【作者】 丰国栋；

【导师】 周荷琴；

【作者基本信息】 中国科学技术大学 ， 模式识别与智能系统， 2009， 博士

【摘要】 数字化X线摄影成像(Digital Radiography,DR)利用平板探测器接收X射线,直接获得数字图像信号,具有图像分辨率高、动态范围宽,成像速度快,对人体辐射小等显著优点,成为当今X线摄影领域最先进的医学成像方法,在临床诊断和科学研究中都得到了越来越广泛的应用。DR成像过程中,噪声尤其是高斯噪声降低了图像质量,X光源的尺寸、人体运动等造成图像边缘模糊,平板探测器像元的非线性响应等使得细节区域对比度不足,这些都增加了医生做出正确诊断的难度.由于以上因素的不确定性,对DR图像进行有效地增强变得十分困难,严重影响和阻碍了DR技术的发展和应用。因此,对DR图像增强方法的研究,成为当前医学成像领域中的研究热点,这方面的任何研究进展都将对DR的发展和应用起到积极的推动作用。本文在深入研究DR成像原理的基础上,分析了DR图像噪声成因和图像细节模糊的原因,提出了几种DR图像去噪、增强和分割算法,有效地提高了DR图像的质量,为组织增强研究打下坚实的基础。本文的研究从分析参数去噪模型入手,在Laplace模型的基础上,提出了适合DR图像的Laplace-Impact混合模型,提高了含噪声图像的质量;再对MUSICA等多尺度对比度增强算法进行了改进,提出了2种有效的对比度增强算法,增强了低对比度细节结构,以及重叠骨骼边缘、小骨骼边缘;最后改进了一种基于图的分割算法,将其应用于DR图像的分割,进一步提高了图像质量。本文完成了以下有新意的研究工作:1)对Laplace去噪模型进行了改进,提出了基于Laplace-Impact模型的最小均方误差估计去噪算法。它利用局部方差建立Laplace-Impact混合模型的概率密度函数来逼近高频子带的系数分布,采用最小均方误差估计调整高频小波系数来进行去噪。该算法解决了参数去噪模型无法描述双树复小波域中DR图像的高频子带系数分布符合重尾分布,在零值处具有大冲击函数特性的问题。本文算法与BLS-GSM模型和Laplace模型等经典去噪算法相比,对DR图像中的高斯噪声有更好的去噪效果。2)提出了2种DR图像对比度增强算法。算法一是在分析视觉特性的基础上,提出了一种改进的塔型多尺度增强算法。它首先通过多尺度分解,得到图像不同尺度的高频和低频信息,然后对图像的高频信息采用非线性变换进行细节部分增强,最后在图像重建过程中,对图像的低频信息采用直方图均衡宋均匀图像直方图的分布。该算法能有效提高DR胸部对比度,同时也改善了图像的整体视觉效果。算法二提出了一种DR图像多尺度多模式增强算法。它利用方向高频子带系数作为方向梯度来估计局部对比度,根据多模式选择准则把图像划分为细节区域、可能的边缘区域和平滑区域,再对可能的边缘区域进行边缘检测,得到更为准确的边界,然后对不同区域的高频子带采用多模式非线性映射函数,最后将调整后的可控金字塔系数进行逆变换,重建图像。实验结果表明,本文算法能更好地增强低对比度细节结构,并在重叠骨骼边缘、小骨骼边缘的增强等方面都取得了很好的效果。3)提出了一种基于Random-Walk算法的DR图像分割方法。它对原图像进行快速Mallat小波分解得到骨干图后,利用其高频子带梯度信息优化边的权重,并在概率阈值的准则下对争议区域做进一步划分,最后把最大到达概率所在类的标签赋予未标定顶点,并扩展到原图像,得到分割边界。该算法用图像的高频子带信息优化了Laplace图构造,并用争议区域的进一步划分取代了Random-Walk中直接将最大到达概率所在的标签赋给待标定数据的方法。用微软GrabCut分割数据库图像和实际DR图像对该算法进行了验证,能快速有效地分割出特定的图像,适用于DR图像分割。更多还原

【Abstract】 Digital Radiography(DR) is a new medical imaging method in the X-ray radiography field during the last decade.It uses flat panel detector to receive X-ray and transforms X-ray to digital image signals directly with many signific(?)nt advantages such as high image resolution,wide dynamic range,fast imaging speed and low-dose exposure on the human body.Now DR is one of the most advanced medical imaging methods,and used in clinical diagnosis more widely.The noises,especially Gaussian noise,reduce images’ quality,X-ray source size and human motion cause the fuzzy edge,and inhomogeneous resp(?)se of the flat panel detector brings about low contrast in detail region.Due to the uncertainty of the reasons above,it’s very difficult to enhance DR image quality which will hinder the development of DR technology.Therefore,DR image enhancement has become the research hotspot in the medical imaging field,and any research in this field will play an important role in the development and application of DR technology.Several DR image denoising,enhancement and segmentation algorithms are proposed in this dissertation to improve the quality of DR images and solve the key problem of tissue enhancement based on the in-depth study of DR imaging principle and the analysis of DR image noise and vague detail.Firstly,We propose Laplace-Impact mixture model to denoise DR image and improve noisy image’s quality based on Laplace model by analyzing parameter denoising method.Then two kinds of contrast enhancement algorithms are proposed to enhance low-contrast details of the structure,as well as the edges of overlapping bones and small bones,by optimizing some multiscale contrast amplification algorithms such as MUSICA.At last,we make better a segmentation algorithm based on graph cuts and apply it to DR images to improve the images’ quality further.Our meaningful and detailed works are organized as follows:1.A novel DR image denoising algorithm based on Laplace-hnpact mixture model in dual-tree complex wavelet domain is proposed to denoise Gaussian noise.It uses local variance to build probability density function of Laplace-Impact model fitted to the distribution of high-frequency subband coefficients well.Within Laplace-Impact framework,we describe a novel method for image denoising based on designing minimum mean squared error(MMSE) estimators,which relies on strong correlation between amplitudes of nearby coefficients.The experimental results show that the MMSE algorithm based on the Laplace-Impact mixture model proposed in this paper outperforms several state-of-art denoising methods such as Bayes least squared Gaussian scale mixture(BLS-GSM) and Laplace prior.2.Two improved DR image contrast enhancement algorithms are proposed.On the basis of analyzing the human vision and the multiscale contrast enhancement algorithms,a new multiscale pyramid image enhancement algorithm is proposed.At first,low frequency and high frequency subbands of different scales are got by multiscale decomposition,then the high frequency subbands are mapped by a nonlinear function,and the low frequency subbands of each level is equalized by histogram equalization during reconstructing process.The experiment shows that both the contrast and visual effect of chest radiography are improved efficiently after the process.The second algorithm proposes a novel multiscale and multi-modality nonlinear enhancement method based on local contrast.Firstly,we use high-frequency subbands’ coefficients in different direction to estimate local contrast measurement. Then the image is divided into detail,possible edge and smooth region according to multi-modality selection criteria and we use edge detection to get better edge.Finally, multi-modality nonlinear mapping function in different regions of high-frequency subband is used to enhance the image quality and reconstructed image is got through inverse steerable pyramid transforming.The experimental results show that the algorithm can enhance the low-contrast details better and have achieved very good edge enhancement results in overlapping bone and small bone structure.3.A novel method which is based on the Random-Walk algorithm is proposed for DR image segmentation.Firstly,the original image is decomposed to build the backbone graph by using Mallat’s fast wavelet transform.And the edge weight is optimized by using gradient information from high-frequency subband.Then a further division of the ambiguous area is done under the probability threshold criteria.Finally, the label with the greatest probability is assigned to each unlabeled vertex,and image segmentation boundaries are obtained by expanding the labeled backbone graph to the original image.The experimental results on Microsoft GrabCut segmentation database images and real DR images demonstrated that our algorithm is able to segment out the expectable part from DR image effectively and last(?).更多还原