CausalFund: Causality-Inspired Domain Generalization in Retinal Fundus Imaging for Low-Resource Screening

The paper introduces CausalFund, a causality-inspired framework that enhances the domain generalization of AI models for glaucoma and diabetic retinopathy screening by disentangling disease-specific features from spurious image factors, thereby enabling reliable diagnosis across diverse clinical and low-resource settings using portable devices.

The Gist

The Big Problem: The "Fancy Camera" vs. The "Smartphone"

Imagine you are trying to teach a student how to identify a specific type of bird.

• The Traditional Way: You show the student 10,000 photos of that bird taken in a professional studio with perfect lighting, a high-end camera, and a white background. The student becomes an expert at spotting the bird in those perfect photos.
• The Real-World Problem: Now, you send that student into a muddy forest with a cheap, shaky smartphone camera. The photos are blurry, the lighting is dim, and there are leaves in the way. The student fails miserably. Why? Because they didn't learn to recognize the bird; they learned to recognize the perfect studio lighting.

This is exactly what happens in eye care today.
Doctors use AI to detect eye diseases like Glaucoma and Diabetic Retinopathy. These AI models are trained on high-quality images from expensive hospital cameras. But in rural or low-income areas, doctors often have to use cheap, portable smartphone cameras. When the AI trained on "hospital photos" sees a "smartphone photo," it gets confused and makes mistakes because the image quality is different.

The Solution: CausalFund (The "Truth Detector")

The researchers created a new method called CausalFund. Think of it as a "Truth Detector" for AI.

Instead of letting the AI memorize the background noise (like the color of the phone case or the specific lighting), CausalFund forces the AI to focus only on the causal truth: the actual physical signs of the disease inside the eye.

The Analogy: The "Intervener" Game

Imagine the AI is a detective trying to solve a crime (the disease).

1. The Old Way (ERM): The detective looks at the crime scene and says, "The suspect must be guilty because they were wearing a red hat." (The red hat is a "spurious" clue—it might just be the lighting making the hat look red, not a real clue).
2. The CausalFund Way: The detective has a magical assistant called the "Intervener."
   • The Intervener takes the detective's view and secretly changes the "spurious" clues. They swap the red hat for a blue one, or they change the lighting from sunny to cloudy.
   • The Test: If the detective still says, "This is a crime scene!" even after the hat changed color and the lighting shifted, then the detective is actually looking at the real evidence (the broken window, the footprint).
   • If the detective changes their mind because the hat changed color, the Intervener scolds them: "Stop looking at the hat! Look at the window!"

CausalFund trains the AI to play this game millions of times until it learns to ignore the "hats" (bad lighting, blurry phones, patient demographics) and only focus on the "broken windows" (the actual damage to the eye).

What Did They Find?

The researchers tested this on two major eye diseases using data from both fancy hospital cameras and cheap smartphone cameras.

1. It Works Better: When they switched from hospital photos to smartphone photos, the old AI models crashed (their accuracy dropped significantly). The CausalFund models, however, stayed strong. They didn't panic when the image quality got worse.
2. It's "Model Agnostic": This means CausalFund isn't a specific type of AI; it's a training technique that can be added to almost any existing AI brain. They tested it on seven different types of AI "brains," and it helped all of them.
3. It Handles "Bad" Photos: They intentionally made the smartphone photos worse (blurry, dark, noisy) to simulate real-world disasters. Even when the photos were terrible, CausalFund held its ground much better than the traditional methods.

Why Does This Matter?

It brings high-tech medicine to the people who need it most.

Right now, if you live in a remote village without a big hospital, you might not get screened for blindness-causing diseases. You might only have a community health worker with a smartphone.

• Before CausalFund: The AI would likely fail on that smartphone, giving false alarms or missing the disease.
• With CausalFund: The AI can look at that shaky, low-quality photo and say, "I see the specific signs of Glaucoma here," with high confidence.

The Bottom Line

The paper proposes a smarter way to train AI. Instead of teaching AI to recognize "hospital-quality images," they teach it to recognize disease-causing features that stay the same no matter what camera you use.

It's like teaching a chef to recognize the taste of a perfect tomato, rather than teaching them to recognize a tomato only when it's sitting on a specific white plate. Now, the chef can find that perfect taste even if the tomato is in a muddy basket.

Note: The authors admit this is a preprint (a draft before final peer review) and that real-world testing in actual clinics is the next big step. But the results so far suggest a huge leap forward for affordable eye care.

Technical Summary

1. Problem Statement

The paper addresses a critical bottleneck in deploying AI for ophthalmology: the domain gap between high-quality hospital-grade fundus images and low-cost images acquired via portable devices (e.g., smartphones) in low-resource settings.

• The Challenge: Current state-of-the-art AI models for detecting Glaucoma and Diabetic Retinopathy (DR) are trained on curated, high-quality datasets from clinical cameras (e.g., Topcon, Zeiss). These models often fail to generalize when applied to smartphone-captured images due to variations in field of view, illumination, sharpness, color response, and compression artifacts.
• The Consequence: Models trained on hospital data tend to rely on spurious correlations (e.g., device signatures, specific lighting conditions, or demographic cues) rather than true disease-causal features. When deployed in uncontrolled, low-resource environments, this leads to significant performance degradation and unreliable screening.
• The Limitation of Current Solutions: Retraining or fine-tuning on target-domain data is often impractical in low-resource settings due to the lack of labeled data and consistent acquisition protocols.

2. Methodology: CausalFund

The authors propose CausalFund, a model-agnostic, causality-inspired learning framework designed to enforce Domain Generalization (DG). The core philosophy is to disentangle disease-relevant causal features from spurious, domain-dependent nuisance factors.

Core Mechanism

CausalFund utilizes an intervention-based approach to train models to be invariant to domain shifts. The framework consists of the following components:

1. Shared Feature Extractor: A deep learning backbone (e.g., ResNet, ViT, MobileNet) maps input images to a latent causal representation $Z$.
2. Intervener (MLP): A three-layer Multi-Layer Perceptron generates a perturbation $\Delta$ conditioned on the label (true or predicted). This perturbation simulates domain shifts in non-causal (spurious) factors while preserving the semantic class information.
3. Intervened Representation: The model creates an intervened latent representation $Z' = Z + \Delta$.
4. Consistency Constraint: A shared classifier predicts labels from both the original representation ($Z$) and the intervened representation ($Z'$). The model is optimized to ensure predictions remain consistent ($\hat{y} \approx \hat{y}'$) despite the perturbation.

Optimization Objective

The framework optimizes a combined loss function (Equation 1) to balance predictive performance with invariance:
$$L = L_{cls} + \lambda_{int}L_{int} + \lambda_{cons}L_{cons} + \lambda_{reg}L_{reg} + \lambda_{kl}L_{kl}$$

• $L_{cls}$: Standard cross-entropy classification loss on causal features.
• $L_{int}$: Classification loss on intervened features to ensure correctness under perturbation.
• $L_{cons}$: Consistency loss (MSE) forcing predictions from $Z$ and $Z'$ to be identical.
• $L_{reg}$: Regularization to constrain the magnitude of the intervention $\Delta$.
• $L_{kl}$: Kullback–Leibler divergence to encourage stable latent representations.

3. Experimental Setup

• Datasets:
  • Glaucoma: Hospital domain (Multichannel Glaucoma Benchmark, 12,316 images) vs. Smartphone domain (Brazil Glaucoma/BrG, 2,000 images).
  • Diabetic Retinopathy (DR): Hospital domain (Aggregated APTOS, DDR, IDRiD, EyePACS, Messidor, 45,327 images) vs. Smartphone domain (Mobile Brazilian Retinal Dataset/mBRSET, 4,884 images).
• Backbones: Evaluated across seven architectures, including standard models (ResNet, DenseNet, EfficientNet, VGG, ViT) and lightweight models for edge deployment (MobileNet, SqueezeNet).
• Protocol: Models were trained exclusively on the hospital domain and tested on the smartphone domain (zero-shot domain generalization).
• Robustness Testing: Controlled degradation levels (mild, moderate, severe) were applied to smartphone images to simulate real-world quality issues (blur, noise, compression, color distortion).

4. Key Results

CausalFund consistently outperformed standard Empirical Risk Minimization (ERM) baselines across all metrics and backbones.

Generalization Performance

• Glaucoma:
  • Smartphone Domain AUC: ERM ranged from 0.649 to 0.885; CausalFund improved this to 0.757 to 0.889.
  • Significant Gains: EfficientNet improved from 0.649 to 0.757 ($p < 0.001$); ViT improved from 0.854 to 0.881.
  • Sensitivity/Specificity: CausalFund achieved a more favorable trade-off, particularly in sensitivity for smartphone screening.
• Diabetic Retinopathy:
  • Smartphone Domain AUC: CausalFund showed consistent improvements, e.g., DenseNet improved from 0.938 to 0.945, and EfficientNet from 0.907 to 0.935 ($p < 0.001$).
  • Lightweight Models: MobileNet and SqueezeNet showed notable gains in averaged sensitivity/specificity on smartphone data (e.g., MobileNet improved from 0.838 to 0.912).

Robustness to Image Degradation

• Performance Stability: As image quality degraded from mild to severe, ERM models suffered steep performance drops. CausalFund exhibited flatter degradation curves, maintaining higher AUC and sensitivity.
• Attention Analysis (Grad-CAM):
  • ERM: Attention maps shifted unpredictably across degradation levels, often focusing on background artifacts or lighting changes.
  • CausalFund: Attention remained stable and focused on the optic nerve head (cup-to-disc ratio, neuroretinal rim thinning), demonstrating that the model learned causal disease features rather than spurious correlations.

5. Key Contributions

1. Causal Framework for Low-Resource Screening: Introduced CausalFund, a novel framework that explicitly disentangles causal disease features from domain-specific nuisance factors using intervention-based learning.
2. Model Agnosticism: The framework is compatible with various deep learning backbones, including lightweight architectures (MobileNet, SqueezeNet) essential for on-device deployment in resource-constrained areas.
3. Empirical Validation: Demonstrated significant improvements in domain generalization for both Glaucoma and DR screening, specifically bridging the gap between hospital-grade and smartphone-captured images.
4. Robustness to Quality Shifts: Proved that CausalFund models are significantly more robust to severe image quality degradations (blur, noise, compression) compared to conventional ERM models.

6. Significance and Impact

• Clinical Accessibility: This work enables the deployment of reliable AI screening tools in underserved, rural, and low-resource populations where only portable devices are available, potentially preventing irreversible vision loss.
• Reduced Reliance on Labeled Target Data: By improving generalization from hospital data alone, the need for expensive and difficult-to-obtain labeled smartphone data in target regions is reduced.
• Interpretability: The framework encourages models to learn biologically plausible features (e.g., optic disc structures) rather than dataset artifacts, increasing trust in AI-driven diagnostics.
• Future Directions: While the study is retrospective, it lays the groundwork for prospective clinical trials and the extension of causal learning to other ophthalmic diseases and imaging modalities.

Limitations Noted by Authors: The study relies on retrospective public datasets and synthetic degradation pipelines, which may not fully capture the complexity of real-world clinical workflows. Future work requires prospective validation, fairness audits across demographics, and real-world deployment analysis.