Abstract
Purpose :
Visualization of choroid remains limited with spectral-domain optical coherence tomography (SDOCT). We propose to enhance the visualization of the choroid from SDOCT scans, using a deep-learning (DL) model, trained from matched pairs of SDOCT and swept-source OCT (SSOCT) images.
Methods :
A total of 117110 images from SDOCT (Zeiss Cirrus 5000) and SSOCT (Zeiss PlexElite) were used to develop the DL model, including B-scans from healthy subjects as well as subjects with high myopia (239 eye pairs from 155 patients). We developed a novel DL Generative Adversarial Network which learns deep anatomical features from SSOCT B-scans and translates enhanced choroid properties to SDOCT images. To quantitatively evaluate the quality of the generated synthesized predictions, we analyzed the peak signal-to-noise ratio (PSNR) and the structural similarity index metric (SSIM) of the enhanced images. To further assess the anatomical plausibility, we manually measured the choroid thickness and compared the results with the original values by computing the statistical correlation and measurement errors against reference measurements.
Results :
A total of 10535 image pairs have been used to test the DL model. The DL-generated images preserve the anatomical shape of Cirrus retina layers and show enhanced deep morphological structures as the propagation of choroidal anatomical information from PlexElite scans. We found the synthesized images to have PSNR of 28dB (SDOCT: 19dB), and SSIM with SSOCT data 3.1 times higher than SDOCT. Furthermore, the results illustrate that using our translational DL model significantly improved choroid visibility and thickness measurements. (Correlation coefficient of synthesized with respect to the reference: 0.91 vs SDOCT 0.10, mean standard error: 0.001 vs 0.017, mean absolute error: 0.024 vs 0.101, mean absolute percentage error: 10% vs 39%, root mean squared error: 0.040 vs 0.133).
Conclusions :
We developed a DL model to enhance the visibility of deep posterior structures including choroid from standard clinical SDOCT images. As a result, the choroid layer visualization was improved, and the choroid thickness of the generated B-scans extensively matches the measurements of the reference SSOCT. The proposed DL methodology is robust and crucial to study processes and mechanisms of the choroid in normal eyes and in pathological conditions which were not assessable before in a clinical setting.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.