Enhanced Convolutional Neural Networks for Segmentation of Retinal Blood Vessel Image

Abstract

Segmentation of fundus retinal blood vessels is one of the most challenging tasks in computer vision because of the difficulty in precisely capturing every minute detail in the tiny vessels necessary for accurate and early diagnosis and prognosis. These problems are attributed to noise, artifacts, low contrast associated with medical images especially fundus retinal images as a result of uneven illumination and the approach used in acquiring them. Other difficulties are the interference and distinction of pathologies (background) from blood vessels (foreground) during segmentation. To address these problems, this paper proposes an Enhanced Deep Convolutional Networks for Segmentation of Retinal Blood Vessel to explore the availability of huge channels and usage of global location and context in the U-net model. Enhancement techniques are also applied to input images at data pre-processing stage to enhance brightness and visibility before passing them to CNN for segmentation. The combination of these two effective tools will boost the sensitivity accuracy since it is a key factor that echoes a truthful valuation of blood vessel pixels, which is the principal goal in vessel segmentation for fundus image analysis. Our proposed method is evaluated on the Digital Retinal Images for Vessel Extraction (DRIVE) dataset and the performance achieves the state-of-the-art for segmentation of retinal blood vessels. with accuracy of 94.47, sensitivity of 70.92, specificity of 98.20 and area under the ROC curve of 97.56.