1. Medical Image Synthesis via Monte Carlo Simulation James Z. Chen, Stephen M. Pizer, Edward L. Chaney, Sarang Joshi Medical Image Display & Analysis Group, University of North Carolina, Chapel Hill, U.S.A. Methodology Overview References Gerig, G., M. Jomier, M. Chakos (2001). “Valmet: A new validation tool for assessing and improving 3D object segmentation.” Proc. MICCAI 2001, Springer LNCS 2208: 516-523. Cootes, T. F., A. Hill, C.J. Taylor, J. Haslam (1994). “The Use of Active Shape Models for Locating Structures in Medical Images.” Image and Vision Computing 12(6): 355-366. Rueckert, D., A.F. Frangi, and J.A. Schnabel (2001). “Automatic Construction of 3D Statistical Deformation Models Using Non-rigid Registration.” MICCAI 2001, Springer LNCS 2208: 77-84. Christensen, G. E., S.C. Joshi and M.I. Miller (1997). “Volumetric Transformation of Brain Anatomy.” IEEE Transactions on Medical Imaging 16: 864-877. Joshi, S., M.I. Miller (2000). “Landmark Matching Via Large Deformation Diffeomorphisms.” IEEETransactions on Image Processing. Maes, F., A. Collignon, D. Vandermeulen, G. Marchal, P. Suetens (1997). “Multi-Modality Image Registration by Maximization of Mutual Information.” IEEE-TMI 16: 187-198. Pizer, S.M., J.Z. Chen, T. Fletcher, Y. Fridman, D.S. Fritsch, G. Gash, J. Glotzer, S. Joshi, A. Thall, G. Tracton, P. Yushkevich, and E. Chaney (2001). “Deformable M-Reps for 3D Medical Image Segmentation.” IJCV, submitted. Fiducial Point Models (FPM) and Their Construction FPM is a 3D point distribution model (PDM) used to represent the diffeomorphism across a class of target anatomic objects. The minimal number and the optimal locations of the surface points in a FPM are determined algorithmically to give a compact descriptor for shape modeling. FPM is free from subjective bias and can be used for multi-object complex modeling. Algorithm for FPM Construction 1.Register I 0 to each I t via I t = H t (I 0 ). Initialize {F m } with a few geometrically salient points on S 0 ; 2.Apply the training warp function H t on {F m } to get the warped fiducial points: F m,t = H t (F m ); 3.Reconstruct the diffeomorphic warp field H' t for the entire image volume based on the displacements {F m,t – F m }; 4.For each training case t, locate the point p t on the surface of S 0 that yields the largest discrepancy between H t and H' t ; 5.Find most discrepant point p over the point set {p t } established from all training cases. Add p to the fiducial point set; 6.Return to step 2 until a pre-defined optimization criterion has been reached. Example of FPM: the Human Right Kidney 36 clinical CT images in the training set. Optimization criterion: inter-human segmentation comparison – 94% volume overlap, 0.58 surface distance (in voxel units). Clinical or synthetic? Classify the following images. ABC FED Key for quiz in lower left: B,C,D,E are synthetic. The Goal Develop a methodology for generating realistic, synthetic medical images for segmentation method evaluation. Premise Any medical image of a selected anatomic region can be formed by deforming the geometry of an atlas together with a selection of image intensities from a training set. 1.Represent 3D anatomic object(s) compactly using a Fiducial Point Model (FPM). 2.Train the statistics of FPMs using clinical images. One image forms the template that has been carefully segmented. 3.Synthesize shape deformations by sampling the distribution of the FPM and interpolating a template diffeomorphism H s. 4.Borrow image intensity pattern (contrast, noise, etc.) by deforming a randomly selected training image to the template. 5.Deform the borrowed intensity pattern by H s to produce synthetic image. Deform the template’s segmentation by H s to produce image’s ground-truth segmentation. FPM: 88 points Template: I 0, S 0 Training: {I t } Template transformation H s,t = H s ·H t -1 Synthetic images from template I t : I s ' = H s,t (I t ) S s ' = S s = H s (S 0 ) Registration I t  H t (I 0 ) Analyze {H t } via PCA on {FPM displacements} to yield Pr(FPM) Generate random diffeomorphic H s : Sample Pr(FPM) and interpolate Synthetic images from template I 0 : I s = H s (I 0 ) S s = H s (S 0 ) Training Sampling Template Synthesized from H s (I 0 ) Synthesized from H s,t (I t ) Synthesis of CT images of kidney region Axial Sagittal Coronal Principal component analysis: first seven modes of FPM cover 88% of the total variation. Results: Generate FPM