Image Processing
NiChart Image Processing Pipelines is a set of tools that can be used to extract features from multi-modal MRI data. The pipelines are made up of individual components that can be installed and run independently. Users can choose the specific components that they need for their analyses. NiChart pipelines are available both as installation packages and software containers. This ensures reproducibility across different platforms, and allows users to easily install and run the pipelines without having to worry about installing any additional software dependencies.
NiChart uses a combination of both established and state of the art techniques to extract imaging features that quantify both normal and abnormal brain structures. Alongside conventional atlas-based segmentation methods for segmenting intra-cranial volume (ICV), anatomical regions of interest (ROIs), and white matter lesions (WMLs), we also offer an alternative parcellation method using non-negative matrix factorization for generating multi-resolution data-driven structural covariance components. Our atlas-based segmentation methods are facilitated by deep learning networks that enable rapid segmentation.
Deep Learning Intra Cranial Volume (DLICV) is a new deep learning (DL)-based tool to accurately segment the intracranial volume (ICV) from a raw T1-weighted MRI image. It's easy to use, requires minimal data preprocessing, and is robust against various factors that can affect segmentation accuracy. DLICV specifically segments the overall cerebrospinal fluid (CSF) surrounding the brain, rather than just the brain tissues themselves, providing an ICV estimation that is not influenced by overall cortical atrophy due to aging or disease.
Figure 1. Example segmentation using DLICV (green) for cases with significant cortical atrophy
Deep Learning MUlti-atlas region Segmentation utilizing Ensembles of registration algorithms and parameters (DLMUSE) is a tool for automatically segmenting T1-weighted brain MRI scans. It is accurate, robust, easy to use, and fast. DLMUSE is built on a 3D convolutional neural network (CNN) architecture that has been extensively validated for various neuroimaging segmentation tasks. DLMUSE model was trained on a large and diverse training set, with ROI labels derived using a computationally intensive multi-atlas segmentation method.
Deep Learning White Matter Lesion Segmentation (DLWMLS) is a multi-modal segmentation method for segmenting white mater hyper-intensities (brain lesions) from T1-weighted and FLAIR MRI images. DLWMLS model was trained on a large and diverse training set, with semi-automatically segmented labels for lesions.
Figure 2. Example segmentatation using DLICV, DLMUSE and DLWMLS
Stochastic Orthogonally Projective Non-negative Matrix Factorization (sopNMF) is an algorithm for large-scale multivariate structural analysis of human brain data. Using sopNMF, the MuSIC atlas parcellates the human brain by structural covariance in MRI data over the lifespan and a wide range of disease populations, allowing to explore the phenotypic landscape and genetic architecture of the human brain. You can find out more in the Github repository of the python implementation.
Figure 3. Multi-resolution MuSIC atlas parcellation
QSIPrep is a specialized software platform designed for the preprocessing of diffusion MRI datasets, ensuring the deployment of adequate workflows for the task. It primarily focuses on diffusion-weighted magnetic resonance imaging (dMRI), a pivotal method for non-invasively examining the organization of white matter in the human brain. QSIPrep stands out for its integrative nature, being compatible with nearly all dMRI sampling schemes, thus providing a broad spectrum of utility in diffusion image processing.
The platform employs an automated approach, configuring pipelines for processing dMRI data. It adheres to a BIDS-app methodology for preprocessing, which encompasses a variety of modern diffusion MRI data types. The preprocessing pipelines generated by QSIPrep are automatic, accurately grouping, distortion correcting, motion correcting, denoising, and coregistering the data, among other operations, to ensure the integrity and quality of the processed images.
Figure 4. QSIPrep flowchart
Functional MRI processing combines well-established and extensively validated tools for image preprocessing, feature extraction, and calculation of functional networks.
fMRIPrep is a robust preprocessing pipeline tailored for the analysis of functional Magnetic Resonance Imaging (fMRI) data. The pipeline leverages a combination of well-regarded software packages including FSL, ANTs, FreeSurfer, and AFNI to ensure optimal software implementation for each preprocessing stage. Designed to minimize manual intervention, fMRIPrep facilitates a transparent workflow that enhances the reproducibility of fMRI data analyses. It is suited for handling both task-based and resting-state fMRI data, adapting to the nuances of different datasets to provide high-quality preprocessing without requiring manual intervention.
fMRIPrep is a NiPreps (NeuroImaging PREProcessing toolS) application for the preprocessing of task-based and resting-state functional MRI (fMRI).
Figure 5. fMRIPrep flowchart
The XCPEngine, or XCP imaging pipeline, is an open-source software package engineered for processing multimodal neuroimages. Utilizing a modular design, it integrates analytic routines from leading MRI analysis platforms like FSL, AFNI, and ANTs. This engine offers a configurable, modular, and agnostic platform for neuroimage processing and quality assessment, encapsulating a variety of high-performance denoising approaches while computing regional and voxelwise values for each modality.
Figure 6. XCPEngine flowchart
Personalized Functional Network Modeling (pNet) is designed to provide a user-friendly interface to perform personalized functional network (pFN) computation and visualization.
It is open-source, cross-platform, and expandable. The toolbox is built with support for MATLAB and Python users. The MATLAB version offers GUI and code scripts. The Python version uses NumPy for simple code development, and PyTorch for high computation performance. And it provides a step-by-step guide in terminal command. pNet provides streamlined workflow to carry out computation and visualization of pFNs.
It also integrates several statistical methods to investigate the relationship between pFNs and behavior data. In addition, quality control is available to check the quality of pFN modeling results. This toolbox can be downloaded from YuncongMa/pNet and MLDataAnalytics/pNet.
Figure 7. pNet network model