Researchers from a joint Skoltech-University of Sharjah laboratory and AIRI Institute have automated the analysis of retina images used to diagnose diabetic retinopathy. This refers to retinal damage in diabetes patients that can potentially cause permanent blindness. Depending on the details of the case and the skill of the physician, it can take them anywhere from 10 to 40 minutes to examine the blood vessel network in the retina image and make a diagnosis. Presented in Pattern Recognition Letters, the team’s AI solution delivers the result instantaneously, leaving it to the eye doctor to review and confirm the findings.

Eye care professionals use specialized cameras to take retina images and study them, manually segmenting the photo. This involves differentiating between the background and blood vessels of varying length, width, and tortuosity — the latter refers to a swelling pattern. Features of the retinal blood vessel network can point to diabetic retinopathy, as well as other eye and cardiovascular diseases — even atherosclerosis. However, manual image segmentation is very difficult, time-consuming, and error-prone.

Now, researchers have automated this daunting task in a way that promises not just to save time to the eye doctors but possibly to eliminate some human errors. By training their AI system on a highly reliable sample of double-checked cases studied by top physicians, the team has achieved exceptionally good performance in tests on three state-of-the-art datasets. That includes an accuracy of over 97% and a sensitivity of over 84% on the industry-standard database called DRIVE.

“For this research, achieving 97% accuracy is not that difficult due to the nature of the data. It is the sensitivity that matters the most. It reflects the ability of the model to identify microvessels, which the previous models struggled with,” the paper’s lead author Melaku Getahun, a Skoltech PhD student in the Engineering Systems program, explained.

What makes this kind of segmentation particularly challenging are the fine details in the retinal photos, which have to be accounted for and yet often elude both the neural networks developed for the task earlier and some of the eye specialists analyzing these images manually.

“In this study, we propose a neural network architecture different from those used by prior approaches, which tend to overlook the microvessels found in the retina,” Getahun said. “We also introduced an algorithm for tuning the output of the model by understanding the underlying nature of the retina vessel image data. This helps avoid the misclassification of vessel pixels as background.”

One of the challenges faced by the team was the limited size of the dataset: While the images twice segmented by experts and used to train the neural network were quite reliable, there weren’t as many of them available as one would ideally want.

“This hindered the model’s ability to generalize effectively to unseen data. However, through the careful application of data augmentation and processing techniques, we managed to significantly improve the model’s performance,” said the study’s principal investigator on the Russian side, Senior Research Scientist Oleg Rogov from Skoltech AI, who heads the Reliable and Secure Intelligent Systems group at AIRI. “Also, even with our new neural network architecture, the issue with certain microvessel pixels getting misclassified as background persisted. To address this, we implemented an adaptive threshold algorithm, which delivered a substantial improvement in sensitivity and accuracy.”

Asked about the solution’s future prospects, the team commented that the ability to spot tiny unhealthy blood vessels should be valuable for clinical use. As the system continues to develop, the researchers said, it could become a standard tool for eye disease screening, helping ophthalmologists diagnose conditions faster and more accurately. The work opens new possibilities for early detection of eye diseases and could lead to better patient outcomes through earlier intervention, because the small vessels often show the first signs of eye-related pathologies.

“This can help in the early diagnosis and prevention of eye diseases that are difficult to treat, such as diabetic retinopathy, which is prevalent in populations with high incidence of diabetes, as well as other related microvessel eye diseases,” study co-author and University of Sharjah Professor Rifat Hamoudi added.

The study reported in this story was carried out by the Biomedically Informed Artificial Intelligence Laboratory (BIMAI-Lab), which is a Skoltech-University of Sharjah research laboratory jointly headed by Assistant Professor Maxim Sharaev from Skoltech and Professor Rifat Hamoudi from UoS. BIMAI-Lab’s team includes Professor Ahmed Bouridane, the co-investigator of the project at the University of Sharjah, who has extensive expertise in applying artificial intelligence to medical data analytics.