1. Subject Specific Sparse Dictionary Learning for Atlas Based Brain MRI Segmentation

Publication:
S. Roy, A. Carass, J. L. Prince, D. L. Pham
"Subject specific sparse dictionary learning for atlas based brain MRI segmentation", Machine Learning in Medical Imaging, pp. 248-255, 2014.
S. Roy, Q. He, E. Sweeney, A. Carass, D. S. Reich, J. L. Prince, D. L. Pham
"Subject specific sparse dictionary learning for atlas based brain MRI segmentation", IEEE Journal of Biomedical and Health Informatics, vol. 19, no. 5, 1598-1609, 2015.

More Info

S3DL



S3DL

2. PET Attenuation Correction using Synthetic CT from Ultrashort Echo-time MRI

Publication:
S. Roy, W. T. Wang, A. Carass, J. L. Prince, J. A. Butman, D. L. Pham
"PET Attenuation Correction Using Synthetic CT from Ultrashort Echo-Time MR Imaging" Journal of Nuclear Medicine, vol. 55, no. 12, 2071-2077, 2014.

More Info

3. Magnetic Resonance Image Example Based Contrast Synthesis (MIMECS)

Abstract: The tissue contrast of a magnetic resonance (MR) neuroimaging dataset has an impact on the performance of image analysis tasks such as registration and segmentation. The type of pulse sequence, its implementation, and the scanner type and calibration determine the tissue contrast that one will observe, and these details are difficult to control in large cross-sectional or longitudinal studies. It is also common to encounter a dataset in which a desired tissue contrast is missing, i.e., never actually acquired, which may prevent certain image processing steps from being carried out or their results applied to alternate data may be inconsistent with the rest of the study. This paper introduces a sparse prior based technique that uses image patches from an atlas to synthesize contrasts not present or not intensity normalized in the original dataset. The proposed image synthesis technique is demonstrated using two applications. First, it is used to normalize the intensity of images acquired using the same pulse sequence but with different parameters or on different scanners. Second, it is used to synthesize images with a different tissue contrast than that which was acquired. Unlike previous synthesis methods, the proposed method does not require images to be acquired using a particular pulse sequence or set of pulse sequences or to carry out an image registration procedure as part of the process. The method is shown to yield more consistent segmentations and to work in the presence of mild pathologies.

Publication: S. Roy, A. Carass, J. L. Prince, "A compressed sensing approach for MR tissue contrast synthesis", IPMI, pp. 371-383,2011. (Best Poster Award )
S. Roy, A. Carass, J. L. Prince, "Magnetic Resonance Image Example based Contrast Synthesis", IEEE Trans. Medical Imaging, vol. 32, no. 12, pp. 2348 - 2363, 2013.

More Info

Example: Suppose we want to synthesize MPRAGE contrast of an SPGR image using atlases. We can certainly use deformable registration to register the subject image to an atlas SPGR image, and transfer the deformation to the MPRAGE scan of the atlas. However, registration might fail due to pathologies (such as lesions) between the subject and the atlas. Our method can overcome this issue.

Big Boat

There might be changes in the anatomy that a registration can not reproduce. An example is shown below where the subject has much larger ventricles than the atlas. Registration can not successfully recover the large change. However, MIMECS is not dependent on such differences.

Big Boat

4. Fuzzy C Means with Variable Compactness

Abstract: Fuzzy c-means (FCM) clustering has been extensively studied and widely applied in the tissue classification of biomedical images. Previous enhancements to FCM have accounted for intensity shading, membership smoothness, and variable cluster sizes. In this paper, we introduce a new parameter called "compactness" which captures additional information of the underlying clusters. We then propose a new classification algorithm, FCM with variable compactness (FCMVC), to classify three major tissues in brain MRIs by incorporating the compactness terms into a previously reported improvement to FCM. Experiments on both simulated phantoms and real magnetic resonance brain images show that the new method improves the repeatability of the tissue classification for the same subject with different acquisition protocols.

Publication: S. Roy, H. Agarwal, A. Carass, Y. Bai, D. L. Pham, J. L. Prince, "Fuzzy c-means with variable compactness", ISBI, pp. 452-455, 2008.

Software: Download executable

Example: An SPGR and MPRAGE acquisition of the same subject, when segmented with different methods, give different results, in general. Using FCMVC, the segmentations are made closer.
Big Boat


The cortical surfaces (CRUISE), generated from the fuzzy segmentations, match more closely to a set of manually picked landmarks.

Big Boat

5. A Rician Mixture Model Classification Algorithm

Abstract:. Tissue classification algorithms developed for magnetic resonance images commonly assume a Gaussian model on the statistics of noise in the image. While this is approximately true for voxels having large intensities, it is less true as the underlying intensity becomes smaller. In this paper, the Gaussian model is replaced with a Rician model, which is a better approximation to the observed signal. A new classification algorithm based on a finite mixture model of Rician signals is presented wherein the expectation maximization algorithm is used to find the joint maximum likelihood estimates of the unknown mixture parameters. Improved accuracy of tissue classification is demonstrated on several sample data sets. It is also shown that classification repeatability for the same subject under different MR acquisitions is improved using the new method.

Publication: S. Roy, A. Carass, P. L. Bazin, S. Resnick, J. L. Prince, "Consistent segmentation using a Rician classifier",
Medical Image Analysis, vol. 16, no. 2, pp. 524-535, 2011.

Software: Download executable

Example: An SPGR and MPRAGE acquisition of the same subject, in general, have different intensity histograms. As the noise in the MR intensities are primarily of Rician nature, a mixture of Ricians fit the histograms better than a mixture of Gaussians.

Big Boat

Better fit in intensity distributions is reflected in more consistent segmentation between these two contrasts.

Big Boat