Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Mapping the structural network topology of the human brain is a fundamental challenge in science, and one that may advance our understanding and treatment of neurological and neuropsychiatric disorders. Diffusion-weighted magnetic resonance imaging (DWI) is currently the only non-invasive technique for probing the structural connectivity in the brain in vivo. Its principle is based on indirect measurement of the diffusion anisotropy of water, which is correlated with tissue microstructure. Estimating the local geometry of axonal fibres then hinges on biophysical models of diffusion in white matter. Such local estimates can subsequently be integrated along the image to reconstruct global structural connections in the brain, a process known as tractography. Many state-of-the-art DWI analysis methods build on strong model assumptions about the signal in white matter, which are hard to validate and may not generalize to other tissues and pathology. In this thesis, we therefore aim to reconstruct the local and global fibre configuration in brain white matter with as few prior assumptions about the microstructure as possible. Instead, we develop data-driven methodology, informed by spatial priors, population priors, and the signal itself. From this perspective, we develop a global tractography framework that integrates a spatial prior on the continuity of white matter fibres into a minimal convolutive multi-tissue model for DWI in the brain, based on a fibre response function that is estimated from and thus adapted to the data at hand. In this method, local fibre orientation estimates inform the global track configuration and vice versa, hence integrating local and global scales into one Markov chain Monte Carlo optimization framework. Results show improved specificity of valid connections and maintain a quantitative correspondence between track density and the apparent fibre density in the data. Secondly, we introduce population priors in the form of atlases of the local fibre orientation or of the global white matter bundle label to which individual fibres belong. As such, tractography in individual subjects is informed by common structure found across a cohort. Results indicate that such priors can reduce false positive tracks, thus improving specificity. Finally, we develop a blind source separation technique for multi-shell DWI, decomposing the data as a convolutive mixture of nonnegative tissue orientation distribution functions and corresponding response functions, without assuming the latter as known thus fully unsupervised. In healthy human brain data, the resulting components are associated with white matter fibres, grey matter, and cerebrospinal fluid. This factorization is on par with state-of-the-art supervised methods, as demonstrated also in Monte-Carlo simulations evaluating accuracy and precision. In animal data and in the presence of edema, our method is able to recover unseen tissue structure, fully data-driven. In summary, we developed local and global fibre reconstruction methods for DWI that improve over the state-of-the-art and extend to applications in preclinical data and pathology.

Theses

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

This book constitutes the refereed joint proceedings of the First International Workshop on Smart Ultrasound Imaging, SUSI 2019, and the 4th International Workshop on Preterm, Perinatal and Paediatric Image Analysis, PIPPI 2019, held in conjunction with the 22nd International Conference on Medical Imaging and Computer-Assisted Intervention, MICCAI 2019, in Shenzhen, China, in October 2019. The 10 full papers presented at SUSI 2019 and the 10 full papers presented at PIPPI 2019 were carefully reviewed and selected. The SUSI papers cover a wide range of medical applications of B-Mode ultrasound, including cardiac (echocardiography), abdominal (liver), fetal, musculoskeletal, and lung. The PIPPI papers cover the detailed scientific study of volumetric growth, myelination and cortical microstructure, placental structure and function.

Artificial intelligence. --- Optical data processing. --- Application software. --- Computer organization. --- Artificial Intelligence. --- Computer Imaging, Vision, Pattern Recognition and Graphics. --- Computer Applications. --- Computer Systems Organization and Communication Networks. --- Organization, Computer --- Electronic digital computers --- Application computer programs --- Application computer software --- Applications software --- Apps (Computer software) --- Computer software --- Optical computing --- Visual data processing --- Bionics --- Electronic data processing --- Integrated optics --- Photonics --- Computers --- AI (Artificial intelligence) --- Artificial thinking --- Electronic brains --- Intellectronics --- Intelligence, Artificial --- Intelligent machines --- Machine intelligence --- Thinking, Artificial --- Cognitive science --- Digital computer simulation --- Logic machines --- Machine theory --- Self-organizing systems --- Simulation methods --- Fifth generation computers --- Neural computers --- Optical equipment --- Diagnostic ultrasonic imaging --- Image processing --- Computer vision. --- Computer engineering. --- Computer networks. --- Computer and Information Systems Applications. --- Computer Engineering and Networks. --- Digital techniques. --- Communication systems, Computer --- Computer communication systems --- Data networks, Computer --- ECNs (Electronic communication networks) --- Electronic communication networks --- Networks, Computer --- Teleprocessing networks --- Data transmission systems --- Digital communications --- Electronic systems --- Information networks --- Telecommunication --- Cyberinfrastructure --- Network computers --- Machine vision --- Vision, Computer --- Artificial intelligence --- Pattern recognition systems --- Digital image processing --- Digital electronics --- Distributed processing --- Design and construction

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

The influence of hybrid modelling on deep learning-based MRI reconstruction performance. Magnetic Resonance imaging is a crucial diagnostic tool, particularly valued for its superior soft tissue contrast and safety compared to other modalities like Computed Tomography imaging. However, long acquisition times pose challenges for patient comfort and healthcare efficiency. Unfortunately, naive acceleration of the Magnetic Resonance imaging process often results in image artifacts, compromising the diagnostic quality of the images. Traditional image reconstruction methods such as GRAPPA, SENSE, and CS have been the standard in clinical practice to mitigate this, but are limited by moderate acceleration factors and long iteration times. Recent advancements in deep learning have shown promise in addressing these limitations. Nonetheless, such models remain inherently data-hungry. In an attempt to deal with this, the novel approach of hybrid modelling has emerged, integrating classical reconstruction techniques with deep learning models to potentially improve reconstruction performance for the same amount of training data available. This work employed the fastMRI brain dataset to train and evaluate three different hybrid neural network models: GRAPPA U-Net, SENSE U-Net, and CS U-Net. Additionally, two conventional deep learning baseline models were created: ZF (ACS) U-Net and ZF (no ACS) U-Net, trained with and without the presence of an autocalibration signal region in k-space respectively. These models were evaluated on test scans accelerated with acceleration factors of R=4 and R=8. Used evaluation metrics include traditional measures like the NMSE, PSNR, MSSIM index, and VGG loss, as well as a newly proposed Singular Value Decomposition Metric to assess error structure. Experimental results reveal that hybrid models significantly outperform conventional deep learning and classical reconstruction methods, especially in the absence of an autocalibration signal region in k-space. The CS U-Net model demonstrated the highest resilience, effectively handling an acceleration factor of at least R=4 without noticeable artifacts in inspected test scans. These findings underscore the potential of hybrid modelling to enhance Magnetic Resonance image reconstruction, suggesting that further optimization and exploration of hybrid architectures could yield even greater improvements.