Copyright 2021 - CDS-QuaMRI consortium

Prev: Project Objectives                                                         Next: Open Source Packages

Project Results

Algorithm development was a predominant part of the CDSQuaMRI project and yielded a number of advanced methods and tools for the analysis of various MR imaging modalities. These methods were implemented as individual stand-alone software packages which in a second step were integrated in processing workflows in the framework platform.

 

Anatomical MRI: myelin & relaxation

Methods have been developed that enable the extraction of quantitative information about tissue micro-structure from qualitative anatomical brain MRI scans. These methods were initially based on regression techniques. More recently, methods based on artificial intelligence have been developed. Results demonstrated that whenever specialized quantitative protocols are not available, approximate myelin and relaxation indices can be obtained from qualitative scans (F These findings demonstrate that the CDS-QuaMRI clinical decision system could be used for the retrospective analysis of large existing data bases of clinical scans, providing important insight on long-term changes to brain micro-structure in neurological disorders.

fig2
Figure 1: Examples of synthetic Magnetization Transfer Ratio (MTR, i.e. a myelin-sensitive metric) produced via U-Nets (U-MTR) from qualitative MRI scans that are routinely performed in radiology departments. The images in the top-left, top-right and bottom-left corner show qualitative images, ground-truth MTR and U-MTR results for one control, primary progressive multiple sclerosis (PPMS) and secondary progressive multiple sclerosis (SPMS) test-subjects. U-MTR appears very similar to the ground-truth, with the exception of the noise characterizing MTR maps. In the bottom-right corner, one relapsing remitting multiple sclerosis (RRMS) subject
is shown, which we excluded from the cohort due to the corrupted ground-truth; U-MTR produced from the qualitative maps, however, lacks such artifacts. This subject is an example of the usefulness of this method when MT-data is corrupted by noise or otherwise missing or unusable.


Microstructural MRI: diffusion


Several novel methods for microstructural MRI have been developed throughout the course of the project:
1) Microscopic susceptibility anisotropy mapping: a novel technique that tears apart the two principal effects conflated in gradient-echo measurements, (a) the susceptibility properties of tissue micro-environments, especially the myelin microstructure, and (b) the axon orientation distribution relative to the magnetic field.
2) Neural Soma Imaging: To capture the heterogeneous morphology of grey matter, it is imperative to disentangle cylindrical and spherical geometries commonly attributed to neurites and neural soma, respectively, but also to regress out orientation heterogeneity in nervous tissue. This is achieved by leveraging the latest advances in B-tensor diffusion encoding and deep-learning techniques and present microstructural feature maps of neurites and neural soma in-vivo in the human brain
3) Spherical Mean Technique (SMT): this is a clinically feasible method for microscopic diffusion anisotropy imaging. The purpose is to map microscopic features unconfounded by the effects of fiber crossings and orientation dispersion, which are ubiquitous in the brain. SMT comes in two flavors, using a microscopic tensor model and a multi-compartment microscopic model. The former model maps the microscopic diffusion coefficients parallel and perpendicular to the axons and then derives the microscopic fractional anisotropy, which were only available with clinically prohibitive scans thus far. The second SMT model provides estimates of microscopic features specific to the intra- and extra-neurite compartments, such as an index of neurite density.
fig1

 
Figure 2: Neurite Density and Orientation Dispersion Index (NODDI) and Spherical Mean Technique (SMT) metrics as obtained from an MS patient. Top: NODDI (left: intra-neurite volume fraction vin; orientation dispersion index ODI; isotropic volume fraction viso); bottom: SMT (intra-neurite volume fraction vin; orientation dispersion entropy ODE; neural diffusivity D).

 

 

 

 

 

 

 

 

 

Functional MRI

Methods for data analysis and feature extraction from resting-state and task-based fMRI were developed:
1) A new method for comprehensive data visualization called “Brainglance”. The method allows all analysis steps to be carried out in the native subject space, thus supporting “single subject analysis”. The results are summarized on the group level, allowing to assess the heterogeneity within a group without destroying information about individual subjects. In a clinical context, it is vital to preserve such individual information.

fig3
Fig 3: Illustration of the brain glance visualization. Each colored square corresponds to a single brain area from a single subject. (A) Subjects are shown in rows, here we show four subjects. (B) Brain areas are shown in columns. Here, six brain regions within the insular cortex are visualized. (C) The color corresponds to the measured value of interest, this could for instance be the strength of an activation. The color mapping is customizable, here we show positive values in orange and negative values in blue. (D) This figure only visualizes illustrative data from the insular cortex.

 

 

 

2) A new method called “Local Indicators of Spatial Association (LISA)” for statistical inference of fMRI data. LISA is a method to statistically detect local activation in the human brain, using a non-parametric and threshold-free framework. The key component is a non-linear filter for incorporating spatial context without sacrificing spatial precision. Multiple comparison correction is achieved by controlling the false discovery rate in the filtered maps. Compared to widely used other methods, LISA shows a boost in statistical power and allows to find small activation areas that have previously evaded detection, which then can be used as features for further analysis.
fig4

Figure 4: Statistical inference with the newly developed LISA method has a much higher reproducibility as compared to standard methods such as SPM, AFNI and FSL. The reproducibility across 100 tests was based on randomly drawn samples of size 20 from the human connectome project motor data. The colors represent reproducibility scores, i.e., the number of tests in which a given voxel consistently passed the significance threshold.


 

 

 

 

 

 

3) A new method called “Bipartite connectivity mapping (BCM)” for analyzing connectivity in resting state fMRI between two brain regions. The main idea is to represent connectivity between the regions as a bipartite graph and analyse it using bipartite network projections. The advantage of this approach is that it allows to fully preserve spatial precision in both ROIs. This sets it apart from traditional seed-based connectivity mapping where the seed ROI must be averaged so that its spatial information is lost.


4) Two major improvements to a previously published method called "Eigenvector centrality mapping" (ECM). The improved version brings a significant improvement in computational speed and allows to apply ECM to ultrahigh resolution images.

5) A new method for revealing task-induced edge density (TED). TED considers transient networks in response to external stimuli. A major advantage compared to other methods is that it does not depend on any specific hemodynamic response model. It also does not require a pre-segmentation of the data for dimensionality reduction as it can handle large networks consisting of tens of thousands of voxels.

 

Brain Perfusion: arterial spin labeling

A novel arterial spin labeling /(ASL) acquisition technique was developed. It interleaves the standard ASL acquisition sequence with a measurement of labeling efficiency which occurs during ‘dead space’ naturally present in ASL acquisitions. With this approach, the labeling efficiency can be measured without any time-penalty. The technique was validated and the importance of labelling efficiency was proven in an subject study, in which sub-optimal labeling efficiency resulted in underestimation of cerebral perfusion within a single flow territory.
A correction scheme for motion artefacts in ASL images acquired with the Multiband technique was developed and used to further increase the robustness of the ASL analysis.

ASL – cortical architecture and network analysis
Partial volume correction algorithms are essential for discriminating between perfusion changes and atrophy (loss of brain tissue). This is especially important in the context of the CDS-QUAMRI project focusing on a clinical support system for personalized treatment.
Determination of main flow territories from traditional ASL scans based on the assumption that each flow territory will exhibit unique signal fluctuations was studied. It was demonstrated that natural occurring fluctuations are not large enough to identify the main flow territories, but that by inducing tiny additional fluctuations into the ASL-signal, joint estimation of perfusion images and the main flow territories becomes feasible.


Metabolic MRI: proton and non-proton magnetic resonance spectroscopy

MRSI reconstruction of under-sampled data
Several novel magnetic resonance spectroscopic imaging reconstruction methods have been developed for accelerated undersampled 1H MRSI: (i) overdiscrete static magnetic field inhomogeneity (B0) correction and sensitivity encoding (SENSE) reconstruction; (ii) neural-network based GeneRalized Autocalibrating Partial Parallel Acquisition (GRAPPA) reconstruction and (iii) a hybrid SENSE and compressed sensing reconstruction. The neural network reconstruction is superior to both classical parallel imaging and compressed sensing reconstruction. It can be combined with overdiscrete SENSE for optimal SNR and enhanced spatial resolution.

fig5

Figure 5: Comparison between metabolite maps for 4 major brain metabolites derived from fully k-space sampled data (left) and four fold k-space under sampled data with conventional GRAPPA (mid) and neuronal network based reconstruction (right). The neural network reconstruction removes the lipid artefacts and yields metabolite maps that correspond to the non-accelerated MRSI data.

 

 

 

 

 

 

 

 

 

 

 

MRSI quantification

A quantification pipeline correcting MRSI spectral fitting results for relaxation times of metabolites and tissue water was implemented. To that segmentation of high-resolution anatomical images into tissue fraction maps of grey matter, white matter, and cerebrospinal fluid and recently measured T1-relaxation times of water and metabolites were utilized. The T1-relaxation time for each metabolite is corrected for each voxel from the MRSI acquisition considering its specific voxel composition by utilizing the tissue probability maps and GM and WM T1-relaxation time estimates for 9.4 T. T1-relaxation times for water in GM, WM, and CSF were taken from 9.4 T human results from Hagberg, et al. . Metabolite maps are co-registered to high resolution MP2RAGE images via rigid body transformations. Then, quantification of metabolite maps using an internal water reference is performed by utilizing the method as described by Gasparovic et al.5; hence, concentrations are reported in molal.

fig6
Figure 6: Quantitative metabolite maps for 12 brain metabolites in the human brain derived from 1H MRSI data acquired at 9.4T.

 

 

 

 

Spectral Fitting Algorithms & Spectral Models
A versatile spectral fitting algorithm – ProFit 1D - for modelling of 1D spectra was developed. It allows for integration of spectral models of all detectable metabolites and macro-molecules and considers imperfections with respect to line width, line shape, chemical shift offsets, phase and baseline. The prior knowledge based model spectra of all relevant metabolites is fitted to the experimental spectra in an iterative matter with increasing degrees of freedom per iteration but tightened boundaries.
A ProFit 2D spectral fitting algorithm was further developed to fit data acquired with an adiabatic J-resolved semiLASER sequence by implementing a dedicated fit model and to support the internal water reference standard.
Metabolite spectra are overlapped by broad resonance lines. This so-called macromolecular signal origins from amino acids of protein and peptide chains. In order to account for the macromolecular signal contribution to metabolite spectra during the spectral fitting procedure a relaxation corrected macromolecular model was developed. This enables automatized simulation of a macromolecular spectrum for any given sequence parameters inside the spectral fitting software ProFit 1D that was developed during the CDS-QUAMRI project (Figure 11 / left). A respective manuscript is currently under review. As input to the development of the relaxation corrected MM model T1 and T2 relaxation times of macromolecular peaks up-field and downfield of water have been measured in the human brain at 9.4 T for the first time (Murali-Manohar et al 2020 & 2021). Estimation of the relaxation times of macro-molecules not only helped develop the above-discussed relaxation corrected MM simulation model and thus improve the estimation of tissue metabolite concentrations, but also helped in quantification of macromolecule concentrations itself.

fig7
Figure 7: (left) Spectral fit of MC-semiLASER data that includes a relaxation corrected macromolecular baseline model (MMB) in the spectral basis vector. (right) A sample downfield spectrum (Data+Fit) with all quantified metabolites/peaks. The pH dependence of the chemical shift of homocarnosine (hCs) was investigated (green arrows). The red arrow points to the means of measuring the line width of each peak. From the quantified properties, several peaks could be assigned to metabolites, with the most important finding shown by the purple arrows.

 

 

 

 

 

 

 

 

 

Retrospective Classification by Machine Learning

One of the project goals was to develop a retrospective multi-parametric classification and clustering approach based on features extracted from previously acquired multi-modal quantitative MRI data sets and machine learning. For the development and testing of clustering and classification algorithms pre-existing input patient data acquired with consistent scan protocols and same scanner vendor from cohorts of patients with major depressive disorder (MDD) and multiple sclerosis (MS) were used. All data were consistently processed and prepared for classification trials; subsequently, the implementation of different machine learning methods such as support vector machines (SVM), random forest (RF), k-Nearest Neighbors and multiple kernel learning (MKL) approaches were performed to investigate their performance in different prediction tasks. Early stage efforts demonstrated that distinguishing the patient groups from healthy volunteers was possible for both diseases, MDD and MS. In a second phase more challenging tasks were studied (differential diagnostics; treatment responders versus non-responders; patients with different clinical subtypes; subjects with different disease progression) and provided good specificity and sensitivity. Finally, the focus shifted on refining the classification and clustering methods for the more difficult prediction tasks, on optimizing the combination and selection of multi-modal approaches, and on the integration of the classification module into the software framework.


Prospective Clinical Trial and Decision Support System
The first objective was to acquire a consistent and large data set from multiple MRI modalities in MDD patients.The second objective was the development of a classification module that was to be integrated into the software framework. The implementation was accomplished based on the implemented classifier implementations. The system was then applied as a prospective clinical decision support for MDD and MS based on the most successful classifiers, on data acquired both prior and during the project run time for multiple sclerosis and major depressive disorders. It was demonstrated that anatomical, functional and metabolic MRI yield features that predict therapy response to electroconvulsive, Ketamine therapy and psychotherapy in patients with major depressive disorder. Furthermore, the correct assignment of patients to different types of multiple scleroses with distinct disease progression was possible based on a combination of anatomical and microstructural MRI. Microstructural MRI shows strong predictive power for disability scores in multiple sclerosis patients. In both patient groups it was found that anatomical imaging data that are routinely acquired yield predictive power if a quantitative analysis is performed instead of the qualitative evaluation that is the current clinical standard.

Prospective Clinical Trial in Major Depressive Disorder
Prospective multimodal QuaMRI dataset including MDD patients and healthy control participants were acquired at one partner site. Acquisition sequences included a 2D J-semiLASER magnetic resonance spectroscopy (MRS) sequence, an arterial spin labeling (ASL) sequence and a diffusion tensor imaging (DTI), all developed in the CDS-QuaMRI project. The multimodal MRI data was acquired at multiple time points (baseline, 24 h post treatment, 4 weeks post treatment, 6 months post treatment), to develop classification routines that take into account early brain changes during different antidepressant treatments and to make predictions on long-term outcomes based on these early changes. Long-term follow-up measurements will be conducted beyond the project end. The status at the end of the project duration is given in Table 1.

fig8
Table 1: Final status of the data acquisition. The numbers depict the N that has been acquired for each data set and modality.




 

 

 

 

 

 

 

 

 

 

 

 

 

 

Demonstration of Clinical Decision Support in MDD

Next to diagnostic classification tools, classifications routines for the prediction of treatment responses have been developed. Pre-existing datasets of MDD patients that received Electroconvulsive Therapy (ECT), or ketamine treatment have been processed. The next paragraphs summarize the most important findings.
The Major Depression Study includes multiple samples of patients affected by Major Depressive Disorder (MDD) and Healthy Controls (HC) subjects, acquired in different modalities (Structural, resting-state and task-based fMRI), at different time points (before and after treatment) and at different sites (Berlin and Zurich). The main focus was the development of an optimal and interpretable prediction pipeline, that is applicable for the diverse set of input data and clinical tasks. The goal was to obtain a robust and accurate predictive model, providing interpretable results while identifying the informative modalities for a specific clinical task.

Clinical task: patients vs controls
In the first phase of the project, focus lay on the development of the best classification approach, with the aim of distinguishing patients and controls. The disease diagnosis task was based on a dataset of 118 samples of task-based fMRI data. This effort resulted in the development of a multi-voxel pattern classification pipeline. The machine learning analysis employed feature selection and linear Support Vector Machines to distinguish between healthy controls and depressed patients based on whole-brain fMRI data. The post processing cluster analysis identified brain regions that mostly contributed to the classification, i.e. activation patterns in patients or controls. The results of this approach showed good classification performance, with accuracy reaching 71%. Details on the analysis and discussion of the corresponding findings can be found in the published manuscript.

Subsequently, the focus shifted to studying whether a multi-modal approach is superior over a single-modality study. For this task, resting-state and Structural MRI data from the same cohort of MDD and HC subjects was used. The results of Multiple Kernel Learning (MKL) techniques and feature concatenation strategies were compared. It was observed that a combination of contrast images obtained from task-based fMRI provide the best results with linear classification methods. When non-linear kernels are used, the single best performing modality dominates over the multi-modal counterparts. With the same method, it has also been observed that combining resting state and task-based fMRI improves over the individual modalities. Overall, there was no clear benefit in using complex MKL techniques over simple feature concatenation strategies. In general, the improvements observed with modality combinations are not consistent across methods and modalities. The outcome is certainly influenced by the limitations given by a small sample size. This analysis suggests that combining multiple modalities is promising and has the potential to improve the automated clinical decision system. Still, more data should be acquired in order to provide a final complete assessment. Enlarging the sample size would also be beneficial for model exploration, allowing the use of deep learning based strategies or transfer learning techniques.

Clinical task: response prediction
In personalized medicine, a highly relevant task is the identification of patients that positively respond to a given treatment based on their clinical characteristics. In the final stage of the CDS-QuaMRI project the aim was to assess response prediction to Electroconvulsive Therapy (ECT) treatment using MRI data. The question was investigated both from a classification and regression perspectives, while structural MRI features were used for the analysis. Initially, a whole-brain approach based on Support Vector Machines was developed to classify responders versus non responders. A good predictive performance was obtained with 26/39 responders and 23/32 non-responders being correctly identified. The post processing analysis showed a cluster in the right anterior parahippocampal gyrus (aPHCr) region which provided the most informative contribution in the characterization of ECT response. Then, in the second step the same region was used within a regression model to predict the percentage of symptom reduction (PSR). The results showed a significant correlation between predicted and true PSR, indicating that the aPHCr area contains the relevant information to identify responders. The results of this pioneer study confirm that there is predictive power in structural brain images of MDD patients to predict ECT response and that this method can be used to predict the treatment response in single patients. Future work should focus on integrating the multi-modal analysis applied for the patients versus control task with the ECT classification task, and extend it to the regression problem.


Demonstration of prospective Clinical Decision Support in MS
Multiple Sclerosis is a neurological disease of the central nervous system, whose cause and progression is still unknown. Patients can be diagnosed into three subtypes, having different disease course and severity stages: Relapsing Remitting (RR), Secondary Progressive (SP), and Primary Progressive (PP). Correct MS diagnosis, as well as subgroup identification and progression prediction, are of crucial relevance for the clinicians in order to provide the patient with personalized and efficient therapies. In addition clinically isolated syndrome (CIS) is a central nervous system demyelinating event isolated in time that is compatible with the possible future development of multiple sclerosis (MS). Early risk stratification for conversion of CIS to MS helps with treatment decisions. These tasks were all investigated. In particular a multimodal data set that included MRI data from Diffusion Weighted Images (DWI) and Magnetization Transfer Ratio (MTR) images was used. Diffusion metrics were extracted from DWI images: Fractional Anisotropy (FA); Mean Diffusivity (MD); Radial Diffusivity (RD); Axial Diffusivity (AD). Region of Interest (ROI) features were also derived. ROI measures were obtained by averaging the voxel values in predetermined brain regions, resulting in a low-dimensional brain representation.
In another data set, the extracted MRI features were: relaxometry measurements: quantitative proton density (qPD), T2 (qT2) and T1 (qT1); diffusion-derived measurements, namely intra-neurite volume fraction (intra), intrinsic diffusivity (diff) and neurite orientation dispersion entropy from the Spherical Mean Technique; cortical thickness (CT), region-of-interest (ROI) volume (vol) and quantitative total sodium concentration (TSC). The weighted mean of relaxometry, diffusion and sodium features over a probabilistic lesion map and normal appearing white matter were also added, for a total of 693 features. This cohort consisted of a total of 123 subjects: 29 HC, 18 CIS, 63 relapsing-remitting MS (RRMS) and 13 secondary progressive MS (SPMS) patients with same disease duration.

Classification task: patients vs controls
For the Multiple Sclerosis study, whole brain high-dimensional features were initially considered, resulting in poor predictive performances. Therefore, ROI features were used for the subsequent analysis. Based on single modalities, good classification results were obtained, with the Random Forest being the best performing classifier. This suggests that non-linear interactions among ROI based features are informative to predict MS.
The next step was to perform a multi-modal analysis. At this stage, it was confirmed that linear classifiers are not suitable for this task. Therefore, Multiple Kernel Learning techniques based on Gaussian Kernels as well as feature concatenation strategies with a Random Forest classifier were investigated. The results showed that a combination of features is often beneficial to improve the predictive performance. This approach in combination with Multiple Kernel Learning techniques is suitable to classify single patients into the correct peer group as needed for clinical decision support.

Classification task: disease progression
Predicting the progression of Multiple Sclerosis is known to be one of the most challenging tasks for both clinicians and data analysts. As there is no uniform consensus on the clinical assessment of disease severity itself, it is non-trivial to define exact class labels for the prediction task. On the available cohort, state-of-the-art classification models showed limited predictive performance on the disease progression task when tested to predict evolvement of patients with CIS to MS. A further investigation is recommended after a larger number of scans has been acquired, investigating the potential of complex deep learning based techniques and modern computer vision approaches. However our recent open access article in Frontiers (Johnson et al, Frontiers in neurology 2021) demonstrates that metrics from SMT (a technique for microstructure characterization developed within the CDS-QUAMRI project) correlate strongly with disability scores in MS, providing a simple non-invasive predictive marker that we plan to use widely in clinical studies and trials.

Classification task: Multiple Sclerosis subtypes
It was also investigated on the first cohort whether unsupervised algorithms were able to successfully identify clusters of patients reflecting the MS subtypes. As for the classification task, ROI features from diffusion images and MTR metrics were used as input. Various approaches were compared including DBSCAN, k-means, hierarchical clustering, and spectral clustering. When distinguishing between the three groups of MS (PP, SP, RR) a poor separation boundary was obtained. However, we observed that merging the groups of PP and SP patients helped to generally improve the results. In particular, both k-means and hierarchical clustering resulted in significant performances gains.  As a final step, the potential to apply Principal Component Analysis (PCA) and cluster the subjects based on the first two components was investigated. This analysis resulted in a good separation of two groups of MS subtypes (SP + PP; RR) and can thus be used to classify single patients into the correct group as needed for clinical decision support. Furthermore, it was observed that only a subset of the ROI features was associated with the MS phenotype, suggesting that a lower dimensional representation would be beneficial. Given the sample size limitation, the number of clusters was considered as a fixed parameter. Future work should primarily focus on acquiring a larger data cohort. A sufficiently big sample size, improved hyperparameter optimization, and features combinations should be strived forto further enhance the clustering analysis.


The above is only an overview of the work and results achieved in the CDS-QuaMRI project. Detailed methods and results description can be found in the peer-reviewed publications that resulted from the project. The list of publications is available here.