Modality-adaptive learning for glaucoma diagnosis with incomplete availability of fundus photographs and visual field data

Multimodal diagnostic models often assume that all input modalities are available, whereas clinical datasets commonly contain incomplete examination records. We developed CGMA-Net to evaluate glaucoma diagnosis under three modality-availability conditions: fundus photography only, visual field testing only, and paired fundus and visual field data. CGMA-Net combines modality-specific encoders, an explicit availability mask, and gated adaptive fusion to prevent absent inputs from contributing to the fused representation without imputing missing examinations. The model was evaluated in a retrospective two-center cohort comprising 5,181 samples from 3,262 participants, with participant-level separation of training, validation, and test sets. On the internal test set, CGMA-Net achieved an area under the receiver operating characteristic curve of 0.997, with 97.6% accuracy, 98.9% sensitivity, and 94.9% specificity, and maintained high discrimination across modality-availability conditions. In ablation analyses, removing gated fusion produced a larger performance decrease than removing cross-attention. Post hoc temperature scaling improved probability calibration without changing classification metrics. Both learned uncertainty and Monte Carlo dropout uncertainty were higher for misclassified than correctly classified samples. These results demonstrate the technical feasibility of modality-adaptive glaucoma classification across CFP-only, VF-only, and paired CFP–VF settings in a selected retrospective cohort, but do not establish performance in an unselected clinical population. Validation in a consecutive cohort that includes diagnostically ambiguous cases, relevant differential diagnoses, and examinations of variable quality is required before clinical use.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-18
DOI
https://doi.org/10.1038/s41598-026-81016-9
Primary Topic
Retinal Imaging and Analysis
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Modality-adaptive learning for glaucoma diagnosis with incomplete availability of fundus photographs and visual field data

Ke Si, Jiajia Chen, Wei Gong, Juan Ye et al.
4 links
Scientific Reports
Retinal Imaging and Analysis
article

Modality-adaptive learning for glaucoma diagnosis with incomplete availability of fundus photographs and visual field data

Ke Si, Jiajia Chen, Wei Gong, Juan Ye, Yao Zheng, Jiazhu Zhu
article en

Abstract

Multimodal diagnostic models often assume that all input modalities are available, whereas clinical datasets commonly contain incomplete examination records. We developed CGMA-Net to evaluate glaucoma diagnosis under three modality-availability conditions: fundus photography only, visual field testing only, and paired fundus and visual field data. CGMA-Net combines modality-specific encoders, an explicit availability mask, and gated adaptive fusion to prevent absent inputs from contributing to the fused representation without imputing missing examinations. The model was evaluated in a retrospective two-center cohort comprising 5,181 samples from 3,262 participants, with participant-level separation of training, validation, and test sets. On the internal test set, CGMA-Net achieved an area under the receiver operating characteristic curve of 0.997, with 97.6% accuracy, 98.9% sensitivity, and 94.9% specificity, and maintained high discrimination across modality-availability conditions. In ablation analyses, removing gated fusion produced a larger performance decrease than removing cross-attention. Post hoc temperature scaling improved probability calibration without changing classification metrics. Both learned uncertainty and Monte Carlo dropout uncertainty were higher for misclassified than correctly classified samples. These results demonstrate the technical feasibility of modality-adaptive glaucoma classification across CFP-only, VF-only, and paired CFP–VF settings in a selected retrospective cohort, but do not establish performance in an unselected clinical population. Validation in a consecutive cohort that includes diagnostically ambiguous cases, relevant differential diagnoses, and examinations of variable quality is required before clinical use.

Scientific Reports
Second Affiliated Hospital of Zhejiang University (CN), Zhejiang University (CN)
Fundamental Research Funds for the Central Universities, Key Research and Development Program of Zhejiang Province
Good health and well-being
Retinal Imaging and Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon