Bayesian Ultrasound Image Analysis on Graphics Hardware

In this thesis, we investigate using the new generation of programmable Graphics Processing Units (GPUs) which support floating point computations for statistical restoration of ultrasound images. Deconvolution is a widely used method of recovering an image from degradation caused by blurring, and thus increasing the image quality. We present a modified Bayesian 2D deconvolution method which provides better parameter estimation and improves the speed performance over the previous approach. This method lies within the Joint Maximum A Posteriori (JMAP) framework and involves three steps. First is the Point Spread Function (PSF) estimation, which uses the Homomorphic method; second, reflectance field estimation uses a Conjugate Gradient (CG) optimization algorithm; and third , variance field estimation uses a Markov Chain Monte Carlo (MCMC) sampling algorithm. We implement the 2D method entirely on programmable floating-point graphics hardware, and results are achieved at an interactive rate. In order to avoid readback from GPU to CPU, we adopt a multi-pass rendering method to realize the iterative model. This indicates the possibility of using the GPU as a coprocessor in the ultrasound imaging system, to improve image quality in real time. Due to the special architecture of GPUs, not all models are suitable for mapping onto them. We also discuss which structures and schemes GPUs favor and which they do not. Experimental results are presented on synthetic and real ultrasound images acquired by a typical diagnostic ultrasound machine. We believe our research opens the door for many other image processing methods that are otherwise currently impractical, due to time consuming and complicated computations. This is especially important for medical image processing applications.