Study Of Automatic Detection And Quantification Algorithms For Weak Signals In Magnetic Resonance Spectroscopy

Based on magnetic resonance spectroscopy (MRS) which is anon-invasive imaging informatics methology for the detection of specificchemical characteristics in living tissues, our paper studied and designed aset of automatic detection and quantification algorithms aimed at weakMRS signals, and realized the estimation and quantification of low SNRspectral signal, providing possibilities to its clinical application.The algorithms in our research involve the signal pre-processing andparameter estimation. The former step includes water suppression, noisereduction and baseline correction water signal, and the latter mainly usethe time-domain frequency-domain (TDFD) method combined by ER filter(ER-filter) and Hankel singular value decomposition (HSVD). In theparameter estimation section, we did research into the critical issue onselection of the Lorentzian model order K, which represents the signalcomponent number, tried different methods and testes their performancesthrough the synthesis simulation and semi-synthesis simulation with a testsignal inserted into real human brain MRS data. Keeping the algorithms’false alarm rate below5%, we measured their detection and estimationperformances through the detection rate and the relative mean square error(RRMSE), and gave out the signal to noise ratio (SNR) requirements inconsideration of the clinical applicability.On the moder order K selection, we firstly compared the performancedifferences between the Gaussian curve fitting, kurtosis analysis and fixedvalue methods, the results of which show that the Gaussian curve fittinghas advantages on the detection rate and RRMSE over the other two. Afterthis, we further investigated the methods based on the normalized fit quality number (Qfit) and minimum description length (MDL), andproposed an approach called minimum description length with condition(MDLcon), which shows better estimation performance and the relativelyhighest detection rate. Considering the upper bound on the estimation errorin clinical application, MDLcon performs better in accuracy, especially forthe quantification of the signal with low damping factors, but performs outlower detection rate, compared to the Gaussian curve fitting method.Finally, we give prospect of the later study work. It is expected thatour work can provide more accurate and concrete spectroscopicinformation for clinical MRS detection applications on weak spectralpeaks, and improve on the algorithms’ practicability and user friendliness,realizing its clinical application goal.