OCT-AttenNet: Developing An Improved Deep Learning Framework for Multi-Class Eye Disease Detection

Abstract

It becomes really difficult to work without a sight. We have to save our sight before its too late. For this we need early detection of diseases. We developed a novel OCTAttenNet model based on InceptionV3 and added BAM with ECA attention mechanism. We also applied several preprocessing and data enhancement techniques. Our proposed model OCT-AttenNet achieved an accuracy of 92% on a 10 class dataset collected from Bangladesh. It outperforms its backbone InceptionV3 by 2%. We did a comparative study of several CNN and transformers. Our proposed model outperforms all. We applied XAI like GradCAM++, IG to make it reliable to doctors and have a better understand of how the model is thinking. The model performed well with all diseases except early glaucoma and non-pathological myopia. The paper also covers how to improve this prediction.