A Hybrid CNN-RBF Approach for Classification of Diabetic Retinopathy

According to WHO health statistics, maintaining high blood glucose levels causes damage to blood vessels in many parts of the body and can bring on various complications. Some complications of diabetes include loss of sight and blindness from problems with the eye’s blood vessels, sores and loss of feet from nerve problems and kidney problems [1]. According to the 2025 edition of the International Diabetes Federation (IDF) Diabetes Atlas, diabetes affects 11.1% of adults aged 20-79 years worldwide—equivalent to roughly one in every nine individuals. Alarmingly, more than 40% of these adults remain undiagnosed. Projections suggest that by 2050 the prevalence will rise to about one in eight adults, representing nearly 853 million people, which marks a 46% increase compared to current figures. Approximately one-third of diabetic patients are at risk of developing vision loss during their lifetime [2]. Retinal fundus imaging is a convenient way to study these disorders and their progression. Particularly for diabetics, color fundus images highlight blood vessels to identify any early signs of diabetes-related retinal disease. When the disease is found and treated early, it can be more successfully managed and prevented from getting worse [3]. AI makes it possible to accurately assess the severity of DR using image pre-processing and deep learning algorithms [4]. The ongoing advancements in this area have made it possible to combine AI into automated retinal vessel detection [5]. Those with uncontrolled diabetes are more likely to develop DR since excess diabetes can injure the retinal vessels. It is necessary to find and separate the blood vessels in the retina for DR diagnosis and to prevent premature vision loss [6]. Doctors now rely on retinal vascular segmentation technology which makes the job easier for both experienced and beginner ophthalmologists [7].

Therefore, dividing retinal vessels by programs is very important for both diagnosing and treating DR at various stages. In clinical imaging, there is a need for new ways to automate and improve how retinal problems like DR are recognized. We present a strategy that combines Multi-Scale Discriminative Robust Local Binary Patterns (MS-DRLBP) with a Convolutional Neural Network Radial Basis Function (CNN-RBF) based classifier. The method aims to simplify the segmentation of retinal blood vessels and avoid drawbacks of manual segmentation; hence it greatly increases the accuracy and effectiveness of diagnosing DR. By combining deep learning and pattern recognition, our method improves the reliability and accurately diagnoses DR. We demonstrate in the study that by using this approach instead of previous ones, it improves the way different retinal illnesses can be identified. Correspondingly, numerous research works include AI in the identification of age-related macular degeneration, glaucoma and DR. In previous work, researchers used machine learning and deep learning approaches to organize retinal diseases. Local binary patterns offer a practical approach for feature extraction; however, since the features are local, they may not be reliable for identifying disease classes. Additionally, to work well, CNNs often need access to a big set of data that has been annotated and a computer that can handle the workload. Better tools for correctly spotting and grouping retinal disorders are necessary so that scientific research can advance in this field. Our suggested research explains an automatic threshold technique for diagnosing DR using blood vessel segmentation. In addition, to fix earlier issues, we use a classifier that pairs CNN-RBF and we modify DRLBP. The new search algorithm looking for boundaries in objects helps the classifier focus on their texture and shape for better identification. Segmented fundus image features allow us to change the usual RBF training, so the model can use data that is not fully labeled for improved classification.

Contributions of this paper include:

• We use three well-known public datasets APTOS 2019, EyePACS, and Messidor to cover a wide range of retinal images.

• To better highlight disease patterns, we apply a multi-scale feature extraction method called MS-DRLBP, which captures detailed retinal textures.

• For classification, we propose a hybrid model that combines the feature learning power of CNN with the decision-making ability of RBF classifiers.

Although RBF have many benefits, incorporating them into contemporary CNN designs can be difficult because of their nonlinear activation, which can obstruct effective gradient flow, and the presumption of fixed MS-DRLBP features with preset cluster centers at the beginning.

The structure of the remaining sections is as follows: We examine earlier studies in Section 2, and Section 3 gives the dataset and methodology, including our suggested strategy. Section 4 contains the experimental findings that indicate the efficacy of our model versions on the dataset. Section 5 contains the paper's conclusion.

1.png

Figure 1. Developed architecture

3.1 Data acquisition

The retinal fundus images and their related annotations or diagnoses are normally gathered for research on DR. These are the Diabetic Retinopathy Datasets from EyePACS, APTOS, APTOS (Gaussian Filtered) and Messidor. This combining dataset increases diversity in imaging conditions, acquisition devices, and patient demographics, thereby reducing overfitting to a single domain. In total, the data set contains 92,501 jpg files which were divided randomly into train (60%) and test (40%) [22]. About 55% more data was added to the dataset by manually altering it. All pictures have been adjusted to 600x600 pixels. Therefore, the size of the data set was cut by more than half (18.5GB to 3.8GB), helping lower resource use for tasks such as data augmentation and resizing in the training process.

2.png

Figure 2. Types of DR based on fundus features

For clearer understanding, refer to the retinal fundus images that represent various phases of DR. Figure 2 displays example images corresponding to the five graded severity levels of DR. Below Table 1. gives the image distribution between 3 benchmark datasets. The decision to use data augmentation led to 60:40 as the training and test split and total number of images was raised to 143,669 and data augmentation was performed exclusively on the training set, with care taken to avoid any overlap with the test samples.

3.2 Dataset preprocessing

The presence of noise in the red and blue channels is diminished by removing the green channel from the fundus images. This green channel provides optimal balance between brightness and contrast, allowing clearer visualization of blood vessels and retinal lesions. In retinal imaging, the green channel enhances visibility by highlighting blood vessels and lesions, including microaneurysms and hemorrhages, against the background. The information from the green channel is improved by using CLAHE, which concentrates the area to a small, tiny region for better contrast. CLAHE increases the visibility of arteries and veins, most prominently in areas where things are darker in the retina [23]. The CLAHE algorithm is represented by Eq. (1):

$\mathrm{I}_{\text {enhanced }}=$ CLAHE(I, clipLimit, tileGridSize)                      (1)

where, I correspond to the input image, clipLimit controls the contrast enhancement, tileGridSize defines the size of the contextual regions.

A Gaussian filter is an image-blurring filter that uses a Gaussian function for weighting pixel values [24]. It's widely used in image processing to reduce noise and detail, acting as a smoothing filter and it can operate according to Eq. (2):

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

The pixel coordinates (x, y) are relative to the kernel's center, and σ is the standard deviation. Images after all the preprocessing steps. The sequence of preprocessing steps applied to retinal fundus images is illustrated in Figure 3.

3.png

Original Image                  Green Channel                     Denoising                    CLAHE Filter

Figure 3. Retinal fundus image preprocessing framework

3.3 Morphological operations

Morphological image processing involves a set of non-linear techniques used to analyze and modify the shape or structure of objects within an image. Morphological processing methods are commonly used on DR images to enhance and isolate significant structural features such as blood vessels and lesions [25]. These techniques, which include dilatation, erosion, opening, and closing, are employed in the preprocessing step to reduce noise, fill gaps, and increase the clarity of retinal structures. For example, morphological opening reduces small light artifacts and background noise, whereas closing can reconnect fractured vessel segments. Such changes increase the contrast of diseased regions, allowing for better segmentation and feature extraction. These improved images form a more accurate foundation for subsequent categorization tasks in automated DR diagnosis systems.

3.3.1 Erosion

Erosion of a set A by a structuring element B is the collection of all positions z for which B, when shifted so that its origin aligns with z, remains entirely within A. Shrinks bright regions and removes small noise.

$A \ominus B=\left\{z \mid(B)_z \subseteq A\right\}$                          (3)

3.3.2 Dilation

Dilation of a set A by a structuring element B is defined as the set of all positions z where, when the origin of B is placed at z, the shifted B intersects with A. Expands bright regions and fills small gaps.

$A \oplus B=\left\{z \mid(\widehat{B})_z \cap A \neq \Phi\right\}$                       (4)

3.3.3 Opening

Applying erosion first and then dilation, known as morphological opening, helps remove small protrusions, separate narrow connections, and smooth the boundaries of objects.

$A \circ B=(A \ominus B) \oplus B$                      (5)

3.3.4 Closing

When dilation is applied before erosion, a process known as morphological closing, it smooths object boundaries, bridges narrow gaps or elongated depressions, removes small voids, and fills discontinuities along the contour.

$A \cdot B=(A \oplus B) \ominus B$                      (6)

The application of morphological operations such as erosion, dilation, opening, and closing on retinal fundus images is demonstrated in Figure 4.

4.png

Original                   Erosion                  Dilation                  Opening                  Closing

Figure 4. Images after morphological operation

3.4 Otsu’s thresholding

Otsu's Thresholding is an image binarization method used in image processing to automatically perform clustering-based thresholding [26]. It was introduced by Nobuyuki Otsu in 1979 and is particularly effective for grayscale images with bimodal histograms. The Otsu algorithm determines the best threshold by enhancing the separation between background and foreground intensity levels in the histogram, making it an effective tool for medical image analysis such as detecting DR lesions.

Compute histogram of the grayscale image.

Iterate through all possible thresholds (from 0 to 255 for 8-bit images).

Divide the pixels into two groups: Class 0 background (below threshold) and class 1 foreground (above threshold).

Compute the mean and weight (probability) of each group.

Compute Intra-class variation for each criterion.

Choose the threshold that produces the lowest intra-class variance or the maximum inter-class variance.

Let t be the threshold ω₀(t), ω₁(t) be the weights (probabilities) of the background and foreground classes.

$\omega_0(t)=\sum_{i=0}^t p_i$                      (7)

$\omega_1(t)=\sum_{i=t+1}^{L-1} p_i$                       (8)

Let μ₀(t), μ₁(t) be the means of the two classes

$\mu_0(t)=\frac{1}{\omega_0(t)} \sum_{i=0}^t i p_i$                      (9)

$\mu_1(t)=\frac{1}{\omega_1(t)} \sum_{i=t+1}^{L-1} i p_i$                       (10)

μ_T be the total mean of the image

$\mu_T(t)=\sum_{i=0}^{L-1} i p_i$                      (11)

Then the between-class variance is:

$\sigma_b^2(t)=\omega_0(t) \cdot \omega_1(t)\left(\mu_0(t)-\mu_1(t)\right)^2$                      (12)

3.5 Feature extraction

Feature extraction in this work is carried out using a hybrid strategy that combines MS-DRLBP and CNN.

3.5.1 MS-DRLBP

MS-DRLBP (Multiscale-Discriminative Robust Local Binary Pattern) is a texture feature extraction technique designed for robust image analysis, particularly in tasks like face recognition, texture classification, and medical image analysis [27]. It builds upon and improves the classical Local Binary Pattern (LBP) by addressing some of its limitations, such as sensitivity to rotation and inability to capture directional and multi-scale texture information effectively. It can Improve microaneurysm, hemorrhage, and exudate detection.

The standard Local Binary Pattern (LBP) operator compares a pixel with its neighbours and encodes the differences as a binary pattern:

$L B P_{P, R}\left(x_c, y_c\right)=\sum_{p=0}^{p-1} s\left(g_p-g_c\right) 2^p$                         (13)

where, g_p and g_c are the intensity of the neighbors and central pixel P is the number of neighbors, R is the radius.

$s(z)= \begin{cases}1, & z \geq 0 \\ 0, & z \leq 0\end{cases}$                        (14)

3.5.2 Robust and Discriminative Extension (DRLBP)

Instead of directly using binary differences, DRLBP applies a discriminative weighting to reduce sensitivity to noise and illumination changes. Positive and negative differences are separated, creating two complementary histograms that capture both bright and dark lesion patterns more effectively.

3.5.3 Multi-Scale Extension (MS-DRLBP)

Features are extracted across multiple neighbourhood radii R1, R2..., Rm.

This allows detection of small lesions (microaneurysms) at fine scales and larger structures (exudates, hemorrhages) at coarser scales. Descriptor is:

$M S-D R L B P\left(x_c, y_c\right)=\bigcup_{i=1}^m D R L B P_{P, R_i}\left(x_c, y_c\right)$                       (15)

where, $\cup$ denotes the histogram concatenation across multiple scales.

3.5.4 VGG16

VGG16 is a CNN model introduced in 2014 by the Visual Geometry Group at the University of Oxford. It became famous for its simplicity, depth, and strong performance in classification tasks. VGG16 is structured into 5 convolutional blocks followed by dense layers. Each block increases the feature richness while reducing spatial resolution. The VGG16 model is composed of 16 learnable layers, including 13 convolutional layers followed by 3 fully connected layers. Uses only 3×3 convolution filters and 2×2 max-pooling layers for feature extraction. Its architecture employs successive convolutional layers arranged in blocks, each followed by max-pooling, with the feature depth gradually increasing at deeper stages of the network [28]. The detailed layer architecture of the VGG framework is shown in Figure 5. This design allows the network to capture both low-level and high-level features, making it highly effective for extracting detailed image representations. Unlike ResNet, it follows a simple sequential structure. Widely used in pre-trained models for medical imaging, object detection, etc. A widely used model with simple yet effective convolutional layers. The block1 provides the foundation for detecting blood vessel edges and optic disc boundaries in retinal images. The block2 enhances detection of small lesions like microaneurysms in DR images.  The block 3 useful for identifying hemorrhages and intermediate lesion structures. The block4 is crucial for differentiating between moderate and severe DR, where lesion groupings matter and block5 can help distinguish between proliferative DR and advanced stages where neovascularization patterns appear. Dense layers used for classification into ImageNet classes.

5.png

Figure 5. VGG16 layers

3.6 Classification methodology

The features extracted through MS-DRLBP and CNN are applied to the classifier for DR classification. The pretrained VGG16 network is employed to extract deep hierarchical features, while MS-DRLBP provides robust texture descriptors.

These complementary features capture both high-level semantic representations and fine-grained local patterns from retinal fundus images.

The extracted features are normalized and concatenated into a joint feature vector.

$F=\left\{f_{-} C N N \| f_{M S-D R L B P}\right\}$                       (16)

where, "$\|$" denotes concatenation operator.

This vector retains spatial, textural, and structural information critical for DR detection. Instead of using a conventional softmax, we employ an RBF network where each hidden neuron applies a Gaussian kernel:

$h_j(x)=\exp \left(-\frac{\left|x-c_j\right|^2}{2 \sigma_j^2}\right)$                     (17)

where,

x is input feature vector

c_j is the centre of the j^th RBF neuron

σ_j is parameter controlling the width of the Gaussian

The final output is:

$y_k(x)=\sum_{j=1}^M w_{k j} h_j(x)+b_k$                      (18)

where, M is number of hidden RBF neurons.

W_kj is the weight connecting to j^th hidden neuron to the k^th output classes b_k is the bias

To thoroughly evaluate classification performance and enhance the DR detection rate, we employed various classifiers, including RBF, SVM (Support Vector Machine) and CNN-RBF. The CNN-RBF classifier benefits from random initialization of network weights and RBF centers. This process incorporates a controlled level of randomness, optimizing the diversity of the neural network’s internal representations and enhancing its capacity to learn complex patterns. Such initialization mimics the essence of non-iterative randomization-based methods, facilitating a more efficient solution for space exploration. Integrating MS-DRLBP features further enriches this model, enabling the discrimination of subtle variances between normal and DR-affected retinal images. This method marks a significant breakthrough in medical analysis, emphasizing the capability of hybrid models to attain high precision in disease classification tasks. Instead of using a conventional softmax, we employ an RBF network where each hidden neuron applies a Gaussian kernel.

3.7 Network model and training

In this study, the training hyperparameters such as learning rate and batch size were optimized to maximize classification accuracy, highlighting the effectiveness of hybrid architectures for medical image analysis. Within the CNN, neurons are spatially arranged in tiled patterns, each corresponding to a localized region of the input field. The proposed framework integrates the CNN with a RBF classifier, where stochastic learning strategies are utilized to simplify the training process and minimize computational overhead.

Unlike traditional iterative optimization methods like back propagation, randomization-based approaches focus on non-iterative processes or stochastic initialization to efficiently learn decision boundaries. In our model, the RBF layers integrate stochastic modeling to enhance feature extraction and classification without the need for extensive iterative tuning. This results in faster training and improved efficiency, which is especially important for applications requiring scalable and accessible solutions. The MS-DRLBP features are integrated with the features extracted by CNN-RBF in this study. The present work utilizes a CNN model (VGG16) with hyperparameter settings as per Table 3 including maximum of 100 epochs, a learning rate of 0.0001, and a momentum factor of 0.9, with batch sizes tested at 8, 16 and 32. At batch size 32, model achieve higher accuracy in retinal image classification. Based on the training, the tested optimizer includes Stochastic Gradient Descent (SGDM) with yielding superior results.

Table 3. Hyper parameters of CNN

3.8 PSO optimization

In the CNN-RBF hybrid model, PSO helps fine-tune hyperparameters, optimize RBF kernel values, and automatically find the best settings, reducing manual trial-and-error. In this work, PSO is applied to optimize model hyperparameters. The configuration of PSO hyperparameters used for tuning the proposed model is summarized in Table 4. A swarm of 20 particles explores the search space for 30 iterations. The inertia weight (w = 0.7) maintains a balance between exploration and convergence. The cognitive parameter (c1 = 1.5) guides particles using their own best experience, while the social parameter (c2 = 1.5) directs them toward the global best solution. Together, these settings ensure efficient convergence and improved classification performance.

Table 4. Configuration of PSO hyperparameters

$\begin{aligned} v_i(t+1)=\omega v_i(t) & +c_1 r_1\left(p_i^*-x_i(t)\right)+c_2 r_2\left(g^*\right.\left.-x_i(t)\right)\end{aligned}$                   (19)

$x_i(t+1)=x_i(t)+v_i(t+1)$                      (20)

Here, ω denotes the inertia weight, c₁ and c₂ represent the acceleration coefficients, while r₁ and r₂ are uniformly distributed random values in the range [0,1].

PSO is applied to minimize the loss function (categorical cross-entropy) on the validation set, defined as:

$\mathcal{L}=-\sum_{i=1}^N \sum_{j=1}^K y_{i j} \log \left(\hat{y}_{i j}\right)$                       (21)