India is a diabetic capital of the world, with the prevalence of diabetes in adults being 8.9% in the rural region. Diabetes may cause irreversible damage to the retina, resulting in vision impairment and sometimes even blindness. In fact, reports state that the total number of people with diabetes is projected to rise to 439 million in 2030, with four out of ten diabetic individuals requiring some kind of eye treatment. Thus, it requires early detection and treatment that can stop the damage.
However, the lack of trained eye doctors to identify diabetic retinopathy affects the care, especially in the rural and semi-urban areas, which involves an enormous scarcity of experienced eye specialists. To resolve that issue, Bengaluru-based Sankara Eye Hospital decided to collaborate with Singapore-based Leben Care, to deploy a cloud-based AI software platform — Nethra AI in order to diagnose retinal conditions in patients.
To understand the solution and its benefits in-depth, we got in touch with the Ocularists of Sankara Eye Hospital, Dr Divyansh K Mishra. “Eye examination provides an effective clue for detecting eye disease. Diabetic eye disease can cause permanent loss of visions, including blindness,” said Dr Mishra.
Tech Behind Nethra AI
Developed by Leben Care, Netra.AI is a comprehensive retina risk assessment software-as-a-service platform that is available over the cloud. Explaining the tech, Gurunath Bhuvanagiri, the chief architect and AI scientist at Leben.ai stated, patients’ images of the fundus of the eye are captured using cameras and are then uploaded on to the machine learning-based platform, which can receive these anonymised patient data either via a web portal or through API.
“With the retinal images captured, the AI-based solution Nethra AI can diagnose retinal conditions like diabetic retinopathy, age-related macular degeneration, glaucoma and various other retinal pathologies requiring referral to an eye doctor,” said Bhuvanagiri.
The solution uses cutting edge algorithms, developed in guidance with some of the leading experts in retina globally, with a four-step deep convolutional neural network (DCNN). It helps in detecting retinal photographs from non-retinal images; detecting generic quality distortion for automated image quality assessment; detecting diabetic retinopathy (DR) stage; and annotating the lesions based upon pixel density in the fundus images.
These lesions could be microaneurysms, hard exudates, cotton wool spots, superficial and deep haemorrhages, neovascularisations and fibrovascular proliferation.
Further, factors like level of operator’s expertise, type of equipment used, patient conditions, low-quality image retrieval can lead to incorrect analytics. Therefore it was critical to creating an algorithm that can automatically assess the quality of the fundus image for reliable lesion detection for an AI-based screening system.
Traditional image quality assessment algorithm relies on handcrafted features, based on either generic or structural quality parameters, such as global histogram features, textural features, vessel density, local non-reference perceptual sharpness etc. These do not generalise well and work on a different dataset, cameras and field of view (FOV), resulting in biases. However, the image quality assessment algorithm in Nethra AI is based on a convolution neural network, which uses a custom 17-layer CNN network to train the data. The training dataset consisted of 103,578 good quality (gradable) retina images and 8911 not-so-gradable retina images.
Additionally, “data augmentation was used to avoid overfitting and making the model more robust. About 72% of the training images were from mydriatic, and the remaining 28% were from non-mydriatic cameras,” said Bhuvanagiri.
While grading of diabetic retinopathy, the images were graded using the international clinical DR severity scale (ICDRS), both for training and validation. The grading included — No DR; Nonproliferative DR, which can be mild, moderate or severe; and Proliferative DR, which can cause retinal/vitreous haemorrhage.
The model has been tested on a dataset consisting of 1533 independent images, reviewed by retina specialists, which were not assigned patient-wise and not overlapping with the training dataset. Alongside the statistical analysis was done using Microsoft Excel 2016 and MedCalc Statistical Software version 18.11.6.
According to Dr Mishra, the solution also triages the urgency for the referral of the patient. “Such advancement has helped in bridging the gap of the sparse availability of the trained eye specialist at various levels, from a medical diagnostic centre for a blood test to the rural or semi-urban areas. This can enable millions across our country, getting the much-needed treatment in time.”
Nethra AI also comes with advanced tools for blood vessel segmentation and image post-processing that helps doctors with more significant insights and accurate diagnosis. Nethra Track enables the doctors to manage and view retinal changes over time. It also maintains a record of all the retinal images to automatically detect and notify the changes in the patient. Such an additional tool provides better visualisation and control over the disease progression.
Till date, Netra AI has screened 3000+ patients worldwide and identified 5% at-risk patients. “The solution, Nethra AI is currently available in both online and offline modules like a stand-alone box,” said Dr Mishra.
With Nethra AI, the report generated not only provides timely identification and referral of the individual patients but also allows the doctor to educate the patient using the photo of the retina mentioning lesions present, thereby reinforcing the need for treatment to avoid blindness or loss of vision.
The proposed solution also showcased excellent sensitivity and accuracy while detecting any DR, i.e., 99.7% and 98.5% respectively, thus making it suitable for an accurate screening tool.