🤖 machine learning

EFIQA: Explainable Fundus Image Quality Assessment via Anatomical Priors

The paper introduces EFIQA, a novel, label-free framework for fundus image quality assessment that leverages anatomical priors through masked inpainting and feature distillation to generate explainable spatial quality maps, outperforming supervised methods in generalization and interpretability.

Original authors: Pengwei Wang, José Morano, Qian Wan, Hrvoje Bogunović

Published 2026-06-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Pengwei Wang, José Morano, Qian Wan, Hrvoje Bogunović

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Subjective" Eye Doctor

Imagine you are taking a photo of a complex city map (the retina) to check for traffic jams or broken roads (diseases). Before you can analyze the map, you need to make sure the photo is clear.

Currently, computers check photo quality by being taught by humans. A human looks at thousands of photos and says, "This one is good," or "This one is blurry." The computer memorizes these specific examples.

The Flaw: If the computer learns from a team that hates blurry photos, it will reject a photo that is slightly blurry but still useful for a different doctor who only cares about brightness. The computer is rigid; it only knows the rules of the specific teacher it had.
The Second Flaw: If the computer says "Bad Photo," it can't tell you why or where. It's like a teacher giving you a failing grade on a test without circling the specific mistakes.

The Solution: EFIQA (The "Missing Puzzle Piece" Detective)

The authors propose a new way called EFIQA. Instead of asking the computer, "What does a bad photo look like?" (which requires human labels), they ask, "What should be in this photo?"

They use a clever trick based on anatomy (the natural structure of the eye). They know that a healthy eye must have a specific network of blood vessels, just like a city map must have roads.

How It Works (The Two-Stage Process)

Stage 1: The "Blindfolded" Architect (VUAD)
Imagine you have a blueprint of a city's road network. You cover up random parts of the blueprint with a black marker.

The computer's job is to look at the visible roads and guess what the missing roads should look like to complete the pattern.
If the computer sees a gap where a road should be, but the photo is too blurry or dark to show it, the computer realizes, "Hey, I can't see the road here. This part of the photo is low quality."
Key Point: The computer never saw a "bad photo" label. It just learned what a complete road network looks like. If it can't "fill in the blanks," it knows the photo is bad.

Stage 2: The "Translator" (Distillation)
The first stage is great at finding missing roads, but it's slow and clunky.

The authors take the "knowledge" the first computer learned (the rules of the road network) and teach a smaller, faster computer (an "adapter") how to spot these missing roads instantly.
This small computer looks at the whole eye photo and produces a heat map.
- Green areas: "I see the roads clearly. Good quality."
- Red areas: "I can't see the roads here. Bad quality."

Why This Is Better (The Results)

1. It's a "Universal" Detective
Because EFIQA doesn't memorize human opinions, it works on photos taken by different cameras, in different lighting, or by different people.

Analogy: If you teach a child to recognize a "dog" by showing them only Golden Retrievers, they might think a Poodle isn't a dog. But if you teach them "dogs have four legs and a tail," they recognize any dog. EFIQA learns the "four legs and tail" (the anatomy), not the specific breed (the specific dataset).
The Paper's Claim: When tested on new datasets that the computer had never seen before, EFIQA outperformed the state-of-the-art methods that were trained on specific human labels.

2. It Gives You a Map, Not Just a Score
Old methods give you a single number: "This photo is 60% good."
EFIQA gives you a map: "The top-left corner is blurry, and the bottom-right is too dark, but the center is perfect."

Analogy: Instead of a teacher saying "You failed," EFIQA highlights the exact sentences in your essay that need fixing.

3. It Works Without "Grading" Data
The system was trained on high-quality images where the computer tried to "fill in the blanks." It didn't need thousands of photos labeled "Good" or "Bad" by doctors. It just needed to understand the anatomy of the eye.

The Limitations (What the Paper Admits)

The authors are honest about where their system might stumble:

The "Fovea" Problem: There is a tiny spot in the center of the eye (the fovea) that naturally has very few blood vessels. Because EFIQA looks for missing vessels, it might mistakenly think this natural empty spot is a "bad quality" area.
The "Disease" Problem: If a disease destroys the blood vessels, the computer might think the photo is low quality, when actually the photo is clear, but the patient is sick. The system confuses "missing structure due to bad photo" with "missing structure due to disease."

Summary

EFIQA is a new way to check eye photos. Instead of learning from human opinions on what is "good," it learns the rules of anatomy. It asks, "Can I see the blood vessels clearly?" If the answer is no, it marks that spot as low quality. This makes the system more flexible, more accurate on new data, and able to show exactly where the photo is blurry or dark.

Technical Summary: EFIQA – Explainable Fundus Image Quality Assessment via Anatomical Priors

Problem Statement
Medical image quality assessment (IQA) is critical for downstream clinical applications and AI preprocessing pipelines. Current state-of-the-art methods for Color Fundus Photography (CFP) are predominantly supervised classifiers trained on dataset-specific quality labels. This approach suffers from two fundamental limitations:

Lack of Generalization: Quality definitions are subjective and task-dependent (e.g., an underexposed macula may be acceptable for glaucoma screening but critical for AMD screening). Models trained on specific labeling criteria fail to generalize to datasets with different criteria or acquisition conditions.
Lack of Explainability: Existing methods output a single image-level score or class, providing no spatial feedback on where the quality is degraded. Post-hoc attribution methods (e.g., GradCAM) often activate on discriminative regions rather than degradation regions and lack precise, grounded clinical meaning.

Methodology: EFIQA Framework
The authors propose EFIQA, a framework that assesses fundus image quality by measuring deviations from expected anatomical structures rather than learning "what degradation looks like" from human labels. The approach is label-free regarding quality supervision and operates in two stages:

Stage 1: Vessel Unsupervised Anomaly Detection (VUAD)
- Objective: Train an unsupervised anomaly detector to learn the topology of retinal vasculature, a ubiquitous anatomical structure highly susceptible to common image degradations.
- Mechanism: The model is trained on vessel segmentation maps derived from high-quality images using a masked anatomical inpainting objective.
- Progressive Relaxation: To prevent "identity collapse" (where the model simply copies unmasked pixels), a progressive masking strategy is employed. Training begins with high masking ratios ( $\gamma \in [0.8, 1.0]$ ) to force structural learning, then linearly expands to $\gamma \in [0.0, 1.0]$ .
- Output: The network predicts the residual (missing structure) for a given input. The magnitude of this residual forms an anomaly map, highlighting regions where anatomical visibility is compromised (e.g., missing or blurred vessels).
Stage 2: Distillation to a Foundation Model Adapter
- Objective: Distill the anatomical prior knowledge from the VUAD model into a lightweight adapter that operates on features from a frozen Vision Foundation Model (FM). This improves generalization and reduces reliance on the intermediate vessel segmentation step during inference.
- Mechanism:
  - A frozen FM extracts dense patch-level features ( $Z$ ) from the input fundus image.
  - The VUAD model generates pseudo-labels (binary anomaly maps) for a mixed-quality dataset.
  - A learnable adapter ( $g_\psi$ ) is trained to predict these pseudo-labels from the FM features using a class-balanced Binary Cross-Entropy (BCE) loss.
- Inference: The adapter produces a patch-level quality score map ( $Q$ ). This map can be aggregated via a lightweight linear pooling layer to generate an image-level score or used directly for spatial localization.

Key Contributions

Novel Framing: The authors reframe IQA as a deviation from expected anatomical appearance, grounding the task in objective anatomy rather than subjective labels.
Label-Free, Explainable Approach: EFIQA requires no explicit quality-related supervision (manual labels or synthetic degradation) and produces spatial quality maps by design.
Two-Stage Architecture: The combination of an unsupervised vessel reconstruction network and a distillation process into a foundation model adapter allows for robust, generalizable performance.
Superior Generalization: External evaluations demonstrate that this approach outperforms supervised methods across benchmarks with varying quality criteria.

Experimental Results
The method was evaluated on four public datasets (MSHF, mBRSET, DRIMDB, EyeQ) and an internal dataset, comparing against supervised fundus-specific methods (e.g., MCF-Net, FGR-Net), general supervised IQA methods (MANIQA, TOPIQ), and unsupervised baselines (NIQE, IL-NIQE).

External Datasets (MSHF, DRIMDB, mBRSET): EFIQA achieved the best average performance across all metrics (Balanced Accuracy, F1, MCC, AUROC, AUPRC). Notably, it achieved a Matthews Correlation Coefficient (MCC) of 73.14%, surpassing the second-best supervised method (FGR-Net, 69.71%). It demonstrated robustness even on the challenging mBRSET dataset, where other methods struggled with class imbalance.
EyeQ Benchmark: EFIQA achieved competitive accuracy (87.55%), slightly below fully supervised models trained end-to-end on EyeQ (e.g., FGR-Net at 89.58%) but outperforming other adapted methods like TOPIQ and ARNIQA. Crucially, it is the only method in this comparison that provides a precise quality map.
Qualitative Analysis: Visualizations show EFIQA produces stable, low scores for high-quality images and accurately localizes degraded regions (e.g., illumination artifacts, blurring) in low-quality images. In contrast, VUAD alone sometimes misses degraded areas where vessels are visible but blurred, and supervised methods like MCF-Net often provide coarse, scattered activations.
Ablation Study: The study confirmed that while simple vessel density is a strong baseline, the VUAD model adds explainability (local maps) at a slight cost to raw performance, which is fully recovered and improved upon by the distillation step.

Significance and Claims
The paper claims that EFIQA offers a generalizable and explainable alternative to black-box supervised IQA. By leveraging anatomical priors, the framework avoids the brittleness associated with shifting labeling criteria. The authors emphasize that while the method provides local-level severity scores, it is designed to be adaptable to specific clinical criteria with minimal tuning.

The authors note limitations, specifically that using vessels as the sole prior may cause false activations in naturally vessel-sparse regions (e.g., the fovea) or ischemic lesions. They suggest future work could incorporate additional anatomical priors (macula, optic disc) or pathological structures to improve specificity. The paper concludes that this framework is applicable to any imaging scenario where anatomical visibility defines quality, such as OCT retinal layers or chest radiography landmarks.