EMAtlasBrainClassifier By Kilian Maria Pohl
Kilian M. Pohl Pipeline 2 Atlas Alignment 3 EM Segmentation 1 Intensity Normalization
Kilian M. Pohl Intensity Normalization BeforeAfter
Kilian M. Pohl Registration Non-Rigid Registration Based on Maxwell Demons [Guimond 99] Accurate but computationally expensive Affine Registration Based on Mutual Information [van Leemput 99] Robust but generally not as accurate Joint Registration & Segmentation Based on work by [Pohl 05] Robust, accurate, and computationally expensive
Kilian M. Pohl Atlas Alignment Pre-selected Subject Training Subjects Segmentations Spatial Prior Registration Resampling
Kilian M. Pohl Segmentation
Kilian M. Pohl Requirements Minimum of 1 Gig of RAM (1.5 h on PC) Aligned T1 and T2 of the whole head Optimized for –isotropic T2 (Dimension mm x mm x 3mm ) –SPGR (Dimension mm x mm x 1.5mm ) For further information see
Kilian M. Pohl Segmentation of 31 Structures
Kilian M. Pohl Define Model Labelmap T Inhomogeneity B Parameter Data Image I Registration R Shape S
Kilian M. Pohl Define Model Inhomogeneity B Labelmap T Inhomogeneity B Parameter Data Image I Registration R Shape S Hierarchy H
Kilian M. Pohl Observed Data (ROI) EM Hierarchical Implementation Image Prior HierarchyLabelmap
Kilian M. Pohl Level 1 Prior Information IMAGE BGICC CSFGMWM
Kilian M. Pohl Level 2 IMAGE ICC Current Parameter CSFGMWM ROI
Kilian M. Pohl Segmentation of 31 Structures
Kilian M. Pohl The EM Algorithm Expectation Step: Calculate lower bound Maximization Step: Find maxima of bound Graph from Minka