Convolutional Graph Isomorphism Network to Detect Glaucomatous Visual Field Defects

To evaluate the performance of a deep learning (DL) model based on graph isomorphism networks (GINs) for detecting glaucomatous visual field defects on 24-2 standard automated perimetry (SAP) and to compare it against traditional diagnostic criteria, a dense neural network (NN) model, and a convolutional neural network (CNN) model.

A cross-sectional retrospective study. Participants: 1874 reliable SAP tests (Humphrey Field Analyzer, Carl-Zeiss Meditec Inc.) from 1009 eyes of 676 patients.

Standard automated perimetry tests were classified as normal or abnormal due to glaucomatous damage by two glaucoma specialists, with adjudication by a third. A GIN architecture was developed to classify tests using full 54-point spatial SAP data modeled as graphs, with node features comprising sensitivity, total deviation, and pattern deviation values. The dataset was split at the patient level (60% training/validation, 40% testing). The GIN model's diagnostic performance was compared to the Anderson criteria, the glaucoma hemifield test/pattern standard deviation (GHT/PSD) criteria, a fully connected dense NN, and a CNN model.

Area under the receiver operating characteristic curve (AUC), precision-recall curve, sensitivity at 95% specificity, F1-score, repeatability, and model explainability.

Among the 1874 SAP tests, 70.0% were graded as abnormal. The GIN model achieved an AUC of 0.982, significantly outperforming the Anderson criteria (AUC: 0.906, P < 0.001), GHT/PSD (AUC: 0.936, P = 0.006), the NN model (AUC: 0.941, P = 0.007), and the CNN model (AUC: 0.941, P = 0.027). At 95% specificity, the GIN model reached the highest sensitivity of 94.1%, surpassing the NN model (88.3%), CNN model (92.0%), GHT/PSD (90.1%), and Anderson criteria (85.1%). The GIN model also achieved the highest average precision (0.952) among evaluated criteria. Explainability analysis using GraphNOSE demonstrated that the GIN model emphasized clinically relevant regions associated with glaucomatous loss, offering interpretability advantages over conventional DL approaches.

By modeling SAP as a graph and incorporating spatial relationships among test points, the GIN model provided superior diagnostic performance and interpretability relative to traditional criteria and standard NNs. This graph-based approach offers a promising tool for accurate and explainable detection of glaucomatous visual field defects in clinical practice.

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.