Robust registration using multiple OCT/OCTA en face images in advanced disease cases

Abstract

Purpose : Robust registration of longitudinal data is required to measure the changes of OCT/OCTA slabs and retinal thicknesses over time. The pathological changes in the retina from one visit to another makes the registration of two OCT/OCTA slabs a challenging task. This abstract proposes a robust registration method that utilizes multiple slabs.

Methods : The OCT registration consists of aligning the corresponding keypoints in a pair of images. The well distributed keypoints matches were ensured by extracting the keypoints from OCT/OCTA slabs at different depth. OCT images were generated from 1) superficial and deep retinal layer, 2) RPE and 3) choroidal slabs. OCTA images were generated from 1) superficial (SRL), 2) deep capillary plexus and 3) RPE and choriocapillaris slabs. An exhaustive search method was used to fit best rigid transformation to a subset of matched keypoints. 15, 25, and 5 pairs of OCT volumes over 6x6 mm using CIRRUS^TM HD-OCT 5000 (ZEISS, Dublin, CA), PLEX^® Elite 9000 SS-OCT (ZEISS, Dublin, CA), and CIRRUS^TM HD-OCT 5000 with AngioPlex^® OCT Angiography (ZEISS, Dublin, CA) were used with eye diseases such as diabetic retinopathy and age-related macular degeneration. The performance of 5 registration methods were measured using keypoints extracted from 1) RPE, 2) all 3 OCT, 3) SRL, 4) all 3 OCTA and 5) all OCT and OCTA images. The nonparametric statistics of the RMSE of the matched keypoints along with the number and keypoints image area coverage for each pair of OCT volumes were reported for all 5 methods.

Results : Fig 1 shows examples using multiple slabs for registration. Fig 2 shows the box plots for RMSE of the matched keypoints, the number and keypoints image area coverage for 5 registration methods. The number and the keypoints area coverage increase significantly when multiple OCT/OCTA slabs were used for keypoints extraction, which is important for robust image registration with small RMSE. The RMSE using multiple slabs shows tighter distribution with a median value around 10 μm indicating robustness and higher success rate. This may meet the accuracy requirement of OCT registration with A-scan spacing of 12 to 20 μm.

Conclusions : We introduced a novel registration approach based on matched keypoints extraction from multiple slabs. We showed that using multiple slabs outperforms using a single slab for registration which is essential for measuring changes in longitudinal data.

This is a 2021 ARVO Annual Meeting abstract.