Может ли базовая модель, обученная на естественных изображениях, превзойти специализированную ретинальную модель в выявлении глазных и системных заболеваний?

Can a Natural Image-Based Foundation Model Outperform a Retina-Specific Model in Detecting Ocular and Systemic Diseases?

The advent of foundation models (FMs) have ushered in a paradigm shift in medical deep learning approaches.¹, ², ³, ⁴ Driven by the availability of large-scale high-quality unstructured data sets and advanced self-supervised learning methods, FMs excel in performing multiple tasks with high adaptability and generalizability to new unseen tasks, often requiring minimal fine-tuning.^5,6 In ophthalmology, the introduction of RETFound, the first retina-specific FM developed by Zhou et al⁷ in 2023, led to a significant transformation in deep learning approaches for medical image analysis. Using a Vision Transformer (ViT)^8,9 backbone and a masked autoencoder–based¹⁰ self-supervised learning approach, RETFound was pretrained on a diverse dataset of 1.4 million natural images (ImageNet-1K), ∼900 000 color fundus photographs (CFPs), and ∼730 000 OCT scans. This pretraining approach enabled RETFound to effectively extract and learn retinal features and representations, enhancing diagnostic performance, data efficiency, and generalizability across ophthalmic and systemic clinical tasks.¹¹, ¹², ¹³