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

AI Generated Summary*

RETFound-Green is a colour fundus foundation model adapted with a new self-supervised Token Reconstruction objective, where a network reproduces a frozen copy's output tokens from corrupted inputs. Pretraining used 75,000 public images, half the number behind DERETFound and a twelfth of RETFound-MEH's 900,000, with about 400 times less compute than RETFound-MEH and a cost under $100 against estimates of $10,000 for RETFound-MEH and $14,000 for DERETFound. It downloads 14 times faster, computes embeddings 2.7 times faster, and stores them in 2.6 times less space. Across six geographically diverse datasets it took 36 significant wins versus 11 and 7 for the comparators, 41 in cross-dataset diabetic retinopathy transfer, and roughly matched a fully finetuned RETFound-MEH using only linear probing.

*This summary was generated by AI and is published unedited. Oku does not alter these summaries. It may contain errors or omissions and is provided for general informational purposes only. Oku does not guarantee its accuracy, completeness, or reliability. For authoritative information, please refer to the original, peer-reviewed article.