COMPRESSED SENSING BASED IMAGE PROCESSING WITH APPLICATION TO 3D MRI

Abstract

In the present age of technology, the buzzwords are low-power, energy efficient and compact systems. This directly leads to the data processing and hardware techniques employed in the core of these devices. One of the most power-hungry and space-consuming schemes are that of image or video processing, due to its high quality requirements. In current design methodologies, a point has been reached in which physiological and physical impacts restrict the capacity to simply encode information faster. These limits have led to search for strategies to diminish the amount of obtained data without losing information and increase vitality and time effectiveness. Compressive sensing (CS) is an emerging technology which is based on an efficient signal compression and reconstruction technique. It can be used to efficiently reduce the data acquisition and processing. It utilizes the sparsity of a signal in a different transform domain for sampling and stable reconstruction. It is another opportunity to conventional data processing and is robust in nature. Unlike the conventional methods, CS provides an information capturing paradigm with both sampling and compression. It allows signals to be sampled below the Nyquist rate, and still allowing optimal reconstruction of the signal. The required measurements for reconstruction are far less than those of conventional methods, and the process is non-adaptive, making the sampling process faster and universal. In this thesis, CS method is applied to magnetic resonance imaging (MRI), which is popularly used imaging technique in clinical applications and image 12 compression. Over the years, MRI has enhanced significantly in both imaging quality and working speed. This has further advanced the field of diagnostic medication. In any case, imaging speed, which is fundamental to numerous MRI applications, still remains a major question. Moreover, the fast growing 3D MRI images and real-time MRI scanning has become more complex leading to a further increase in data acquisition times. On the other hand, to improve the speed, it becomes necessary that the processing algorithms are also computationally sped up. There have been various attempts where Graphic processing unit (GPUs), Field programmable gate array (FPGAs), clusters and central processing unit (CPUs) are used for this purpose, which have time durations ranging from minutes to hours. Again, though, there are solutions for reduction in computation time, the desire for the least computational time needs to be ascertained. Considering the requirements discussed above, the work in this thesis is presented in two parts. In the first part, a scheme for 3D MRI reconstruction is proposed by taking the benefit from collaborative sparsity, which at the same time enforces nonlocal 3D sparsity and local 2D sparsity in a hybrid transform domain. An efficient and improved augmented Lagrangian technique is used for the solution of CS optimization problem which enhances the MRI performance when compared with the conventional sampling. Experimental results for all slices are combined to form a 3D view. This work reveals the efficacy of already proposed 2D recovery algorithm by reconstructing image in 3D. As, 3D MRI experiences long scan time, and CS can significantly reduce the encoding time for 2D magnetic resonance images, therefore, for faster 3D MRI 13 reconstruction the algorithm is parallelized. So, in the second part this problem is tackled by using the benefits of multicore architecture and GPU. The sequential Compressed Sensing reconstruction algorithm is made parallel using different parallelizing techniques and compared on the basis of Peak signal to noise ratio (PSNR), correlation coefficients, Relative error (RE), Root mean square error (RMSE), histogram and topography measure. The effect of GPU based implementation is not significant because the application of complex collaborative sparsity on small blocks of images forms the algorithm repetitive in structure. This bottleneck lies because of a lot of time in data transfer from host to GPU and GPU to host. An optimization is introduced for sequential implementation by using combination of lower level and higher level languages. In this work, multicore architecture is proven to be most efficient and accelerated for 3D CS reconstruction because of repetitive structure of algorithm. When compared to sequential CS recovery via Collaborative sparsity this design is energy-efficient, fast and has lower complexity.