PVE for MRI Brain Tissue Classification Zeng Dong SLST, UESTC 6-9
PVE Partial Volume Effect
Contents Overview Method Nei. PVE Model MAP Result Discussion Conclusion
Overview 1-role Roles of Segmentation in qualitative in visualization
Overview 2 – difficulty difficulty inhomogeneous PVE hard soft segmentation Statistic MAP MRF
Overview 3 - PVE IEE95 NOISE MODELS based Sampling noise Material-dependent Noise PVE direct indirect
Direct Determine PVC directly IEEE91 multichannel
continuous IEE03 Only tow types
continuous 02 Fuzzy Markov discrete PVC
Indirect PV class Determine PVC based on PV voxels
continuous More Tow Not Multi-channal Discrete PVC Boundary voxels More accurate, more efficient
Method-Nei. PVE N: numbers of pixels Assume: mask
continuous K: numbers of pure g(.,.): Gaussian function
continue Observed is mixed with its nei.s meanly during sampling Nei. Size M
continue
continue L kinds of mixed types: Mixed set Assume A mixed type
continue
PVE Segmentation MAP Y: observed images X: segmentation images
Likelihood term Assume that the intensity at voxel i does not depend on the tissue content of the other voxels.
Prior Assume X is MRF on nei. System C, and x is a realization RF X Z: Partition function U(x): Potential energy function
continue
continue
continue
ICM Iterative Constrained Mode local combination
continue
continue
continue 1. Init X (beta = 0, M=1) 2. Update mixel mean and variance 3. ICM 4. goto 2
Result Brain tissue classification K=3: CSF, GM, WM N=1,3,7,27
continue Generate Mixed types: for CSF=M:0 for GM=(M-CSF):0 { WM=M-CSF-GM … }
continue Example(K=3,M=7) Reduce: Not tow maximization( 3) CSF !=0 && GM != 0 (18) CSF GM WM
continue Ori seg Ori Seg with b Seg without b
continue compare 数据 100_23 1_24 11_3 110_3 111_2 112_2 12_3 13_3 15_3* 17_3* CSF GM WM 数据 17_3 191_3 202_3 205_3 CSF GM WM Mean: csf 20% GM 86% WM :83%
Discussion 4-1 Mixel of CSF and WM 0: : : : : : : : : : : mix mean and var 0: : : : : : : : : : : : : : : : : : : : : : : :
continous Mixel of CSF and WM
Discussion 4-2 Measure Parameter estimation csf mean : csf var : gm mean : gm var : wm mean : wm var : csf mean : csf var : gm mean : gm var : wm mean : wm var : Result para:Ori para Csf mean: Csf var: gm mean: gm var: wm mean: wm var: Init para(ML)
Discussion 4-3 Prior Parameter estimation
Discussion 4-4 Intensity inhomogeneous
Conclusion More accurate, more efficient Unify framework generalization
Thanks