A Novel Deep Learning Approaches for Early Detection of DR and DME Using Fundus and OCT Images

Abstract

Diabetic patients suffered from various eye diseases such as Diabetic Macular Edema (DME) and Diabetic Retinopathy (DR) which causes optic disc damage and vision loss. Hence, early diagnosis of DR and DME is necessary for active treatment, but the slow movement of disease signs leads to difficulty in early detection. For early detection and severity grading of DR and DME, we proposed a novel deep learning approach. This research includes three major processes namely image quality assessment, bi-level segmentation, and multi-feature-based DR and DME detection. In first process, the image quality is estimated by generating threshold, if it has less quality then it performs preprocessing and data augmentation, otherwise it directly forward to the data augmentation process for increasing training efficiency. In preprocessing, we perform two process such as noise filtering and contrast enhancement. Noise removal is performed by Adaptive Threshold based Bilateral Filter (ATBF) which provides noise removed edge preserved image which increases the resolution of the image. For contrast enhancement, we proposed Honey Badger Optimization (HBO) which increases the image brightness by adjusting the intensity of the images. After completed image quality assessment, we perform image segmentation by using U-Net³⁺ which segments both optic disc and lesion by considering various features. The accuracy of segmentation is evaluated by Chi-square which reduces the false positive rate. Finally, feature extraction and DR and DME detection is performed by Attention based Hybrid SqueezeNet-VGG16 (AHS-VGG16) which extracts the features from segmented region and classified the images into three classes such as normal, DR and DME. The simulation of this research is conducted by Matlab simulation tool, and the performance of this research is evaluated based on different performance metrics by considering four types of datasets namely, Indian Diabetic Retinopathy Image Dataset (IDRID), Retinal Fundus Multi-Disease Image Dataset (RFMiD), Fundus and Fundus and OCT images, MESSIDOR