Detection and exploratory severity grading of age-related macular degeneration using machine learning

Age-related Macular Degeneration (AMD) is a leading cause of irreversible vision loss among the elderly population worldwide, making early detection and reliable severity assessment essential for timely intervention and prevention of severe visual impairment. This study presents a novel machine-learning–based framework for automated AMD detection and exploratory severity grading using Optical Coherence Tomography (OCT) images. The proposed approach employs deep learning architectures, including ResNet and VGG, trained on publicly available datasets. An unsupervised learning strategy is introduced for dry AMD exploratory severity grading through feature extraction based on drusen size. Advanced AMD manifestations, such as geographic atrophy and choroidal neovascularization, are identified using clinically relevant retinal features derived from OCT images. The proposed detection framework achieves high accuracy (98%) on the primary dataset and demonstrates robust generalization with an accuracy of 83% for cross-dataset evaluation using transfer learning. Furthermore, the unsupervised severity grading model attains a silhouette score close to 0.5, reflecting meaningful cluster structure consistent with the continuous nature of drusen progression. Overall, the proposed framework offers a scalable and automated solution for AMD detection and exploratory severity grading, with the potential to support clinical decision–making and enhance personalized disease management.

According to the World Health Organization (WHO), AMD is one of the leading causes of vision loss in developed countries (World Health Organization, 2025), primarily affecting aging individuals and severely impacting quality of life. The global aging population of approximately 962 million people is expected to reach 2 billion by the year 2050 (Tan et al., 2018). Age-Related Macular Degeneration (AMD) is one of the leading causes of vision loss, especially among elderly populations (Lim et al., 2012). The prevalence of AMD is rising, placing increased demand on healthcare systems for timely diagnosis and monitoring of the disease's progression. Unfortunately, AMD is not a curable disease, making early detection and accurate grading even more critical. Without proper timely intervention, AMD can progress from the early dry stages, which are characterized by drusen, to advanced wet AMD, which poses a significant threat to vision.

Advancements in medical imaging and machine learning (ML), particularly deep learning (Miotto et al., 2018), present a promising solution for scalable and precise AMD assessment. Computer Vision (CV) have the ability to assess medical images such as X-rays, CT scans and reveal issues that humans might miss (Akkus et al., 2021). By applying computer vision and deep learning techniques on detailed Optical Coherence Tomography (OCT) images, automated models could significantly enhance the efficiency and accuracy of diagnosis. The process of manual interpretation of OCT images for detecting and grading AMD remains time-consuming and relies heavily on the expertise of ophthalmologists. The development of ML models for AMD classification is challenging due to several factors such as annotation scarcity, where the creation of pixel-level labels and disease severity requires retinal specialists that is time-consuming, expensive and can be subject to inter-observer variability. OCT scans obtained from different devices from different clinics or imaging protocols can have different image quality or resolution that lead to domain shift and models trained on different datasets can perform poorly on other datasets. Another major challenge is the lack of labeled OCT datasets for severity grading. Most publicly available datasets focus on detection and binary classification of AMD without providing severity grading labels.

Therefore, a successful model for automated AMD detection and severity grading, without pre-determined severity labels, would support early intervention and help preserve vision for at-risk individuals and reduce healthcare costs. The aim of this research was to design and develop a novel two-stage deep learning-based model to automate AMD detection and, motivated by the absence of severity grading labeling, perform unsupervised exploratory severity grading of dry-AMD using OCT images. The main contribution of this paper is to address detection of AMD using ResNet and VGG in combination with exploratory severity grading of AMD using an unsupervised approach, to overcome the challenges of unlabelled datasets for severity grading, contributing to improved AMD classification.

Background and related work

Overview of eye diseases and age-related macular degeneration

Vision-related diseases and conditions are a major concern worldwide. Globally, at least 1 billion people have some form of vision impairment that could have been prevented or is yet to be addressed (World Health Organization, 2025). These diseases not only affect the quality of life of individuals but can also lead to substantial economic burdens. Some of the most common causes of blindness globally include Cataract, Glaucoma, uncorrected refractive errors, and Age-related Macular Degeneration (Johnson et al., 2012). While some vision impairments are correctable with medical procedures or surgery, several other eye diseases, such as glaucoma and age-related macular degeneration (AMD), remain incurable. Glaucoma affects a person's peripheral vision, whereas AMD primarily impacts their central vision. AMD is especially concerning as it is a leading cause of vision loss in developed countries, primarily affecting older adults (Tan et al., 2018). Almost 200 million people worldwide have some form of AMD (Wong et al., 2014). The progressive nature of AMD, from early stages with minimal symptoms to advanced stages where vision loss becomes profound, underscores the need for early detection and precise classification to manage the disease effectively.

AMD can be classified as early-stage (medium-sized drusen and retinal pigmentary changes) to late-stage (neovascular and geographic atrophy) (Mitchell et al., 2018). AMD can be separated as dry and wet AMD, with dry AMD being the most common type of AMD. Dry AMD, which accounts for 80%–85% of cases, is characterized by drusen, which are yellow deposits beneath the retina, and retinal atrophy. Wet AMD (exudative/neovascular) is the type with developed neovascularization (abnormal blood vessel growth under the retina) leading to hemorrhaging and leakage of fluid in the inner retinal layers. It affects around 15%–20% of cases. Overall around 10%–15% of dry AMD progresses to wet AMD (Urbanska et al., 2024). Although age is the main factor for AMD development, other factors such as increased body mass index, cigarette smoking, hyperlipidaemia, hypertension, and genetics are also important risk factors (Clemons et al., 2005; Velilla et al., 2013).

Optical coherence tomography (OCT) is a non-invasive ophthalmology imaging technique. OCT-based imaging of the eye is widely accessible to a large population due to its low cost and zero side effects. The use of inexpensive and non-radioactive optical contrast agents can offer safe, highly sensitive, and targeted imaging of regions of interest (Thomas et al., 2021). Portability, low instrumentation cost, and operational cost are other advantages of this mechanism (Thomas et al., 2021). There are two types of OCT images, time-domain OCT images and spectral-domain OCT (SD-OCT) images, the latter being more extensively used because of its superior quality over the former. Using retinal OCT images for diagnosis purposes is advantageous over fundus images. While fundus photography captures a two-dimensional view of the retina, OCT provides cross-sectional details, allowing for the identification of subtle structural changes as shown in Figure 1. OCT provides a clear view of the nine retinal layers, whose varying thickness and irregularities encode a lot of information that is helpful in the diagnosis of several diseases (Oakley, 2014), including choroidal neovascularization (CNV), diabetic macular edema (DME), and drusen.

Besides ophthalmic diagnosis, recent studies have shown that OCT images contain biomarkers for many other diseases such as Parkinson's disease (Nunes et al., 2019) and Alzheimer's disease (Cabrera DeBuc et al., 2019). AMD can be classified into two major types, dry AMD and wet AMD.

Dry AMD

Dry AMD affects the central portion of the retina, called the macula, which is essential for detailed vision. It is characterized by the accumulation of drusen, small yellow deposits beneath the retina, and gradual thinning or atrophy of the retinal pigment epithelium (RPE). This form of AMD progresses slowly. Dry AMD is generally categorized into early, intermediate, and advanced stages, based primarily on drusen size and retinal pigment changes. This three-stage framework is formally defined by the Beckman Clinical Classification, in which early AMD is characterized by medium drusen (63 to < 125 μm) without pigmentary abnormality, intermediate AMD by large drusen (125 μm) or pigmentary abnormalities with at least medium drusen, and late AMD by neovascular disease or geographic atrophy (Ferris et al., 2013). Figure 2 shows a cross-section of dry AMD where corrugations can be seen in the RPE membrane layer, which has very high reflectance, shown as reddish pixels. As it advances, it can lead to geographic atrophy, a severe stage marked by the loss of retinal cells, which leads to central vision impairment.

Wet AMD

Wet AMD is a severe form of macular degeneration that involves the growth of abnormal blood vessels beneath the retina, a process known as choroidal neovascularization (CNV). These fragile, leaky blood vessels disrupt the integrity of the Bruch's membrane, a key structure in the retina, leading to fluid and blood leakage into the macula, which is responsible for central vision (Kulkarni and Kuppermann, 2005). Figures 3, 4 show a color SD-OCT image of a wet AMD and normal eye respectively. This leakage causes swelling, scarring, and ultimately damage to the macula, resulting in rapid and severe vision loss, particularly in the center of the visual field.

Unlike dry AMD, which progresses slowly and involves the build-up of drusen, wet AMD can lead to rapid deterioration of vision. The presence of CNV, which leads to retinal hemorrhage, fluid accumulation, and the formation of scar tissue, distinguishes wet AMD from other forms of retinal degeneration. The process of manual interpretation of OCT images for detecting and grading AMD remains time-consuming and relies heavily on the expertise of ophthalmologists.

Applications of machine learning in medical imaging

The healthcare industry has greatly benefited from the adoption of AI and machine learning models. Currently, ML models support a physician or analyst in fulfilling their roles, identifying healthcare trends, and developing disease prediction models (Habehh and Gohel, 2021). In large medical organizations, ML based approaches can be adapted for tasks such as identification of irregularities in blood samples (Siddiqui et al., 2020), organization of electronic health records (Rao et al., 2011), robot assisted surgeries (Kaouk et al., 2019), among others. During the COVID-19 pandemic, hospitals were able to organize, share, and track patients, beds, rooms, ventilators, EHRs (Electronic Health Records), and even staff using a deep learning system by GE (General Electric) called the Clinical Command Center (Roth, 2020). ML based approaches have been used for detection of eye conditions such as glaucoma (Krishnakumar and Soobhany, 2025), diabetic retinopathy (Emara et al., 2026), among others.

Convolutional Neural Networks (CNN)s, which are part of supervised machine learning, are multilayer networks that rely on feature recognition and identification and are useful for anomaly detection, image recognition, and identification (Leijnen and Veen, 2020). CNN is a mathematical construct that is typically composed of three types of layers: convolution, pooling, and fully connected layers. The first two, convolution and pooling layers, perform feature extraction, whereas the third, a fully connected layer, maps the extracted features into final output, such as classification (Yamashita et al., 2018). CNNs have been successfully applied in many healthcare fields such as Pneumonia Detection on Chest X-Rays (Yang, 2019), skin cancer classification (Esteva et al., 2017), Diabetic retinopathy detection (Jansson et al., 2018), among others. CNNs come in different forms, and two prominent architectures, ResNet and VGG, have gained widespread adoption due to their effectiveness in dealing with the complexity of medical images (Xu et al., 2023; Majib et al., 2021).

These architectures, while fundamentally CNNs, are engineered to address challenging tasks, such as detecting subtle anomalies in high-dimensional scans. ResNet, or Residual Network, introduces residual connections that allow the network to bypass certain layers, enabling training of very deep architectures, often exceeding dozens or hundreds of layers, without degradation in performance (He et al., 2022b). This capability enhances its sensitivity to intricate patterns, such as those observed in retinal scans. VGG, developed by (Visual Geometry Group, n.d.), uses a sequence of small 3 × 3 convolutional filters applied in deep stacks, increasing the network capacity to extract fine-grained features (Simonyan and Zisserman, 2014). This design is particularly advantageous for analyzing high-resolution images. Together, ResNet and VGG have demonstrated superior performance in healthcare applications. These attributes make them reliable foundations for developing diagnostic tools, especially in fields of ophthalmology where detailed imaging is prominent. Transformer-based models have been applied to capture global retinal context, with architectures like Vision Transformers (Dosovitskiy et al., 2021) that can improve performance in classification and detection in fundus and OCT images. Self-supervised learning approaches such as SimCLR (Chen et al., 2020) and MAE (He et al., 2022a) learn robust representations from unlabelled retinal data using contrastive or reconstruction-based objectives.

Related work

Recent research has focused on binary classification of AMD (normal vs. AMD) using OCT images, using architectures like Inception and VGG for their robustness in feature extraction. For instance, (He et al., 2022b) combined ResNet50 with Local Outlier Factor (LOF) on UCSD and Duke datasets, achieving up to 99.87% accuracy, though with high computational cost. Similarly, (Kadry et al., 2022) applied VGG16 on a private dataset, reaching 97.5% accuracy, limited by data quality needs. (Thomas et al., 2021) used a 7-layer CNN across multiple datasets (Mendeley, OCTID), hitting 99.73% accuracy but struggled with grading differentiation. Table 1 summarizes these and other efforts, highlighting performance and limitations (El-Den et al., 2024).

Beyond detection, studies have tackled multiclass AMD grading (e.g., normal, drusen, CNV), often using CNNs and vision transformers. Azizi et al. (2024) employed MedViT on NEH and UCSD datasets for 3-class grading, achieving 97.7% accuracy, though requiring high-quality images. El-Den et al. (2024) developed an xAI system with ResNet50 and DeepLabV3+ for 7-class grading (normal to inactive wet), reaching 90.82% accuracy but struggling with early vs. intermediate AMD. (Akça et al., 2024) used Mobile-ViT on the Mendeley dataset, matching this study's data, achieving 99.17% for 4-class grading.

Several additional studies further illustrate the range of methodologies and limitations in multiclass AMD grading. Gueddena et al. (2024) applied a CNN to both a private dataset and the Duke dataset for 3-class classification (Normal, AMD, DME), achieving accuracies of 99.6 and 99.53% respectively, though the approach was prone to bias, overfitting, and class imbalance. Diao et al. (2023) used a CNN on the UCSD dataset for 3-class grading (Normal, CNV, Drusen), reaching 96.93% accuracy, but similarly required high image quality and showed bias toward the source dataset. Collectively, these grading-focused studies span 3-class to 7-class formulations and report accuracies ranging from approximately 83% to over 99%, with performance generally declining as the number of severity classes increases and as datasets shift from private to public sources. A recurring limitation across papers is the dependence on high-quality input images and a persistent susceptibility to dataset-specific bias, both of which constrain real-world generalisability and motivate the transfer learning strategy adopted in our study.

These studies represent a subset of the extensive body of research demonstrating the effective detection of AMD in OCT images using both public and private datasets. It is important to emphasize that such models are not intended to replace ophthalmologists; rather, they are designed to support clinical decision-making. By assisting in the rapid assessment of disease presence, grade, and severity, these tools can substantially reduce diagnostic time and optimize the use of clinical resources.

Datasets and challenges in machine learning for eye disease analysis

In Ophthalmology, machine learning models use image datasets of fundus photography and Optical Coherence Tomography (OCT) to address eye diseases such as Age-related Macular Degeneration (AMD). Fundus images, capturing the retina's surface, and OCT scans, detailing its cross-sectional layers, form the backbone of datasets like AREDS (NICE, 2018) and Mendeley OCT (Kermany et al., 2018), respectively. Public datasets such as Mendeley OCT, AREDS, and Duke (Farsiu et al., 2014) are widely utilized for detection and grading tasks. Mendeley OCT provides spectral-domain OCT (SD-OCT) images with labels (normal, drusen, DME, and CNV) across thousands of samples, while AREDS offers fundus images with detailed severity annotations based on clinical standards. Private datasets dominate much of the literature (Kadry et al., 2022; El-Den et al., 2024), offering larger, tailored samples but limiting reproducibility due to restricted access.

Data quality poses significant challenges. SD-OCT images, while superior to time-domain OCT, vary in noise levels, resolution, and contrast due to device differences or patient factors like motion (Thomas et al., 2021). Subtle AMD features such as drusen size variations or choroidal neovascularization (CNV) demand high-quality inputs, as noted in Akça et al. (2024), where Mobile-ViT achieved 99.17% accuracy on the Mendeley dataset. Differences in acquisition methods across datasets introduce variability, a challenge reflected in studies relying on private datasets (Wang et al., 2024).

Annotation scarcity is another critical barrier. While binary labels (normal vs. AMD) are common, severity grading labels are rare in public datasets. Mendeley OCT, for instance, lacks annotations for early, intermediate, or advanced AMD stages, unlike AREDS, which focuses only on fundus data. There is a lack of labeled severity grading public datasets for OCT images. Manual labeling by ophthalmologists is labor-intensive and subjective, leading to small annotated sets—e.g. (El-Den et al., 2024) used a private dataset for 7-class grading but faced precision issues.

These challenges impact the effectiveness of ML models. Convolutional Neural Networks (CNNs) like ResNet and VGG, prevalent in AMD detection, require large, balanced, high-quality datasets to avoid overfitting, a frequent limitation in the studies reviewed above. Scarcity of severity labels affects multi-class grading, pushing reliance on other types of optic imagery datasets like AREDS (fundus images) or private collections, though even these vary in scope and quality.

Our approach integrates deep learning for initial classification and unsupervised clustering for severity estimation, leveraging the Mendeley OCT dataset. The step-by-step process, as shown in Figure 5, for classifying and grading AMD from OCT images begins with an OCT image input, capturing cross-sectional scans of the retina. The images are then subjected to a pre-processing stage before undergoing feature extraction to identify key attributes indicative of AMD. These features are processed by a CNN and are then classified, where the model determines whether the image shows signs of AMD. The primary classification stage categorizes each image as Dry AMD, Wet AMD, or No AMD. If Dry AMD is detected, the model proceeds to a secondary classification phase to further grade the disease based on severity. The graded classification aims to provide detailed insights into disease progression, which can guide potential interventions.

Dataset

The “Large Dataset of labeled Optical Coherence Tomography (OCT) and Chest X-Ray Images” dataset from Mendeley was used (Kermany et al., 2018) for this study. The dataset is widely recognized in ophthalmic research and provides high-quality OCT images necessary for training models for AMD detection. The dataset contains approximately 108,000 labeled images for the following categories: Normal, Choroidal Neovascularization (CNV), Diabetic Macular Edema (DME), and Drusen. Initial exploration of the Mendeley OCT dataset was conducted to assess its structure, quantify its distribution, and verify its quality and four classes; CNV, DME, Drusen, and Normal, were confirmed. For the proposed classification classes (Figure 5), there were three classes used instead of the four classes from the datasets. The CNV class was mapped to the Wet AMD class, the DME and Normal classes were mapped to the No AMD class and the Drusen class was mapped to the Dry AMD class respectively.

The training set distribution revealed an imbalance, with approximately 51,000 Normal, 37,000 CNV, 11,000 DME, and 8,616 Drusen images, while the test set was uniform across all classes (250 images each). The training data imbalance suggests the need for pre-processing adjustments, such as data augmentation for under-represented classes or Synthetic Minority Oversampling Technique (SMOTE). The dataset also lacks explicit severity labels (e.g. early, intermediate, advanced AMD), posing challenges for direct severity grading and requiring unsupervised ML methods to infer progression.

Sample images from the training set (four per class) were visualized in grayscale, as shown in Figures 6–9, highlighting retinal features like drusen deposits and CNV irregularities, validating the dataset's relevance for AMD detection. Image analysis revealed variability in image dimensions across classes, necessitating image resizing as a pre-processing step. No corrupted files were identified, ensuring data integrity for all subsequent processing stages.

Preprocessing

The Mendeley OCT dataset was pre-processed to ensure compatibility with pre-trained convolutional neural networks (ResNet-18 and VGG-16) and to address challenges identified in Section 3.1, such as varying image dimensions and training set imbalance. Preprocessing was implemented with distinct pipelines for training and testing data. For the training set, images were first resized to 224 × 224 pixels to match the input dimensions of ResNet-18 and VGG-16, mitigating the size variability observed across classes. Data augmentation was performed to address the dataset imbalance (i.e., 51,000 Normal vs. 9,000 Drusen images) by flipping horizontally and applying rotation up to ±20 degrees on randomly selected images. These transformations enhanced the model validity by simulating variations in OCT scan orientations. Images were then converted to tensors, while duplicating the greyscale channel into three identical channels to form an RGB format compatible with the models' 3-channel input. Finally, pixel values were normalized in the range of [−1, 1] per channel to stabilize training gradients.

For the test set, pre-processing omitted augmentation to ensure consistent evaluation. Images were resized to 224 × 224 pixels, converted to 3-channel tensors, and normalized similarly to the training set. The pre-processed datasets were loaded in batch size of 32, shuffling enabled for training, and 24 worker threads for efficient GPU utilization on CUDA. No corrupted images had to be removed, as data integrity was confirmed during exploratory data analysis. This data pre-processing pipeline produced a standardized input suitable for the subsequent classification and clustering stages.

Classification model training

This section details the classification model with two convolutional neural networks (CNNs), ResNet-18 and VGG-16, for classifying OCT images from the pre-processed Mendeley OCT dataset into four classes: Normal, CNV, DME, and Drusen. The training process was executed using PyTorch and CUDA-enabled GPU acceleration, and incorporated a number of optimisation techniques and libraries to improve the accuracy. Both models were fine-tuned from ImageNet pre-trained weights, with their final layers adapted to the four-class task, and trained over multiple epochs with detailed monitoring of loss and accuracy metrics.

Model architecture and initialization

ResNet-18 and VGG-16 were chosen for their complementary strengths in medical imaging tasks (Xu et al., 2023; Majib et al., 2021). ResNet-18, with 18 layers and residual connections, balances computational efficiency (11.7M parameters) and depth, mitigating vanishing gradients via skip connections. VGG-16, with 16 layers of 3 × 3 convolutions and 138M parameters, offers greater feature extraction capacity but at a higher computational cost. Both models were initialized with ImageNet pre-trained weights to leverage general feature extraction capabilities, then fine-tuned for OCT-specific patterns such as retinal fluid in DME and drusen deposits. For ResNet-18, the original fully connected (fc) layer was replaced with a new layer configured to output four classes. Similarly, in VGG-16, the final layer of the classifier stack was adjusted to produce four outputs. These modifications enabled both architectures to support the four-class OCT classification task while preserving their pre-trained convolutional feature representations.

Training dynamics

Cross-entropy loss was used to penalize wrong classifications among the four classes. The Adam (ADAptive Moment estimation) optimisation method was selected over SGD (Stochastic Gradient Descent) because of its adaptive learning rate and incorporates momentum to improve convergence speed and stability. Both models were trained for 10 epochs, since there was no significant improvements after the 10 epochs. For the ResNet-18 model, a step learning rate scheduler was used to lower the learning rate by a factor of 0.1 after five epochs, helping in fine-tuning the model's weights later in training. Gradient clipping was applied to prevent very large gradients, enhancing training stability.

ResNet-18 showed faster convergence, with validation accuracy rising from 0.9560 (Epoch 1) to 0.9830 (Epoch 7), then plateauing at 0.9820 (Epoch 10). An early stopping mechanism (patience = 3) halted training if validation loss did not improve, saving the best model at Epoch 7. VGG-16 progressed more gradually, peaking at 0.9820 (Epoch 7) and maintaining gains through Epoch 10 without early stopping. Epoch-wise metrics for ResNet-18 are shown in Table 2.

VGG-16's metrics followed a similar trend, with a best validation accuracy of 0.9820 at Epoch 7 (train loss = 0.0725, train acc = 0.9757). The slight edge in ResNet-18's peak accuracy (98.3 vs. 98.2%) and significantly lower training time made it suitable as the primary model for feature extraction in clustering. ResNet-18's 512-dimensional penultimate layer features were selected for extraction for K-means clustering to grade Drusen severity due to its better accuracy and efficiency.

Transformer-based baseline (ViT-B/16)

To provide an experimental comparison against the transformer-based architectures discussed in Section 2.2, a Vision Transformer (ViT-B/16) (Dosovitskiy et al., 2021), pre-trained on ImageNet-1k, was fine-tuned on the same four-class Mendeley OCT classification task (Normal, CNV, DME, Drusen). To isolate the effect of architecture, training conditions were matched as closely as possible to the CNN baselines: identical train/test split, 224 × 224 input resolution, batch size of 32, horizontal-flip and ±20° rotation augmentation, and a 10-epoch budget with early stopping (patience = 3).

Two aspects of the training recipe were adapted specifically for the transformer, reflecting established differences between fine-tuning convolutional and attention-based architectures. The network was optimized with AdamW rather than Adam, decoupling weight decay from the adaptive gradient update, with a lower learning rate (3 × 10 −5) and cosine annealing in place of the step-decay schedule used for the CNNs, since pre-trained attention weights are more sensitive to large early-training updates than pre-trained convolutional filters. Input images were normalized using the ImageNet mean and standard deviation [(0.485, 0.456, 0.406)]/(0.229, 0.224, 0.225)] matching ViT-B/16's pre-training statistics, rather than the normalization applied for ResNet-18 and VGG16. All other pre-processing and evaluation procedures followed Sections 3.2 and 4.1.

Unsupervised severity grading framework

Due to the absence of a labeled dataset with Drusen severity annotations, an unsupervised clustering-based severity grading framework, using features derived from the trained ResNet-18 model, was developed to enable exploratory blind severity grading. Clinically, the severity of dry AMD is closely linked to drusen size. Features such as Drusen size, number of Drusen, and total area increase monotonically with disease progression. Features that encode drusen-related information can be leveraged for unsupervised severity estimation. An exploratory analysis was therefore conducted to examine whether the deep features extracted from the ResNet-18 classification model could reveal underlying severity-related patterns in the absence of pre-defined labels, offering a data-driven representation of underlying drusen severity.

Features were extracted from ResNet-18's penultimate layer, utilizing its fine-tuned weights. All 8,616 Drusen images from the training set were processed to maximize the data pool for unsupervised exploration. The final classification head was replaced with an identity layer to output raw 512-dimensional vectors. Images were pre-processed with ImageNet normalization and resized to 224 × 224 pixels.

The dimensionality reduction stage was performed using UMAP (Uniform Manifold Approximation and Projection), which was selected over linear alternatives such as PCA because the 512-dimensional ResNet feature vectors lie on a non-linear manifold. PCA captures only directions of maximum variance and can fail to preserve the local neighborhood structure. UMAP, in contrast, learns the topological structure of the data and produces embeddings that preserve both local relationships between nearby samples and the global cluster geometry, making it well-suited to exploratory medical imaging tasks where class boundaries are not linearly separable (McInnes et al., 2018). It is also considerably faster and more scalable than t-SNE. A parameter sweep across n_components (2, 10, 50), n_neighbors (5, 15, 30), and min_dist (0.01, 0.1, 0.5) was conducted to identify the UMAP configuration that produced the most compact and well-separated clusters, as quantified by the silhouette score. K-means clustering algorithm with k = 3 was applied to the best UMAP output. The choice of k = 3 is from the established clinical staging of dry AMD, the Beckman Classification, which formally defines three progressive stages: early, intermediate, and advanced. Each of which corresponds to increasing drusen burden (Ferris et al., 2013). This clinical evidence provides a biological basis for exploring three latent groupings in the feature space, and is further supported by the elbow plot analysis, which showed the within-cluster sum of squares flattening beyond k = 3.

A comprehensive volumetric assessment of AMD severity ideally incorporates drusen volume, total area, and geographic atrophy derived from full OCT volumes, but individual B-scans with pathology present, can carry sufficient structural information to support preliminary severity grading. Each B-scan provides a cross-sectional view of the retinal layers at a given position, directly revealing drusen height, sub-RPE deposits, and RPE irregularities that are primary indicators of AMD stage Thomas et al. (2021). To address the absence of clinically interpretable drusen quantification, drusen area and equivalent diameter were measured directly from the raw OCT B-scans for all 8,616 Drusen images. Each image was processed through a classical segmentation pipeline: the lower 60% of the image was isolated to focus on the sub-RPE region where drusen deposits are located, a Gaussian filter suppressed speckle noise, and Otsu thresholding separated hyper-reflective drusen candidates from surrounding tissue. Connected-component analysis then identified individual drusen blobs, discarding noise artifacts below 50 pixels and non-drusen bright regions exceeding 30% of image area. For each image, total drusen area (summed blob area in pixels) and equivalent diameter (d = 2 A / π) were computed, the latter chosen to mirror the clinical drusen-size convention used in the Beckman Classification (Ferris et al., 2013).

These per-image measurements were merged with the cluster assignments produced by the UMAP + K-means framework, and the three clusters were re-ordered by ascending mean total drusen area, yielding a clinically motivated correspondence: the cluster with the smallest mean drusen area was labeled Early, the intermediate cluster Intermediate, and the largest-area cluster Advanced. A Kruskal-Wallis test (non-parametric, appropriate given the non-normal distribution of drusen area) was used to test whether drusen area differed significantly across the three clusters.

Cluster interpretation

The three resulting clusters are designed to correspond, in principle, to the clinically established stages of dry AMD (early, intermediate, and advanced). The grouping is based on deep features that capture drusen-related characteristics and drusen size generally increases with AMD progression. Therefore, clusters corresponding to greater drusen burden can be interpreted as representing increasing disease severity. However, the precise mapping of each cluster to a clinical stage cannot be confirmed without either (a) ophthalmologist annotation of a representative subset of images from each cluster, or (b) validation against a ground-truth labeled dataset such as AREDS NICE (2018), which provides explicit severity annotations. In the absence of such reference data, the clinical assignment remains inferential. This clustering strategy addresses the lack of labeled datasets for the flexibility of data-driven exploration, enabling preliminary grading where annotated severity data are unavailable. The combination of UMAP for manifold projection and K-means for partitioning aligns well with this exploratory objective. However, the assumption of spherical cluster boundaries inherent to K-means may inadvertently oversimplify the continuous and heterogeneous progression of drusen-related pathology.

Transfer learning on the Duke dataset

While the Mendeley OCT dataset was used for the core classification task, the model adaptability was tested on the Duke OCT dataset (Farsiu et al., 2014) to assess whether the trained ResNet-18 and VGG16 models could generalize beyond the learning data distribution to an external dataset with distinct characteristics (e.g.,.mat format, AMD vs. Control classes).

Each Duke.mat file stores one patient's OCT volume as an image array of dimensions (height × width × num_slices), where the third dimension indexes individual B-scans across the macula. Volumes were parsed with SciPy's loadmat, and every slice was extracted via images[:,:, i] for i up to num_slices. All slices in each volume were retained, including peripheral scans, to maximize data for the binary AMD/Control task. Each greyscale slice was converted to 3-channel RGB by channel replication (PIL. convert('RGB')), resized to 224 × 224, and normalized with ImageNet statistics to match the pre-trained ResNet-18/VGG-16 input requirements. No quality or motion-artifact filtering was applied as data integrity was assumed given the curated nature of the public release.

An 80/20 split train/test was performed at the slice level, not the patient level. All B-scans from both classes were pooled and divided, so slices from the same patient volume could appear in both training and test sets. This is a limitation as slice-level splitting overstates the independence between train and test samples, since slices from one volume share anatomy and acquisition conditions, and the reported post-transfer-learning accuracies may therefore be somewhat optimistic relative to a strict patient-wise split. The result across datasets still has value as a demonstration of domain adaptation, since the Mendeley–Duke domain shift (different scanners, protocols, and class structures) is a genuine source of variance regardless of split strategy.

The results focused on the classification and clustering of the proposed model using the Mendeley OCT dataset and evaluation across datasets using the Duke dataset with transfer learning. Evaluation metrics, including accuracy, precision, recall, F1-score, and specificity, are reported alongside visualizations such as confusion matrices, t-SNE plots, and elbow curves. These findings demonstrate that ResNet-18 and VGG16 serve as effective baselines for retinal OCT classification, capable of adapting to different data sources.

Evaluation metrics

Several evaluation metrics [“Equations 1–5”] are used to assess the model's effectiveness across AMD classifications, including:

Accuracy: measures the overall proportion of correct predictions made by the model.

Precision: important for understanding the model's effectiveness in correctly identifying AMD cases without over-prediction.

Recall (Sensitivity): evaluates the model's ability to detect AMD cases accurately, critical for identifying true positive rates.

Specificity: measures the model's ability to correctly identify non-AMD cases, reducing false positives.

F1 Score: this harmonic mean of precision and recall is particularly useful for imbalanced datasets where some AMD stages may have fewer examples.

where TP = True Positives (correctly identified AMD cases), TN = True Negatives (correctly identified non-AMD cases), FP = False Positives (incorrectly labeled as AMD cases), and FN = False Negatives (AMD cases incorrectly labeled as non-AMD cases).

Evaluation with Mendeley dataset

The fine-tuned ResNet-18 and VGG16 models were evaluated on a balanced Mendeley test set (1,000 images: 250 each of CNV, DME, Drusen, Normal). ResNet-18 achieved 98% accuracy and VGG16's achieved 96% accuracy, both reflecting strong generalization from training. Table 3 shows the per-class metrics. For ResNet-18, precision ranged from 0.94 for CNV, 0.99 for DME, to 1.00 for Drusen and Normal. Recall reached 1.00 for CNV, DME and Normal and 0.94 for Drusen. The F1-scores were 0.97 for CNV and Drusen and 1.00 for DME and Normal respectively.

VGG16 showed similar results, with precision ranging from 0.88 for CNV, 0.99 for DME, to 1.00 for Drusen and Normal. Recall started from 0.86 for Drusen, 0.99 for CNV, to 1.00 for DME and Normal. The F1-scores were 0.92 for Drusen, 0.93 for CNV and 1.00 for DME and Normal. Specificity was between 0.98 to 1.00 for both ResNet-18 and VGG16, though ResNet-18 misclassified fewer Drusen (16 vs. VGG16's 34). These results cement the models' reliability for the four-class task, setting a high bar.

ViT-B/16 was evaluated on the same balanced Mendeley test set (1,000 images) used for ResNet-18 and VGG16. ViT-B/16 achieved an overall accuracy of 95%, positioning it after VGG16 and ResNet-18. Precision remained high across all classes and recall for CNV, DME, and Normal was likewise high. Consistent with the pattern already observed for ResNet-18 and VGG16, the principal source of error for ViT-B/16 was Drusen recall (0.85), with the corresponding drop in CNV precision (0.87) and specificity (0.95) indicating that most Drusen errors were predicted as CNV. This Drusen recall is slightly lower than both ResNet-18 and VGG16, indicating that the transformer's global attention-based feature aggregation did not resolve the Drusen/CNV visual overlap already present in the CNN baselines.

The confusion matrices for the 4 classes shown in Figure 10 demonstrate that for ResNet-18 the classification performance was excellent across all classes, with a small proportion (0.06) of Drusen samples misclassified as CNV, reflecting the inherent visual similarity between these conditions in OCT imaging. For the VGG-16 model, the classification performance was excellent across the 3 classes (CNV, DME, Normal) with limited misclassification (0.14) of Drusen cases as CNV cases.

Unsupervised severity grading

The unsupervised clustering of 8,616 Drusen images produced three clusters using the best UMAP configuration (n_components=2, n_neighbors=30, min_dist=0.01) and K-means (k = 3). Cluster sizes were 3,428,2,155, and 3,033 respectively with a silhouette score of 0.49 that indicates moderate separation across all samples as shown in Table 4.

The t-SNE Visualization shown in Figure 11A confirms distinct groupings among 3 clusters with minimal overlapping in 2D space. As no ground-truth severity labels are available, assigning the identified patterns to clinical stages (e.g., early, intermediate, and advanced AMD) requires expert clinical annotation or labeled reference data. Consequently, the current findings should be interpreted as exploratory rather than definitive. The elbow plot shown in Figure 11B supports the choice of k = 3 clusters, with within-cluster sum of squares flattening beyond this point.

To provide exploratory clinical interpretation of the unsupervised clusters, drusen area was measured independently for all 8,616 images and compared across the three clusters to investigate whether they represented distinct levels of disease severity. The results, summarized in Table 5, show a consistent monotonic increase in both mean total drusen area and mean equivalent diameter from the Early to the Advanced cluster: mean total area rose from 36,425.78 px (Early) to 38,556.40 px (Intermediate) to 46,391.46 px (Advanced), with median values following the same ordering (32,623, 33,760, and 40,982 px respectively). Mean equivalent diameter increased similarly, from 134.20 px (Early) to 135.64 px (Intermediate) to 151.11 px (Advanced). A Kruskal-Wallis test confirmed that these differences were statistically significant (H = 270.39, p < 0.001), indicating that drusen area is not uniformly distributed across the three clusters but varies systematically with cluster assignment.

Figure 12 illustrates this relationship, showing the distribution of total drusen area across the three severity clusters. The separation between Early and Intermediate clusters was comparatively modest (a 5.8% increase in mean area), with overlapping interquartile ranges reflecting the continuous and gradual nature of drusen progression in its earlier stages. In contrast, the separation between Intermediate and Advanced was more pronounced (a 20.3% increase in mean area), suggesting that the clustering more reliably distinguishes advanced disease from earlier stages than it does the early-to-intermediate transition. This pattern is consistent with the clinical understanding that drusen growth accelerates non-linearly as AMD progresses (Urbanska et al., 2024), and supports interpreting advanced stage as a distinct, more separable advanced-disease grouping, while early and intermediate stages occupy a more continuous region of the feature space.

Evaluation with Duke dataset

The evaluation of Mendeley-trained models on the Duke dataset assessed their performance on a binary AMD/Control classification task using an external OCT dataset. An initial testing performed without transfer learning yielded low accuracy, with ResNet-18 achieving 44.8% and VGG16 getting 43.9% respectively on an unbalanced test set as displayed in Table 6.

Following transfer learning, performance improved significantly on a balanced test set (2289 Control, 2289 AMD), with ResNet-18 attaining 82.24% accuracy and VGG16 reaching 83.88% respectively as shown in Table 7. VGG16 has a higher recall than ResNet-18 for Control (0.8943 vs. 0.8056) and higher precision for AMD (0.8811 vs. 0.8119). The results demonstrate that both models can serve as effective baselines for AMD detection across different datasets.

Discussion

The evaluation on the Mendeley OCT dataset demonstrates strong generalization performance for both deep learning architectures, with classification accuracies of 98% for ResNet-18 and 96% for VGG16. While both models performed competitively, ResNet-18 exhibited superior recall for the Drusen class (0.94 compared to 0.86 for VGG16). This improvement is likely attributable to the residual learning framework of ResNet-18, which facilitates more effective feature extraction and gradient propagation, particularly for subtle and heterogeneous pathological patterns such as drusen. The achieved performance is comparable to, and in some cases rivals, state-of-the-art approaches that did not perform validation across datasets and severity grading, including the method proposed in (Kadry et al., 2022)(97%) and the vision-transformer-based model reported in (Akça et al., 2024) (99%). These findings suggest that relatively lightweight convolutional architectures, when appropriately fine-tuned, remain highly effective for OCT-based retinal disease classification. For ResNet-18, most Drusen samples are correctly classified, while 6% are confused with CNV, suggesting some overlap in structural features between Drusen and CNV cases respectively in OCT images. This overlap can be expected, since intermediate to advanced Drusen can share visual characteristics with neovascular changes. A similar pattern is observed from the results for VGG16, with nearly 14% of Drusen cases are predicted as CNV, reflecting substantial visual similarity between Drusen and neovascular features in OCT images.

The inclusion of a Vision Transformer baseline provides a direct experimental comparison, for whether transformer architectures offer an advantage over convolutional networks on this task. ViT-B/16 attained 95% accuracy which is comparable to, but not exceeding, the CNN baselines (98% for ResNet-18, 96% for VGG16). This is broadly consistent with the wider transformer literature, which notes that the absence of convolutional inductive biases such as locality and translation equivariance can make transformers more reliant on large-scale pre-training data or extended fine-tuning schedules to match CNN performance on modest-sized, domain-specific datasets (Dosovitskiy et al., 2021). The value of this comparison is evidentiary: it situates the transformer literature against this dataset, and shows that the ResNet-18/VGG16 baselines used for the primary detection task hold up against an untested modern alternative rather than being superseded by one.

The evaluation on the Duke dataset highlights the adaptability of the proposed model under domain shift. After applying transfer learning, classification performance improved substantially to 82.24% and 83.88%, for ResNet-18 and VGG16 respectively. VGG16 demonstrated slightly more balanced classification behavior, achieving higher F1-scores (0.8473 and 0.8293 respectively), suggesting improved robustness across classes when fine-tuned on heterogeneous data.

The silhouette score of 0.49 for the K-means clustering for the exploratory severity grading indicates moderate cluster separation, which is appropriate for unsupervised drusen severity grading. Drusen progression occurs along a continuous spectrum and intermediate forms (between early and advanced AMD) often share overlapping structural characteristics with adjacent severity levels, which can limit sharp cluster boundaries. The proposed exploratory severity grading model relies solely on intrinsic feature similarities rather than annotated severity labels, thus it is expected to obtain some resulting overlap between clusters. Overall, this silhouette score of close to 0.5 supports the suitability of the proposed unsupervised framework for exploratory drusen severity grading while emphasizing the need for domain expertise for clinical validation.

Beyond the silhouette score, drusen area measurements provide independent, clinically grounded evidence for the validity of the proposed clustering. The statistically significant monotonic increase in mean drusen area and diameter across the three clusters (Kruskal–Wallis p < 0.001) supports the interpretation of the clusters as a severity-ordered progression consistent with the Beckman Classification (Ferris et al., 2013), addressing the concern that the clustering is purely latent and clinically unanchored. However, the modest separation observed between the Early and Intermediate clusters indicates that the framework is more effective at isolating advanced disease than at finely discriminating between the earliest stages of drusen progression, a limitation consistent with the continuous and overlapping nature of early AMD pathology.

The results demonstrate that lightweight deep learning architectures remain highly effective for OCT-based retinal disease analysis, with ResNet-18 showing particular strength in detecting drusen-related features. Transfer learning evaluation across datasets confirms the robustness of the proposed framework under domain shift, while the moderate silhouette score observed for unsupervised severity grading reflects the continuous and overlapping nature of drusen progression, supporting its use for exploratory analysis with appropriate clinical interpretation.

Limitations and directions

Despite advancements in AMD detection and grading, several limitations persist in applying machine learning (ML) to ocular imaging (Sengupta et al., 2020). As discussed in Section 2.4, the scarcity of large, annotated datasets remains a critical challenge. Studies like (Kadry et al., 2022) and (El-Den et al., 2024) relied on small private datasets, not only leading to the labor-intensive and subjective task of manual annotations but also increasing the risk of over-fitting. Another challenge is the unavailability of standardized KPIs (Key Performance Indicators) for evaluating the performance of a particular model. Different researchers use different metrics to measure their work, making it difficult to compare various machine learning models for the same disease condition. Domain shift poses another significant barrier. This occurs when there is a mismatch between the data distribution used for training and the data encountered in real-world applications. In the case of OCT images, domain shift may arise due to variations in equipment, such as different OCT devices, camera settings, or imaging protocols, which can lead to differences in image quality, resolution, and contrast. These variations can impact the model's ability to generalize, resulting in reduced performance when the model is applied to real-world clinical data.

Some potential improvement in the direction of research will involve the development of advanced data augmentation techniques and the use of synthetic data can be helpful in addressing the problem of small datasets in medical imaging, especially ocular images. Generating synthetic OCT images through techniques like GANs (Generative Adversarial Networks) can help to supplement the limited annotated data (Frid-Adar et al., 2018). To address domain shift, future research can explore domain adaptation techniques that allow models to perform well across different OCT imaging systems and settings. Unsupervised domain adaptation methods that can align feature distributions between different domains have the potential to enhance the performance of models (Xie et al., 2024). Another promising area for future research is the fusion of multimodal data, where OCT images are combined with other types of medical data, such as genetic, clinical, or demographic information, to provide a more detailed approach to AMD classification and grading. Integrating these data sources using deep learning models can improve diagnostic accuracy (Wang et al., 2024), especially when dealing with early-stage AMD, where OCT images alone may not capture all relevant information.

AMD remains a major cause of irreversible vision loss, underscoring the need for timely detection and reliable severity assessment. This study presented a two-stage automated framework that combines deep learning-based AMD detection with unsupervised severity grading using OCT images, addressing the limitations of time intensive and expert dependent manual interpretation. The results demonstrate that lightweight convolutional architectures are highly effective for OCT based retinal analysis, with both ResNet-18 and VGG16 showing good strength in capturing drusen related features. Furthermore, transfer learning evaluation across datasets confirmed the robustness of the proposed approach under domain shift, while the moderate silhouette score achieved for unsupervised severity grading reflects the continuous and overlapping nature of drusen progression. Collectively, these findings support the potential of the proposed framework as a scalable decision support tool for AMD detection and exploratory severity assessment, while emphasizing the importance of clinical interpretation when translating unsupervised outputs into practice.

Despite these promising results, several limitations should be acknowledged. Firstly, the unsupervised severity grading lacks direct link with clinically defined severity stages due to the absence of severity labels, requiring post-hoc interpretation by clinicians. Secondly, public dataset reliance limits real-world testing. Thirdly, K-means clustering simplifies the cluster shape that leads to overlapping of clusters. Future efforts will focus on addressing these limitations and could integrate expert-annotated severity to use quantitative validation. Alternate clustering methods will be explored as well as extending the proposed framework to include longitudinal data to enable modeling of the progression of AMD over time.