Reconstrução de imagens de ultrassom através de problemas inversos usando processamento paralelo em GPUs

Abstract

This project aims to improve the ultrasound image reconstruction methods based on inverse problems for parallel implementation on GPU. These methods are recent innovations for ultrasound and have demonstrated to achieve better image quality upon conventional reconstructions such as, beamforming and delay-and-sum. However, the computing time of these algorithms are still high, not allowing reconstructions in real-time, which is an already established demand in ultrasound. For this, we have modified the algorithms for parallel execution on graphics processing units (GPU). These processors have a very high processing power, much larger than a CPU. But the algorithms can take advantage of this benefit only if they are parallel or parallelizable. The modified algorithms are parallel and can be executed on GPU. However, they require a large amount of memory to store the system matrix. This fact limits the maximum image resolution that can be obtained by reconstruction on the GPU. This work uses a new and alternative way to reconstruct images through the exploitation of the symmetry of the point spread function, halving the memory size required by the system matrix. Thus, it is possible to reconstruct an ultrasound image with better quality then the current images, and virtually at the same time.