Real Time Reconstruction processor for Cone-Beam X-ray CT

2005 Fiscal Year Final Research Report Summary

• Project/Area Number: 16560367
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Single-year Grants
• Section: 一般
• Research Field: Measurement engineering
• Research Institution: OKAYAMA UNIVERSITY
• Principal Investigator: MORIKAWA Yoshitaka OKAYAMA UNIVERSITY, Graduate School of Natural Science and Technology, Professor, 大学院・自然科学研究科, 教授 (30033252)
• Co-Investigator(Kenkyū-buntansha): YAMANE Nobumoto OKAYAMA UNIVERSITY, Graduate School of Natural Science and Technology, Assistant Professor, 大学院・自然科学研究科, 助教授 (80174762) ; MURAKAMI Junichi TAKUMA NATIONAL COLLEGE of TECHNOLOGY, Professor, 教授 (70200269)
• Project Period (FY): 2004 – 2005
• Keywords: X-ray CT / reconstruction algorithm / real time / cone beam / DSP

Research Abstract

Convolution-backprojection (CB) method has been known as a general 2-dimensional (2-D) reconstruction algorithm for X-ray computer tomography (CT). In CB method, each projection signal is emphasized in high frequency by convolution and then backprojected and added in image plane for reconstruction. Convolution in the first part and backprojection in the second part are the reason why CB method is slow. In the proposed method, high-frequency emphasis in the first part is realized by the Hilbert transform using allpass filters and difference operation, and backprojection in the second part is realized by tree-structured filter bank. By adoption of these two new realizing methods the computation amount can be sharply reduced, and high-speed reconstruction processing is possible. According to the simulation, in the 2-D case, proposed method attained about 12 times as much improvement in the speed as the conventional method, and in the 3-D case about 280 times improvement was attained.
In th is research, the proposed method was mounted in TMS 320C6416 (Texas Instruments) which is fixed-point arithmetic DSP, and ADSP-TS101/TS201S (Analog Devices) which is a floating point arithmetic DSP. In mounting to fixed-point DSP, the accuracy comparable to floating point arithmetic was realized by distributing bit shift operations to the filter processing part and image output part, and processing speed (33 m seconds) in which real time observation is almost possible was attained. On the other hand, in mounting to a floating point DSP, although there was generally a problem that processing speed was slower than fixed-point arithmetic, improvements in processing speed were tried by using optimization or parallelization techniques, and the processing speed in which real time observation is possible was attained.
3-D real-time reconstruction is considered possible by performing direct memory access (DMA) and distributed processing. It is estimated that 256x256x256-pixel reconstruction takes 60 seconds about 200 times as many as that in 2-D reconstruction, and, in 16 ideal parallel processing, reconstruction time can be shortened at about 3.8 seconds. It was also estimated that about 2000 parallel processing is acquired for a real-time 3D reconstruction of such a size.
One of the feature of the proposed method is accessing pattern of various directions in image memory array, and the discussion on cash memory is an important subject. The same trial simulation as the above was performed by Pentium4 (2.6GHz). It required about 20 seconds for reconstruction, which was a 3 time high speed compared with DSP. In recent years, not only improvement in the speed of CPU but also in the speed of memory accessing was progressing. The investigators consider that the 3-dimensional picture of such a size can be reconstructed in a subsecond by one processor likethe 64 bit Xeon.