In the future, photos of the fundus of the eye could help to identify early signs of atherosclerosis and peripheral arterial disease – an impending blockage of the blood vessels in the legs. Researchers have trained self-learning software to recognize tiny vascular changes in the veins of the eye – even before those affected show the first symptoms. This could help to take countermeasures at an early stage and thus prevent deaths and a deterioration in the quality of life.
Many older people suffer from atherosclerosis, the most common form of arteriosclerosis. Deposits of fat and calcium form on the inner wall of the arteries. This narrows and hardens the affected blood vessels. Consequences include an increased risk of thrombosis, heart attacks and strokes as well as painful circulatory disorders in the legs, the so-called peripheral arterial occlusive disease (PAOD). Even before the first symptoms appear, the disease is associated with increased mortality. So far, it has often only been diagnosed when consequential damage has already occurred.
Insights through the eyes
A team led by Simon Müller from the University Hospital in Bonn has now developed a method that could be used in the future to identify atherosclerosis at an early stage: They trained self-learning software to recognize vascular changes using photos of the fundus of the eye. The back of the eye is traversed by numerous fine arteries and veins that supply blood to the photoreceptors of the retina and the nerve cells connected to them. These blood vessels can be easily observed and photographed through the pupil.
“We photographed 97 eyes of women and men suffering from PAD,” explains Müller’s colleague Maximilian Wintergerst. “In more than half of them, the disease was still at a stage where it was not causing any symptoms.” In addition, the team recorded the background of 34 eyes of healthy controls with the camera. They used these images to train an artificial neural network – software that works like the human brain.
training the software
Such software usually requires tens of thousands of training photos – significantly more than were available in the study. “We therefore first carried out pre-training with another disease that attacks the vessels in the eye,” explains Thomas Schultz from the University of Bonn. To do this, the researchers used a data set of more than 80,000 additional photos. These came from patients with diabetic retinopathy, a condition in which the blood vessels in the retina change as a result of diabetes. “In a way, the algorithm learns from them what to pay particular attention to,” says Schultz. “We therefore also speak of transfer learning.”
Based on this pre-training and the training with the images that the researchers took in patients with peripheral arterial occlusive disease, the software learned to differentiate between unknown images and whether they came from PAOD patients or healthy people. “A good 80 percent of all those affected were correctly identified if we accepted 20 percent false-positive cases – i.e. healthy people who the algorithm incorrectly classified as sick,” explains Schultz. “This is astonishing, because even trained ophthalmologists cannot identify PAOD on the basis of fundus images.”
early detection important
In further analyses, the researchers were able to show that the neural network pays particular attention to the large vessels in the back of the eye when making its assessment. For the best possible result, however, the process required digital images with a sufficiently high resolution. “Many artificial neural networks work with very low-resolution photos,” says Schultz. “That’s enough to see major changes. For our PAOD classification, on the other hand, we need a resolution at which details of the vascular structures remain recognizable.”
With the current study, the researchers have demonstrated that their approach basically works. Now they want to continue training their software with additional recordings so that its performance improves and it may actually establish itself as a diagnostic tool in the long term. “The only available causal therapy against PAD is the treatment of risk factors such as smoking, high cholesterol or high blood pressure,” the researchers explain. The sooner such measures begin, the more effective they are. New, uncomplicated options for early detection could thus contribute to counteracting the progression of the disease at an early stage.
Source: Simon Müller (Bonn University Hospital) et al., Scientific Reports, doi: 10.1038/s41598-022-05169-z