Improved Retinal Vessel Segmentation

The retina is the part of the body located near the optic nerve, which allows visualization. Color fundus images are captured from the fundus camera using different angles. Many diseases like diabetic retinopathy, glaucoma, hypertension, diabetes, and coronary heart disease can be detected from the

2025-06-28 16:27:48 - Adil Khan

Project Title

Improved Retinal Vessel Segmentation

Project Area of Specialization Artificial IntelligenceProject Summary

The retina is the part of the body located near the optic nerve, which allows visualization. Color fundus images are captured from the fundus camera using different angles. Many diseases like diabetic retinopathy, glaucoma, hypertension, diabetes, and coronary heart disease can be detected from the changes in the vascular system of the fundus image. Manually inspecting the vessels in the fundus image takes a long time for an ophthalmologist. They can spot anomalies in vascular structure more readily with improved automated vessel segmentation. In low-income areas, finding an ophthalmologist is difficult. Computer-assisted diagnostic methods, which are less time-consuming, are more advantageous to these communities.

Because the fraction of thin vessels in the retina is limited, it is challenging to segment them due to poor contrast and lesion regions. Furthermore, in deep learning approaches, networks pay greater attention to segmenting thick vessels because thin vessels' segmentation accuracy is less affected. As a result, a framework is developed in which thick and thin vessels are trained individually. Two additional training data sets were prepared. With three distinct prediction probability maps, three different outputs will be generated. We will not utilize a basic threshold to achieve the final result; instead, we will employ a fusion approach.

About one in three people with diabetes who are older than age 40 already have some signs of diabetic retinopathy [1]. Automated retinal image analysis is an important diagnostic tool for this purpose because more accurate and efficient results are achieved. By detecting diabetic retinopathy at early stage vision loss can be prevented. This will be achieved by creating two different pipelines for both thick and thin vessels.

Project Objectives

Blood vessels with varying contrast have thick and thin vessels that need to be extracted with high accuracy. Changes in the vessel tree structure can indicate a variety of eye abnormalities.

Andriod Application will assist opthomologists in order to detect different diseases like diabetic retinopathy more easily using this application.

Project Implementation Method

A deep learning approach for segmenting retinal vessels has been suggested. We do multiple objective optimizations for the training goal by separating thick and thin vessels. From three datasets of thick, thin and original, three distinct prediction probability maps were created. A smaller network was applied to locations that were mistakenly identified also. Finally, the fusion process was applied to all prediction probability maps to get the final binary segmentation image output.

Network Architecture:

UNET is a more advanced version of convolutional neural networks used initially and built-in in 2015 to segment biomedical images. Every pixel in the image is classified. It is comprised of an encoder and a decoder. To extract different features, convolution is used. The contracting process occurs on the left side of the U using Max pooling, while the expanding route is formed by transpose convolution layers on the right. U-net additionally employed skip-connection to concatenate characteristics of the respective encoder and decoder layers on the channel dimension, allowing deep semantic and shallow representation information to be integrated to improve segmentation results.

Network Architecture:

UNET is a more advanced version of convolutional neural networks used initially and built-in in 2015 to segment biomedical images. Every pixel in the image is classified. It is comprised of an encoder and a decoder. To extract different features, convolution is used. The contracting process occurs on the left side of the U using Max pooling, while the expanding route is formed by transpose convolution layers on the right. U-net additionally employed skip-connection to concatenate characteristics of the respective encoder and decoder layers on the channel dimension, allowing deep semantic and shallow representation information to be integrated to improve segmentation results.

Separation of Thick and Thin Vessels:

We match a threshold width for each pixel in blood vessels; if the vessel width is less than the threshold, it is categorized as thin; otherwise, it is labeled thick. To calculate the vascular width of each pixel, we define twice the distance from the white pixels to the nearest black pixels. The vascular width of skeletal pixels along centerlines is then determined.  The vascular width of the skeleton pixel closest to them is employed. Later steps include removing spur pixels and small objects from a binary image.  The skeleton image's branch points are dilated and eliminated from the original, skeleton, and threshold images, respectively. To achieve the final output, several AND OR operations are performed.

Benefits of the Project Technical Details of Final Deliverable Final Deliverable of the Project Software SystemCore Industry MedicalOther Industries IT , Health Core Technology Artificial Intelligence(AI)Other Technologies Big DataSustainable Development Goals Good Health and Well-Being for PeopleRequired Resources
Item Name Type No. of Units Per Unit Cost (in Rs) Total (in Rs)
Total in (Rs) 75000
D-Eye Portable Ophthalmoscope Equipment17000070000
Consultion with opthomologist Miscellaneous 150005000

More Posts