médical IA

Five retinal diseases, "experts" that specialise: what a sparse Mixture-of-Experts reveals on fundus photographs

Published on August 12, 2026 · 9 min read

Nagur Shareef Shaik and colleagues propose an architecture that does not apply the same computation to every fundus image: a "sparse" routing mechanism sends each photograph to a specialised sub-network — an "expert" — depending on the pathology present, preceded by a lesion-guided attention module. On a five-class set (healthy eye, diabetic retinopathy, AMD, epiretinal membrane, glaucoma) evaluated with patient-disjoint cross-validation, the model reaches a macro AUC of 0.912 and a macro F1 of 0.653, and, crucially, its experts genuinely specialise by disease (p < 0.001). It is an elegant contribution to interpretability, but the gap between the flattering AUC and the modest F1, together with the absence of external validation, is a reminder that a neat mechanism is not yet a screening tool.

The context

Fundus photography is a common, inexpensive examination: a colour image of the retina, optic nerve and blood vessels, captured in a few seconds. It is one of the few places in the body where vessels and neural tissue can be observed directly, which makes it fertile ground for automated screening — diabetic retinopathy being the most mature example.

The problem: in real life, a single eye often accumulates several conditions. An elderly diabetic patient may present with diabetic retinopathy, age-related macular degeneration (AMD) and early glaucoma all at once. Yet most deep-learning classifiers process every image with exactly the same computation, whatever it contains: a single "monolithic" network expected to do everything. That uniformity sits poorly with a reality in which the signs of AMD (deposits at the centre of the macula) have nothing in common with those of glaucoma (cupping of the optic nerve) or an epiretinal membrane (a thin fibrous veil on the retinal surface).

The Mixture-of-Experts (MoE) idea answers exactly this: rather than a single network, several are trained — the "experts" — and a routing module decides, for each image, which ones to activate. When routing is sparse, only a small fraction of experts work on a given image: you gain capacity and specialisation without paying the compute cost of the whole network every time. This machinery, already popular in large language models, is imported here into retinal screening — coupled with the team’s earlier work, Guided Context Gating, an attention module that learns to focus on the salient lesions in the image.

The method

The architecture chains two blocks. At the input, Guided Context Gating (GCG): a spatial attention mechanism that weights image regions by relevance — it learns, in effect, where to look, foregrounding lesioned areas rather than the uniform retinal background. The filtered representations are split into tokens (feature tokens, the equivalent of small encoded image patches), then passed to the Mixture-of-Experts block.

There, a gating network assigns each token to a restricted subset of experts — this is sparse conditional computation: the path taken through the network depends on the image content. The authors’ hypothesis is that, if training goes well, different experts should end up specialising in different pathologies, offering an interpretable view of what the model "sees".

Evaluation covers five classes: healthy eye, diabetic retinopathy, AMD, epiretinal membrane and glaucoma. It is run as patient-disjoint five-fold cross-validation: the data are split into five parts, training on four and testing on the fifth, rotating — and, crucially, images from the same patient never appear in both training and test. This detail matters: mixing one patient’s eyes across train and test is a classic form of data leakage that artificially inflates scores. Here the authors guard against it.

Two metrics are reported. AUC (area under the ROC curve) measures the model’s ability to rank a positive case above a negative one; 0.5 is chance, 1.0 perfection. F1 combines precision (of the cases flagged, how many are correct) and recall (of the true cases, how many are found). The macro prefix means averaging equally across the five classes: a rare disease counts as much as a common one, which is more demanding — and more honest — than a global average dominated by the majority class.

The results

The model reaches a macro AUC of 0.912 (± 0.008) and a macro F1 of 0.653 (± 0.014). The small standard deviations across folds indicate stable performance from one split to another.

The central result, though, is not the raw score but the specialisation of the experts: expert allocation depends significantly on disease (p < 0.001, meaning such an association would be very unlikely by chance alone). The healthy eye and morphologically distinct pathologies — epiretinal membrane, AMD — are assigned dedicated experts. Visualisations via Grad-CAM++ (a heatmap of the regions that drove the decision) and t-SNE (a two-dimensional projection of internal representations) confirm that routing aligns with localised lesions, and that cases where several pathologies coexist sit between the clusters of their components — exactly what one would expect of a model that "understands" co-occurrence rather than forcing it into a single box.

Clinical translation, cautiously. An AUC of 0.912 looks excellent, but the macro F1 of 0.653 tells another story: averaged over the five classes, the precision/recall balance remains mediocre, betraying numerous errors on at least some categories — probably the less frequent or more subtle ones, such as early glaucoma. In other words, out of 1,000 screened images, this model would correctly sort most healthy eyes and clear-cut cases, but would miss or misclassify a non-trivial share of the discreet pathologies. The abstract does not provide the per-class breakdown, which prevents saying precisely which ones suffer — a limitation in itself.

What is good

Interpretability anchored in the structure, not bolted on afterwards. Many works add a saliency map at the end to "explain" a network that remains opaque. Here interpretability is constitutive: routing to experts is an internal mechanism, and its dependence on disease is established statistically (p < 0.001). This is a measurable explanation, not an illustration.

Taking co-occurrence seriously. Treating retinal pathologies as mutually exclusive is a comfortable but false simplification. Showing, via t-SNE, that mixed cases sit between their components is a concrete way to account for the multi-pathology reality of an eye, which most classifiers flatten.

Sound experimental hygiene. Patient-disjoint five-fold validation avoids the trap of per-patient data leakage, a source of spectacular but illusory AUCs in the imaging literature. It is no feat, but it is rigour worth crediting because it is too often neglected.

What is less good

No external validation. Everything rests on internal cross-validation, on a single five-class set. Nothing is known about performance on other cameras, other populations, other countries. This is population bias par excellence: a model calibrated on a single distribution routinely collapses as soon as the camera or the demographics change. Without an external cohort, the 0.912 AUC says nothing about generalisation.

A wide gap between AUC and F1, and aggregate figures that hide. To see 0.912 AUC and rejoice would be to fall for the misleading metric. On an imbalanced multi-class task, AUC can stay high while F1 — closer to real use — stays middling. And reporting only aggregate macro scores conceals which pathologies the model misses. For screening, knowing that early glaucoma is missed is no detail: it is the decisive information, and it is absent.

Interpretable does not mean correct. A routing that specialises and pretty Grad-CAM++ maps do not prove the model relies on the true pathological signal. Saliency maps are known to be fragile, and an expert "dedicated to AMD" could just as well capture a dataset artefact correlated with AMD — shortcut learning, where the model learns a spurious shortcut rather than the disease. The apparent coherence between routing and lesions is encouraging, but it does not replace a check against this failure mode. Add a grey area over reproducibility: the dataset size and source, and the availability of code, are not established at this stage.

What it changes

For the research community, the paper illustrates an interesting direction: making Mixture-of-Experts not merely a scaling tool but an instrument of interpretability for multi-disease screening. Analysing routing as a diagnostic of what a model has actually learned is an idea reusable beyond the retina. It is a methodological contribution, not a clinical breakthrough.

For clinicians, the immediate impact is nil. This is an internally validated prototype, with no external cohort, no prospective real-world evaluation, no regulatory status. Retinal screening tools already exist in practice — notably for diabetic retinopathy — but this work sits upstream, at the stage of architectural proof of concept. As with other retinal imaging approaches we have decrypted (see our analysis of inner-retinal layer segmentation in OCT), the distance between a good score and clinical use remains considerable.

For patients and the public, the interest is conceptual: we move from a "one model, one disease" logic to systems that recognise an eye can accumulate several problems at once — closer to the reality of a consultation. But a benchmark result is not a deployed device, and the question of population bias — will the model work as well across all populations and all devices? — remains wide open, as for much imaging AI (see our decryption of dataset-origin signatures in mammography).