Abstract
Purpose :
OCTA provides clear visualization of the retinal capillary networks. As it requires separate hardware, it has not been integrated into widespread clinical workflow. Hence, generating OCTA images from OCT through generative machine learning can be a promising alternative. The purpose of this study was to demonstrate a quantitative comparison of vascular features on Optical Coherence Tomography Angiography (OCTA) images translated from OCT vs commercially available OCTA.
Methods :
We used a public dataset containing 200 pairs of 3 mm OCT-OCTA volumes to develop an algorithm to translate OCTA images from OCT. The algorithm was built upon a 3D conditional generative adversarial network (GAN) which involved a 3D generator-discriminator working together to produce high-quality translations of OCT to OCTA. There were 2 modules to tackle accumulated information loss during the generation of 3D OCTA: a retinal vascular segmentation model to predict vessel pixels in OCTA projection maps enabling the network to focus more on vascular regions and another module to generate 2D projection maps from paired OCT projection maps retaining contextual information. These 2 modules were integrated into the 3D image translation baseline and trained end-to-end. The generated translated OCTA maps were then compared on four quantitative features to the ground truth (GT) projection maps for qualitative comparison: blood vessel density (BVD), caliber (BVC), tortuosity (BVT), and vessel perimeter index (VPI).
Results :
We observed reasonable structural similarity, SSIM, mean 0.48 (0.29~0.60). This was expected since only vascular regions are prioritized during translation and the SSIM considers the whole projection map which includes noise. In t-tests, BVD and BVC showed very similar values in the translated OCTA and GT OCTA, respectively: BVD (0.249±0.107 vs 0.258±0.11) and BVC (22.80±.81 vs 22.75±0.41). VPI and BVT, which rely on the skeletonization of vessels in OCTA, did not show similar results: VPI (26.91±5.474 vs. 31.432±2.35) and BVT (1.08565±0.00650 vs. 1.08905±0.00617) in the translated OCTA and GT, respectively.
Conclusions :
This study was successful in translating high-resolution OCTA images from OCT and generating vascular features for blood vessel density and caliber which were similar to GT OCTA. Future studies should test and validate this promising approach in clinical OCT data for multi-modal analysis across different retinal diseases.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.