To obtain accurate classifiers for assessing spectral quality of short and long-TE MRSI data from brain tumour patients to be used in clinical routine.

Data was acquired at 1.5T (Siemens Aera, Avanto) using PRESS, CHESS water-suppression, and a 32x32 grid (interpolated from 12x12). In each imaging study, short (30ms) and long (135ms) TE MRSI were recorded sequentially in the same localization. A total of 78 MRSI-recordings from 12 different brain-tumour patients (19032 spectra) acquired pre- and post-operatively, were included in the study. The measurements were performed conforming to local and national ethical regulations. Only spectra from within the PRESS-box were considered. Residual-water-peak-removal was performed prior to feature extraction (jMRUI’s HLSVD¹).

The spectra were manually labelled by two expert spectroscopists in either acceptable or non-acceptable, using jMRUI’s SpectrIm plug-in^1,2. The features for rejecting spectra were: “ghosting” artifacts, bad-shimming, low-SNR, lipid-contaminations, strongly deviating phase, and post-operative-derived artifacts. First, the experts labelled the spectra independently. Then, they revised together the spectra in which there was disagreement, reaching a consensus-labelling. The resulting consensus-labels constituted the ground-truth used for training and testing the automatic classifiers.

A total of 47 features were extracted from the magnitude time-domain (TD) and frequency-domain (FD) signals. The following types of features were used:

1. Maximum-peak-SNR in given range (FD)

2. Mean-SNR in given range (FD, TD)

3. Relative-change in given range - (TD)

4. Global features (maximum, time-point/ppm-value maximum, mean, standard-deviation, skewness, kurtosis) (FD, TD)

The strategy used for validation was Leave-Patient-Out-Cross-Validation (LPOCV), where at each time the complete dataset of one patient was excluded from the training dataset and used as the testing dataset. Short- and long-TE spectra were handled separately.

A random-forest³ (RF) classifier (“R” implementation, 500 trees, maximum depth) was used for the automatic assessment. To evaluate the relative-importance of the input-features in the classification task of both short- and long-TE spectra, the mean-decrease-in-accuracy³ after feature-permutation was measured.

The agreement between the initial labels of the experts (before reaching a consensus) was 88.78% for long-TE, and 85.04% for short-TE. On average, the experts required 3min-36sec for labelling each MRSI grid (~245 spectra).

The performance indicators for both short- and long-TE data are shown in Figure 1. In Figure 2 the Tukey boxplots of the error-rate per MRSI-grid, are presented. The feature-importance plots for both short- (blue) and long-TE (green) are shown in Figure 3. Finally, in Figure 4 several maps are shown for two example-cases: ground-truth and classifier’s prediction (probability of acceptable) for short- and long-TE, the most important feature for short-TE (TD Mean SNR in the range between 75 and 100 ms) and the two most important features for long-TE (FD-skewness and kurtosis). In order to select representative cases, the two examples presented here were chosen such that they had error-rates close to the median error-rates of both classifiers.

The results show that the error of the classifiers is at the same level as the average disagreement between the experts.

Regarding short- and long-TE data, it was shown that is more challenging to assess quality of short-TE than of long-TE spectra. This is confirmed by the higher disagreement of the experts in the assessment of short-TE spectra as well as by the higher error-rate observed in the automatic classification of short-TE data. A possible reason for this is that, besides the higher SNR of short-TE data, artifacts have also a higher SNR, which leads to a smaller number of acceptable spectra (Figure 1). These broad artifacts relax faster than metabolite signals, therefore having a smaller impact in long-TE spectra. Moreover, the variance added by the characteristic short-TE macromolecular baseline (affecting almost every spectral feature) makes the automatic classification of this data more difficult. This might also explain why features such as TD-skewness and kurtosis, that are the two most important features for the classification of long-TE data, drastically lose their importance in short-TE.

Finally, the maps presented for the two cases of Figure 4 show the higher agreement between the classifier’s prediction and the ground-truth. In the same figure the high correlation of the values of the features presented with spectral quality is also visible.

A novel method for automatic-assessment of both short- and long-TE MRSI-spectra was presented. The method shows a level-of-accuracy comparable with the one of an expert and uses a new set of spectral features with high correlation with spectral quality. The method minimizes the experts’ time needed for clinical routine MRSI-analysis.