WorksComputer Science
Auto-GlaKG: Automated Biomarker Reasoning for Explainable Glaucoma Detection
Abstract
Interpretable glaucoma detection requires a transparent connection between fundus images, clinically meaningful biomarkers, and diagnostic predictions. Knowledge graph (KG) reasoning provides this structure, but performance measured using annotated biomarkers does not establish performance when only images are available. We present Auto-GlaKG, an automated framework that combines image-based biomarker prediction with fixed clinical-prior rules. A ResNet50 network predicts the cup-to-disc ratio (CDR) and six morphological signs using biomarker supervision, and a KG reasoner aggregates their signed rule contributions into a glaucoma score. On 689 fundus images evaluated using stratified, group-aware five-fold cross-validation, Auto-GlaKG achieves a pooled area under the receiver operating characteristic curve (AUC) of 0.951 (95% confidence interval: 0.931–0.969), with sensitivity of 0.970 and specificity of 0.779. The matched image classifier also achieves an AUC of 0.951, while the rule stage adds 0.2~ms per image. Replacing predicted CDR with its annotation increases fold-mean AUC from 0.953 to 0.995. The framework provides an explicit rule-level account of image-based predictions and identifies CDR estimation as a priority for validation using independently measured biomarkers.
**Clinical Relevance—**Auto-GlaKG exposes predicted optic-nerve findings and their contributions to an automated assessment, enabling inspection of the evidence and of errors arising during biomarker estimation.
Introduction
Glaucoma causes irreversible vision loss, motivating automated methods for identifying suspicious optic-nerve changes in fundus images [1]. Deep learning has achieved promising discrimination [2], [3], but a diagnostic score alone provides limited information about the anatomical findings driving the prediction [4]. Linking image analysis to recognizable biomarkers, including the cup-to-disc ratio (CDR), neuroretinal rim thinning, and inferior–superior–nasal–temporal (ISNT) rule violation, can make the computational decision process easier to inspect [3], [5].
Concept bottleneck models represent decisions through intermediate concepts [6], [7], while knowledge graphs (KGs) organize biomarkers and their diagnostic relationships. Our previous GlaKG framework [8] explored structured biomarker reasoning using annotations available at inference. That setting establishes what the reasoner can do with supplied biomarker information; it does not assess the complete image-based pipeline. This distinction is especially important when biomarker descriptions are AI-generated and closely aligned with diagnosis labels [9].
We introduce Auto-GlaKG to evaluate automated biomarker reasoning under image-only inference. The framework predicts seven biomarkers and applies a fixed, graph-structured rule set to generate a glaucoma score and explicit rule contributions. Our contributions are:
An automated image-to-biomarker-to-score pipeline that requires no biomarker annotations at inference.
A controlled comparison of annotation-based and image-based reasoning under group-aware cross-validation.
An analysis of biomarker substitution, rule activation, and explanation agreement that identifies the main limitations of the automated pipeline on the evaluated dataset.
Methodology
Biomarker-Supervised Image Network
Let denote a fundus image, its diagnosis with glaucoma as the positive class, and its supplied biomarker targets. The targets comprise CDR , ISNT violation, rim thinning, notching, vessel bayoneting, rim pallor, and the laminar dot sign. An ImageNet-pretrained ResNet50 [10] with a seven-output head predicts , consisting of CDR and six sign probabilities .
The network is optimized using
where is smooth L1 loss and is weighted binary cross-entropy. Positive-class weights are derived from the training-fold biomarker prevalence; missing targets are masked. The CDR weight of 5 balances the smaller scale of the regression loss against the classification terms. This network uses biomarker supervision without an explicit diagnosis-classification loss.
Graph-Structured Clinical Rules
The KG contains image nodes, seven biomarker types, and ten rule nodes. Biomarker-to-rule relations encode the rule definitions, while weighted image-to-rule edges represent activation strengths (Figure 1). The graph organizes the evidence; prediction is computed by fixed additive rule aggregation.

Figure 1. Construction of an image-level evidence subgraph in Auto-GlaKG. Step 1: the fundus image enters the graph as an image node. Step 2: the biomarker network, trained without a diagnosis loss, predicts seven biomarkers, the CDR and six morphological signs, which become biomarker nodes. Step 3: each biomarker is linked to the clinical rules defined on it through schema-level relations, yielding ten candidate rules. Step 4: each rule activations are computed softly from the predicted biomarkers (red: glaucoma evidence; blue: normal evidence; dashed: ). Step 5: activated rules contribute strength-weighted evidence to the glaucoma-directed and normal-directed sums, whose signed difference defines the KG score and the probability . Step 6: the subgraph joins the global KG through the shared rule and biomarker nodes (the example is marked by a star). Highlighted rules and edges provide a compact visualization; all soft contributions, including those below the visualization threshold, remain in the score.
| Rule | Condition | Dir. | |
| CDR high | G | 3 | |
| CDR moderate | G | 2 | |
| CDR normal | N | 3 | |
| Rim thinning | Present | G | 3 |
| Notching | Present | G | 3 |
| ISNT violated | Present | G | 2 |
| ISNT followed | No violation | N | 2 |
| Bayoneting | Present | G | 2 |
| Rim pallor | Present | G | 1 |
| Laminar dot sign | Present | G | 1 |
Table 1. FIXED RULE SET. G AND N INDICATE GLAUCOMA-DIRECTED () AND NORMAL-DIRECTED () EVIDENCE.
Each rule has a condition , an evidence direction , and a strength (Table I). These settings encode simplified clinical priors and are fixed for the experiments. A CDR in activates no CDR rule under hard activation and is treated as indeterminate.
Soft Activation and Rule Contributions
Hard activation applies the rule conditions directly. To reduce discontinuities near CDR thresholds, soft activation uses
where is the threshold in Table I (0.7 and 0.5), is the logistic function, and . The moderate-CDR rule uses
Away from the CDR boundaries, these sigmoid memberships approach their hard-threshold counterparts as . Binary-sign rules use their predicted probabilities directly. is fixed for all experiments, and its effect is reported in Table III(a).
| Setting | Hard | Soft |
| Rule weighting | ||
| Clinical 3/2/1* | 0.938 | 0.951 |
| Uniform 1/1/1 | 0.933 | 0.949 |
| Learned (logistic, cross-fold) | 0.925 | 0.947 |
| Rule subset (number of rules) | ||
| CDR only (3) | 0.935 | 0.953 |
| + Rim thinning (4) | 0.936 | 0.953 |
| + ISNT (6) | 0.936 | 0.953 |
| + Notching (7) | 0.937 | 0.952 |
| All seven biomarkers (10)* | 0.938 | 0.951 |
| Image fusion | ||
| Without image* | 0.938 | 0.951 |
| With image | 0.949 | 0.946 |
| Temperature (soft) | AUC | ECE |
| 0.005 | 0.951 | 0.082 |
| 0.01 | 0.951 | 0.085 |
| 0.03* | 0.951 | 0.090 |
| 0.05 | 0.951 | 0.098 |
| 0.08 | 0.951 | 0.106 |
| 0.12 | 0.951 | 0.121 |
| Setting | Hard | Soft |
| Leave-one-rule-out (AUC, pooled) | ||
| CDR high | −0.0192 | −0.0050 |
| CDR moderate | −0.0003 | +0.0009 |
| CDR normal | −0.0108 | −0.0009 |
| Rim thinning | +0.0001 | −0.0001 |
| Notching | −0.0002 | +0.0012 |
| ISNT violated | +0.0001 | −0.0001 |
| ISNT followed | +0.0001 | −0.0001 |
| Bayoneting | −0.0011 | +0.0006 |
| Rim pallor | −0.0002 | +0.0003 |
| Laminar dot sign | −0.0001 | +0.0005 |
| Annotation substitution (fold-mean AUC, gap recovered) | ||
| No substitution | ||
| Annotated CDR | (67%) | (89%) |
| Annotated ISNT violation | (63%) | (71%) |
| Annotated rim thinning | (57%) | (58%) |
| Annotated notching | (40%) | (37%) |
| Annotated bayoneting | (38%) | (30%) |
| Annotated rim pallor | (34%) | (17%) |
| Annotated laminar dot sign | (25%) | (13%) |
| All annotated (oracle) |
Table III. KG ablations for hard and soft activation. (a) Pooled out-of-fold AUC under different rule weights, rule subsets, soft-threshold temperatures, and image fusion. (b) Change in pooled AUC when a single rule is removed, and fold-mean AUC when a single predicted biomarker is replaced by its annotation; percentages give the fraction of the gap between no substitution and annotation-based reasoning recovered by that biomarker. Fold-mean substitution results must be compared with the fold-mean baseline (no substitution). Defaults are marked with *.
The signed contribution of rule and the resulting prediction are
with and a positive decision when . Because this decision is equivalent to , affects calibration only, not AUC or threshold-dependent metrics. The reasoner has no trainable parameters, and its logistic output is evaluated as a prediction score.
The complete contribution vector exactly reconstructs . For visualization, rules with are highlighted and ordered by ; all remaining contributions are retained in scoring.
Image Fusion and Analysis Settings
An optional fusion combines the KG score with a separate diagnosis-supervised ResNet50:
The weight is selected by inner-validation AUC. Rule contributions describe the KG branch; they do not fully explain the image classifier or the fused prediction.
The annotation-based, or oracle, setting applies the same current reasoner to instead of . Here oracle denotes access to the supplied annotations, not independently verified clinical measurements. Single-biomarker substitution replaces one prediction with its annotation to measure the corresponding change in AUC. Hard-rule flips compare rule conditions evaluated on predicted and annotated biomarkers. Explanation agreement is the Jaccard index of the two sets of highlighted rules:
with when both sets are empty. These annotation-dependent analyses are retrospective diagnostics.
Experimental Setup
Dataset and Evaluation Protocol
Dataset
1) Dataset: We use a public dataset1 containing 689 fundus images: 427 glaucoma and 262 normal images drawn from ACRIMA [11], RIM-ONE [12], and Drishti-GS [13]. Seven biomarker targets are parsed from the accompanying AI-generated descriptions. Risk categories and annotation confidence scores are not inputs to the automated reasoner.
Evaluation Protocol
2) Evaluation Protocol: Perceptual-hash grouping with Hamming distance yields 652 groups, including 37 two-image groups. Stratified, group-aware five-fold cross-validation keeps detected near-duplicates in the same fold. Within each outer training fold, a group-aware split reserves 20% for fusion-weight selection, and both networks are trained on the remaining 80%. Outer test folds are excluded from training and fusion selection.
Training and Evaluation
Training
1) Training: Images are resized to and augmented using rotations (), scaling (0.9–1.1), translation (), horizontal flipping, and brightness and contrast jitter. Vertical flipping is excluded to preserve superior–inferior orientation. Both networks use AdamW with learning rate , weight decay , batch size 32, cosine learning-rate decay, and 15 training epochs. Experiments use PyTorch, seed 42, and four NVIDIA A100 GPUs (80 GB). Each ResNet50 has approximately 23.5 million parameters; reported inference times are 6.3 ms for one network and 0.2 ms for rule evaluation. Fusion additionally requires the separate image classifier.
Evaluation
2) Evaluation: We compare the image classifier, hard and soft Auto-GlaKG, and image fusion. Threshold-dependent metrics use a cutoff of 0.5, and AUC confidence intervals are obtained from 1,000 group-level bootstrap resamples. Fold-mean AUC is distinguished from pooled AUC throughout. Calibration is assessed by expected calibration error (ECE; ten bins). Exploratory AUC comparisons use paired DeLong tests [14] with Holm correction. Biomarker agreement is measured using CDR mean absolute error (MAE), Pearson correlation, and Cohen’s for binary signs.
Results and Discussion
Automated and Annotation-Based Performance
Table 2 reports pooled out-of-fold metrics. With image-only inference, soft Auto-GlaKG achieves a pooled AUC of 0.951 (fold mean 0.953 ± 0.023), matching the image classifier (0.951). The DeLong test detects no difference (, ), and the 95% confidence interval of the difference, [−0.020, 0.020], excludes changes larger than 0.02 in either direction. This bounds the difference but does not formally establish equivalence. At the default threshold, Auto-GlaKG favors sensitivity (0.970 versus 0.920; 13 versus 34 missed glaucoma cases) at the cost of specificity (0.779 versus 0.893; 58 versus 28 false positives), a trade-off that suits screening but increases the referral burden. This advantage does not extend to high-specificity operation: at 95% specificity, soft Auto-GlaKG reaches a sensitivity of 0.658 versus 0.724 for the image classifier, and its ECE (0.090) is the highest among all settings.
| Method | AUC [95% CI] | Fold AUC | AUPRC | Acc | BalAcc | Sens | Spec | Prec | F1 | MCC | Brier↓ | ECE↓ | Sens@95Spec |
| ResNet50 | 0.951 [0.929, 0.970] | 0.952 ± 0.028 | 0.955 | 0.910 | 0.907 | 0.920 | 0.893 | 0.934 | 0.927 | 0.810 | 0.075 | 0.058 | 0.724 |
| Auto-GlaKG (hard) | 0.938 [0.914, 0.960] | 0.938 ± 0.025 | 0.943 | 0.900 | 0.888 | 0.937 | 0.840 | 0.905 | 0.921 | 0.786 | 0.088 | 0.075 | 0.604 |
| Auto-GlaKG (soft) | 0.951 [0.931, 0.969] | 0.953 ± 0.023 | 0.962 | 0.897 | 0.874 | 0.970 | 0.779 | 0.877 | 0.921 | 0.782 | 0.087 | 0.090 | 0.658 |
| Auto-GlaKG (hard) + Image | 0.949 [0.927, 0.968] | 0.952 ± 0.026 | 0.961 | 0.903 | 0.897 | 0.920 | 0.874 | 0.923 | 0.922 | 0.794 | 0.082 | 0.070 | 0.700 |
| Auto-GlaKG (soft) + Image | 0.946 [0.925, 0.967] | 0.955 ± 0.025 | 0.955 | 0.912 | 0.901 | 0.944 | 0.859 | 0.916 | 0.930 | 0.811 | 0.072 | 0.044 | 0.600 |
Table 2. Diagnostic performance on pooled out-of-fold predictions. AUC intervals are 95% group-bootstrap confidence intervals; fold AUC is mean ± std over the five test folds. Best value per column in bold (lower is better for Brier and ECE). Acc: accuracy; BalAcc: balanced accuracy; Sens: sensitivity; Spec: specificity; Prec: precision; MCC: Matthews correlation coefficient; ECE: expected calibration error; Sens@95Spec: sensitivity at 95% specificity.
Soft fusion yields an AUC of 0.946, a marginally higher MCC (0.811 versus 0.810), and the lowest ECE (0.044) among all settings, but it does not improve discrimination. The selected fusion weight also varies widely across folds (), so no stable fusion preference emerges under this protocol.
At the rule level, the co-activation networks in Figure 2 contrast annotated and predicted biomarkers: annotated biomarkers activate two disjoint, near-deterministic rule families, whereas predicted biomarkers produce weaker and more heterogeneous co-activations.

Figure 2. Rule co-activation networks. A rule is counted as activated when . Node size is the activation frequency of a rule; edge width is the fraction of images in which two rules are activated together (red: both glaucoma evidence; blue: both normal evidence; grey: mixed). Annotated biomarkers yield two disjoint, near-deterministic rule families, whereas predicted biomarkers yield weaker and more heterogeneous co-activations.
Biomarker Prediction and Ablation
The predicted CDR has an MAE of 0.080 and a Pearson correlation of 0.81; Cohen’s ranges from 0.50 (laminar dot sign) to 0.77 (ISNT violation). Soft activation outperforms hard activation in every rule configuration (see Table III(a); 0.951 versus 0.938 by default), and affects calibration (ECE 0.082 to 0.121) rather than AUC. The CDR rules alone already reach 0.953, and no single rule removal changes the soft-activation AUC by more than 0.005, so the additional signs enrich the explanation more than the discrimination. Fusion helps only under hard activation (0.938 to 0.949). Annotation-based reasoning reaches a fold-mean AUC of 1.000, which defines the upper end of the gap in Table III(b). Substituting the annotated CDR raises the fold-mean AUC from 0.953 to 0.995, recovering about 89% of the gap, followed by ISNT violation (71%) and rim thinning (58%). These fractions overlap and are not a causal decomposition, but they identify the CDR as the main bottleneck.
Rule Agreement & Error Analysis
For soft Auto-GlaKG + Image, flips of CDR, ISNT, and rim-thinning rules are associated with more errors than flips of weaker rules. Mean Jaccard agreement between predicted and annotation-derived rule sets is 0.76: 0.82 for the 628 correctly classified images and 0.13 for the 61 errors. This association is descriptive, since agreement concerns the KG branch and the outcome concerns the fused model, and it requires reference annotations, so it cannot serve as a deployment-time error detector.
Conclusion
Auto-GlaKG combines automated biomarker prediction with fixed, graph-structured clinical rules for explainable glaucoma detection. It achieves pooled AUC 0.951 on the evaluated dataset and provides an exact decomposition of the KG score. Comparing annotation-based and automated reasoning exposes the effects of biomarker prediction errors, with CDR substitution producing the largest improvement in this dataset. These findings motivate further evaluation with expert biomarker measurements and independent data sources.
Limitations & Future Work
Limitations
The AI-generated biomarker targets are strongly associated with the diagnosis and have not been validated as clinical measurements, so omitting a diagnosis loss does not guarantee label-independent learning. The small, pooled dataset and hash-based grouping limit conclusions about patient-independent and cross-source performance. Paired tests do not fully account for within-group dependence, and inference times are specific to our hardware and implementation.
Future Work
Future work will evaluate CDR derived from optic disc and cup segmentation [15], [16], using expert annotations and documented grading procedures. Independent-source evaluation and ophthalmologist reader studies are needed to assess generalization and the usefulness of the rule-level evidence.
References
References
- [1]Vision Loss Expert Group of the Global Burden of Disease Study and GBD 2019 Blindness and Vision Impairment Collaborators, “Global estimates on the number of people blind or visually impaired by glaucoma: A meta-analysis from 2000 to 2020,” Eye, 2024.
- [2]D. Meedeniya, T. Shyamalee, G. Lim, and P. Yogarajah, “Glaucoma identification with retinal fundus images using deep learning: Systematic review,” Informatics in Medicine Unlocked, 2025.DOI
- [3]A. M. Abd El-Gawad, S. Hassan, M. Elsharkawy, S. Al Hamadani, A. Shivel, T. Couch, I. Saleh, E. A. Atallah, M. Ghazal, G. A. Girdharan, H. M. Amer, A. T. Khalil, and A. El-Baz, “Explainable ai for glaucoma detection and classification: A comprehensive review,” Artificial Intelligence Review, 2026.
- [4]C. Huang, “AI-Driven Biomarker Discovery & Progression Modeling for Precision Diagnosis of Glaucoma,” Ph.D. dissertation, Southern Methodist University, Dallas, TX, USA, 2026.
- [5]W.-H. Yang, Y.-W. Xu, X.-H. Sun et al., “Guidelines for glaucoma imaging classification, annotation, and quality control for artificial intelligence applications,” International Journal of Ophthalmology, 2025.DOI
- [6]Y. Liu, T. Zhang, and S. Gu, “Hybrid concept bottleneck models,” in 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).DOI
- [7]C. Huang, Z. Han, W. Xie, K. Kooner, T. Lee, J.-K. Wang, and J. Zhang, “Glaucoma: A multimodal structured framework for glaucoma risk stratification,” in Proceedings of the 48th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2026.arxiv.org/abs/2508.03750
- [8]C. Huang, J. Zhang, Y. Jiang, Y. Liu, K. Kooner, Y. Liu, T. Lee, Y. Xie, W. Shi, and G. Xiao, “Glakg: A biomarker-centric fundus knowledge graph for explainable glaucoma diagnosis and risk assessment,” in 2026 IEEE International Conference on Bioinformatics and Biomedicine (BIBM).arxiv.org/abs/2607.04673
- [9]M. Havasi, S. Parbhoo, and F. Doshi-Velez, “Addressing leakage in concept bottleneck models,” in Advances in Neural Information Processing Systems (NeurIPS), 2022.DOI
- [10]K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).arxiv.org/abs/1512.03385
- [11]A. Díaz-Pinto, S. Morales, V. Naranjo, T. Köhler, J. M. Mossi, and A. Navea, “Cnns for automatic glaucoma assessment using fundus images: An extensive validation,” Biomedical Engineering Online, 2019.
- [12]F. Fumero, S. Alayón, J. L. Sanchez, J. Sigut, and M. Gonzalez-Hernandez, “Rim-one: An open retinal image database for optic nerve evaluation,” in 2011 IEEE 24th International Symposium on Computer-Based Medical Systems (CBMS).DOI
- [13]J. Sivaswamy, S. R. Krishnadas, G. D. Joshi, M. Jain, and A. U. S. Tabish, “Drishti-gs: Retinal image dataset for optic nerve head segmentation,” in 2014 IEEE International Symposium on Biomedical Imaging (ISBI).DOI
- [14]E. R. DeLong, D. M. DeLong, and D. L. Clarke-Pearson, “Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach,” Biometrics, 1988.pubmed.ncbi.nlm.nih.gov/3203132
- [15]H. Fu, J. Cheng, Y. Xu, D. W. K. Wong, J. Liu, and X. Cao, “Joint optic disc and cup segmentation based on multi-label deep network and polar transformation,” in Medical Image Computing and Computer Assisted Intervention – MICCAI 2018.DOI
- [16]J. I. Orlando, H. Fu, J. Barbosa Breda, K. van Keer, D. R. Bathula, A. Diaz-Pinto et al., “Refuge challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs,” Medical Image Analysis, 2020.DOI