Kyoko Fujimoto¹, Jonathan R Polimeni¹, Andre J van de Kouwe¹, Tobias Kober², Thomas Benner¹, Bruce Fischl^1,3, and Lawrence L Wald^1,4
1Athinoula A. Martinos Center for Biomedical Imaging, Department of Radiology, Harvard Medical School, Massachusetts General Hospital, Charlestown, MA, United States,²Laboratory for Functional and Metabolic Imaging, Ecole Polytechnique Fédérale de Lausanne, Advanced Clinical Imaging Technology, Siemens Suisse SA - CIBM, Lausanne, Switzerland, ³Computer Science and AI Lab (CSAIL), Massachusetts Institute of Technology, Cambridge, MA, United States, ⁴Harvard-MIT Division of Health Sciences and Technology, Massachusetts Institute of Technology, Cambridge, MA, United States

Here we demonstrate that accurate surface models can be generated from 7T anatomical data with the recently introduced MP2RAGE pulse sequence with some additional preprocessing steps prior to using the FreeSurfer software package. We compared the surfaces generated from 7T MP2RAGE data with those generated from 3T MP2RAGE data and 3T MEMPRAGE data. We performed a test-retest analysis with the 3T data to quantify the reproducibility of the surface models and to estimate the precision of the surface reconstruction across the two acquisition methods.

Delphine Ribes^1,2, Tobias Kober^1,2, Giulio Gambarota³, Reto Meuli⁴, and Gunnar Krueger^2
1Laboratory for functional and metabolic imaging, Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland, ²Advanced Clinical Imaging Technology, Siemens Suisse SA - CIBM, Lausanne, Switzerland, ³Clinical Imaging Center, GSK, Imperial College, London, United Kingdom, ⁴Department of Radiology, Centre Hospitalier Universitaire Vaudois, Lausanne, Switzerland

Being a preliminary step for many clinical applications and analyses, accurate skull-stripping is a key challenge in MR brain imaging. One of its major difficulties arises from the contrast similarities at brain/non-brain tissue interfaces. Multispectral imaging may help to mitigate this problem. Specifically, the acquisition of multiple echoes in a MP-RAGE sequence as shown in the work of van der Kouwe et al. (2008) can be used for this purpose. We combine their approach with the classical Dixon method to obtain an additional contrast depicting only the fat signal. This work investigates whether the thus generated additional information can improve the outcome of an unsupervised intensity-based skull-stripping algorithm.

Judd M Storrs^1,2, and Jing-Huei Lee^1,3
1Center for Imaging Research, University of Cincinnati, Cincinnati, OH, United States, ²Department of Psychiatry and Behavioral Neuroscience, University of Cincinnati, Cincinnati, OH, United States, ³School of Energy, Environmental, Biological and Medical Engineering, University of Cincinnati, Cincinnati, OH, United States

Online spatial normalization to the ICBM452 T₁-weighted template was integrated with acquisition of high-resolution 3D anatomic images. The low-spatial frequency components are used for affine spatial normalization during acquisition of the high-spatial frequencies. Online normalization completes prior to the end of scanning and the atlas coordinate system is available immediately for use by the next queued scan.

Ziad Serhal Saad¹, Andrej Vovk², Janez Stare³, Dusan Suput², and Robert W Cox^1
1SSCC, NIMH/NIH, Bethesda, MD, United States, ²Institute of Pathophysiology, University of Ljubljana, Ljubljana, Slovenia, ³Institute for Biostatistics and Medical Informatics, University of Ljubljana, Ljubljana, Slovenia

We present a novel approach for generating a voxel's tissue class membership based on its signature; a collection of spatial texture statistics calculated over a set of spherical neighborhoods around that voxel. We produce tissue class priors that can initialize and regularize image segmentation much in the way population-based priors do as a function of spatial location in standard template space. The signature-based approach is a distinct departure from location-based methods by not requiring population-derived spatial template, registration to template's space, and bias field estimation. It is also suitable where location-based templates are not available or appropriate.

M Mallar Chakravarty^1,2, Matthijs Christiaan van Eede¹, and Jason P Lerch^1
1Mouse Imaging Centre (MICe), The Hospital for Sick Children, Toronto, Ontario, Canada, ²Rotman Research Institute, Baycrest, Toronto, Ontario, Canada

In human MRI experiments, segmentation of neuroanatomy is often accomplished using a single atlas based nonlinear transformation estimation. The accuracy of this technique is limited by errors in the nonlinear transformation estimated, differences in the neuroanatomy between the template brain and the subject, or label resampling errors. Recent work demonstrates improvement of these segmentation techniques through the use of a manually generated template library. In this methodology, instead of using a single expertly labeled MRI template, a number of different templates are manually labeled, and transformations are estimated to match a single subject to each of these templates. After the nonlinear transformations are applied to the anatomical labels, a histogram of labels generated at each voxel can be used to inform the final segmentation on a voxel-by-voxel basis. This template library approach thus improves segmentation accuracy by accounting for varying anatomy through the use of different templates and compensating for registration algorithm inaccuracy by virtue of the multiple registrations needed from each MRI in the template to the target. In the segmentation of MRI data from inbred laboratory mice strains, however, the confounds of variable neuroanatomy are limited, and segmentation errors therefore result from registration inaccuracy and resampling errors. We hypothesize that segmentations can be improved if resampling and nonlinear transformation errors are reduced. Here, we test this hypothesis by implementing a multi-atlas segmentation scheme using automatically generated atlases (instead of manually labeled ones) and verified the accuracy of the segmentation using manually derived gold standards of the neuroanatomy.

Yajing Zhang¹, Jiangyang Zhang², Jun Ma³, Kenichi Oishi², Andreia V. Faria², Michael I. Miller^1,3, and Susumu Mori^2,4
1Department of Biomedical Engineering, Johns Hopkins University School of Medicine, Baltimore, MD, United States, ²Department of Radiology and Radiological Science, Johns Hopkins University School of Medicine, Baltimore, MD, United States, ³Center for Imaging Science, Johns Hopkins University, Baltimore, MD, United States, ⁴F.M. Kirby Research Center for Functional Brain Imaging, Kennedy Krieger Institute, Baltimore, MD, United States

An MR based brain atlas is a key component in modern image analysis process. In this study, a population averaged brain atlas was generated from 20 subject MRI/DTI data using a continuous fluid dynamic model based on image metric distance. This estimated atlas presents a group averaged shape that minimize its anatomical bias while preserves sharp image contrast for accurate structure delineation and image mapping. The characteristics of the estimated atlas were examined with respect to a single subject atlas and two group averaged atlases.

Shannon Agner¹, Jun Xu¹, Sudha Karthigeyan¹, and Anant Madabhushi^1
1Biomedical Engineering, Rutgers University, Piscataway, New Jersey, United States

Accurate lesion segmentation is an important component of determining quantitative features for lesions on MRI. In this study, we develop an automated segmentation method for delineating lesions on DCE-MRI using spectral embedding which serves as alternative image representation upon which to perform an active contour lesion segmentation. We demonstrate on a cohort of 50 breast DCE-MRI datasets that the automated spectral embedding based active contour (SEAC) provides lesion segmentations that are more comparable to the manual segmentation performed by a radiologist than the popular automated fuzzy c-means segmentation method. While we demonstrate the use of SEAC with breast DCE-MRI data, SEAC could be easily applied to segmenting structures on other high dimensional, time-series imaging data as well.

Hassan Bagher-Ebadian^1,2, Siamak P Nejad-Davarani^1,3, Ramesh Paudyal¹, Tom Mikkelsen⁴, Quan Jiang^1,2, and James R Ewing^1,2
1Neurology, Henry Ford Hospital, Detroit, MI, United States, ²Physics, Oakland University, Rochester, MI, United States, ³Biomedical Engineering, University of Michigan, Ann Arbor, MI, United States, ⁴Neurosurgery, Henry Ford Hospital, Detroit, MI, United States

Estimating the longitudinal relaxation time, T1, from spoiled-gradient-recalled-echo (SPGR) images is challenging and susceptible to the level of noise-to-signal ratio (SNR) in acquisition. Methods such as Simplex-Optimization, Weighted-Non-Linear-Least-Squares, Linear-Least-Square, and Intensity-based-Linear-Least-Square have been employed to estimate T1. In linear and non-linear methods, the estimated T1 values are dependent on defining the weighting factors, which may result in a biased estimation. Herein, an adaptive neural network is trained and compared with different techniques using an analytical model of the SPGR signal in the presence of different levels of SNR. Receiver-Operator-Characteristic analysis and K-fold-cross-validation were employed for validation, testing, and network optimization.

William D Rooney¹, James R Pollaro¹, Sean C Forbes², Dah Jyuu Wang³, Krista Vandenborne², and Glenn A Walter^4
1Advanced Imaging Research Center, Oregon Health and Science University, Portland, OR, United States, ²Department of Physical Therapy, University of Florida, Gainesville, Florida, United States, ³Department of Radiology, The Children's Hospital of Philadelphia, Philadelphia, Pennsylvania, United States, ⁴Department of Physiology and Functional Genomics, University of Florida, Gainesville, Florida, United States

Quantitative transverse relaxography (qT₂) of proton MR signals has shown sensitivity for pathology of tissues such as muscles in patients with DMD. Standardization across multiple sites, as well as imperfections in multi-echo imaging sequences has led to contamination of the desired primary echo decay. Crushing gradient schemes have been developed, but these can be difficult to implement, especially in multi-slice acquisitions. Extended phase graphs applied during post-processing can isolate the primary echo to improve accuracy of qT₂ mapping.

Stephen LaConte¹, Jonathan Lisinski¹, and Bradley Sutton^2
1School of Biomedical Engineering and Sciences, Virginia Tech, Blacksburg, VA, United States, ²Bioengineering, University of Illinois, Urbana-Champaign, Urbana, IL, United States

We imaged the oropharyngeal cavity at 15.8 frames per second using a recently developed multi-shot, field corrected, dynamic spiral FLASH sequence. We explored the extent to which speech-related information is captured by this sequence. During imaging, we asked a subject to perform a visually guided speech task, consisting of alternating 20 sec. blocks of slow and fast counting. Support vector machine analysis used the soft palate, lips, and tongue and resulted in 88% prediction accuracy, demonstrating that it is possible to classify individual frames as either “fast� or “slow� speech. This achievement has potential applications in speech therapy and diagnosis.