Abstract
Diabetic retinopathy is a major source of blindness among individuals with diabetes. It is a global health concern affecting millions of people. Traditional diagnostic methods for treating this disease are effective but often time-consuming. They rely heavily on manual analysis which can introduce human error and subjectivity. Additionally, most of the existing approaches depend on handcrafted features which can limit their accuracy and generalization. Thus in order to tackle these challenges, this research aims to develop an automated and effective approach for accurate and early detection of this disease with a focus on reducing both time and errors in diagnosis. In this study, we proposed an ensemble classification framework utilizing transfer learning and the concept of data fusion for more accurate detection of diabetic retinopathy. This research focuses on two types of classification, multi-class and binary. We developed a custom convolutional neural network model named DRSC-Net. We also fine-tuned three pre-trained models and integrated them into the framework. The pre-trained models used are EfficientNetB2, ResNet50, and VGG19. These models together with DRSC-Net formed a stacked ensemble model. By leveraging their combined strengths, we improved classification performance. We used three benchmark DR datasets: APTOS2019, IDRiD, and Messidor-2. These datasets were fused into a single dataset for training. To enhance image quality, we applied preprocessing techniques such as Contrast Limited Adaptive Histogram Equalization (CLAHE) and Gaussian blur. CLAHE improved image contrast while Gaussian blur helped in noise reduction. Data augmentation techniques like zooming, flipping and many other were used to address the problem of class imbalance and to increase the dataset's diversity. For classification, we built a stacked ensemble learning model. It combined predictions from all the individual models to enhance accuracy and generalization. The ensemble model achieved outstanding results with a multi-class classification accuracy of 99.33% and a binary classification accuracy of 99.92%. The model also outperformed the existing state-of-the-art methods on comparison. This demonstrates its potential as a reliable and efficient solution for diabetic retinopathy detection.
Replies