Enhanced choroidal thickness, total choroidal area and choroidal vascularity index from deep‐learning generated OCT images
Valentina Bellemo,Ankit Kumar,Damon Wong,Jacqueline Chua,Xinyu Liu,Xinxing Xu,Yong Liu,Leopold Schmetterer
DOI: https://doi.org/10.1111/aos.15979
2024-01-01
Acta Ophthalmologica
Abstract:Aims/Purpose: Our goal is to enhance the visualization of the choroid in clinical standard spectral‐domain OCT (SDOCT) scans using a deep‐learning (DL) approach that uses matching swept‐source OCT (SSOCT). We aim to improve the measurement of the sub‐foveal choroidal thickness (SFCT), total choroidal area (TCA), and choroidal vascularity index (CVI) in SDOCT images. Methods: A total of 105 350 SDOCT and SSOCT B‐scan images (Zeiss Cirrus 5000 and Zeiss PlexElite) were used to develop the DL model. The data set included B‐scans from healthy (238 eye pairs from 155 patients) and glaucoma (192 eye pairs from 122 patients) subjects. The DL model learned deep anatomical features from SSOCT scans and generated enhanced SDOCT images. The peak signal‐to‐noise ratio (PSNR) and structural similarity index metric (SSIM) were used to assess the quality of generated images. In addition, we compared the anatomical parameters derived from the generated and ground‐truth SSOCT images, including SVCT, TCA and CVI values. Results: We tested our method using 20 580 B‐scan image pairs from 84 eyes. The DL‐generated images effectively preserved the anatomical structure of the retina layers and improved the visualization of the choroid. The enhanced SDOCT images had higher image quality than the original SDOCT images, as demonstrated by the higher PSNR (31 dB vs. 18 dB) and SSIM values (3.4 times higher SSIM than the original SDOCT images). SVCT, TCA, and CVI measurements from the DL‐generated SDOCT images showed strong correlations with corresponding measurements using SSOCT with high Pearson's coefficients (0.93, 0.99, and 0.93), high intra‐class correlation indices (0.96, 0.99 and 0.96), and high coefficient of determination r 2 values (0.85, 0.98 and 0.86). Conclusions: Our DL model enhances the visibility of the choroid in images, surpassing the capabilities of standard clinical SDOCT images. As a result, accurate measurement of SVCT, TCA and CVI, which were previously limited due to imaging depth constraints of SDOCT, becomes possible. This technology opens up new possibilities for utilizing affordable SDOCT devices to study the choroid in healthy and pathological conditions.
ophthalmology