Efficient architectures of heterogeneous fpga-gpu for 3-d medical image compression

• Main Author: Muharam, Azlan
• Format: Thesis
• Language: English; English; English
• Published: 2019
• Subjects: QA76 Computer software
• Online Access: http://eprints.uthm.edu.my/55/1/24p%20AZLAN%20MUHARAM.pdf; http://eprints.uthm.edu.my/55/2/AZLAN%20MUHARAM%20COPYRIGHT%20DECLARATION.pdf; http://eprints.uthm.edu.my/55/3/AZLAN%20MUHARAM%20WATERMARK.pdf; http://eprints.uthm.edu.my/55/
• Institution: Universiti Tun Hussein Onn Malaysia

The advent of development in three-dimensional (3-D) imaging modalities have generated a massive amount of volumetric data in 3-D images such as magnetic resonance imaging (MRI), computed tomography (CT), positron emission tomography (PET), and ultrasound (US). Existing survey reveals the presence of a huge gap for further research in exploiting reconfigurable computing for 3-D medical image compression. This research proposes an FPGA based co-processing solution to accelerate the mentioned medical imaging system. The HWT block implemented on the sbRIO-9632 FPGA board is Spartan 3 (XC3S2000) chip prototyping board. Analysis and performance evaluation of the 3-D images were been conducted. Furthermore, a novel architecture of context-based adaptive binary arithmetic coder (CABAC) is the advanced entropy coding tool employed by main and higher profiles of H.264/AVC. This research focuses on GPU implementation of CABAC and comparative study of discrete wavelet transform (DWT) and without DWT for 3-D medical image compression systems. Implementation results on MRI and CT images, showing GPU significantly outperforming single-threaded CPU implementation. Overall, CT and MRI modalities with DWT outperform in term of compression ratio, peak signal to noise ratio (PSNR) and latency compared with images without DWT process. For heterogeneous computing, MRI images with various sizes and format, such as JPEG and DICOM was implemented. Evaluation results are shown for each memory iteration, transfer sizes from GPU to CPU consuming more bandwidth or throughput. For size 786, 486 bytes JPEG format, both directions consumed bandwidth tend to balance. Bandwidth is relative to the transfer size, the larger sizing will take more latency and throughput. Next, OpenCL implementation for concurrent task via dedicated FPGA. Finding from implementation reveals, OpenCL on batch procession mode with AOC techniques offers substantial results where the amount of logic, area, register and memory increased proportionally to the number of batch. It is because of the kernel will copy the kernel block refer to batch number. Therefore memory bank increased periodically related to kernel block. It was found through comparative study that the tree balance and unroll loop architecture provides better achievement, in term of local memory, latency and throughput.