计算机辅助肝纤维化无创诊断

The accuracy of quantitative noninvasive diagnosis of liver fibrosis has not yet been significantly improved, even though many pioneering researches of qualitative noninvasive diagnosis have been developed in the international community. Currently most research put emphasis on the extraction of indirect feature from the original ultrasonic image, and ignores the inherent feature contained in the image. We will attempt to obtain the diagnostic mathematic model based on sparse representation of the original information from multi-view image set, and investigate the fast algorithm to resolve it. We will construct the error item of the mathematic model which adapts to the Rayleigh statistical distribution; and investigate the regularization item on the basis of L1 penalty, L2 penalty and high order statistics, and investigate the fast algorithm to resolve the mathematic model. The constructed noninvasive diagnostic model employs the inherent feature of the multi-view image set, adopts the sparse representation, therefore the robustness and performance will be improved. Split bregman algorithm improved the efficiency of the system, so the real time diagnosis is guaranteed. Our investigation is favorable to the early detection of liver fibrosis benefiting from the objective foundation of the on line diagnosis system and has a promising clinical application future.

国内外针对无创计算机辅助肝脏疾病定性诊断的研究已获突破，但在纤维化量化分期方面仍未取得显著进步。目前的研究思路多集中于应用特征提取算子，从原始超声图像中提取多种特征，利用间接加工的特征进行诊断，而不注重图像本身所蕴含的内在结构特征。本项目拟研究基于多视角图像集原始信息稀疏表示的肝纤维化分期模型及模型快速求解算法。研究内容包括构建适于超声图像Rayleigh统计分布规律的数学模型误差项；基于L1范数、L2范数以及高阶统计量构建数学模型规则项；数学模型快速求解算法。项目直接利用了肝脏超声多视角图像集内蕴含的内在结构特征，构建基于稀疏表示的肝纤维化无创定量诊断模型,能够提高系统的鲁棒性和准确率；采用线性分裂Bregman迭代算法加速了系统模型的求解,为实时诊断提供保障。本项目的研究将为肝纤维化在线诊断提供客观依据，其无创性更利于疾病早期诊断，具有现实的临床应用前景。

大量临床研究表明,肝脏疾病的早期诊断和治疗可以使早期肝纤维化或肝硬化减轻或逆转。自20世纪90年代以来，计算机辅助诊断（Computer Aided Diagnosis, CAD）获得较快发展，并应用于临床实践，在辅助提高影像科医生的诊断准确率方面取得了可喜的成就。.本研究主要在特征提取，分类模型以及快速算法方面进行了研究和改进，经实验，无创诊断取得满意结果。.由于肝脏超声图像具有回声不均匀、边缘模糊等缺点，肝脏疾病的无创诊断易受到影响。研究中提出了改进的LBP算法并提取了相应的SLBP特征，该特征较传统的纹理特征更准确地描述了B超图像中肝病变的特征，结合二维Gabor变换，解决了上述难题。由于传统的机器学习方法训练时间较长，采用基于超限学习机的训练方法并首次将其应用于肝病识别。实验表明，对测试集的分类准确率达到95.4%，在时间效率上较传统方法有很大提高，有助于肝病的临床诊断。