Dictionary Learning for Automated Cell Tracking in Magnetic Resonance Imaging (MRI)

Standing

Undergraduate

Type of Proposal

Oral Research Presentation

Challenges Theme

Safeguarding Healthy Great Lakes

Your Location

Windsor

Faculty

Faculty of Science

Faculty Sponsor

Dr. Dan Xiao

Proposal

Magnetic Resonance Imaging (MRI) is a powerful imaging modality with excellent soft tissue contrast. Contrast agents can be used to “label” cells creating signal voids, allowing individual cells to be imaged. Time-lapse MRI can be used to track the motion of tagged cells, providing insights in the studies of inflammatory diseases and metastasis of cancer. Counting cells manually is cumbersome, motivating the development of an automated technique.

A dictionary learning based method has been developed for this feature extraction problem. In dictionary learning, a set of “atoms”, representing features of an image, are “fit” to an image. A digital image is simply a set of values representing pixel brightness’s, an atom in dictionary learning is the set of values which create a certain feature. The “fit” is an optimization of parameters, which determine the weighting of each atom (feature), in a given patch of the image. Using the information from the fit parameters, the locations of features, such as signal voids, can be obtained. The method was tested on human brain images with simulated signal voids of various contrast. The signal to noise ratio (SNR) was also varied by adding gaussian noise. With sufficient contrast and SNR, the algorithm was able to extract the signal void features in simulated images with complex features. The method will be applied to process the images acquired in cell tracking MRI experiments, to release the burden of manual counting. The automatic image feature detection technique can also be translated to other systems.

Share

COinS
 

Dictionary Learning for Automated Cell Tracking in Magnetic Resonance Imaging (MRI)

Magnetic Resonance Imaging (MRI) is a powerful imaging modality with excellent soft tissue contrast. Contrast agents can be used to “label” cells creating signal voids, allowing individual cells to be imaged. Time-lapse MRI can be used to track the motion of tagged cells, providing insights in the studies of inflammatory diseases and metastasis of cancer. Counting cells manually is cumbersome, motivating the development of an automated technique.

A dictionary learning based method has been developed for this feature extraction problem. In dictionary learning, a set of “atoms”, representing features of an image, are “fit” to an image. A digital image is simply a set of values representing pixel brightness’s, an atom in dictionary learning is the set of values which create a certain feature. The “fit” is an optimization of parameters, which determine the weighting of each atom (feature), in a given patch of the image. Using the information from the fit parameters, the locations of features, such as signal voids, can be obtained. The method was tested on human brain images with simulated signal voids of various contrast. The signal to noise ratio (SNR) was also varied by adding gaussian noise. With sufficient contrast and SNR, the algorithm was able to extract the signal void features in simulated images with complex features. The method will be applied to process the images acquired in cell tracking MRI experiments, to release the burden of manual counting. The automatic image feature detection technique can also be translated to other systems.