Algorithm Research On CBCT Images Denosing And Registration

With the rapid progress in physics and biological theory, radiation therapy technology has undergone great development and become one of the three major tumor treatment methods. The CBCT (Cone-Beam CT) based IGRT (Image Guided Radio Therapy) Imaging System can reach the goal of improving the accuracy of the radiotherapy. CBCT images have the advantages of good real-time, high sensitivity, and convenience. Their superiority is more prominent when they are registered and fused with Planning CT images. However, as the CBCT images suffers from the disadvantanges of weak edges contaminated by noises and great resolution degradation in low-density regions, the difficulty of diagnosis is increased. Meanwhile, the clinical registration between CBCT images and Planning CT images is carried out manually, which inevitably reduces radiation accuracy to some extent.Based on the analysis and comparison of existing medical image denoising and multi-modality registration methods, we proposed two improvement methods. One is a new CBCT images denoising method based on coefficient classification. And the other is a novel CBCT and Planning CT elastic registration method based on attribute vector.The main contributions and innovations of this thesis are as follows:1. CBCT images denoising based on coefficient classification(1) According to principle of CBCT imaging and characteristics of noise distribution, we proposes a fast CBCT denoising method based on coefficient classification which relies on the Wavelet Modulus Maxima denoising theory. Experimental results for a test image and clinical CBCT images show that this algorithm can suppress noise in CBCT images effectively with important structure details preserved. In addition, the proposed method doesn't need any improvements of clinical equipments, which increases the feasibility of clinical application.(2) After we analyse the noise distribution characteristics of dyadic wavelet, the equation for noise standard deviation relationship across scale 2-D dyadic wavelet transform is deduced and then the noise variance estimation equation for the CBCT image is proposed. Experimental results show that the proposed formula can efficiently improves the edge protection effect of denoising algorithm, and is superior to the median absolute deviation of wavelet coefficients method proposed by Donoho.(3) According to specific directional characteristics of detail bands of dyadic wavelet, different Wiener filters are adopted at all levels based on direction window for different coefficient types. This method can preserve the edges information of CBCT images, reduce the denoising artifacts, achieve higher PSNR(Peak Signal to Noise Ratio), and provide a better vision effect.2. CBCT and Planning CT elastic registration based on attribute vector(1) The existence and corruption of noises in CBCT images is one of the main interfering factors for the accuracy of registration. We analyse the shortage of modified HAMMER, and propose a new structure of attribute vector, in which we use canny operator which has better edge detection and positioning performance to replace the noise-sensitive gradient amplitude. Therefore, a new attribute vector consisted of the intensity, the second order derivative and canny operator is generated. Experimental results show that compared with gradient amplitude, canny operator is superior to identify the strong edges and avoid the noise interference resulted from scattering lines, which leads the improved attribute vector to achieve higher registration accuracy.(2) We present an adaptive feature-point selection method, which can choose the most important feature points adaptively, decrease the required feature point numbers and reduce the computation redundancy. It can also set parameters based on our expected registration effect and achieve the elastic registration of CBCT and Planning CT rapidly and accurately.(3) We provide the choice criteria of attribute vector weights, namely, to increase the proportion of strong edge in the attribute vector in order to locate the edge accurately. Experimental results show that increasing the proportion of strong edge operator can get a higher mutual information value after registration.