Deconvolution of ultrasound biomicroscopy images using generative adversarial networks to visualize and evaluate localization of ocular structures

High frequency ultrasound biomicroscopy (UBM) images are used in clinical ophthalmology due to its ability to penetrate opaque tissues and create high resolution images of deeper intraocular structures. Because these inexpensive, high frequency (50 MHz) systems use single ultrasound elements, there is a limitation in visualizing small structures and anatomical landmarks, especially outside focal area, due to the lack of dynamic focusing. The wide and axially variant point spread function degrade image quality and obscure smaller structures. We created a fast, generative adversarial network (GAN) method to apply axially varying deconvolution for our 3D ultrasound biomicroscopy (3D-UBM) imaging system. Original images are enhanced using a computationally expensive axially varying deconvolution, giving paired original and enhanced images for GAN training. Supervised generative adversarial networks (pix2pix) were trained to generate enhanced images from originals. We obtained good performance metrics (SSIM = 0.85 and PSNR = 31.32 dB) in test images without any noticeable artifacts. GAN deconvolution runs at about 31 msec per frame on a standard graphics card, indicating that near real time enhancement is possible. With GAN enhancement, important ocular structures are made more visible.