Retinal vasculitis is an uncommon condition in which inflammation damages the blood vessels of the retina, the light-sensitive tissue lining the back of the eye. The inflammatory process breaks down the blood-retina barrier, the tightly regulated interface that normally keeps fluid and immune cells from seeping into retinal tissue. When that barrier fails, blood plasma leaks from the vessels and accumulates within the retina, often producing cystoid macular edema, or CME, a swelling of the macula, the small central region responsible for sharp, detailed vision. If left untreated, chronic CME can cause irreversible retinal damage and permanent vision loss, making its timely detection one of the most consequential tasks in managing these patients.

Diagnosing CME typically relies on optical coherence tomography, or OCT, which provides direct structural evidence of fluid-filled spaces within the retina, and on fluorescein angiography, or FA, an imaging technique in which a fluorescent dye is injected into the bloodstream and photographed as it circulates through the retinal vessels. In late-phase FA images, captured roughly three minutes after injection, leaking vessels glow with hyperfluorescence, and experienced specialists look for patterns such as para- and perifoveal hyperfluorescence within the temporal vascular arcades, diffuse punctate hyperfluorescence outside the macular region, and late-phase optic disc hyperfluorescence. But interpreting these images is subjective, time-consuming, and demands expertise that is not evenly distributed across clinical settings, particularly in resource-limited environments where access to seasoned retinal specialists is restricted.

The research team, drawn from the Massachusetts Eye Research and Surgery Institution and collaborators, assembled a retrospective dataset of 207 de-identified late-phase FA images with a 55-degree field of view, obtained from 117 patients with idiopathic retinal vasculitis using a Heidelberg Spectralis imaging system. The cohort included 73 women and 44 men, and 90 patients contributed images from both eyes while 27 contributed images from only one eye. Ground truth was established through specialist review of clinical records and OCT scans obtained at the same clinical visit as the angiography. Crucially, the team defined the outcome at the patient level: a patient was labeled CME-positive if the swelling was present in at least one eye, a rule that yielded 49 CME-positive and 68 CME-negative patients. Patients whose CME predated their retinal vasculitis diagnosis were excluded, since the swelling could not be attributed to the inflammatory disease itself.

At the heart of the study sits a ResNet-18, a compact but powerful convolutional neural network whose architecture uses residual, or skip, connections that allow signals to bypass layers during training. These connections mitigate the vanishing gradient problem, a phenomenon in which the error signals that guide learning become exponentially attenuated as they propagate backward through deep networks, and they enable effective feature reuse across layers. Rather than training from scratch, which typically demands large labeled datasets and invites overfitting on small cohorts, the researchers employed transfer learning: the network was initialized with weights pretrained on ImageNet, a massive database of millions of natural images. Although natural photographs differ radically from angiograms, pretrained networks learn general-purpose visual features such as edges, textures, and spatial compositions that transfer remarkably well to medical imaging tasks. The team converted the FA images to single-channel grayscale, averaged the pretrained RGB filter weights across the three color channels to initialize the modified first convolutional layer, and replaced the original classification head with a dropout layer followed by a fully connected layer producing two logits, one representing evidence for CME and one against it.

One of the study’s most technically interesting design choices is its treatment of each patient as a variable-size multiple-instance learning bag. Because some patients had images from both eyes and others from only one, the model processed all available images from a patient through a shared encoder and computed an image-level CME log-odds score, the difference between the positive and negative logits. The maximum image-level score across all of a patient’s images then served as the patient-level score, mirroring the clinical rule that pathology in either eye dictates patient-level management. Left-eye images were mirrored before encoding to standardize laterality, random horizontal flipping with a probability of 0.5 was applied only to training images, and normalization statistics were computed exclusively from each training partition to prevent information from leaking between data splits.

Perhaps the study’s most rigorous feature is its evaluation framework. The team used five-fold nested, patient-level cross-validation, a scheme in which an outer loop carves out untouched test cohorts of patients while an inner loop searches among 144 prespecified hyperparameter configurations, spanning learning rates, learning-rate schedulers, input resolutions, fine-tuning depths, and loss functions. Within each outer-development cohort, candidates were ranked by mean inner-validation AUROC, the area under the receiver operating characteristic curve, and the winning configuration was trained, checkpointed, and assigned an operating threshold by maximizing Youden’s J statistic, all without any access to the outer-test fold. Each patient appeared in an outer-test fold exactly once, and the 117 pooled patient-level predictions were then used to evaluate the entire model-selection pipeline rather than a single fixed model.

The results were solidly moderate. The pipeline achieved a pooled outer-test AUROC of 0.708, with a 95 percent confidence interval of 0.610 to 0.802, and an AUPRC, the area under the precision-recall curve, of 0.626, comfortably above the roughly 0.42 baseline expected from random ranking given the cohort’s positive-class prevalence. At the validation-derived thresholds, accuracy reached 0.726, sensitivity 0.694, and specificity 0.750, with an F1 score of 0.680. In concrete terms, the model correctly classified 85 of 117 patients, identifying 34 true positives and 51 true negatives while producing 15 false negatives and 17 false positives. The slightly higher specificity than sensitivity reflects a model somewhat better at ruling out CME than at catching every case, a balance with real clinical trade-offs, since missed cases may delay treatment while false alarms generate unnecessary follow-up.

To peer inside the network’s decision-making, the researchers applied Gradient-weighted Class Activation Mapping, or Grad-CAM, an explainability technique that highlights which image regions contributed most strongly to a given prediction. In a true-positive example, the strongest activation formed a localized hotspot over the macular and perifoveal region, overlapping areas of angiographic hyperfluorescence, an anatomically plausible pattern. Yet the maps also revealed the method’s limits: a true-negative patient still showed macular activation, a false-positive case exhibited broad peripapillary and vascular activation rather than a macula-centered hotspot, and a false-negative example concentrated its CME-related evidence on the optic disc and superior-temporal retina while leaving the macula comparatively quiet. Activation location, the authors caution, does not map simply onto classification correctness, and the visualizations should be regarded as qualitative maps rather than validated localizations of disease.

The authors are explicit that these findings do not support clinical deployment. The modest single-center cohort, moderate discrimination, residual classification errors, retrospective design, and absence of external validation all preclude real-world use for now. Internal benchmarks did suggest that transfer learning helped: the ImageNet-pretrained ResNet-18 outperformed a randomly initialized ResNet-18 and a simple CNN on development data, and the nested hyperparameter analysis showed that learning rate and fine-tuning depth, particularly training beyond the classifier layer, were the most influential design factors. The path forward, the researchers argue, lies in external validation on multi-center cohorts, multimodal models that fuse FA with OCT and other imaging modalities, longitudinal modeling to predict CME development and recurrence, and collaborative data repositories with standardized annotation. The greatest clinical value of artificial intelligence in retinal imaging, they conclude, may lie not in replacing clinician judgment but in augmenting it, flagging examinations with suspicious angiographic patterns so specialists can prioritize their attention where it matters most.

Subject of Research: Deep learning classification of cystoid macular edema from fluorescein angiography images in retinal vasculitis

Article Title: Classification of cystoid macular edema status from retinal vasculitis fluorescein angiography images using a convolutional neural network

Article References: Wang, C., Chakrabarty, K., Anesi, S. D., & Chang, P. Y. (2026). Classification of cystoid macular edema status from retinal vasculitis fluorescein angiography images using a convolutional neural network. Frontiers in Ophthalmology, 6, Article 1901237. https://doi.org/10.3389/fopht.2026.1901237

Image Credits: AI Generated

DOI: 10.3389/fopht.2026.1901237

Keywords: artificial intelligence, convolutional neural network, cystoid macular edema, fluorescein angiography, retinal vasculitis, transfer learning, Grad-CAM, nested cross-validation, ophthalmology, medical imaging, ResNet, multiple-instance learning