Diabetic retinopathy detection using modified U-Net architecture and artificial metaplasticity algorithm

Diabetic retinopathy (DR) is an ocular disease caused by long-term diabetes. It is the leading cause of blindness among working-age people. DR is an asymptomatic pathology, that causes changes in the area, diameter, and tortuosity of thin blood vessels. Therefore, analyzing the morphological changes of the retinal vasculature is critical for helping ophthalmologists in the diagnosis of DR at an early stage and preventing vision loss. In this work, we present an automatic approach for diabetic retinopathy screening from color retinal images based on semantic segmentation of retinal blood vessels. First, a preprocessing and data augmentation steps are integrated to extract the green channel and increase the number of images. Then, the segmentation step is carried out using a modified U-Net network. After that, a set of geometrical parameters are calculated from segmented images such as area, perimeter and tortuosity. Finally, Artificial Metaplasticity Multi-Layer Perceptron 'AMMLP' classifier is employed to discriminate between healthy and DR images. The proposed method for vessel segmentation was evaluated on the DRIVE database, and it achieved an accuracy of 99.65%. For DR classification, AMMLP got a sensitivity of 98.45% on Messidor dataset. The experimental results show that our approach outperforms the existing methods, which can assist ophthalmologists in making an early and accurate DR diagnosis.