EFIQA: Explainable Fundus Image Quality Assessment via Anatomical Priors

Traditional deep learning methods for fundus image quality assessment (IQA) are typically trained on human-annotated labels, which limits their generalizability across different clinical criteria and fails to provide spatial explanations for quality degradation. The authors ask whether it is possible to assess image quality by learning the underlying anatomical structure of the eye, thereby removing the need for subjective quality labels and providing inherent explainability.

The authors propose EFIQA, a two-stage framework that treats IQA as an anomaly detection problem. In the first stage, a Vessel Unsupervised Anomaly Detection (VUAD) network is trained to reconstruct vessel maps from masked inputs, forcing the model to learn the expected anatomical topology of the fundus. In the second stage, this knowledge is distilled into a shallow adapter that maps features from a frozen foundation model to precise, spatial quality maps. By focusing on the visibility of vasculature—a key indicator of image quality—the model identifies regions where anatomical structures are missing or obscured, which serves as a proxy for poor image quality.

EFIQA demonstrates superior generalization compared to state-of-the-art supervised methods across multiple external datasets. By learning anatomical priors rather than dataset-specific quality criteria, the model maintains high performance even when evaluated on datasets with different quality definitions. Furthermore, the framework provides spatial quality maps by design, allowing clinicians to see exactly which regions of an image are degraded, unlike black-box classifiers that only output a single global score.

This approach addresses the brittleness of current medical AI systems, which often fail when deployed in clinical environments that differ from their training data. By grounding quality assessment in anatomy, EFIQA offers a more robust and interpretable tool for automated quality control in ophthalmology, potentially reducing the need for expensive, subjective manual labeling and improving the reliability of downstream diagnostic pipelines.