Real-time tomographical reconstructions using Neural Networks

Tomographical algorithms can be separated into two classes: analytical “one-step” methods and iterative reconstruction algorithms. Analytical methods are fast, but require projection data of high quality and are impossible to adapt to use prior knowledge about the reconstructed object. Iterative algorithms have less strict requirements for the used data and are more flexible, but their computation time impedes real-time reconstructions. By reformulating the reconstruction problem as a classification problem, a third option becomes available: machine learning.

Tomographic segmentation

Many segmentation techniques exist in the literature. Well known is global thresholding with automated threshold selection based on the image histogram, e.g. Otsu's method. Other commonly used methods are region growing, watershed segmentation, active contours, etc. In the setting of CT, these techniques base themselves exclusively on the tomographic reconstruction or the tomogram. In practice, however, these tomograms will not be completely accurate because of reconstruction error and artefacts (e.g.

Quantification of brain connectivity using Diffusion Weighted MRI

Diffusion Weighted (DW) MRI is a unique and noninvasive method to characterize tissue microstructure, based on the random thermal motion of water molecules. Of particular interest is its potential for inferring the orientation of the coherently oriented fiber bundles within brain white matter tissue, as this opens up the possibility of investigating brain connectivity in vivo using so-called fiber-tracking algorithms. This relatively new technique is becoming a valuable diagnostic tool for a large number of neuropathological diseases.

GPU-based iterative tomography

Even though iterative reconstruction algorithms give very accurate results, their long computation time inhibits practical applications. The most computationally intensive steps of the algorithms used by the ASTRA research group are very well parallelizable. Modern GPU cores consist of many processing units and are optimized for operations also required during tomographic reconstructions. With the release of NVIDIA CUDA, GPUs became fully programmable.

Subscribe to RSS - Reconstruction