| In recent years, with the rising incidence of cancer as well as science and technology, radiation therapy gradually evolved into the precise radiotherapy from the traditional common radiotherapy. The core of precise radiotherapy treatment is precise, accurate positioning, optimization plan, as much as possible to improve the exposure of tumor target area and to reduce the received radiation of normal tissues to improve the therapeutic gain ratio. Currently, the emerging image-guided radiation therapy (IGRT) is another new radiation therapy technique after three-dimensional conformal radiotherapy radiation therapy (3D-CRT) and intensity modulated radiation therapy (IMRT), but, the Cone Beam Computed Tomography is widely used in IGRT because of its faster scan speed, the lower radiation dose, the higher used ratio of the radiation and the solution of the problem of fractionated radiotherapy in positioning. Compared with the traditional X-ray computed tomography (CT), it has the advantage of lower radiation dose on patient, less time consuming based on CBCT image to plan. However, due to the low radiation dose of CBCT images, it exist amount of scattered artifacts and noise that will reduce the precise HU value of the reconstruction images and make a huge influence on the following image processing.In the use of CBCT imaging system, since the X-rays will generate a series of interactions passing the detected object, the X-ray photons reaching the detector contain the original ray and anther scattered rays and noise, which the existing scatter and noise can seriously degrade the image quality. It will not only affect doctors accurate diagnosis, but also have a serious impact on the subsequent image registration, image segmentation and other operations. Although increasing the X-ray radiation dose can reduce the generation of scattered radiation and received a relatively high quality of CBCT images, but it will also affect the patient’s health and leads to a series of complications. Therefore, it becomes a popular research topic currently about how to improve the quality of the CBCT images and insuring the patient’s health at the meantime.Recently, the research on CBCT image artifacts scattering can be divided into two categories:Hardware-based scatter correction method and software-based scatter correction method. The hardware-based scatter correction method denotes that adding some correction tools on each instrument components in X-ray imaging system to reduce the reaching scattering of the X-ray detector to achieve the purpose of scatter correction. The commonly used methods based on hardware correction are the air gap method, the collimator method and the filter grid method. The software-based scatter correction method denotes that the obtained X-ray projection images are analyzed the essence and estimated the property of the contrast to conduct the scatter correction after that the drawing a scattering profile is done based on the digital image processing methods. The commonly used methods based on software correction are the de-convolution method and the Monte Carlo simulation method.Through theoretical analysis we found that the original noise signal is enhanced after the scatter correction. Therefore, it has the important value in research that modify the de-noise of current CBCT images method to reduce the noise impact on image precise. It is essential to suppress the noise after scatter correction because of the uneven distribution of the s scatter ray that the scatter ray around the bone may has larger value than the original ray. At the meantime, the method of the noise suppression for the CBCT images can also be divided into hardware-based noise suppression method and software-based noise suppression method. The hardware-based noise suppression method denotes that a series of correction tools are used to refine the CBCT imaging system to achieve the purpose of noise suppression based on the flaw of the CBCT imaging system. The software-based noise suppression method is considered on two categories:the space of de-noising and the type of the noise. Based on the side of space of de-noising, the filter of the CBCT images can be divided into spatial domain filtering and transform domain filtering. As for the spatial domain filtering, it also includes Gaussian filtering, median filtering, Wiener filter and non-local means filter. However, the Gaussian filtering, median filtering and Wiener filter can’t fulfill the demand of the CBCT images de-noising due to the following flaw:the tradition median filtering can be well enough to preserve the edge information of the images but can be degrade the image resolution and be seriously impacted the doctor’s accurate diagnosis; compared with the median filtering, the Gaussian filtering obtains the corresponding smooth image with the smooth edge which will make it difficult for doctors to sketch the gross tumor volume; the Wiener filter obtains great de-noising effect for the part of high frequency but it doesn’t fit the low contrast tissue, and it control the minimum mean square deviation between the de-noising images and the raw images but fail to preserve the edge information. Compared with the mentioned three de-noising methods, the non-local means filter is preferable to add into the research of the CBCT image noise suppression. The traditional Fourier transform filtering algorithm in domain filtering can’t achieve absolutely perfect de-noising effect. Thus, compared to the traditional Fourier transform filtering algorithm, the Wiener filter which has the properties of good time frequency, changeable feature and good resolution is the mainstream algorithm in CBCT image noise suppression. It can be divided into two big categories based on the noise in the CBCT images:one category is the electronic noise (Gaussian noise) generated by CT tube current and voltage that is irrelevant to the images;the Gaussian noise’s probability density function subject to the Gaussian distribution (normal distribution); the Gaussian white noise including the hot noise and the shot noise has a uniform power spectral density; another category is the quantum noise(Poisson noise) generated by the X-ray photons which is relevant to the images. The core research in our paper is focus on de-noising the Poisson noise in the CBCT images.The Monte Carlo simulation approach used in our paper is a class of statistical probability theory based on numerical method. The Monte Carlo method can precisely simulate the movement of particle. However, the realistically simulating actual physical interaction of the Monte Carlo method results in huge computational complexity and slow simulated speed. That’s why The Monte Carlo method can’t be widely used. With the development of compute capacity of the computer, the computer industry develop from simply using the Central Processing Unit(CPU) to using he Central Processing Unit(CPU) and the Graphics Processing Unit (GPU) cooperative processing. Recently, a novel parallel computing platform named Compute Unified Device Architecture (CUDA) on GPU is widely used.It is owned by the NVIDIA Company that has the powerful parallel computing capacity and supporting the supercomputing. The operating principle of the CUDA is that:firstly, the CPU sends the instructions to the GPU, the GPU accepts the instructions to process the task parallel, when computing is done, the processing result will be send back to the CPU which dramatically reduce the work time. The problem of huge computational complexity in Monte Carlo simulation approach can be solved by using the property of the parallel computing in CUDA.In this paper, firstly, we introduce the scattered artifacts in CBCT images, the impact of noise and the research situation between domestic and foreign. Secondly, we research the scatter correction and noise suppression separately. Thirdly, we deeply introduce the research based on the Monte Carlo scatter correction noise suppression method. We refine and achieve the follow methods:1.Currently based on the Monte Carlo scatter simulation method:it is commonly used imaging data contains the scattering artifacts as motif and sketched the parameters like material substance, type and density in the Monte Carlo scatter simulation through the CT value(Hounsfield Unit, HU) of motif. Due to the imprecise HU value of scattering artifact images, it’s imprecise to conduct one time simulation. In order to make up this problem, we use the iterating Monte Carlo simulation method to make up, that is the current phantom data is the result of last time scatter image reconstruction. It can be obtained the precise image through two or three iterations but the simulation time is raised. In this paper, in order to avoiding the repeatedly Monte Carlo simulation iterations, we the gMCDRR tools to propose using the original plan of CT image as the motif image in the image guided radiation therapy. The original plan of CT image as the input data of the Monte Carlo simulation can be obtained more precise HU value and be simulated precise scatter signal. Firstly, we use the raw CBCT projection to conduct the fast FDK reconstruction. And then we use the original plan of CT image and the CBCT image obtained from fast FDK reconstruction in the form of fast convolution rigid registration.Finally, we use the deviate registration original plan of CT image as the input data to conduct the fast Monte Carlo scatter simulation. In the simulation of scatter correction, we use the following accelerating methods:down sampling the motif image using the Monte Carlo simulation,degrade the number of photons in simulation and simulate the scatter signal sparsely in the angel direction, prior smooth constrain and interpolate the scatter signal. Finally, we minus the predicted scatter signal from the original projection, and conduct the FDK reconstruction to obtain the reconstruction image of the scatter artifact correction.2.Currently based on the Poisson noise simulation method:firstly, we demonstrate the noise enhancement of the CBCT image after scatter correction through formula derivation. And then, we modify the root of the penalized least squares method formula, and obtain the objective function of projected smooth de-noising based on the property of the Poisson distribution. Calculating derivative of the objective function and using the continuous super relaxation procedure to approximate the optimal value. In this paper, experiments was carried out by using the true data and the simulating data separatly. One group data is the CBCT projection of the Pelvis phantom, the resolution is 512 x 385 x 656,the voxel is 0.0977×0.0977×0.25cm3. In the simulation experiment, the signal to noise ratio(SNR) is improve from the 19.8 dB before de-noising to the optimal value 26.8dB,and the root of mean square error(RMSE) degrade from the 50.7 before de-noising to the optimal value 22.5. In the true data experiment, the RMSE degrade from the 64.1 before de-noising to the optimal value 58.2. At last, the finish time of the whole process including the scatter correction and de-noising within 40 seconds, which make it probable to complete the clinic application.The quality of medical image will be directly impact on doctor’s diagnosis and the following image processing, thus, obtaining better scatter correction result and noise suppression become more and more essential. In our paper, we propose our method to perform the scatter correction and the noise suppression method in the CBCT system and obtain an appreciable result. In future, we will research deeper to modify the shortage of our method. |
PDF Full Text Request