A novel hybrid model for automatic diabetic retinopathy grading and multi-lesion recognition method based on SRCNN & YOLOv3

Automatic grading and lesion identification of diabetic retinopathy (DR) is important for researchers because it is the leading effect of diabetes. Due to diabetes, the tiny blood vessels within the fundus are damaged and multiple lesions such as microaneurysms, hemorrhages, hard exudates, and soft exudates appear in the retina and cause multiple vision-related complications, which can drive to total vision loss without early examination and treatment. For clinical screening and diagnosis of DR, retinal fundus images are commonly used. Fundus images taken by operators with different levels of experience, however, have a broad variance in quality. Low-resolution images of the fundus raise the risk of misdiagnosis which makes it more difficult to observe clinically. In order to avoid low resolution fundus images and to be able to diagnose DR carefully, a new hybrid structure is developed in our proposed system to ensure that DR detection and classification processes become much more precise and faster compared with existing models. In the image pre-processing stage, the proposed model adopts a super-resolution convolutional-neural-network to enhance the pixel density of low-quality fundus images. In the next step, to identify the DR grade, an advanced deep-learning model called You-Only-Look-Once-Version 3 is used. Finally, another You-Only-Look-Once-Version 3 network stage is applied using a bounding box to recognise the multiple lesions in the fundus images. The proposed system is evaluated on an openly accessible MESSIDOR dataset, and the results show that the system achieves 96.89% overall accuracy for DR grading and 97.6% accuracy for lesion detection with a high detection speed of 5.6 s.