Program Information
Rapid Internal Normalization of Spectroscopic MRI Maps Using a Gaussian Mixture Model
S Gurbani1*, E Schreibmann1, S Sheriff2, C Holder3, L Cooper4, A Maudsley2, H Shim1,3, (1) Department of Radiation Oncology, Winship Cancer Institute of Emory University, Atlanta, GA, (2) Department of Radiology, University of Miami Miller School of Medicine, Miami, FL, (3) Department of Radiology and Imaging Sciences, Emory University, Atlanta, GA, (4) Department of Biomedical Informatics & Winship Cancer Institute, Emory University, Atlanta, GA
Presentations
TU-AB-601-10 (Tuesday, August 1, 2017) 7:30 AM - 9:30 AM Room: 601
Purpose: Spectroscopic MRI (sMRI) enables detection of metabolic changes in tissue. Key metabolites in the brain include: choline (Cho), a tumor marker; and N-acetylaspartate (NAA), a neuronal marker. However, there are variations in baseline metabolism and the Cho/NAA ratio between patients and within different lobes of the brain. Therefore, for comparisons between studies it is necessary to apply a signal normalization to the sMRI maps, often implemented using a value derived from the same patient’s contralateral normal-appearing white matter (NAWM). Here, we present an algorithm which uses information in the Cho/NAA map and a corresponding T1-weighted (T1w) MRI to automatically compute the normalization factor and identify regions of abnormality.
Methods: The algorithm takes as input the Cho/NAA map and a co-registered T1w MRI. A histogram analysis is performed to approximate the bimodal distribution of voxels (healthy, tumor) in the ratio map. Using expectation maximization, a two-source Gaussian mixture model (GMM) is computed to separate these two populations, and filter out voxels greater than 3 standard deviations from either Gaussian source. The cerebral hemispheres are segmented using a mid-sagittal plane extraction method based on hemispheric symmetry. A 3D connected-component analysis is done on the voxels classified as tumor, and the hemisphere with the largest single component is selected as ipsilateral. White matter voxels in the contralateral hemisphere are then selected and the GMM is applied to determine which of these voxels are healthy; the mean is used as the normalizing factor for the entire map.
Results: The normalizing factor was compared to values determined by manual segmentation of NAWM by a neuroradiologist, and was found to be within 3% (n=16 scans), with each computation taking 2 seconds.
Conclusion: We report a practical algorithm for the quantitative normalization of spectroscopy MRI, which represents an important step towards sMRI standardization in radiotherapy planning.
Contact Email: