Blood Vessel Segmentation in Retinal Images Using Machine Learning Approach

Abstract

A segmented vessel network can be beneficial for the diagnosis, therapy planning, coordination, and evaluation of eye-related illnesses such as glaucoma, vein occlusions, and diabetic retinopathy (DR). Since manually segmenting vessels is a time-consuming and challenging task, many ways for autonomously segmenting retinal blood vessels have been presented over the years. However, most known retinal vascular segmentation algorithms still have limitations such as low generalization capacity and poor accuracy due to a lack of consideration given to dataset preparation and processing. This research offers a fully supervised method for segmenting and extracting blood vessels from retinal fundus images using machine learning techniques along with appropriate data processing and dataset enhancement strategies to obtain a robust framework and achieve better performance while reducing computation time. The proposed method has two main components: Extracting feature maps from modified U-net and Segmenting the images using Multilayer Perceptron (MLP). We tested the framework quantitatively and qualitatively on three publicly available data sets, STARE, DRIVE, and HRF. The results were compared to ground truth images and other methodologies from previous research. The framework received an average accuracy of 99.78%, 98.34%, and 98.85% on these datasets, respectively.