¹H_Magnetic-Resonance-Spectroscopy (MRS) provides non-invasive measurements of metabolite profiles with the potential to aid diagnosis and management of brain tumours. To permit the analysis of tumour specific MRS patterns and their use for tumour classification, several pattern-recognition methods are available. These techniques can be applied to the pre-processed complete spectra or quantified metabolite-concentrations. The main challenge in analysis of the tumour patterns is the imbalanced distribution of the cases in the tumour classes. In a given learning task, data size has an important role in building reliable classification-algorithms. Skewed data distribution, both within and across class, introduces difficulty in classification and results in learning performance deterioration¹.

This study assesses different imbalanced-multi-class learning techniques and objectively compares the use of more than two input features for childhood brain tumour classification using MRS

Single-centre MRS data were collected retrospectively from children with a suspected brain tumour prior to treatment. The enrolled cohort consisted of 90 patients (age 6.9±4.29 y, 42 female, 48 male) with 3 different tumour types, including 38 medullobastoma, 42 pilocytic-astrocytoma and 10 ependymomas from all regions of the brain. Histopathological, clinical and radiological features, as available, were used for diagnosis.

Single-voxel-MRS data were acquired on 1.5T MRI scanners (Siemens and GE) using a standard protocol (PRESS, TE 30 ms, TR 1500 ms, spectral resolution 1 Hz/point). Raw spectroscopy data were analysed using TARQUIN². Pre-defined quality control criteria were applied. Frequency alignment, zero order phase correction, baseline-correction and water removal using HSVD methods were applied by TARQUIN. Both spectra and metabolite-concentrations were used to develop the imbalanced learning-algorithms and their influence on classification performance was evaluated.

To correct for the effect of class-imbalance, the synthetic minority oversampling technique (SMOTE) ³ was used to overpopulate the original ependymoma group by 200 % and correct for the skewness in the original spectra and metabolite-concentration data (figure1). Principal-component-analysis (PCA) was used for dimension reduction. Classifiers were trained using Support-Vector-Machine (SVM), Linear-Discriminative-Analysis (LDA), Artificial-Neural-Networks (ANN) and Random-Forest (RF). Ten-fold cross-validation was used to evaluate the learning-algorithm performance. Balanced-classification-accuracy (BCA) was calculated as an overall measure of success. F-measure and G-mean metrics were used for performance evaluation in class imbalance learning ¹.

Developed classifiers using the SMOTE-overpopulated sets were tested using the original data and the validity of the classifier was checked by comparison with the known histological diagnosis. Different pattern-recognition approaches were compared in view of their performance.

PCA was performed and PCs accounting for 90 % of variance extracted, giving 33 and 13 principal-components for spectra and metabolite concentrations respectively. Separation between the tumour-groups using metabolite concentrations increased when the SMOTE-overpopulated set was used (figure 2).

Overall classifier performance and F-measure¹ along with the G-mean¹ for ependymoma group for all 4 training sets are illustrated in figure 3. Using complete original spectra all classifiers reached a BCA of more than 0.8 except for LDA which had an accuracy of 0.76 (due to failure in discriminating ependymoma, F-measure=0.25). In contrast when original metabolite concentrations were used for classification, LDA compared favourably with the other methods (BCA=0.91) (figure 3-a). The higher BCA obtained with metabolite-concentration compared with original complete spectra is the result of better classification in medullobastoma and pilocytic-astrocytoma groups (F-measure in figure 3-d).

Performance of all classifiers for discriminating ependymomas increased with SMOTE overpopulated data-sets compared to original data-sets (Mean ependymoma F-measure for complete spectra: SMOTE=0.96, original=0.5 ,p-value=0.01 ; Mean ependymoma F-measure for metabolite-concentration: SMOTE=0.94, original=0.55 ,p-value=0.02) (figure 3-c), resulting in an overall better classification outcome with complete SMOTE complete spectra and SMOTE metabolite-concentration .

Classification of paediatric brain tumours from ¹H_MRS has many desirable characteristics. However the imbalanced nature of the data introduces difficulties in uncovering regularities within the small rare tumour type group and attempts to train learning algorithms without correcting the skewed distribution may be premature.

Here it has been shown that by fusing oversampling and classification techniques together, improved classification-accuracy for all methods with higher discrimination for ependymoma class was obtained . Average classification-accuracy increase varied by input, complete spectra improved by 23% and metabolite-concentration by 10%.

Moreover a good classification performance can be obtained using oversampled complete spectra. This indicates that in the presence of sufficient data samples, processing of MRS data can be simplified for classification purposes.

The choice of learning algorithm, use of oversampling technique and classifier input depends on the data distribution, required accuracy in discriminating specific groups and degree of post-processing complexity.