Improving Diabetic Retinopathy Classification by Employing Convolutional Block Attention Mechanism in Modified InceptionV3 Model

Diabetic retinopathy (DR) is a serious effect of diabetes in human eye, which is considered as a key reason of blindness globally. Early detection and accurate staging of DR has great importance for timely intervention and treatment plan. In this present, we have proposed a framework for categorizing five classes of diabetic retinopathy that combines the convolutional block attention module (CBAM) mechanism with the strong InceptionV3 deep learning architecture. The CBAM module dynamically increases informative regions inside feature maps, allowing the model to concentrate on relevant image regions while suppressing irrelevant ones and hence resulting higher classification performance. The customized InceptionV3 network is used as a feature extractor to collect hierarchical representations of retinal images. The publicly available widely used APTOS dataset is used to test and train the proposed framework. Dataset balancing has done using random sampling technique. The combination of the CBAM attention mechanism with InceptionV3 shows potential for improving automated DR diagnosis systems from its base model with a noteworthy accuracy of 95.37%, which is also compatible with related recent works. The proposed framework might be useful in detection, staging, and management of DR related vision-threatening illness.