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

The paper proposes a novel architecture for multi-label retinal fundus image classification that addresses the challenge of co-occurring pathologies. The proposed model combines a Guided Context Gating (GCG) spatial attention front-end with a sparsely-routed Mixture-of-Experts (MoE) block. The GCG module recalibrates backbone features based on global lesion saliency before tokenization, ensuring the router receives clinically relevant evidence. The MoE block uses top-2 routing among 8 experts, with a DeepSeek-style always-on shared expert that anchors common retinal anatomy, allowing routed experts to specialize on disease-specific variation. The model is trained with a dynamically weighted Binary Cross-Entropy loss and includes an Ontological Decoding step that enforces clinical priors: any pathology strictly negates the healthy state, and under extreme uncertainty, the most probable class is forced.

On a five-class (DR, AMD, ERM, Glaucoma, Normal) patient-disjoint 5-fold cross-validation benchmark derived from the ODIR 2019 dataset, the model achieves a macro AUC of 0.912 ± 0.008, macro F1 of 0.653 ± 0.014, and accuracy of 0.649 ± 0.011. This outperforms state-of-the-art baselines including VGG16 Baseline (0.852 mAUC), TL-CNN (0.868), EfficientNet Ensemble (0.875), DCNet (0.882), and DKCNet (0.895), with improvements of +0.017 mAUC and +0.038 mF1 over the best baseline. Ablation studies show that GCG provides a strong front-end gain, the sparse MoE contributes the largest F1 improvement, and the shared expert further improves both metrics by stabilizing routing and lowering routing entropy.

Qualitative evaluations demonstrate that expert routing is significantly disease-dependent (p < 0.001), with morphologically focal pathologies (ERM, AMD, Glaucoma) isolating to dedicated experts, while diffuse states (DR, Normal) route broadly. Grad-CAM++ visualizations confirm that gated features attend to clinically relevant structures: DR shows scattered focal activations, AMD and ERM activate the macular region, and Glaucoma localizes to the optic disc. t-SNE analysis shows the silhouette score over five classes rises from −0.098 (pre-MoE) to +0.075 (post-MoE), and co-occurring cases settle in corridors between their constituent single-disease clusters. For dual-label images, joint routing strongly matches the union of constituent disease prototypes (similarities 0.978 for DR+ERM and 0.969 for DR+Glaucoma), proving that sparse routing decomposes multi-label presentations into precise blends of experts rather than generic averages.

Improvements to AI systems:

1. Add a spatial attention front-end (GCG) that recalibrates features based on global lesion saliency before tokenization. This improves the router’s input quality, leading to more clinically relevant evidence extraction and higher AUC (+0.017 over best baseline) in multi-label medical imaging.

2. Integrate a sparsely-routed Mixture-of-Experts (MoE) block with top-2 routing among 8 experts, plus an always-on shared expert. This stabilizes routing (lower routing entropy), increases F1 (+0.038), and allows the model to decompose multi-label presentations into precise blends of disease-specific experts rather than generic averages—critical for co-occurring pathologies.

3. Implement a dynamically weighted Binary Cross-Entropy loss to handle class imbalance and co-occurrence, improving macro F1 and accuracy without manual tuning.

4. Add an Ontological Decoding step that enforces clinical priors: (a) any pathology strictly negates the healthy state, and (b) under extreme uncertainty, force the most probable class. This reduces false positives for “Normal” and improves clinical plausibility of predictions.

5. Use disease-dependent expert routing as an interpretability mechanism. Since routing is significantly disease-dependent (p<0.001), the system can output which experts fired for a given image, providing a human-readable explanation of which pathological prototypes contributed to the decision.

6. Leverage the t-SNE/clustering improvement (silhouette score from −0.098 to +0.075) to build a more separable feature space. This enables better downstream tasks like patient stratification, disease progression tracking, or retrieval of similar cases.

7. Exploit the joint routing similarity for dual-label images (0.978 for DR+ERM, 0.969 for DR+Glaucoma) to build a system that can predict co-occurring diseases by combining single-disease expert activations, enabling compositional generalization to unseen multi-label combinations.

What the improved AI system can do:

• Clinically accurate multi-label diagnosis on retinal fundus images, outperforming state-of-the-art (mAUC 0.912, mF1 0.653) with robust 5-fold cross-validation.

• Explainable predictions via expert routing maps and Grad-CAM++ visualizations, showing exactly which disease-specific features (e.g., macular region for AMD/ERM, optic disc for Glaucoma, scattered foci for DR) drove the decision.

• Handle co-occurring pathologies by decomposing them into blends of single-disease experts, avoiding the “average” failure mode of standard classifiers.

• Enforce clinical consistency (e.g., never predicting “Normal” alongside a pathology) and provide a fallback under high uncertainty, reducing risky misclassifications.

• Generalize to new multi-label combinations without retraining, because the sparse MoE learns compositional disease prototypes.

• Provide a routing entropy metric as a confidence score—low entropy indicates high diagnostic certainty, high entropy flags ambiguous cases for human review.

Retinal fundus images frequently exhibit multiple co-occurring pathologies, yet standard deep learning classifiers apply static, identical computation to every image regardless of the underlying disease distribution. We propose a novel architecture that resolves this via sparse conditional computation, pairing a Guided Context Gating (GCG) spatial attention front-end with a sparsely-routed Mixture-of-Experts (MoE) block operating over feature tokens. Crucially, this routing yields an interpretable, data-driven decomposition. Expert allocation is significantly disease-dependent (p < 0.001), with the healthy Normal state and morphologically distinct pathologies (e.g., ERM, AMD) isolating to dedicated experts. On a five-class, patient-disjoint 5-fold cross-validation benchmark, our model achieves 0.912 +/- 0.008 macro AUC and 0.653 +/- 0.014 macro F1. Furthermore, Grad-CAM++ and post-MoE t-SNE visualizations confirm that expert routing aligns with localized lesions and geometrically maps co-occurring cases between their constituent clusters, positioning sparse MoE as an interpretable approach to multi-disease retinal screening.