Segmentation of medical images using U-Net++

Abstract

Many eye diseases can be detected by segmenting retinal vessels, and their accurate computerization is becoming increasingly important in automated retinal disease screening systems. The segmentation of retinal vessels can be performed using many algorithms, but these methods improve accuracy levels. Low contrast vessels require more attention in the segmentation process due to low sensitivity due to improper segmentation. A new pre-processing step is proposed in this paper to extract retinal blood vessels with greater precision. Observing their impact on other existing algorithms is another aspect of testing these preprocessing steps. To segment retinal blood vessels, we suggest two steps. In the first step, the preprocessing module must be implemented and validated. We then apply these preprocessing stages to the proposed binarization steps to extract retinal blood vessels. This preprocessing phase provides a much-improved segmented vessel image using the traditional image-processing method. The image coherence technique was used to binarize retinal blood vessels. Identification and localization of retinal vessels separate anatomical structures from background fundus images, such as optic discs and maculae. Due to the noninvasive fundus imaging and the crucial information in the vascular network, retinal vessel identification studies are excellent for detecting and diagnosing diabetic retinopathy, glaucoma, hypertension, and age-related macular degeneration (AMD). As computer-aided techniques for segmenting retinal vessels have developed over the past two decades, they have become increasingly important and are becoming routine in clinical practice. Additionally, retinal vessel segmentation results are evaluated and validated. A final objective assessment of retinal vessel identification techniques is presented, along with a discussion of future developments and trends.