CAMEL: Cross-Architecture Masked Embedding Learning for Ultra-Lightweight Retinal Foundation Models

Vision foundation models for retinal imaging are large and computationally demanding, limiting point-of-care deployment. We ask how much of such a model's downstream utility survives compression into a smaller network. We propose a topology-aware discretization framework that distills a continuous foundation model into a discrete codebook, then uses it as the reconstruction target for masked image modeling (MIM). It combines gradient-free codebook learning via Growing Neural Gas, which yields an adjacency graph over the codes, with a graph-distance soft cross-entropy loss that replaces one-hot MIM targets with topologically graded supervision. With student pretraining on 383 public Duke iAMD optical coherence tomography (OCT) volumes, the pipeline distills an 86M-parameter teacher into a 22M-parameter student with at least 3.6x fewer FLOPs. The student performs close to, and in some settings above, its teacher across three classification and two segmentation benchmarks, with the magnitude and direction of the difference depending on the teacher, task, metric, and student capacity. On the external OLIVES benchmark, it surpasses its MIRAGE teacher on every metric (98.06% vs. 95.22% balanced accuracy). It works with two Vision Transformer (ViT-Base) teachers, retinal (MIRAGE) and natural-image (DINOv3); with the out-of-domain teacher, the student outperforms it on retinal-layer segmentation, indicating that the added value of retina-specific pretraining is task-dependent: largest for segmentation, while a general-domain teacher is competitive for classification. These results yield compact retinal models potentially suitable for deployment, transferring across OCT datasets, devices, and clinical tasks, including three datasets not used in the student's pretraining.

Foundation models are used to extract transferable representations from large amounts of unlabeled data, typically via self-supervised learning (SSL). However, many of these models rely on architectures that offer limited interpretability, a critical issue in high-stakes domains such as medical imaging. We propose DualIFM, a foundation model that is interpretable-by-design via a BagNet backbone whose small receptive fields generate class evidence maps that are faithful to the model's decision-making process. Additionally, DualIFM incorporates a $2D$ projection layer during pretraining that enables direct visualization of the representation space, providing a dataset-level view of the learned structure including meaningful clinical clusters as well as potential spurious correlations. We trained DualIFM on over 800,000 color fundus photographs from various sources to learn generalizable representations for different downstream tasks. Our model achieves performance comparable to RETFound, which has $16\times$ more parameters, while providing interpretable predictions on out-of-distribution data. These results suggest that large-scale SSL pretraining paired with inherent interpretability can lead to robust representations for retinal imaging. Code and pretrained models are available at github.com/berenslab/interpretable_FM.

Despite the widespread adoption of foundation models as feature extractors for medical imaging, relatively little is understood about how different pretraining strategies influence the transferability of learned representations to weakly supervised ophthalmic imaging tasks. We investigate this question in ultra-widefield (UWF) retinal imaging by evaluating foundation model representations within a patch-based multiple instance learning (MIL) framework for disease classification on UWF images. We compare Vision Transformer encoders pretrained with supervised, Masked Autoencoder (MAE), and self-distillation objectives, while keeping the downstream aggregation architecture unchanged. Within a controlled comparison of ViT-B encoders pretrained on ImageNet-1k, the choice of pretraining objective substantially influenced frozen representation transfer, with supervised and self-distillation-based models outperforming MAE. A contemporary DINOv3 model pretrained at a larger scale achieved the strongest overall performance, with a quadratic weighted kappa of 0.863 for five-class diabetic retinopathy grading, comparable with DINOv1. Attention analysis further revealed distinct patch-aggregation behaviours associated with the different pretrained representations, while partial fine-tuning substantially reduced the performance gap for MAE. These findings suggest that pretraining strategy influences both representation transferability and the subsequent aggregation of patch-level evidence within MIL, resulting in differences in downstream classification performance.