Enhancing Diabetic Retinopathy Detection with Weighted Sum Ensemble Models

Recently, the breakthrough in healthcare technologies has been brought by artificial intelligence (AI) with its application and potential to transform medical diagnostics notably evolved [1]. The introduction of AI techniques, primarily Image analysis using CAD (Computer Aided Diagnosis) has revolutionized the improvement in diagnostic specificity and efficiency. Considering the burden posed by chronic conditions, particularly higher prevalence rates of diabetic retinopathy (DR), a leading blinding disease worldwide; this is an important step forward. In this regard, the deployment of deep learning (DL) models in general and prominent architectures such as ResNet50 in particular have been at the front line making important contributions to timely diagnosis and appropriate control strategies for these diseases [2].

Various aspects of healthcare delivery have been transformed through AI in medical diagnostics. Despite how it may sound, using the traditional approach manual interpretation of clinical data for medical diagnostics created a host challenge in regards to speed, accuracy and scalability. But the rise of AI has sparked a change in mindsets towards increased automation and sophistication for analysis [3]. In particular, the ability of AI to analyze and interpret great volumes of clinical imaging data more accurately has been a boon. It is not simply an improvement of the current but a reimagining of what medical diagnostics can be. Perhaps, for the handling of heavy diagnostic tasks - so complex that few specialists are competent in them anywhere on this globe - Watson-like tools will do a marvelous job.

A medical condition like diabetic retinopathy, a common complication of diabetes that can lead to blindness if untreated, perfectly typifies the kind of diagnosis AI could greatly improve. The diabetics population at risk of having DR have been reported to subsequently increase significantly in many countries and are very high burden among the working aged adults who if not detected, do occur visual impairment or blindness [4]. Current DR diagnostic pipeline focuses on ophthalmologists scrutinizing fundus images to detect signs of retinal damage, which represents the traditional approach. This approach is not only time-consuming but there are also limitations because of the availability of medical professional required for these tests and resources. Second, due to the subjective nature of manual examination, diagnostic results may vary among examiners - less than ideal for diseases that often require early detection before irreversible damage is done.

Deep learning is subset of machine learning (also known as deep structured) and opulent with complex neural networks changes the landscape of medical image analysis specifically in detection of diabetic retinopathy. In the family of deep learning architectures, ResNet50 has been highlighted as a powerful model capable to learn effective features from massive datasets called in medical imaging domain and is resilient against problems such vanishing gradients for very deep networks [5]. How: This architecture uses residual learning, a shortcut that makes it easier for the network to train deeper nets by enabling training of those additional layers without needing them all to be learned on their own. In particular ResNet50 has been found to identify early DR better than any previous methodology, as it is capable of detecting unique and subtle patterns present in the retinal images which exist before traditional methods [6] suggested they are able.

Deep learning models such as ResNet50 have been used for automated classification of retinal images to detect presence or absence of microaneurysms, haemorrhages and exudates which are the key determinants in diagnosis and grading DR [4]. These advancements have greatly improved the diagnostic capacity and time required for diagnoses of these models. The advantage of the deep learning models is that they can process this patient quickly rather than manually screening volume, which further helps in bridging a gap between patient requirement and medical resources today [4].

The addition of reference alternatives including deep learning models such as ResNet50 into cardiovascular events, diabetic retinopathy and many other conditions diagnostic assures revolutionary evolution in the implementation status of AI across health care. Through the automation of such complex medical imaging analyses, these technologies not only enhance diagnostic accuracy and efficiency but also make crucial services scalable to wider populations. With the ever-growing shot of diabetes around-the-globe, it is unlikely for AI and deep learning not to be included in managing complications from this debilitating condition unveiling a future where technology and healthcare meet delivering better patient outcomes [7-9].

While ensemble learning, in combination with deep learning has great potential in improving the predictive performance especially for complex tasks such as medical diagnostics. In this method, we combine various deep learning models together and form an ensemble, which usually provides more accurate predictions with better generalization than any single model can provide alone. To reduce overfitting and improve robustness of the predictions, ensemble methods combine outputs from different trained models - e.g., diverse flavors of convolutional neural networks or architectures like ResNet50 [10]. This is especially important in medical image analysis as the heterogeneity of data and subtle nature of important features necessitate serial diagnostic systems with extremely high reliability and precision. It can be thought of as a philosophy that combines the best features of different learning algorithms together in order to provide us with an overall framework which compounds from various biases and variances present in individual models. Consequently, not only does this union improve diagnostic accuracy but also confidence in clinical decision-making which results in better patient outcomes for conditions like diabetic retinopathy [11, 12].

3.1 General algorithm

It is concluded that an efficient approach has been developed for preprocessing and grading of retinal images aiming at improving image quality to make it suitable a more accurate classification. This starts with Image Enhancement to enhance the visual characteristics and remove noise from the retinal images. This step uses a intensity of Gaussian Blur and Image Resizing with Blending algorithms. The image is first scaled to a set size, in order to maintain same dimensions across all the images. Upon resizing, a Guassian blur is used to soften the image reducing noise while still protecting important signifiant information. The blurred image is then added to the original image using certain weights that emphasize important patterns while reducing noise and make it ready for subsequent analysis.

Then the process moves on to Histogram Equalization, particularly using Contrast Limited Adaptive Histogram Equalization (CLAHE). Cellular image processing is an optional step used to increase the contrast and dynamic range of the figure. CLAHE rescales intensity values in the image to use a full dynamic range, improving both appearance and analysis. We strive to retain the black appearance of retinal features in images, and this is key for depicting these findings adequately.

The proposed methodology follows a systematic four-phase approach as illustrated in Figure 1. The process begins with image pre-processing to standardize the input, followed by enhancement techniques to improve image quality. The third phase incorporates deep learning for feature extraction and initial classification, culminating in the final phase of ensemble learning to combine multiple model predictions for improved accuracy. This structured pipeline ensures thorough processing of retinal images before making diagnostic predictions.

1.png

Figure 1. The four-phase approach for advanced medical diagnostics

3.2 Image enhancement

GaussianBlur and Image Resizing with Blending is a commonly used computer vision algorithm in the image processing to alter images. Algorithm 1 includes several steps such as resizing the image, blurring to the resized image with Gaussian blur and blending blurred because of original images. This algorithm is use to decrease the noise pixels available in an image along with preservation of vital features. This algorithm starts by entering an image and measuring this height, width. To the end of this, it measures both parameters: how tall or wide is the figure and accordingly resizes. After having found the target size dimensions, this is what it does behind the scenes to make sure these are exact sizes: This is done by generate a new image of the target dimension, and using interpolation algorithm to substitute pixels. The resulting image is then Gaussian blurred in the next step of the algorithm This operation smooths the image, decreasing noise and simplifying identification of useful features. The algorithm uses a kernel of given size and sigma for applying Gaussian blur. sigma: Value, Sigma modulates amount of smoothing to do in the image (Smaller value gives fewer smooth images). Finally, these weights are used to blend the blurred image with its corresponding original. These weights control what percentage of each image is to be included in the final blended image. The gamma is a fixed weight that is added to all the pixels in image, and alpha & beta weights determines how original (alpha=1) or smoothed (out of focus looks like labels9(alpha=0.5)) are linearly combined from source image + blurred version respectively

For example, we can use the algorithm to diagnose and treatment diabetic retinopathy from analyzing retina images. It can be applied to image preprocessing before doing lesion/abnormality matching etc. Use of Gaussian blur operation to smoothen the image and reduce noise so that required features for diagnostic purposes stand out.

In the image enhancement phase, the selection of Gaussian blur parameters was crucial for optimizing feature extraction. The Gaussian blur kernel size and sigma values were determined through systematic experimentation, considering their impact on both noise reduction and preservation of DR-relevant features.

Based on this analysis, we selected a 3×3 kernel with σ=1.0 as optimal parameters. As shown in Table 1, this combination provides an effective balance between noise reduction and feature preservation, particularly crucial for early DR detection. The 3×3 kernel preserves fine vessel structures while the sigma value of 1.0 ensures sufficient smoothing without excessive loss of small lesions characteristic of early DR stages.

3.2.1 Histogram equalization

Histogram equalization is a technique used in digital image processing to improve the contrast and dynamic range of an image. The technique works by adjusting the intensity values of the image to span the full range of possible values, which can help to make the image more visually appealing and easier to analyze. The basic idea behind histogram equalization is to transform the image's intensity histogram into a more uniform distribution. The intensity histogram is a plot of the number of pixels in the image at each intensity value. In an ideal case, a histogram that is uniformly distributed would indicate that the image has a good contrast and all intensities are well represented.

Table 1. Effect of Gaussian blur parameters on feature preservation and noise reduction in DR images

Table 2. Impact of CLAHE parameters on image enhancement across DR severity levels

For example, we can use the algorithm to diagnose and treatment diabetic retinopathy from analysing retina images. For example, it can be applied to image preprocessing before doing lesion/abnormality matching etc. Use of Gaussian blur operation to smoothen the image and reduce noise so that required features for diagnostic purposes stand out.

2.jpg

Figure 2. Example of histogram equalization

CLAHE what stands for Contrast Limited Adaptive Histogram Equalization is a powerful digital image processing now which to tell the truth needs some pieces of work to take full advantages of. This is an extended version of the traditional histogram equalization which tends to over-amplify different parts of the image and thus it enhances contrast within regions but also increases their noise. Histogram equalization - CLAHE is a variant of histogram equalization, that constrains the contrast enhancement locally to prevent from amplification of noise. As the black background can affect the results of CLAHE, this algorithm checks for pixels in which there is at least one pixel with a black color and keeps it that way on the processed image.

Figure 2 demonstrates the effectiveness of histogram equalization in enhancing image contrast. The comparison shows both the visual effect on a sample image and the corresponding histogram distributions before and after equalization. The right histogram (old) shows an uneven distribution of pixel intensities, while the left histogram (new) displays a more balanced distribution across the available range, resulting in improved contrast and feature visibility. This enhancement step is crucial for highlighting subtle features that may be indicative of different stages of diabetic retinopathy. We then present the pseudocode of CLAHE in Algorithm 2.

The effectiveness of CLAHE in enhancing DR images heavily depends on two key parameters: grid size and contrast limit threshold. Based on extensive testing, as shown in Table 2, we selected an 8x8 grid size with a clip limit of 2.0, which provided optimal results across all DR severity levels. The 8x8 grid size offers sufficient local contrast enhancement for subtle DR features while preserving global image structure and maintaining balanced enhancement across different image regions. Additionally, the clip limit of 2.0 ensures controlled histogram equalization, prevents over-enhancement in high-contrast regions, and reduces noise amplification while maintaining lesion visibility. Smaller grid sizes (2x2, 4x4) led to over-enhancement of normal vessels and excessive contrast in lesions, while larger sizes (16x16) resulted in loss of fine vessel detail and reduced sensitivity to small lesions. Similarly, higher clip limits (3.0) introduced artifacts and over-enhanced lesions, particularly in moderate and severe DR cases.

3.2.2 Class imbalance mitigation

To address the class imbalance inherent in DR datasets. in Table 3, we implemented class weighting during model training. The weight for each class was calculated using the formula: $class_{weight}=\frac{total_{samples}}{\left(num_{classes\, samples_{{per}_{class}}}\right)}$. This approach assigns higher weights to underrepresented classes (Severe and Proliferative DR) and lower weights to overrepresented classes (No DR and Mild DR) during the loss calculation phase, effectively balancing the model's learning process across all DR severity grades.As shown in Table 3, higher weights were assigned to underrepresented classes such as Proliferative DR (5.71) and Severe DR (2.94), while lower weights were given to overrepresented classes like No DR (0.37) and Mild DR (0.81). This weighting scheme effectively balanced the model's learning process across all DR severity grades during the loss calculation phase.

Table 3. Calculated class weights for addressing class imbalance in DR grade classification

3.2.3 Ensemble based classification

Algorithm 3 uses a weighted sum approach to predict class labels for a set of testing feature vectors, based on the predicted probabilities of each class obtained from a set of classifiers. The algorithm initializes an empty array of weights and an empty array of final predictions. The weights for each classifier in the ensemble are determined through a performance-based optimization process on a validation dataset. Initially, each classifier is assigned equal weights (1/N, where N is the number of classifiers). These weights are then refined based on each classifier's performance on the validation set, specifically their accuracy in detecting different DR stages. For each classifier, a weight is calculated as the ratio of its classification accuracy for a particular class to the sum of all classifiers' accuracies for that class (weight[i]=classifier_accuracy[i] / sum(all_classifier_accuracies)). This creates a weight matrix where better-performing classifiers for specific DR stages receive higher weights, effectively allowing the ensemble to leverage the strengths of individual classifiers. The weighted probabilities are then combined during prediction to produce the final classification, with the weights dynamically adjusting the contribution of each classifier based on their demonstrated reliability for different DR severity levels.

For each classifier, the algorithm trains the classifier on the training feature vectors, uses the trained classifier to obtain the predicted probabilities for each class for the testing feature vectors, and saves the predicted probabilities in an array called predictions. For each testing feature vector, the algorithm initializes the array of weights with all elements set to zero and calculates the weights for each class based on the predicted probabilities obtained from the classifiers. The algorithm then finds the index of the class with the highest weight and appends the corresponding class label to the final predictions array. The output of the algorithm is an array of final predictions for the class labels of the testing feature vectors.

Diabetic Retinopathy is a growing cause of blindness in diabetic patients, and early detection and treatment are essential for effectively preventing DR-related disability. This paper aimed to investigate deep learning schemes on enhancing their ability in DR diagnosis as well as their superiority over traditional quantification methodology in recent decades. Specifically, the paper emphasized strengthening diagnostic ability using a Weighted Sum Ensemble method with ResNet50 as the basic architecture. Among the first experiments implemented on ResNet50 model without any modification, it was observed that prediction accuracy is inconsistent across DR stages, with relatively poor performance in predicting Mild and Proliferative conditions. The challenge here is that the phenomenology of these phases has subtle and diverse forms, not well-represented by standard processing methods.

After modifying the consensus model by implementing an ensemble method, performance improved for all stages of DR: AUC [No DR] increased from 0.87 to 0.90, and a particularly dramatic improvement in accuracy was achieved for detection of Severe DR (AUC=0.99). This suggests that the ensemble method is better at balancing class bias and improves feature extraction to enhance overall diagnostic accuracy.

The success of the ensemble model illustrates its potential application in real-world clinical practice where timely identification of any DR stage is vital for appropriate treatment interventions. It also provides evidence on how preprocessing and ensemble techniques become key to working around the limited performance of traditional CNNs in medical imaging cases characterized by high variability, simplifying otherwise cumbersome data.

While we acknowledge the importance of model robustness, several aspects remain to be explored in future work. These include evaluating performance under different illumination conditions, testing with varying image qualities, analyzing model robustness against different noise levels, and assessing performance with images from different equipment sources. Additionally, further advancement in these ensemble strategies will enable their application to different expert review of Medical Imaging tasks to improve health diagnosis outcomes.

Finally, for the detection of DR stages in retinal images, a good balance between computational complexity and diagnostic accuracy can be achieved using the weighted sum ensemble model. The study carries significant potential for advancing AI expertise in eye care and improving the diagnosis, management, and treatment of complications that diabetic practitioners encounter.