The automatic analysis of medical images has played a key role in many discoveries in neuroscience over the past two decades. Magnetic resonance imaging (MRI) maintains a central role in this scientific process as well as in clinical neuroimaging because of its ability to use different pulse sequences that can provide alternate contrasts capable of revealing subtle tissue differences in both normal and diseased tissues. Yet there are three widely recognized problems in the reliable and consistent application of automatic image processing algorithms to MR data. First, the lack of a standardized scale in the image measurements means that results obtained on different scanners or at different times are not necessarily comparably quantified for individual studies or reliably pooled for population studies. For example, T1-weighted images are routinely acquired, but differences in the pulse sequences can cause significant differences in the brain tissue contrasts. Second, tissue contrasts that are ideal for certain steps in automatic processing are not always acquired in a given study or at a given imaging center. For example, although double-echo PD/T2-weighted images are routinely acquired, FLAIR images are often omitted for time considerations unless white matter lesions are expected or directly under study. Third, images often have intensity shading artifacts caused by spatially varying coil sensitivity patterns. These problems are worse at higher field strengths, preventing consistent analysis of these data without correction. All three of these problems will be addressed in this research project by investigation and further development of the method called Magnetic Resonance Image Example-based Contrast Synthesis (MIMECS). MIMECS is a post processing method that uses a standardized atlas with multiple images in order to synthesize contrasts that are consistent with the atlas given one or more subject images. The strategy is quite different than past approaches, which have focused on rich data acquisition, nonlinear atlas registration, or histogram modification techniques. MIMECS focuses on image synthesis using patches that index into an atlas thousands of times in order to learn an optimal synthesis formula at each voxel. It uses anatomical information from the atlas while avoiding the time-consuming process that would be required of a multi-atlas nonlinear registration approach. The research plan comprises three specific aims: 1) The theory of example-based image synthesis will be studied in order to optimize MIMECS;2) The computational approach will be refined and optimized for different applications;3) The use of MIMECS in synthesizing both FLAIR images for white matter lesion detection and optimized T1-weighted images for cortical surface extraction will be thoroughly evaluated on large existing data sets. The software will be thoroughly tested and then released as open source software within the Java Image Science Toolkit (JIST) for widespread availability to the neuroscience community.
Automated image analysis of magnetic resonance images plays a central role in neuroscience, yet it is very challenging to obtain consistent results when data is acquired from different scanners or at significantly different times. This exploratory research project will develop, validate, and make freely available as an open source software tool a post processing method called Magnetic Resonance Image Example-based Contrast Synthesis (MIMECS), which addresses these standardization issues using a novel atlas-based strategy.
|Roy, Snehashis; Carass, Aaron; Prince, Jerry L et al. (2014) Subject Specific Sparse Dictionary Learning for Atlas based Brain MRI Segmentation. Mach Learn Med Imaging 8679:248-255|
|Jog, Amod; Carass, Aaron; Pham, Dzung L et al. (2014) RANDOM FOREST FLAIR RECONSTRUCTION FROM T 1, T 2, AND PD -WEIGHTED MRI. Proc IEEE Int Symp Biomed Imaging 2014:1079-1082|
|Roy, Snehashis; Wang, Wen-Tung; Carass, Aaron et al. (2014) PET attenuation correction using synthetic CT from ultrashort echo-time MR imaging. J Nucl Med 55:2071-7|
|Bilgel, Murat; Carass, Aaron; Resnick, Susan M et al. (2014) Deformation field correction for spatial normalization of PET images using a population-derived partial least squares model. Mach Learn Med Imaging 8679:198-206|
|Roy, Snehashis; Carass, Aaron; Jog, Amod et al. (2014) MR to CT Registration of Brains using Image Synthesis. Proc SPIE Int Soc Opt Eng 9034:|
|Jog, Amod; Carass, Aaron; Prince, Jerry L (2014) IMPROVING MAGNETIC RESONANCE RESOLUTION WITH SUPERVISED LEARNING. Proc IEEE Int Symp Biomed Imaging 2014:987-990|
|Roy, Snehashis; Carass, Aaron; Prince, Jerry (2013) Magnetic Resonance Image Example Based Contrast Synthesis. IEEE Trans Med Imaging :|