Dynamic Convolutional Attention for Classification of Diabetic Retinopathy

Deep learning algorithms in medical image processing have showed considerable promise, notably in the identification of DR. However, due to the disease’s complexity and variability, precisely identifying DR remains difficult. A recently developed deep learning approach known as DCA has shown promising outcomes in a range of image classification applications, including medical imaging. The DCA model excels at recognising essential parts and patterns by using attention mechanism to focus on significant areas and exclude irrelevant ones. Our findings add to the body of knowledge about deep learning approaches in medical picture processing. We discovered that combining DCA approaches with ResNet-50 and ensembling it with other state of art models improved DR classification accuracy considerably. The ensemble model, with weights assigned to each architecture, outperformed the standalone ResNet-50-DCA model’s accuracy of 73.59%, precision of 75.52%, and recall of 74.37%. The ensemble technique maximises performance while minimising weaknesses by harnessing the strengths of diverse models.