Development and Dissemination of Robust Brain MRI Measurement Tools (1R01EB006733) Dinggang Shen IDEA Department of Radiology and BRIC UNC-Chapel Hill Team • UNC-Chapel Hill - Dinggang Shen - 1/2 Postdoctoral fellow(s) • UPenn - Christos Davatzikos • GE - Jim Miller - Xiaodong Tao Goal of this project • To further develop HAMMER registration and white matter lesion (WML) segmentation algorithms, for improving their robustness and performance. • To design separate software modules for these two algorithms and incorporate them into the 3D Slicer. Overview of Our Brain Measurement Tools • To further develop HAMMER registration and WML segmentation algorithms, for improving their robustness and performance. • To design separate software modules for these two algorithms and incorporate them into the 3D Slicer. Format Converter PACS Database Data importer Data processing Skull Stripping Multimodality Registration Tissue Classification Data processing Learn Best Features Models Complexity Levels MI Q-MI Skull Stripping Intensity Normalization HAMMER Deformation Constraints Parameter Tuning Registration Tissue Density Maps ROI Labeling Manual Segmentation Voxel-wise Segmentation Training SVM Classifier False-Positive Elimination Application Training WML Atlas SPM Group Analysis ROI-based Analysis Applications HAMMER Registration Algorithm WML Segmentation Algorithm Visualization Engine HAMMER Matching attribute vectors Image registration and warping Shen, et al., “HAMMER: Hierarchical Attribute Matching Mechanism for Elastic Registration”, IEEE Trans. on Medical Imaging, 21(11):1421-1439, Nov 2002. (2006 Best Paper Award, IEEE Signal Processing Society) Registration – HAMMER (1) Formulated as correspondence detection Individual: Model: How can we detect correspondences? Difficulty: High variations of brain structures Solution: Use both global and local image features to represent anatomical structures, such as using wavelets or geometrical moments. Xue, Shen, et al., “Determining Correspondence in 3D MR Brain Images Using Attribute Vectors as Morphological Signatures of Voxels”, IEEE Trans. on Medical Imaging, 23(10): 1276-1291, Oct 2004. Distinctive character of attribute vector: toward an anatomical signature of every voxel Brain A Brain B Similarity Map Examples of attribute vector similarity maps, and point correspondences HAMMER (2) Hierarchical registration – reliable points first To minimize the effect of local minima Few driving voxels Smooth approximation of the energy function Roots of sulci Crowns of gyri Voxels with distinct attribute vectors. Many driving voxels Complete energy function All boundary voxels HAMMER (2) Hierarchical registration – reliable points first Beginning of registration End of registration 158 brains we used to construct average brain 158 subjects Average Template 3D renderings Model brain A subject before warping and after warping HAMMER in labeling brain structures: Model HAMMER Subject HAMMER - Cross-sectional views Model Subject Registration – HAMMER - Label cortical surface Inner cortical surface Outer cortical surface Model Subject Simulating brain deformations for validating registration methods Template Simulated Xue, Shen, et al., “Simulating Deformations of MR Brain Images for Evaluation of Registration Algorithms”, Neuroimage, Vol. 33: 855-866, 2006. Successful applications of HAMMER: 10+ large clinical research studies and clinical trials involving >8,000 MR brain images: • One of the largest longitudinal studies of aging in the world to date, (an 18-year annual follow-up of 150 elderly individuals) • A relatively large schizophrenia imaging study (148 participants) • A morphometric study of XXY children • The largest imaging study of the effects of diabetes on the brain to date, (650 patients imaged twice in a 8-year period) • A large study of the effects of organolead-exposure on the brain • A study of effect of sustained, heavy drinking on the brain Improving: Learning Best Features for Registration Best-scale moments: Criteria for selecting best-scale moments of each point: • Maximally different from those of its nearby points. (Distinctiveness) • Consistent across different samples. (Consistency) • Best scales, used to calculate best-scale features, should be smooth spatially. (Regularization) Moments w.r.t. scales: Wu, Qi, Shen, “Learning Best Features for Deformable Registration of MR Brains”, MICCAI, 2005. Improving: Learning Best Features for Registration Results: • Visual improvement: Model Ours HAMMER’s • Average registration error: Histogram of deformation estimation errors 0.07 0.06 0.05 0.04 0.03 Improved method HAMMER Wu, Qi, Shen, “Learning-Based Deformable Registration of MR Brain Images”, IEEE Trans. Med. Imaging, 25(9):1145-1157, 2006. Wu, Qi, Shen, “Learning Best Features and Deformation Statistics for Hierarchical Registration of MR Brain Images”, IPMI 2007. 0.02 0.01 0 Error 2mm 0.66mm 0.95mm Improving: Statistically-constrained HAMMER HAMMER Registration Template Statistical Model of Deformations, using waveletPCA Subject Normal brain deformation captured from 150 subjects Xue, Shen, et al., “Statistical Representation of High-Dimensional Deformation Fields with Application to Statistically-Constrained 3D Warping”, Medical Image Analysis, 10:740-751, 2006. Improving: Statistically-constrained HAMMER Results: • More smooth deformations: • Detection on simulated atrophy: Comparison of Histograms of Jacobian Determinants 2.0% HAMMER SMD+HAMMER Percentage 1.5% 1.0% 0.5% 0.0% 0 1 2 3 4 Jacobian Determinant HAMMER SMD+HAMMER White Matter Lesion (WML) Segmentation WML Segmentation • WMLs are associated with cardiac and vascular disease, and may lead to different brain diseases, such as MS. • Manual delineation • Computer-assisted segmentation - Fuzzy-connection Multivariate Gaussian Model Atlas based normal tissue distribution model KNN based lesion detection • Lao, Shen, et al "Computer-Assisted Segmentation of White Matter Lesions in 3D MR images Using Support Vector Machine", Academic Radiology, 15(3):300-313, March 2008. Our approach • Image property: serious intensity overlap in WMLs T2 T1 WML PD FLAIR Attribute Vector • Attribute vector for each point v FLAIR PD T2 T1 Neighborhood Ω (5x5x5mm) F v I tm | tm vm , m {T1 , T2 , PD, FLAIR} • SVM To train a WML segmentation classifier. • Adaboost To adaptively weight the training samples and improve the generalization of WML segmentation method. Overview of Our Approach Co-registration Manual Segmentation Skull-stripping Training SVM model via training sample and Adaboost Intensity normalization Pre-processing False positive elimination Post-processing Training Voxel-wise evaluation & segmentation Testing Results Results – 45 Subjects 10 for training, and 35 for testing • Paired Spearman Correlation (SC) Gold standard (rater 1) Gold standard (rater 1) Rater 2 Computer Mean+dev. of the lesion volume 1.0 0.95 0.79 1494+/-3416 mm3 mm3 Rater 2 0.95 1.0 0.74 2839+/-6192 Computer 0.79 0.74 1.0 1869+/-3400 mm3 • Coefficient of variation (CV) Coefficient of Variation Rater 1 189% Rater 2 218% Double To investigate the variation of the lesion load’s distribution of the 35 evaluated subjects Defined as CV=/. Close Computer 182% • Lao, Shen, et al "Computer-Assisted Segmentation of White Matter Lesions in 3D MR images Using Support Vector Machine", Academic Radiology, 15(3):300-313, March 2008. Improvement in this project • Improve the robustness of multi-modality image registration (for T1/T2/PD/FLAIR) by using a novel quantitative and qualitative measurement for mutual information, where salient points will be considered more during the registration. • Design region-adaptive classifiers, in order to allow each classifier for capturing relative simple WML intensity pattern in each region; we will also develop a WML atlas for guiding the WML segmentation. • Lao, Shen, et al "Computer-Assisted Segmentation of White Matter Lesions in 3D MR images Using Support Vector Machine", Academic Radiology, 15(3):300-313, March 2008. Conclusion Further develop HAMMER registration and WML segmentation algorithms improve their robustness and performance 3D Slicer Thank you! http://bric.unc.edu/IDEAgroup/ http://www.med.unc.edu/~dgshen/ IDEA