Quantitative Analysis of Retinal Fluid by a Deep Learning Model in Uveitic Macular Edema.

Abstract

To assess whether artificial intelligence (AI)-derived fluid volume provides prognostic value for visual outcomes in uveitic macular edema (UME) and to compare model performance to central macular thickness (CMT) measurements alone. Secondary subanalysis of patients with UME in the First-line Antimetabolites as Steroid-sparing Treatment (FAST) clinical trial using a deep learning segmentation model trained on patients with age-related macular degeneration (AMD) and patients with retinal vein occlusion. Patients with UME secondary to noninfectious uveitis from the FAST Uveitis Trial. A 2-dimensional U-Net model, trained on patients with AMD and retinal vein occlusion using the RETOUCH data set, was applied to segment intraretinal fluid (IRF) and subretinal fluid (SRF) in Heidelberg Spectralis OCT scans from the FAST Uveitis Trial. Model performance was validated against binary fluid gradings. Linear mixed-effects models evaluated fluid resolution during treatment. Likelihood ratio tests and leave-one-subject-out cross-validation assessed whether baseline IRF and SRF volumes provided improved model fit for visual acuity (VA) change over 6 months of treatment compared to CMT alone. Segmentation accuracy, correlation of fluid volume with VA, and association of VA change with baseline fluid volumes and CMT versus CMT only. The model achieved Dice scores of 0.61 for IRF and 0.74 for SRF on the RETOUCH data set. In the FAST data set, fluid segmentation correlated with binary gradings (biserial correlation: IRF 0.39, SRF 0.63; Mann-Whitney <i>U</i> test: IRF and SRF <i>P</i> &lt; 0.01). Longitudinal fluid modeling showed no significant treatment differences in IRF or SRF resolution. However, baseline IRF volume had a significant interaction with treatment assignment (<i>P</i> = 0.02) when modeling VA change, suggesting that baseline IRF was more strongly associated with worse outcomes in the methotrexate group than the mycophenolate mofetil group. The likelihood ratio test confirmed that adding IRF and SRF improved fit over a model using CMT alone (<i>P</i> &lt; 0.01) while leave-one-subject-out cross-validation showed a small, nonsignificant difference in prediction error. Artificial intelligence-based segmentation of IRF and SRF enabled quantitative fluid measurements in UME and may provide additional prognostic signal for VA change in UME compared to CMT alone, although larger cohorts will be needed to determine the magnitude of generalizable performance gains, if any. These findings support integrating AI-driven fluid analysis in clinical workflows and suggest that future clinical trials should consider stratifying by baseline fluid characteristics using AI fluid analysis. Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.