Comparison of baseline correction algorithms for in vivo 1H-MRS.

Proton MRS is used clinically to collect localized, quantitative metabolic data from living tissues. However, the presence of baselines in the spectra complicates accurate MRS data quantification. The occurrence of baselines is not specific to short-echo-time MRS data. In short-echo-time MRS, the baseline consists typically of a dominating macromolecular (MM) part, and can, depending on B0 shimming, poor voxel placement, and/or localization sequences, also contain broad water and lipid resonance components, indicated by broad components (BCs). In long-echo-time MRS, the MM part is usually much smaller, but BCs may still be present. The sum of MM and BCs is denoted by the baseline. Many algorithms have been proposed over the years to tackle these artefacts. A first approach is to identify the baseline itself in a preprocessing step, and a second approach is to model the baseline in the quantification of the MRS data themselves. This paper gives an overview of baseline handling algorithms and also proposes a new algorithm for baseline correction. A subset of suitable baseline removal algorithms were tested on in vivo MRSI data (semi-LASER at TE = 40 ms) and compared with the new algorithm. The baselines in all datasets were removed using the different methods and subsequently fitted using spectrIm-QMRS with a TDFDFit fitting model that contained only a metabolite basis set and lacked a baseline model. The same spectra were also fitted using a spectrIm-QMRS model that explicitly models the metabolites and the baseline of the spectrum. The quantification results of the latter quantification were regarded as ground truth. The fit quality number (FQN) was used to assess baseline removal effectiveness, and correlations between metabolite peak areas and ground truth models were also examined. The results show a competitive performance of our new proposed algorithm, underscoring its automatic approach and efficiency. Nevertheless, none of the tested baseline correction methods achieved FQNs as good as the ground truth model. All separately applied baseline correction methods introduce a bias in the observed metabolite peak areas. We conclude that all baseline correction methods tested, when applied as a separate preprocessing step, yield poorer FQNs and biased quantification results. While they may enhance visual display, they are not advisable for use before spectral fitting.

Analysis and compensation of the effects of analog VLSI arithmetic on the LMS algorithm.

Analog very large scale integration implementations of neural networks can compute using a fraction of the size and power required by their digital counterparts. However, intrinsic limitations of analog hardware, such as device mismatch, charge leakage, and noise, reduce the accuracy of analog arithmetic circuits, degrading the performance of large-scale adaptive systems. In this paper, we present a detailed mathematical analysis that relates different parameters of the hardware limitations to specific effects on the convergence properties of linear perceptrons trained with the least-mean-square (LMS) algorithm. Using this analysis, we derive design guidelines and introduce simple on-chip calibration techniques to improve the accuracy of analog neural networks with a small cost in die area and power dissipation. We validate our analysis by evaluating the performance of a mixed-signal complementary metal-oxide-semiconductor implementation of a 32-input perceptron trained with LMS.