1. NA-MIC National Alliance for Medical Image Computing Georgia Tech Contributions to NAMIC Allen Tannenbaum

2. Georgia Tech Researchers Delphine Nain (GRA): Shape analysis, rule-based and stochastic methods in segmentation, validation methods (Received Ph.D. 2006) John Melonakos (GRA): Knowledge-based segmentation; rule- based segmentation, DTI, fMRI Eric Pichon: DTI Tractography, Statistically-based segmentation, validation, visualization (Received Ph.D. 2005) Shawn Lawton (GRA): Conformal and optimal transport methods for image registration Marc Niethammer: Active contour methods (Received Ph.D. 2004); posdoc 2005; now working with Martha Shenton Ramsey Al-Hakim (Undergraduate researcher): Rule-based segmentation methods Vandana Mohan (GRA): DTI, directional based segmentation Yi Gao (GRA): Shape based analysis and representations Xavier LeFaucheur (GRA): KPCA, GPCA

3. NAMIC Summary Our Contributions: –Finding white matter tracts and blood vessels –Shape representation –Geometric and shape driven segmentation –Conformal and optimal transport methods for visualization and registration –Rule-based segmentation –Statistical/PDE methods –Stochastic methods for curvature based flows in medical imaging Benefits to us: –Imaging for neuroengineering and neuroscience research –Work directly with clinicians –Opportunity to have our methods directly impact medical imaging technology

4. Cost functional: Resulting maximizing flow: –unidirectional flow, efficient implementation –part of open-source software 3D Slicer favors large regions favors homogeneous regions normalized histogram Region-based segmentation Pichon, Tannenbaum & Kikinis, MICCAI and MedIA

5. Region-based Segmentation Results: White and gray matter (MRI) comparison to ground truth from [kaus01] 2-d sagittal slice of image and proposed segmentation 3-d rendering proposed ground truth

6. Region-based Segmentation Results: Brain ventricle (MRI) comparison to ground truth from [kaus01] 2-d axial slice of image and proposed segmentation 3-d rendering proposed ground truth

7. Define a direction-dependent local cost: Examples –Tractography: using high angular resolution dataset, define  (p,d) to be small if there is a fiber at position p in direction d, large otherwise –Segmentation: define  (p,d) from a direction-dependent pattern detector local cost Direction-dependent segmentation (Active Contours in a Finsler Metric)

8. position tangent direction local cost global cost curve local cost Direction-dependent segmentation Define the global cost of a curve by integrating the local cost. Minimal cost curves can be obtained using dynamic programming or calculus of variations.

9. Applications proposed technique streamline technique (based on tensor field) 2-d axial slide of tensor field (based on S/S 0 ) Diffusion MRI tractography Vessel Segmentation

10. We represent shapes with spherical wavelet basis functions localized in space and scale Localized Shape Analysis Spherical wavelet functions Resolution 2Resolution 4 To describe a shape in a population, each wavelet coefficient encodes variation from the mean shape at a particular scale & location Original Caudate Mean Caudate Low Resolution Wavelet coeffs Low and High Resolution coeffs =+ …+ Characterization of local variations could be important for shape analysis since a disease, such as cancer, could affect only a portion’s of an organ’s surface

11. Our technique learns a shape prior from the distribution of the wavelet coefficients In an estimation task, our prior incorporates local details that a previous technique (PCA) does not encode and significantly improves the approximation of shapes. Localized Shape Analysis Used multiscale shape prior in a segmentation framework and apply it to the task of shape classification Ground Truth PCA estimationWavelet prior estimation Comparison of techniques for estimation of a test shape (not used to learn shape prior)

12. Evolve Shape evolve ,p Data PriorShape Prior Shape Representation Segmentation Region-Based Active Contour Pose: Rotations, Translations, Scaling shape Prior: represent shape in eigenvector subspace (  coordinates) constrain  value (+/- 4 std) Shape Driven Segmentation

13. Automatic Brain Registration Given a segmented surface Use deep sulci as landmarks Use conformal map to flatten surface to an annulus Use mass preserving map to register the two annuli This technique could help perform 3D atlas registration extremely quickly

14. Conformal Mapping of Neonate Cortex

15. Area Preserving Surface Warping of Minimal Distortion Optimal transport allows one to find area correcting flattening. After conformally flattening surface, apply area correcting map to find area-preserving flattening of minimal distortion.

16. Segmentation via Stochastic Flows

17. Bayesian Classifier Image Filter: This filter performs Bayesian image segmentation. John Melonakos, Georgia Institute of Technology Luis Ibanez, Kitware (software) Karthik Krishnan, Kitware (software) Algorithm: The basic idea is to incorporate prior knowledge into the segmentation through Bayes’ rule. Image noise is removed via an affine invariant anisotropic smoothing of the posteriors AlgorithmTeam Publication Illustration of Algorithm Clinical Applications: test by segmenting brain volumes into white, gray and background Open Source Software ITK code developed during the NAMIC Programming Week in 2006 has been ported to an ITK filter and is in the NAMIC Sandbox and ITK CVS repository. J. Melonakos, K. Krishnan, and A. Tannenbaum. “An ITK Filter for Bayesian Segmentation: itkBayesianClassifierImageFilter”. Insight Journal, 2006 Raw ImageManual SegmentationITK Filter Output

18. Conformal Flattening: ITK Conformal Flattening Filter. This is useful for the visualization of irregular surfaces. Yi Gao, Georgia Institute of Technology John Melonakos, Georgia Institute of Technology Jim Miller, GE (software) Luis Ibanez, Kitware (software) Algorithm: Use conformal mapping to map an irregular surface onto a sphere while preserving the angle. class: itkConformalFlatteningFilter APIs: filter->setPointP(cellId); filter->mapToPlane( ); filter->setScale( scaleFactor ); Y. Gao, J. Melonakos, and A. Tannenbaum. “Conformal Flattening ITK Filter.” ISC/NA-MIC Workshop on Open Science at MICCAI 2006 ITK code developed during the NAMIC Programming Week in 2006 has been ported to an ITK filter and is in the NAMIC Sandbox Clinical Applications: This is useful for the visualization of irregular surfaces. In the special case of FMRI visualization, the flattened view facilitates the understanding of the mapping of function to spatial location. AlgorithmTeam Publication Open Source Software Brain Surface Conformal Mapping of Brain Surface Illustration of Algorithm

19. 3D shape analysis using Spherical Wavelets: Develop algorithm and ITK software module for Spherical Wavelet computation Xavier Le Faucheur, Georgia Institute of Technology Delphine Nain, Georgia Institute of Technology Yi Gao, Georgia Institute of Technology Martin Styner, University of North Carolina John Melonakos, GT /GE Luis Ibanez, Kitware (software) Algorithm: decomposition of a scalar signal defined on a spherical mesh into spherical wavelet coefficients and vice-versa. This algorithm can be used to represent shape via a conformal mapping to the sphere, and then an encoding of the (x,y,z) signal into wavelet coefficients. Clinical Applications: The multiscale shape representation can be used for segmentation and shape analysis [1-2] AlgorithmTeam [3] X. LeFaucheur, Y. Gao, D. Nain, A. Tannenbaum “Spherical Wavelet ITK Filter”. To be submitted to Insight journal. ITK code developed during the NAMIC Programming Week in 2006 is in the NAMIC Sandbox. Publications Open Source Software Illustration of Algorithm LPHP LPHP LPHP Spherical WaveletObject Spherical WaveletObject Signal-to- Coeffs Filter Coeffs-to- Signal Filter [1] D. Nain, et al. “Multiscale 3D Shape Analysis using Spherical Wavelets”. In MICCAI, 2005 [2] D. Nain, et al. “3D Segmentation using Spherical Wavelets”. In MICCAI, 2006

20. Rule Based Segmentation: Semi-Automatic Segmentation of Brain Structures using rules created by neuroanatomists Ramsey Al Hakim, Georgia Institute of Technology Shawn Lankton, Georgia Institute of Technology Tauseef Rehman, Georgia Institute of Technology John Melonakos, Georgia Institute of Technology Delphine Nain, Georgia Institute of Technology Jim Levitt, Brigham and Women’s Hospital Jim Fallon, UCI Striatum Segmentation: Manually input most superior/dorsal point on putamen and anterior commissure; striatum is then delineated automatically based on rules of Dr. James Levitt. Dorso-Lateral Prefrontal Cortex Segmentation: Manually input temporal lobe tip and frontal pole points based on Fallon-Kindermann Rules; DLPFC is then segmented automatically using the bayesian classification tool Clinical Applications: Study the effects of schizophrenia on shape. AlgorithmsTeam Rule-Based Segmentation Module is in Slicer 2 Publications Open Source Software Illustration of Algorithm [1] R. Al-Hakim, J.Fallon, D. Nain, J. Melonakos, A. Tannenbaum. A DLPFC semi-automatic segmenter. In SPIE Medical Imaging, 2006. [2] J. Melonakos, R. Al-Hakim, J. Fallon, and A. Tannenbaum. Knowledge-Based Segmentation of Brain MRI Scans Using the Insight Toolkit. Insight Journal, 2005. Anterior commissure Post Putamen Pre Caudate Post Caudate Nucleus Accumbens Pre Putamen superior/dorsal point on putamen Anterior commissure

21. NAMIC Publications-I

22. NAMIC Publications-II

23. NAMIC Publications-III