A novel approach for early detection and grading of diabetic retinopathy by using ensemble model

Diabetic retinopathy (DR) is a one of the causes of blindness of the aged 20 to 79 adults, damaging retinal blood vessels. Early detection using traditional methods is challenging. However, machine learning algorithms can improve DR detection and grading. This study combines features of deep learning models such as CNN, ResNet152v2, and VGG16 for enhanced accuracy for the IDRiD dataset. Image processing techniques such as colour negation, greyscale conversion, circular crop, and Gaussian blur filter are employed. Processed images are analysed by a 13-layer custom CNN, ResNet152V2, and VGG16. The CNN achieves 92% accuracy, ResNet152v2 90%, and VGG16 93%. An ensemble model combining these three yields a 95% accuracy. This method offers a reliable DR detection approach, beneficial for medical organisations with limited ophthalmology specialists.