Jiachen Zhuo

I work with a group of researchers at the University of Maryland School of Medicine who have been at the forefront of MR imaging–guided focused ultrasound (MRgFUS) development to treat movement disorders including essential tremor, Parkinson disease, and neuropathic pain and develop other techniques including blood-brain barrier opening and neuromodulation. A key question that is in most people’s minds during this procedure is how to improve the workflow for the patients and minimize any additional burden to the patients. Inspired by the developments in radiation oncology to use MRI to create pseudo-CT images for attenuation correction, it occurred to us that something similar may be done for MRgFUS procedures, where one could obtain skull density information from MRI rather than CT and the result would be that patients would not have to undergo yet another procedure that involved radiation exposure.

Ultrashort TE (UTE) imaging has been a promising MRI method that provides good skull and bone information. In this paper where we used a deep learning method and convolutional neural networks, we extend our previous work where we established the relationship between UTE MRI signal intensity and CT Hounsfield units. UTE imaging has been very promising for us to obtain good, high-resolution synthetic CT skull images with high accuracy.

There has been a lot of interest among the academicians to use this technique not only for human work but also to extend it to the preclinical imaging setting. We are also working with our industry partners to further develop the technology and integrate it into the present clinical workflow.

As always, there is room for further improvement. We are currently exploring a new direction of research to further improve the accuracy of the synthetic CT, including utilization of multimodal MRI data to further improve and validate the deep learning network to not only synthesize the skull but also accurately identify calcification regions. And lastly, we plan on validating the developed algorithms at multiple sites to assess the robustness of the model.

Read this article at AJNR.org …