磁共振波谱微弱信号的自动检测和量化算法研究

发布时间：2018-04-27 09:21

本文选题：磁共振波谱 + 低信噪比　；参考：《上海交通大学》2014年硕士论文

【摘要】：基于磁共振波谱（MRS）这种无创地检测活体组织特定化学特性的影像信息学方法，本文研究和设计了一套针对微弱谱峰信号的自动检测的量化算法，实现了对低信噪比信号的量化估计，为其临床应用提供了可能。本课题研究的算法包涵了信号预处理和量化估计，前者内容包括水信号的抑制、降噪和基线校正，后者主要运用ER滤波（ER-filter）和基于Hankel矩阵的奇异值分解法（HSVD）相结合的时频结合方法。在量化估计部分，针对洛仑兹模型中代表信号成分个数的模型阶数K的选取这一关键问题做了不同方法的研究测试，并通过全仿真测试实验和在真实人脑MRS数据加入测试信号的半仿真测试实验对算法的性能进行测试。在保证算法误码率在5%以下的前提下，我们通过检出率和相对均方误差（RRMSE）来衡量算法的检测和估计性能，并给出了算法适合临床应用的信噪比（SNR）要求。在模型阶数K的选取上，我们首先比较了高斯曲线拟合法、峰度分析法和固定值法的性能差别，结果显示高斯拟合法在检出率和RRMSE上都有优势。在此基础上，我们进一步研究了基于归一化拟合质量数Qfit和最小信息长度MDL的估计方法，并提出一种有条件的最小信息长度方法MDLcon，其估计性能较好，且检出率最高。考虑临床应用的误差上限条件，MDLcon与高斯拟合法相比，其准确性较高，，尤其是针对低衰减系数的信号的量化估计有较大的优势，检出率略低于高斯拟合。最后，本文对后期的工作进行了展望，希望本研究能为临床MRS微弱谱峰检测应用提供更加准确和直观的波谱信息，并提高算法的实用性和用户友好性，以实现其临床应用目标。
[Abstract]:Based on magnetic resonance spectroscopy (MRS), a non-invasive image informatics method for detecting the specific chemical characteristics of living tissues, a quantization algorithm for automatic detection of weak peak signals is studied and designed in this paper. The quantization estimation of low signal-to-noise ratio (SNR) signal is realized, which provides the possibility for its clinical application. The algorithms studied in this paper include signal preprocessing and quantization estimation. The former includes water signal suppression, noise reduction and baseline correction. In the latter case, the combination of ER filter and singular value decomposition (SVD) based on Hankel matrix is mainly used. In the part of quantization estimation, the key problem of selecting the order K of the model representing the number of signal components in the Lorentz model is studied and tested by different methods. The performance of the algorithm is tested by full-simulation test and semi-simulation test of adding test signal to real human MRS data. On the premise that the error rate of the algorithm is less than 5%, the detection rate and RRMSE are used to evaluate the detection and estimation performance of the algorithm, and the SNR requirements of the algorithm for clinical application are given. In the selection of model order K, we first compare the performance differences of Gao Si curve fitting method, kurtosis analysis method and fixed value method. The results show that Gao Si fitting method has advantages in detection rate and RRMSE. On this basis, we further study the estimation method based on normalized fitting quality number (Qfit) and minimum information length (MDL), and propose a conditional minimum information length method (MDLcon), which has better estimation performance and the highest detection rate. The accuracy of MDLcon is higher than that of Gao Si fitting, especially the quantization estimation of the signal with low attenuation coefficient has a great advantage, and the detection rate is slightly lower than that of Gao Si fitting. Finally, this paper looks forward to the later work, hoping that this study can provide more accurate and intuitionistic spectrum information for clinical application of MRS weak spectral peak detection, and improve the practicability and user-friendliness of the algorithm in order to achieve its clinical application goal.
【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：R445.2

【共引文献】