Data driven discovery of anti-VEGF treatment response groups based on fully automatic vitreomacular interface segmentation

Abstract

Purpose : To study the effect of vitreomacular adhesion (VMA) on functional response to anti-vascular endothelial growth factor (VEGF) therapy using fully automated image processing of optical coherence tomography (OCT) data in large scale clinical trials populations.

Methods : Baseline SD-OCT scans (Heidelberg Spectralis, 6x6mm, 49 B-scans) of 238 resp. 187 patients included in prospective trials evaluating anti-VEGF therapy for macular edema secondary to branch resp. central retinal vein occlusion (BRVO resp. CRVO) were processed using a fully automated 3D segmentation algorithm yielding an inner limiting membrane (ILM) and posterior vitreous boundary (VIT) surface.
The two surfaces were used to generate distance maps which were grouped into multiple clusters using unsupervised machine learning algorithms . The progression of best-corrected visual acuity (BCVA) scores and the BCVA change relative to baseline were analyzed for each patient up to month 6 resp. month 12.
Furthermore the patient population was split by different percentiles of BCVA change at month 6 resp. month 12, and the mean distance maps at baseline were analyzed.

Results : Patients closest to cluster centers resembling VMA demonstrated significantly higher gains in BCVA at month 6 resp. month 12 (+14.0 vs. +9.1 letters, p = 0.007 for BRVO, +16.9 vs. +11.5 letters, p = 0.029 for CRVO). This finding was stable in respect to the number of clusters into which the patient population was split, however statistical significance was lost when splitting into more clusters.
The analysis based on BCVA change percentiles revealed that the that the mean distance maps of patients with larger BCVA gains clearly demonstrated features of VMA, as opposed to the VMI maps of patients with smaller BCVA gains.

Conclusions : The configuration of the vitreomacular interface can be efficiently analyzed in detail using ILM/VIT distance maps obtained by our fully automated segmentation method. Unsupervised machine-learning based clustering of the vitreous configuration maps revealed robust morphologic subgroups of patients with clinically distinct functional response patterns to anti-VEGF therapy in branch resp. central vein occlusion.

This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.

 

top: B-scan with ILM and VIT segmentation, middle: ILM-VIT distance maps, bottom: closest cluster centers