AI in Diabetic Retinopathy

The primary goal of using AI for diabetic retinopathy (DR) is to screen for proliferative diabetic retinopathy (PDR) and diabetic macular edoema (DME), the two leading causes of significant vision loss in individuals with DR. For the majority of algorithms, the most significant predictor is the detection of referable diabetic retinopathy (RDR). RDR is defined as moderate nonproliferative diabetic retinopathy (NPDR) or greater and clinically significant macular oedema (CSME). Sight-threatening DR (STDR) is defined as the presence of severe NPDR, proliferative PDR, and/or DME.

• Diabetic patients’ screening for DR changes has developed into an excellent approach of preventing blindness. AI-based screening techniques have been verified and proven to be highly accurate in discriminating between referable and non-referable DR.
• Gulshan and colleagues from Google AI Healthcare described a deep learning system that performed very well in diagnostic tasks. Their approach evaluated 128,175 retinal pictures for DR and DME by a panel of 54 ophthalmologists and ophthalmology residents in the United States. Seven board-certified ophthalmologists assessed a test data set of about 10,000 photographs obtained from publically accessible sources. The area under the receiver operating characteristic curve (AUC) for both datasets was close to 0.991.

Another research employing a different DL method produced an AUC of 0.980, with a sensitivity and specificity of 96.8 and 87.0 percent, respectively, in detecting referable DR. These studies demonstrate the potential use of DL in early detection of referable DR. Ting et al. conducted a large investigation in Singapore to validate DL using several retinal pictures acquired with conventional fundus cameras. This research demonstrated a high degree of sensitivity and specificity for DR detection. The possible hurdles and uncertainties include validating these DL systems in real-world DR screening programmes and determining their generalizability when applied to populations of various ethnic origins and employing retinal pictures obtained by various fundus cameras.

• IDx is the first AI device authorised by the US Food and Drug Administration for use in DR screening in 2018. This gadget is equipped with a Topcon NW400 fundus camera, which is used to upload photos to the programme. If the DR is more than mild, the condition is referred to an ophthalmologist. If it was not’more than moderate DR,’ the programme contacts the patient 12 months later for rescreening. Sensitivity and specificity of 87.3 percent and 89.5 percent, respectively, were observed in a multicenter study including 900 persons with diabetes. Because the gadget is capable of making a screening judgement, it can be used by individuals who are not ophthalmologists.

EyeArt by Eyenuk was utilised to train the AI algorithms for DR screening using the EyePACS tele-screening system and demonstrated a sensitivity and specificity of 90% and 63.2 percent, respectively. Additionally, it identified microaneurysms with a sensitivity of 100%. The algorithm analysed 40,542 pictures obtained from 5084 patients. Tufail et al. found that EyeArt was 94.7 percent sensitive for any DR, 93.8 percent sensitive for referable retinopathy, and 99.6 percent sensitive for PDR. Additionally, it assessed the findings of Retmarker, which shown sensitivities of 73.0% for any retinopathy, 85.00% for RDR, and 97.9% for PDR. Furthermore, Google Health revealed that it created a collection of 128,000 photos that scientists used to train a deep learning network for diabetic retinopathy.

OCT angiography is a novel technique with enormous potential in the areas of DR and DME. Numerous research publications have focused on the use of OCT angiography to quantify foveal avascular zone (FAZ) regions, changes in retinal microangiopathy (such as capillary tortuosity and dropouts), and retinal vascular density indices. The use of artificial intelligence to OCT angiography pictures remains in its infancy. There are just a few published publications describing the use of DL algorithms to determine vascular alterations detected by OCT angiography. Guo et al. devised a deep learning approach for segmenting and quantifying the capillary density of the superficial FAZ. A correlation coefficient of 0.997 was found between the area computed by the DL method and the area determined manually.

Heisler et al. classified DR on OCT angiography using ensemble learning approaches in conjunction with DL. After analysing 380 eyes, the authors determined that ensemble learning improves the prediction accuracy of CNNs for identifying RDR on OCT angiography. Lo et al. also used OCT angiography pictures to examine the superficial and deep capillary plexus of the retina using CNNs. Using CNNs, the programme accurately assessed the retinal microvasculature. Thus, OCT angiography is a valuable tool, and segmentation of the images using CNNs is an exciting field of study.