Contrast Enhancement in Deep Convolutional Neural Networks for Segmentation of Retinal Blood Vessels

Abstract

The segmentation of blood vessels from the retinal fundus image is known to be complicated. This difficulty results from visual complexities associated with retinal fundus images such as low contrast, uneven illumination, and noise. These visual complexities can hamper the optimal performance of Deep Convolutional Neural Networks (DCNN) based methods, regardless of its ground-breaking success in computer vision and segmentation tasks in the medical domain. To alleviate these problems, image contrast enhancement becomes inevitable to improve every minute objects’ visibility in retinal fundus images, particularly the tiny vessels for accurate analysis and diagnosis. This study investigates the impact of image contrast enhancement on the performance of DCNN based method. The network is trained with the raw DRIVE dataset in RGB format and the enhanced version of DRIVE dataset using the same configuration. The take-in of the enhanced DRIVE dataset in the segmentation task achieves a remarkably improved performance hit of 2.60% sensitivity, besides other slight improvements in accuracy, specificity and AUC, when validated on the contrast-enhanced DRIVE dataset.