This work presents a single-step deep-learning framework for longitudinal image analysis, coined Segis-Net. To optimally exploit information available in longitudinal data, this method concurrently learns a multi-class segmentation and nonlinear registration. Segmentation and registration are modeled using a convolutional neural network and optimized simultaneously for their mutual benefit. An objective function that optimizes spatial correspondence for the segmented structures across time-points is proposed. We applied Segis-Net to the analysis of white matter tracts from N=8045 longitudinal brain MRI datasets of 3249 elderly individuals. Segis-Net approach showed a significant increase in registration accuracy, spatio-temporal segmentation consistency, and reproducibility comparing with two multistage pipelines. This also led to a significant reduction in the sample-size that would be required to achieve the same statistical power in analyzing tract-specific measures. Thus, we expect that Segis-Net can serve as a new reliable tool to support longitudinal imaging studies to investigate macro- and microstructural brain changes over time.
Objective: To identify if whole-brain structural network alterations in patients with temporal lobe epilepsy (TLE) and focal to bilateral tonic-clonic seizures (FBTCS) differ from alterations in patients without FBTCS. Methods: We dichotomized a cohort of 83 drug-resistant patients with TLE into those with and without FBTCS and compared each group to 29 healthy controls. For each subject, we used diffusion MRI to construct whole-brain structural networks. First, we measured the extent of alterations by performing FBTCS-negative (FBTCS-) versus control and FBTCS-positive (FBTCS+) versus control comparisons, thereby delineating altered sub-networks of the whole-brain structural network. Second, by standardising networks of each patient using control networks, we measured the subject-specific abnormality at every brain region in the network, thereby quantifying the spatial localisation and the amount of abnormality in every patient. Results: Both FBTCS+ and FBTCS- patient groups had altered sub-networks with reduced fractional anisotropy (FA) and increased mean diffusivity (MD) compared to controls. The altered subnetwork in FBTCS+ patients was more widespread than in FBTCS- patients (441 connections altered at t>3, p<0.001 in FBTCS+ compared to 21 connections altered at t>3, p=0.01 in FBTCS-). Significantly greater abnormalities-aggregated over the entire brain network as well as assessed at the resolution of individual brain areas-were present in FBTCS+ patients (p<0.001, d=0.82). In contrast, the fewer abnormalities present in FBTCS- patients were mainly localised to the temporal and frontal areas. Significance: The whole-brain structural network is altered to a greater and more widespread extent in patients with TLE and FBTCS. We suggest that these abnormal networks may serve as an underlying structural basis or consequence of the greater seizure spread observed in FBTCS.
Diffusion magnetic resonance imaging (dMRI) is a relatively modern technique used to study tissue microstructure in a non-invasive way. Non-Gaussian diffusion representation is related to the restricted diffusion and can provide information about the underlying tissue properties. In this paper, we analytically derive $n$-th order statistics of the signal considering a stretched-exponential representation of the diffusion. Then, we retrieve the Q-space quantitative measures such as the Return-To-the-Origin Probability (RTOP), Q-space mean square displacement (QMSD), Q-space mean fourth-order displacement (QMFD). The stretched-exponential representation enables the handling of the diffusion contributions from a higher $b$-value regime under a non-Gaussian assumption, which can be useful in diagnosing or prognosis of neurodegenerative diseases in the early stages. Numerical implementation of the method is freely available at https://github.com/TPieciak/Stretched.
The surface Boundary Element Method (BEM) is one of the most commonly employed formulations to solve the forward problem in electroencephalography, but the applicability of its classical incarnations is lamentably limited to piece-wise homogeneous media. Several head tissues, however, are strongly anisotropic due to their complex underlying micro-structure. This implies that standard boundary integral formulations oversimplify the electrical properties of the head and produce unrealistic solutions, something that drastically limits the suitability and impact of BEM technologies to brain imaging. This contribution addresses this issue by observing that the brain anisotropy in the white matter is due to the presence of neuronal wire-like structures. We then extend the well known wire integral equations used for high frequency problems to the imperfectly conducting quasi-static case and we propose a new hybrid wire/surface/volume integral equation. When applied on multimodal magnetic resonance images combined with tractography, this new approach can flexibly and realistically handle the conductivity anisotropy in any head compartment providing high level of accuracy and efficiency. The beneficial properties of the new formulation together with its impact on brain imaging is demonstrated via numerical results on both canonical and realistic case scenarios.
We propose a model-based deep learning architecture for the reconstruction of highly accelerated diffusion magnetic resonance imaging (MRI) that enables high resolution imaging. The proposed reconstruction jointly recovers all the diffusion weighted images in a single step from a joint k-q under-sampled acquisition in a parallel MRI setting. We propose the novel use of a pre-trained denoiser as a regularizer in a model-based reconstruction for the recovery of highly under-sampled data. Specifically, we designed the denoiser based on a general diffusion MRI tissue microstructure model for multi-compartmental modeling. By using a wide range of biologically plausible parameter values for the multi-compartmental microstructure model, we simulated diffusion signal that spans the entire microstructure parameter space. A neural network was trained in an unsupervised manner using an autoencoder to learn the diffusion MRI signal subspace. We employed the autoencoder in a model-based reconstruction and show that the autoencoder provides a strong denoising prior to recover the q-space signal. We show reconstruction results on a simulated brain dataset that shows high acceleration capabilities of the proposed method.
Brain activation mapping using functional MRI (fMRI) based on blood oxygenation level-dependent (BOLD) contrast has been conventionally focused on probing gray matter, the BOLD contrast in white matter having been generally disregarded. Recent results have provided evidence of the functional significance of the white matter BOLD signal, showing at the same time that its correlation structure is highly anisotropic, and related to the diffusion tensor in shape and orientation. This evidence suggests that conventional isotropic Gaussian filters are inadequate for denoising white matter fMRI data, since they are incapable of adapting to the complex anisotropic domain of white matter axonal connections. In this paper we explore a graph-based description of the white matter developed from diffusion MRI data, which is capable of encoding the anisotropy of the domain. Based on this representation we design localized spatial filters that adapt to white matter structure by leveraging graph signal processing principles. The performance of the proposed filtering technique is evaluated on semi-synthetic data, where it shows potential for greater sensitivity and specificity in white matter activation mapping, compared to isotropic filtering.
We evaluate a new approach for achieving diffusion MRI data with high spatial resolution, large volume coverage, and fast acquisition speed. A recent method called gSlider-SMS enables whole-brain sub-millimeter diffusion MRI with high signal-to-noise ratio (SNR) efficiency. However, despite the efficient acquisition, the resulting images can still suffer from low SNR due to the small size of the imaging voxels. This work proposes to mitigate the SNR problem by combining gSlider-SMS with a regularized SNR-enhancing reconstruction approach. Illustrative results show that, from gSlider-SMS data acquired over a span of only 25 minutes on a 3T scanner, the proposed method is able to produce 71 MRI images (64 diffusion encoding orientations with $b=$1500 s/mm$^2$, and 7 images without diffusion weighting) of the entire \emph{in vivo} human brain with nominal 0.66 mm spatial resolution. Using data acquired from 75 minutes of acquisition as a gold standard reference, we demonstrate that the proposed SNR-ehancement procedure leads to substantial improvements in estimated diffusion parameters compared to conventional gSlider reconstruction. Results also demonstrate that the proposed method has advantages relative to denoising methods based on low-rank matrix modeling. A theoretical analysis of the trade-off between spatial resolution and SNR suggests that the proposed approach has high efficiency. The combination of gSlider-SMS with advanced regularized reconstruction enables high-resolution quantitative diffusion MRI from a relatively fast acquisition.
Developmental dyslexia is characterized by persistent reading and spelling deficits. Partly due to technical challenges with investigating subcortical sensory structures, current research on dyslexia in humans by-and-large focuses on the cerebral cortex. These studies found that dyslexia is typically associated with functional and structural alterations of a distributed left-hemispheric cerebral cortex network. However, findings from animal models and post-mortem studies in humans suggest that developmental dyslexia might also be associated with structural alterations in subcortical sensory pathways. Whether these alterations also exist in developmental dyslexia in-vivo and how they relate to dyslexia symptoms is currently unknown. Here we used ultra-high resolution structural magnetic resonance imaging (MRI), diffusion MRI and probabilistic tractography to investigate the structural connections of the visual sensory pathway in dyslexia in-vivo. We discovered that individuals with developmental dyslexia have reduced structural connections in the direct pathway between the left visual thalamus (LGN) and left middle temporal area V5/MT, but not between the left LGN and left primary visual cortex (V1). In addition, left V5/MT-LGN connectivity strength correlated with rapid naming abilities - a key deficit in dyslexia [14]. These findings provide the first evidence of specific structural alterations in the connections between the sensory thalamus and cortex in developmental dyslexia. The results challenge current standard models and provide novel evidence for the importance of cortico-thalamic interactions in explaining dyslexia.
The increasing quantity and quality of the publicly available human cerebral diffusion MRI data make possible the study of the brain as it was unimaginable before. The Consensus Connectome Dynamics (CCD) is a remarkable phenomenon that was discovered by continuously decreasing the minimum confidence-parameter at the graphical interface of the Budapest Reference Connectome Server (\url{http://connectome.pitgroup.org}). The Budapest Reference Connectome Server depicts the cerebral connections of $n=418$ subjects with a frequency-parameter $k$: For any $k=1,2,...,n$ one can view the graph of the edges that are present in at least $k$ connectomes. If parameter $k$ is decreased one-by-one from $k=n$ through $k=1$ then more and more edges appear in the graph, since the inclusion condition is relaxed. The surprising observation is that the appearance of the edges is far from random: it resembles a growing, complex structure, like a tree or a shrub (visualized on \url{https://www.youtube.com/watch?v=yxlyudPaVUE}). Here we examine the robustness of the CCD phenomenon, and we show that it is almost independent of the particular choice of the set of underlying individual connectomes, yielding the CCD phenomenon. This result shows that the CCD phenomenon is very likely a biological property of the human brain and not just a property of the data sets examined. We also present a simulation that well-describes the growth of the CCD structure: in our random graph model a doubly-preferential attachment distribution is found to mimic the CCD: a new edge appear with a probability proportional to the sum of the degrees of the endpoints of the new edge.
Human brain structural networks contain sets of centrally embedded hub regions that enable efficient information communication. However, it remains largely unknown about categories of structural brain hubs and their microstructural, functional and cognitive characteristics as well as contributions to individual identification. Here, we employed three multi-modal imaging data sets with structural MRI, diffusion MRI and resting-state functional MRI to construct individual structural brain networks, identify brain hubs based on eight commonly used graph-nodal metrics, and perform comprehensive validation analysis. We found three categories of structural hubs in the brain networks, namely, aggregated, distributed and connector hubs. Spatially, these distinct categories of hubs were primarily located in the default-mode system and additionally in the visual and limbic systems for aggregated hubs, in the frontoparietal system for distributed hubs, and in the sensorimotor and ventral attention systems for connector hubs. Importantly, these three categories of hubs exhibited various distinct characteristics, with the highest level of microstructural organization in the aggregated hubs, the largest wiring cost and topological vulnerability in the distributed hubs, and the highest functional associations and cognitive flexibility in the connector hubs, although they behaved better regarding these characteristics compared to non-hubs. Finally, all three categories of hub indices displayed high across-session spatial similarities and acted as a structural fingerprint with high predictive rates (100%, 100% and 84.2%) for individual identification. Collectively, our findings highlighted three categories of brain hubs with differential microstructural, functional and cognitive associations, which may shed light on the topological mechanisms of the human connectome.
Neural network oscillations are a fundamental mechanism for cognition, perception and consciousness. Consequently, perturbations of network activity play an important role in the pathophysiology of brain disorders. When structural information from non-invasive brain imaging is merged with mathematical modeling, then generative brain network models constitute in-silico platforms for the exploration of causal mechanisms of brain function and clinical hypothesis testing. We here demonstrate along the example of drug-resistant epilepsy that patient-specific virtual brain models derived from diffusion MRI have sufficient predictive power to improve diagnosis and surgery outcome. In partial epilepsy, seizures originate in a local network, the so-called epileptogenic zone, before recruiting other close or distant brain regions. We create personalized large-scale brain networks for 15 patients and simulate the individual seizure propagation patterns. Model validation is performed against the presurgical stereotactic electroencephalography (SEEG) data and the standard-of-care clinical evaluation. We demonstrate that the individual brain models account for the temporal variability in patient seizure propagation patterns and explain the variability in postsurgical success. Our results show that individual variations in structural connectivity, when linked with mathematical dynamic models, have the capacity to explain changes in spatiotemporal organization of brain dynamics as observed in network-based brain disorders, thus opening up avenues towards discovery of novel clinical interventions.
The average human brain volume of the males is larger than that of the females. Several MRI voxel-based morphometry studies show that the gray matter/white matter ratio is larger in females. Here we have analyzed the recent public release of the Human Connectome Project, and by using the diffusion MRI data of 511 subjects (209 men and 302 women), we have found that the relative volumes of numerous subcortical areas and the gray matter of most cortical areas are significantly larger in women than in men. Additionally, we have discovered differences of the strengths of the sexual correlations between the same structures in different hemispheres.
Graph theory in the last two decades penetrated sociology, molecular biology, genetics, chemistry, computer engineering, and numerous other fields of science. One of the more recent areas of its applications is the study of the connections of the human brain. By the development of diffusion magnetic resonance imaging (diffusion MRI), it is possible today to map the connections between the 1-1.5 cm$^2$ regions of the gray matter of the human brain. These connections can be viewed as a graph: the vertices are the anatomically identified regions of the gray matter, and two vertices are connected by an edge if the diffusion MRI-based workflow finds neuronal fiber tracts between these areas. This way we can compute 1015-vertex graphs with tens of thousands of edges. In a previous work, we have analyzed the male and female braingraphs graph-theoretically, and we have found statistically significant differences in numerous parameters between the sexes: the female braingraphs are better expanders, have more edges, larger bipartition widths, and larger vertex cover than the braingraphs of the male subjects. Our previous study has applied the data of 96 subjects; here we present a much larger study of 426 subjects. Our data source is an NIH-founded project, the "Human Connectome Project (HCP)" public data release. As a service to the community, we have also made all of the braingraphs computed by us from the HCP data publicly available at the \url{http://braingraph.org} for independent validation and further investigations.
We have developed a novel method to describe human white matter anatomy using an approach that is both intuitive and simple to use, and which automatically extracts white matter tracts from diffusion MRI volumes. Further, our method simplifies the quantification and statistical analysis of white matter tracts on large diffusion MRI databases. This work reflects the careful syntactical definition of major white matter fiber tracts in the human brain based on a neuroanatomist's expert knowledge. The framework is based on a novel query language with a near-to-English textual syntax. This query language makes it possible to construct a dictionary of anatomical definitions that describe white matter tracts. The definitions include adjacent gray and white matter regions, and rules for spatial relations. This novel method makes it possible to automatically label white matter anatomy across subjects. After describing this method, we provide an example of its implementation where we encode anatomical knowledge in human white matter for 10 association and 15 projection tracts per hemisphere, along with 7 commissural tracts. Importantly, this novel method is comparable in accuracy to manual labeling. Finally, we present results applying this method to create a white matter atlas from 77 healthy subjects, and we use this atlas in a small proof-of-concept study to detect changes in association tracts that characterize schizophrenia.
The human connectome is the object of an intensive research today. In these graphs, the vertices correspond to the small areas of the gray matter, and two vertices are connected by an edge, if a diffusion-MRI based workflow finds connections between those areas. One main question of the field is discovering the directions of the edges. In a previous work we have reported the construction of the Budapest Reference Connectome Server http://connectome.pitgroup.org from the data recorded in the Human Connectome Project of the NIH. After the server had been published, we recognized a surprising and unforeseen property of it: The server can generate the braingraph of connections that are present in at least $k$ graphs out of the 418, for any value of $k=1,2,...,418$. When the value of $k$ is changed from $k=418$ through 1 by moving a slider at the webserver from right to left, more and more edges appear in the consensus graph. The astonishing observation is that the appearance of the new edges is not random: it is similar to a growing tree. We hypothesize that this movement of the slider in the webserver may copy the development of the connections in the human brain in the following sense: the connections that are present in all subjects are the oldest ones, and those that are present in a decreasing fraction of subjects are gradually the newer connections in the individual brain development. An animation on the phenomenon is available at https://youtu.be/EnWwIf_HNjw. Based on this hypothesis, we can assign directions to the edges of the connectome as follows: Let $G_i$ denote the consensus connectome where each edge is present in at least $i$ graphs. Suppose that vertex $v$ is isolated in $G_{k+1}$, and becomes connected to a vertex $u$ in $G_k$, where $u$ was connected to other vertices already in $G_{k+1}$. Then we direct this $(v,u)$ edge from $v$ to $u$.
Deep graph-theoretic ideas in the context with the graph of the World Wide Web led to the definition of Google's PageRank and the subsequent rise of the most-popular search engine to date. Brain graphs, or connectomes, are being widely explored today. We believe that non-trivial graph theoretic concepts, similarly as it happened in the case of the World Wide web, will lead to discoveries enlightening the structural and also the functional details of the animal and human brains. When scientists examine large networks of tens or hundreds of millions of vertices, only fast algorithms can be applied because of the size constraints. In the case of diffusion MRI-based structural human brain imaging, the effective vertex number of the connectomes, or brain graphs derived from the data is on the scale of several hundred today. That size facilitates applying strict mathematical graph algorithms even for some hard-to-compute (or NP-hard) quantities like vertex cover or balanced minimum cut. In the present work we have examined brain graphs, computed from the data of the Human Connectome Project, recorded from male and female subjects between ages 22 and 35. Significant differences were found between the male and female structural brain graphs: we show that the average female connectome has more edges, is a better expander graph, has larger minimal bisection width, and has more spanning trees than the average male connectome. Since the average female brain weights less than the brain of males, these properties show that the female brain is more "well-connected" or perhaps, more "efficient" in a sense than the brain of males.
Models of diffusion MRI within a voxel are useful for making inferences about the properties of the tissue and inferring fiber orientation distribution used by tractography algorithms. A useful model must fit the data accurately. However, evaluations of model-accuracy of some of the models that are commonly used in analyzing human white matter have not been published before. Here, we evaluate model-accuracy of the two main classes of diffusion MRI models. The diffusion tensor model (DTM) summarizes diffusion as a 3-dimensional Gaussian distribution. Sparse fascicle models (SFM) summarize the signal as a linear sum of signals originating from a collection of fascicles oriented in different directions. We use cross-validation to assess model-accuracy at different gradient amplitudes (b-values) throughout the white matter. Specifically, we fit each model to all the white matter voxels in one data set and then use the model to predict a second, independent data set. This is the first evaluation of model-accuracy of these models. In most of the white matter the DTM predicts the data more accurately than test-retest reliability; SFM model-accuracy is higher than test-retest reliability and also higher than the DTM, particularly for measurements with (a) a b-value above 1000 in locations containing fiber crossings, and (b) in the regions of the brain surrounding the optic radiations. The SFM also has better parameter-validity: it more accurately estimates the fiber orientation distribution function (fODF) in each voxel, which is useful for fiber tracking.
Diffusion-weighted magnetic resonance imaging (DWI) is the only noninvasive method for quantifying microstructure and reconstructing white-matter pathways in the living human brain. Fluctuations from multiple sources create significant additive noise in DWI data which must be suppressed before subsequent microstructure analysis. We introduce a self-supervised learning method for denoising DWI data, Patch2Self, which uses the entire volume to learn a full-rank locally linear denoiser for that volume. By taking advantage of the oversampled q-space of DWI data, Patch2Self can separate structure from noise without requiring an explicit model for either. We demonstrate the effectiveness of Patch2Self via quantitative and qualitative improvements in microstructure modeling, tracking (via fiber bundle coherency) and model estimation relative to other unsupervised methods on real and simulated data.
Analysis of longitudinal changes in imaging studies often involves both segmentation of structures of interest and registration of multiple timeframes. The accuracy of such analysis could benefit from a tailored framework that jointly optimizes both tasks to fully exploit the information available in the longitudinal data. Most learning-based registration algorithms, including joint optimization approaches, currently suffer from bias due to selection of a fixed reference frame and only support pairwise transformations. We here propose an analytical framework based on an unbiased learning strategy for group-wise registration that simultaneously registers images to the mean space of a group to obtain consistent segmentations. We evaluate the proposed method on longitudinal analysis of a white matter tract in a brain MRI dataset with 2-3 time-points for 3249 individuals, i.e., 8045 images in total. The reproducibility of the method is evaluated on test-retest data from 97 individuals. The results confirm that the implicit reference image is an average of the input image. In addition, the proposed framework leads to consistent segmentations and significantly lower processing bias than that of a pair-wise fixed-reference approach. This processing bias is even smaller than those obtained when translating segmentations by only one voxel, which can be attributed to subtle numerical instabilities and interpolation. Therefore, we postulate that the proposed mean-space learning strategy could be widely applied to learning-based registration tasks. In addition, this group-wise framework introduces a novel way for learning-based longitudinal studies by direct construction of an unbiased within-subject template and allowing reliable and efficient analysis of spatio-temporal imaging biomarkers.
Current brain white matter fiber tracking techniques show a number of problems, including: generating large proportions of streamlines that do not accurately describe the underlying anatomy; extracting streamlines that are not supported by the underlying diffusion signal; and under-representing some fiber populations, among others. In this paper, we describe a novel autoencoder-based learning method to filter streamlines from diffusion MRI tractography, and hence, to obtain more reliable tractograms. Our method, dubbed FINTA (Filtering in Tractography using Autoencoders) uses raw, unlabeled tractograms to train the autoencoder, and to learn a robust representation of brain streamlines. Such an embedding is then used to filter undesired streamline samples using a nearest neighbor algorithm. Our experiments on both synthetic and in vivo human brain diffusion MRI tractography data obtain accuracy scores exceeding the 90\% threshold on the test set. Results reveal that FINTA has a superior filtering performance compared to conventional, anatomy-based methods, and the RecoBundles state-of-the-art method. Additionally, we demonstrate that FINTA can be applied to partial tractograms without requiring changes to the framework. We also show that the proposed method generalizes well across different tracking methods and datasets, and shortens significantly the computation time for large (>1 M streamlines) tractograms. Together, this work brings forward a new deep learning framework in tractography based on autoencoders, which offers a flexible and powerful method for white matter filtering and bundling that could enhance tractometry and connectivity analyses.