Deep learning pipeline for quality filtering of MRSI spectra.

With the rise of novel 3D magnetic resonance spectroscopy imaging (MRSI) acquisition protocols in clinical practice, which are capable of capturing a large number of spectra from a subject's brain, there is a need for an automated preprocessing pipeline that filters out bad-quality spectra and identifies contaminated but salvageable spectra prior to the metabolite quantification step. This work introduces such a pipeline based on an ensemble of deep-learning classifiers. The dataset consists of 36,338 spectra from one healthy subject and five brain tumor patients, acquired with an EPSI variant, which implemented a novel type of spectral editing named SLOtboom-Weng (SLOW) editing on a 7T MR scanner. The spectra were labeled manually by an expert into four classes of spectral quality as follows: (i) noise, (ii) spectra greatly influenced by lipid-related artifacts (deemed not to contain clinical information), (iii) spectra containing metabolic information slightly contaminated by lipid signals, and (iv) good-quality spectra. The AI model consists of three pairs of networks, each comprising a convolutional autoencoder and a multilayer perceptron network. In the classification step, the encoding half of the autoencoder is kept as a dimensionality reduction tool, while the fully connected layers are added to its output. Each of the three pairs of networks is trained on different representations of spectra (real, imaginary, or both), aiming at robust decision-making. The final class is assigned via a majority voting scheme. The F1 scores obtained on the test dataset for the four previously defined classes are 0.96, 0.93, 0.82, and 0.90, respectively. The arguably lower value of 0.82 was reached for the least represented class of spectra mildly influenced by lipids. Not only does the proposed model minimise the required user interaction, but it also greatly reduces the computation time at the metabolite quantification step (by selecting a subset of spectra worth quantifying) and enforces the display of only clinically relevant information.

Factorizer: A scalable interpretable approach to context modeling for medical image segmentation.

Convolutional Neural Networks (CNNs) with U-shaped architectures have dominated medical image segmentation, which is crucial for various clinical purposes. However, the inherent locality of convolution makes CNNs fail to fully exploit global context, essential for better recognition of some structures, e.g., brain lesions. Transformers have recently proven promising performance on vision tasks, including semantic segmentation, mainly due to their capability of modeling long-range dependencies. Nevertheless, the quadratic complexity of attention makes existing Transformer-based models use self-attention layers only after somehow reducing the image resolution, which limits the ability to capture global contexts present at higher resolutions. Therefore, this work introduces a family of models, dubbed Factorizer, which leverages the power of low-rank matrix factorization for constructing an end-to-end segmentation model. Specifically, we propose a linearly scalable approach to context modeling, formulating Nonnegative Matrix Factorization (NMF) as a differentiable layer integrated into a U-shaped architecture. The shifted window technique is also utilized in combination with NMF to effectively aggregate local information. Factorizers compete favorably with CNNs and Transformers in terms of accuracy, scalability, and interpretability, achieving state-of-the-art results on the BraTS dataset for brain tumor segmentation and ISLES'22 dataset for stroke lesion segmentation. Highly meaningful NMF components give an additional interpretability advantage to Factorizers over CNNs and Transformers. Moreover, our ablation studies reveal a distinctive feature of Factorizers that enables a significant speed-up in inference for a trained Factorizer without any extra steps and without sacrificing much accuracy. The code and models are publicly available at https://github.com/pashtari/factorizer.

Post-acquisition water-signal removal in 3D water-unsuppressed 1 H-MR spectroscopic imaging of the prostate.

PURPOSE: To develop a robust processing procedure of raw signals from water-unsuppressed MRSI of the prostate for the mapping of absolute tissue concentrations of metabolites. METHODS: Water-unsuppressed 3D MRSI data were acquired from a phantom, from healthy volunteers, and a patient with prostate cancer. Signal processing included sequential computation of the modulus of the FID to remove water sidebands, a Hilbert transformation, and k-space Hamming filtering. For the removal of the water signal, we compared Löwner tensor-based blind source separation (BSS) and Hankel Lanczos singular value decomposition techniques. Absolute metabolite levels were quantified with LCModel and the results were statistically analyzed to compare the water removal methods and conventional water-suppressed MRSI. RESULTS: The post-processing algorithms successfully removed the water signal and its sidebands without affecting metabolite signals. The best water removal performance was achieved by Löwner tensor-based BSS. Absolute tissue concentrations of citrate in the peripheral zone derived from water-suppressed and unsuppressed 1 H MRSI were the same and as expected from the known physiology of the healthy prostate. Maps for citrate and choline from water-unsuppressed 3D 1 H-MRSI of the prostate showed expected spatial variations in metabolite levels. CONCLUSION: We developed a robust relatively simple post-processing method of water-unsuppressed MRSI of the prostate to remove the water signal. Absolute quantification using the water signal, originating from the same location as the metabolite signals, avoids the acquisition of additional reference data.

High-resolution metabolic mapping of gliomas via patch-based super-resolution magnetic resonance spectroscopic imaging at 7T.

OBJECTIVES: To demonstrate the feasibility of 7 T magnetic resonance spectroscopic imaging (MRSI), combined with patch-based super-resolution (PBSR) reconstruction, for high-resolution multi-metabolite mapping of gliomas. MATERIALS AND METHODS: Ten patients with WHO grade II, III and IV gliomas (6/4, male/female; 45 ±â€¯9 years old) were prospectively measured between 2014 and 2018 on a 7 T whole-body MR imager after routine 3 T magnetic resonance imaging (MRI) and positron emission tomography (PET). Free induction decay MRSI with a 64 × 64-matrix and a nominal voxel size of 3.4 × 3.4 × 8 mm³ was acquired in six minutes, along with standard T1/T2-weighted MRI. Metabolic maps were obtained via spectral LCmodel processing and reconstructed to 0.9 × 0.9 × 8 mm³ resolutions via PBSR. RESULTS: Metabolite maps obtained from combined 7 T MRSI and PBSR resolved the density of metabolic activity in the gliomas in unprecedented detail. Particularly in the more heterogeneous cases (e.g. post resection), metabolite maps enabled the identification of complex metabolic activities, which were in topographic agreement with PET enhancement. CONCLUSIONS: PBSR-MRSI combines the benefits of ultra-high-field MR systems, cutting-edge MRSI, and advanced postprocessing to allow millimetric resolution molecular imaging of glioma tissue beyond standard methods. An ideal example is the accurate imaging of glutamine, which is a prime target of modern therapeutic approaches, made possible due to the higher spectral resolution of 7 T systems.

Tensor-Based Method for Residual Water Suppression in 1H Magnetic Resonance Spectroscopic Imaging.

OBJECTIVE: Magnetic resonance spectroscopic imaging (MRSI) signals are often corrupted by residual water and artifacts. Residual water suppression plays an important role in accurate and efficient quantification of metabolites from MRSI. A tensor-based method for suppressing residual water is proposed. METHODS: A third-order tensor is constructed by stacking the Löwner matrices corresponding to each MRSI voxel spectrum along the third mode. A canonical polyadic decomposition is applied on the tensor to extract the water component and to, subsequently, remove it from the original MRSI signals. RESULTS: The proposed method applied on both simulated and in-vivo MRSI signals showed good water suppression performance. CONCLUSION: The tensor-based Löwner method has better performance in suppressing residual water in MRSI signals as compared to the widely used subspace-based Hankel singular value decomposition method. SIGNIFICANCE: A tensor method suppresses residual water simultaneously from all the voxels in the MRSI grid and helps in preventing the failure of the water suppression in single voxels.

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.

Semi-automated brain tumor segmentation on multi-parametric MRI using regularized non-negative matrix factorization.

BACKGROUND: Segmentation of gliomas in multi-parametric (MP-)MR images is challenging due to their heterogeneous nature in terms of size, appearance and location. Manual tumor segmentation is a time-consuming task and clinical practice would benefit from (semi-) automated segmentation of the different tumor compartments. METHODS: We present a semi-automated framework for brain tumor segmentation based on non-negative matrix factorization (NMF) that does not require prior training of the method. L1-regularization is incorporated into the NMF objective function to promote spatial consistency and sparseness of the tissue abundance maps. The pathological sources are initialized through user-defined voxel selection. Knowledge about the spatial location of the selected voxels is combined with tissue adjacency constraints in a post-processing step to enhance segmentation quality. The method is applied to an MP-MRI dataset of 21 high-grade glioma patients, including conventional, perfusion-weighted and diffusion-weighted MRI. To assess the effect of using MP-MRI data and the L1-regularization term, analyses are also run using only conventional MRI and without L1-regularization. Robustness against user input variability is verified by considering the statistical distribution of the segmentation results when repeatedly analyzing each patient's dataset with a different set of random seeding points. RESULTS: Using L1-regularized semi-automated NMF segmentation, mean Dice-scores of 65%, 74 and 80% are found for active tumor, the tumor core and the whole tumor region. Mean Hausdorff distances of 6.1 mm, 7.4 mm and 8.2 mm are found for active tumor, the tumor core and the whole tumor region. Lower Dice-scores and higher Hausdorff distances are found without L1-regularization and when only considering conventional MRI data. CONCLUSIONS: Based on the mean Dice-scores and Hausdorff distances, segmentation results are competitive with state-of-the-art in literature. Robust results were found for most patients, although careful voxel selection is mandatory to avoid sub-optimal segmentation.

An advanced MRI and MRSI data fusion scheme for enhancing unsupervised brain tumor differentiation.

Proton Magnetic Resonance Spectroscopic Imaging (1H MRSI) has shown great potential in tumor diagnosis since it provides localized biochemical information discriminating different tissue types, though it typically has low spatial resolution. Magnetic Resonance Imaging (MRI) is widely used in tumor diagnosis as an in vivo tool due to its high resolution and excellent soft tissue discrimination. This paper presents an advanced data fusion scheme for brain tumor diagnosis using both MRSI and MRI data to improve the tumor differentiation accuracy of MRSI alone. Non-negative Matrix Factorization (NMF) of the spectral feature vectors from MRSI data and the image fusion with MRI based on wavelet analysis are implemented jointly. Hence, it takes advantage of the biochemical tissue discrimination of MRSI as well as the high resolution of MRI. The feasibility of the proposed frame work is validated by comparing with the expert delineations, giving mean correlation coefficients for the tumor source of 0.97 and the Dice score of tumor region overlap of 0.90. These results compare favorably against those obtained with a previously proposed NMF method where MRSI and MRI are integrated by stacking the MRSI and MRI features.

Hierarchical non-negative matrix factorization applied to three-dimensional 3 T MRSI data for automatic tissue characterization of the prostate.

In this study non-negative matrix factorization (NMF) was hierarchically applied to simulated and in vivo three-dimensional 3 T MRSI data of the prostate to extract patterns for tumour and benign tissue and to visualize their spatial distribution. Our studies show that the hierarchical scheme provides more reliable tissue patterns than those obtained by performing only one NMF level. We compared the performance of three different NMF implementations in terms of pattern detection accuracy and efficiency when embedded into the same kind of hierarchical scheme. The simulation and in vivo results show that the three implementations perform similarly, although one of them is more robust and better pinpoints the most aggressive tumour voxel(s) in the dataset. Furthermore, they are able to detect tumour and benign tissue patterns even in spectra with lipid artefacts. Copyright © 2016 John Wiley & Sons, Ltd.

Classifying Glioblastoma Multiforme Follow-Up Progressive vs. Responsive Forms Using Multi-Parametric MRI Features.

Purpose: The purpose of this paper is discriminating between tumor progression and response to treatment based on follow-up multi-parametric magnetic resonance imaging (MRI) data retrieved from glioblastoma multiforme (GBM) patients. Materials and Methods: Multi-parametric MRI data consisting of conventional MRI (cMRI) and advanced MRI [i.e., perfusion weighted MRI (PWI) and diffusion kurtosis MRI (DKI)] were acquired from 29 GBM patients treated with adjuvant therapy after surgery. We propose an automatic pipeline for processing advanced MRI data and extracting intensity-based histogram features and 3-D texture features using manually and semi-manually delineated regions of interest (ROIs). Classifiers are trained using a leave-one-patient-out cross validation scheme on complete MRI data. Balanced accuracy rate (BAR)-values are computed and compared between different ROIs, MR modalities, and classifiers, using non-parametric multiple comparison tests. Results: Maximum BAR-values using manual delineations are 0.956, 0.85, 0.879, and 0.932, for cMRI, PWI, DKI, and all three MRI modalities combined, respectively. Maximum BAR-values using semi-manual delineations are 0.932, 0.894, 0.885, and 0.947, for cMRI, PWI, DKI, and all three MR modalities combined, respectively. After statistical testing using Kruskal-Wallis and post-hoc Dunn-Sidák analysis we conclude that training a RUSBoost classifier on features extracted using semi-manual delineations on cMRI or on all MRI modalities combined performs best. Conclusions: We present two main conclusions: (1) using T1 post-contrast (T1pc) features extracted from manual total delineations, AdaBoost achieves the highest BAR-value, 0.956; (2) using T1pc-average, T1pc-90th percentile, and Cerebral Blood Volume (CBV) 90th percentile extracted from semi-manually delineated contrast enhancing ROIs, SVM-rbf, and RUSBoost achieve BAR-values of 0.947 and 0.932, respectively. Our findings show that AdaBoost, SVM-rbf, and RUSBoost trained on T1pc and CBV features can differentiate progressive from responsive GBM patients with very high accuracy.

Hierarchical non-negative matrix factorization to characterize brain tumor heterogeneity using multi-parametric MRI.

Tissue characterization in brain tumors and, in particular, in high-grade gliomas is challenging as a result of the co-existence of several intra-tumoral tissue types within the same region and the high spatial heterogeneity. This study presents a method for the detection of the relevant tumor substructures (i.e. viable tumor, necrosis and edema), which could be of added value for the diagnosis, treatment planning and follow-up of individual patients. Twenty-four patients with glioma [10 low-grade gliomas (LGGs), 14 high-grade gliomas (HGGs)] underwent a multi-parametric MRI (MP-MRI) scheme, including conventional MRI (cMRI), perfusion-weighted imaging (PWI), diffusion kurtosis imaging (DKI) and short-TE (1)H MRSI. MP-MRI parameters were derived: T2, T1 + contrast, fluid-attenuated inversion recovery (FLAIR), relative cerebral blood volume (rCBV), mean diffusivity (MD), fractional anisotropy (FA), mean kurtosis (MK) and the principal metabolites lipids (Lip), lactate (Lac), N-acetyl-aspartate (NAA), total choline (Cho), etc. Hierarchical non-negative matrix factorization (hNMF) was applied to the MP-MRI parameters, providing tissue characterization on a patient-by-patient and voxel-by-voxel basis. Tissue-specific patterns were obtained and the spatial distribution of each tissue type was visualized by means of abundance maps. Dice scores were calculated by comparing tissue segmentation derived from hNMF with the manual segmentation by a radiologist. Correlation coefficients were calculated between each pathologic tissue source and the average feature vector within the corresponding tissue region. For the patients with HGG, mean Dice scores of 78%, 85% and 83% were obtained for viable tumor, the tumor core and the complete tumor region. The mean correlation coefficients were 0.91 for tumor, 0.97 for necrosis and 0.96 for edema. For the patients with LGG, a mean Dice score of 85% and mean correlation coefficient of 0.95 were found for the tumor region. hNMF was also applied to reduced MRI datasets, showing the added value of individual MRI modalities.

Tumour Relapse Prediction Using Multiparametric MR Data Recorded during Follow-Up of GBM Patients.

PURPOSE: We have focused on finding a classifier that best discriminates between tumour progression and regression based on multiparametric MR data retrieved from follow-up GBM patients. MATERIALS AND METHODS: Multiparametric MR data consisting of conventional and advanced MRI (perfusion, diffusion, and spectroscopy) were acquired from 29 GBM patients treated with adjuvant therapy after surgery over a period of several months. A 27-feature vector was built for each time point, although not all features could be obtained at all time points due to missing data or quality issues. We tested classifiers using LOPO method on complete and imputed data. We measure the performance by computing BER for each time point and wBER for all time points. RESULTS: If we train random forests, LogitBoost, or RobustBoost on data with complete features, we can differentiate between tumour progression and regression with 100% accuracy, one time point (i.e., about 1 month) earlier than the date when doctors had put a label (progressive or responsive) according to established radiological criteria. We obtain the same result when training the same classifiers solely on complete perfusion data. CONCLUSIONS: Our findings suggest that ensemble classifiers (i.e., random forests and boost classifiers) show promising results in predicting tumour progression earlier than established radiological criteria and should be further investigated.

Automatic segmentation and volumetry of multiple sclerosis brain lesions from MR images.

The location and extent of white matter lesions on magnetic resonance imaging (MRI) are important criteria for diagnosis, follow-up and prognosis of multiple sclerosis (MS). Clinical trials have shown that quantitative values, such as lesion volumes, are meaningful in MS prognosis. Manual lesion delineation for the segmentation of lesions is, however, time-consuming and suffers from observer variability. In this paper, we propose MSmetrix, an accurate and reliable automatic method for lesion segmentation based on MRI, independent of scanner or acquisition protocol and without requiring any training data. In MSmetrix, 3D T1-weighted and FLAIR MR images are used in a probabilistic model to detect white matter (WM) lesions as an outlier to normal brain while segmenting the brain tissue into grey matter, WM and cerebrospinal fluid. The actual lesion segmentation is performed based on prior knowledge about the location (within WM) and the appearance (hyperintense on FLAIR) of lesions. The accuracy of MSmetrix is evaluated by comparing its output with expert reference segmentations of 20 MRI datasets of MS patients. Spatial overlap (Dice) between the MSmetrix and the expert lesion segmentation is 0.67 ± 0.11. The intraclass correlation coefficient (ICC) equals 0.8 indicating a good volumetric agreement between the MSmetrix and expert labelling. The reproducibility of MSmetrix' lesion volumes is evaluated based on 10 MS patients, scanned twice with a short interval on three different scanners. The agreement between the first and the second scan on each scanner is evaluated through the spatial overlap and absolute lesion volume difference between them. The spatial overlap was 0.69 ± 0.14 and absolute total lesion volume difference between the two scans was 0.54 ± 0.58 ml. Finally, the accuracy and reproducibility of MSmetrix compare favourably with other publicly available MS lesion segmentation algorithms, applied on the same data using default parameter settings.

Integrating diffusion kurtosis imaging, dynamic susceptibility-weighted contrast-enhanced MRI, and short echo time chemical shift imaging for grading gliomas.

BACKGROUND: We assessed the diagnostic accuracy of diffusion kurtosis imaging (DKI), dynamic susceptibility-weighted contrast-enhanced (DSC) MRI, and short echo time chemical shift imaging (CSI) for grading gliomas. METHODS: In this prospective study, 35 patients with cerebral gliomas underwent DKI, DSC, and CSI on a 3 T MR scanner. Diffusion parameters were mean diffusivity (MD), fractional anisotropy, and mean kurtosis (MK). Perfusion parameters were mean relative regional cerebral blood volume (rrCBV), mean relative regional cerebral blood flow (rrCBF), mean transit time, and relative decrease ratio (rDR). The diffusion and perfusion parameters along with 12 CSI metabolite ratios were compared among 22 high-grade gliomas and 14 low-grade gliomas (Mann-Whitney U-test, P < .05). Classification accuracy was determined with a linear discriminant analysis for each MR modality independently. Furthermore, the performance of a multimodal analysis is reported, using a decision-tree rule combining the statistically significant DKI, DSC-MRI, and CSI parameters with the lowest P-value. The proposed classifiers were validated on a set of subsequently acquired data from 19 clinical patients. RESULTS: Statistically significant differences among tumor grades were shown for MK, MD, mean rrCBV, mean rrCBF, rDR, lipids over total choline, lipids over creatine, sum of myo-inositol, and sum of creatine. DSC-MRI proved to be the modality with the best performance when comparing modalities individually, while the multimodal decision tree proved to be most accurate in predicting tumor grade, with a performance of 86%. CONCLUSIONS: Combining information from DKI, DSC-MRI, and CSI increases diagnostic accuracy to differentiate low- from high-grade gliomas, possibly providing diagnosis for the individual patient.

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.

Unsupervised nosologic imaging for glioma diagnosis.

In this letter a novel approach to create nosologic images of the brain using magnetic resonance spectroscopic imaging (MRSI) data in an unsupervised way is presented. Different tissue patterns are identified from the MRSI data using nonnegative matrix factorization and are then coded as different primary colors (i.e. red, green, and blue) in an RGB image, so that mixed tissue regions are automatically visualized as mixtures of primary colors. The approach is useful in assisting glioma diagnosis, where several tissue patterns such as normal, tumor, and necrotic tissue can be present in the same voxel/spectrum. Error-maps based on linear least squares estimation are computed for each nosologic image to provide additional reliability information, which may help clinicians in decision making. Tests on in vivo MRSI data show the potential of this new approach.

Exploiting spatial information to estimate metabolite levels in two-dimensional MRSI of heterogeneous brain lesions.

MRSI provides MR spectra from multiple adjacent voxels within a body volume represented as a two- or three-dimensional matrix, allowing the measurement of the distribution of metabolites over this volume. The spectra of these voxels are usually analyzed one by one, without exploiting their spatial context. In this article, we present an advanced metabolite quantification method for MRSI data, in which the available spatial information is considered. A nonlinear least-squares algorithm is proposed in which prior knowledge is included in the form of proximity constraints on the spectral parameters within a grid and optimized starting values. A penalty term that promotes a spatially smooth spectral parameter map is added to the fitting algorithm. This method is adaptive, in the sense that several sweeps through the grid are performed and each solution may tune some hyperparameters at run-time. Simulation studies of MRSI data showed significantly improved metabolite estimates after the inclusion of spatial information. Improved metabolite maps were also demonstrated by applying the method to in vivo MRSI data. Overlapping peaks or peaks of compounds present at low concentration can be better quantified with the proposed method than with single-voxel approaches. The new approach compares favorably against the multivoxel approach embedded in the well-known quantification software LCModel.