Interpretable Fundus Image Classification via Ring-Based Retinal Vasculature Features

Every day, millions of people rely on retinal fundus photography, a specialized camera technique that captures the back of the eye, to screen for sight-threatening diseases. This simple image holds a wealth of information about the body's health, particularly regarding conditions like diabetic retinopathy and glaucoma, which are leading causes of blindness worldwide. For decades, doctors have examined these photographs by eye, looking for subtle changes in the delicate network of blood vessels that supply the retina. They check for vessels that have become too wide, too twisted, or changed color, using these visual clues to diagnose disease. However, this human judgment can vary from one doctor to another, and the sheer volume of images makes consistent, rapid screening difficult. In recent years, powerful computer programs known as deep learning models have been trained to read these images automatically, often matching or exceeding human accuracy. Yet, these programs operate like black boxes; they make a diagnosis based on complex, hidden patterns that are impossible for a human to see or understand, leaving doctors unsure of why the computer reached its conclusion.

A new study from researchers at the University of Toronto, Duke Kunshan University, and York University offers a different path forward. Instead of relying on hidden patterns, they developed a method that breaks the eye image down into a series of concentric rings, much like the rings of a tree, centered on the optic disc where the main blood vessels enter the eye. By measuring specific, understandable characteristics of the blood vessels within each ring—such as their thickness, how they branch, their color, and how they contrast with the surrounding tissue—the team created a clear, transparent map of the eye's vascular health. This approach does not just tell a doctor whether a patient has a disease; it explains exactly which parts of the blood vessel network and which specific physical traits led to that conclusion. The researchers found that this method, which relies on direct measurements rather than hidden algorithms, can classify eye diseases with high accuracy, matching the performance of the most advanced, complex computer models while remaining fully understandable to a human observer.

The core of this work lies in how the researchers organized the information. They divided the fundus image into circular zones radiating outward from the optic disc. In each zone, they calculated four distinct types of information. First, they measured the geometry of the vessels, counting how many branches appeared and how much of the area was covered by blood vessels. Second, they analyzed the color of the vessels, looking at the balance between red and green light, which can indicate how much oxygen the blood is carrying. Third, they examined the texture of the area immediately surrounding the vessels, checking for subtle changes in the background that might signal disease. Finally, they combined these measurements into a single, compact set of numbers for each image. This process transforms a complex, colorful photograph into a structured list of facts that a simple computer classifier can read. The result is a system that is not only accurate but also interpretable, allowing a doctor to see exactly which ring or which feature contributed most to the diagnosis.

When the researchers tested this ring-based system on three different public collections of eye images, the results were striking. On one dataset containing high-quality images with expert-drawn vessel maps, the method achieved perfect accuracy. On another dataset with more challenging, lower-quality images, it still performed exceptionally well, matching the accuracy of RETFound, a state-of-the-art artificial intelligence model that was trained on over 1.6 million retinal images. This comparison is significant because it suggests that the researchers' simple, transparent method can compete with massive, complex deep learning models without needing enormous amounts of data or computing power. In fact, the study revealed that these powerful deep learning models might be relying on shortcuts. When the researchers altered the background of the images or changed the size of the visible field of view, the performance of the deep learning models dropped significantly. This suggests that these advanced models were sometimes picking up on clues related to how the photo was taken—such as the camera settings or the specific framing—rather than just the biological signs of disease. The ring-based method, by contrast, remained stable because it focused strictly on the physical properties of the blood vessels themselves.

The study also explored how the quality of the initial vessel maps affected the final diagnosis. In cases where the computer had to guess the location of the blood vessels automatically, the accuracy was slightly lower than when experts provided the maps, but it was still very strong. This indicates that while precise vessel detection is important, the method is robust enough to handle real-world variations in image quality. Furthermore, the researchers found that different types of features mattered in different ways. For instance, measurements related to the color and oxygenation of the blood provided unique information that geometry alone could not capture. When they removed these color-based features from their model, the accuracy dropped noticeably, proving that the combination of shape, color, and texture is essential for a complete picture.

Perhaps most importantly, the researchers demonstrated that their method could be understood. They created visual heatmaps that highlighted exactly which rings of the eye supported the diagnosis. For a healthy eye, the system pointed to specific rings where the vessel patterns matched a healthy norm. For a diseased eye, it highlighted rings where the vessels were too thin, too twisted, or had the wrong color. This level of transparency is something that deep learning models struggle to provide. While a deep learning model might say "this is glaucoma," it often cannot explain why. The ring-based method, however, can say "this is glaucoma because the vessels in the ring closest to the optic disc are significantly thinner than usual." This ability to link a diagnosis directly to a measurable physical change makes the tool valuable not just for screening, but for understanding the progression of disease.

The researchers were careful to note the limitations of their work. The method depends on the ability to accurately identify the blood vessels in the first place; if the image is too blurry or the vessels are too faint, the measurements may be less reliable. They also acknowledged that their color measurements are relative and not calibrated to absolute physiological standards, meaning they are best used for comparing images rather than measuring exact oxygen levels. Despite these constraints, the study suggests a promising direction for the future of medical imaging. By returning to the fundamental, observable properties of the eye and organizing them in a way that mirrors human anatomical understanding, the researchers have created a tool that is both powerful and transparent. It offers a way to harness the speed of computers without losing the ability to trust and understand the diagnosis, bridging the gap between complex data and clinical insight. In an era where artificial intelligence is becoming increasingly sophisticated, this work reminds us that sometimes the most effective tool is one that speaks the same language as the doctor using it.