Research In Denoising And Reconstruction Acceleration Technology Of Cone Beam CT

Thanks to the development of science and technology, radiation therapy issteering from the traditional radiation therapy to precise radiotherapy. The core ofprecise radiotherapy is accurate positioning, optimized planning and precisetreatment, which aims to optimize the dose distribution to get a higher therapeuticgain ratio through the enlarged target-irradiated volume and the reduced normal tissuedose. Image guided radiation therapy (IGRT) is a new precision radiotherapytechnique following appropriate conformal radiotherapy and intensity modulatedradiation therapy, and cone-beam CT is becoming an important tool in the IGRTbecause of its flexible scanning.The principle of cone-beam CT system is X-ray volume imaging. In the courseof treatment, a series of images are collected by rotating and scanning. After datareconstruction, the images are matched and fused with that from the Planning CT(PCT), then doctors can setup errors and correct the dose distribution accordingly. Thequality of the image and the speed of reconstruction directly affects the accuracy andefficiency of the subsequent matching and fusion, which is related to the treatment ofpatients. The thesis mainly studies the image noise collected from the cone-beam CTand tedious three-dimensional reconstruction process.First, the cone-beam CT image resolution is high, but compared with traditionalCT images, the resolution of soft tissue is still poor, which affects the doctor’sdiagnosis. This section briefly describes the CT imaging principle and structure ofcone-beam CT system, as well as the way of data collection and storage format. Thethesis analyzes the sources of image noise which can be divided into two parts:impulse noise and Gaussian noise. We deal with the former by using traditionalmedian filtering method, and we will mainly discuss the Gaussian noise reduction inthe following part.This paper compares time domain and frequency domain de-noising method, andfound that in the time domain non-local means algorithm for Gaussian noisecancellation works well, but the similarity measurement by its pixel Euclidean distance cannot accurately describe this problem caused by the center pixel of animage block, so the edge blurs. Frequency domain wavelet transform has twofrequency windows, so the ability to extract the image local feature is very strong,which is suitable to deal with the details of the image edge retention problem. But itdoes not have invariant features, which becomes more sensitive to noise. Waveletmoment in polar coordinates is constructed by wavelet transform, which has a scaledisplacement and rotation invariant characteristics. It can not only improve localfeature analysis of the image, but is also less sensitive to image noise. The thesis willcombine the wavelet moment with non-local algorithm, making the wavelet momentas the similarity measure of non-local means algorithm on CBCT image denoising.The algorithm adopts the C language programming, and uses Shepp-Logan headphantom and clinical CBCT images as the test objectives.Second, de-noised CBCT images need three-dimensional reconstruction beforeplan image registration, but the three-dimensional reconstruction is a huge project.The most widely used way in the reconstruction is FDK algorithm, which are roughlydivided into two processes: filtering and inverse projection and the latter consumesmore than80%of the whole time. Inverse projection needs a large amount ofcomputation, thus will increase the patient’s X-ray radiation dose if reconstructiontime is too long. The main objective of this section is to speed up the design for theinverse projection, which could be resolved by software acceleration or hardwareacceleration. Software acceleration includes finding a geometric parameter tablemethod, single instruction multiple data optimization method and symmetry method;hardware acceleration includes a cluster of computers, digital signal processor, imageprocessor, CBE (Cell Broadband Engine) and FPGA (field-programmable gate array).FPGA internal integration has a very rich storage, logical resources, the algorithmmodule, embedded hard core module, and the huge memory bandwidth. The parallelprocessing capability is very strong, which is very suitable for the reconstruction ofthe CBCT images.The thesis selects Xilinx Virtex-5series XC5VLX110chip as the hardware target,optimizes the design of the lookup table, multiplier, divider and buffer interface. We use bilinear difference method, and have developed a pipelined parallel computationsystem. Finally we test the standard head model Shepp-Logan data reconstructionwith Verilog HDL language, and compared with the PC software reconstruction.Result shows that the image deviation of the two reconstruction outcomes is less than1%, and in the volume of the same scale, the processing speed of the FPGA can be upto256times faster than PC software process.Finally, we summary the work above, and point out the weakness of the research.We also look forward to the bright future of CBCT.