Neuro-symbolic glaucoma detection meets the harder test of cross-hospital imaging
A glaucoma detection system that can perform well inside one dataset but falter at another hospital is of limited bedside value. This study tackles that problem directly with glaucoma screening from fundus images, pairing a convolutional backbone with anatomical graph reasoning and a differentiable symbolic constraint layer.
The central question is not whether a model can memorize retinal images. It is whether it can recognize disease when the camera, color profile, and acquisition setting change.
Investigators built the neuro-symbolic framework, or NSC, around three components: EfficientNet-B4 with dual spatial-channel attention, a six-node anatomical graph attention network, and a symbolic module encoding four fuzzy clinical rules tied to cup-to-disc ratio, rim thinning, retinal nerve fiber layer defects, and vessel density. The architecture was trained on 9,100 fundus images drawn from a curated pool of 12,619 images across eight datasets and eight countries. The remaining non-holdout images were divided 75% for training and 12.5% each for validation and testing, while a 485-image RIM-ONE DL set was withheld entirely as a never-seen cross-hospital test set.
Cross-hospital performance lagged behind in-distribution results
The sharpest finding comes from the external setting. Among the evaluated variants, the best cross-hospital AUC was 0.6997 for EfficientNet-B4 plus the symbolic module, with a 95% confidence interval of 0.649 to 0.750. The differences between the B2, B3, and B4 variants on the cross-hospital holdout were small and not statistically significant, which is an important signal for any model intended for real-world deployment across institutions.
In other words, the model still struggled when it left the environment in which it was trained.
Internal performance looked much stronger. The full NSC model reached a test AUC of 0.946, with sensitivity of 0.900, and the B3 variant achieved the highest in-distribution AUC among the evaluated models at 0.9612. Across three independent runs using seeds 42, 1337, and 2024, B3 maintained a mean test AUC of 0.9602 ± 0.0002, compared with 0.9492 ± 0.0015 for B2. Those repeated runs suggest some stability inside the study’s own split, but they do not erase the external gap.
Anatomical reasoning improved internal discrimination more than external transportability
The study’s mechanistic premise is clinically sensible. Glaucoma damages the optic nerve head and surrounding retinal architecture, so a model that explicitly reasons about cup-to-disc ratio, rim thinning, RNFL defects, and vessel density should be better aligned with how clinicians inspect fundus images. The graph attention network adds anatomical structure, while the symbolic layer tries to enforce clinically plausible relationships among features that are otherwise learned implicitly by a deep net.
Yet the cross-hospital result shows the limits of that strategy. When the symbolic component was combined with the graph network, cross-hospital performance fell by 8.1 percentage points in the GNN-symbolic interaction analysis, with z = −3.08 and p = 0.002. That reduction is a warning sign: formalizing anatomy does not automatically make a model more portable across devices or sites.
The authors also provide a practical deployment argument for the B3 configuration, favoring it over the full NSC model on the basis of in-distribution performance and inference cost rather than cross-hospital AUC. In a screening workflow, computational efficiency matters, but so does the price of missing glaucoma in a different hospital’s imaging stream. A model that is fast but brittle can still fail the clinical use case it was built to serve.
The cup-to-disc proxy remains biologically unresolved
One especially important detail is the mismatch between the model’s cup-to-disc ratio proxy and expert measurement on images excluded from training. Agreement was weak and inconsistent across datasets, with rho = 0.549 in DRISHTI-GS on 26 images and rho = 0.081 in REFUGE2 on 100 images. Those values do not establish anatomical grounding for the proxy.
This matters because cup-to-disc ratio is not just another numeric output. It is a structural marker of optic nerve head excavation, and if the learned representation does not reliably match expert measurement, the model’s interpretability is incomplete. The same concern extends to the latent representations for CDR, RNFL, rim, and vessel features, which still require validation against segmentation ground truth.
The architecture therefore occupies a useful but unsettled space. It is more clinically informed than a black-box classifier, but the study does not yet prove that each internal representation corresponds robustly to the retinal anatomy it is meant to encode.
Why external generalization remains the decisive endpoint
The external holdout used images from a different camera pipeline, moving from Canon CR-2 to Nidek AFC-210, and the model was tested without segmentation masks. That setting is closer to how glaucoma screening would have to function in practice. Performance drops under those circumstances are common in medical imaging, especially when color balance, resolution, and acquisition protocols shift from site to site.
The broader implication is straightforward: glaucoma AI still needs cross-hospital validation that is strict enough to expose domain shift rather than mask it. The study shows that a neuro-symbolic framework can improve internal performance and can be built without segmentation supervision, but the best external AUC remained modest. For now, the most convincing result is not that the model solves generalization. It is that the problem has been measured honestly.
The study also leaves open a practical question for future work: whether stronger anatomical supervision, better calibration across devices, or more diverse multicenter training data can lift the cross-hospital AUC without sacrificing efficiency. Until then, the model’s strongest claim is methodological rather than clinical deployment-ready, and its weakest point is still the same one that has challenged many retinal AI systems—transporting accuracy from one hospital to another.