⚡ electrical engineering

Disentangling Co-Occurring Retinal Pathologies with Saliency-Guided Sparse Expert Routing

This paper proposes a novel deep learning architecture that combines Guided Context Gating with a sparsely-routed Mixture-of-Experts (MoE) block to achieve high-performance, interpretable classification of co-occurring retinal pathologies by dynamically allocating disease-specific experts based on image content.

Original authors: Nagur Shareef Shaik, Jeongwoo Park, Yeong-Jin Kim, Jaeuk Jung, Hyunjung Oh, Dong Hye Ye

Published 2026-08-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Nagur Shareef Shaik, Jeongwoo Park, Yeong-Jin Kim, Jaeuk Jung, Hyunjung Oh, Dong Hye Ye

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

Imagine a world where your eyes are like a busy city, and doctors are the traffic controllers trying to spot accidents. For years, the tools used to scan these cities (retinal images) have been like a single, giant security camera that takes one photo of the whole street and tries to guess what's wrong with every car, building, and person in it all at once. This approach works okay if there's just one problem, like a single flat tire. But in the real world, especially as people get older or have diabetes, a single eye can have multiple problems happening at the same time—like a flat tire and a broken headlight and a pothole all in one spot.

The big challenge in this field of science, known as medical image analysis, is teaching computers to untangle these mixed-up problems. Standard computer programs usually treat every part of the eye image the same way, using the same "brain power" for a healthy spot as they do for a sick one. This paper tackles that by asking: What if we could build a computer system that doesn't just look at the whole picture, but sends different parts of the image to different specialized experts? Think of it like a hospital where a patient with a broken leg goes to an orthopedist, while a patient with a fever goes to an infectious disease doctor, rather than everyone seeing the same general practitioner. The goal is to make the computer smarter, faster, and easier to understand when it's diagnosing complex eye diseases.

The "Smart Hospital" for Eye Scans

In this paper, the researchers from Georgia State University and Piehealthcare Inc. propose a new way to analyze retinal fundus images (the colorful photos of the back of your eye). They noticed that standard deep learning models are a bit like a student trying to study for five different exams at the same time using the exact same study notes for every single subject. When diseases like Diabetic Retinopathy (DR), Age-Related Macular Degeneration (AMD), Epiretinal Membrane (ERM), and Glaucoma happen together, the computer gets confused because it tries to mash all those different "disease shapes" into one big, messy pile of information.

To fix this, the team built a system they call Saliency-Guided Sparse Expert Routing. That's a mouthful, so let's break it down with a fun analogy: imagine a massive, high-tech library where books (eye images) arrive every second.

1. The Spotlight (Saliency-Guided Tokenization)
First, the system uses a "spotlight" called Guided Context Gating (GCG). Instead of reading every single page of every book equally, this spotlight scans the cover and the table of contents to find the most important, "sick" parts of the image. It ignores the boring, healthy background (like the white of the eye) and zooms in on the messy, colorful spots where the disease actually is. This ensures the computer only pays attention to the clues that matter.

2. The Team of Specialists (Sparse Mixture-of-Experts)
Once the spotlight has found the important clues, the image is broken into tiny pieces called "tokens." These tokens are then sent to a Mixture-of-Experts (MoE) block. Think of this as a team of eight different doctors (experts) sitting in a room.

The Router: A smart traffic cop (the router) looks at each tiny piece of the image and decides which two doctors are best suited to examine it. If a piece looks like a hemorrhage, it goes to the "Bleeding Expert." If it looks like a specific type of scar, it goes to the "Scar Expert."
The Shared Doctor: To make sure the system doesn't forget the basics, there is one "Shared Expert" who looks at every piece of the image. This doctor handles the common stuff that all eyes have, like the blood vessels and the optic disc, so the other specialists can focus entirely on the weird, disease-specific stuff.
The Result: Instead of one giant brain trying to do everything, the system uses a "sparse" approach, meaning it only wakes up the specific experts needed for that specific image. This saves energy and, more importantly, keeps the different diseases from getting mixed up in the computer's memory.

3. The Final Diagnosis (Structural Decoding)
After the experts have done their work, the system combines their opinions. However, the researchers added a clever rule: if the computer thinks a patient is sick, it automatically knows they aren't "Normal." This prevents the system from giving contradictory answers, like saying a patient has a broken leg and is perfectly healthy at the same time.

What They Found

The team tested their new "Smart Hospital" system on a dataset of eye images containing five different categories: Diabetic Retinopathy, AMD, ERM, Glaucoma, and Normal. They compared their method against other top-tier computer models using a strict test where the computer had to guess the diseases for patients it had never seen before.

The results were impressive. Their new model achieved a macro AUC of 0.912 ± 0.008 and a macro F1 score of 0.653 ± 0.014. In plain English, this means the model was significantly better at spotting multiple diseases at once than the previous best methods, which hovered around 0.895 AUC.

But the coolest part isn't just the score; it's that the system is interpretable. The researchers could actually see how the computer was thinking.

The "Traffic" Map: When they looked at which "doctor" (expert) was chosen for which disease, they found a clear pattern. Diseases that look very different from each other (like AMD or Glaucoma) consistently sent their image pieces to specific, dedicated experts.
The "Mix" Test: When a patient had two diseases at once (like DR and ERM), the computer didn't just guess randomly. Instead, it sent parts of the image to the "DR Expert" and other parts to the "ERM Expert." The computer's internal map showed these mixed cases sitting right in the middle between the two separate disease groups, proving it had successfully untangled the two problems.
Visual Proof: Using a technique called Grad-CAM++, they showed that the computer was looking exactly at the right spots—like the optic disc for Glaucoma or the macula for AMD—just like a human doctor would.

Why This Matters

This paper suggests that by moving away from "one-size-fits-all" computer brains and toward a system that dynamically routes different parts of an image to specialized experts, we can get better at diagnosing complex eye diseases. The researchers measured that the routing of these experts is strongly linked to the specific disease present (with a statistical significance of p < 0.001), meaning the computer isn't just guessing; it's genuinely learning to separate the different conditions.

While the model is still a computer program and not a replacement for a human doctor, this approach offers a promising new way to screen for multiple eye diseases at once, potentially catching more problems earlier and helping doctors understand why the computer made its decision. The authors conclude that this "sparse" method, which mimics how we might send different parts of a problem to different specialists, is a powerful tool for untangling the messy reality of co-occurring medical conditions.

Technical Summary: Disentangling Co-Occurring Retinal Pathologies with Saliency-Guided Sparse Expert Routing

Problem Statement
Retinal fundus images frequently present multiple co-occurring pathologies (e.g., diabetic retinopathy alongside glaucoma), yet standard deep learning classifiers typically apply static, identical computation to every image. This approach forces networks to encode distinct, and often conflicting, disease morphologies within a single dense embedding. Consequently, when pathologies co-occur, their representations become entangled, degrading both classification accuracy and interpretability. While attention mechanisms can highlight salient regions, they do not alter the fundamental static nature of the computation; every spatial location is processed by the same feed-forward weights regardless of the underlying disease distribution.

Methodology
The authors propose a novel architecture that shifts retinal analysis from static processing to pathology-aware conditional computation. The pipeline consists of three primary stages:

Saliency-Guided Tokenization (Front-end):
To ensure the model focuses on clinically relevant evidence rather than background anatomy, the authors employ a Guided Context Gating (GCG) module. This acts as a spatial information bottleneck prior to tokenization. It utilizes:
- Global Context Attention (GCA): Aggregates spatial features weighted by pathological relevance to create a global context vector.
- Attention Gate (AG): Fuses local features with the global context to generate a selective spatial gating map.
  The resulting gated features are flattened into a sequence of lesion-centric tokens, ensuring the router receives evidence weighted by lesion saliency.
Sparse Mixture-of-Experts (MoE) with Shared Capacity:
The token sequence is processed by a sparsely-routed MoE transformer block. Key components include:
- Top-k Routing: A lightweight router dispatches each token to a subset of $k=2$ specialized experts out of $E=8$ total experts.
- Shared Expert: Inspired by DeepSeekMoE, an "always-on" shared expert processes every token. This acts as a structural anchor for common retinal anatomy (e.g., optic disc, vasculature), freeing the routed experts to specialize in the residual pathological variations that distinguish specific diseases.
- Architecture: The block uses RMSNorm, Multi-Head Self-Attention (MHSA), and SwiGLU feed-forward networks for the experts.
Multi-Task Learning with Structural Decoding:
Following Global Average Pooling (GAP), the embedding is passed to independent Multi-Layer Perceptron (MLP) heads for each of the five clinical states (DR, AMD, ERM, Glaucoma, Normal).
- Loss Function: The model is optimized using a dynamically weighted Binary Cross-Entropy (BCE) loss to address extreme class imbalance.
- Ontological Decoding: To respect clinical constraints (e.g., a patient cannot be both "Normal" and have a pathology), the continuous probabilities are projected onto a clinically valid boolean manifold post-prediction. This enforces mutual exclusivity between the healthy state and any pathology without penalizing the model during training for logical contradictions.

Key Contributions

Conditional Computation for Medical Vision: The paper adapts sparse MoE layers, typically used in Large Language Models, to the medical vision domain to disentangle co-occurring pathologies.
Interpretable Routing: The authors demonstrate that expert allocation is significantly disease-dependent ( $p < 0.001$ ). Morphologically distinct pathologies (e.g., ERM, AMD) isolate to dedicated experts, while diffuse states utilize shared pathways.
Saliency-Guided Tokenization: The integration of GCG ensures that the routing mechanism operates on features recalibrated for lesion saliency, rather than raw spatial patches.
Shared Expert Mechanism: The inclusion of a shared expert stabilizes training and prevents the "ubiquitous background" from dominating specific experts, allowing for cleaner disease specialization.

Experimental Results
The model was evaluated on a five-class, patient-disjoint 5-fold cross-validation benchmark derived from the ODIR 2019 dataset.

Performance: The proposed method achieved a macro AUC of 0.912 ± 0.008 and a macro F1 of 0.653 ± 0.014. This outperforms state-of-the-art baselines, including DKCNet (0.895 AUC) and EfficientNet Ensembles (0.875 AUC).
Ablation Study: The study confirmed that the GCG front-end provides a strong initial gain, while the sparse MoE block yields the largest improvement in F1 scores. The addition of the shared expert further improved both metrics by stabilizing routing entropy.
Qualitative Analysis:
- Grad-CAM++: Visualizations confirmed that the gated features attend to clinically relevant structures (e.g., optic disc for Glaucoma, macula for AMD) prior to routing.
- t-SNE Embeddings: Post-MoE embeddings showed a significant reduction in feature entanglement (silhouette score improved from -0.098 to +0.075). Co-occurring cases were observed to fall geometrically between their constituent single-disease clusters, indicating the model decomposes multi-label presentations rather than averaging them.
- Routing Analysis: Quantitative analysis showed that co-occurring pathologies activate a precise blend of constituent experts rather than a generic average.

Significance
The paper claims that sparse MoE offers an interpretable approach to multi-disease retinal screening. By transforming the routing distribution into a diagnostic signal, the model not only achieves state-of-the-art performance but also provides a data-driven decomposition of co-occurring conditions. The work suggests that dynamic conditional computation is better suited for disentangling complex medical pathologies than static dense networks, bridging the gap between high-performance deep learning and the need for clinical interpretability.

The "Smart Hospital" for Eye Scans

What They Found

Why This Matters

More like this