Abstract
Glaucoma is an ocular disease that deteriorates the optic nerve of the eye. It happens when the pressure
in the eye increases over time due to a blockage in the
ow of the aqueous humour. It is estimated that
about 76 million people had the disease globally in 2020, and approximately 111.8 million people will be
a ected by 2040. The damage caused by the condition is irreversible and can lead to blindness if not
treated early. Furthermore, most people with the disease don't have obvious symptoms or pain. Hence,
the need to have systems that can detect the disease early.
Detection of glaucoma starts with a careful segmentation of the areas of interest from the fundus images
(the image of the inner eye). These areas of interest include the optic cups, optic discs and the blood
vessels. First, we present a comprehensive review of the existing methods of fundus images segmentation.
The existing segmentation techniques can be grouped into two wide classes: techniques based on pixel
classi cation and techniques based on deep learning architectures. The pixel classi cation techniques
need a lot of human intervention to tune the algorithm's parameters to t the fundus image being
processed. Also, the segmentation technique is based on the assumption that the brightest spot on the
fundus image is the optic nerve head. However, this is not true for some fundus images with lesions or
haemorrhages. The deep learning segmentation techniques do not require human intervention but are
usually very cumbersome and incur extra training cost. A segmentation method that overcomes the cons
of both techniques is proposed. The proposed algorithm: U-Net Lite, does not require human intervention
and does not have the underlying assumption that the brightest spot on the fundus image will always be
the optic nerve head. Therefore, the novel U-Net Lite model is adequate for segmenting fundus images
with lesions or haemorrhages.
The proposed U-Net Lite model is designed to have very few parameters (6.8 x105), which is about 45
times fewer than the traditional U-Net model. Also, the U-Net Lite model is designed to have a training
time which is about 10 times less than the conventional U-Net model. The conventional U-Net model
has a training time that ranges between 2 to 10 hours depending on the speci cations of the machine
v
used, while the U-Net Lite model has a training time of 30 minutes to 1 hour. This reduces the cost
of training and allows the model to be deployed easily on the web and for mobile applications (since
the size is just about 6 MB). The novel U-Net Lite model achieves state-of-the-art results in the optic
discs, optic cups and blood vessels segmentation. The novel model performs better than the traditional
U-Net model, the Visual Geometry Group (VGG-16) model and the Fully Connected Network (FCN)
models. The novel model achieves dice scores of up to 0.99 in the optic discs and cups segmentation.
For the blood vessel segmentation process, the novel model achieves sensitivity and speci city of 0.8059
and 0.9826, respectively, on the High Resolution Fundus (HRF) database. The achieved speci city and
sensitivity is above the widely accepted score for a good model. A model with a speci city and sensitivity
score greater than 0.75 is generally considered good. The novel method segments 1,000 fundus images in
about 20 minutes while it takes a typical segmentation model about 3 hours.
Glaucoma has been detected from segmented fundus images using the Cup-to-Disc Ratio (CDR) and the
Inferior Superior Nasal and Temporal (ISNT) rule methods. In the CDR method, a speci c threshold
value is used to classify fundus images as glaucomatous or non-glaucomatous. Fundus images with CDR
values below the chosen threshold value are classi ed as non-glaucomatous, and fundus images with CDR
values above the threshold value are classi ed as glaucomatous. The major drawback of this method is
that the CDR value to be used as a threshold is subjective. CDR values that have been used in the
literature range from 0.3 to 0.6.
Furthermore, a fundus image may have a high CDR value (e.g. 0.7) but not be glaucomatous. This can
be due to hereditary factors. The ISNT rule uses the size of the neuro-retinal rim quadrants to detect
glaucoma. However, the method is unstable and does not detect glaucoma effectively.