Classification of cystoid macular edema status from retinal vasculitis fluorescein angiography images using a convolutional neural network.
- 1Milton Academy, Milton, MA, United States.
- 2Massachusetts Eye Research and Surgery Institution, Waltham, MA, United States.
- 3Department of Biomedical Engineering, University at Buffalo, Buffalo, NY, United States.
Abstract
Introduction
Cystoid macular edema (CME) is a vision-threatening complication of retinal vasculitis, a rare inflammatory disease for which large imaging datasets are difficult to obtain. Detection of CME on fluorescein angiography (FA) requires expert interpretation, motivating automated image-based approaches to support clinical decision-making. Deep learning methods, particularly convolutional neural networks (CNNs), have shown strong performance in medical image analysis but are constrained by limited labeled data in rare diseases. Transfer learning may mitigate this limitation by adapting pretrained models to small, task-specific datasets.Methods
We analyzed 207 de-identified late-phase 55° FA images from 117 patients with retinal vasculitis. The primary target was concurrent patient-level CME status, established through specialist review of clinical records and optical coherence tomography obtained at the same clinical visit as FA. A patient was classified as CME-positive when CME was present in at least one eye. An ImageNet-pretrained ResNet-18 was evaluated using five-fold nested, patient-level cross-validation. Within each outer-development cohort, 144 candidate configurations were compared by four-fold inner cross-validation and selected according to mean inner-validation AUROC. Each selected configuration was evaluated only on its corresponding untouched outer-test fold.Results
Across 117 pooled patient-level outer-test predictions, the model-selection pipeline achieved an AUROC of 0.708 (95% CI, 0.610-0.802), an AUPRC of 0.626 (95% CI, 0.493-0.755), and an accuracy of 0.726 (95% CI, 0.641-0.803). Sensitivity was 0.694 (95% CI, 0.558-0.822), specificity was 0.750 (95% CI, 0.643-0.853), and the F1 score was 0.680 (95% CI, 0.563-0.780). Exploratory Grad-CAM analysis visualized regions associated with model predictions.Discussion
The nested model-selection pipeline showed moderate discrimination with a relatively balanced operating profile and slightly higher specificity than sensitivity. Clinically relevant false-positive and false-negative errors, together with the absence of external validation, preclude clinical deployment.
Related
- Longitudinal Quantitative Analysis of Ultra-Widefield Fluorescein Angiography Leakage Metrics: Analysis of the KINGFISHER Clinical Trial
- CBAM-enhanced convolutional networks for diabetic retinopathy classification in the ODIR dataset
- Feature-aware attention convolutional neural network for efficient pulmonary fibrosis classification from chest CT images with FPGA deployment
- MedFuse: dual-stream fusion of convolutional and vision transformer-based features for enhanced medical image classification
- COVLIAS 3.5: integration of attention-based segmentation technique with fuzzy dilated convolutional neural networks for improved classification of chest X-ray scans for multiclass pneumonia diagnosis
BibTeX
@article{wang2026classification,
title = {{Classification of cystoid macular edema status from retinal vasculitis fluorescein angiography images using a convolutional neural network.}},
author = {Celina Wang and Koushani Chakrabarty and Stephen D Anesi and Peter Y Chang},
journal = {Frontiers in ophthalmology},
year = {2026},
month = oct,
doi = {10.3389/fopht.2026.1901237},
url = {https://europepmc.org/article/MED/42908598},
note = {CC BY},
}
RIS
TY - JOUR
TI - Classification of cystoid macular edema status from retinal vasculitis fluorescein angiography images using a convolutional neural network.
AU - Celina Wang
AU - Koushani Chakrabarty
AU - Stephen D Anesi
AU - Peter Y Chang
PY - 2026
DA - 2026/10/07
JO - Frontiers in ophthalmology
DO - 10.3389/fopht.2026.1901237
UR - https://europepmc.org/article/MED/42908598
AB - Introduction Cystoid macular edema (CME) is a vision-threatening complication of retinal vasculitis, a rare inflammatory disease for which large imaging datasets are difficult to obtain. Detection of CME on fluorescein angiography (FA) requires expert interpretation, motivating automated image-based approaches to support clinical decision-making. Deep learning methods, particularly convolutional neural networks (CNNs), have shown strong performance in medical image analysis but are constrained by limited labeled data in rare diseases. Transfer learning may mitigate this limitation by adapting pretrained models to small, task-specific datasets.Methods We analyzed 207 de-identified late-phase 55° FA images from 117 patients with retinal vasculitis. The primary target was concurrent patient-level CME status, established through specialist review of clinical records and optical coherence tomography obtained at the same clinical visit as FA. A patient was classified as CME-positive when CME was present in at least one eye. An ImageNet-pretrained ResNet-18 was evaluated using five-fold nested, patient-level cross-validation. Within each outer-development cohort, 144 candidate configurations were compared by four-fold inner cross-validation and selected according to mean inner-validation AUROC. Each selected configuration was evaluated only on its corresponding untouched outer-test fold.Results Across 117 pooled patient-level outer-test predictions, the model-selection pipeline achieved an AUROC of 0.708 (95% CI, 0.610-0.802), an AUPRC of 0.626 (95% CI, 0.493-0.755), and an accuracy of 0.726 (95% CI, 0.641-0.803). Sensitivity was 0.694 (95% CI, 0.558-0.822), specificity was 0.750 (95% CI, 0.643-0.853), and the F1 score was 0.680 (95% CI, 0.563-0.780). Exploratory Grad-CAM analysis visualized regions associated with model predictions.Discussion The nested model-selection pipeline showed moderate discrimination with a relatively balanced operating profile and slightly higher specificity than sensitivity. Clinically relevant false-positive and false-negative errors, together with the absence of external validation, preclude clinical deployment.
SN - 2674-0826
ER -