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

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