1, 4)	2 (0, 4)	0.054	44/54	10/13	0.709
IT	2 (0, 2)	2 (0, 3)	0.059	8/54	1/13	0.499
CI	1 (1, 3)	1 (0, 3)	0.403	24/54	2/13	0.054
IN	3 (2, 4)	2 (0, 3)	0.330	32/54	5/13	0.176
Superiorc	3 (2, 4)	2 (0, 3)	<0.001	31/54	5/13	0.219
Inferiord	2 (2, 3)	2 (1, 3)	0.002	33/54	5/13	0.139
Temporale	1 (1, 2)	2 (0, 3)	0.007	12/54	2/13	0.586
Nasalf	3 (2, 5)	3 (1, 3)	<0.001	44/54	10/13	0.709

Data are presented as median (interquartile range, IQR) for hard drusen counts and as proportions for presence of confluent hard drusen.

ST superior temporal, CS central superior, SN superior nasal, CT central temporal, CN central nasal, IT inferior temporal, CI central inferior, IN inferior nasal.

^aMann-Whitney U tests were used to compare differences in hard drusen counts between progressors and non-progressors.

^bChi-square tests or Fisher’s exact tests were used to compare proportions of confluent hard drusen presence between groups. All statistical tests were two-sided, with significant P-values highlighted in bold. No adjustments were made for multiple comparisons.

^cThe count of hard drusen in superior quadrant was defined by averaging the counts in the superior temporal, central superior, and superior nasal quadrants.

^dThe count of hard drusen in inferior quadrant was defined by averaging the counts in the inferior temporal, central inferior, and inferior nasal quadrants.

^eThe count of hard drusen in temporal quadrants was defined by averaging the counts in the superior temporal, central temporal, and inferior temporal.

^fThe count of hard drusen in nasal quadrants was defined by averaging the counts in the superior nasal, central nasal, and inferior nasal.

Table 7.

Logistic Regression Analysis of Drusen Counts as Predictors of Progression to Neovascular AMD

Variables	Univariate		Multivariate
	OR (95%CI)	Pa	OR (95%CI)	Pa
Basic information
Age	1.037 (0.980, 1.098)	0.204
Sex	0.851 (0.427, 1.693)	0.645
Follow-up durationb	0.995 (0.977, 1.013)	0.586
nAMD diagnosis in fellow eye at baseline	1.039 (0.477, 2.262)	0.924
Drusen counts
ST	1.183 (0.972, 1.439)	0.094
CS	1.410 (1.137, 1.749)	0.002	1.324 (1.013, 1.731)	0.040
SN	1.237 (1.057, 1.446)	0.008	1.204 (0.910, 1.593)	0.193
CT	1.051 (0.813, 1.358)	0.703
CN	1.150 (0.990, 1.336)	0.067
IT	1.213 (1.007, 1.462)	0.042	0.898 (0.660, 1.223)	0.495
CI	1.094 (0.883, 1.356)	0.409
IN	1.069 (0.916, 1.247)	0.396

Extreme values have been removed from regression analyses, which are defined as values less than Q1−1.5 * IQR or greater than Q3 + 1.5 * IQR, where Q1 is the first quartile and Q3 is the third quartile of the data distribution.

ST superior temporal, CS central superior, SN superior nasal, CT central temporal, CN central nasal, IT inferior temporal, CI central inferior, IN inferior nasal.

^aLinear regression analyses. Variables demonstrating a P-value < 0.05 in univariate models were included in the multivariate model. All statistical tests were two-sided, with significant P-values highlighted in bold. No adjustments were made for multiple comparisons.

^bFollow-up duration: time from baseline ICGA to last follow-up, in months.

Discussion

By integrating manual identification with machine learning, this hospital-based retrospective study of ICGA images from 1562 fellow eyes revealed that the prevalence and count of hard drusen increased with age. The most significant increase in drusen counts was commonly observed in the 50 s age group across all nine quadrants, with earlier occurrences in the nasal and inferior quadrants. Notably, the nasal quadrants consistently showed significantly higher drusen counts and the frequency of confluent type compared to the temporal quadrants, whereas no substantial differences were observed between the superior and inferior quadrants. Among all quadrants, central superior hard drusen counts consistently emerged as an independent correlate in all three analyses (macular hard drusen counts, macular soft drusen volume, and nAMD progression). It should be noted that the observed association with nAMD progression is exploratory and hypothesis-generating, constrained by the retrospective design, limited longitudinal subgroup, and lack of multimodal confirmation (e.g., colour fundus photography); whether peripheral drusen burden in particular quadrants constitutes a reliable prognostic marker merits prospective investigation in larger cohorts.

Drusen were first described by Franciscus Donders in 1855 and termed “drusen” by Heinrich Müller in 1856⁵. It is now widely recognized that age-related declines in Bruch’s membrane hydraulic conductivity impair RPE-choroid nutrient and waste exchange, resulting in drusen formation²⁶. Since larger drusen (>63−125 μm) serve as AMD diagnostic criteria, subjects with only small hard drusen have been consistently excluded or used as controls in previous research^27–29. Such study designs provide limited information about the age-related and topographic distribution of hard drusen in otherwise normal eyes. Although small hard drusen may not significantly affect vision, their presence may signal early AMD risk, as previous research indicates that having 20 or more small hard drusen in the macular area is associated with a higher probability of developing soft drusen^16,17. Due to the lack of normal eye samples with comprehensive fundus examinations, as well as the subjectivity of drusen assessment on fundus photography and autofluorescence imaging, no studies have yet demonstrated the correlation between peripheral and macular drusen in normal eyes.

Advances in machine learning and tools like ImageJ have improved image quantification efficiency and consistency. By marking objects in images to create a training dataset, a classifier can be refined through direct feedback and then applied to classify similar images^30–32. The Trainable Weka Segmentation (TWS) plugin in FIJI has been widely applied across diverse fields, including mineralogy and microbiology^33,34. In late-phase ICGA images, hard drusen, due to their high phospholipid content, bind effectively with ICG molecules, displaying well-defined hyperfluorescent regions^5,14. This creates a stark contrast against the background of the surrounding normal chorioretinal tissue, which to some extent simplifies the training process, allowing for improved ROI segmentation results with relatively smaller sample sizes. We pioneered the application of TWS methodology to ophthalmic imaging, with comprehensive procedural details provided in the Methods section as a reference for future studies. Validation on 80 images yielded acceptable performance (Dice coefficient: 0.697), which is consistent with expectations for small lesion segmentation and supports its applicability for large-scale population-level drusen quantification in ICGA images acquired under standardised conditions. Conclusions drawn from this study reflect the performance of TWS on single-centre ICGA data and should be interpreted within this methodological context.

We observed age-dependent increases in peripheral hard drusen across all quadrants, with age 50 serving as a critical threshold. This finding parallels the current age-based AMD diagnostic criteria²⁵. However, the age at which significant drusen accumulation occurs varied across peripheral quadrants, suggesting regional differences in aging rates. This indicates that drusen formation is influenced not only by age and systemic risk factors (including elevated serum triglycerides, reduced high-density lipoprotein [HDL] cholesterol¹⁶, and genetic variants such as CFH and ARMS2)^35–38, but also by quadrant-specific anatomical features. Our study suggested that the nasal side is likely the earliest area to show drusen, while the temporal side is the latest. This nasal-predominant, temporal-sparing pattern is consistent with findings from several studies. Corbelli et al. and Lengyel et al. both reported that peripheral drusen frequently spare the temporal region^11,39, while Munch et al. found drusen prevalence primarily on the nasal side of the optic disc¹⁷. However, Sharma et al., using ultra-widefield fundus imaging, reported no significant quadrant-specific distribution differences⁴⁰. These discrepancies likely reflect methodological differences in imaging modalities and drusen definitions across studies. We employed late-phase ICGA, which offers superior sensitivity for drusen detection compared to conventional color fundus imaging¹⁵, potentially revealing early nasal drusen distribution patterns before they become visible on fundoscopy. Additionally, color fundus imaging often suffers from lower resolution in the nasal periphery due to stretching artifacts and defocus effects, which may underestimate nasal drusen burden.

We hypothesize that the greater distribution of hard drusen on the nasal side compared to the temporal side may reflect regional variations in lipoprotein clearance capacity. The widely accepted mechanism of drusen formation posits that RPE-secreted lipoproteins accumulate in the sub-RPE-basal lamina space due to impaired clearance across aged Bruch’s membrane (BrM)-choriocapillaris complex^41,42. Therefore, structural and functional differences in the BrM-choriocapillaris complex across peripheral quadrants may be key determinants of regional drusen distribution patterns. BrM surface roughness around the optic disc increases significantly with age in the nasal, while temporal changes remain modest⁴³, suggesting earlier structural remodeling. Furthermore, the landmark Beaver Dam Eye Study reported that pigmentary abnormalities were most likely to occur in the nasal or superior quadrants⁴⁴, findings corroborated by Heussen et al.,⁴⁵ indicating earlier metabolic stress in the nasal RPE-Bruch’s membrane complex. OCTA has enabled non-invasive assessment of peripheral choriocapillaris perfusion. Using 12 × 12 mm swept-source OCTA, Borrelli et al. demonstrated that the temporal region exhibited the highest perfusion in the near/mid periphery⁴⁶. Studies employing wider scanning protocols (20 × 23 mm SS-OCTA) further revealed significantly higher choroidal thickness in the temporal versus nasal field^47,48. Collectively, we propose that the nasal-predominant distribution of peripheral hard drusen observed in our study may result from the synergistic effects of (1) reduced choriocapillaris perfusion compromising metabolic waste clearance, and (2) earlier Bruch’s membrane structural remodeling disrupting trans-membrane lipoprotein transport. However, direct histopathological evidence is needed to validate these proposed mechanisms.

Prior studies have preliminarily investigated the correlation between peripheral hard drusen and AMD using subjective assessment methods. Sharma et al. and Seddon et al. reported significant associations between peripheral hard drusen and AMD presence based on UWF imaging and SD-OCT analyses^38,40. Heussen et al. found a higher prevalence of abnormal peripheral autofluorescence patterns in AMD patients versus controls (70% vs. 18%)⁴⁵. However, these studies focused on AMD eyes rather than normal eyes. The correlation between peripheral hard drusen and several macular pathologies was further explored in current study. Our results revealed that the central superior quadrant showed consistent associations with all forms of macular changes, including hard drusen counts, soft drusen (or drusenoid PED) volume, and progression to nAMD. This finding is difficult to explain given the scarcity of relevant mechanistic studies. Whether the superior quadrant shares similar pathophysiological mechanisms with the macula, or whether superior images provide better quality for drusen quantification remains uncertain. Nevertheless, our findings provide preliminary evidence that superior peripheral hard drusen counts may be associated with subsequent nAMD progression, suggesting the superior quadrant merits further investigation. A long-term prospective follow-up study is underway to track peripheral drusen evolution and their correlation with macular changes. We believe these longitudinal data will provide deeper insights into the prognostic significance of the superior quadrant and help elucidate the mechanisms linking peripheral and macular drusen patterns.

Multiple limitations need to be acknowledged. First, although longitudinal analysis was performed, the retrospective study design is subject to selection bias, and the limited follow-up data cannot capture the complete disease progression profile of the entire study population. Particularly in the soft drusen analysis, the small sample size with countable scattered hard drusen (n = 12) precluded further subgroup analyses or adjustment for potential confounding factors. No correction for multiple comparisons was applied across the nine quadrant-wise analyses. Accordingly, the reported associations should be regarded as exploratory and interpreted with caution. Second, the counting of hard drusen may not perfectly align with their actual quantity due to variations in fluorescence intensity of drusen with different internal components, potentially leading to the omission of some drusen with lower fluorescence intensity during machine recognition. However, the consistency between manual and machine counting reinforces the reliability of our data. Furthermore, the TWS model was trained and validated on single-centre ICGA images; performance on images from other acquisition systems or protocols has not been assessed, and quantitative findings should be interpreted within this methodological context. Additionally, eyes with confluent hard drusen were excluded from quantitative counting analyses, which may result in conservative estimates of the drusen–nAMD associations. Third, the older subjects in this study were limited due to significant refractive media opacities and low cooperation. Future studies should employ advanced imaging techniques with improved signal quality to include elderly patients who were excluded from our current analysis due to suboptimal ICGA image quality. Fourth, the longitudinal subgroup was limited to 246 followed eyes with 49 progression events and was retrospectively assembled. The association between superior peripheral hard drusen and nAMD progression thus warrants validation in well-designed prospective studies before its prognostic value can be fully established. Future prospective studies with comprehensive systemic and genetic profiling would provide more comprehensive insights. Fifth, the absence of matched color fundus or ultra-widefield retinal photographs represents an additional limitation. Although late-phase ICGA provides high-contrast visualization of hard drusen, the possibility that some hyperfluorescent spots detected by the algorithm may represent other retinal lesions cannot be entirely excluded. Future studies incorporating multimodal imaging would help validate the specificity of ICGA-based drusen detection across different quadrants and age groups. Despite these limitations, the strength of this research lies in its pioneering use of a TWS-based counting method to quantify hard drusen in normal eyes via ICGA images, revealing distribution characteristics across ages and quadrants, and providing valuable population-based evidence for future studies.

Conclusion

In conclusion, by training a machine learning model for drusen segmentation in late-phase ICGA using the TWS plugin, our study revealed three key findings. First, age 50 represents a critical demarcation point, with significant increases in drusen counts commonly observed across all quadrants. Second, the nasal quadrants demonstrated earlier and more prominent drusen accumulation compared to the temporal quadrants, suggesting regional differences in the aging process of the RPE-Bruch’s membrane-choriocapillaris complex. Third, the superior quadrant showed consistent associations with macular pathology. This exploratory finding, however, requires confirmation in larger prospective studies before any clinical utility can be established. In summary, these findings provide insights into the age-related and topographic distribution patterns of peripheral hard drusen and highlight the potential value of comprehensive peripheral retinal evaluation during ICGA examination.

Author contributions

Study concept and design: J.X. Pu and F. Wen. Acquisition, analysis, or interpretation of data: J.X. Pu, Y.Y. Ji, L. Mi, M.L. Li, X.Z. Zhang, C.G. Zuo, X.N. Zhuang, X.L. Hao, G.Q. He. Drafting of the manuscript: J.X. Pu. Critical revision of the manuscript for important intellectual content: F. Wen and L. Mi. Statistical analysis: J.X. Pu, L. Mi, Y.H. Gan, Y.Y. Su, Y.N. Zhang, R.J. Yang, X.L. Chen. Administrative, technical, or material support: Y.Y. Ji, L. Mi, M.L. Li, X.Z. Zhang, C.G. Zuo, X.N. Zhuang, X.L. Hao, G.Q. He and F. Wen. Study supervision: F. Wen.

Peer review

Peer review information

Communications Medicine thanks Tae Keun Yoo and Arun Govindaiah for their contribution to the peer review of this work.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Data availability

The source data for Fig. 5 is in Supplementary Data 1, and the source data for Fig. 6 is in Supplementary Data 2. Additional data supporting the findings of this study are not publicly available due to restrictions imposed by the institutional ethics committee of Zhongshan Ophthalmic Center and applicable data governance regulations. De-identified data may be made available to qualified researchers upon reasonable request and subject to institutional approval. Requests for data access should be directed to the corresponding author, Dr. Feng Wen MD, PhD (wenfeng208@foxmail.com).

Code availability

The machine learning code used for drusen segmentation and quantification in this study is not publicly available as it is the subject of a pending patent application. Further information regarding the methodology is available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s43856-026-01737-w.