Next Article in Journal
A Novel Method for Analyzing the Kinetics of Convective/IR Bread Drying (CIRD) with Sensor Technology
Previous Article in Journal
SMILE: Segmentation-Based Centroid Matching for Image Rectification via Aligning Epipolar Lines
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Applied Sciences
Volume 15
Issue 9
10.3390/app15094963
applsci-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Perspective

Alzheimer’s Disease Detection from Retinal Images Using Machine Learning and Deep Learning Techniques: A Perspective

by
Adilet Uvaliyev
and
Leanne Lai Hang Chan
*
Department of Electrical Engineering, City University of Hong Kong, Kowloon, Hong Kong SAR, China
*
Author to whom correspondence should be addressed.
Appl. Sci. 2025, 15(9), 4963; https://doi.org/10.3390/app15094963
Submission received: 18 February 2025 / Revised: 17 April 2025 / Accepted: 17 April 2025 / Published: 30 April 2025
Versions Notes

Abstract

Alzheimer’s disease (AD) is a neurodegenerative disease that results in a loss of cognitive functions. The early discovery of it can potentially stop or decrease the severity of AD. Extensive research has been conducted to find AD biomarkers. In recent years, due to the development of AI technologies and the ease of obtaining retinal images, various machine learning (ML)- and deep learning (DL)-based methods of identifying AD patients from these images have been proposed. These models are significant as they represent a potential screening tool for AD and a tool for identifying biomarkers from retinal images. This paper reviews the recent progress in this direction. It presents an overview of relevant methods and analyzes their strengths and limitations. Also, it discusses common challenges and possible future directions related to this topic.

1. Introduction

Alzheimer’s disease (AD) is a brain disease that results in the loss of cognitive functions. Aggregation of protein amyloid and tau in the brain is considered a distinguishing characteristic of AD [1]. Its symptoms include problems with memory such as difficulties recalling recent dialogues, issues with communication, and impaired decision-making [1]. At an advanced level, daily activities such as walking and eating are also affected by AD [1]. It is estimated that after AD diagnosis, patients live around 5.7 years on average [2]. The World Alzheimer’s Report 2018 calculated that the overall cost due to Alzheimer’s disease is USD 1 trillion per year. This encompasses costs such as nursing and hospital care. The total number of people with AD was estimated to be 416 million [3]. This number is much greater than estimates in other studies because it includes people with preclinical AD in the calculation. Overall, these statistics show the significance of the problems that have arisen due to AD.
There is evidence suggesting that brain changes due to AD begin 20 years before the first clinical symptoms [4]. This shows the possibility of early detection of AD before the start of first symptoms. Early discovery of AD in patients, before the onset of symptoms, can allow for preventive measures that can alleviate AD-related problems. In a recent review, Yu et al. [5] identified 10 factors such as cognitive activity and stress that have strong evidence of their effectiveness in AD intervention. Furthermore, it has been estimated that a 10–25% decrease in AD risk factors such as smoking and obesity might even reduce 1.1–3 million AD cases globally [6]. These studies provide a lot of motivation for discovering biomarkers for early detection of AD.
Since AD is a global problem affecting millions of people, affordable and easy methods for screening AD are needed. The retina is a potential option for achieving this goal. This is because the retina is considered a visible extension of the central nervous system (CNS) with regard to embryology and anatomy [7,8]. The retina’s ganglion cells exhibit similar characteristics to CNS neurons, and it’s axons constitute the optic nerve [9]. Various neurodegenerative diseases exhibit symptoms in the retina [7,10]. There are studies showing that retina is affected in patients with AD [11]. For example, decreased retinal nerve fiber layer (RNFL) thickness in patients with AD and MCI has been reported [12]. Also, there are studies showing the presence of amyloid plaques in the retina of individuals with AD [13]. These studies illustrate the relationship between the brain and retina, making the retina a potential biomarker for AD detection. Images of the retina are easy to acquire. For example, optical coherence tomography (OCT) machines are available in many ophthalmologic centers. OCT is a non-invasive imaging modality that generates cross-sectional images and can be used to visualize retinal layers [14]. Moreover, fundus images can be acquired even with a phone camera.
In the past, AI techniques have shown good performance in the detection of various ocular diseases from retinal images such as diabetic retinopathy [15] and glaucoma [16,17]. Moreover, retinal images have also been used to predict other diseases such as kidney disease [18]. However, the development of AI techniques for AD detection from retinal images is still in progress. Various techniques based on machine learning algorithms have been developed over the past 5–10 years. For example, Tian et al. [19] trained the support vector machine (SVM) algorithm from fundus images to distinguish between healthy and AD individuals. Similarly, Wang et al. [20] proposed a technique based on the XGBoost algorithm to identify AD patients by using features extracted from OCT images. Another type of AD detection method is based on deep learning. For instance, in a collaborative study, a DL model to screen AD patients was built by utilizing fundus images [21]. The strength of this study was using a relatively large number of patients for the development of the model. In order to capture richer information about the retina, DL models based on multi-modal images were also developed. For example, Shi et al. [22] developed a deep learning model based on fundus and OCT images to distinguish healthy and diseased people. Apart from human studies, techniques using mice models of AD have also been developed. For example, ref. [23] built a deep learning model to predict transgenic AD mice from OCT scans. The key advantage of using mice models is that it is possible to track AD in an early asymptomatic stage. In humans, in contrast, it is challenging to find asymptomatic AD patients due to late diagnosis.
Contributions: The main contributions of this paper are as follows.
  • An overview of deep learning and machine learning techniques for AD detection: this paper presents a thorough analysis of the relevant methods and identifies their strengths and limitations.
  • Directions for future research: this paper discusses various approaches to the common challenges of building ML and DL models; possible directions for future research are also discussed.
The remaining part of the article contains the following sections. In Section 2 and Section 3, the search strategy and an overview of methods are presented. In Section 4, the strengths and limitations of these methods and various approaches for typical problems are discussed. Finally, Section 5 contains key directions for future research.

2. Method

2.1. Eligibility Criteria and Data Extraction

The main eligibility criteria were determined based on the objectives of the paper. Papers were selected if they focused on detecting Alzheimer’s disease from retinal images by using machine learning or deep learning techniques. Studies of both human and animal models of AD were accepted, with no age restriction on study participants. Retinal images from any type of imaging modality, including OCT and fundus imaging, were considered. The papers were excluded if the text wasn’t in English or was unavailable. Review papers and unpublished papers were also excluded. The verification of papers and data collection were performed by one reviewer initially. The second reviewer independently performed the final verification. In the beginning, titles and abstracts were screened to check their relevance based on the eligibility criteria. Finally, the whole paper was examined to confirm its eligibility. The collected papers were divided into machine learning and deep learning groups for further analysis. The following information was collected during data extraction: methodology, imaging modality, dataset, evaluation metrics, results, and year of publication. Sample size is one of the key factors in assessing the bias of studies, and Table 1 includes the sample size information for the reviewed studies.

2.2. Search Strategy

The objective of the study was to review papers screening Alzheimer’s disease from retinal images. The search terms “Alzheimer’s disease”, and “retina” were used for this purpose. Next, in this study, the methodologies were limited to machine learning- and deep learning-based methods, and the terms “artificial intelligence”, “deep learning”, “machine learning” were employed. Finally, Google Scholar and PubMed were used to search for papers using combinations of search terms using logical operators AND and OR. The search for relevant papers was conducted between 1 May 2024 and 28 May 2024 without restrictions on the date of publication. After the process of screening papers based on eligibility criteria, 16 papers were included in this study.

3. Results

3.1. Traditional Machine Learning Techniques

This section contains the overview of machine learning-based AD detection algorithms, as shown in Table 2. Various algorithms based on support vector machine (SVM) algorithms have been developed over the last 5–6 years. Nunes et al. [27] built an SVM-based algorithm to differentiate between healthy, AD, and Parkinson’s disease (PD) groups. The dataset comprised volumetric OCT scans acquired from both eyes of 27 healthy, 20 AD, and 28 PD individuals. Texture features were extracted from these images using techniques such as the gray level co-occurrence matrix. An average sensitivity of 79.5% and specificity of 92.5% were attained for AD patients upon k-fold cross-validation evaluation. However, a consideration for this study is that the model has not yet been evaluated on an external test set. Additionally, classification results between PD and AD were not reported, and only the average values were given. The similar approach was taken by Bernardes et al. [28] to classify transgenic AD mice and wild-type mice from OCT images. An average accuracy of 90% was achieved upon k-fold cross-validation evaluation. The advantage of using mice instead of humans was the possibility of tracking the retina at the initial stages of AD. In another similar study, an SVM model demonstrated 92% accuracy in distinguishing between wild-type and genetically modified AD mice [34].
Moreover, Tian et al. [19] developed an SVM-based model for the identification of AD patients from fundus images. The training and validation set was composed of 244 images acquired from 174 subjects. The blood vessels were segmented to extract the relevant information. Statistical t-tests were conducted for individual pixels to filter out pixels that showed differences between the two groups. Only filtered pixels were used as features for the SVM model. The model attained 84.8% specificity, 79.2% sensitivity, and 81.5% F-1 score based on 5-fold cross-validation. In contrast to other methods using conventional retinal images, Sharafi et al. [26] built an SVM model based on hyperspectral imaging to classify AD patients. A total of 20 AD and 26 healthy individuals were used in this study. Various features such as the diameter of vessels and textural features as a contrast were calculated and used for the SVM model. The model demonstrated 82% sensitivity and 86% specificity in the 10-fold cross-validation. A challenge of this approach is that making the model widely accessible may be difficult, as hyperspectral imaging is not commonly used.
Alongside SVM, other ML-based models have also been proposed. Recently, a light gradient boosting machine (LightGBM)-based model was developed to identify AD patients from OCTA images [24]. Some 170 images taken from both eyes of 48 healthy and 37 AD individuals were used for training and testing. The geometric characteristics of the foveal avascular zone such as area and eccentricity along with medical record data including age and gender were used for the model. The test set achieved a sensitivity of 54.4%, a specificity of 83.7%, and an AUC of 72%. While this approach is quite novel, its performance has room for improvement before it can be suitable for clinical application. In another study, OCT images were utilized to identify AD patients using the XGBoost algorithm [20]. Some 299 healthy and 159 AD subjects were recruited for this study. Retinal characteristics that have statistical differences between two groups were used as features. The test set achieved an F1 score of 70% and an AUC of 69%. Furthermore, Lemmens et al. [25] developed a linear discriminant analysis (LDA) model to identify AD patients by using hyperspectral and OCT images. The dataset consisted of 22 subjects with normal cognitive function and 17 subjects with AD. A 74% AUC was achieved in the validation set.

3.2. Deep Learning Techniques

The overview of deep learning techniques is presented in Table 3. In order to leverage the information contained in different retinal images, multi-modal deep learning techniques were proposed. For example, Wisely et al. [33] proposed a convolution neutral network (CNN)-based model to classify AD patients using OCT, optical coherence tomography angiography (OCTA), ultra-widefield (UWF) scanning laser ophthalmoscopy (SLO) color, fundus autofluorescence images, and patient medical record data. The network consisted of a five-layer CNN followed by an FC layer. The information from different images was fused by taking the average of output of the network for each image. A total of 1136 images from 159 individuals were used for training and testing the model, achieving an AUC of 83.6% in the test set. In another study, Shi et al. [22] developed an ensemble deep learning method to differentiate between cognitively impaired and healthy individuals. OCT images and fundus photographs centered on the macula and optic disc were used in this work. The information from different images was fused by concatenating feature vectors that were obtained independently by feature encoders. The joined feature vector was fed into a fully connected layer to obtain a classification result. Four different feature encoders were used, such as VGG-19 and ResNet-50, and the final result was the average of these different networks. The training set comprised 2356 subjects, while the test set contained 295 subjects, with six images acquired per person. The test set achieved 77.1% sensitivity, 70.2% specificity, and an AUC of 78.5%. Differing from previous methods that use basic techniques to merge information from different modalities, Gao et al. [30] trained a deep learning model that uses an attention module to combine information from feature extractors. A five-layer CNN was used as a feature extractor. By merging activations in different layers through this module, a richer combination of information was achieved, resulting in a reported AUC of 96.8% in the test set. This study used 5266 OCT and fundus images to classify 38 Alzheimer’s disease (AD) patients, 29 mild cognitive impairment (MCI) patients, and 50 healthy subjects for training and testing the model. Such a high number of images was reached by using 3D OCT and using augmentation techniques for fundus images.
Apart from multi-modal techniques, single-modal techniques were also introduced. Cheung et al. [21] developed a deep learning model to screen AD patients from fundus images. EfficientNet-b2 network was used as feature extractor. Multiple fundus images were taken for each person, with images of both eyes. A total of 12,949 images from 3888 people were used. The features were concatenated to obtain the final result. The model evaluated multiple test sets, and the average performance was 90.4% sensitivity, 93.5% specificity, and an 80.6% AUC. In another study, Ferreira et al. [23] built a classification model to distinguish between AD transgenic mice and wild-type mice from OCT images. A modified Inception-v3 network was used for the classification task. A total of 1144 OCT volume scans were acquired from 57 wild-type mice and 57 AD mice at different ages. OCT volumes were converted to mean-value fundus images, and therefore the dataset was composed of 1144 images. The model demonstrated 80.4% sensitivity, 86.5% specificity, and 83.3% F1 on the test set. Furthermore, in a different study, the DenseNet-121 network was used to classify 41 AD patients, 33 MCI patients, and 39 healthy subjects from fundus images, achieving a reported 97% F1 score, 99% sensitivity, and 90% specificity [29]. Similarly, a modified MobileNetV3 network was trained to differentiate 111 AD patients from 111 healthy individuals using fundus images, achieving 83.7% sensitivity, 89.1% specificity, and an AUC of 92.9% [32].

4. Discussion

The performance of classification models in terms of quantitative metrics are shown in Table 1.

4.1. Machine Learning Techniques

Extracting features from images and training machine learning models might produce better performance than directly using images to build deep learning models. For example, in [24] an ML model using FAZ area characteristics from OCTA images performed better than a DL model using OCTA images. This illustrates the importance of ML models for AD detection. Another advantage of these models is their explainability compared to building DL models. On the other hand, in many studies, models are trained and validated on the same dataset, and evaluation on an external test set is important in accurately assessing the generalization capability of these models.
One of the challenges of building these models is the identification of potentially useful features. This might require medical expertise or the selection of appropriate image-processing techniques. For instance, in [27], various techniques such as the gray level co-occurrence matrix were used to calculate relevant features so that the model showed good classification performance. Another difficulty is filtering out the important features for training the ML model. In order to build a good ML model, only relevant features should be included in the model. Various approaches, such as incorporating features that show significant statistical differences between AD and healthy groups [19], and dimensionality reduction algorithms such as principial component analysis (PCA) [35] could be employed to address this problem. Class imbalance is a common challenge in building ML models, and one possible way of addressing it is to use the synthetic minority over-sampling technique (SMOTE) [36].

4.2. Deep Learning Techniques

Overall, DL methods perform better than ML methods in terms of quantitative metrics, as shown in Table 1. Deep learning models are more complex than machine learning models and can better approximate the theoretically best solution. However, they are more prone to overfitting and expensive to train. One of the potential ways of enhancing the capability of these models is to include the patient’s medical data such as age and gender. In the context of AD detection techniques, inclusion of patient data has shown contradictory results. While some studies have shown improvement on model performance, others have reported no difference [21,33]. Further research is needed on how to merge patient data into the model effectively. On the other hand, using images from multiple modalities has demonstrated improved performance with all methods. As an example, in [20], a DL model using OCT and fundus images performed better than a model using OCT or fundus images alone.
The potential problem in training a deep learning model is overfitting due to limited sample size. Overfitting means the dataset size is too small for the given model capacity, and the model cannot generalize well to unseen data. There are various ways to address this, including decreasing the complexity of a model by using parameter norm penalties. Also, it is possible to increase the dataset size by collecting more data or using techniques such as data augmentation and semi-supervised learning, among others [37]. Data augmentation is used in many models to tackle limited dataset size, but this approach might lead to the generation of unrealistic features that will not be useful for predicting AD. Moreover, in many studies, pre-trained models based on natural images have been used, and they might have limited capacity for AD detection.
One of the limitations of DL techniques is the approach taken to merge the information from different modalities. It is questionable at which level the information should be merged. The data could be merged at the hidden layer, the classifier layer, or the prediction layer. Many studies use simple techniques such as averaging or concatenating at the classifier layer. Other approaches might give better results, such as merging the activations of different images at each CNN layers [30].
Another issue is the interpretability of DL models, which hinders the identification of biomarkers and their deployment in clinical settings. Deep learning models contain millions of parameters, and the initial input goes through many non-linear transformations. For example, the popular Resnet-50 network contains around 26 million parameters [38]. Unlike machine learning models which input hand-crafted explainable features, deep learning models complete the task in an end-to-end manner, from raw data to output prediction, making interpretability challenging. One approach to this problem is to use class activation mapping (CAM) to generate heatmaps highlighting areas of an image contributing to the decision [30]. Another technique is to investigate the correlation between explainable biomarkers and quantitative metrics derived from feature maps at different layers of the CNN. For instance, in a recent study, the neuron activation pattern (NAP) score was found to be correlated with cardiovascular risk score in a CNN model which predicts blood pressure from fundus images [39]. Apart from interpretability, there are other challenges with clinical implementation, such as the initial investment needed, regulatory compliance, etc.

5. Future Directions

Publicly available datasets are necessary to compare the performance of different models effectively. Presently, models are tested on custom datasets, which makes comparison challenging. While most studies report satisfactory classification performance on quantitative metrics, there are still some limitations. Many studies test their model’s capability by comparing healthy subjects to those with AD. However, in clinical settings, the population is more diverse and includes individuals with various diseases. Therefore, models should aim to discriminate AD from similar conditions with high specificity. Furthermore, ML and DL models may not perform consistently on images obtained from various imaging machines. Currently, most models are typically tested on images from a single machine. Therefore, it is important to evaluate these models on images from different machines as well. Various approaches such as the domain adaption technique can be explored to increase the adaptability of a model for different images [21].
Another interesting direction for future research is the application of foundational models for AD detection. Foundational models are a type of AI model that are trained on massive datasets, mainly in a self-supervised manner, and can be tailored to various specific tasks [40]. These models can address the need for extensive labelled datasets and have strong generalization capability. However, in the medical domain, they have challenges such as varying types of medical images and the scarcity of large available datasets [41,42]. In a recent study, a self-supervised foundational model was trained based on 1.6 million OCT and color fundus images. The fine-tuned model showed good performance in the classification of eye disease but had lower performance in predicting non-ocular diseases such as heart failure, showing the need for further research [43].

6. Conclusions

This paper presented an overview of ML and DL techniques for AD detection from retinal images. The development of such techniques is important in terms of identifying new AD biomarkers and screening AD patients. These methods have shown promising performance on various quantitative metrics. However, there are still challenges, such as a lack of publicly available datasets, limited dataset size, and the interpretability of DL models. These problems can be addressed in future research.

Author Contributions

A.U.: Conceptualization, Methodology, Investigation, Writing—original draft, Writing—reviewing and editing. L.L.H.C.: Supervision, Writing—reviewing and editing. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the City University of Hong Kong (7020058).

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Association, A. Alzheimer’s disease facts and figures. Alzheimer’s Dement. 2023, 19, 1598–1695. [Google Scholar]
  2. Waring, S.C.; Doody, R.S.; Pavlik, V.N.; Massman, P.J.; Chan, W. Survival among patients with dementia from a large multi-ethnic population. Alzheimer Dis. Assoc. Disord. 2005, 19, 178–183. [Google Scholar] [CrossRef]
  3. Gustavsson, A.; Norton, N.; Fast, T.; Frölich, L.; Georges, J.; Holzapfel, D.; Kirabali, T.; Krolak-Salmon, P.; Rossini, P.M.; Ferretti, M.T.; et al. Global estimates on the number of persons across the Alzheimer’s disease continuum. Alzheimer’s Dement. 2023, 19, 658–670. [Google Scholar] [CrossRef] [PubMed]
  4. Bateman, R.J.; Xiong, C.; Benzinger, T.L.; Fagan, A.M.; Goate, A.; Fox, N.C.; Marcus, D.S.; Cairns, N.J.; Xie, X.; Blazey, T.M.; et al. Clinical and biomarker changes in dominantly inherited Alzheimer’s disease. N. Engl. J. Med. 2012, 367, 795–804. [Google Scholar] [CrossRef] [PubMed]
  5. Yu, J.T.; Xu, W.; Tan, C.C.; Andrieu, S.; Suckling, J.; Evangelou, E.; Pan, A.; Zhang, C.; Jia, J.; Feng, L.; et al. Evidence-based prevention of Alzheimer’s disease: Systematic review and meta-analysis of 243 observational prospective studies and 153 randomised controlled trials. J. Neurol. Neurosurg. Psychiatry 2020, 91, 1201–1209. [Google Scholar] [CrossRef]
  6. Barnes, D.E.; Yaffe, K. The projected effect of risk factor reduction on Alzheimer’s disease prevalence. Lancet Neurol. 2011, 10, 819–828. [Google Scholar] [CrossRef]
  7. London, A.; Benhar, I.; Schwartz, M. The retina as a window to the brain—from eye research to CNS disorders. Nat. Rev. Neurol. 2013, 9, 44–53. [Google Scholar] [CrossRef] [PubMed]
  8. Nguyen, C.T.; Acosta, M.L.; Di Angelantonio, S.; Salt, T.E. Seeing beyond the eye: The brain connection. Fronti. Neurosci. 2021, 15, 719717. [Google Scholar] [CrossRef]
  9. Czakó, C.; Kovács, T.; Ungvari, Z.; Csiszar, A.; Yabluchanskiy, A.; Conley, S.; Csipo, T.; Lipecz, A.; Horváth, H.; Sándor, G.L.; et al. Retinal biomarkers for Alzheimer’s disease and vascular cognitive impairment and dementia (VCID): Implication for early diagnosis and prognosis. Geroscience 2020, 42, 1499–1525. [Google Scholar] [CrossRef]
  10. Ptito, M.; Bleau, M.; Bouskila, J. The retina: A window into the brain. Cells 2021, 10, 3269. [Google Scholar] [CrossRef]
  11. Valenti, D.A. Alzheimer’s disease: Visual system review. Optom.-J. Am. Optom. Assoc. 2010, 81, 12–21. [Google Scholar] [CrossRef] [PubMed]
  12. Ascaso, F.J.; Cruz, N.; Modrego, P.J.; Lopez-Anton, R.; Santabárbara, J.; Pascual, L.F.; Lobo, A.; Cristóbal, J.A. Retinal alterations in mild cognitive impairment and Alzheimer’s disease: An optical coherence tomography study. J. Neurol. 2014, 261, 1522–1530. [Google Scholar] [CrossRef] [PubMed]
  13. Koronyo-Hamaoui, M.; Koronyo, Y.; Ljubimov, A.V.; Miller, C.A.; Ko, M.K.; Black, K.L.; Schwartz, M.; Farkas, D.L. Identification of amyloid plaques in retinas from Alzheimer’s patients and noninvasive in vivo optical imaging of retinal plaques in a mouse model. Neuroimage 2011, 54, S204–S217. [Google Scholar] [CrossRef] [PubMed]
  14. Fujimoto, J.; Drexler, W. Introduction to optical coherence tomography. In Optical Coherence Tomography: Technology and Applications; Springer: Berlin/Heidelberg, Germany, 2008; pp. 1–45. [Google Scholar]
  15. Ting, D.S.W.; Cheung, C.Y.L.; Lim, G.; Tan, G.S.W.; Quang, N.D.; Gan, A.; Hamzah, H.; Garcia-Franco, R.; San Yeo, I.Y.; Lee, S.Y.; et al. Development and validation of a deep learning system for diabetic retinopathy and related eye diseases using retinal images from multiethnic populations with diabetes. JAMA 2017, 318, 2211–2223. [Google Scholar] [CrossRef]
  16. Liu, H.; Li, L.; Wormstone, I.M.; Qiao, C.; Zhang, C.; Liu, P.; Li, S.; Wang, H.; Mou, D.; Pang, R.; et al. Development and validation of a deep learning system to detect glaucomatous optic neuropathy using fundus photographs. JAMA Ophthalmol. 2019, 137, 1353–1360. [Google Scholar] [CrossRef]
  17. Ran, A.R.; Tham, C.C.; Chan, P.P.; Cheng, C.Y.; Tham, Y.C.; Rim, T.H.; Cheung, C.Y. Deep learning in glaucoma with optical coherence tomography: A review. Eye 2021, 35, 188–201. [Google Scholar] [CrossRef]
  18. Joo, Y.S.; Rim, T.H.; Koh, H.B.; Yi, J.; Kim, H.; Lee, G.; Kim, Y.A.; Kang, S.W.; Kim, S.S.; Park, J.T. Non-invasive chronic kidney disease risk stratification tool derived from retina-based deep learning and clinical factors. NPJ Digit. Med. 2023, 6, 114. [Google Scholar] [CrossRef]
  19. Tian, J.; Smith, G.; Guo, H.; Liu, B.; Pan, Z.; Wang, Z.; Xiong, S.; Fang, R. Modular machine learning for Alzheimer’s disease classification from retinal vasculature. Sci. Rep. 2021, 11, 238. [Google Scholar] [CrossRef]
  20. Wang, X.; Jiao, B.; Liu, H.; Wang, Y.; Hao, X.; Zhu, Y.; Xu, B.; Xu, H.; Zhang, S.; Jia, X.; et al. Machine learning based on Optical Coherence Tomography images as a diagnostic tool for Alzheimer’s disease. CNS Neurosci. Ther. 2022, 28, 2206–2217. [Google Scholar] [CrossRef]
  21. Cheung, C.Y.; Ran, A.R.; Wang, S.; Chan, V.T.; Sham, K.; Hilal, S.; Venketasubramanian, N.; Cheng, C.Y.; Sabanayagam, C.; Tham, Y.C.; et al. A deep learning model for detection of Alzheimer’s disease based on retinal photographs: A retrospective, multicentre case-control study. Lancet Digit. Health 2022, 4, e806–e815. [Google Scholar] [CrossRef]
  22. Shi, X.H.; Ju, L.; Dong, L.; Zhang, R.H.; Shao, L.; Yan, Y.N.; Wang, Y.X.; Fu, X.F.; Chen, Y.Z.; Ge, Z.Y.; et al. Deep Learning Models for the Screening of Cognitive Impairment Using Multimodal Fundus Images. Ophthalmol. Retin. 2024, 8, 666–677. [Google Scholar] [CrossRef]
  23. Ferreira, H.; Serranho, P.; Guimarães, P.; Trindade, R.; Martins, J.; Moreira, P.I.; Ambrósio, A.F.; Castelo-Branco, M.; Bernardes, R. Stage-independent biomarkers for Alzheimer’s disease from the living retina: An animal study. Sci. Rep. 2022, 12, 13667. [Google Scholar] [CrossRef] [PubMed]
  24. 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. Enhancing foveal avascular zone analysis for Alzheimer’s diagnosis with AI segmentation and machine learning using multiple radiomic features. Sci. Rep. 2024, 14, 1841. [Google Scholar] [CrossRef] [PubMed]
  25. Lemmens, S.; Van Craenendonck, T.; Van Eijgen, J.; De Groef, L.; Bruffaerts, R.; de Jesus, D.A.; Charle, W.; Jayapala, M.; Sunaric-Mégevand, G.; Standaert, A.; et al. Combination of snapshot hyperspectral retinal imaging and optical coherence tomography to identify Alzheimer’s disease patients. Alzheimer’s Res. Ther. 2020, 12, 144. [Google Scholar] [CrossRef] [PubMed]
  26. Sharafi, S.M.; Sylvestre, J.P.; Chevrefils, C.; Soucy, J.P.; Beaulieu, S.; Pascoal, T.A.; Arbour, J.D.; Rhéaume, M.A.; Robillard, A.; Chayer, C.; et al. Vascular retinal biomarkers improves the detection of the likely cerebral amyloid status from hyperspectral retinal images. Alzheimer’s Dementia Transl. Res. Clin. Interv. 2019, 5, 610–617. [Google Scholar] [CrossRef]
  27. Nunes, A.; Silva, G.; Duque, C.; Januario, C.; Santana, I.; Ambrosio, A.F.; Castelo-Branco, M.; Bernardes, R. Retinal texture biomarkers may help to discriminate between Alzheimer’s, Parkinson’s, and healthy controls. PLoS ONE 2019, 14, e0218826. [Google Scholar] [CrossRef]
  28. Bernardes, R.; Silva, G.; Chiquita, S.; Serranho, P.; Ambrósio, A.F. Retinal biomarkers of Alzheimer’s disease: Insights from transgenic mouse models. In Proceedings of the Image Analysis and Recognition: 14th International Conference, ICIAR 2017, Proceedings 14, Montreal, QC, Canada, 5–7 July 2017; Springer: Berlin/Heidelberg, Germany, 2017; pp. 541–550. [Google Scholar]
  29. Luengnaruemitchai, G.; Kaewmahanin, W.; Munthuli, A.; Phienphanich, P.; Puangarom, S.; Sangchocanonta, S.; Jariyakosol, S.; Hirunwiwatkul, P.; Tantibundhit, C. Alzheimer’s Together with Mild Cognitive Impairment Screening Using Polar Transformation of Middle Zone of Fundus Images Based Deep Learning. In Proceedings of the 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), Sydney, Australia, 24–27 July 2023; pp. 1–4. [Google Scholar]
  30. Gao, H.; Zhao, S.; Zheng, G.; Wang, X.; Zhao, R.; Pan, Z.; Li, H.; Lu, F.; Shen, M. Using a dual-stream attention neural network to characterize mild cognitive impairment based on retinal images. Comput. Biol. Med. 2023, 166, 107411. [Google Scholar] [CrossRef]
  31. Hao, J.; Kwapong, W.R.; Shen, T.; Fu, H.; Xu, Y.; Lu, Q.; Liu, S.; Zhang, J.; Liu, Y.; Zhao, Y.; et al. Early detection of dementia through retinal imaging and trustworthy AI. NPJ Digit. Med. 2024, 7, 294. [Google Scholar] [CrossRef]
  32. Lim, Y.J.; Park, J.H.; Sunwoo, M.H. Efficient Deep Learning Algorithm for Alzheimer’s Disease Diagnosis using Retinal Images. In Proceedings of the 2022 IEEE 4th International Conference on Artificial Intelligence Circuits and Systems (AICAS), Incheon, Republic of Korea, 13–15 June 2022; pp. 254–257. [Google Scholar]
  33. Wisely, C.E.; Wang, D.; Henao, R.; Grewal, D.S.; Thompson, A.C.; Robbins, C.B.; Yoon, S.P.; Soundararajan, S.; Polascik, B.W.; Burke, J.R.; et al. Convolutional neural network to identify symptomatic Alzheimer’s disease using multimodal retinal imaging. Br. J. Ophthalmol. 2022, 106, 388–395. [Google Scholar] [CrossRef]
  34. Sayeed, F.; Rafeeq Ahmed, K.; Vinmathi, M.; Priyadarsini, A.I.; Gundupalli, C.B.; Tripathi, V.; Shishah, W.; Sundramurthy, V.P. Classification of transgenic mice by retinal imaging using SVMS. Comput. Intell. Neurosci. 2022, 2022, 9063880. [Google Scholar] [CrossRef]
  35. Greenacre, M.; Groenen, P.J.; Hastie, T.; d’Enza, A.I.; Markos, A.; Tuzhilina, E. Principal component analysis. Nat. Rev. Methods Prim. 2022, 2, 100. [Google Scholar] [CrossRef]
  36. Pradipta, G.A.; Wardoyo, R.; Musdholifah, A.; Sanjaya, I.N.H.; Ismail, M. SMOTE for handling imbalanced data problem: A review. In Proceedings of the 2021 Sixth International Conference on Informatics and Computing (ICIC), Virtual, 3–4 November 2021; pp. 1–8. [Google Scholar]
  37. Goodfellow, I.; Bengio, Y.; Courville, A.; Bengio, Y. Deep Learning; MIT Press: Cambridge, UK, 2016; Volume 1. [Google Scholar]
  38. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2016; pp. 770–778. [Google Scholar]
  39. An, S.; Squirrell, D. Validation of neuron activation patterns for artificial intelligence models in oculomics. Sci. Rep. 2024, 14, 20940. [Google Scholar] [CrossRef] [PubMed]
  40. Bommasani, R.; Hudson, D.A.; Adeli, E.; Altman, R.; Arora, S.; von Arx, S.; Bernstein, M.S.; Bohg, J.; Bosselut, A.; Brunskill, E.; et al. On the opportunities and risks of foundation models. arXiv 2021, arXiv:2108.07258. [Google Scholar]
  41. Zhang, S.; Metaxas, D. On the challenges and perspectives of foundation models for medical image analysis. Med. Image Anal. 2024, 91, 102996. [Google Scholar] [CrossRef]
  42. Moor, M.; Banerjee, O.; Abad, Z.S.H.; Krumholz, H.M.; Leskovec, J.; Topol, E.J.; Rajpurkar, P. Foundation models for generalist medical artificial intelligence. Nature 2023, 616, 259–265. [Google Scholar] [CrossRef]
  43. Zhou, Y.; Chia, M.A.; Wagner, S.K.; Ayhan, M.S.; Williamson, D.J.; Struyven, R.R.; Liu, T.; Xu, M.; Lozano, M.G.; Woodward-Court, P.; et al. A foundation model for generalizable disease detection from retinal images. Nature 2023, 622, 156–163. [Google Scholar] [CrossRef]
Table 1. Performance of machine learning- and deep learning-based AD detection methods.
Table 1. Performance of machine learning- and deep learning-based AD detection methods.
RefMethodDataset Size
(# Images)		Sensitivity
(%)		Specificity
(%)		AUC
(%)		F-1
(%)
[24]Machine learning17054.483.772.0-
[20]Machine learning458--69.070.0
[19]Machine learning24479.284.8-81.5
[25]Machine learning78--74.0-
[26]Machine learning13882.086.0--
[27]Machine learning15079.592.5--
[28]Machine learning77----
[22]Deep learning1590677.170.278.5-
[29]Deep learning22599.090.0-97.0
[30]Deep learning5266-94.696.890.4
[31]Deep learning575181.6-90.082.9
[21]Deep learning1294990.493.580.6-
[23]Deep learning114480.486.5-83.3
[32]Deep learning44583.789.192.9-
[33]Deep learning1136--83.6-
Table 2. Summary of traditional machine learning-based methods for AD detection.
Table 2. Summary of traditional machine learning-based methods for AD detection.
RefYearModalityAlgorithmFeaturesType of StudyAge Range
[24]2024OCTALightGBMGeometric characteristics of FAZ zone, patient dataHuman65.70 (7.90)
[20]2022OCTXGBoostRetinal layer thicknesses, macular volumeHuman63.03 (9.06)
[34]2022OCTSupport vector machineTexture featuresMice—
[19]2021Fundus photographySupport vector machinePixel intensitiesHuman65.17 (4.16)
[25]2020Hyperspectral imaging, OCTLinear discriminant analysisReflectance values, RNFL thicknessHuman55–85
[26]2019Hyperspectral imagingSupport vector machineVasculature characteristics, texture featuresHuman60–85
[27]2019OCTSupport vector machineTexture featuresHuman53–77
[28]2017OCTSupport vector machineTexture featuresMice4–8 months
Table 3. Summary of deep learning-based methods for AD detection.
Table 3. Summary of deep learning-based methods for AD detection.
RefYearModalityNetwork TypeBiomarkersType of StudyAge Range
[31]2024OCTACNN and GNNFAZ area and neighboring vesselsHuman66.42 (8.58)
[22]2024Fundus photography, OCTVGG-19, ResNet-50 … (ensemble model)Optic nerve, macular regionsHuman63.88 (9.63)
[29]2023Fundus photographyDenseNet-121Superior, inferior quadrantsHuman—
[30]2023OCT, Fundus photography5-layer CNN with an attention moduleVascular bifurcations, retinal layersHuman—
[21]2022Fundus photographyEfficientNet-b2 with fusion module—HumanMultiple studies
[23]2022OCTModified Inception-v3—Mice1–12 months
[32]2022Fundus photographyModified MobileNetV3—Human—
[33]2022OCT, OCTA, UWF SLO, FAF5-layer CNN—Human71.08 (8.83)
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Uvaliyev, A.; Chan, L.L.H. Alzheimer’s Disease Detection from Retinal Images Using Machine Learning and Deep Learning Techniques: A Perspective. Appl. Sci. 2025, 15, 4963. https://doi.org/10.3390/app15094963

AMA Style

Uvaliyev A, Chan LLH. Alzheimer’s Disease Detection from Retinal Images Using Machine Learning and Deep Learning Techniques: A Perspective. Applied Sciences. 2025; 15(9):4963. https://doi.org/10.3390/app15094963

Chicago/Turabian Style

Uvaliyev, Adilet, and Leanne Lai Hang Chan. 2025. "Alzheimer’s Disease Detection from Retinal Images Using Machine Learning and Deep Learning Techniques: A Perspective" Applied Sciences 15, no. 9: 4963. https://doi.org/10.3390/app15094963

APA Style

Uvaliyev, A., & Chan, L. L. H. (2025). Alzheimer’s Disease Detection from Retinal Images Using Machine Learning and Deep Learning Techniques: A Perspective. Applied Sciences, 15(9), 4963. https://doi.org/10.3390/app15094963

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop