A Single Photograph to Diagnose Diabetic Retinopathy: Artificial Intelligence in Ophthalmology
SciFocus/June 11, 2024 -- Diabetes, a global epidemic, is a chronic condition characterized by elevated blood sugar levels due to the body's inability to produce or effectively use insulin. This metabolic disorder can lead to countless complications, and while many are aware of its impact on overall health, few know that diabetes can lead to blindness. At first glance, the idea itself is disconcerting. Diabetes results from a lack of regulation of one’s blood sugar, how could this impact our eyes??
Simply put, a build-up of sugar in blood can block capillaries (the smallest blood vessels of the human body) that nurture the retina, which comprises light-sensitive layers of nerve tissue at the back of the eye that receives images and transmits them as electric signals through the optic nerve to the brain. The eyes often grow new blood vessels in an attempt to fix this. However, this might lead to retinal neovascularization, a pathological condition where the new blood vessel growth is abnormal and has varying effects on the patient’s vision. This is called Diabetic Retinopathy (DR).?
As this disease progresses, the abnormal blood vessels can cause leakage of blood into a region of the eye called the vitreous humor, a clear gel-like substance behind the lens of the eye. This prevents light from reaching the retina, preventing clarity of vision. That’s not all, if the new blood vessels prevent movement of fluid across the eye, the consequent buildup of pressure can damage the ocular nerve, causing glaucoma and blindness. One significant cause of vision loss from diabetes is glaucoma—a group of eye conditions that damage the optic nerve.
Nearly one-third of diabetic patients over 40 years of age develop diabetic retinopathy, with over 16% of them facing a threat to their vision. Despite its high risk, most patients are unaware of this complication and let symptoms progress until it's too late. It is vital to adopt a proactive approach to tackle DR, and understanding the intersection of diabetes and glaucoma is crucial. Artificial intelligence (AI) has taken huge leaps towards accomplishing this, and advancements in AI are increasingly being utilized to improve diagnosis and treatment outcomes for these interconnected health issues.
Traditionally, screening for this disease required a specialist doctor, a hospital, and equipment for retinal photography. Such necessities made DR checkups rare, or even non-existent in rural areas. However, this is no longer the case since the advent of AI in therapeutic modalities. A branch of AI called Deep Learning (DL) has seen immense growth over the past decade, especially in ophthalmology. Consequently, a single snapshot of an eye can now reveal not just the present condition but also predict the likelihood of diabetic retinopathy in a patient (Figure 1).?
AI-DL has reduced the DR screening process from a doctor’s consultation leading to a hospital visit for specialized tests to merely taking a picture of the patient's eye; specifically, a fundus photograph, an image of the back of the eye, and an evaluation of the same with AI.
Researchers at Phelcom Technologies have led this breakthrough through their use of AI to form a network called PhelcomNet. At the core of their digital architecture lie layers of algorithms and machine-learning techniques that make this possible.
So how does it work?
Upon receiving a fundus photograph, submitted by an operator, PhelcomNet starts to scrutinize every pixel of the image for signs of diabetic retinopathy. This step of processing the input image is done using a Convolutional Neural Network (CNN). A CNN is made up of layers of code that meticulously search the image for patterns, extracting features indicative of retinal damage, or diabetic retinopathy in particular.?
CNNs are trained in advance with information, usually as images, telling them which symptoms can be signs for diabetic retinopathy. These pieces of code replicate the functionality of the human brain by following a similar series of steps when interpreting information: ‘Perception’, ‘Feature extraction’, ‘Abstraction’, ‘Recognition’, and ‘Learning’.
In the case of diabetic retinopathy, the DL model searches for abnormalities in the blood vessels, looking for micro-aneurysms, abnormal growth, and other symptoms such as floaters in the eye. The model uses its layers to automatically learn and extract such features from the fundus photograph. Each image undergoes a transformation, becoming a canvas where patterns of hemorrhages (blood leakage), microaneurysms, or macular edema are clearly mapped out. PhelcomNet then matches these patterns to different severities of the disease, classifying and diagnosing the state of Diabetic Retinopathy for the provided Fundus photograph.
And it doesn’t end there! The process by which AI diagnoses such diseases is quite similar to that of a human. Thus, using a technique called GradCam, the researchers created heatmaps that reveal the underlying thought process that guides the decisions of the AI.
Figure 2 is a diagrammatic representation indicating the difference before and after the GradCam heatmap. The highlighted regions in Figure 2b indicate what the AI model classifies as “regions of interest”. This information serves as a visual aid, revealing the telltale signs of Diabetic Retinopathy and thus equipping healthcare providers with a deeper insight into the pathology of DR.
How does this help us?
This potent technology was packed into a hand-held device that is now used to diagnose patients completely independently. The FDA has also approved the use of AI for the automated diagnosis of DR using two fundus photographs of each eye, but the latest research has shown great accuracy with just one.
AI-based diagnosis systems for DR are a long-term investment. Although the development and production may be costly; with time, such devices have the potential to be much cheaper than traditional methods. This is primarily because they can analyze large amounts of images quickly, reducing the number of specialist referrals.?
领英推荐
The research team tested the precision of the AI against a human “specialist reader”, an experienced radiologist, and the AI showed that it was up to par, and thus has significant potential in the real world.
Showing high accuracy and sensitivity, the groundbreaking work by the researchers allows the screening of this disease to be accessible and easy. Rural areas with minimal access to doctors, hospitals, and medical equipment are now able to have large-scale screening camps with a single handheld device and one operator. This is especially promising for developing nations like India, where patients from rural regions may not be able to get proper screening for conditions like DR.
This has already begun in China, where operators utilize AI for DR screening. Thus, rising awareness and proactive early diagnosis of this disease may save thousands, even millions, from blindness and improve their quality of life.
References:
Authors: Sarthak Grover and Shalini Sanyal, Ph.D. * (*[email protected] ).
If you're enjoying this content, sign up to SciFocus daily newsletter by Biopatrika for all the biggest stories in Science, healthcare, STEM, startups, and how we're advocating to "Bring Science to Society." #SciFocus #Biopatrika
For partnership or advertisement opportunities, feel free to contact us here or visit www.biopatrika.com .
Career Opportunities: Biopatrika Career Hub https://www.dhirubhai.net/groups/13946324/
Follow Virender Singh and Biopatrika on LinkedIn.