Advancing Visual Perception Through VCANet-Crossover Osprey Algorithm: Integrating Visual Technologies

Abstract

Diabetic retinopathy (DR) is a significant vision-threatening condition, necessitating accurate and efficient automated screening methods. Traditional deep learning (DL) models struggle to detect subtle lesions and also suffer from high computational complexity. Existing models primarily mimic the primary visual cortex (V1) of the human visual system, neglecting other higher-order processing regions. To overcome these limitations, this research introduces the vision core–adapted network-based crossover osprey algorithm (VCANet-COP) for subtle lesion recognition with better computational efficiency. The model integrates sparse autoencoders (SAEs) to extract vascular structures and lesion-specific features at a pixel level for improved abnormality detection. The front-end network in the VCANet emulates the V1, V2, V4, and inferotemporal (IT) regions to derive subtle lesions effectively and improve lesion detection accuracy. Additionally, the COP algorithm leveraging the osprey optimization algorithm (OOA) with a crossover strategy optimizes hyperparameters and network configurations to ensure better computational efficiency, faster convergence, and enhanced performance in lesion recognition. The experimental assessment of the VCANet-COP model on multiple DR datasets namely Diabetic_Retinopathy_Data (DR-Data), Structured Analysis of the Retina (STARE) dataset, Indian Diabetic Retinopathy Image Dataset (IDRiD), Digital Retinal Images for Vessel Extraction (DRIVE) dataset, and Retinal fundus multi-disease image dataset (RFMID) demonstrates superior performance over baseline works, namely EDLDR, FFU_Net, LSTM_MFORG, fundus-DeepNet, and CNN_SVD by achieving average outcomes of 98.14% accuracy, 97.9% sensitivity, 98.08% specificity, 98.4% precision, 98.1% F1-score, 96.2% kappa coefficient, 2.0% false positive rate (FPR), 2.1% false negative rate (FNR), and 1.5-s execution time. By addressing critical limitations, VCANet-COP provides a scalable and robust solution for real-world DR screening and clinical decision support.

Introduction

Background Knowledge About Lesion Recognition from Retinal Images

Diabetic retinopathy (DR) is a leading cause of vision impairment among individuals with diabetes. It is characterized by a variety of retinal lesions, including microaneurysms, hemorrhages, and exudates. The early symptoms include blurred vision and eye floaters leading to complete vision loss over time in untreated conditions [1]. Hence, the early diagnosis of DR is crucial in preventing its progression across various stages and reducing the severe outcomes [2]. Moreover, the detection of DR in the early stages allows for timely interventions such as laser treatment or blood sugar control. Color fundus images are instrumental in detecting and classifying DR which enables timely diagnosis to prevent vision impairment or loss [3].

With various advancements in the field of medical imaging technology, the role of medical images present in clinical diagnosis has become more critical. Medical image analysis plays a major role and indispensable part in diagnosing various conditions and guiding clinical therapy, significantly boosting the precision and effectiveness of patient care [4]. In recent years, an emerging strategy to enhance DR screening involves the use of artificial intelligence (AI) to analyze digital fundus images [5]. It has significantly transformed medical imaging by providing robust solutions for the automated analysis and early detection of DR from fundus images [6]. Traditionally, the healthcare systems screen for DR through direct clinical exams and nonmydriatic fundus images. However, the manual assessment of retinal images by healthcare professionals was time-consuming and prone to subjectivity.

In contrast, DL techniques such as convolutional neural networks (CNNs) have demonstrated remarkable capabilities in lesion recognition that are often difficult to interpret by healthcare professionals. It imitates the human visual system and has shown various promises in identifying diverse features associated with DR [7]. Furthermore, DL algorithms also support clinicians by providing decision-support tools and enabling more timely and accurate diagnoses. Despite everything, traditional models struggle to capture subtle lesion features and generalize across diverse datasets due to high computational costs and suboptimal hyperparameter tuning. Moreover, the integration of DL with bionic vision principles has the potential to revolutionize DR screening and diagnosis by offering improved accuracy and reliability [8]. However, earlier works have mainly focused on mimicking the V1 cortex, neglecting the crucial roles of V2, V4, and IT regions. This limited approach reduces the model’s ability to extract comprehensive subtle features, impairing lesion detection in retinal images [9].

To resolve these problems, the VCANet-COP model is designed for accurate lesion recognition in retinal images. Primarily, SAEs are deployed to simultaneously extract vascular features (vessel structure, thickness, tortuosity) and lesion-specific features (microaneurysms, hemorrhages, exudates) from the retinal images by focusing on pixel-level patterns for precise lesion localization. Moreover, the VCANet-COP model resolves the problem of subtle lesion feature detection by utilizing a two-part structure. The front-end processing network modeled based on the primate visual system (V1, V2, V4, IT regions) is used to capture hierarchical features at multiple levels, enhancing the model’s ability to identify subtle lesion features. The back-end CNN further refines these features for accurate lesion recognition. Particularly, the COP algorithm which integrates the OOA with a crossover strategy optimizes hyperparameters and network configurations by diversifying genetic material, enhancing search efficiency, and mitigating premature convergence. This results in reduced computational costs and improved execution efficiency, offering a scalable and robust solution for real-world DR diagnosis.

Motivation

The motivation behind this research arises from the need to overcome the limitations of traditional DL models in performing subtle, small, or irregular lesion detection in DR images with better computational efficiency. Traditional models often fail to capture fine-grained features and struggle to generalize across diverse datasets due to high computational complexity. In recent years, some research works have focused on neural network models that replicate the visual processing of the V1 in the human brain, often overlooking the crucial roles of the V2, V4, and IT cortical regions. This narrow focus limits the model’s ability to capture comprehensive visual features, hindering accurate lesion detection in retinal images. Inspired by the hierarchical structure of the primate visual system, the current research addresses these gaps by emulating the visual processing of all crucial V1, V2, V4, and IT cortical regions in its front-end architecture to enhance feature extraction and recognition of complex patterns in retinal images. Additionally, the incorporation of the COP optimization algorithm ensures efficient exploration of the solution space, reducing computational overhead while maintaining model performance. Ultimately, VCANet-COP aims to bridge these gaps in existing methods to enable more accurate and efficient clinical DR diagnosis.

Contributions

The major contributions made by this research work are illustrated below.

Development of VCANet-COP Model

This research work introduces the VCANet-COP model a DL architecture inspired by the biological visual processing system. Unlike traditional methods, VCANet-COP excels in detecting complex lesion patterns by learning hierarchical representations and enhances diagnostic accuracy in diverse clinical scenarios.

Vascular and Lesion-Specific Feature Extraction

The study introduces the use of SAEs to simultaneously extract vascular (vessel structure, thickness, tortuosity) and lesion-specific features (microaneurysms, hemorrhages, exudates) from retinal images. By focusing on pixel-level patterns, SAE effectively distinguishes abnormalities by deriving meaningful representations from complex data.

Hierarchical Visual Processing

The VCANet architecture is divided into a two-part structure. The front-end processing network leverages the hierarchical structure of the primate visual system (V1, V2, V4, IT regions) for feature extraction and a back-end CNN for lesion recognition, providing a more efficient and precise way of recognizing subtle lesion features.

Hyperparameter Optimization

The COP algorithm combines OOA with a crossover strategy to optimize hyperparameters and network configurations and ensure computational efficiency. This method promotes exploration of the solution space, avoids premature convergence, and helps find optimal or superior solutions by diversifying genetic material.

Comprehensive Evaluation

A rigorous evaluation of the VCANet-COP model on various DR datasets and metrics illustrated the supremacy of the model in comparison with existing models to accurately detect lesions in retinal images with better computational efficiency.

Literature Survey

Palaniswamy et al. [10] implemented an Internet of Things (IoT) and DL-enabled DR diagnosis model (IoTDL-DRD) using retinal fundus images. The IoT devices were used by the IoTDL-DRD approach to collect and transfer the data to the cloud server for further processing. Then, the preprocessing of retinal images was performed by noise removal and contrast enhancement approaches. The lesion regions in the fundus images were detected by applying the segmentation approach called Mayfly optimization-based region growing (MFORG). The effective DR diagnosis is carried out by using the long short-term memory (LSTM) classifier. The results revealed that the LSTM-MFORG model surpassed baseline models in DR diagnosis. However, the model’s effectiveness was evaluated with a limited dataset. Islam et al. [11] introduced a supervised contrastive learning (SCL) framework for DR detection and severity grading using fundus images. The model employed a two-stage SCL-based training method with a supervised contrastive loss function. The contrast-limited adaptive histogram equalization (CLAHE) enhanced image quality and a pre-trained Xception CNN with transfer learning was used as the encoder. The model outperformed conventional CNN by obtaining better classification performance. Despite superior feature discrimination and classification, the model suffered from computational complexity.

Bilal et al. [12] developed an AI-driven vision-threatening DR (VTDR) detection framework that integrated multiple models using a majority voting strategy to enhance classification accuracy. The model utilized hybrid CNN-singular value decomposition (SVD) for feature extraction and integrated improved support vector machine-radial basis function (SVM-RBF) with K-nearest neighbor (KNN) for classification. The experimental validation with the IDRiD dataset resulted in higher accuracy, specificity, and sensitivity. Overall, this fusion of models enabled superior robustness and precision in VTDR detection. Despite its superior performance, the VTDR detection model suffered from high computational requirements. To resolve this challenge, Bilal et al. [13] developed a nature-inspired multi-enhanced quantum-self attention capsule network (NIMEQ-SACNet) that significantly improves upon their prior work for the accurate categorization of VTDR. The NIMEQ-SACNet surpassed its prior works in multiple key areas, including model architecture, optimization strategy, classification capability, and clinical application. The model integrated self-attention capsule networks (SACNet) with enhanced quantum-inspired binary grey wolf optimization (EQI-BGWO) for multi-stage VTDR classification (binary, 5-stage, and 7-stage). Unlike the CNN-SVD approach, the NIMEQ-SACNet leveraged self-attention mechanisms which reduce redundancy by focusing only on relevant spatial hierarchies and minimized convolutional operations. Additionally, the hyperparameter tuning by EQI-BGWO diminished the computational overhead and ensured faster convergence. Despite everything, the model’s dependence on high-quality retinal images and potential generalization issues across diverse datasets limited real-world applicability.

Al-Fahdawi et al. [14] introduced a multi-label deep learning classification system (Fundus-DeepNet) for multiple ocular disease detection in fundus images. Initially, the preprocessing steps, including image resizing, noise removal, circular body cropping, data augmentation, and contrast enhancement, are performed. Feature extraction was conducted using the high-resolution network (HRNet) and the discriminative restricted Boltzmann machine (DRBM) performed effective classification. The comprehensive evaluation demonstrated the superior performance of Fundus-DeepNet in detecting multiple ocular diseases effectively. Nonetheless, the model struggles to derive lesion-specific features. Mondal et al. [15] developed an automated ensemble DL approach (EDLDR) for accurate categorization of DR. The EDLDR model for DR combined modified DenseNet101 and ResNeXt to enhance classification accuracy. The ResNeXt employed a split–transform–merge strategy for improved feature extraction and DenseNet improved superior feature reuse efficiency. The model effectively handled class imbalance using generative adversarial network (GAN)-based augmentation and enhanced image contrast with the CLAHE method. It achieved superior classification accuracy and outperformed existing methods. However, it struggled to capture tiny and subtle lesions in complex images.

Xu et al. [16] developed feature fusion U-Net (FFU-Net) a novel model designed to enhance U-Net performance in lesion segmentation. In this model, the pooling layer is replaced with a convolutional layer to minimize spatial loss, and multiscale feature fusion (MSFF) block is integrated into the encoder for efficient multiscale feature learning. A contextual channel attention (CCA) mechanism is applied to fuse relevant information through skip connections and the lower-resolution decoder. Additionally, we address data imbalance and misclassification issues with a balanced focal loss function. The experimental results on the IDRID dataset showed that the FFU-Net outperformed the baseline U-Net by achieving significant margins on key metrics. Despite everything, the FFU-Net model struggles to handle subtle lesions. Kalyani et al. [17] presented a reformed capsule network for DR detection and categorization. The model extracted features using a convolution and primary capsule layer, followed by a class capsule layer and softmax function to determine class probabilities. Particularly, the capsule network improved spatial hierarchies and rotational invariance with its advanced structure to improve classification robustness. The comprehensive experimental evaluation of the network revealed superior performance for stages 1, 2, and 3 of DR. However, higher computational costs and complex training processes remain challenges for large-scale deployment in real-world clinical applications.

Xu et al. [18] presented a prior-guided attention fusion transformer encoder (AFTE) for multi-lesion segmentation in DR that incorporated self-attention for global lesion correlations and cross-attention with vessel masks to improve detection accuracy. Moreover, a shift block was employed to enhance sensitivity for small structures and scale-adaptive attention to adaptively fuse features across different scales. The experimental results demonstrated superior performance compared to baseline methods. While the model excels in multi-lesion segmentation, potential limitations include reliance on high-quality vessel masks and possible computational complexity affected by real-time performance. Qiu et al. [19] designed a novel one-shot anomaly detection framework designed to identify DR from fundus images. By employing a single annotated DR sample, the model synthesized a large number of artificial DR images through the augmentation of DR lesion patches. These synthetic images were generated by pasting lesion patches into normal images using prior knowledge–based strategies. A classification network was trained on these synthetic abnormal images alongside true normal images to detect anomalies. The experimental results showed that the model surpassed leading unsupervised and few-shot methods by achieving performance close to fully supervised methods. However, the model was limited in deriving lesion-specific features.

Luo et al. [20] established a DL framework to effectively capture long-distance dependencies between lesion features (hemorrhages and exudates) in retinal fundus images. The model captured both local and long-range global dependencies in retinal fundus images for DR detection. A residual structure for the long-range block is introduced to analyze the relationship between pixel patches to ensure small and scattered lesion information. The experimental results demonstrated that this approach significantly improved detection accuracy and provided a robust solution for DR detection by considering both local and global image features. However, it achieves poor outcomes with subtle small lesions. Romero-Oraá et al. [21] developed an innovative end-to-end DL model for the automatic grading of DR with five severity levels. The model separated the attention of dark (red) and bright (exudates) lesions in fundus images to allow for optimized classification by independently processing these features. Using the Xception architecture and focal loss to handle data imbalance, the framework generated interpretable attention maps that highlighted microaneurysms, hemorrhages, and hard exudates effectively. The model exhibited better performance but struggled to extract lesion-specific features. Furthermore, the critical analysis of the existing DR detection models against the proposed model is given in Table 1.

Table 1.

Critical analysis of the existing DR detection models against the proposed model

Existing study			Current study
Author	Methodology	Limitations
Palaniswamy et al. [10]	LSTM-MFORG for DR diagnosis	Evaluated with a limited dataset which leads to potential overfitting	The robust optimization, hierarchical visual processing, and large dataset compatibility of the VCANet-COP model ensure better generalization and accuracy across diverse datasets
Islam et al. [11]	SCL with Xception CNN	High computational complexity	The model reduces computational overhead through efficient hyperparameter optimization with the COP algorithm, ensuring faster convergence and less resource consumption
Bilal et al. [12]	Hybrid CNN-SVD model with SVM-RBF and KNN	High computational requirements	VCANet-COP incorporates a hierarchical architecture that boosts computational efficiency and reduces overhead using the COP algorithm
Bilal et al. [13]	NIMEQ-SACNet	Relies on high-quality retinal images and suffers from generalization issues	The hierarchical visual processing (V1, V2, V4, IT) of the model captures complex features, thus improving generalization across diverse datasets and less reliance on high-quality images
Al-Fahdawi et al. [14]	Fundus-DeepNet for multiple ocular diseases	Struggles to detect tiny and subtle lesions in complex images	By focusing on pixel-level patterns, the SAE in the model extracts both vascular and lesion abnormalities for accurate lesion recognition
Mondal et al. [15]	EDLDR with DenseNet101 and ResNeXt	Struggles with small and subtle lesions	VCANet-COP improves fine-grained feature recognition with a biologically inspired hierarchical model that excels at detecting subtle lesions
Xu et al. [16]	FFU-Net	Struggles to handle subtle lesions	The multi-stage hierarchical processing of the model improves the recognition of small and subtle lesions by combining low-level and high-level feature extraction
Kalyani et al. [17]	Capsule network for DR detection	High computational costs and complex training	The model utilizes the COP algorithm to reduce computational overhead and complexity while maintaining accuracy, making it more efficient for large-scale deployment
Xu et al. [18]	AFTE Transformer Encoder for multi-lesion segmentation	Relies on high-quality vessel masks and computational complexity	The model eliminates the reliance on vessel masks and reduces computational demands through efficient hyperparameter tuning and optimized feature extraction
Qiu et al. [19]	One-shot anomaly detection framework for DR	Limited in deriving lesion-specific features	The model employs SAEs to perform lesion-specific feature recognition by deriving meaningful and discriminative features from complex data
Luo et al. [20]	DL framework for long-range dependencies in lesions	Struggles with small, subtle lesions	The model with its multi-stage hierarchical architecture enhances both local and global feature extraction to perform subtle lesion detection
Romero-Oraá et al. [21]	End-to-end DL model for DR grading with severity levels	Struggles to extract lesion-specific features	The robust lesion detection capabilities of the model especially for subtle and small lesions, address the issue of extracting detailed lesion-specific features

Research Gap

The accurate detection and classification of DR lesions from retinal images remains a significant challenge for small, irregular, or subtle lesions from DR images. Traditional CNNs struggle to capture lesion-specific or vascular features and often fail to generalize well across varied datasets due to high execution time, computational complexity, and inefficient hyperparameter optimization.

Inefficient Feature Representation

Conventional CNNs struggle with fine-scale lesion recognition, especially in small or complex lesions. The VCANet model addresses this by incorporating a hierarchical architecture that mimics the visual processing in primate brains (V1, V2, V4, IT regions). This allows the model to progressively refine features from low-level to high-level lesion representations, improving the detection of subtle and irregular lesions.

Computational Complexity

Many existing models suffer from high computational overhead, which makes them impractical for real-time applications. By integrating the COP algorithm with VCANet, the VCANet-COP model accelerates the search for optimal hyperparameters, thus balancing computational efficiency and recognition performance.

Execution Time

Traditional approaches often require excessive time for training and evaluation which limits their clinical utility. The VCANet-COP addresses this problem by optimizing the hyperparameter search process using the COA algorithm which accelerates convergence by eliminating redundant evaluations.

Overall, the VCANet-COP model effectively addresses the critical gaps in DR detection by combining a biologically inspired architecture with an optimization algorithm. This fusion not only enhances the accuracy of lesion classification in subtle and irregular lesion regions but also maintains the computational efficiency of the model.

Proposed Methodology

This study primarily aims to enhance DR recognition using a DL approach inspired by a bionic vision system. Figure 1 illustrates the procedural overview of the proposed algorithm. Initially, data collection is performed to gather retinal images, followed by preprocessing steps to prepare the data for effective feature extraction and accurate lesion recognition. The preprocessed data are fed to SAEs which learn vascular and lesion-specific features from retinal images by focusing on relevant pixel-level patterns to highlight color contrasts, shape, and intensity for accurate lesion recognition. Following that, the extracted features are fed to the novel VCANet-COP model for lesion recognition. The VCANet-COP model leverages the power of VCANet combined with an OOA to enhance lesion recognition in DR images. The VCANet architecture is structured in two parts, namely a front-end processing network inspired by the hierarchical structure of the primate visual system (V1, V2, V4, IT cortical regions) and a back-end CNN for lesion recognition. The V1, V2, V4, and IT modules progressively process visual information at different levels starting from low-level feature extraction (edges and orientations) and refining to high-level lesion recognition (shape, size, and context). This approach allows VCANet to focus on subtle lesion features and improve robustness, particularly for small or irregular lesions. Moreover, the development of the COP algorithm by combining OOA with crossover strategy employs genetic operations, including selection, crossover, and mutation to optimize hyperparameters and network configurations and maintain computational efficiency. This enables effective exploration of the solution space to prevent premature convergence and encourage diversity in the search for optimal solutions.

Fig. 1.

Workflow of the proposed framework

Data Pre-processing

The following pre-processing steps are applied to those five datasets for improving the quality of the data in lesion detection and its visualization is provided in Fig. 2.

Fig. 2.

Visualization of the pre-processing outcomes: a input, b pre-processed output

Image Normalization

This process scales the pixel values to a consistent range typically of [0, 1] and helps in faster convergence during training.

Contrast Enhancement

It is used to enhance important details without amplifying noise, making subtle features more distinguishable in retinal images.

Noise Removal

Bilateral filtering is used for noise removal that helps retain edge details and is effective in removing noise without blurring edges.

Data Augmentation

It is employed to address class imbalance in this research by artificially increasing the diversity of the training data. Typically, the number of images in the common classes (healthy images or mild lesions) is higher but with lesser images for severe lesions. This imbalance causes the model to perform well in majority classes but struggle with minority ones, causing biased and inaccurate predictions. Hence, techniques like rotations, flipping, zooming, and brightness adjustments are performed to create a more balanced dataset and simulate real-world variations. This enhances the model’s robustness and generalization in lesion detection.

Image Resizing

It is applied to ensure that all input images have the same dimensions of 224 × 224 pixels. This process helps optimize memory usage and speeds up training by providing uniform input sizes across the dataset.

To perform robustness analysis, destructive operations such as adding random noise, Gaussian blurring, splash masking, and oversaturation are applied to test images as given in Fig. 3. These operations simulate real-world conditions where noise and variations in image quality are common.

Fig. 3.

Artificial image distortions for robustness evaluation: A original image, B image with point noise, C Gaussian blurred image, D splash-masked image, and E over-saturated image

Added Point Noise (Salt-and-Pepper Noise)

It involves randomly setting the image’s pixels to either the maximum (salt) or the minimum (pepper) pixel values, resulting in black and white spots scattered across the image. The noisy pixel value $I_{\mathit{noisy}}\left(y,,,z\right)$ is then calculated as:

$$I_{\mathit{noisy}}\left(y,,,z\right)I\left(y,,,z\right)+\eta \left(y,,,z\right)$$ 1

In Eq. (1), $I\left(y,,,z\right)$ indicates the original pixel value within the coordinates $\left(y,,,z\right)$, and $\eta \left(y,,,z\right)$ represents the added noise which is either a “salt’ (maximum intensity) or “pepper” (minimum intensity) value.

Gaussian Blurred Image

It is used to smooth the image by reducing fine details and noise while preserving the overall structure of the image. To achieve Gaussian blurring, a Gaussian kernel $G_b\left(y,,,z\right)$ with a standard deviation σ is convolved with the original image where y and z are the coordinates of the kernel relative to the pixel being processed. It is defined as:

$$G_b\left(y,,,z\right)=\frac{1}{{2\pi \sigma }^2}exp\left(-,\frac{y^2+z^2}{{2\sigma }^2}\right)$$ 2

The blurred image $I_B\left(y,,,z\right)$ is then obtained by convolving the original image $I\left(y,,,z\right)$ with the Gaussian kernel, which is expressed as:

$$I_B\left(y,,,z\right)={\sum }_{y\prime ,z\prime }I\left(y,,,z\right)G_b\left(y,-,y,\prime ,,,z,-,z,\prime \right)$$ 3

In Eq. (3), y’ and z’ refer to the coordinates of the surrounding pixels.

Splash-Masked Image

The goal is to retain color in specific regions and convert the rest to grayscale. The Splash Masked Image $\left(I_m,\left(y,,,z\right)\right)$ is represented as follows:

$$l_m\left(y,,,z\right)=\left\{\begin{array}{cc}I\left(y,,,z\right)\hfill & ifN\left(y,,,z\right)=1\hfill \\ Grayscale\left(I,\left(y,,,z\right)\right)\hfill & ifN\left(y,,,z\right)=0\hfill \end{array}\right)$$ 4

In Eq. (4), $Grayscale\left(I,\left(y,,,z\right)\right)$ refers to the grayscale version of the original image. The term $N\left(y,,,z\right)$ refers to the binary mask where $N\left(y,,,z\right)=1$ for regions to retain color (lesions) and $N\left(y,,,z\right)=0$ for regions to convert to grayscale.

Over-Saturated Image

It enhances color vibrancy by increasing the saturation of each pixel in the Hue, Saturation, and Value (HSV) color space. The over-saturated image $I_s\left(y,,,z\right)$ is represented as:

$$I_s\left(y,,,z\right)=HSV\left(H,u,e,\left(y,,,z\right),,,m,i,n,\left(S,a,t,u,r,a,t,i,o,n,\left(y,,,z\right),\times ,l,,,1\right),,,v,a,l,u,e,\left(y,,,z\right)\right)$$ 5

In Eq. (5), $\text{l}$ refers to the saturation scaling factor ( k > 1 increases saturation intensity).

Feature Extraction Using Sparse Autoencoder (SAE)

Unlike traditional autoencoders that aim for faithful reconstruction, SAE prioritizes activating only a subset of neurons in the hidden layer during encoding. This selective activation helps focus on the most relevant features, making the feature representation more efficient and accurate. In this research work, the SAE is employed to learn vascular features (vessel structure, thickness, tortuosity, and branching) and lesion-specific features (microaneurysms, hemorrhages, and exudates) simultaneously from retinal images. By focusing on pixel-level patterns, the SAE extracts both types of features distinguishing vascular and lesion abnormalities for accurate lesion recognition. It employs sparse representations to learn and extract lesion-specific and vascular features from retinal images by deriving meaningful and discriminative features from complex data. In the case of microaneurysms, the SAE is trained to recognize these features by focusing on small, distinct patterns of color (bright red) that appear against the darker background of the retina. Similarly, it derived darker irregular regions (hemorrhages) and yellowish spots (exudates) through unsupervised learning by prioritizing color contrasts, shape, and intensity while discarding irrelevant noise in the input images. Moreover, the SAE efficiently extracts both blood vessel morphology features by recognizing variations in thickness and branching patterns and analyzing the spatial relations between pixels along the vessels efficiently.

During training, SAEs optimize both the reconstruction fidelity of input images and the sparsity of hidden layer activations. This dual objective ensures that the learned representations are both compact and informative, facilitating efficient processing and interpretation by medical professionals. The unsupervised learning nature of sparse autoencoders allows them to derive features directly from unlabeled medical image datasets, mitigating the need for extensive manual annotation which is labor-intensive and expensive. Figure 4 represents the structure of the SAE network. In SAE, the number of decoding and encoding layers is equal. For SAE networks, a symmetric configuration is required. The processing unit of both the decoder and encoder is $N$ neurons, the centralized layer has $L$ neurons that are utilized for the characterization of structural signal $y$ that consists of $N$ data points. The input variables are transformed into new variables with low dimensions, represented by $L<<N$. It is necessary to observe the feature vector $i$, utilized in assessing the appropriate and reliable data [22].

Fig. 4.

Structure of SAE network

Lesion Recognition Using VCANet-COP

The VCANet-COP model is designed by integrating a modified CNN architecture with an OOP to enhance lesion recognition in DR images. The VCANet architecture is used to perform lesion recognition by mimicking the primate visual system. Moreover, the COP algorithm is used to optimize the hyperparameters of the VCANet model through genetic operations, thus maintaining computational efficiency. The detailed explanation and working of the proposed approach are provided below:

VCANet Model

The proposed VCANet is a novel modified CNN model designed to improve lesion recognition in DR images by emulating the hierarchical structure of the primate visual system which is provided in Fig. 5. Unlike traditional CNNs, which often struggle with subtle lesion features, VCANet integrates a front-end architecture that simulates the visual processing of the V1, V2, V4, and IT cortical regions, followed by a CNN back-end network. The V1 module processes low-level features such as edges and orientation through a VVBlock which is inspired by VVNet [23] through the incorporation of a Gabor filter bank, simple and complex cell nonlinearities, and a stochasticity generator. The V2 module refines these features and the V4 focuses on the color, size, and structural information of lesions which is vital for lesions in varied locations, or at different magnifications in DR detection. The IT module supports high-level lesion recognition and memory by providing invariant representations crucial for accurate diagnosis even in noisy or corrupted images. Overall, this approach enhances lesion recognition accuracy and model robustness for small-sized lesions. The simulation modeling of V1, V2, V4, and IT are detailed below:

Fig. 5.

Architecture of the VCANet model

Primary Visual Cortex (V1)

Inspired by VVNet, the V1 module is designed in this paper. The VVNet consists of two parts, namely the VVBlock and a modified CNN back-end network. In our design, the first layer of the CNN (which is usually a stack of convolutional, normalization, nonlinear, and pooling layers) is replaced with V1 which comprises four layers, namely VV block and transition layer (layer 1), InceptionE-like module (layer 2), structural composite layer (layer 3), and feature refinement layer (layer 4). As illustrated in Fig. 6, the VVBlock includes multiple nonlinear layers that simulate image information processing by simple and complex cells in the V1 region of the primate visual system. The VVBlock maintains the spatial map dimensions of the replaced block but has more channels. Hence, it is followed by a transition layer (1 × 1 conv) that acts as a bottleneck to reduce the number of channels and match the depth of the original block. Inspired by the linear-nonlinear-Poisson (LNP) model of V1, the VVBlock consists of three consecutive processing stages namely convolution, nonlinearity, and a stochasticity generator. Each channel contains two distinct types of neurons (simple and complex cells) where each having a specific number of units per spatial location. The layers of the VV block are described below:

Fig. 6.

Architecture of V1

Convolutional Layer (Biologically Constrained Gabor Filter Banks (GFB))

The first layer of the VVBlock is a GFB which is parameterized to approximate the response of neurons in the primate V1 region. It uses multi-directional, multi-scale, and multi-spatial frequency Gabor filters to process the RGB input image. Here, the span is set to 4 to create a 56 × 56 activation map. Since the number of channels in the first convolution of most CNNs is relatively small, we increased the number of channels in the VVBlock that allows the Gabors to cover a larger parameter space and more accurately approximate primate V1. Hence, the VV block is configured with 512 channels by equally splitting between simple and complex cells. Moreover, each channel in the GFB processes an individual color channel from the input image. The Gabor function consists of a two-dimensional grating with a Gaussian envelope defined as:

$$G_{\theta ,g\mathrm{\Phi },n_a{n}_b}\left(a,,,b\right)=\frac{1}{{2\pi \sigma }_a{\sigma }_b}exp\left[-,0.5,\left(a_{\mathit{rot}}^2,/,{\sigma }_a^2,+,b_{\mathit{rot}}^2,/,{\sigma }_b^2\right)\right]\text{cos}\left(2,\pi ,g,+,\varphi \right)$$ 6

$$\begin{array}{cc}{a}_{\mathit{rot}}=a\text{cos}\left(\theta \right)+b\text{sin}\left(\theta \right)& b_{\mathit{rot}}=a\text{sin}\left(\theta \right)+b\text{cos}\left(\theta \right)\end{array}$$ 7

$$\begin{array}{cc}{\sigma }_a=\frac{n_a}{g}& {\sigma }_y=\frac{n_y}{g}\end{array}$$ 8

In the above equation, a_rot and a_rot represent the orthogonal and parallel orientations to the grating, respectively. The term $\theta$ is the angle of the grating orientation, g is the spatial frequency of the grating, and φ is the phase of the grating with respect to the Gaussian envelope. The terms ${\sigma }_a$ and ${\sigma }_b$ are the standard deviation of the Gaussian envelope correspond to the orthogonal and parallel directions to the grating, respectively, and are defined as multiples (na and nb) of the grating period (the inverse of the frequency). Even though, the GFB reduces the number of parameters needed for the V1 neuron’s linear spatial components, five parameters per channel are still required. Using neurophysiological data from primate V1, we constrained these parameters by sampling from an empirically defined distribution, with spatial frequency (g < 5.6 cpd) and number of cycles (n > 0.1) based on the input image resolution.

Non-linear Layer (Simple and Complex Cells)

The VVBlock contains two distinct nonlinear layers where each layer corresponds to one of the two cell types and is applied to the respective channels. This design simulates the rectified linear transformations of simple cells and the spectral power of orthogonal phase pairs of complex cells.

$$S_{\theta ,f,\varphi ,n_x,n_y}^{n1}=\left\{\begin{array}{cc}{S}_{\theta ,f,\varphi ,n_x,n_y}^l& if{S}_{\theta ,f,\varphi ,n_x,n_y}^l>0\\ 0,& \mathit{otherwise}\end{array}\right\}$$ 9

$$C_{\theta ,f,\varphi ,n_x,n_y}^{n1}=\frac{1}{\sqrt{2}}\sqrt{{\left(C_{\theta ,f,\varphi ,n_x,n_y}^l\right)}^2+{\left(C_{\theta ,f,\varphi ,n_x,n_y}^l\right)}^2}$$ 10

In Eqs. (8) and (9), S^l and S^nl represent the linear and nonlinear responses of a simple neuron, respectively. The term C^l and C^nl represent the linear and nonlinear responses of a complex neuron, respectively.

Stochasticity Layer and Bottleneck Layer

Neuronal responses exhibit inherent randomness, with spike sequences differing even when the same visual stimulus is presented to a neuron multiple times. To capture this characteristic, independent Gaussian noise was added to each cell in the VVBlock, with the variance of the noise set equal to the activation of the corresponding unit. Additionally, to closely replicate the neuronal properties of the primate V1 region, the GFB model was designed using an existing empirical distribution. After the nonlinear layer, a 1 × 1 convolutional bottleneck layer was introduced.

Following the VV block, the InceptionE-like module [24] enables multi-scale feature extraction through four parallel processing paths: (1) a 1 × 1 Conv-Batch normalization (BN) -LReLU block for capturing fine-grained details, (2) a juxtaposed convolutional path that sequentially applies 1 × 1, 1 × 3, and 3 × 1 convolutional kernels to extract directional features, (3) a hybrid convolutional path that combines 1 × 1 and 3 × 3 convolutions, followed by 1 × 3 and 3 × 1 filters to enhance complex feature representations, and (4) a pooling-enhanced path integrating an average pooling layer with a 1 × 1 Conv-BN-LReLU block utilized to retain spatial information while reducing redundancy. This structured architecture effectively captures both localized and global retinal features from the input images. Layer 3 incorporates a 3 × 3 convolutional kernel, BN, and LeakyReLU activation to enhance feature learning and improve spatial representation. Layer 4 employs a 1 × 1 convolutional kernel to reduce dimensionality and preserve essential spatial details.

V2

The V2 module is used to refine and integrate the basic visual features processed by V1. It handles a broader range of attributes, including orientation, spatial frequency, color, and texture. It is structured as a modular system with parallel layers where each layer is specialized to process distinct visual features. Moreover, V2 enhances feature representation by combining basic visual inputs from V1 and adding spatial and contextual relationships between them. This integration enables the model to recognize more complex patterns, such as object parts and textures. Ultimately, the V2 module prepares the visual data for higher-level processing in modules like V4 and IT. Its architecture is provided in Fig. 7.

Fig. 7.

Architecture of V2

V4

The V4 module is used to recognize high-level features such as color, texture, and complex shapes from the input DR images. This module as given in Fig. 8 incorporates a convolutional layer, channel-wise attention, spatial attention, and global pooling mechanisms to efficiently process and highlight the most relevant features. It begins by applying convolutions with larger kernels to extract intricate patterns and textures from the output of the V2 module. Following this, a channel-wise attention mechanism (squeeze-and-excitation block) recalibrates the importance of each channel in the feature map to emphasize the most relevant features.

$$z_c=\frac{1}{H_{}^{\ast }W}\sum _{i=1}^H\sum _{J=1}^W{y}_{\mathit{ijc}}$$ 11

Fig. 8.

Architecture of V4

In Eq. (11), $y_{\mathit{ijc}}$ represents the value of the feature map at the spatial location (i,j) for the c^th channel, and z_c is the scalar output for the channel c representing its global information (squeezed feature). After the squeeze operation, the next step is to apply a fully connected layer to generate channel-wise attention scores. Moreover, the channel-wise attention vector $\left(s_c,\in ,R^C\right)$ is expressed as:

$$s_c=\sigma \left(W_2,\bullet ,R,e,L,U,\left(W_1,\bullet ,z_c,+,b_1\right),+,b_2\right)$$ 12

In Eq. (12), $W_1\in R^{C/r\times C}$ and $W_1\in R^{C\times C/r}$ are weight matrices with r being the reduction ratio (typically r = 16). The terms b₁ and b₂ are bias terms, and σ denotes the sigmoid activation function which ensures the attention scores are between 0 and 1. The attention scores (s_c) are then multiplied with the feature map y to reweight the channels and enhance the importance of relevant features.

A spatial attention mechanism is then used to focus on important spatial regions of the image. Here, the input feature map $Y\in R_{}^{H\ast W\ast C}$ is processed by applying a 1 × 1 convolution with a learnable weight matrix $W_{\mathit{spatial}}\in R_{}^{1\ast 1\ast C}$, followed by a sigmoid activation function to generate a spatial attention map $A_{\mathit{spatial}}\in R_{}^{H\ast W\ast 1}$. The attention map assigns weights to different spatial locations, and these weights are applied to the original feature map Y through element-wise multiplication. It is expressed as:

$$Y\prime =A_{\mathit{spatial}}\bullet Y$$ 13

In Eq. (13), Y’ represents the resulting output feature map that emphasizes relevant spatial regions and suppresses less important ones. Finally, global average pooling is used to reduce the spatial dimensions. This enables the V4 module to enhance the recognition of complex visual attributes, improving lesion detection in DR.

Inferiotemporal Cortex (IT)

As shown in Fig. 9, the IT region is modeled as a complex layer-block structure that receives three independent inputs from V1, V2, and V4. After processing through customized convolutional blocks, the V1 and V2 outputs are merged and then combined with the processed V4 output to form the final Vm result. Following this, the processed V4 output passes through an extended convolutional block and merges with Vm. The resulting intermediate data is fed into the IT bottleneck section which includes four conv-BN-ReLU blocks each having specific kernel sizes, step sizes, and dilations corresponding to the six transformations of inputs V1, V2, V4, and Vm. Following that, four upsampling blocks are used to restore feature maps to their original resolution and output predictions, realizing jump connections between stages. Consequently, the kernel size of the IT module is larger than in the preceding regions.

Fig. 9.

Architecture of IT

Crossover Osprey Optimization (COP) Algorithm

The COP algorithm is designed through the combination of OOA with a crossover strategy to optimize the VCANet architecture for improved lesion detection and classification in retinal images. Its flow chart is given in Fig. 10. By initializing a diverse population of candidate solutions each with unique hyperparameters and configurations, the algorithm iteratively refines these solutions through genetic operations, namely selection, crossover, and mutation. This process ensures that only high-performing solutions are passed to the subsequent, promoting exploration and preventing premature convergence. The incorporation of genetic optimization not only enhances classification accuracy but also significantly boosts computational efficiency by minimizing the need for exhaustive searches.

Fig. 10.

The flowchart of the COP algorithm for DR lesion recognition

Traditional Osprey Optimization Algorithm (OOA)

The OOA [25] approach and its mathematical model are presented in this section.

Initialization

The OOA is a population-based method where each osprey representing a candidate solution determines problem variables based on its position in the search space. The population is modeled as a matrix, and initial positions are randomly set. The algorithm iteratively updates the positions of the ospreys to find optimal solutions. It is expressed as follows:

14

$$y_{j,k}=l{b}_k+s_{j,k}.\left(u,b_k,-,l,b_k\right),k=1,2,....,n$$ 15

From the above equations, the population matrix of osprey position is represented by $Y$ where each $j\text{th}$ osprey is denoted by $Y_j$ having $j\text{th}$ dimension is denoted by $y_{j,k}$. The total number of osprey is denoted by $M$, and the number of problem variables is represented by $n$. The random numbers are denoted by $s_{j,k}$ which ranges from $0to1$. The upper and lower bounds of the $i\text{th}$ problem variable are represented by $u{b}_k$ and $l{b}_k$, respectively. The estimated values for the objective function are calculated using the following vector-based equation:

$$G={\left[\begin{array}{c}{G}_1\\ ⋮\\ G_j\\ ⋮\\ G_M\end{array}\right]}_{M\times 1}={\left[\begin{array}{c}G\left(Y_1\right)\\ ⋮\\ G\left(Y_j\right)\\ ⋮\\ G\left(Y_M\right)\end{array}\right]}_{M\times 1}$$ 16

In Eq. (16), the vector of objective function value is represented by $G$, and the determined objective function value for $j\text{th}$ osprey is represented by $G_j$.

Stage 1 (Exploration Phase)

Inspired by the hunting behavior of ospreys, the OOA population update is modeled based on their ability to detect fish in the water. The mathematical formulation for this set of fish for each osprey can be expressed as:

$$G{Q}_j=\left\{Y_l,|l\in ,\left\{1,,,2,,,.,.,.,.,,,M\right\},\wedge ,G_l,<,G_j\right\}\cup \left\{Y_{\mathit{best}}\right\}$$ 17

In Eq. (17), the set of fish positions with respect to $j\text{th}$ osprey is represented by $G{Q}_j$, and the best osprey is represented by $Y_{\mathit{best}}$.

Stage 2 (Exploitation Phase)

In OOA, the osprey’s behavior of carrying fish to an optimal location is simulated to update the population. This process refines positions in the search space to enhance local exploitation and convergence. A new position is determined as:

$$y_{j,k}^2=y_{j,k}+\frac{l{b}_k+s.\left(u,b_k,-,l,b_k\right)}{u},j=1,2,....,n,j=1,2,...,U$$ 18

$$y_{j,k}^2=\left\{\begin{array}{c}{y}_{j,k}^2,l{b}_k\le y_{j,k}^2\le u{b}_{k}l;\\ l{b}_k,y_{j,k}^2<l{b}_k;\\ u{b}_k,y_{j,k}^2>u{b}_k;\end{array}\right)$$ 19

$$Y_j=\left\{\begin{array}{cc}{Y}_j^2,G_j^2& <G_j\\ Y_j,& else;\end{array}\right\}$$ 20

From the above equations, the new position of the $j\text{th}$ osprey is represented by $y_{j,k}^2$ on the basis of the second phase. The objective function value is represented by$G_j^2$, and the $j\text{th}$ dimension is represented by $y_{j,k}^2$. Total number of iterations is denoted by$U$, and iteration counter of a model is represented by$u$.

Stage 3

The OOA follows an iterative process to update the osprey positions and refine the best candidate solution by evaluating the objective function. This cycle continues until the final iteration, ensuring that the algorithm efficiently converges toward an optimal solution.

Cross-Over Strategy

The hybridization of two parent solutions is performed using a crossover operator to generate an offspring that integrates features from both parents [26]. It is mathematically expressed as follows:

$$x_t=\left\{\begin{array}{ccc}{y}_a^t& \mathit{if}& f_a\le F_R\\ e_t^a& & \mathit{otherwise}\end{array}\right)$$ 21

In Eq. (22), the crossover rate is signified by $F_R$; the two parent solutions are denoted by the $e_t^a$ and $y_a^t$. Hence, the probability of crossover is $0.3$. In this research work, the OOA algorithm employs a crossover strategy that recombines traits from parent solutions to diversify the genetic material, thereby enhancing search efficiency and reducing the risk of premature convergence.

Practical Implications

The COP algorithm diversifies genetic material through crossover and introduces random mutations to avoid premature convergence leading to suboptimal solutions, enabling a broader exploration of the design space. This strategy optimizes critical VCANet parameters such as layer depth, convolutional filter sizes, activation functions, dropout rates, and learning rates to ensure superior lesion recognition performance. Ultimately, the integration of crossover with OOP allows the algorithm to exploit promising design configurations, overcome local optima, and converge toward solutions that achieve or exceed defined performance criteria. In practice, how the COP algorithm optimizes the VCANet parameters is provided below:

Diversification

By combining elements of successful parent configurations (different filter sizes, activation strategies, and dropout rates), the crossover mechanism generates novel offspring architectures that yield improved lesion recognition. This diversification helps VCANet explore underutilized network configurations, thus enhancing feature extraction and classification accuracy in retinal image analysis.

Avoiding Premature Convergence

The COP algorithm avoids premature convergence by introducing random perturbations in hyperparameters (learning rate adjustments and dropout modifications) to ensure a dynamic search process mechanism. As a result, the VCANet improves its adaptability to varying image resolutions, lighting conditions, and lesion complexities.

Escaping Local Optimal Solution

The crossover mechanism in COP helps VCANet to escape local optima by balancing network depth, complexity, and regularization techniques. This strategic refinement enables the model to identify robust feature representations of diabetic lesions across multiple datasets.

Tournament Selection for Robust Feature Learning

The COP algorithm employs a tournament selection process where multiple candidate VCANet configurations compete based on their classification accuracy. The best-performing models (detecting microaneurysms, hemorrhages, and exudates with higher accuracy) are chosen as parents to generate improved offspring architectures. This evolutionary selection ensures that VCANet continuously learns optimal feature extraction techniques.

Iterative Convergence for Optimized Lesion Detection

Here, the COP algorithm fine-tunes hyperparameters of the VCANet model by iteratively selecting, crossing over, mutating, and refining model configurations through multiple optimization cycles. Over time, the algorithm converges to an architecture that maximizes lesion recognition accuracy, reducing false positives and false negatives.

Handling Computational Efficiency

The COP algorithm enhances computational efficiency through crossover and mutation by exploring hyperparameter configurations and avoiding exhaustive searches. This enables effective optimization while minimizing resource use.

VCANet-COP Algorithm

The VCANet-COP model employs the COP algorithm to enhance the capability of VCANet in accurately detecting retinal lesions. It begins by initializing a diverse population of candidate solutions where each solution represents a unique set of hyperparameters and architectural choices. The algorithm evaluates each candidate’s fitness using detection accuracy, incorporating the EMLLoss function with the binary cross-entropy loss to enhance optimization. Through an iterative process, the algorithm refines solutions using genetic operations, namely selection, crossover, and mutation. The selection prioritizes high-performing solutions which ensure that the best traits are passed to the next generation. The crossover enables the combination of beneficial attributes from different parent configurations to promote exploration within the solution space. Moreover, the mutation introduces random variations, preventing premature convergence and enhancing diversity. To ensure convergence, the algorithm monitors performance improvement over successive generations. The termination criteria include a predefined number of generations, fitness score stabilization, or reaching an optimal performance threshold. Ultimately, the algorithm returns the best-performing VCANet-COP configuration, maximizing lesion recognition accuracy and robustness for clinical applications.

Initialization

Begin by initializing a population of candidate solutions. Each solution represents a configuration of VCANet-COP, which includes parameters such as neural network architecture (number of layers, types of layers), activation functions, learning rates, and other hyperparameters.

Fitness Evaluation

Define a fitness function that quantitatively measures the performance of each candidate solution in lesion recognition from retinal fundus images. This can be based on detection accuracy derived from validation or test datasets. In this paper, we utilize the EML Loss function [27] with the binary cross-entropy loss to enhance optimization.

$$DSC=\frac{2TP}{2TP+FP+FN}$$ 22

$$Loss=EMLLoss+BCELsoss$$ 23

$$Loss=\frac{-{\alpha }_t{\left(1,-,p_t\right)}^{\gamma }\text{log}\left(p_t\right)}{N}+\text{log}\left(1,-,D,S,C\right)+BCELoss$$ 24

Parent Selection

Use a selection mechanism (tournament selection) to choose pairs of parent solutions from the population based on their fitness scores. Solutions with higher fitness values are more likely to be selected for reproduction, ensuring that well-adapted configurations have a higher probability of contributing to the next generation.

Crossover Operation

Perform crossover between selected parent solutions to generate new offspring solutions. The crossover operation involves selecting random crossover points within the solutions and combining segments from each parent to create new configurations of VCANet-COP.

Inheritance of Traits

Through crossover, offspring inherit beneficial traits from both parents, potentially enhancing their ability to extract and classify features relevant to retinal lesions.

Mutation

Apply mutation operators to introduce random changes into offspring solutions. This step ensures the exploration of new solution spaces beyond those explored by crossover. Mutation involves perturbing parameter values or introducing new architectural features.

Evaluation of Offspring

Assess the fitness of offspring solutions based on the defined evaluation metric. Evaluate how well each offspring performs in recognizing retinal lesions compared to the original parent solutions.

Replacement

Replace less fit solutions in the current population with the newly generated offspring solutions. This step ensures that the population evolves over successive generations toward configurations that better recognize and classify retinal lesions.

Termination Condition

Repeat the above steps iteratively until a termination condition is met. This could be a maximum number of generations, convergence of fitness scores, or achievement of a desired level of performance improvement.

Optimized Solution

Once the termination condition is satisfied, the algorithm returns the best-performing solution found. This solution represents an optimized configuration of the VCANet-COP for lesion recognition on retinal fundus images, maximizing accuracy and effectiveness in clinical applications. The pseudocode of the VCANet-COP Algorithm is given Algorithm 1.

Algorithm 1.

Pseudocode of the VCANet-COP algorithm

Experimental Results and Discussions

The lesion recognition is performed with the novel VCANet-COP method comprising various stages. The effectiveness of the VCANet-COP approach is evaluated with various performance measures. Five datasets are chosen to obtain the retinal images for the verification process such as the DRIVE dataset, STARE dataset, DR-Data, IDRiD, and RFMID datasets. The experimental results of the VCANet-COP method are compared with existing models such as EDLDR [15] CNN_SVD [12], FFU_Net [13], LSTM_MFORG [10], and fundus-DeepNet [14]. The results are described deeply below.

Experimental Setup

The proposed VCANet-COP for recognizing the lesions is implemented using Python and Keras framework. The experiments are conducted by utilizing the HD Computer specified with Intel i5 3rd Gen Processor, 8 GB RAM, and 3.2 GHz CPU speed.

Hyper Parameter Configuration

This sub-section provides the key hyperparameters of the VCANet-COP model which is depicted in Table 2.

Table 2.

Hyperparameters of the VCANet-COP methods and their values

Hyperparameters	Ranges
Learning rate	0.01
Batch size	50
Epochs	500
Drop out value	0.1
Hidden size	156
Activation function	ReLU
Attention dropout rate	0.1
Weight decay	0.01

Dataset Description

DR-Data

It includes 42,545 high-resolution images which are divided into three parts such as training, testing, and validation parts. Each part has mild, moderate, no DR, proliferate DR, and severe classes (https://www.kaggle.com/datasets/medqsdqsdqsdqszd/diabetic-retinopathy-data). In testing, each class contains 10 images, and the validation part includes 800 images in each class. In the training part, the mild class includes 7660 images, the moderate includes 7722 images, no-DR and Proliferate DR classes include 7603 and 7733 images each, and the severe class constitutes 7777 images.

Stare_dataset

It is established to diagnose the diseases associated with the human eyes automatically. It comprises 800 images where the training part includes 760 images and the testing part comprises 40 images each (https://www.kaggle.com/datasets/vidheeshnacode/stare-dataset?select=im0001.ppm).

IDRiD Dataset

This dataset comprises pixel-level labeled general DR lesion images and normal retinal structures (https://www.kaggle.com/datasets/aryaadithyan/idrid-dataset). It provides detailed data on the DR severity and the diabetic macular edema as shown in Table 3. It accounts for 519 images distributed in three splits of training testing and validation. Each split comprises five classes based on DR severity namely, Class A (No DR), Class B (Mild DR), Class C & D (Moderate and Severe DR), and Class E (Proliferative DR).

Table 3.

Images belong to the individual classes of the IDRiD dataset

Class	A (no DR)	B (mild DR)	C (moderate DR)	D (severe DR)	E (proliferative DR)
Training	107	16	108	59	39
Testing	34	5	32	19	13
Validation	27	4	28	15	10

DRIVE Dataset

It is a well-known dataset used for retinal vessel segmentation tasks (https://www.kaggle.com/datasets/yurekanramasamy/drive-dataset). Of the total 340 images, 220 images are included in the training set and the rest of the 120 images belong to the testing set.

RFMID

The RFMID dataset is a publicly available collection designed for the automatic categorization of ocular diseases, including both common and rare pathologies (https://ieee-dataport.org/open-access/retinal-fundus-multi-disease-image-dataset-rfmid). It contains a total of 3200 retinal fundus images captured using three different fundus cameras across 46 scenarios, and these images have been labeled by two senior scholars with an adjudicated consensus. The dataset is split into 1920 training images and 640 images each for testing and validation. The distribution of retinal fundus images across various datasets is detailed in Table 4.

Table 4.

Distribution of datasets selected for the evaluation process

Dataset	Total images	Training	Testing	Validation
DR-Data	42,545	38,495	50	4000
Stare_dataset	800	760	40	-
IDRiD dataset	519	329	103	86
DRIVE dataset	340	220	120	-
RFMID	3200	1920	640	640

Evaluation Measures

To evaluate the performance of the proposed VCANet-COP model, certain experiments are conducted. The metrics including, kappa coefficient, execution time, sensitivity, accuracy, specificity, area under the receiver operating characteristic curve (AUC-ROC), precision, area under the precision-recall curve (AUC-PR), and F1-score are employed.

Execution Time

It refers to the amount of time it takes for the VCANet-COP model to process the input retinal images, detect lesions, and output the final results.

False Positive Rate (FPR)

It refers to the proportion of non-lesion pixels (true negatives) that are incorrectly identified as lesions (false positives).

False Negative Rate (FNR)

It refers to the proportion of actual lesions (true positives) that are incorrectly identified as healthy areas (false negatives).

Kappa Coefficient

It is also known as Cohen’s kappa which is used to evaluate the agreement between the predicted and true class labels:

$$Kappa=\frac{P_{\mathrm{O}\mathrm{A}}-P_{\mathrm{E}C}}{1-P_{\mathrm{E}C}}$$ 25

In Eq. (25), $P_{\mathrm{O}\mathrm{A}}$ is the observed agreement which is the proportion of times the predicted class labels match the true class labels, and $P_{\mathrm{E}C}$ represents the expected agreement by chance which is calculated based on the frequency of the class labels in both the predicted and true values.

Performance Evaluation

The confusion matrix of the VCANet-COP model across five datasets is given in Fig. 11a–e. A confusion matrix is a vital tool to measure the classification performance of the model by comparing the prediction made by the model with the actual labels. It summarizes the four components such as TP, FP, TN, and FN. All five figures indicated that the prediction rate of the VCANet-COP model is closely compatible with the ground truths. The values presented in the confusion matrices are normalized percentages, where each decimal value represents the proportion of correctly or incorrectly classified instances within a given class. This normalization ensures a consistent comparison across datasets, particularly when dealing with imbalanced class distributions. Moreover, the figures confirmed that the model is performing with high accuracy and low error rates, thus confirming its reliability and effectiveness in lesion detection across various datasets.

Fig. 11.

Confusion matrix of a RFMID dataset, b DRIVE dataset, c STARE dataset, d IDRiD dataset, e DR_Data

Figure 12 illustrates the recognition results across various datasets to illustrate model performance in lesion detection. The visualization results demonstrate the model’s ability to accurately detect and classify lesions, highlighting its performance and effectiveness across diverse retinal image datasets.

Fig. 12.

Recognition results of various classes: a RFMID dataset, b DRIVE dataset, c STARE dataset, d IDRiD dataset, e DR_Data

Figure 13 illustrates that the model consistently demonstrates strong performance across multiple datasets, including RFMID, DRIVE, STARE, IDRiD, and DR_Data. On the IDRiD dataset, it achieves an accuracy of 98.8%, with a specificity of 99.0%, indicating its high ability to correctly identify non-DR cases. In terms of all other metrics, the VCANet-COP model showed strong performance on all other datasets which highlights the model’s robustness in detecting lesions effectively and makes it a reliable tool for medical diagnosis in clinical settings.

Fig. 13.

Analysis of lesion recognition performance of the VCANet-COP method across diverse datasets

Table 5 provides a detailed comparison of five baseline models with the proposed model assessed on five widely used datasets based on critical performance metrics. Among these, VCANet-COP consistently demonstrated superior performance across all datasets by achieving the highest accuracy, sensitivity, and specificity while maintaining the lowest FPR, FNR, and execution time. This makes it the most reliable model for lesion detection in DR images.

Table 5.

Comparison study of the different methods across diverse datasets

Model	Accuracy (%)	Sensitivity (%)	Specificity (%)	Precision (%)	F1-score (%)	Kappa coefficient (%)	FPR (%)	FNR (%)	Execution time (s)
RFMID dataset
VCANet-COP	97.1	96.8	97.52	97.5	97.2	95.2	2.48	3.2	1.2
EDLDR	96.3	95.9	95.2	96.4	96.1	92.6	4.8	4.1	1.8
FFU_Net	95.9	95.5	94.0	95.8	95.6	91.2	6.0	4.5	3.1
LSTM_MFORG	93.5	96.2	93.8	96.7	96.4	90.3	6.2	3.8	2.1
fundus-DeepNet	92.1	95.7	92.7	96.0	95.8	91.9	7.3	4.3	3.0
CNN_SVD	90.0	95.6	91.5	94.9	95.2	89.8	8.5	4.4	2.6
DRIVE dataset
VCANet-COP	98.1	98.0	98.42	98.4	98.2	95.8	1.58	2.0	2.0
EDLDR	97.2	97.0	91.94	97.6	97.3	93.3	8.0	3.0	3.1
FFU_Net	96.9	96.7	97.9	97.2	96.9	94.1	2.1	3.3	2.8
LSTM_MFORG	97.0	96.8	94.8	97.1	96.9	92.7	5.2	3.2	2.8
fundus-DeepNet	96.8	96.5	95.7	97.0	96.7	90.4	4.3	3.5	3.6
CNN_SVD	96.7	96.4	93.6	96.9	96.6	91.5	6.4	3.6	4.5
STARE dataset
VCANet-COP	98.5	98.2	98.48	98.7	98.4	96.6	1.52	1.8	1.3
EDLDR	97.6	97.4	92.36	97.9	97.6	94.4	7.64	2.6	4.0
FFU_Net	97.4	94.2	95.2	97.8	95.9	93.9	4.8	5.8	1.8
LSTM_MFORG	96.5	91.3	91.0	97.5	94.3	94.0	9.0	8.7	2.7
fundus-DeepNet	97.3	97.0	95.9	97.6	97.3	93.1	4.1	3.0	3.1
CNN_SVD	94.2	96.9	93.8	97.3	97.1	92.3	6.2	3.1	3.9
IDRiD dataset
VCANet-COP	98.8	98.6	99.0	98.9	98.7	97.2	1.0	1.4	1.4
EDLDR	98.0	97.7	94.8	98.2	97.9	95.4	5.2	2.3	1.9
FFU_Net	97.8	97.6	97.7	98.1	97.8	94.8	2.3	2.4	3.6
LSTM_MFORG	96.9	97.7	95.6	98.3	98.0	92.0	4.4	2.3	4.1
fundus-DeepNet	97.7	92.5	96.5	97.9	95.1	93.8	3.5	7.5	2.3
CNN_SVD	97.6	97.4	93.4	97.8	97.6	97.1	6.6	2.6	2.8
DR_Data
VCANet-COP	98.2	97.9	97.0	98.5	98.2	96.4	3.0	2.1	1.6
EDLDR	97.3	97.0	95.6	97.6	97.3	94.1	4.4	3.0	2.0
FFU_Net	98.0	96.7	94.4	97.3	97.0	93.9	5.6	3.3	3.4
LSTM_MFORG	97.1	93.9	96.2	97.4	95.6	90.5	3.8	6.1	4.8
fundus-DeepNet	96.9	96.6	93.8	97.2	96.9	89.2	6.2	3.4	2.8
CNN_SVD	96.8	91.5	93.5	97.0	94.2	92.9	6.5	8.5	3.9

The average performance of the six models across all five datasets is given in Table 6. The VCANet-COP model achieves high accuracy, sensitivity, specificity, and F1-score alongside efficient execution time. Moreover, EDLDR and FFU_Net also demonstrate competitive performance by maintaining strong accuracy and sensitivity but showing slightly lower specificity compared to VCANet-COP. This illustrates the impressive performance of the VCANet-COP model in lesion detection.

Table 6.

Average outcomes of the different methods across diverse datasets

Model	Accuracy (%)	Sensitivity (%)	Specificity (%)	Precision (%)	F1-score (%)	Kappa coefficient (%)	FPR (%)	FNR (%)	Execution time (s)
VCANet-COP	98.14	97.9	98.08	98.4	98.1	96.2	2.0	2.1	1.5
EDLDR	97.28	97. 0	93.98	97.5	97.2	94.0	6.20	3.0	2.56
FFU_Net	97.20	96.14	95.84	97.2	96.6	93.6	4.16	3.86	2.94
LSTM_MFORG	96.20	95.4	94.28	97.4	96.4	91.9	5.68	4.82	3.30
fundus-DeepNet	96.16	95.66	94.92	97.1	96.4	92.08	5.08	4.34	3.0
CNN_SVD	95.06	95.56	93.16	96.8	96.1	92.7	6.84	4.44	3.54

Figure 14 presents the AUC-ROC curve of the VCANet-COP model. The AUC-ROC curve shows strong discriminative power across both normal and abnormal classes, ensuring robust classification. This robust discriminative power ensures that the model performs well even when the dataset has a skewed distribution of positive and negative samples.

Fig. 14.

AUC-ROC curve analysis of the VCANet-COP model across different datasets: a RFMID dataset, b DRIVE dataset, c STARE dataset, d IDRiD dataset, e DR_Data

Moreover, the AUC-PRC curve given in Fig. 15 highlights the model’s ability to maintain high precision and recall for minority classes, such as severe and proliferative DR which are often underrepresented in imbalanced datasets. This suggests that the VCANet-COP model effectively prioritizes the identification of these minority classes, thus minimizing false positives and false negatives.

Fig. 15.

AUC-PRC curve analysis of the VCANet-COP model across different datasets: a RFMID dataset, b DRIVE dataset, c STARE dataset, d IDRiD dataset, e DR_Data

Figure 16 presents the FPR and FNR across five datasets to evaluate the model’s reliability in handling imbalanced datasets. The low FPR indicates minimal misclassification of healthy cases, and the low FNR demonstrates effective detection of retinal diseases, ensuring accurate diagnosis across severity levels.

Fig. 16.

Error rate analysis of the VCANet-COP model across different datasets: a FPR and b FNR

Figure 17 presents the F1-score analysis of different datasets which highlights the model’s effectiveness in handling imbalanced datasets. Despite class disparities, the model maintains high F1-scores across all disease severity levels, ensuring balanced performance in detecting both common and rare retinal abnormalities.

Fig. 17.

F1-score analysis of the VCANet-COP model across different datasets

The robustness analysis given in Table 7 evaluates the performance of the VCANet-COP under various image distortions. The baseline scenario without distortions achieves the highest accuracy of 98.14% and F1-score of 98.1%. Point noise results in the most notable accuracy drop of 95.42% due to random pixel disruptions. Moreover, Gaussian blurring and over-saturation introduce moderate effects and preserve overall performance by achieving an accuracy of 96.85% and 96.47%, respectively. The splash-masked image remains relatively stable by achieving an accuracy of 97.02%, indicating the model’s adaptability to selective grayscale transformations. Overall, the model demonstrates high robustness, with minimal degradation across different distortions.

Table 7.

Robustness analysis of the VCANet-COP model

Various scenarios	Accuracy (%)	Precision (%)	F1-score (%)	Kappa coefficient (%)	FPR (%)	FNR (%)	Execution time (s)
w/o image distortions	98.14	98.4	98.1	96.2	2.0	2.1	1.5
Point noise	95.42	95.8	95.6	91.3	3.8	4.2	1.7
Gaussian blurred image	96.85	97.1	96.9	93.6	3.1	3.4	1.8
Splash-masked image	97.02	94.3	94.1	94.1	2.8	3.0	1.6
Over-saturated image	96.47	96.7	96.5	92.8	3.4	3.6	1.7

Here, the various optimization algorithms are employed and compared with the developed COP algorithm to analyze the efficiency of the algorithm in optimizing VCANet for lesion detection. The algorithms used for the comparison study are OOA, Harris hawk optimization (HHO) [28], moth–flame optimization (MFO) [29], whale optimizer algorithm (WOA) [30], and particle swarm optimization (PSO) [31]. These algorithms are inspired by the natural behaviors of respective animals, and they are employed to optimize VCANet for improved lesion detection accuracy and efficiency. Table 8 compares the optimization properties of various algorithms. The COP algorithm in the VCANet utilizes a crossover strategy and random mutations. This design prevents premature convergence to suboptimal solutions and enables the algorithm to explore more search space effectively. As a result, VCANet-COP achieves the lowest best fitness value of 0.00012, the fastest convergence speed of 45 iterations, and the shortest execution time of 1.5 s. This ability to overcome local optima and converge toward optimal solutions significantly enhances the model’s speed, accuracy, and computational efficiency, surpassing traditional optimization techniques.

Table 8.

Analyzing the optimization properties of the COP algorithm against optimization algorithms

Model	Best fitness value	Mean fitness	Standard deviation	Convergence speed (iterations)	Success rate (%)	Execution time (s)
VCANet-COP	0.00012	0.00035	0.00008	45	98.5	1.5
VCANet + OOA	0.00015	0.00040	0.00010	50	97.8	1.6
VCANet + HHO	0.00018	0.00055	0.00012	55	96.5	1.8
VCANet + MFO	0.00022	0.00060	0.00014	60	94.9	2.0
VCANet + PSO	0.00019	0.00053	0.00011	57	95.4	1.9
VCANet + WWO	0.00025	0.00068	0.00016	65	92.7	2.1

Table 9 presents a comparative analysis of COP algorithm in lesion recognition across diverse metrics. The VCANet-COP model demonstrates superior performance with the highest accuracy of 98.14%, sensitivity of 97.9%, specificity of 99.1%, and precision of 98.3%. In contrast, the existing methods exhibit a noticeable decline in all major metrics and have higher false positive and false negative rates. Overall, it is clear that the superior COP algorithm enables VCANet-COP to outperform other baseline models in lesion detection tasks with better performance.

Table 9.

Analyzing the role of the COP algorithm in lesion recognition against baseline optimization algorithms

Model	Accuracy (%)	Sensitivity (%)	Specificity (%)	Precision (%)	F1-score (%)	Kappa coefficient (%)	FPR (%)	FNR (%)
VCANet-COP	98.14	97.9	99.1	98.3	98.1	96.2	2.0	2.1
VCANet + OOA	97.9	97.2	91.7	97.8	97.5	93.9	6.5	3.3
VCANet + HHO	96.6	96.8	95.4	94.5	95.6	94.3	5.8	3.5
VCANet + MFO	97.2	95.5	94.2	94.0	94.7	95.8	3.1	5.9
VCANet + PSO	95.5	96.9	93.3	95.3	96.1	90.6	2.9	2.4
VCANet + WWO	93.8	93.2	90.9	92.7	92.9	92.5	4.3	4.7

Figure 18 visualizes the features extracted by the SAE and compares them with existing methods. The visualization outcomes reveal that SAE highlights lesion-specific features (microaneurysms, hemorrhages, and exudates) and vascular features by employing sparse representations. It captures subtle and discriminative features that aid in accurate lesion detection, especially for small and irregular patterns. In contrast, existing methods focus on broader patterns or pre-defined features and struggle to identify small and complex lesions. Therefore, the unsupervised learning capability of the SAE model enables it to extract more meaningful and detailed features, thus improving performance.

Fig. 18.

Visualization of features extracted by the sparse autoencoder and other existing methods: a Microaneurysms, b hemorrhages, and c exudates

Figure 19 visualizes the model’s lesion detection ability through gradient-weighted class activation mapping (Grad-CAM). This technique highlights the regions of the image that contribute most to the model’s decision to provide valuable insights into the predictions made by the model. By visualizing lesions, the heatmap improves transparency and trust in the model’s predictions. This is crucial for clinical applications to understand the rationale behind the model’s predictions and ensure alignment with medical expertise.

Fig. 19.

Visualization of Grad CAM output to analyze the efficiency of the model in lesion recognition: a input, b Grad-CAM output

Figure 20 reveals that the VCANet-COP model obtains 98.14% accuracy, 97.9% sensitivity, and 99.1% specificity, showing strong overall performance in lesion detection tasks. Meanwhile, EDLDR and FFU_Net showed recognition accuracy of 97.28% and 97.20%, respectively. Additionally, the LSTM_MFORG model achieves 96.20% accuracy while fundus-DeepNet and CNN_SVD report slightly lower results with 96.16% and 95.06% accuracy, reflecting the higher efficiency of the VCANet-COP model.

Fig. 20.

Lesion recognition accuracy analysis of the different methods

Figure 21 provides the precision analysis of the different methods across where VCANet-COP achieves a precision rate of 98.4%, reflecting its superior ability to identify true positives. The existing methods revealed lower precision in comparison with the VCANet-COP model, demonstrating the supremacy of the model in lesion detection.

Fig. 21.

Analysis of the precision of the VCANet-COP model against baseline approaches

Figure 22 shows that the VCANet-COP model obtains the highest F1-score of 98.1% with the shortest execution time of 1.5 s. However, LSTM_MFORG obtains an F1-score of 97.36% but takes 3.30 s for execution. In the case of all other methods, the F1-score for the lesion recognition is lower, and the execution time is higher when compared with the proposed model. The figure illustrates that the VCANet-COP model with its advanced architecture efficiently recognized lesions in color fundus images with improved computational efficiency, making it highly suitable for real-time medical applications.

Fig. 22.

F1-score analysis of the different approaches employed for lesion recognition

Figure 23 indicates that the VCANet-COP model achieves the lowest FPR and FNR values of 2.0% and 2.1% respectively. Moreover, the VCANet-COP model exhibits a significantly reduced error rate in comparison with other baseline methods. The result illustrates that the VCANet-COP model accurately distinguished lesions with minimized false alarms and missed detection rates, thus providing an ideal solution for improving the precision and robustness of automated retinal disease diagnosis.

Fig. 23.

Analysis of the FPR and FNR generated for different methods during lesion identification

Figure 24 presents the execution times of different lesion recognition models. VCANet-COP is the fastest with a time of 1.5 s while baseline methods take slightly longer time than the proposed model.

Fig. 24.

Comparison of the execution time of various DR lesion recognition techniques

Table 10 presents a Friedman performance ranking of various models based on diverse metrics with the VCANet-COP model leading in all categories and achieving the best average rank of 1.00. The EDLDR model follows with an average rank of 2.75 and the CNN_SVD model ranks the lowest with an average rank of 5.63.

Table 10.

Friedman performance-based ranking table for different methods across diverse metrics

Model	Accuracy (%)	Sensitivity (%)	Specificity (%)	Precision (%)	F1-score (%)	Kappa coefficient (%)	FPR (%)	FNR (%)	Average rank
VCANet-COP	1	1	1	1	1	1	1	1	1
EDLDR	2	2	5	2	2	2	5	2	2.75
FFU_Net	3	3	2	4	5	3	2	3	3.13
LSTM_MFORG	4	6	4	3	3	6	4	5	4.36
fundus-DeepNet	5	4	3	5	4	5	3	4	4.13
CNN_SVD	6	5	6	6	6	4	6	6	5.63

Table 11 summarizes the statistical analysis of the model performance using the Friedman test by comparing different models. The chi-square values are higher than the critical value, with corresponding p-values all less than 0.05, indicating statistical significance. The critical value at a significance level of 0.05 is 11.07 for degrees of freedom 5(6–1) based on the chi-square distribution table, and since all models show p-values below this threshold, the null hypothesis (that there are no significant differences between the models) is rejected for each model. This suggests that the models differ significantly in their performance across the evaluated metrics, with each outperforming the others in certain aspects.

Table 11.

Statistical analysis using Friedman test for model comparison

Model	Chi-square Value	p-value	Critical value (σ = 0.05)	Null hypothesis Status
VCANet-COP	15.99	0.0081	11.07	Rejected
EDLDR	20.83	0.0072	11.07	Rejected
FFU_Net	14.98	0.0099	11.07	Rejected
LSTM_MFORG	36.82	0.0012	11.07	Rejected
fundus-DeepNet	19.05	0.0092	11.07	Rejected
CNN_SVD	28.73	0.0034	11.07	Rejected

Discussion

This section critically evaluates how the VCANet-COP model addresses the limitations made by the prior approaches and also highlights its novel contributions to DR detection. The experimental results as summarized in Table 6 illustrate the average comparative performance of the VCANet-COP model by achieving an accuracy of 98.14%, sensitivity of 97.9%, specificity of 98.08%, and execution time of 1.5 s against existing DR detection models. Furthermore, a critical evaluation of existing literature reveals that while many models achieve high accuracy in DR detection, they often fail in specific areas such as small lesion recognition, feature representation efficiency, and real-time clinical applicability.

Deriving Efficient Feature Representation

Since small and irregular lesions are often challenging to recognize, extracting comprehensive feature representation is crucial for accurate lesion recognition. Traditional models such as LSTM-MFORG [10] and Fundus-DeepNet [14] have shown promising results in DR classification. However, these models fail to capture complex spatial hierarchies in retinal images. The FFU-Net model [16] attempted to improve feature representation through multi-scale feature fusion and contextual channel attention mechanisms but still struggled with capturing subtle lesions effectively. In contrast, the VCANet-COP model introduces a hierarchical architecture that mimics the primate visual processing system (V1, V2, V4, IT regions). This multi-stage structure allows the model to progressively refine features, ensuring that small and irregular lesion patterns are accurately detected. By leveraging SAEs for vascular and lesion-specific feature extraction, VCANet-COP significantly outperforms prior models in terms of lesion differentiation and recognition accuracy.

Improving Computational Efficiency and Execution Time

High computational complexity remains a significant barrier to the clinical deployment of DR detection models. The AI-driven VTDR detection framework based on CNN-SVD by Bilal et al. [12] has demonstrated enhanced classification accuracy but at the cost of increased computational overhead. The VCANet-COP overcomes this limitation by combining the OOA algorithm with a crossover strategy. This approach enhances hyperparameter tuning efficiency, accelerating convergence and reducing computational overhead. As shown in Table 6, VCANet-COP achieves an execution time of 1.5 s and significantly outperformed other models like EDLDR and FFU-Net having execution times of 2.56 and 2.94 s, respectively. This substantial reduction in processing time makes VCANet-COP a more viable approach for real-time clinical applications.

Enhancing Generalization Across Diverse Datasets

Many existing DR detection models lack generalizability. Models such as EDLDR [15] and Fundus-DeepNet [14] have demonstrated promising results but were primarily evaluated on small or homogeneous datasets. Therefore, VCANet-COP resolves these problems by evaluating the framework with various publicly available DR datasets. Additionally, the SAE-based feature extraction further enhances the generalization ability of the model by learning a broader range of vascular and lesion characteristics. As a result, VCANet-COP achieves the highest Kappa coefficient of 96.2%, indicating superior agreement between model predictions and ground truth annotations across diverse datasets.

Overall, the VCANet-COP model demonstrates the best balance of high performance and computational efficiency by effectively addressing gaps in prior methods and setting a new standard for similar applications.

Conclusion

In this study, we proposed the VCANet-COP model to enhance subtle lesion recognition in DR screening. At first, data preprocessing such as image normalization, noise reduction, contrast enhancement, data augmentation, and image resizing are performed. Particularly, data augmentation techniques such as rotation, flipping, and scaling were applied to handle imbalanced datasets. The use of SAEs in VCANet-COP enables the model to learn efficient representations at a pixel level, allowing it to simultaneously capture vascular structures and lesion-specific features. The extracted features are fed to VCANet which simulates the hierarchical structure of the primate visual system and allows the model to capture subtle lesion features, which are critical for improving the accuracy and robustness of DR detection across diverse datasets. The COP algorithm is used to optimize the hyperparameters and network configurations of the VCANet by efficiently exploring the solution space. This process avoids premature convergence, ensures faster convergence, and significantly reduces the training time. This leads to a more computationally efficient model without compromising detection performance, making VCANet-COP well-suited for real-world DR screening. The extensive evaluations on multiple DR Datasets such as DRIVE, STARE, DR-Data, RFMID, and IDRiD datasets revealed outstanding results compared to existing models with accuracy of 98.14%, precision of 98.4%, sensitivity of 97.9%, specificity of 98.08%, F1-score of 98.1%, and kappa coefficient of 96.2%. In particular, the VCANet-COP model demonstrates significant improvements in both accuracy and execution time. By achieving 98.14% accuracy, it outperforms EDLDR by 0.86%, FFU_Net by 0.94%, LSTM_MFORG by 1.94%, fundus-DeepNet by 2.0%, and CNN_SVD by 3.08%. In terms of execution time, VCANet-COP performs significantly faster with 1.5 s, outperforming EDLDR (2.56 s), FFU_Net (2.94 s), LSTM_MFORG (3.30 s), fundus-DeepNet (3.0 s), and CNN_SVD (3.54 s). Overall, the proposed model outperformed existing models in both detection and computational speed, making it a highly suitable solution for real-world clinical applications.

Limitations

While data augmentation enhances generalization, it may not account for all rare or extreme variations in DR images, affecting its performance on atypical lesions. Additionally, biases in the training datasets such as underrepresentation of certain demographic groups or variations in image quality may impact the model’s generalizability and fairness across diverse patient populations. Although VCANet-COP has shown impressive results on the selected DR datasets, scalability to larger, more complex datasets or real-world clinical settings may pose challenges.

Future Work

Future work could focus on refining the VCANet-COP model by incorporating advanced transfer learning techniques to improve performance on smaller datasets and exploring semi-supervised learning to reduce reliance on labeled data. Implementing real-time processing, integrating with diagnostic systems like OCT, and validating across diverse ethnic groups will ensure broader applicability and fairness. Enhancing scalability for larger datasets, extending the model to multi-disease classification, and optimizing for real-world clinical settings will significantly broaden its potential, transforming it into a comprehensive solution for retinal disease screening.

Data Availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Consent to Participate

Informed consent was obtained from all individual participants included in the study.

Human and Animal Rights

This article does not contain any studies with human or animal subjects performed by any of the authors.

Conflict of Interest

The authors declare no competing interests.