Introduction

MR Imaging is a powerful technique for the diagnosis, treatment planning, and progression monitoring of brain tumors. Complementarily, MR Spectroscopy (MRS) and Spectroscopic Imaging (MRSI) measure the chemical composition of tissues non-invasively, providing metabolic markers that characterize the phenotype of healthy brain tissue and different types of tumor tissue. In the case of brain tumors, MRS has been used to measure 2-hydroxy-glutarate (2HG) non-invasively^1,2,3, an oncometabolite of gliomas with isocitrate dehydrogenase (IDH) mutations. Thus, metabolic changes as measured by MRS could provide a non-invasive accurate diagnosis and stratify patients for a more patient-specific treatment of the disease. However, given the low concentration of some relevant metabolites such as 2HG, lactate (Lac), or glutamate (Glu), the sensitivity and specificity achieved with the evaluation of individual metabolites is still limited. This work proposes a standard reference frame for the evaluation of MRS scans of brain tumors. The method combines multiple metabolites to identify phenotypes⁴ of normal-appearing tissue and tumor tissue and characterize their metabolic composition using high-SNR single-voxel MRS scans. The most relevant metabolites are combined to improve the sensitivity and specificity of MRS scans of brain tumors compared to the evaluation of single metabolites, such as 2HG.

Methods

Data collection: 364 single-voxel MRS scans (40 normal-appearing tissue samples, 324 tumor-tissue samples) scans of 125 independent subjects (116 glioma patients, 9 healthy subjects), were analyzed to identify the major tissue classes in which healthy and tumor brain tissue could be classified. All MRS scans were acquired at 3T (Siemens Skyra, Prisma, and Verio, Erlangen, Germany) using PRESS localization, TE/TR=97/2000ms, and volumes in the range of 3.4-25mL. The distributions of the tissue types and tumor types in the dataset are shown in Fig.-1.
Data Processing and Quantification: The MRS datasets were reconstructed using OpenMRS Lab⁵ and quantified using LCModel⁶. The quantified metabolite ratios to total creatine (tCr) were used to build a decision tree that optimally classified the dataset into the different labels. The metabolites used as features for the decision tree were chosen from relevant tissue signatures of healthy brain tissue and tumor tissue, these included: total NAA (tNAA). total choline (tCho), lactate (Lac), glutamate (Glu), glutamine (Gln), and the combined Glu+Gln (Glx), and 2HG. The distribution of these metabolites is shown in Fig.-2.
Decision Tree Classification: Two classification models were shown: (1) a decision tree for the classification of MRS samples into normal‑appearing tissue and tumor tissue, and (2) a decision tree for the classification of tumor MRS samples into IDH-Mutant (IDH) and IDH‑wildtype (IDH-WT) tumors. The model was generated with a maximum of 5 features and a maximum depth of 4 nodes.

Results and Discussion

The receiver operating characteristic (ROC) curves of the generated classification models is shown in Fig.-3. The obtained area under the curve (AUC) were 0.93 and 0.98 for the model (1) and (2) respectively. Fig.-4 and Fig.-5 show the decision tree with thresholds that best classifies the datasets into normal and tumor tissue and IDH and IDH-WT. Given the unbalanced classes in both models, overfitting may occur for both the case of IDH vs IDH-WT tumors, as the sensitivity of 2HG is lower.

Conclusion

A standard reference frame for the classification of tissue scans using MRS was built. The methodology has the goal of improving the robustness for the identification of tumor types compared to univariate analysis, especially for the identification of IDH-mutated tumors. A standard reference like the one built here can potentially improve patient care and clinical decision making in neuro‑oncology.

Acknowledgements

No acknowledgement found.