Dataset

For the purpose of this study, the OCT dataset was gathered from two research centres and Labeled Optical Coherence Tomograph dataset [42]. AVL cases were collected from Eye Clinic, Department of Public Health, Federico II University in Italy and Chair and Department of General and Pediatric Ophthalmology at the Medical University of Lublin in Poland. These data were thoroughly verified. All illegible images were excluded. In total, 125 images were prepared for further study. From the OCT dataset [42] Drusen and normal cases were gathered. As a result, 577 images of Drusen, and 578 normal cases were assembled (Fig. 2).

The study was carried in accordance with the relevant guidelines and regulations, as well as, the Helsinki Declaration. The scope and method of conducted study were approved by the Ethics Committee of the Medical University of Lublin (no. KE-0254/260/12/2022). An informed consent was obtained from all study participants or their legal guardians. All data available as part of the conducted studies were anonymized to prevent obtaining personal patient data.

The gaussian and median filters were applied to decrease speckle noise and to keep essential structures of all OCT images.

Gaussian filtering was used to suppress high-frequency speckle noise by locally smoothing pixel intensities. In contrast, median filtering was used to suppress isolated intensity outliers without blurring diagnostically significant retinal layer boundaries. Both filters were applied with small local kernels before resizing and normalisation.

In case of AVL distortions, the region with the macula was manually selected in order to indicate the relevant area of this disease. Next, all images were resized to a fixed resolution of 224x224 pixels to be appropriate to input layer of CNN architectures. The intensities of each image were scaled to range between 0 and 1 in order to normalize pixel intensities and reduce variability of various imaging sessions. The dataset involved in this study consisted of 125 AVL images, 577 Drusen images and 578 images of normal cases. The whole dataset was divided into training, validation, and testing subsets, with a ratio of 70%, 15%, and 15%, respectively. The proportions of individual classes were maintained (Tab. 1).

In order to improve the generalization capabilities of machine learning architecture and its performance, the augmentation process was involved. It was utilized in model training to increase samples diversity and to reduce the risk of overfitting. The following image transformations were involved [43,44]: rotation, scaling, randomly cropping and resizing, colour jittering, gaussian noise, elastic deformation, random erasing, random masking by square regions, horizontal, and vertical flipping.

Deep CNN-GRU

Deep CNN-GRU architecture [17] integrates a deep convolutional backbone with a recurrent module so that both fine-scale structural cues and wider contextual dependencies present in optical‑coherence‑tomography (OCT) B‑scans are captured within a single end‑to‑end trainable network. The convolutional front‑end is organised as five consecutive blocks, each composed of two (3 ×3) convolutions employing zero padding and unit stride, rectified linear unit (ReLU) activations, and a (2×2) max‑pooling operation that halves the spatial resolution. Filter counts grow in a doubling scheme: (64, 128, 256, 512, 512), to enlarge the receptive field while preserving computational tractability. This design mirrors the philosophy of CNN‑style architectures, allowing deeper representational hierarchies without excessively large kernels. After the last pooling stage, the feature tensor has spatial dimensions (7 ×7) with 512 channels. Instead of flattening these maps directly, the tensor is first reshaped into a sequence of 49 feature vectors (one per spatial position), each 512-dimensional, and forwarded to a gated recurrent unit (GRU) layer.

The GRU processes the sequence row‑by‑row, using its update and reset gates to decide how much of the previously aggregated contextual information to retain and how much to overwrite with the current spatial sample. The network effectively models long‑range co‑occurrences between distant retinal regions that may be diagnostically linked yet spatially disjoint on a single B‑scan. The recurrent block therefore complements the strictly local modelling of the convolutional filters. The final hidden state emitted by the GRU serves as a global representation of the entire image. This vector is fed to a 64-unit fully connected layer that applies dropout regularisation during training to mitigate overfitting. A terminal Softmax layer with four logits transforms the network output into posterior probabilities over the target classes allowing straightforward optimisation via categorical cross‑entropy. Information therefore flows through the model as: [I_224×224×3→^{Conv-Block_1...5}F_7×7×512→^reshape{f_t}_t = 1⁴⁹→^GRUh→^FC₆₄z→^Softmaxp

where I denotes the input OCT slice, F the deepest feature maps, f_t the sequential embeddings, h the GRU state, z the fully connected activations, and p the predicted class distribution. Each stage is fully differentiable, enabling simultaneous tuning of convolutional kernels, recurrent weights, and classification parameters through back‑propagation and stochastic‑gradient descent.

Convolutional GRU U-Net

Convolutional GRU U-Net [18] fuses the symmetrical encoder–decoder topology of the original U‑Net with bidirectional convolutional gated‑recurrent units (BiConvGRU). Thereby enabling the model to exploit both multi‑scale spatial context and long‑range feature dependencies when analysing ultra‑wide‑field fundus images.

In the contracting branch, the network ingests a 300 ×375 pixel image that has been intensity‑normalised to the range [0,1]. Four successive resolution levels each apply a pair of 3 ×3

convolutions with zero padding and rectified‑linear‑unit activations, immediately followed by a 2 ×2 max‑pooling operation that halves the spatial dimensions while doubling the number of feature maps, thus generating a hierarchy of increasingly abstract representations.

At the bottleneck, the plain convolutions of the vanilla U‑Net are replaced by a stack of densely connected convolutional blocks in the spirit of DenseNet. Every block receives, as additional channels, the concatenation of all feature maps produced by its predecessors, so that gradient signals propagate unimpeded and redundant filters are suppressed through feature reuse. This design widens the effective receptive field without inflating the parameter count and provides rich semantic tensors for the subsequent decoding path.

The expansive branch begins by up‑sampling the encoded tensor through a transposed convolution that doubles the spatial resolution and halves the channel depth. Concurrently, the feature maps cropped from the mirror level of the encoder are resized by an identical up‑convolution to ensure perfect alignment.

Rather than concatenating the two streams outright, the architecture interposes a bidirectional convolutional GRU that processes the aligned tensors as a spatio‑sequential signal scanned in both raster directions. Inside each GRU cell, the reset and update gates, as well as the candidate activation, are implemented with 3 ×3

convolutions so that state transitions preserve the spatial layout. The forward and backward hidden states are finally concatenated, giving a context‑enriched feature block that is passed to the next decoding stage.

A batch‑normalisation layer follows every up‑convolution to stabilise the distribution of activations and accelerate convergence. Decoding proceeds through four such up‑sampling BiConvGRU stacks, progressively restoring full image resolution while blending coarse semantic cues from the bottleneck with fine localisation cues from shallow encoder layers.

The last decoder output feeds a 1 ×1 convolution that compresses the channel dimension to three logits, one per target class. A spatial softmax converts these logits into pixel‑wise posterior maps, and averaging these maps over the field of view yields an image‑level class distribution that is optimised end‑to‑end with categorical cross‑entropy.

Overall information flow can be summarised as:

where I is the input image, E the multiscale encoder output, B the densely connected bottleneck representation, D the reconstructed feature map, Z the class‑logit tensor, and ŷ the predicted class distribution. All operations are differentiable, allowing joint training of convolutional, recurrent, and classification parameters via back‑propagation and stochastic‑gradient descent.

Residual Attention Convolutional Neural Network

The Residual Attention Network (RAN) proposed in [45] combines deep residual learning with stacked soft–attention blocks so that diagnostically salient structures in OCT‑angiography B‑scans of the optic nerve head are selectively amplified while irrelevant background is suppressed. Each Attention Module (AM) contains a trunk branch, implemented as a pre‑activation residual unit and a mask branch that follows a bottom‑up / top‑down scheme to generate a spatially and channel‑wise adaptive weight map, which is applied multiplicatively to the trunk features before the residual skip connection is added back.

Input images of 352 ×352 ×3 are first processed by a 3 ×3 convolution (stride 1, padding 1) followed by a 2 ×2 max‑pooling layer, yielding feature maps of size 176 ×176. The core of the encoder then unfolds as four resolution stages. Each stage consists of one residual unit that halves the spatial resolution, immediately succeeded by an AM that refines the down‑sampled representation. Consequently, the feature tensor dimensions progress as 176 ×176 →88 ×88 →44 ×44 →22 ×22 →11 ×11. Filter depths grow in proportion to preserve capacity, while identity shortcuts inside every residual unit guarantee unobstructed gradient flow.

Within an AM, the mask branch performs successive max‑pooling operations to aggregate global context and then upsamples via bilinear interpolation, inserting skip links from matching encoder levels so that fine localisation cues are reinstated. Two 1 ×1 convolutions and a sigmoid activation restrict the soft mask M(z) to [0,1]. The final output of the module is H_i,c = M_i,c T_i,c, where T denotes the trunk features and the indices i,c enumerate spatial positions and channels, respectively. To prevent attenuation of residual signals when many AMs are stacked, the authors employ an augmented formulation H_i,c = (1 + M_i,c) F_i,c that preserves identity mapping when the mask goes to zero.

The mask branch itself is a fully convolutional encoder–decoder whose receptive field expands rapidly through repeated pooling, enabling capture of long‑range dependencies; symmetric upsampling ensures that the mask aligns pixel‑wise with the trunk output. A pair of skip connections between corresponding bottom‑up and top‑down levels further stabilises training and sharpens the attention contours. Three such AMs, parameterised by (x,m,n) = (1,2,1), denoting one pre‑processing residual unit, two trunk residual units and one residual unit between each pooling layer in the mask path, are cascaded, allowing progressive refinement of attention as depth increases.

After the final residual–attention pair at 11 ×11, global average pooling collapses the spatial dimensions, producing a feature vector that feeds a fully connected. A softmax transforms outputs into class posterior probabilities and the entire model is optimised end‑to‑end with categorical cross‑entropy.

The complete forward mapping can be formalised as:

where F_k denotes feature maps of spatial size k ×k, h the global descriptor generated by GAP, z the pre‑activation logits and p the predicted class distribution, GAP stands for global average pooling, FC denotes fully connected layers. All convolutional, residual, and attention parameters are updated simultaneously by back‑propagation, ensuring that spatial attention masks, residual filters, and classifier weights co‑adapt to maximise discriminative power.

Ensemble models

Bagging generates diversity by repeatedly resampling the original training set with replacement, so every bootstrap subset contains a unique mosaic of duplicated and omitted instances. A separate copy of the same base algorithm is fitted on each subset. Because the individual learners are trained on overlapping yet distinct views of the data, their errors tend to be weakly correlated. Final predictions are obtained by simple majority vote, in case of classification. The ensemble therefore drives down variance without inflating bias, yielding a model that is markedly less prone to over‑fitting than any single constituent learner. In ophthalmic scenarios a bagged ensemble ensures that no single high‑prevalence pattern determines the decision boundary, leading to more stable detection of rare findings.

Boosting takes an almost opposite route: instead of sampling the data, it reweights it. A first “weak” learner is trained on the full dataset; misclassified samples then receive higher weights, forcing the next learner to concentrate on precisely those difficult cases. This sequential reweighting repeats for a user‑set number of rounds. The final output is an error‑weighted vote of all intermediate models, so that better learners exert greater influence. By systematically focusing attention on outliers and borderline instances, boosting reduces bias while maintaining low variance. When confronted with subtle phenotypic variations, for example, faint AVL dystrophy patterns that can be overlooked by conventional classifiers, boosting iteratively sharpens the ensemble’s sensitivity, often correcting images that earlier rounds mislabelled.

Stacking introduces a meta‑learning layer on top of a heterogeneous collection of first‑level models. Each base learner, often drawn from fundamentally different families, first produces predictions for every instance. Those predictions, typically expressed as class probabilities, are then assembled into a new dataset on which a second‑level “meta‑learner” is trained. Because the meta‑learner observes where individual models agree or diverge, it can discover non‑linear relationships among their outputs and learn how to weight them optimally. Unlike bagging or boosting, stacking is free to exploit complementary inductive biases: an attention networks might excel at local texture cues, whereas a GRU based methods could leverage global colour statistics. The meta‑learner reconciles both. This flexibility is especially advantageous in multimodal or highly heterogeneous eye‑disease datasets (fundus photographs, autofluorescence, OCT), where no single algorithm consistently dominates across all imaging modalities or disease stages.

Explainable

Gradient‑weighted Class Activation Mapping is an attribution technique that produces coarse, human‑interpretable heatmaps showing which image regions most strongly influence a convolutional neural network’s prediction for a given class.

Given a trained model, let A^k denotes the k^th feature map of a selected convolutional layer and y^c the logit associated with class c. Grad‑CAM computes the class‑specific importance coefficient α_k^c = 1/Z∑_i,j(∂y^c)/(∂A_ij^k), where i,j stands for index spatial locations and Z is the total number of pixels in the map.

A weighted combination of the feature maps is formed, L_Grad‑CAM^c = ReLU(∑_kα_k^cA^k), and then up‑sampled to input resolution to yield the final localization map. Because gradients are taken with respect to activations that preserve spatial structure, the heatmap highlights regions that provide positive evidence for the target class, while the ReLU operation suppresses negative contributions [23].

Metrics

The CNN architectures are evaluated using the commonly applied metric assessing the performance of machine learning models, such as: accuracy (1), precision (2), recall (3), and F1-score (4). These measures provide deepen understanding of model’s predictive power and its ability to correctly classify samples. In eq. (1)-(4) TP, FP, TN, and FN, correspond to true positive, false positive, true negative, and false negative, respectively.

Accuracy measures the number of correctly predicted samples across all samples (1). Precision assesses the model accuracy of the positive predictions (2). This metric is of great importance in medical diagnosis. The higher the value is, the more correct the model prediction is obtained. Recall provides information about all positive cases, which is extremely important in medical diagnosis where all distortions should be recognized (3). F1-score combines precision and recall to obtain one measure indicating the model’s performance (4).

Single Convolutional Neural Networks

The performance for each class for the Deep CNN-GRU Network, Convolutional Gated Recurrent Units U-Net (CGRU U-Net), and Residual Attention Network (RAN) are gathered in Tab.2. The 5-fold cross validation is utilized in order to enhance the reproducibility and stability of the obtained results. The confuses matrices are presented in Fig. 3-5. The training, validation accuracies and loss functions for three CNN architectures are depicted in Fig. 6-8. These results prove that the proposed architectures have the ability to distinguish normal cases from these affected by Drusen or AVL. The high values of recall, exceeding 85%, and moderate precision results for AVL detection state that the model catches most positive cases, but it also incorrectly labels many negative cases as positive. For Drusen and heatlty images, the models are more effective.

The Deep CNN-GRU Network achieved 90.20% average model accuracy.

The analysed CNN architectures have high average accuracy for AVL, Drusen, and normal cases detection, above 89% (Fig. 3-5). The Deep CNN-GRU Network identifies AVL with 94.74% accuracy. This disease is rarely misclassified with Drusen, up to 5.26%. Drusen images are classified with the accuracy of 85.06%. They are confused with other classes only up to 8.06%. In case of healthy cases, they are very rarely misclassified with images presenting retinal diseases. The Convolutional Gated Recurrent Units U-Net detects all classes with high accuracy, ranging between 88.51% for normal cases and 90.90% for Drusen. Residual Attention Network identifies AVL and Drusen with the highest accuracy, exceeding 94%. Slightly worse result is obtained for normal images. As in case of Deep CNN-GRU, the Convolutional Gated Recurrent Units U-Net and Residual Attention Network confuse AVL only with Drusen, while Drusen and healthy images are rarely misleading with other classes.

The accuracies obtained while training and validation and loss function provide the effectiveness of three CNN-based architectures (Fig. 6-8). The learning dynamics are similar for all models. The architectures reach a high level of accuracy, reaching 20 epochs or even earlier. The difference between the training and validation curves remains low. In the context of the difficult, sparse AVL class, high validation accuracy suggests that the network effectively recognizes even relatively rare disease patterns. At the same time, the loss function decreases smoothly for both training and validation, stabilizing within the range of 0.1-0.2.

All obtained results confirm the effectiveness of the deep learning approach for AVL, Drusen, and normal cases identification.

Explainability of CNN-based architectures

The Grad-CAM heatmaps are applied to highlight the image regions that are the most relevant for AVL, Drusen, and healthy cases identification. The warmer colour is the part of the image of the greatest importnt for the model’s decision is. The colder tones indicate the lower significante of the particular regiion.

The Grad-CAM heatmaps for Drusen identification by the single CNN-based models are depicted in Fig. 9 b)-d). The symptoms of this disease are shown in the Fig. 9 a), marked with a red border by a professional ophthalmologist. It should be emphasized that all models considered in this study focus on the image region containing retinal distortions in the form of Drusen. However, in the case of the Deep CNN-GRU Network and Convolutional Gated Recurrent Units U-Net, it can be observed that the leftmost part of the distortions is omitted as important for the diagnosis of this disease. The Residual Attention Network model, on the other hand, increases the image area, based on which it makes decisions about classifying Drusen. As a result, the most significant region, dark red, concentrates all the distortions associated with this retinal disease. Therefore, RAN is the most suitable model for identification of this type of retinal lesions. The Grad-CAM heatmaps play a crucial role in improving the explainability of the models, which is especially important in diseases identification. This technique proves the models involved in this study to make the decision based on relevant regions. Moreover, this technique highlights the distortions and stay as an additional a tool for ophthalmologists to indicate the relevant areas on OCT imaging.

Ensemble learning

The experiments involving bagging, boosting, and stacking ensemble learning techniques are performed in order to enhance the performance of the analysed architectures in this study. The new models are created utilizing all three single CNN-based networks. The final response of the bootstrap model is determined by majority voting. In stacking technique, the XGBoost was chosen as a metamodel. In case of all types of ensemble learning the performance of retinal diseases identification is improved. Bagging enhanced the detection rate of AVL, Drusen, and normal cases by 6.91%, 7.82%, and 5.29% compared to the Deep CNN-GRU Network, Convolutional Gated Recurrent Units U-Net, and Residual Attention Network, respectively. Boosting method increased the mean accuracy by 6.39%, 7.30%, and 4.77% as opposed to the Deep CNN-GRU Network, Convolutional Gated Recurrent Units U-Net, and Residual Attention Network, respectively. The stacking model achieved the highest accuracy, enhancing the model’s performance by 7.43%, 8.34%, and 5.81% of the Deep CNN-GRU Network, Convolutional Gated Recurrent Units U-Net, and Residual Attention Network, respectively. The obtained results in Tab. 3 clearly prove that bagging, boosting, and stacking ensemble learning methods are suitable techniques to improve the model's ability to recognize AVL, Drusen and that discriminates them from healthy cases.

Comparison with the state-of-the-art

In order to compare the results obtained in this study for the proposed single CNN-based architectures and their combinations utilizing ensemble learning techniques, the state-of-the-art deep learning models, which are often applied in disease identification based on OCT imaging, were selected. The VGG-16, ResNet-18, InceptionV3, and DenseNet-121 were chosen due to their high-performance achieved. Their ability to identify the AVL, Drusen, and healthy cases using the dataset from this study, are computed (Tab. 4). The ResNet-18, InceptionV3, and DenseNet-121 demonstrates good model performance, obtaining average model accuracy of 82.86%, 82.64%, and 82.80%, respectively. Slightly worse accuracy, 78.64%, was obtained for the VGG-16 model. It should be noted that these models achieved the lowest precision for AVL identification, which also resulted in the lowest F1-score values for this disease.

All studies were conducted using a computer equipped with a 24-core AMD Ryzen Threadripper PRO 7965WX processor clocked at 4.20 GHz. The computer uses 512 GB of RAM. The system also benefits from the support of a NVIDIA GeForce RTX 4090 graphics card. The equipment runs in a 64-bit operating system environment. The comparison of computational complexity and average processing time is presented in Tab. 5.

At the current level of development of retinal diseases detection, it can be seen that the Deep CNN-GRU Network, Convolutional Gated Recurrent Units U-Net, and Residual Attention Network provide significantly better ability to recognize AVL, Drusen, and healthy cases based on OCT imaging, achieving average models accuracy in the range of 89.59%-92.12%. None of the state-of-the-art architectures supplies these distortions identification at a similar level using the same dataset. The VGG-16, ResNet-18, InceptionV3, and DenseNet-121 achieved the highest ability to recognize healthy cases. The same situation occurs with the Deep CNN-GRU Network. However, the Convolutional Gated Recurrent Units U-Net, Residual Attention Network, as well as boosting and stacking models identify retinal distortions with the highest performance, which is extremely important in medical diagnosis. Moreover, the ensemble learning methods prove to enhance ability to recognize all retinal diseases and healthy cases. Due to overcoming individual model limitations and delivering more reliable results, the new models are characterized by better generalization to new samples.

The performance of all architectures, involved in this study, were verified using accuracy and loss curves (Fig. 8, 10). All accuracy curves increase over time gradually, enhancing the model’s performance. All networks learn from the training data properly, in stable and consistent manner up to reach convergence. It means that all models understood important patterns to make accurate predictions. The loss curves indicate that the models improve their performance while the loss gradually decreases.

The Deep CNN-GRU Network, Convolutional Gated Recurrent Units U-Net, and Residual Attention Network, as well as proposed new ensemble learning models outperform the state-of-the-art architectures.