Non-negative matrix factorisation methods for the spectral decomposition of MRS data from human brain tumours

Advances in the Use of Deep Learning for the Analysis of Magnetic Resonance Image in Neuro-Oncology

Machine Learning is entering a phase of maturity, but its medical applications still lag behind in terms of practical use. The field of oncological radiology (and neuro-oncology in particular) is at the forefront of these developments, now boosted by the success of Deep-Learning methods for the analysis of medical images. This paper reviews in detail some of the most recent advances in the use of Deep Learning in this field, from the broader topic of the development of Machine-Learning-based analytical pipelines to specific instantiations of the use of Deep Learning in neuro-oncology; the latter including its use in the groundbreaking field of ultra-low field magnetic resonance imaging.

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.

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.

Establishing Imaging Biomarkers of Host Immune System Efficacy during Glioblastoma Therapy Response: Challenges, Obstacles and Future Perspectives

This hypothesis proposal addresses three major questions: (1) Why do we need imaging biomarkers for assessing the efficacy of immune system participation in glioblastoma therapy response? (2) Why are they not available yet? and (3) How can we produce them? We summarize the literature data supporting the claim that the immune system is behind the efficacy of most successful glioblastoma therapies but, unfortunately, there are no current short-term imaging biomarkers of its activity. We also discuss how using an immunocompetent murine model of glioblastoma, allowing the cure of mice and the generation of immune memory, provides a suitable framework for glioblastoma therapy response biomarker studies. Both magnetic resonance imaging and magnetic resonance-based metabolomic data (i.e., magnetic resonance spectroscopic imaging) can provide non-invasive assessments of such a system. A predictor based in nosological images, generated from magnetic resonance spectroscopic imaging analyses and their oscillatory patterns, should be translational to clinics. We also review hurdles that may explain why such an oscillatory biomarker was not reported in previous imaging glioblastoma work. Single shot explorations that neglect short-term oscillatory behavior derived from immune system attack on tumors may mislead actual response extent detection. Finally, we consider improvements required to properly predict immune system-mediated early response (1–2 weeks) to therapy. The sensible use of improved biomarkers may enable translatable evidence-based therapeutic protocols, with the possibility of extending preclinical results to human patients.

Anti-PD-1 Immunotherapy in Preclinical GL261 Glioblastoma: Influence of Therapeutic Parameters and Non-Invasive Response Biomarker Assessment with MRSI-Based Approaches

Glioblastomas (GBs) are malignant brain tumours with poor prognosis even after aggressive therapy. Programmed cell death-1 (PD-1) immune checkpoint blockade is a promising strategy in many types of cancer, but its therapeutic effects in GB remain low and associated with immune infiltration. Previous work suggests that oscillations of magnetic resonance spectroscopic imaging (MRSI)-based response pattern with chemotherapy could act as a biomarker of efficient immune system attack onto GBs. The presence of such oscillations with other monotherapies such as anti-PD-1 would reinforce its monitoring potential. Here, we confirm that the oscillatory behaviour of the response biomarker is also detected in mice treated with anti PD-1 immunotherapy both in combination with temozolomide and as monotherapy. This indicates that the spectral pattern changes observed during therapy response are shared by different therapeutic strategies, provided the host immune system is elicited and able to productively attack tumour cells. Moreover, the participation of the immune system in response is also supported by the rate of cured animals observed with different therapeutic strategies (in the range of 50–100% depending on the treatment), which also held long-term immune memory against tumour cells re-challenge. Taken together, our findings open the way for a translational use of the MRSI-based biomarker in patient-tailored GB therapy, including immunotherapy, for which reliable non-invasive biomarkers are still missing.

Embedding MRI information into MRSI data source extraction improves brain tumour delineation in animal models

Glioblastoma is the most frequent malignant intra-cranial tumour. Magnetic resonance imaging is the modality of choice in diagnosis, aggressiveness assessment, and follow-up. However, there are examples where it lacks diagnostic accuracy. Magnetic resonance spectroscopy enables the identification of molecules present in the tissue, providing a precise metabolomic signature. Previous research shows that combining imaging and spectroscopy information results in more accurate outcomes and superior diagnostic value. This study proposes a method to combine them, which builds upon a previous methodology whose main objective is to guide the extraction of sources. To this aim, prior knowledge about class-specific information is integrated into the methodology by setting the metric of a latent variable space where Non-negative Matrix Factorisation is performed. The former methodology, which only used spectroscopy and involved combining spectra from different subjects, was adapted to use selected areas of interest that arise from segmenting the T2-weighted image. Results showed that embedding imaging information into the source extraction (the proposed semi-supervised analysis) improved the quality of the tumour delineation, as compared to those obtained without this information (unsupervised analysis). Both approaches were applied to pre-clinical data, involving thirteen brain tumour-bearing mice, and tested against histopathological data. On results of twenty-eight images, the proposed Semi-Supervised Source Extraction (SSSE) method greatly outperformed the unsupervised one, as well as an alternative semi-supervised approach from the literature, with differences being statistically significant. SSSE has proven successful in the delineation of the tumour, while bringing benefits such as 1) not constricting the metabolomic-based prediction to the image-segmented area, 2) ability to deal with signal-to-noise issues, 3) opportunity to answer specific questions by allowing researchers/radiologists define areas of interest that guide the source extraction, 4) creation of an intra-subject model and avoiding contamination from inter-subject overlaps, and 5) extraction of meaningful, good-quality sources that adds interpretability, conferring validation and better understanding of each case.

Analysis of metabolic abnormalities in high-grade glioma using MRSI and convex NMF

Clinical use of MRSI is limited by the level of experience required to properly translate MRSI examinations into relevant clinical information. To solve this, several methods have been proposed to automatically recognize a predefined set of reference metabolic patterns. Given the variety of metabolic patterns seen in glioma patients, the decision on the optimal number of patterns that need to be used to describe the data is not trivial. In this paper, we propose a novel framework to (1) separate healthy from abnormal metabolic patterns and (2) retrieve an optimal number of reference patterns describing the most important types of abnormality. Using 41 MRSI examinations (1.5 T, PRESS, T_E 135 ms) from 22 glioma patients, four different patterns describing different types of abnormality were detected: edema, healthy without Glx, active tumor and necrosis. The identified patterns were then evaluated on 17 MRSI examinations from nine different glioma patients. The results were compared against BraTumIA, an automatic segmentation method trained to identify different tumor compartments on structural MRI data. Finally, the ability to predict future contrast enhancement using the proposed approach was also evaluated.

The successive projection algorithm as an initialization method for brain tumor segmentation using non-negative matrix factorization

Non-negative matrix factorization (NMF) has become a widely used tool for additive parts-based analysis in a wide range of applications. As NMF is a non-convex problem, the quality of the solution will depend on the initialization of the factor matrices. In this study, the successive projection algorithm (SPA) is proposed as an initialization method for NMF. SPA builds on convex geometry and allocates endmembers based on successive orthogonal subspace projections of the input data. SPA is a fast and reproducible method, and it aligns well with the assumptions made in near-separable NMF analyses. SPA was applied to multi-parametric magnetic resonance imaging (MRI) datasets for brain tumor segmentation using different NMF algorithms. Comparison with common initialization methods shows that SPA achieves similar segmentation quality and it is competitive in terms of convergence rate. Whereas SPA was previously applied as a direct endmember extraction tool, we have shown improved segmentation results when using SPA as an initialization method, as it allows further enhancement of the sources during the NMF iterative procedure.

Characteristic Gene Selection Based on Robust Graph Regularized Non-Negative Matrix Factorization

Many methods have been considered for gene selection and analysis of gene expression data. Nonetheless, there still exists the considerable space for improving the explicitness and reliability of gene selection. To this end, this paper proposes a novel method named robust graph regularized non-negative matrix factorization for characteristic gene selection using gene expression data, which mainly contains two aspects: Firstly, enforcing L₂₁-norm minimization on error function which is robust to outliers and noises in data points. Secondly, it considers that the samples lie in low-dimensional manifold which embeds in a high-dimensional ambient space, and reveals the data geometric structure embedded in the original data. To demonstrate the validity of the proposed method, we apply it to gene expression data sets involving various human normal and tumor tissue samples and the results demonstrate that the method is effective and feasible.

Comparison of unsupervised classification methods for brain tumor segmentation using multi-parametric MRI

Tumor segmentation is a particularly challenging task in high-grade gliomas (HGGs), as they are among the most heterogeneous tumors in oncology. An accurate delineation of the lesion and its main subcomponents contributes to optimal treatment planning, prognosis and follow-up. Conventional MRI (cMRI) is the imaging modality of choice for manual segmentation, and is also considered in the vast majority of automated segmentation studies. Advanced MRI modalities such as perfusion-weighted imaging (PWI), diffusion-weighted imaging (DWI) and magnetic resonance spectroscopic imaging (MRSI) have already shown their added value in tumor tissue characterization, hence there have been recent suggestions of combining different MRI modalities into a multi-parametric MRI (MP-MRI) approach for brain tumor segmentation. In this paper, we compare the performance of several unsupervised classification methods for HGG segmentation based on MP-MRI data including cMRI, DWI, MRSI and PWI. Two independent MP-MRI datasets with a different acquisition protocol were available from different hospitals. We demonstrate that a hierarchical non-negative matrix factorization variant which was previously introduced for MP-MRI tumor segmentation gives the best performance in terms of mean Dice-scores for the pathologic tissue classes on both datasets.

•

Unsupervised classification algorithms are applied for brain tumor segmentation on multi-parametric MRI datasets.

•

Reported mean Dice-scores are in the range of state-of-the-art segmentation algorithms.

•

Hierarchical NMF obtained the best segmentation results in terms of mean Dice-scores for most of the tissue classes.

An NMF-L2,1-Norm Constraint Method for Characteristic Gene Selection

Recent research has demonstrated that characteristic gene selection based on gene expression data remains faced with considerable challenges. This is primarily because gene expression data are typically high dimensional, negative, non-sparse and noisy. However, existing methods for data analysis are able to cope with only some of these challenges. In this paper, we address all of these challenges with a unified method: nonnegative matrix factorization via the L_2,1-norm (NMF-L_2,1). While L_2,1-norm minimization is applied to both the error function and the regularization term, our method is robust to outliers and noise in the data and generates sparse results. The application of our method to plant and tumor gene expression data demonstrates that NMF-L_2,1 can extract more characteristic genes than other existing state-of-the-art methods.

Discrete Wavelet Transform-Based Whole-Spectral and Subspectral Analysis for Improved Brain Tumor Clustering Using Single Voxel MR Spectroscopy

Many approaches have been considered for automatic grading of brain tumors by means of pattern recognition with magnetic resonance spectroscopy (MRS). Providing an improved technique which can assist clinicians in accurately identifying brain tumor grades is our main objective. The proposed technique, which is based on the discrete wavelet transform (DWT) of whole-spectral or subspectral information of key metabolites, combined with unsupervised learning, inspects the separability of the extracted wavelet features from the MRS signal to aid the clustering. In total, we included 134 short echo time single voxel MRS spectra (SV MRS) in our study that cover normal controls, low grade and high grade tumors. The combination of DWT-based whole-spectral or subspectral analysis and unsupervised clustering achieved an overall clustering accuracy of 94.8% and a balanced error rate of 7.8%. To the best of our knowledge, it is the first study using DWT combined with unsupervised learning to cluster brain SV MRS. Instead of dimensionality reduction on SV MRS or feature selection using model fitting, our study provides an alternative method of extracting features to obtain promising clustering results.

Manifold Learning in MR spectroscopy using nonlinear dimensionality reduction and unsupervised clustering

PURPOSE

To investigate whether nonlinear dimensionality reduction improves unsupervised classification of (1) H MRS brain tumor data compared with a linear method.

METHODS

In vivo single-voxel (1) H magnetic resonance spectroscopy (55 patients) and (1) H magnetic resonance spectroscopy imaging (MRSI) (29 patients) data were acquired from histopathologically diagnosed gliomas. Data reduction using Laplacian eigenmaps (LE) or independent component analysis (ICA) was followed by k-means clustering or agglomerative hierarchical clustering (AHC) for unsupervised learning to assess tumor grade and for tissue type segmentation of MRSI data.

RESULTS

An accuracy of 93% in classification of glioma grade II and grade IV, with 100% accuracy in distinguishing tumor and normal spectra, was obtained by LE with unsupervised clustering, but not with the combination of k-means and ICA. With (1) H MRSI data, LE provided a more linear distribution of data for cluster analysis and better cluster stability than ICA. LE combined with k-means or AHC provided 91% accuracy for classifying tumor grade and 100% accuracy for identifying normal tissue voxels. Color-coded visualization of normal brain, tumor core, and infiltration regions was achieved with LE combined with AHC.

CONCLUSION

The LE method is promising for unsupervised clustering to separate brain and tumor tissue with automated color-coding for visualization of (1) H MRSI data after cluster analysis.

Data analysis and tissue type assignment for glioblastoma multiforme

Glioblastoma multiforme (GBM) is characterized by high infiltration. The interpretation of MRSI data, especially for GBMs, is still challenging. Unsupervised methods based on NMF by Li et al. (2013, NMR in Biomedicine) and Li et al. (2013, IEEE Transactions on Biomedical Engineering) have been proposed for glioma recognition, but the tissue types is still not well interpreted. As an extension of the previous work, a tissue type assignment method is proposed for GBMs based on the analysis of MRSI data and tissue distribution information. The tissue type assignment method uses the values from the distribution maps of all three tissue types to interpret all the information in one new map and color encodes each voxel to indicate the tissue type. Experiments carried out on in vivo MRSI data show the feasibility of the proposed method. This method provides an efficient way for GBM tissue type assignment and helps to display information of MRSI data in a way that is easy to interpret.

Hierarchical non-negative matrix factorization (hNMF): a tissue pattern differentiation method for glioblastoma multiforme diagnosis using MRSI

MRSI has shown potential in the diagnosis and prognosis of glioblastoma multiforme (GBM) brain tumors, but its use is limited by difficult data interpretation. When the analyzed MRSI data present more than two tissue patterns, conventional non-negative matrix factorization (NMF) implementation may lead to a non-robust estimation. The aim of this article is to introduce an effective approach for the differentiation of GBM tissue patterns using MRSI data. A hierarchical non-negative matrix factorization (hNMF) method that can blindly separate the most important spectral sources in short-TE ¹H MRSI data is proposed. This algorithm consists of several levels of NMF, where only two tissue patterns are computed at each level. The method is demonstrated on both simulated and in vivo short-TE ¹H MRSI data in patients with GBM. For the in vivo study, the accuracy of the recovered spectral sources was validated using expert knowledge. Results show that hNMF is able to accurately estimate the three tissue patterns present in the tumoral and peritumoral area of a GBM, i.e. normal, tumor and necrosis, thus providing additional useful information that can help in the diagnosis of GBM. Moreover, the hNMF results can be displayed as easily interpretable maps showing the contribution of each tissue pattern to each voxel.