Disease Spectrum Analysis and AI-Assisted Screening of Fundus Disorders Using Portable Handheld Fundus Imaging in Health Examination Populations

Seeing the world clearly is a fundamental part of being human, yet for billions of people, that clarity is fading due to eye diseases that often show no warning signs until it is too late. The back of the eye, known as the retina, is a delicate layer of tissue that acts like a camera sensor, capturing the light that allows us to see. When this tissue is damaged by conditions like diabetes, high blood pressure, or aging, the damage is often permanent. Because these problems start quietly, doctors rely on taking pictures of the retina to find them early. Traditionally, these pictures are taken with large, expensive machines that require a patient to sit still and look into a fixed device. However, a newer type of camera is changing this landscape. These are small, handheld devices that a doctor can hold in one hand, making it possible to take pictures of the retina almost anywhere, from a busy community clinic to a routine physical exam room. The challenge has been that these portable cameras are harder to use perfectly; the images can be blurry or poorly lit, and it is unclear if artificial intelligence, the computer software designed to spot diseases, can handle these imperfect pictures as well as it handles the perfect ones taken in specialist clinics.

A team of researchers set out to solve this problem by testing whether these portable cameras could be used to screen large groups of healthy people for a wide range of eye diseases, not just one specific condition. They worked with a massive collection of images taken over five years from nearly thirty thousand people who visited a hospital in Beijing for routine health check-ups. In total, they gathered more than fifty-eight thousand pictures of retinas. Because the cameras were handheld and used on people who were not patients waiting for an eye exam, the quality of the pictures varied wildly. Some were crystal clear, while others were too dark, too blurry, or misaligned to be useful. The researchers first built a computer system to act as a strict gatekeeper, sorting the images into three groups: those that were perfect, those that were good enough to use, and those that had to be thrown away. This step was crucial because it ensured that the artificial intelligence models were only trained on images that actually showed the eye clearly. After this cleaning process, they were left with over thirty-five thousand high-quality images to study the actual health of the population.

What the researchers found was a clear picture of how eye health changes as people get older. In the group of images they analyzed, about one-third showed some sign of eye trouble. The most common issue, by far, was changes related to nearsightedness, which affected nearly twenty percent of the people. This makes sense given that the study included many young and middle-aged adults. However, the type of disease found depended heavily on age. Younger people mostly had issues linked to nearsightedness, but as the group got older, the problems shifted. In people over sixty, the most common findings were diseases of the blood vessels in the eye and degeneration of the central part of the retina, which are conditions typically associated with aging and overall health. This discovery is important because it suggests that eye screening programs should not be one-size-fits-all; what doctors should look for in a thirty-year-old is different from what they should look for in a seventy-year-old.

To see if computers could help doctors find these problems faster, the team tested twelve different types of artificial intelligence models. They asked these models to do two things: first, to decide if a picture showed a healthy eye or an unhealthy one, and second, to identify exactly what kind of disease was present if the eye was unhealthy. The results were encouraging for the first task. The computer models were very good at spotting when something was wrong, correctly identifying abnormal eyes in more than ninety percent of cases. This means that a handheld camera combined with this kind of software could reliably flag people who need to see a specialist. However, the second task was much harder. When the models tried to name the specific disease, their accuracy dropped significantly. They often confused different types of blood vessel problems or struggled to distinguish between different kinds of tissue damage. This suggests that while computers are excellent at acting as a first filter to catch potential issues, they are not yet ready to replace a human doctor in diagnosing the exact condition from a handheld photo.

The researchers also tested a new generation of advanced computer models that had been trained on millions of standard, high-quality eye pictures from big clinics. They hoped these "foundation models" would be smarter and better at handling the messy, real-world images from the handheld cameras. Surprisingly, these advanced models did not perform better than the simpler, lighter models. In fact, the specialized models sometimes struggled more, likely because the images from the small handheld cameras looked too different from the perfect images they were trained on. The simpler models, which are faster and require less computing power, actually did the best job. This finding is a practical guide for the future: for widespread screening with portable devices, we do not need the most complex, heavy-duty artificial intelligence. We need tools that are robust enough to handle imperfect pictures and fast enough to process thousands of images a day.

Ultimately, this study proves that it is possible to use small, portable cameras to screen large numbers of people for many different eye diseases, provided there is a system in place to check the quality of the pictures first. The research shows that eye diseases follow a predictable pattern based on age, with nearsightedness dominating in the young and vascular diseases taking over in the old. It also shows that while artificial intelligence can be a powerful partner in spotting these problems, the technology is not yet perfect at naming every specific disease. The work lays the foundation for a new way of protecting vision, where routine health check-ups could include a quick, painless look at the retina, catching problems early before they lead to blindness. By combining a simple handheld device with smart software and a clear understanding of how eye diseases change over a lifetime, we can move toward a future where eye health is monitored as routinely as blood pressure or weight.