An automated analytics approach for diabetic retinopathy detection with ensemble deep learning models in healthcare

Diabetic Retinopathy (DR) is a leading complication of prolonged diabetes, which poses a significant threat to vision and may lead to permanent blindness. Early identification and timely intervention are crucial to preventing disease progression. Traditionally, DR diagnosis relies on medical examination of retinal fundus images by expert ophthalmologists, which is time-consuming and resource intensive. However, deep learning techniques, particularly medical imaging, have demonstrated remarkable performance in the automated detection and classification of DR. This study proposes an ensemble-based deep learning framework using feature-level fusion stacking, which integrates four complementary convolutional neural networks named ReXInDen and three complementary convolutional neural networks named ReXDen for automated DR detection from retinal fundus images. These frameworks extract high-level features from each backbone, concatenate them into a unified representation, and classify using a feedforward neural network. Three datasets were utilized to validate the model including a region-specific dataset collected from Bangladeshi medical sources. The proposed ReXInDen model achieved accuracies of 98.27%, and 98.69% on Dataset 1 and Dataset 2, while ReXDen achieved the highest accuracy of 99.05% on Dataset 3. These results indicate a substantial improvement over individual models and demonstrate the potential of the ensemble approach to support early-stage DR detection. Moreover, these models show promise for integration into automated DR screening tools that can aid in reducing the global burden of diabetic vision loss. • Develop an ensemble method combining different convolutional models for medical image-based diagnosis. • Apply advanced image filtering techniques to improve retinal image clarity before analysis. • Merge features from multiple models and classify them using a feedforward neural network. • Validate the method using three datasets, including one from local healthcare providers. • Achieve 99.05 percent accuracy with the ensemble method on the combined dataset.