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

Systematic comparison of unsupervised feature extraction methods for enhancing discrimination accuracy of brown rice flour adulteration in LIBS

In this article an accurate and efficient technique for tissue typing is presented. The proposed technique is based on Canonical Correlation Analysis, a statistical method able to simultaneously exploit the spectral and spatial information characterizing the Magnetic Resonance Spectroscopic Imaging (MRSI) data. Recently, Canonical Correlation Analysis has been successfully applied to other types of biomedical data, such as functional MRI data. Here, Canonical Correlation Analysis is adapted for MRSI data processing in order to retrieve in an accurate and efficient way the possible tissue types that characterize the organ under investigation. The potential and limitations of the new technique have been investigated by using simulated as well as in vivo prostate MRSI data, and extensive studies demonstrate a high accuracy, robustness, and efficiency. Moreover, the performance of Canonical Correlation Analysis has been compared to that of ordinary correlation analysis. The test results show that Canonical Correlation Analysis performs best in terms of accuracy and robustness.

SummaryBackgroundProcesses leading to the onset of neurodegeneration in amyotrophic lateral sclerosis (ALS) are largely unknown. To gain insight into disease mechanisms, we measured brain metabolism in vivo in asymptomatic C9orf72 mutation carriers and patients with ALS.MethodsWe enroled 15 asymptomatic family members with (AFM C9+) and 18 without a C9orf72 mutation (AFM C9−), 4 patients with ALS with (ALS C9+) and 35 without this mutation (ALS C9−), and 25 population-based controls (CO). Two-dimensional proton (1H) and whole-brain phosphorus (31P) magnetic resonance spectroscopic imaging (MRSI) data were obtained using a 7T MR scanner. 11 brain metabolites were compared between groups using weighted Bayesian linear multilevel models.FindingsCompared to AFM C9−, AFM C9+ showed evidence of neuronal dysfunction (decreased total N-acetyl aspartate/total creatine (tNAA/tCr)), widespread increased membrane breakdown product (glycerol phosphorylethanolamine/phosphocreatine (GPE/PCr)), glutamate excitotoxicity (increased glutamate + glutamine/tNAA (Glx/tNAA)) and, in frontoparietal regions, an increase in the glycogen precursor uridine diphosphoglucose/PCr (UDPG/PCr). Compared to AFM C9+, neuronal dysfunction and membrane breakdown are similar in ALS C9+, but glutamate excitotoxicity and increased glycogen precursor are more severe and widespread, also involving the primary motor region. Moreover, lower total adenosine triphosphate/PCr (tATP/PCr) emerged in ALS C9+, and signs of disturbed membrane synthesis, intracellular second messenger system and glial pathology (myo-inositol + glycine/tCr (mI + Gly/tCr)). ALS C9− is characterised by glutamate excitotoxicity, increased tATP/PCr, and lower phospholipid levels.Interpretation1H and 31P 7T MRSI can detect evolving patterns of altered brain metabolism in asymptomatic mutation carriers and patients with ALS. Abnormalities in patients with ALS C9+ appeared to be different from those in patients with ALS C9−. Metabolic markers, measured in vivo, can serve as biomarkers for inclusion or stratification as well as for drug-target engagement in clinical trials. This method can facilitate identification of new and personalised drug targets to prevent or treat this devastating disease.FundingALS Foundation Netherlands.

The application of proton (1H) magnetic resonance spectroscopic imaging (MRSI) allows for noninvasive, localized analyses of brain biochemistry; however, minimal work has been devoted to the evaluation of 1H MRSI reproducibility. This study examined the reproducibility of 1H MRSI from five normal subjects on two occasions, separated by 10 days. Reproducibility of the MR signal was evaluated in the context of automated shimming, automated processing, and accurate subject repositioning. Reliability measures for physicochemical indices (choline moieties, creatine, N-acetylaspartate, and myo-inositol) were moderately concordant across repeat studies. Gain variation and repositioning results were excellent. It has been concluded that 1H MRSI reproducibility is adequate for serial studies of brain metabolism.

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.

To enable simultaneous high-resolution mapping of brain function and metabolism. An encoding scheme was designed for interleaved acquisition of functional MRI (fMRI) data in echo volume imaging trajectories and MR spectroscopic imaging (MRSI) data in echo-planar spectroscopic imaging trajectories. The scheme eliminates water and lipid suppression and utilizes free induction decay signals to encode both functional and metabolic information with ultrashort TE, short TR, and sparse sampling of -space. A subspace-based image reconstruction method was introduced for processing both the fMRI and MRSI data. The complementary information in the fMRI and MRSI data sets was also utilized to improve image reconstruction in the presence of intrascan head motion, field drift, and tissue susceptibility changes. In-vivo experimental results were obtained from healthy human subjects in resting-state fMRI/MRSI experiments. In these experiments, the proposed method was able to simultaneously acquire metabolic and functional information from the brain in high resolution. For scans of 6.5 minutes, we achieved 3.0 × 3.0 × 1.8 mm3 spatial resolution for fMRI, 1.9 × 2.5 × 3.0 mm3 nominal spatial resolution for MRSI, and 1.9 × 1.9 × 1.8 mm3 nominal spatial resolution for quantitative susceptibility maps. This work demonstrates the feasibility of simultaneous high-resolution mapping of brain function and metabolism with improved spatial resolution and synergistic image reconstruction.

Proton Magnetic Resonance Spectroscopic Imaging in Newly Diagnosed Glioblastoma: Predictive Value for the Site of Postradiotherapy Relapse in a Prospective Longitudinal Study

To determine the association between H-1 magnetic resonance (MR) spectroscopic imaging and MR imaging differences in subjects with Alzheimer disease (AD) or subcortical ischemic vascular dementia (SIVD) versus control subjects and if both studies combined enable discrimination of AD from control subjects better than either study alone. Measures were obtained in nine AD, eight SIVD, and 11 control subjects with MR imaging segmentation software. Statistically significantly lower N-acetylaspartate/choline-containing metabolites (Cho) and higher Cho/creatine-containing metabolites in posterior mesial gray matter in AD versus control subjects were independent of MR imagining differences. Combined measures allowed correct classification of AD and control subjects, but none of the MR measures allowed accurate discrimination between AD and SIVD subjects. Between-group differences in tissue-type contributions to H-1 MR spectroscopic imaging voxels must be accounted for when reporting H-1 MR spectroscopic imaging data in AD, SIVD, and control subjects. Combined studies allowed more accurate discrimination between AD and control subjects than either study alone.

1H magnetic resonance spectroscopic imaging (MRSI) can improve the accuracy of target delineation for gliomas, but it lacks the anatomic resolution needed for image fusion. This paper presents a simple protocol for fusing simulation computer tomography (CT) and MRSI images for glioma intensity-modulated radiotherapy (IMRT), including a retrospective study of 12 patients. Each patient first underwent whole-brain axial fluid-attenuated-inversion-recovery (FLAIR) MRI (3 mm slice thickness, no spacing), followed by three-dimensional (3D) MRSI measurements (TE/TR: 144/1000 ms) of a user-specified volume encompassing the extent of the tumor. The nominal voxel size of MRSI ranged from 8 x 8 x 10 mm3 to 12 x 12 x 10 mm3. A system was developed to grade the tumor using the choline-to-creatine (Cho/Cr) ratios from each MRSI voxel. The merged MRSI images were then generated by replacing the Cho/Cr value of each MRSI voxel with intensities according to the Cho/Cr grades, and resampling the poorer-resolution Cho/Cr map into the higher-resolution FLAIR image space. The FUNCTOOL processing software was also used to create the screen-dumped MRSI images in which these data were overlaid with each FLAIR MRI image. The screen-dumped MRSI images were manually translated and fused with the FLAIR MRI images. Since the merged MRSI images were intrinsically fused with the FLAIR MRI images, they were also registered with the screen-dumped MRSI images. The position of the MRSI volume on the merged MRSI images was compared with that of the screen-dumped MRSI images and was shifted until agreement was within a predetermined tolerance. Three clinical target volumes (CTVs) were then contoured on the FLAIR MRI images corresponding to the Cho/Cr grades. Finally, the FLAIR MRI images were fused with the simulation CT images using a mutual-information algorithm, yielding an IMRT plan that simultaneously delivers three different dose levels to the three CTVs. The image-fusion protocol was tested on 12 (six high-grade and six low-grade) glioma patients. The average agreement of the MRSI volume position on the screen-dumped MRSI images and the merged MRSI images was 0.29 mm with a standard deviation of 0.07 mm. Of all the voxels with Cho/Cr grade one or above, the distribution of Cho/Cr grade was found to correlate with the glioma grade from pathologic finding and is consistent with literature results indicating Cho/Cr elevation as a marker for malignancy. In conclusion, an image-fusion protocol was developed that successfully incorporates MRSI information into the IMRT treatment plan for glioma.

Purpose: To demonstrate the feasibility of high spatial resolution 1H magnetic resonance spectroscopic imaging (MRSI) of human breast cancer in a clinical setting at 3T. Method and Materials: The pulse sequence consists: outer volume pre‐saturation (OVP), independent CHESS pulses for water and lipid suppression, and standard PRESS pulse sequence with an elliptical weighted k‐space sampling scheme. Ten patient studies were performed in a 3.0 Tesla whole‐body clinical imager (Siemens Trio). A four‐channel breast coil (two for each breast) was used for both MRI and MRS. Patients lay prone with their breasts in the coil wells; when possible, gentle breast compression was applied to minimize motion. After global shimming, clinical DCE‐MRI was performed which defined the volume of interest for MRSI. Data processing consisted of line broadening, standard fast Fourier transformations, and phasing. Results: All ten studies were technically successful. The spectra were acquired with FOV = 120×120mm2; TR of 1500ms, TE of 80ms, slice thickness of 14mm, and 24×24 elliptical weighted k‐space sampling with 2 averages for total acquisition of 11.9minutes. All spectra are good in spectral quality, and Cho signals are clearly visible in the MRI‐lesion area, consistent with malignancy, while there was no detectable Cho in the control area. Interestingly, the distribution of Cho signal was nonuniform across the MRI lesion. Conclusion: To our knowledge, this is the first 2D/3D MRSI study of human breast cancer with short TE (less than 135ms) at 3T and highest spatial resolution (up to 0.25cm3) to date. In summary, we have presented a robust technique for high spatial resolution in vivo 1H MRSI of human breast cancer that uses the combined advantages of high field, short TE, multi‐voxel, and high spatial resolution itself to overcome shimming and lipid suppression difficulty, and demonstrated its potential for routine clinical examination.

Chapter 19 - Dimensionality Reduction and Latent Variable Modeling

At present, manifold learning is mainly applied to dimensionality reduction. However, from viewpoint of dimensionality reduction, manifold learning algorithms are only local feature preserving algorithms. For example, Local Linear Embedding is local linear preserving, Local Tangent Space Alignment is local homeomorphic preserving and Laplacian Eigenmap is local similarity preserving. The community of dimensionality reduction is now pursuing the algorithms which can preserve both local and global features of data during dimensionality reduction. In this paper, a new algorithm of dimensionality reduction, called Hilbert-Schmidt Independence Criterion Regularized Manifold Learning (HSIC-ML for short), is proposed, in which HSIC between the high dimensional data and the dimension-reduced data is added as a regularization term to the objective functions of manifold learning. The addition of HSIC regularization term makes HSIC-ML capable of preserving both local and global features during dimensionality reduction. HSIC is a criterion measuring the statistical dependence between two data sets and has been widely applied to machine learning in recent years. However, since HSIC was first proposed around 2005, there seems to have not been applied directly to dimensionality reduction, not applied as a regularization term either. The proposed HSIC-ML may be the first try in this respect. The experimental results presented in this paper show that the manifold learning with HSIC regularization performs better than that without HSIC regularization.

Locally linear embedding (LLE) is an unsupervised nonlinear manifold learning algorithm. It performs well in visualizing data yet has a very poor recognition rate in facial expression recognition. In this paper, to improve the performance of LLE in facial expression recognition, we first employ the independent component analysis (ICA) technique to preprocess the face images such that they are represented by some independent components and some noise is filtered from them. We then propose a supervised LLE (SLLE) algorithm to learn the hidden manifold. SLLE constructs the neighborhood graphs for the data according to the Euclidean distances between them and the cluster information of them. Its embedding step is the same as that of LLE. Finally, we use a generalized regression neural network (GRNN) to learn the implicit nonlinear mapping from the ICA space to the embedded manifold. Experiments on the JAFFE database show promising results

To investigate the diagnostic value of proton magnetic resonance spectroscopic imaging (MRSI) in patients with breast lesions. Eighteen patients underwent breast MRSI and MRI at 1.5 T. Contrast-enhanced MR was used to identify the lesion, after which single-slice MRSI (TR/TE = 2000/272 msec, 10-mm slice thickness) was performed. Water, lipid, and choline (Cho) images were reconstructed from MRSI data. The area of the Cho was measured in the lesion and expressed relative to the background noise level (signal-to-noise ratio (SNR)), measured between 7.0 and 9.0 ppm. Cho SNRs were compared between benign and malignant lesions as determined by histopathology. Three cases were considered technical failures on MRSI. Of the remaining 15 cases, on histopathology, eight were classified as malignant carcinoma and seven were benign. The Cho SNR from malignant tissue was significantly elevated compared to benign tissue (6.2 +/- 2.1 vs. 2.4 +/- 0.7, P < 0.0008). MRSI measurements of Cho are feasible in the human breast, and the SNR for Cho was significantly different between benign and malignant lesions. The potential advantages of MRSI over SV spectroscopy include the ability to assess multiple lesions as well as tissue with normal MRI appearance, as well as to perhaps gauge lesion borders and infiltration into surrounding tissue.

Manifold learning is a field of study in machine learning and statistics that is closely associated with dimensionality reduction algorithmic techniques is gaining popularity these days. There are two types of manifold learning approaches: linear and nonlinear. Principal component analysis (PCA) and multidimensional scaling (MDS) are two examples of linear techniques that have long been staples in the statistician's arsenal for evaluating multivariate data. Nonlinear manifold learning, which encompasses diffusion maps, Laplacian Eigenmaps, Hessian Eigenmaps, Isomap, and local linear embedding, has seen a surge in research effort recently. A few of these methods are nonlinear extensions of linear approaches. A nearest search, the definition of distances or affinities between points (a crucial component of these methods' effectiveness), and an Eigen problem for embedding high-dimensional points into a lower dimensional space make up the algorithmic process of the majority of these techniques. The strengths and weaknesses of the new method are briefly reviewed in this article. In the field of computer graphics, we utilize a particular manifold learning method was first presented in statistics and machine learning to create a global, Spectral-based shape descriptor.