Paper Key : IRJ************677
Author: Aryan,Piyush Bandal,Vikram Kumar,Siddhesh Dhumal, Prof P. B. Ekatpure
Date Published: 01 Feb 2025
Abstract
The improper circulation of flow of blood inside the retinal vessel in the body is the primary source of most of the optical disorders including partial vision loss and blindness. Accurate blood vessel segmentation of the retinal image is used for biometric identification, computer-assisted laser surgical procedure, automatic screening, and diagnosis of eye diseases like diabetic retinopathy, age-related macular degeneration, hypertensive retinopathy, and so on. Proper identification of retinal blood vessels at its early stage helps medical experts to take convenient treatment procedures which could reduce vision loss. Automatic and proper retinal blood vessel segmentation helps to solve various optic diseases. As the number of patients and the necessity of the vessel segmentation is increasing day by day, an automated system is an alternative to the manual system. Retinal blood vessels have an important role in the diagnosis and treatment of various retinal diseases. For this reason, vasculature extraction is important in order to help specialists for the diagnosis and treatment of systemic diseases. For segmentation various machine learning methods are available such as Support Vector Machines (SVM). But deep learning models perform better than traditional machine learning algorithms like SVM at segmentation tasks. Currently various deep learning models are available such as fully convolutional networks, encoder- decoder based models. U-Net and V-Net are two popular image segmentation architectures used in biomedical image segmentation. In an attempt to provide a highly accurate retinal blood vessel segmentation method, this project includes experiment with transfer learning approach. VGG- 19 is used as a pre-trained encoder for the U-Net model. The objective of the project is to study the impact of transfer learning on retinal blood vessel segmentation. The layers from the encoder section are frozen selectively in layer-by-layer manner. After each layer is frozen the model is trained and statistics are recorded. Using the recorded statistics, the impact of transfer learning is measured.