基于重复采集的活体磁共振波谱的深度学习降噪

Denoising Repeatedly Sampled In Vivo Magnetic Resonance Spectroscopy with Deep Learning

答辩人：胡琬祺

指导老师：屈小波 教授/博导

答辩委员会主席：董继扬 教授     博导 厦门大学电子科学系

答辩委员会成员：包立君 副教授   硕导 厦门大学电子科学系

林建忠 主任医师 硕导 厦门大学附属中山医院影像科

答辩秘书：杨钰 助理教授 厦门大学电子科学系

时间：2021年06月30日上午09:00

地点：厦门大学海韵园科研二-313

———————————————————————————————

摘要

磁共振波谱(Magnetic Resonance Spectroscopy，MRS)是一种无创检测活体代谢产物的方法。在实际应用中，低浓度的代谢物通常会导致相对较低的信噪比，导致从MRS获得可靠的量化信息存在挑战。因此，提高MRS信噪比成为了磁共振学术界和工业界的共同目标。

典型提高MRS信噪比方法是对多次采集的信号进行平均，但这会使得采集时间随重复采集次数线性增长。过长的扫描时间对病人也十分不友好，限制了MRS的临床应用。近年来，深度学习被引入到MRS的欠采样重建和降噪中。由于活体MRS数据较少，这些方法一般通过模拟数据来进行神经网络训练。但是MRS采集系统的不理想使得这类问题难以进行全面的数学建模。本文独辟蹊径地用MRS采集中广泛存在的重复采集数据来提升样本数量，尝试用有限实测活体MRS构建具有一定鲁棒性的MRS深度学习降噪网络。

针对重复采集的活体MRS，本文提出了一种基于深度学习的降噪方法——因果长短记忆网络，将较少次重复采样平均的低信噪比MRS映射到较多次重复采样平均的高信噪比MRS，利用重复采集数据进行排列组合构成不同水平信噪比的MRS作为网络输入和输出，解决了网络训练需要大量数据的瓶颈问题。在模拟与人体脑部数据上的实验结果表明，所提方法可以减少80%重复采集时间，且比传统的低秩降噪方法获得的代谢物量化结果更接近参考的高信噪比MRS。

针对重复采集的多通道线圈MRS，提出平均平滑奇异值分解方法。该方法首先利用奇异值分解获得初步结果，然后用二维卷积在重复采集维优化线圈灵敏度矩阵。在体模和人脑活体数据上的实验结果表明，相比传统多通道线圈合并方法所提方法可以提高合成谱的信噪比。

关键词：磁共振波谱；活体；降噪；深度学习；奇异值分解

ABSTRACT

Magnetic resonance spectroscopy (MRS) is a noninvasive tool to reveal metabolic information. However, due to the low concentration of metabolites, the low signal-to-noise ratio (SNR) of MRS in application prevents it from providing reliable quantitative information. Therefore, achieving fast acquisition of MRS with high SNR has become a common goal of magnetic resonance research in academia and industry.

To improve the SNR, the basic method is to average the repeated samplings. However, the data acquisition time is correspondingly increased, making the scanned subject uncomfortable. This greatly limits the clinical application of MRS. In recent years, deep learning has been introduced into the field of magnetic resonance. Due to the lack of in vivo MRS data, these methods are usually used the simulated data for neural network training. However, the MRS acquisition system is not ideal which makes it difficult to carry out comprehensive mathematical modelling. This paper explores a novel way to use repeated sampling, which is widely existed in MRS acquisition, to increase the number of training data, trying to use the limited in vivo MRS to get a denoising network with certain robustness.

The proposed deep learning MRS denoising method is called Causal Long Short Term Memory. First, the low SNR MRS with fewer averages of repeated sampling is mapped to the high SNR MRS with more averages of repeated sampling. Then, MRS with different SNR levels are formed by the arrangement and combination of repeated sampling data as the input and output of the network, which solves the bottleneck problem that the network training needs a large number of training data. Results on simulated data and in vivo data showed that the proposed method can obtain high quality MRS with the reduction of acquisition time by 80%. When compared with the Low Rank denoising method, the corresponding metabolite quantization results are closer to the reference high SNR spectra.

Moreover, a new method, Average Smoothing Singular Value Decomposition, is proposed to obtain higher SNR signal by combining multi-channel coil data. Based on the preliminary results of singular value decomposition, the sensitivity matrix of the coil is optimized by two-dimensional convolution in the repeated sampling dimension. Experimental results of phantom and in vivo spectra show that, compared with the traditional method, the proposed method further improves the SNR of the averaged spectra of repeated samplings.

Keywords：Magnetic Resonance Spectroscopy, In Vivo, Denoising, Deep Learning, Singular Value Decomposition.

欢迎广大师生莅临指导！