Deconvolution and improved visualization of ocular structures in UBM using deep learning

Ultrasound biomicroscopy is a unique imaging modality used for high-resolution visualization of opaque tissues in the anterior segment of the eye. Due to the lack of dynamic focusing and wide spatially varying point spread function in conventional single element probe, small ocular structures are blurred. Computationally expensive deconvolution approaches can enhance the images but are not suitable for real-time use. We created a supervised generative adversarial network-based deconvolution approach that maps an input image to an enhanced image. We achieved a structural similarity index of 0.85 in our test images without significant degradation. The proposed GAN based deconvolution approach can enhance a single image frame at 31 msec making deconvolution almost real-time.