DenseExudatesNet: a novel approach for hard exudates detection in retinal images using deep learning

Diabetic Retinopathy (DR) affects light sensitive layer that is a prominent reason for visual deterioration. If the patient with Hard Exudates (HE) is not identified promptly, serious irreversible blindness is developed. By employing a computer-aided technique to detect the HE in retinal images early on, an ophthalmologist is able to accurately diagnose the problem of blindness. Recently various HE detection approaches have developed although, they have poor accuracy and high complexity. This research presents a novel DenseExudatesNet model to detect the HE in retinal images to overcome the above challenges and establish an enhanced outcome in the HE detection process. In this research, the data are gathered from the Indian Diabetic Retinopathy Image Dataset (IDRID) and DIARETDB1-Standard Diabetic Retinopathy Database (SDRD). A Dilated Attention ResNet50 is applied for extracting the spatial features. Further, a Channel-Spatial Attention Module (CSAM) is employed to extract the features in the spatial and channel dimensions and the attention is weighted. For detecting the HE, this study applies a DenseExudateNet. This model utilizes an AdaptiveDenseNet (ADN) that enhances the feature reuse capability and mitigates the gradient disappearance issues. To manage the complexity of the model, this research employs an Efficient Channel Attention (ECA). The experimental validation expressed that the proposed DenseExudateNet model accurately and effectively detected the HE in retinal images and achieved a higher accuracy of 98.86% and lower False Alarm rate (FAR) of 0.009%. This research paper concludes that the proposed model achieved better outcomes in the detection of HE when compared to the existing research analyses.