Page 37 - AIH-1-4

Artificial Intelligence in Health Segmentation and classification of DR using CNN

1. Introduction to increase, emphasizing the importance of early detection
and intervention. To address this challenge, Singapore has
Diabetic retinopathy (DR) is a severe complication of implemented various initiatives, including nationwide
diabetes, posing a threat of blindness by damaging the DR screening programs, enhanced diabetic management
delicate blood vessels in the retina. This condition progresses protocols, and the integration of telemedicine technologies
through four distinct stages: mild non-proliferative for remote retinal imaging and diagnosis. Despite these
retinopathy, moderate non-proliferative retinopathy, efforts, barriers to accessing eye care services, especially
severe non-proliferative retinopathy, and proliferative DR. among underserved populations, remain a concern.
Each stage presents unique characteristics, complicating Continued investment in preventive measures, health-care
the diagnostic process, especially in the initial stage where infrastructure, and public awareness campaigns is crucial
warning signs are absent. 1 to effectively managing DR and mitigating its impact on
The potential to reduce new cases of DR by a substantial vision health in Singapore. 5
56% through timely treatment and monitoring emphasizes The stages of DR progress as follows:
the gravity of the situation. However, accurately identifying (i) Mild non-proliferative retinopathy: This is the earliest
the disease’s early stages remains a challenging task for stage of DR and is characterized by the occurrence of
clinicians, even those well trained in the field. Manual microaneurysms (MAs), which have a limited impact
examination of diagnostic fundus images for early-stage on blood vessels and cause minimal distortion.
detection is intricate, and the existing diagnostic methods (ii) Moderate non-proliferative retinopathy: Progression
are plagued by inefficiencies, resulting in disagreements to this stage involves the loss of blood vessels’ ability
among ophthalmologists and the provision of inaccurate to transport blood due to increased distortion and
ground-truth data for research purposes. In response swelling. As abnormalities in the blood vessels become
to these challenges, various algorithms have emerged more pronounced, the distortion and swelling hinder
to improve DR detection. Initially, these algorithms the normal transportation of blood, significantly
were grounded in classical computer vision approaches. impacting overall retinal health.
However, recent years have witnessed the rise of deep (iii) Severe non-proliferative retinopathy: This stage results
learning, with convolutional neural networks (CNNs) in a depleted blood supply to the retina. Increased
demonstrating their prowess in tasks such as classification blockage of blood vessels exacerbates the condition,
and object detection, including the diagnosis of DR. 2 prompting the retina to stimulate the growth of new
This research paper presents a novel approach to blood vessels in an attempt to compensate for the
addressing the complexities of detecting DR. The proposed reduced supply.
method, which employs transfer learning, leverages a (iv) Proliferative DR: This advanced stage is marked by the
single fundus photograph to automatically identify the proliferation of new blood vessels. Growth features
stage of DR. Notably, the approach is designed to learn secreted by the retina activate the proliferation of new
essential features from a dataset that is both limited and blood vessels. These vessels grow along the inside
noisy, presenting itself as a valuable screening tool within covering of the retina and extend into the vitreous gel,
automated solutions, as depicted in Hann et al. 3 filling the eye.
Highlighting the method’s effectiveness, the proposed 1.1. Related work
approach achieved a commendable ranking in the APTOS Numerous research endeavors have focused on the challenge
2019 Blindness Detection Competition, underscoring of early detection of DR. Initially, researchers explored
its capability with a high quadratic weighted kappa score classical computer vision and machine learning methods
of 0.92546. This research aims to significantly advance to develop viable solutions. For example, Priya and Aruna
6
DR detection methodologies, particularly in automated proposed a computer vision-based approach using color
systems, addressing the critical need for early diagnosis fundus images for DR stage detection. Their methodology
and intervention in the fight against DR. 4 involved extracting features from raw images through
DR is a significant concern in Singapore due to its image processing techniques, which were subsequently
high prevalence among individuals with diabetes mellitus, fed into a support vector machine for binary classification.
affecting approximately one in nine adults in the country. They achieved performance results on a testing set of 250
Globally, DR is a leading cause of vision loss and blindness, images, with a sensitivity of 98%, a specificity of 96%, and
imposing a substantial burden on Singapore’s health-care an accuracy of 97.6%. In addition, researchers have also
system and economy. With Singapore’s aging population explored other models for multiclass classification. For
and rising rates of diabetes, the incidence of DR is expected example, employing principal component analysis on

Volume 1 Issue 4 (2024) 31 doi:10.36922/aih.2783