Classification of brain tumours from MR spectra: the INTERPRET collaboration and its outcomes

Machine Learning Analysis of Single-Voxel Proton MR Spectroscopy for Differentiating Solitary Fibrous Tumors and Meningiomas

Solitary fibrous tumor (SFT), formerly known as hemangiopericytoma, is an uncommon brain tumor often confused with meningioma on MRI. Unlike meningiomas, SFTs exhibit a myoinositol peak on magnetic resonance spectroscopy (MRS). This study aimed to develop automated classifiers to distinguish SFT from meningioma grades using MRS data from a 26-year patient cohort. Four classification tasks were performed on short echo (SE), long echo (LE) time, and concatenated SE + LE spectra, with datasets split into 80% training and 20% testing sets. Sequential forward feature selection and linear discriminant analysis identified features to distinguish between meningioma Grade 1 (Men-1), meningioma grade 2 (Men-2), meningioma grade 3 (Men-3), and SFT (the 4-class classifier); Men-1 from Men-2 + 3 + SFT; meningioma (all) from SFT; and Men-1 from Men-2 + 3 and SFT. The best classifier was defined by the smallest balanced error rate (BER) in the testing phase. A total of 136 SE cases and 149 LE cases were analyzed. The best features in the 4-class classifier were myoinositol and alanine at SE, and myoinositol, glutamate, and glutamine at LE. Myoinositol alone distinguished SFT from meningiomas. Differentiating Men-1 from Men-2 was not possible with MRS, and combining higher meningioma grades did not improve distinction from Men-1. Notably, combining short and long echo times (TE) enhances classification performance, particularly in challenging outlier cases. Furthermore, the robust classifier demonstrates efficacy even when dealing with spectra of suboptimal quality. The resulting classifier is available as Supporting Information of the publication. Extensive documentation is provided, and the software is free and open to all users without a login requirement.

A comparison of non-negative matrix underapproximation methods for the decomposition of magnetic resonance spectroscopy data from human brain tumors

Magnetic resonance spectroscopy (MRS) is an MR technique that provides information about the biochemistry of tissues in a noninvasive way. MRS has been widely used for the study of brain tumors, both preoperatively and during follow-up. In this study, we investigated the performance of a range of variants of unsupervised matrix factorization methods of the non-negative matrix underapproximation (NMU) family, namely, sparse NMU, global NMU, and recursive NMU, and compared them with convex non-negative matrix factorization (C-NMF), which has previously shown a good performance on brain tumor diagnostic support problems using MRS data. The purpose of the investigation was 2-fold: first, to ascertain the differences among the sources extracted by these methods; and second, to compare the influence of each method in the diagnostic accuracy of the classification of brain tumors, using them as feature extractors. We discovered that, first, NMU variants found meaningful sources in terms of biological interpretability, but representing parts of the spectrum, in contrast to C-NMF; and second, that NMU methods achieved better classification accuracy than C-NMF for the classification tasks when one class was not meningioma.

Using Single-Voxel Magnetic Resonance Spectroscopy Data Acquired at 1.5T to Classify Multivoxel Data at 3T: A Proof-of-Concept Study

In vivo magnetic resonance spectroscopy (MRS) has two modalities, single-voxel (SV) and multivoxel (MV), in which one or more contiguous grids of SVs are acquired.

PURPOSE

To test whether MV grids can be classified with models trained with SV.

METHODS

Retrospective study. Training dataset: Multicenter multiformat SV INTERPRET, 1.5T. Testing dataset: MV eTumour, 3T. Two classification tasks were completed: 3-class (meningioma vs. aggressive vs. normal) and 4-class (meningioma vs. low-grade glioma vs. aggressive vs. normal). Five different methods were tested for feature selection. The classification was implemented using linear discriminant analysis (LDA), random forest, and support vector machines. The evaluation was completed with balanced error rate (BER) and area under the curve (AUC) on both sets. The accuracy in class prediction was calculated by developing a solid tumor index (STI) and segmentation accuracy with the Dice score.

RESULTS

The best method was sequential forward feature selection combined with LDA, with AUCs = 0.95 (meningioma), 0.89 (aggressive), 0.82 (low-grade glioma), and 0.82 (normal). STI was 66% (4-class task) and 71% (3-class task) because two cases failed completely and two more had suboptimal STI as defined by us.

DISCUSSION

The reasons for failure in the classification of the MV test set were related to the presence of artifacts.

Magnetic resonance spectroscopy ex vivo: A short historical review

Over the last half century, there have been several periods during which magnetic resonance spectroscopy (MRS) has been used ex vivo, for a variety of reasons, on samples such as microorganisms, cells, animal or human tissue, tissue extracts or biological fluids. These studies began in the days before the acronym MRS had been invented, when all such methods were still called nuclear magnetic resonance (NMR), and have extended to the present day. I will describe the historical development of NMR methods used ex vivo, their influences on the development of MRS in vivo, and their longer-term uses. All the interpretations will be personal, based on what I saw, or discussed with colleagues at the time.

Extraction of artefactual MRS patterns from a large database using non-negative matrix factorization

Despite the success of automated pattern recognition methods in problems of human brain tumor diagnostic classification, limited attention has been paid to the issue of automated data quality assessment in the field of MRS for neuro-oncology. Beyond some early attempts to address this issue, the current standard in practice is MRS quality control through human (expert-based) assessment. One aspect of automatic quality control is the problem of detecting artefacts in MRS data. Artefacts, whose variety has already been reviewed in some detail and some of which may even escape human quality control, have a negative influence in pattern recognition methods attempting to assist tumor characterization. The automatic detection of MRS artefacts should be beneficial for radiology as it guarantees more reliable tumor characterizations, as well as the development of more robust pattern recognition-based tumor classifiers and more trustable MRS data processing and analysis pipelines. Feature extraction methods have previously been used to help distinguishing between good and bad quality spectra to apply subsequent supervised pattern recognition techniques. In this study, we apply feature extraction differently and use a variant of a method for blind source separation, namely Convex Non-Negative Matrix Factorization, to unveil MRS signal sources in a completely unsupervised way. We hypothesize that, while most sources will correspond to the different tumor patterns, some of them will reflect signal artefacts. The experimental work reported in this paper, analyzing a combined short and long echo time ¹ H-MRS database of more than 2000 spectra acquired at 1.5T and corresponding to different tumor types and other anomalous masses, provides a first proof of concept that points to the possible validity of this approach.

Deep learning magnetic resonance spectroscopy fingerprints of brain tumours using quantum mechanically synthesised data

Metabolic fingerprints are valuable biomarkers for diseases that are associated with metabolic disorders. 1H magnetic resonance spectroscopy (MRS) is a unique noninvasive diagnostic tool that can depict the metabolic fingerprint based solely on the proton signal of different molecules present in the tissue. However, its performance is severely hindered by low SNR, field inhomogeneities and overlapping spectra of metabolites, which affect the quantification of metabolites. Consequently, MRS is rarely included in routine clinical protocols and has not been proven in multi-institutional trials. This work proposes an alternative approach, where instead of quantifying metabolites' concentration, deep learning (DL) is used to model the complex nonlinear relationship between diseases and their spectroscopic metabolic fingerprint (pattern). DL requires large training datasets, acquired (ideally) with the same protocol/scanner, which are very rarely available. To overcome this limitation, a novel method is proposed that can quantum mechanically synthesise MRS data for any scanner/acquisition protocol. The proposed methodology is applied to the challenging clinical problem of differentiating metastasis from glioblastoma brain tumours on data acquired across multiple institutions. DL algorithms were trained on the augmented synthetic spectra and tested on two independent datasets acquired by different scanners, achieving a receiver operating characteristic area under the curve of up to 0.96 and 0.97, respectively.

Robust Conditional Independence maps of single-voxel Magnetic Resonance Spectra to elucidate associations between brain tumours and metabolites

The aim of the paper is two-fold. First, we show that structure finding with the PC algorithm can be inherently unstable and requires further operational constraints in order to consistently obtain models that are faithful to the data. We propose a methodology to stabilise the structure finding process, minimising both false positive and false negative error rates. This is demonstrated with synthetic data. Second, to apply the proposed structure finding methodology to a data set comprising single-voxel Magnetic Resonance Spectra of normal brain and three classes of brain tumours, to elucidate the associations between brain tumour types and a range of observed metabolites that are known to be relevant for their characterisation. The data set is bootstrapped in order to maximise the robustness of feature selection for nominated target variables. Specifically, Conditional Independence maps (CI-maps) built from the data and their derived Bayesian networks have been used. A Directed Acyclic Graph (DAG) is built from CI-maps, being a major challenge the minimization of errors in the graph structure. This work presents empirical evidence on how to reduce false positive errors via the False Discovery Rate, and how to identify appropriate parameter settings to improve the False Negative Reduction. In addition, several node ordering policies are investigated that transform the graph into a DAG. The obtained results show that ordering nodes by strength of mutual information can recover a representative DAG in a reasonable time, although a more accurate graph can be recovered using a random order of samples at the expense of increasing the computation time.

Cancer metabolism in a snapshot: MRS(I)

The contribution of MRS(I) to the in vivo evaluation of cancer-metabolism-derived metrics, mostly since 2016, is reviewed here. Increased carbon consumption by tumour cells, which are highly glycolytic, is now being sampled by ¹³ C magnetic resonance spectroscopic imaging (MRSI) following the injection of hyperpolarized [1-¹³ C] pyruvate (Pyr). Hot-spots of, mostly, increased lactate dehydrogenase activity or flow between Pyr and lactate (Lac) have been seen with cancer progression in prostate (preclinical and in humans), brain and pancreas (both preclinical) tumours. Therapy response is usually signalled by decreased Lac/Pyr ¹³ C-labelled ratio with respect to untreated or non-responding tumour. For therapeutic agents inducing tumour hypoxia, the ¹³ C-labelled Lac/bicarbonate ratio may be a better metric than the Lac/Pyr ratio. ³¹ P MRSI may sample intracellular pH changes from brain tumours (acidification upon antiangiogenic treatment, basification at fast proliferation and relapse). The steady state tumour metabolome pattern is still in use for cancer evaluation. Metrics used for this range from quantification of single oncometabolites (such as 2-hydroxyglutarate in mutant IDH1 glial brain tumours) to selected metabolite ratios (such as total choline to N-acetylaspartate (plain ratio or CNI index)) or the whole ¹ H MRSI(I) pattern through pattern recognition analysis. These approaches have been applied to address different questions such as tumour subtype definition, following/predicting the response to therapy or defining better resection or radiosurgery limits.

Evaluation of the INTERPRET decision-support system: can it improve the diagnostic value of magnetic resonance spectroscopy of the brain?

Purpose

We evaluated in a clinical setting the INTERPRET decision-support system (DSS), a software generated to aid in MRS analysis to achieve a specific diagnosis for brain lesions.

Methods

The material consisted of 100 examinations of focal intracranial lesions with confirmed diagnoses. MRS was obtained at 1.5 T using TE 20–30 ms. Data were processed with the LCModel for conventional analysis. The INTERPRET DSS 3.1. was used to obtain specific diagnoses. MRI and MRS were reviewed by one interpreter. DSS analysis was made by another interpreter, in 80 cases by two interpreters. The diagnoses were compared with the definitive diagnoses. For comparisons between DSS, conventional MRS analysis, and MRI, the diagnoses were categorised: high-grade tumour, low-grade tumour, non-neoplastic lesion.

Results

Interobserver agreement in choosing the diagnosis from the INTERPRET database was 75%. The diagnosis was correct in 38/100 cases, incorrect in 57 cases. No good match was found in 5/100 cases. The diagnostic category was correct with DSS/conventional MRS/MRI in 67/58/52 cases, indeterminate in 5/8/20 cases, incorrect in 28/34/28 cases. Results with DSS were not significantly better than with conventional MRS analysis. All definitive diagnoses did not exist in the INTERPRET database. In the 61 adult patients with the diagnosis included in the database, DSS/conventional MRS/MRI yielded a correct diagnosis category in 48/32/29 cases (DSS vs conventional MRS: p = 0.002, DSS vs MRI: p = 0.0004).

Conclusion

Use of the INTERPRET DSS did not improve MRS categorisation of the lesions in the unselected clinical cases. In adult patients with lesions existing in the INTERPRET database, DSS improved the results, which indicates the potential of this software with an extended database.

Electronic supplementary material

The online version of this article (10.1007/s00234-018-2129-7) contains supplementary material, which is available to authorized users.

Clinical applications of in vivo magnetic resonance spectroscopy in oncology

Magnetic resonance spectroscopy (MRS) can be performed in vivo using commercial MRI systems to obtain biochemical information about tissues and cancers. Applications in brain, prostate and breast aid lesion detection and characterisation (differential diagnosis), treatment planning and response assessment. Multi-centre clinical trials have been performed in all these tissues. Single centre studies have been performed in many other tissues including cervix, uterus, musculoskeletal and liver. While generally MRS is used to study endogenous metabolites it has also been used in drug studies, for example those that include 19F as part of their structure. Recently the hyperpolarisation of compounds enriched with 13C such as [1-13C] pyruvate has been demonstrated in animal models and now in preliminary clinical studies, permitting the monitoring of biochemical processes with unprecedented sensitivity. This review briefly introduces the underlying methods and then discusses the current status of these applications.

ACRIN 6684: Multicenter, phase II assessment of tumor hypoxia in newly diagnosed glioblastoma using magnetic resonance spectroscopy

A multi-center imaging trial by the American College of Radiology Imaging Network (ACRIN) "A Multicenter, phase II assessment of tumor hypoxia in glioblastoma using 18F Fluoromisonidazole (FMISO) with PET and MRI (ACRIN 6684)", was conducted to assess hypoxia in patients with glioblastoma (GBM). The aims of this study were to support the role of proton magnetic resonance spectroscopic imaging (1H MRSI) as a prognostic marker for brain tumor patients in multi-center clinical trials. Seventeen participants from four sites had analyzable 3D MRSI datasets acquired on Philips, GE or Siemens scanners at either 1.5T or 3T. MRSI data were analyzed using LCModel to quantify metabolites N-acetylaspartate (NAA), creatine (Cr), choline (Cho), and lactate (Lac). Receiver operating characteristic curves for NAA/Cho, Cho/Cr, lactate/Cr, and lactate/NAA were constructed for overall survival at 1-year (OS-1) and 6-month progression free survival (PFS-6). The OS-1 for the 17 evaluable patients was 59% (10/17). Receiver operating characteristic analyses found the NAA/Cho in tumor (AUC = 0.83, 95% CI: 0.61 to 1.00) and in peritumoral regions (AUC = 0.95, 95% CI 0.85 to 1.00) were predictive for survival at 1 year. PFS-6 was 65% (11/17). Neither NAA/Cho nor Cho/Cr was effective in predicting 6-month progression free survival. Lac/Cr in tumor was a significant negative predictor of PFS-6, indicating that higher lactate/Cr levels are associated with poorer outcome. (AUC = 0.79, 95% CI: 0.54 to 1.00). In conclusion, despite the small sample size in the setting of a multi-center trial comprising different vendors, field strengths, and varying levels of expertise at data acquisition, MRS markers NAA/Cho, Lac/Cr and Lac/NAA predicted overall survival at 1 year and 6-month progression free survival. This study validates that MRSI may be useful in evaluating the prognosis in glioblastoma and should be considered for incorporating into multi-center clinical trials.

Exploration of human brain tumour metabolism using pairwise metabolite-metabolite correlation analysis (MMCA) of HR-MAS 1H NMR spectra

METHODS

We quantified 378 HRMAS 1H NMR spectra of human brain tumours (132 glioblastomas, 101 astrocytomas, 75 meningiomas, 37 oligodendrogliomas and 33 metastases) from the eTumour database and looked for metabolic interactions by metabolite-metabolite correlation analysis (MMCA).

RESULTS

All tumour types showed remarkably similar metabolic correlations. Lactate correlated positively with alanine, glutamate with glutamine; creatine + phosphocreatine (tCr) correlated positively with lactate, alanine and choline + phosphocholine + glycerophosphocholine (tCho), and tCho correlated positively with lactate; fatty acids correlated negatively with lactate, glutamate + glutamine (tGlut), tCr and tCho. Oligodendrogliomas had fewer correlations but they still fitted that pattern.

CONCLUSIONS

Possible explanations include (i) glycolytic tumour cells (the Warburg effect) generating pyruvate which is converted to lactate, alanine, glutamate and then glutamine; (ii) an association between elevated glycolysis and increased choline turnover in membranes; (iii) an increase in the tCr pool to facilitate phosphocreatine-driven glutamate uptake; (iv) lipid signals come from cytosolic lipid droplets in necrotic or pre-necrotic tumour tissue that has lower concentrations of anabolic and catabolic metabolites. Additional metabolite exchanges with host cells may also be involved. If tumours co-opt a standard set of biochemical mechanisms to grow in the brain, then drugs might be developed to disrupt those mechanisms.

Quality of clinical brain tumor MR spectra judged by humans and machine learning tools

PURPOSE

To investigate and compare human judgment and machine learning tools for quality assessment of clinical MR spectra of brain tumors.

METHODS

A very large set of 2574 single voxel spectra with short and long echo time from the eTUMOUR and INTERPRET databases were used for this analysis. Original human quality ratings from these studies as well as new human guidelines were used to train different machine learning algorithms for automatic quality control (AQC) based on various feature extraction methods and classification tools. The performance was compared with variance in human judgment.

RESULTS

AQC built using the RUSBoost classifier that combats imbalanced training data performed best. When furnished with a large range of spectral and derived features where the most crucial ones had been selected by the TreeBagger algorithm it showed better specificity (98%) in judging spectra from an independent test-set than previously published methods. Optimal performance was reached with a virtual three-class ranking system.

CONCLUSION

Our results suggest that feature space should be relatively large for the case of MR tumor spectra and that three-class labels may be beneficial for AQC. The best AQC algorithm showed a performance in rejecting spectra that was comparable to that of a panel of human expert spectroscopists. Magn Reson Med 79:2500-2510, 2018. © 2017 International Society for Magnetic Resonance in Medicine.

Magnetic Resonance Spectroscopy and Single-Photon Emission Computed Tomography in the Evaluation of Cerebral Tumors: A Case Report

In their daily clinical practice, physicians have to confront diagnostic dilemmas which cannot be resolved by the application of only one imaging technique. In this case report, we present a 66-year-old woman who was admitted to our institution for the surgical resection of a recently diagnosed brain tumor. The patient had a history of epileptic seizures and was hospitalized in the past for anti-phospholipid syndrome related to a non-Hodgkin lymphoma in remission. Magnetic resonance imaging (MRI) examination revealed an enhancing right parasagittal lesion with significant edema suggestive of a high grade glioma. Advanced MRI techniques including proton magnetic resonance spectroscopy (¹H-MRS) showed findings compatible of glioma. An additional examination was performed as part of a protocol that we are routinely performing in our institution for all brain tumors including not only the gold standard advanced MRI techniques but also single-photon emission computed tomography (SPECT) with technetium-99m (Tc99m). Brain SPECT indicated the presence of a meningioma which was verified by the histopathology of the resected specimen. In conclusion, a multimodality approach for the pre-surgical assessment of brain tumors has significant advantages not only for the diagnosis but also for the evaluation of intracranial tumors histology.

Refined modelling of the short-T2 signal component and ensuing detection of glutamate and glutamine in short-TE, localised, (1) H MR spectra of human glioma measured at 3 T

Short-TE (1) H MRS has great potential for brain cancer diagnostics. A major difficulty in the analysis of the spectra is the contribution from short-T2 signal components, mainly coming from mobile lipids. This complicates the accurate estimation of the spectral parameters of the resonance lines from metabolites, so that a qualitative to semi-quantitative interpretation of the spectra dominates in practice. One solution to overcome this difficulty is to measure and estimate the short-T2 signal component and to subtract it from the total signal, thus leaving only the metabolite signals. The technique works well when applied to spectra obtained from healthy individuals, but requires some optimisation during data acquisition. In the clinical setting, time constraints hardly allow this. Here, we propose an iterative estimation of the short-T2 signal component, acquired in a single acquisition after measurement of the full spectrum. The method is based on QUEST (quantitation based on quantum estimation) and allows the refinement of the estimate of the short-T2 signal component after measurement. Thus, acquisition protocols used on healthy volunteers can also be used on patients without further optimisation. The aim is to improve metabolite detection and, ultimately, to enable the estimation of the glutamine and glutamate signals distinctly. These two metabolites are of great interest in the characterisation of brain cancer, gliomas in particular. When applied to spectra from healthy volunteers, the new algorithm yields similar results to QUEST and direct subtraction of the short-T2 signal component. With patients, up to 12 metabolites and, at least, seven can be quantified in each individual brain tumour spectrum, depending on the metabolic state of the tumour. The refinement of the short-T2 signal component significantly improves the fitting procedure and produces a separate short-T2 signal component that can be used for the analysis of mobile lipid resonances. Thus, in brain tumour spectra, distinct estimates of signals from glutamate and glutamine are possible. Copyright © 2016 John Wiley & Sons, Ltd.