A New Deep Learning model for Retinal Blood Vessels Segmentation

The blood vessels of the eye help doctors to detect various diseases such as Cataract, Hypertensive retinopathy, Diabetic retinopathy, and Glaucoma. To extract important information from retinal images, they must be segmented accurately. The segmentation process by an expert is expensive and slow, so it is necessary to automate retinal segmentation for its high speed and accuracy. Artificial intelligence, especially deep learning, has been used to segment the retina and the results were good, but these techniques must be developed. In this paper, we propose using two paths, Encoder and Decoder to segment the retina. In the encoder path, we used the new architecture XcelNet17 to extract important features in the retina, and in the decoder path, we used a self-attention block to enhance the final binary image. We used the DRIVE (Digital Retinal Images for Vessel Extraction) database, and the accuracy and F1-Score values were 98.05% and 98.74% respectively.