Skip to main content
. 2026 Sep 25;14:149008–149023. doi: 10.1109/ACCESS.2026.3737380

FIGURE 1.

FIGURE 1.

Overview of the patch-based multiple instance learning framework used to evaluate foundation model representations. Ultra-widefield retinal images are divided into patches and processed by a feature extractor. Patch-level features are aggregated using an attention-based MIL module to produce a bag-level representation for classification.