Quantifying the Effect of Synthetic Colour Fundus Photograph Scaling on Foundation Models
This study demonstrates that scaling synthetic color fundus photographs during foundation model pretraining offers a tunable strategy where lower-volume supplementation optimizes the performance–computation balance while larger volumes significantly enhance external generalization, with the optimal approach depending on specific evaluation objectives and available computational resources.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The human eye offers a unique window into the body's overall health. By examining the retina, the light-sensitive layer at the back of the eye, doctors can spot signs of conditions like diabetes, high blood pressure, and even Alzheimer's disease. To do this, they often rely on color fundus photographs, which are detailed pictures of the retina's blood vessels and nerve head. In recent years, artificial intelligence has become a powerful tool for analyzing these images, helping to detect diseases faster and more accurately than ever before. However, these intelligent systems require vast amounts of real patient data to learn effectively. Gathering enough high-quality images from diverse populations is difficult, expensive, and raises serious privacy concerns. This creates a bottleneck: the more advanced the technology becomes, the harder it is to find the real-world examples needed to teach it.
Researchers have begun exploring a solution: creating artificial, or synthetic, images that look and behave like real retinal photographs. These computer-generated images are not photographs of actual people but are crafted by algorithms trained on real data to mimic the patterns of healthy and diseased eyes. The idea is that these synthetic images could supplement real data, allowing artificial intelligence models to learn more without needing to collect millions of new patient photos. But a critical question remained unanswered: does simply adding more fake images make the system smarter, or is there a point where the extra images stop helping and only waste computing power? A team of researchers set out to find the answer by testing how different amounts of synthetic data affect the performance of these medical AI models.
The team started with a collection of 28,430 real retinal photographs gathered from three different public datasets. These images came from patients in China, Brazil, and other regions, covering a variety of eye conditions. Using this real data, they trained a computer program to generate new, synthetic retinal images. They created a massive pool of 568,600 synthetic images, which is twenty times the size of their original real collection. To test how these synthetic images influenced the AI, the researchers designed six different training scenarios. The first scenario used only the real images. The second used only the synthetic images. The remaining four scenarios mixed the real images with synthetic ones in increasing amounts: one synthetic image for every real one, five for every one, ten for every one, and finally, twenty synthetic images for every single real image.
Each of these six groups was used to train a separate artificial intelligence model. These models were then put to the test on a series of tasks, including detecting glaucoma, diabetic retinopathy, and age-related macular degeneration. The researchers evaluated the models in two ways. First, they tested them on data that was similar to what they had seen during training, known as internal validation. Second, they tested them on completely different datasets from other hospitals and regions, known as external validation, to see if the models could generalize their knowledge to new situations. They also carefully tracked the amount of computing power required for each strategy, measuring the total number of mathematical calculations, or floating-point operations, needed to generate the images and train the models.
The results revealed a nuanced picture of how synthetic data works. When the models were tested on familiar data, the best performance came from the group that mixed real images with ten times as many synthetic images. This combination achieved the highest accuracy scores among all the groups. However, when the models were tested on completely new, external data, the story changed. The group that used twenty times as many synthetic images performed the best, showing the strongest ability to handle unseen cases. Interestingly, the group that used only synthetic images performed just as well as the group using only real images when the total number of images was the same. This suggests that synthetic images can stand on their own as a training resource, at least for the tasks tested.
The researchers also looked at the cost of these gains. Generating the synthetic images required a staggering amount of computing power, accounting for about ninety-nine percent of the total energy used in the study. The rest was used for training the models. Because of this, the researchers found that using a smaller amount of synthetic data offered a better balance between performance and cost. The strategy that mixed real images with just one synthetic image for every real one provided a significant boost in accuracy while using the least amount of computing power among the mixed strategies. It improved the models' performance by a meaningful margin without the massive energy cost of generating thousands of extra images.
The study concludes that there is no single "perfect" amount of synthetic data to use. Instead, the best approach depends on the goal. If the priority is to build a model that performs well on new, unseen data from different populations, adding a large volume of synthetic images is beneficial. If the goal is to improve performance efficiently with limited computing resources, adding a smaller amount of synthetic data is the smarter choice. The findings suggest that synthetic images are a valuable tool for training medical AI, either as a way to expand existing datasets or as a standalone resource when real patient data is scarce. By carefully matching the amount of synthetic data to the specific needs of the project and the available computing power, researchers can build more robust and reliable tools for eye care without compromising patient privacy or exhausting digital resources.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.