An estimated 79.6 million people were affected by glaucoma in 2020.
That global figure is projected to rise to 111.8 million by 2040. To address early identification challenges across different hospitals and imaging devices, researchers presented a new neuro-symbolic glaucoma detection framework, known as NSC.
Architecture of the NSC Framework
The newly described system integrates three core components: an EfficientNet-B4 network utilizing dual spatial-channel attention, a six-node anatomical graph attention network called GAT-GNN, and a differentiable symbolic constraint module.
The symbolic constraint module applies four clinical fuzzy logic rules. These rules evaluate the cup-to-disc ratio, rim thinning, retinal nerve fiber layer defects, and vessel density.
Authors Trained the Framework on 9,100 Fundus Images
The study authors trained the framework on 9,100 fundus images drawn from a curated corpus of 12,619 images sourced across eight datasets and eight countries. Out of 12,134 non-holdout images, the dataset was split into 9,100 training images, 1,517 validation images, and 1,517 test images. The 485-image RIM-ONE DL set was held out entirely as a cross-hospital test set transitioning from Canon CR-2 to Nidek AFC-210 equipment without utilizing segmentation masks.
The full NSC model achieved a test area under the receiver operating characteristic curve of 0.946, with a 95% confidence interval ranging from 0.934 to 0.958, and a sensitivity of 0.900. Meanwhile, the EfficientNet-B4 plus GAT-GNN variant, referred to as B3, achieved the highest in-distribution AUC among evaluated variants at 0.9612, with a 95% confidence interval of 0.951 to 0.972.
| Model Configuration | Metric / Value |
|---|---|
| Full NSC Test AUC | 0.946 (95% CI 0.934–0.958) |
| Full NSC Sensitivity | 0.900 |
| Highest Cross-Hospital AUC (B4+Symbolic) | 0.6997 (95% CI 0.649–0.750) |
| B3 In-Distribution AUC | 0.9612 (95% CI 0.951–0.972) |
| B3 Mean Test AUC Across Three Runs | 0.9602 ± 0.0002 |
| B2 Mean Test AUC Across Three Runs | 0.9492 ± 0.0015 |
EfficientNet-B4 Plus Symbolic Achieved Highest Cross-Hospital AUC
Among the evaluated variants, the highest cross-hospital AUC of 0.6997 was obtained by EfficientNet-B4 plus Symbolic. Researchers noted that performance differences between the B2, B3, and B4 models on the cross-hospital holdout were small and not statistically significant.
The research team proposed the EfficientNet-B4 plus GAT-GNN configuration as the deployment setup based on in-distribution performance and inference cost rather than cross-hospital AUC. Ablation studies demonstrated that adding the GAT-GNN to EfficientNet-B4 improved the in-distribution test AUC by 1.4 percentage points in a single-seed ablation, moving from 0.9472 to 0.9612.
According to the published findings, agreement between the cup-to-disc ratio proxy and expert-measured values on images excluded from training was weak and inconsistent across datasets, recording a correlation coefficient of 0.549 across 26 samples on DRISHTI-GS and 0.081 across 100 samples on REFUGE2.
