OCTA Image-Based Machine Learning Models for Discriminating

Alzheimer’s Disease from Neurodegenerative and Ocular Conditions

Cunyi Xu

Leicester Grammar School, London Road, Great Glen, Leicester, LE8 9FL, U.K.

Keywords:

Alzheimer’s Disease, Optical Coherence Tomography Angiography, Machine Learning, Early Diagnosis.

Abstract:

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder that poses a signiﬁcant challenge, par-

ticularly as the global population ages. Timely diagnosis is crucial for managing AD, and this study aims to

contribute to early detection by analyzing Optical Coherence Tomography Angiography (OCTA) images us-

ing machine learning models. In this work, we leverage the structural and functional connections between the

eye and brain to enhance the discrimination of AD from other neurodegenerative and ocular conditions. We

also compiled a comprehensive dataset of OCTA images from various imaging devices, representing a range

of diseases. Using a pre-trained nnU-Net, we segmented vascular structures and calculated vascular density

metrics, while also extracting histogram and Gray-Level Co-occurrence Matrix (GLCM) features for texture

analysis. Various machine learning models were trained and evaluated through ﬁve-fold cross-validation, with

the Random Forest model achieving 78.15% accuracy in classifying multi-disease OCTA images. The model

exhibited high recall for stroke, diabetes, and age-related macular degeneration, but lower recall for AD, con-

genital heart disease, and hypertension, indicating potential misclassiﬁcation. Our ﬁndings emphasize the

utility of OCTA imaging and machine learning for early AD diagnosis, paving the way for future research to

reﬁne image processing and classiﬁcation methods.

1 INTRODUCTION

The burgeoning global elderly population has led to

a proportional rise in the prevalence of Alzheimer’s

disease (AD), a progressive neurodegenerative disor-

der that poses signiﬁcant healthcare challenges. With

current estimates suggesting that over 130 million in-

dividuals will be affected by 2050, the urgency to de-

velop effective diagnostic and therapeutic strategies

is paramount(Nichols et al., 2022). The economic

and social burden of AD is further exacerbated by

its increasing incidence among younger demograph-

ics, highlighting the necessity for early and accu-

rate detection methods to mitigate disease progres-

sion(Hassen et al., 2024). The quest for early diag-

nostic markers has led researchers to explore the intri-

cate relationship between the eye and the brain(Diogo

et al., 2022; Hassen et al., 2024). The retina’s struc-

tural and functional connections to the central nervous

system make it a viable window into brain health,

with preliminary abnormalities often manifesting in

the retina prior to the onset of neurological symp-

toms(O’Bryhim et al., 2018). This insight has spurred

Author of this article is the corresponding author.

interest in non-invasive ocular imaging techniques as

potential diagnostic tools for early AD detection.

Optical Coherence Tomography Angiography

(OCTA) has emerged as a pivotal technology in

this domain, offering high-resolution, non-invasive

imaging of the retinal and choroidal microvascula-

ture(Naseripour et al., 2020; Liu et al., 2024; Turkan

and Tek, 2023). OCTA’s ability to track red blood cell

movement within the vasculature provides a unique

perspective on the microvascular changes that may

precede AD-related neurological symptoms.

Despite the promising results of OCTA, exist-

ing diagnostic methods still face limitations, par-

ticularly in large-scale screenings and among el-

derly populations where coexisting eye conditions can

confound interpretations of OCTA images(Turkan

and Tek, 2022). Differentiating early signs of AD

from other retinal pathologies requires sophisticated

analytical techniques to improve diagnostic accu-

racy(Katsimpris et al., 2022).

Recent advancements in OCTA have facilitated a

deeper understanding of retinal vascular changes as-

sociated with AD. Studies have reported reduced reti-

nal vascular density in AD patients, suggesting a cor-

324

Xu, C.

OCTA Image-Based Machine Learning Models for Discriminating Alzheimer’s Disease from Neurodegenerative and Ocular Conditions.

DOI: 10.5220/0013141300003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 1, pages 324-331

ISBN: 978-989-758-731-3; ISSN: 2184-4305

relation with cerebral vascular alterations(Yoon et al.,

2024). Additionally, OCTA has shown potential in

distinguishing AD from other neurodegenerative dis-

eases and common eye conditions such as diabetic

retinopathy (DR) and age-related macular degenera-

tion (Li et al., 2023).

The academic community has increasingly recog-

nized the importance of retinal biomarkers in demen-

tia diagnosis, with OCTA playing a signiﬁcant role.

Systematic reviews and high-resolution imaging stud-

ies have begun to elucidate the morphological charac-

teristics of the retinal microvasculature in relation to

various diseases.

This paper aims to build on these foundations by

exploring the utility of OCTA in differentiating AD

from other eye diseases and neurodegenerative con-

ditions. We focus on quantitatively assessing the dis-

tinguishability of OCTA features associated with AD

and other diseases to enhance early diagnostic accu-

racy. Through the application of machine learning

algorithms to a comprehensive dataset of OCTA im-

ages, we seek to provide new insights and methods for

early AD detection, potentially alleviating the societal

impact of this pervasive disease.

2 RELATED WORKS

2.1 Vascular Density and Vessel

Segmentation

Optical Coherence Tomography Angiography

(OCTA) often uses Blood Vessel Density (BVD) as a

key metric because it is simple to calculate and useful

for assessing the severity of eye diseases like Diabetic

Retinopathy (DR) and Glaucoma. The process of

deriving BVD from OCTA images is considered a

segmentation task, making deep learning techniques

such as U-Net particularly suitable. For instance,

Convolutional Neural Network (CNN) models have

been successfully applied to segment choroidal blood

vessels, which is more challenging than retinal seg-

mentation due to the higher density of larger vessels

in the choroid and greater signal loss in deeper

tissues. Liu et al. has shown that BVD measurements

obtained through this method are comparable to those

from manual segmentation(Liu et al., 2019).

Furthermore, various U-Net models have helped

standardize BVD measurements across different de-

vices and improve resistance to projection artifacts.

Vessel segmentation can also be improved with im-

age enhancement techniques, as better image quality

increases the contrast between blood ﬂow and back-

ground pixels. Like most OCTA measurement tech-

niques, these methods can be applied to en-face im-

ages, which are essential for accurate anatomical slice

segmentation. Another approach, the Image Projec-

tion Network (IPN), utilizes CNNs to extract fea-

tures and perform projections simultaneously. It has

been tested on two key areas of retinal image seg-

mentation—retinal vessel segmentation and Foveal

Avascular Zone (FAZ) segmentation. Results indicate

that the IPN is an effective implementation of 3D to

2D segmentation networks, outperforming standard

methods(Li et al., 2020).

2.2 Biomarkers and Features for

Developing Computer-Aided

Diagnostic Systems

The combination of biomarkers and unique features

is crucial for creating CAD systems. Traditional

methods like logistic regression have proven effec-

tive, and machine learning models are also viable op-

tions. CAD systems that utilize OCTA data have been

designed to DR at various severity levels (Krittana-

wong et al., 2020). Research has shown that using

a Random Forest model with inputs such as struc-

tural OCT reﬂectivity, curvature and thickness, OCTA

vascular density, the number of vascular bifurcations

and crossings, and the size of the Foveal Avascular

Zone (FAZ) has resulted in an area under the receiver

operating characteristic curve exceeding 95%. This

performance was key in differentiating eyes with DR

from those without and in classifying the conditions

of patients with non-proliferative DR.

Additionally, studies that combined clinical in-

dicators with OCTA features, using a hybrid ap-

proach of a balanced optimizer and Support Vec-

tor Machine for feature selection and classiﬁcation,

have produced promising results for Vogt-Koyanagi-

Harada syndrome (Dhodapkar et al., 2022). Concur-

rently, research has explored deep learning systems

that integrate macular Blood Vessel Density (BVD)

with the thickness of the ganglion cell layer/inner

plexiform layer. The combination of these metrics,

learned through AI, has outperformed individual mea-

surements.

Deep learning has also been applied to the diag-

nosis of Age-Related Macular Degeneration (AMD),

with studies indicating that CNN can accurately pre-

dict the progression of AMD from intermediate to ad-

vanced stages (Peng et al., 2020). Moreover, as pre-

viously mentioned, CNNs’ ability to detect Choroidal

Neovascularization (CNV) makes deep learning a vi-

able option for classifying neovascular patterns in

AMD. The effectiveness of these CAD systems re-

lies on the quality of input feature measurements;

OCTA Image-Based Machine Learning Models for Discriminating Alzheimer’s Disease from Neurodegenerative and Ocular Conditions

325

therefore, continuous improvements in feature quan-

tiﬁcation will support the development of CAD sys-

tems. However, their overall performance remains

constrained by predeﬁned functionalities.

2.3 AI-Enhanced Analysis of OCTA

Images and Disease Diagnosis

AI-based diagnostic systems utilizing OCTA imag-

ing are widely employed for assessing retinal diseases

and identifying various eye conditions. However,

these AI diagnostic tools, despite being validated, en-

counter numerous challenges. Firstly, the resolution

and details of retinal blood vessels captured by fun-

dus images are limited, which hampers the ability to

accurately measure the curvature of microvessels near

the fovea and across different retinal layers (Dhodap-

kar et al., 2022).

Additionally, the quality of OCTA images can

vary due to factors like eye movement, eyelid ob-

struction, and artifacts from optical coherence tomog-

raphy, which signiﬁcantly affects manual annotations

and leads to inconsistencies in subjective assessments

(Lauermann et al., 2019). Consequently, diagnostic

systems that rely on supervised machine learning al-

gorithms struggle with poor performance in quantita-

tive feature analysis and lack of diagnostic accuracy.

Furthermore, understanding the reasoning behind

AI model decisions is crucial for clinical disease di-

agnosis, and providing clear interpretability of al-

gorithms will be essential for the clinical integra-

tion of AI-assisted OCTA disease diagnosis. More-

over, deep learning systems require large and veriﬁ-

able databases, ranging from 100,000 to several mil-

lion images, for effective training and optimization

(Le et al., 2024). Even after successfully training AI

systems, differences among databases from various

imaging centers make it extremely difﬁcult to provide

reliable accuracy metrics (Yang et al., 2023). At the

same time, creating large multicenter databases for

the effective use of AI diagnostic tools presents a sig-

niﬁcant challenge, especially for new retinal imaging

techniques like OCTA. Nevertheless, AI is expected

to help realize the full potential of OCTA imaging.

2.4 Retinal Neurodegeneration and

Vascular Alterations in AD

Studies using OCT have shown considerable retinal

neurodegeneration in individuals with AD. A sys-

tematic review conducted in 2018 found that patients

with AD exhibited thinning in several retinal layers,

including the peripapillary retinal nerve ﬁber layer

(pRNFL), macular ganglion cell inner plexiform layer

(mGCIPL), ganglion cell complex, and choroidal lay-

ers, as well as a decrease in overall macular volume

and thinning in both inner and outer macular regions

(Chan et al., 2019).

The thinning of the mGCIPL has been linked to

the severity of the disease (Ferrari et al., 2017). Var-

ious ﬂuorescence-based fundus imaging techniques

have been used to visualize and measure retinal

pathology associated with AD. For instance, adminis-

tering curcumin, a beta-amyloid binding ﬂuorophore,

intravenously resulted in a twofold increase in reti-

nal β-amyloid levels in AD patients. These retinal

β-amyloid levels have been associated with cortical

β-amyloid levels and reductions in hippocampal vol-

ume (Koronyo et al., 2017). Moreover, blue autoﬂu-

orescence has been employed to measure the surface

area of retinal inclusions related to preclinical cortical

β-amyloid burden (Snyder et al., 2016). Lastly, ﬂuo-

rescence lifetime imaging ophthalmoscopy has shown

differences between AD patients and matched con-

trols.

Retinal vascular alterations have also been ob-

served in fundus images of AD patients. The fractal

dimension (FD), which quantitatively represents the

complexity of vascular branching, can be assessed us-

ing commercial software or expert evaluation. A sys-

tematic review from 2019 reported a decrease in vas-

cular FD among AD patients across four case-control

studies involving fundus imaging (Lemmens et al.,

2020). Additionally, one study noted increased vas-

cular tortuosity and narrowed venous diameters in AD

patients, although another study presented conﬂicting

results (Cheung et al., 2014). It can be seen that in-

dividuals with AD have an enlarged central arterio-

lar avascular zone and reduced overall superﬁcial and

deep retinal vascular density.

3 FRAMEWORK FOR EARLY

DIAGNOSIS OF ALZHEIMER’S

DISEASE

3.1 Overview

The overarching goal of this study is to harness the

potential of OCTA for early diagnosis of AD by

differentiating it from other eye diseases and neu-

rodegenerative conditions. Our methodology encom-

passes the following key steps: First, we assembled

a diverse dataset of OCTA images to ensure repre-

sentation of various imaging devices, disease states,

and image resolutions. Second, we employ a pre-

trained neural network, nnU-Net, we segmented vas-

BIOIMAGING 2025 - 12th International Conference on Bioimaging

326

cular structures from OCTA images to facilitate the

extraction of vascular density metrics. Next, we ex-

tracted both vascular and image features from the

segmented OCTA images, including overall vascu-

lar density and regional vascular density based on

the Early Treatment Diabetic Retinopathy Study (ET-

DRS) grid. Finally, A suite of machine learning algo-

rithms was trained on the extracted features to classify

images into AD and control groups, employing a ﬁve-

fold cross-validation approach to ensure robustness

of the models. The workﬂow of our methodology,

encapsulating dataset compilation, image segmenta-

tion, feature extraction, and machine learning model

training and validation, is schematically represented

in Figure 1.

3.2 Dataset

Our dataset is a curated collection of 2000 OCTA im-

ages sourced from multiple open datasets: OCTA-

500 (Li et al., 2024), OCTAGON (D

ıaz et al., 2019),

Foveal Avascular Zone Image Database (FAZID)

(Agarwal et al., 2020), Soul (Xue et al., 2024) and

DRAC2022 (Qian et al., 2024), providing a wide

spectrum of disease representations. The dataset in-

cludes images from patients with varying conditions,

such as stroke, diabetes, age-related macular degen-

eration (AMD), and AD, as well as a control group.

The image resolutions vary, including 3x3 mm², 6x6

mm², 8x8 mm², and 12x12 mm², to simulate real-

world clinical scenarios.

3.3 Image Segmentation

nnU-Net is an adaptive neural network architecture

that automatically adjusts its structure based on the

speciﬁc task. In this research, a version of the U-

Net architecture tailored for OCTA image segmenta-

tion was chosen. Utilizing nnU-Net to segment OCTA

images from various devices, sizes, and disease types

allows for further feature extraction and pathological

analysis. The segmentation network discussed in this

chapter uses the OCTA-500 dataset, which consists of

500 OCTA images. Loss Function: A combination of

cross-entropy and Dice loss functions was used to ad-

dress class imbalance and improve segmentation ac-

curacy.

The cross-entropy loss function focuses on pixel-

level classiﬁcation, while the Dice loss function em-

phasizes overall shape matching. Optimizer and

Learning Rate: The Adam optimizer was utilized for

model training, starting with a learning rate of 0.001,

along with a learning rate decay strategy to adjust

the learning rate dynamically during training. This

method helps the model converge towards the global

optimum. Training Strategy: The model was trained

using a ﬁve-fold cross-validation approach to evaluate

its stability and generalizability. In each fold, 80% of

the data was used for training, and 20% was set aside

for validation.

3.4 Feature Extraction

The objective of this study is to harness the diagnos-

tic potential of OCTA images for early detection of

AD by differentiating it from other eye diseases. A

critical step in achieving this objective is the extrac-

tion of meaningful features from OCTA images that

can serve as inputs for machine learning algorithms.

This section outlines our approach to feature extrac-

tion, focusing on vascular density measurements and

texture analysis.

3.4.1 Vascular Density Calculation

Following the segmentation of OCTA images, vascu-

lar density is calculated to quantify the vascular pat-

terns associated with different disease states. This

process involves two main steps, which are Overall

Image Vascular Density and ETDRS Regional Vascu-

lar Density. Overall Image Vascular Density is deter-

mined by calculating the ratio of vascular pixels to the

total number of pixels in the OCTA image, providing

a global measure of vascularization. Meanwhile, The

ETDRS grid is used to divide the retina into nine dis-

tinct areas. Vascular density is calculated separately

for each of these regions, enabling a detailed analy-

sis of local vascular patterns. This regional analysis

is crucial for early disease detection and accurate di-

agnosis, as it allows for the identiﬁcation of localized

vascular changes that may not be apparent in overall

density measurements.

3.4.2 Image Feature Extraction

In addition to vascular density, texture analysis plays

a vital role in characterizing OCTA images. We focus

on two types of image features: histogram features

and Gray-Level Co-occurrence Matrix (GLCM) fea-

tures. For histogram feature extraction, Image his-

tograms provide a statistical representation of gray

level distribution, offering insights into brightness and

contrast. In this study, 16 histogram features are de-

rived by quantizing image gray levels into 16 bins and

counting the pixel occurrences within each bin. These

features are indicative of the brightness distribution

and contrast information in OCTA images, which are

essential for disease analysis.

OCTA Image-Based Machine Learning Models for Discriminating Alzheimer’s Disease from Neurodegenerative and Ocular Conditions

327

Figure 1: Schematic representation of the methodology for early diagnosis of Alzheimer’s disease using OCTA images. The

process includes dataset compilation, image preprocessing, segmentation using nnU-Net, feature extraction, model training,

validation, and performance evaluation.

On the other hand, The GLCM is a texture anal-

ysis technique that captures spatial relationships be-

tween gray levels in an image. We extract ﬁve GLCM

features—contrast, similarity, homogeneity, energy,

and correlation—to characterize the texture structure

of OCTA images. These features provide a quantita-

tive description of vascular structures, which is vital

for subsequent machine learning classiﬁcation tasks.

3.4.3 Machine Learning Model

The extracted vascular and image features provide a

rich set of information for machine learning models,

aiding in achieving better performance in classiﬁca-

tion tasks. This study involved training and eval-

uating various machine learning models for multi-

disease classiﬁcation. To ensure optimal model per-

formance, a range of common machine learning mod-

els were selected, and hyperparameter tuning was

performed for each. The models included Random

Forest, Support Vector Machine (SVM), K-Nearest

Neighbors (KNeighbors), Naive Bayes, Logistic Re-

gression, Decision Tree, Gradient Boosting, and Mul-

tilayer Perceptron (MLP). Hyperparameter tuning for

each model was conducted using GridSearchCV.

4 EVALUATION

Each model was trained, and a ﬁve-fold cross-

validation strategy was employed to evaluate model

performance. Metrics such as accuracy, macro pre-

cision, macro recall, macro F1 score, weighted pre-

cision, weighted recall, and weighted F1 score were

used for assessment.

The results indicate that the Random Forest model

outperformed all other models, achieving an accu-

racy rate of 78.37%. To understand these results,

it is essential to analyze the performance of each

model in conjunction with the OCTA modality and

the selected features. The Random Forest algorithm,

which integrates multiple decision trees, effectively

handles high-dimensional data and mitigates the risk

of overﬁtting. The rich vascular and image features

from OCTA images allow the Random Forest to cap-

ture complex relationships between features, thus ex-

celling in classiﬁcation tasks.

The Gradient Boosting model, which optimizes

incrementally, also performs admirably by captur-

ing intricate feature relationships. This model ex-

cels in processing nonlinear features, and when com-

bined with the diverse features from OCTA images, it

achieves commendable classiﬁcation results. In con-

trast, Logistic Regression, which is adept at handling

linear features, exhibits limitations when dealing with

complex nonlinear characteristics. Although the Mul-

tilayer Perceptron possesses strong nonlinear model-

ing capabilities, its performance in this study may be

suboptimal due to the inﬂuence of feature selection

and model complexity on the given feature data.

In this study, we aimed to explore the distin-

guishability between AD and other common oph-

thalmic and chronic diseases. Based on the assess-

ment using the Random Forest model, we provided

a detailed analysis of metrics and a confusion ma-

trix. Analysis of the results, as shown in Table II,

revealed that AD, congenital heart disease, and hy-

pertension are more likely to be confused with other

BIOIMAGING 2025 - 12th International Conference on Bioimaging

328

Table 1: Classiﬁcation results of various machine learning models for OCTA image classiﬁcation.

Model Accuracy (%) Macro Precision (%) Macro Recall (%) Macro F1 (%) Weighted Precision (%) Weighted Recall (%) Weighted F1 (%)

Decision Tree 65.89 58.32 51.34 52.19 61.16 65.89 61.42

KNeighbors 72.57 59.11 55.55 55.29 71.05 72.57 70.97

Gradient Boosting 76.02 69.52 62.13 63.54 74.50 76.02 74.38

Random Forest 78.37 73.97 70.07 69.08 80.17 76.07 76.81

Logistic Reg 67.82 61.14 52.15 49.87 65.74 67.82 63.16

Naive Bayes 59.20 37.04 53.29 40.31 51.23 59.20 53.06

MLP 63.54 43.58 42.95 40.67 55.40 63.54 55.92

SVC 64.51 43.51 39.72 37.53 55.24 64.51 55.53

Table 2: Performance metrics of the Random Forest model

in classifying various diseases from OCTA images.

Disease Precision (%) Recall (%) F1-score (%)

AD 72.73 47.06 57.14

AMD 91.07 94.44 92.73

CHD 100.00 40.00 57.14

Control 93.92 74.73 83.23

DM 68.57 78.69 73.28

DR 63.93 79.59 70.91

HP 62.00 72.09 66.67

Stroke 39.53 73.91 51.52

diseases during classiﬁcation, attributed to their lower

recall rates, suggesting a higher likelihood of these

class samples being misclassiﬁed. Meanwhile, stroke,

diabetes, and AMD demonstrated higher recall rates,

indicating that these diseases are more readily and ac-

curately identiﬁed by the model, with less confusion

with other categories. Notably, the high recall rate for

stroke highlights the model’s superior performance in

identifying this disease, with minimal confusion with

other diseases.

Figure 2 presents the confusion matrix, visually

revealing the relationships between the model’s pre-

dicted categories and the actual categories. An in-

depth analysis of the confusion matrix not only un-

covers the model’s performance variances in identify-

ing various diseases but also exposes potential short-

comings, providing a basis for further model opti-

mization. The confusion matrix also aids in under-

standing the similarities and differences in features

between different diseases, thereby assisting in the

precision of clinical diagnoses. Additionally, we paid

particular attention to the classiﬁcation of the control

group, ﬁnding its recall rate to be relatively low, with

samples from the control group tending to be mis-

classiﬁed into categories such as diabetes, diabetic

retinopathy, hypertension, and stroke.

5 DISCUSSION

AD is a long-term neurodegenerative disorder that of-

ten results in dementia among older adults. How-

ever, diagnosing AD early is difﬁcult due to the lack

AMD

CHD

Control

Stroke

AMD

CHD

Control

Stroke

8 2 0 0 0 0 0 7

1 51 0 0 0 2 0 0

0 0 4 1 3 2 0 0

0 1 0 68 4 13 11 18

0 1 0 2 48 4 6 0

0 0 0 1 6 39 2 1

0 0 0 2 9 1 31 0

2 1 0 3 0 0 0 75

Figure 2: The confusion matrix displays the performance

of the Random Forest model in distinguishing between

Alzheimer’s disease (AD), age-related macular degenera-

tion (AMD), congenital heart disease (CHD), control group,

diabetes mellitus (DM), diabetic retinopathy (DR), hyper-

tension (HP), and stroke.

of deﬁnitive diagnostic tools. The retina, sharing de-

velopmental and structural similarities with the brain,

offers a unique perspective on brain pathology, as

changes in the retina may occur before clinical symp-

toms of AD appear. Thus, creating a model that links

ocular and brain structural characteristics to the dis-

ease could help identify early retinal changes, paving

the way for improved early screening and diagnosis

of AD.

OCTA is vital for examining retinal microvascu-

lature, but the variability of existing image analysis

tools hinders its wider use in research and clinical set-

tings. To address this issue, this study utilized ma-

chine learning algorithms for multi-disease classiﬁca-

tion of OCTA images. The results showed that the

Random Forest model achieved the highest perfor-

mance, with an accuracy of 78.37%. This highlights

the effectiveness of ensemble learning in managing

high-dimensional medical imaging data and minimiz-

ing overﬁtting risks.

Despite these advancements, differentiating be-

tween diseases remains a challenge. Speciﬁcally, AD,

OCTA Image-Based Machine Learning Models for Discriminating Alzheimer’s Disease from Neurodegenerative and Ocular Conditions

329

congenital heart disease, and hypertension showed

lower recall rates, indicating a tendency for mis-

classiﬁcation, possibly due to less distinct features

in OCTA images or similarities with other diseases.

Conversely, stroke, diabetes, and AMD had higher

recall rates, suggesting that these conditions exhibit

more distinct features in OCTA images, resulting in

better model accuracy for their identiﬁcation.

The results of this study not only aid clinicians in

gaining a clearer understanding of disease character-

istics in OCTA images but also open new avenues for

future research. This includes investigating innova-

tive image processing methods to extract more repre-

sentative features and exploring different algorithmic

combinations to further improve classiﬁcation accu-

racy and reliability. These initiatives will establish a

stronger basis for early diagnosis and treatment, ulti-

mately enhancing patient outcomes.

6 THREATS TO VALIDITY

This study faced several challenges while using

OCTA for diagnosing ocular diseases, neurodegener-

ative disorders, and systemic conditions. One major

limitation is the intricate relationship between the su-

perﬁcial and deep retinal blood vessels, which ham-

pers the thorough detection of vascular changes asso-

ciated with various diseases. Additionally, the quality

of OCTA images can vary due to factors like differ-

ences in equipment, patient cooperation, eye move-

ments, and eyelid interference, making accurate im-

age analysis difﬁcult. The presence of similar vascu-

lar changes in OCTA images across different diseases

complicates the ability to distinguish between them.

The effects of neurodegenerative and systemic

diseases extend beyond what OCTA can reveal about

vascular structure. The varying stages of disease pro-

gression may also mean that OCTA does not capture

early or subtle vascular changes. The absence of lon-

gitudinal data restricts the monitoring of disease pro-

gression. The technology for automated quantitative

analysis of vascular changes is still being reﬁned, with

improvements needed in accuracy and reproducibil-

ity. Machine learning models also struggle with gen-

eralizability across different populations or devices.

There is an urgent need for more validation studies

and standardized diagnostic criteria for the clinical

use of OCTA. The high cost of the equipment lim-

its its accessibility in resource-constrained environ-

ments.

When creating multicenter databases, challenges

related to data sharing and patient privacy must be ad-

dressed. Finally, the interpretability of AI-assisted di-

agnostic results and the reliability of clinical decision-

making require further validation by healthcare pro-

fessionals. These limitations indicate that while

OCTA technology is a valuable asset for disease re-

search and diagnosis, additional research and tech-

nological advancements are necessary to address the

current challenges in its practical application.

7 CONCLUSION

In this pilot study, we have explored the early di-

agnostic potential of Optical Coherence Tomography

Angiography (OCTA) imaging enhanced by machine

learning for Alzheimer’s Disease (AD) and its dif-

ferentiation from other neurodegenerative and ocular

disorders. Our analysis of a diverse OCTA dataset us-

ing a pre-trained nnU-Net for segmentation and fea-

ture extraction revealed that the Random Forest model

achieved a commendable classiﬁcation accuracy of

78.15%. While acknowledging the possibility of mis-

classiﬁcation among AD, congenital heart disease,

and hypertension, our results highlight the promising

role of OCTA and machine learning in early AD di-

agnosis.

As future works, we plan to focus on reﬁning im-

age processing techniques to extract more discrimina-

tive features and exploring advanced machine learn-

ing algorithms to enhance classiﬁcation accuracy.

Additionally, addressing the challenges of varying

image resolutions and database inconsistencies across

different imaging centers will be crucial for fully har-

nessing the diagnostic potential of OCTA imaging.

REFERENCES

Agarwal, A., Raman, R., Lakshminarayanan, V., et al.

(2020). The foveal avascular zone image database

(fazid). In Applications of Digital Image Processing

XLIII, volume 11510, pages 507–512. SPIE.

Chan, V. T., Sun, Z., Tang, S., Chen, L. J., Wong, A., Tham,

C. C., Wong, T. Y., Chen, C., Ikram, M. K., Whit-

son, H. E., et al. (2019). Spectral-domain oct mea-

surements in alzheimer’s disease: a systematic review

and meta-analysis. Ophthalmology, 126(4):497–510.

Cheung, C. Y.-l., Ong, Y. T., Ikram, M. K., Ong, S. Y.,

Li, X., Hilal, S., Catindig, J.-A. S., Venketasubra-

manian, N., Yap, P., Seow, D., et al. (2014). Mi-

crovascular network alterations in the retina of pa-

tients with alzheimer’s disease. Alzheimer’s & De-

mentia, 10(2):135–142.

Dhodapkar, R. M., Li, E., Nwanyanwu, K., Adelman, R.,

Krishnaswamy, S., and Wang, J. C. (2022). Deep

learning for quality assessment of optical coherence

tomography angiography images. Scientiﬁc reports,

12(1):13775.

BIOIMAGING 2025 - 12th International Conference on Bioimaging

330

ıaz, M., Novo, J., Cutr

ın, P., G

omez-Ulla, F., Penedo,

M. G., and Ortega, M. (2019). Automatic segmenta-

tion of the foveal avascular zone in ophthalmological

oct-a images. PloS one, 14(2):e0212364.

Diogo, V. S., Ferreira, H. A., Prata, D., and Initiative, A.

D. N. (2022). Early diagnosis of alzheimer’s disease

using machine learning: a multi-diagnostic, general-

izable approach. Alzheimer’s Research & Therapy,

14(1):107.

Ferrari, L., Huang, S.-C., Magnani, G., Ambrosi, A., Comi,

G., and Leocani, L. (2017). Optical coherence tomog-

raphy reveals retinal neuroaxonal thinning in fron-

totemporal dementia as in alzheimer’s disease. Jour-

nal of Alzheimer’s Disease, 56(3):1101–1107.

Hassen, S. B., Neji, M., Hussain, Z., Hussain, A., Alimi,

A. M., and Frikha, M. (2024). Deep learning meth-

ods for early detection of alzheimer’s disease using

structural mr images: A survey. Neurocomputing,

576:127325.

Katsimpris, A., Karamaounas, A., Sideri, A. M., Katsim-

pris, J., Georgalas, I., and Petrou, P. (2022). Opti-

cal coherence tomography angiography in alzheimer’s

disease: A systematic review and meta-analysis. Eye,

36(7):1419–1426.

Koronyo, Y., Biggs, D., Barron, E., Boyer, D. S., Pearlman,

J. A., Au, W. J., Kile, S. J., Blanco, A., Fuchs, D.-T.,

Ashfaq, A., et al. (2017). Retinal amyloid pathology

and proof-of-concept imaging trial in alzheimer’s dis-

ease. JCI insight, 2(16).

Krittanawong, C., Virk, H. U. H., Bangalore, S., Wang,

Z., Johnson, K. W., Pinotti, R., Zhang, H., Kaplin,

S., Narasimhan, B., Kitai, T., et al. (2020). Machine

learning prediction in cardiovascular diseases: a meta-

analysis. Scientiﬁc reports, 10(1):16057.

Lauermann, J., Treder, M., Alnawaiseh, M., Clemens, C.,

Eter, N., and Alten, F. (2019). Automated oct angiog-

raphy image quality assessment using a deep learning

algorithm. Graefe’s Archive for Clinical and Experi-

mental Ophthalmology, 257:1641–1648.

Le, D., Son, T., Kim, T.-H., Adejumo, T., Abtahi, M.,

Ahmed, S., Rossi, A., Ebrahimi, B., Dadzie, A., Ma,

G., et al. (2024). Deep learning-based optical coher-

ence tomography angiography image construction us-

ing spatial vascular connectivity network. Communi-

cations Engineering, 3(1):28.

Lemmens, S., Devulder, A., Van Keer, K., Bierkens, J.,

De Boever, P., and Stalmans, I. (2020). Systematic

review on fractal dimension of the retinal vasculature

in neurodegeneration and stroke: assessment of a po-

tential biomarker. Frontiers in neuroscience, 14:16.

Li, M., Chen, Y., Ji, Z., Xie, K., Yuan, S., Chen, Q., and Li,

S. (2020). Image projection network: 3d to 2d image

segmentation in octa images. IEEE Transactions on

Medical Imaging, 39(11):3343–3354.

Li, M., Huang, K., Xu, Q., Yang, J., Zhang, Y., Ji, Z., Xie,

K., Yuan, S., Liu, Q., and Chen, Q. (2024). Octa-500:

a retinal dataset for optical coherence tomography an-

giography study. Medical image analysis, 93:103092.

Li, Y., El Habib Daho, M., Conze, P.-H., Zeghlache, R.,

Le Boit

e, H., Bonnin, S., Cosette, D., Magazzeni, S.,

Lay, B., Le Guilcher, A., et al. (2023). Hybrid fusion

of high-resolution and ultra-wideﬁeld octa acquisi-

tions for the automatic diagnosis of diabetic retinopa-

thy. Diagnostics, 13(17):2770.

Liu, X., Bi, L., Xu, Y., Feng, D., Kim, J., and Xu, X.

(2019). Robust deep learning method for choroidal

vessel segmentation on swept source optical coher-

ence tomography images. Biomedical Optics Express,

10(4):1601–1612.

Liu, X., Zhu, H., Zhang, H., and Xia, S. (2024). The frame-

work of quantifying biomarkers of oct and octa im-

ages in retinal diseases. Sensors, 24(16):5227.

Naseripour, M., Ghasemi Falavarjani, K., Mirshahi, R., and

Sedaghat, A. (2020). Optical coherence tomography

angiography (octa) applications in ocular oncology.

Eye, 34(9):1535–1545.

Nichols, E., Steinmetz, J. D., Vollset, S. E., Fukutaki, K.,

Chalek, J., Abd-Allah, F., Abdoli, A., Abualhasan, A.,

Abu-Gharbieh, E., Akram, T. T., et al. (2022). Esti-

mation of the global prevalence of dementia in 2019

and forecasted prevalence in 2050: an analysis for the

global burden of disease study 2019. The Lancet Pub-

lic Health, 7(2):e105–e125.

O’Bryhim, B. E., Apte, R. S., Kung, N., Coble, D., and

Van Stavern, G. P. (2018). Association of preclin-

ical alzheimer disease with optical coherence tomo-

graphic angiography ﬁndings. JAMA ophthalmology,

136(11):1242–1248.

Peng, Y., Keenan, T. D., Chen, Q., Agr

on, E., Allot, A.,

Wong, W. T., Chew, E. Y., and Lu, Z. (2020). Predict-

ing risk of late age-related macular degeneration using

deep learning. NPJ digital medicine, 3(1):111.

Qian, B., Chen, H., Wang, X., Guan, Z., Li, T., Jin, Y., Wu,

Y., Wen, Y., Che, H., Kwon, G., et al. (2024). Drac

2022: A public benchmark for diabetic retinopathy

analysis on ultra-wide optical coherence tomography

angiography images. Patterns, 5(3).

Snyder, P. J., Johnson, L. N., Lim, Y. Y., Santos, C. Y.,

Alber, J., Maruff, P., and Fern

andez, B. (2016).

Nonvascular retinal imaging markers of preclinical

alzheimer’s disease. Alzheimer’s & Dementia: Diag-

nosis, Assessment & Disease Monitoring, 4:169–178.

Turkan, Y. and Tek, F. B. (2022). A survey on automated

diagnosis of alzheimer’s disease using optical coher-

ence tomography and angiography. arXiv preprint

arXiv:2209.03354.

Turkan, Y. and Tek, F. B. (2023). Automated diagnosis of

ad using oct and octa: A systematic review. Authorea

Preprints.

Xue, J., Feng, Z., Zeng, L., Wang, S., Zhou, X., Xia, J.,

and Deng, A. (2024). Soul: An octa dataset based on

human machine collaborative annotation framework.

Scientiﬁc Data, 11(1):838.

Yang, D., Ran, A. R., Nguyen, T. X., Lin, T. P., Chen, H.,

Lai, T. Y., Tham, C. C., and Cheung, C. Y. (2023).

Deep learning in optical coherence tomography an-

giography: Current progress, challenges, and future

directions. Diagnostics, 13(2):326.

Yoon, J. M., Lim, C. Y., Noh, H., Nam, S. W., Jun, S. Y.,

Kim, M. J., Song, M. Y., Jang, H., Kim, H. J., Seo,

S. W., et al. (2024). Enhancing foveal avascular zone

analysis for alzheimer’s diagnosis with ai segmenta-

tion and machine learning using multiple radiomic

features. Scientiﬁc Reports, 14(1):1841.

OCTA Image-Based Machine Learning Models for Discriminating Alzheimer’s Disease from Neurodegenerative and Ocular Conditions

331