Deep Learning-Based Quantification of Vitreous Hyperreflective Foci as a Biomarker for Intraocular Inflammation.

OBJECTIVE

To develop and validate an artificial intelligence (AI)-driven pipeline to quantify vitreous hyperreflective foci (vHRF) from OCT images and assess their association with intraocular inflammation (IOI).

DESIGN

A retrospective analysis of a multicenter double-masked placebo-controlled clinical trial cohort.

SUBJECTS: A clinical analysis cohort of 369 patients from the GALLEGO clinical trial (Galegenimab vs. placebo in patients with geographic atrophy, clinical trial ID: NCT03972709).

METHODS

We trained a deep learning segmentation model, U-Net Transformer (UNETR) with a Vision Transformer backbone on 491 OCT B-scans with expert vHRF annotations from the GALLEGO, BURGUNDY (neovascular age-related macular degeneration; NCT04567303), and DOVETAIL (uveitic macular edema; NCT06771271) clinical trials. We applied the optimized model to a clinical analysis cohort from the GALLEGO clinical trial, comprising 1049 OCT volumes (34 images of IOI-positive cases), and evaluated the association between the resulting quantitative vHRF metrics and clinically diagnosed IOI.

MAIN OUTCOME MEASURES: Segmentation performance and the association of quantitative vHRF metrics with clinically diagnosed concurrent IOI, including receiver operating characteristic, precision-recall, predictive value, and eye-clustered generalized estimating equation analyses.

RESULTS

The UNETR model demonstrated strong segmentation performance. In the clinical cohort, IOI-positive eyes showed significantly elevated vHRF metrics (P < 0.001), and vHRF volume density was the strongest biomarker for concurrent IOI detection with an area under the receiver operating characteristic curve of 0.84. In eye-clustered logistic generalized estimating equation models, 5 filtered biomarkers were significantly associated with inflammation, with the strongest association for filtered vHRF density [vHRF/μm3] (odds ratio per standard deviation 1.62, 95% confidence interval 1.26-2.07, P < 0.001). At the optimal threshold for vHRF density, sensitivity was 67.7%, specificity 89.0%, and positive predictive value was 16.9% despite an IOI prevalence of 3.2%.

CONCLUSIONS

Our AI-driven pipeline accurately quantified vHRF from OCT images, and the resulting metrics, particularly vHRF volume density, were significantly associated with concurrent IOI. These findings support automated vHRF quantification as a promising imaging biomarker for inflammation assessment, although further validation in larger and more diverse datasets is needed.

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