Motivation IM & BC SR Results Contributions Limitations
Reconstruction of fetal brain MRI with
intensity matching and
complete outlier removal
M Kuklisova-Murgasova, G Quaghebeur, MA Rutherford,JV Hajnal, JA Schnabel
Med. Image Anal., vol. 16, no. 8, pp. 1550-64, 2012
presented by Michael Ebner
June 10, 2015
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 1
Motivation IM & BC SR Results Contributions Limitations
Agenda
1 Motivation
2 Intensity matching and bias correction
3 Super-resolution reconstruction
4 Results
5 Contributions
6 Limitations
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 2
Motivation IM & BC SR Results Contributions Limitations
Motivation
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 3
Motivation IM & BC SR Results Contributions Limitations
Motivation
1. Inhomogeneity of magnetic field
Relative position to scannerchanges due to motion
Intensity bias
Intensity matchingand
bias correction
2. Motion corrupted slices
Identification
Possible exclusion
Robust statistics
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 4
Motivation IM & BC SR Results Contributions Limitations
Intensity matchingand
bias correction
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 5
Motivation IM & BC SR Results Contributions Limitations
Intensity matching and bias correction
Voxels of observed k-th slice yk = {yk,j}Nkj=1
Voxels of unkown volume x = {xi}Ni=1
Bias field model
Bias fields bk = {bk,j}Nkj=1
Scaling factor sk
y ∗k,j := sk e
−bk,jyk,j
Slice acquisition model
Scaled and bias corrected slices y∗kSpatially aligned discretized PSF Mk
y∗k ≈Mkx
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 6
Motivation IM & BC SR Results Contributions Limitations Optimization problem Robust statistics Algorithm
Super-resolution reconstruction
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 7
Motivation IM & BC SR Results Contributions Limitations Optimization problem Robust statistics Algorithm
Super-resolution reconstruction
Point spread function (Matrix Mk):
In-plane: Gaussian withFWHM = 1.2×in-plane resolution
Through-plane: Gaussian withFWHM = slice thickness
Optimization problem:
ek := y∗k −Mkx→ min
Regularized cost function:∑k,j
e2k,j + λR(x)→ min
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 8
Motivation IM & BC SR Results Contributions Limitations Optimization problem Robust statistics Algorithm
Robust statistics
1 Classification of voxels in inliers and outliers
2 Posterior probability of voxel being classified as inlier pk,j3 Weighted cost function:∑
k,j
pk,je2k,j + λR(x)→ min
4 Gradient descent step:
x(n+1)i = x
(n)i + α
∑k,j
mkj ,i
(n)p(n)k,j e
(n)k,j + αλ
∂
∂xiR(x(n))
5 Classification of slices as inliers and outliers ( pslicek )
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 9
Motivation IM & BC SR Results Contributions Limitations Optimization problem Robust statistics Algorithm
Advantage of EM robust statistics approach
ek = y∗k −Mkx
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 10
Motivation IM & BC SR Results Contributions Limitations Optimization problem Robust statistics Algorithm
Overview of algorithm
y∗k,j = sk e
−bk,j yk,j , y∗k ≈ Mkx, ek = y∗k −Mkx
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 11
Motivation IM & BC SR Results Contributions Limitations Simulation Fetal data Leave-one-out
Results
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 12
Motivation IM & BC SR Results Contributions Limitations Simulation Fetal data Leave-one-out
Evaluation of method
1 Simulated experiments
Preterm neonate with GA = 27 weeksT2-weighted fast-spin echo, 3 TTR = 8620 ms, TE = 169 ms, voxel sizes = (1 mm)3
Adding of artificial motion corruption
2 Reconstruction of clinical fetal data
10 clinical fetal brain MRI: GA between 20 and 37 weeksT2-weighted single-shot turbo spin echo sequence, 1.5 TTR = 32 805 ms, TE = 100 ms, flip angle = 90◦
Voxel sizes = 0.75 mm× 0.75 mm× 3 (or 4) mm
3 Leave-one-out analysis of clinical data reconstruction
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 13
Motivation IM & BC SR Results Contributions Limitations Simulation Fetal data Leave-one-out
Simulation
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 14
Motivation IM & BC SR Results Contributions Limitations Simulation Fetal data Leave-one-out
Simulation
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 15
Motivation IM & BC SR Results Contributions Limitations Simulation Fetal data Leave-one-out
Reconstruction of clinical fetal data
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 16
Motivation IM & BC SR Results Contributions Limitations Simulation Fetal data Leave-one-out
Leave-one-out strategy
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 17
Motivation IM & BC SR Results Contributions Limitations
Contributions
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 18
Motivation IM & BC SR Results Contributions Limitations
Contributions
Comprehensive reconstruction algorithm of fetal brainMRI
Novel intensity matching and bias field correction ofacquired 2D slices
Novel robust statistics to exclude identified misregisteredor corrupted voxels and slices
Multi-resolution approach
Estimation that five stacks of thick-slice data (3 mm to4 mm) with a gap (0.5 mm) sufficient for goodreconstruction of fetal brain
New comparison method: Leave-one-out analysis
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 19
Motivation IM & BC SR Results Contributions Limitations
Limitations
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 20
Motivation IM & BC SR Results Contributions Limitations
Limitations
More rigorous mathematical framework missing
Complexity of algorithms
Precomputed low resolution representation of PSF pervoxel and subsequent linear interpolation to keepcomputational time reasonable
Iteration between volume reconstruction andslice-to-volume computationally expensive
Depends on external slice-to-volume registration algorithm
Target stack chosen and segmented manually
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 21
Motivation IM & BC SR Results Contributions Limitations
Thanks!
Questions?
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 22
Motivation IM & BC SR Results Contributions Limitations
Appendix
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 23
Motivation IM & BC SR Results Contributions Limitations
Simulated stacks
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 24
Motivation IM & BC SR Results Contributions Limitations
Bias field
Reconstruction of fetal brain MRI with intensity matching . . . Kuklisova-Murgasova et al. (2012) 25