Tomography with GPU

The project's goal was making feasibility check for improving the time-consuming algorithm of tomography with a new method of opto-acoustic.

The project’s goal was making feasibility check for improving the time-consuming algorithm of tomography with a new method of opto-acoustic. The large amount of data which requires processing, together with the heavy mathematical operations required for making picture reproduction in this method, require development of parallel processing methods, which will be executed with an advanced graphical processor, edge processors and dedicated hardware.

At the beginning, we should have learned and understand the reproduction diagram of tomographic opto-acoustic, and recognize the places in it which we can be upgraded to parallel run, while taking into consideration the realistic dimensions of the problem and the added value of parallel run in each step.

Our first goal was trying improve the basic multiplication operation by parallel implementation on the GPU.
This part was implemented in two different ways, first with the MATLAB interface and afterwards with a direct programming on the GPU while using CUDA. The reason for that consists of two elements: 1. Efficiency test for the run-time on the two platforms. 2. Providing accessibility of parallel abilities for investigators in this area. During the project we were demanded to deal with problems and wrong characterization of GPU interface in the MATLAB environment. We created contact with the Mathworks in order to fix these issues.

Another main goal was optimizing the time-consuming algorithm of building the model matrix, which are used to get the actual tomographic picture.
During the project, we studied the system capabilities (a server with advanced process abilities and graphic processor) and the advanced performance it supplies. Thus, we could have reached important consequences about preferability of using the CPU over a single GPU and about the need of parallel GPU operation in order to achieve real-time performance for multiplication.
At the end of the project, we supplied a parallel platform which will help a lot to the research in the lab.

Another important achievement was significant improvement of the run-time of building the model-matrix. This matrix are used by investigators to check the influence of detector geometry on the quality of the picture reproduction. This improvement reduce the run-time significantly and therefore it’s an efficient research tool in the new lab of “Medical Image and Sensing” in the Technion, which develops these models.
Moreover, during the project we suggested a solution to one of the most difficult problems In this area, which is the amount of data required to process which pass the limit of the current hardware abilities. We suggested a method which takes into consideration the sparseness of the detected signal and as a result supplies high resolution 3D process of the data (We called this method “NAKNIK”). The system’s abilities we mapped during the project will be important for future projects and research in the lab of “biomedical imaging and sensing” (LBIS). In particular, for development of 3D tomographic-model in high resolution.