Wednesday 22 January 2025
The quest for a perfect AI model has long been a holy grail in the field of computer vision. Recently, researchers have made significant strides in developing foundation models that can learn generalizable features from vast amounts of data. In a new study, scientists have evaluated the performance of these foundation models against traditional convolutional neural networks (CNNs) in detecting various ocular and systemic diseases.
The team trained three different models: RETFound, a retina-specific foundation model; ResNet50, a widely used CNN architecture; and SwinV2, another popular vision transformer. They fine-tuned each model on datasets of varying sizes to detect six ocular diseases – disease-related visual impairment, visually significant cataract, diabetic retinopathy, glaucoma, hypertension, and chronic kidney disease.
The results showed that RETFound outperformed the traditional models in detecting systemic diseases like diabetes, hypertension, and chronic kidney disease when fine-tuned on smaller datasets. This is likely due to its pre-training approach, which involves reconstructing masked retinal images. This process enables the model to capture subtle patterns and features specific to retinal images that may not be visible to human eyes or traditional CNNs.
In contrast, RETFound’s performance was comparable to the traditional models in detecting ocular diseases like diabetic retinopathy, glaucoma, and visually significant cataract when fine-tuned on larger datasets. This suggests that traditional models can still achieve good results in these tasks, especially when large amounts of labeled data are available.
However, RETFound’s superior performance in systemic disease detection highlights its potential to be a valuable tool for oculomics – the application of machine learning and artificial intelligence to eye diseases. By leveraging foundation models like RETFound, researchers may be able to develop more accurate and generalizable AI systems that can detect a wide range of ocular and systemic diseases.
The study’s findings also have important implications for healthcare systems. With the increasing demand for accurate disease detection and diagnosis, AI-powered systems like RETFound could potentially help reduce healthcare costs by identifying high-risk patients earlier and improving treatment outcomes.
While foundation models like RETFound show great promise in oculomics, there are still several challenges to overcome before they can be widely adopted. For instance, these models require large amounts of labeled data, which can be time-consuming and expensive to collect. Additionally, the lack of diversity in training datasets can lead to biased AI systems that perform poorly on underrepresented populations.
Cite this article: “Foundation Models Outperform Traditional CNNs in Detecting Ocular and Systemic Diseases”, The Science Archive, 2025.
Ai Models, Computer Vision, Foundation Models, Retina-Specific Model, Convolutional Neural Networks, Ocular Diseases, Systemic Diseases, Oculomics, Machine Learning, Artificial Intelligence.
