Abstract
Purpose :
Current state-of-the-art Optical coherence technology (OCT) devices, including enhanced depth imaging OCT and swept-source OCT, provide improved depth resolution and facilitate visualization of deeper structures such as choroid. In contrast, outer retinal layers and the choroid are poorly visible in earlier spectral domain OCT (SD-OCT) devices. However, these devices are still in clinical use and have collected significant retrospective data. Against this backdrop, it is imperative to improve the resolution of poor-quality SD-OCT images to facilitate accurate screening and to develop algorithms to quantify. To this end, we proposed a deep-learning (DL) super-resolution approach to enhance the resolution of SD-OCT images to higher resolution and demonstrate that improvise choroid detection on the super-resolved data in comparison to conventional SD-OCT images.
Methods :
A retrospective dataset of 896 OCT images was taken from Cirrus 5000 OCT device (Carl Zeiss Meditec). We adopted a DL-based super-resolution approach to enhance the images. Specifically, we employed a real enhanced super-resolution generative adversarial network (realESRGAN) which is shown to significantly reduce the noise and increase the resolution. We down-sampled the image by four times to perform super-resolution to the original scale. Subsequently, to objectively assess the quality of super-resolution, we attempted to segment the choroid layer from the original and enhanced images using our previously validated exponentiation enhancement method. The accuracy of segmentation was evaluated by an expert in masked fashion where segmentation is graded as accurate or inaccurate.
Results :
RealESRGAN super-resolution significantly improved the visibility of various structures including the choroid layer (Figure 1). Subjective grading results showed choroid segmentation is accurate in 20 % of the original images and in 80% of the super-resolution images.
Conclusions :
The proposed algorithm generated high-quality super-resolved OCT images from conventional SD-OCT images, with significant improvement in accuracy for choroidal segmentation. Future direction includes further improvement in quality for the quantification of vascularity and other choroidal biomarkers.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.