Training a high-performance retinal foundation model with half-the-data and 400 times less compute

Medical artificial intelligence is limited by available training datasets. Foundation models like RETFound from Moorfields Eye Hospital (MEH) can be adapted with small downstream datasets and thus alleviate this issue. RETFound-MEH used 900,000 training images. Recently, “data-efficient” DERETFound achieved comparable performance with 150,000 images. Both require very substantial compute resources for training and use. We propose RETFound-Green trained on only 75,000 publicly available images with 400 times less compute using a novel Token Reconstruction objective. RETFound-MEH and DERETFound training costs are estimated at $10,000 and $14,000, respectively. RETFound-Green cost less than $100, with equally reduced environmental impact. RETFound-Green can be downloaded 14 times faster, computes vector embeddings 2.7 times faster which then require 2.6 times less storage space. On a variety of downstream tasks from geographically diverse datasets, RETFound-Green achieves more than twice as many statistically significant wins than the next best model.