Date post: | 15-Jan-2016 |
Category: |
Documents |
Upload: | octavia-ross |
View: | 215 times |
Download: | 0 times |
IDEA Image Display, Enhancement, and Analysis
Department of Radiology and BRIC, UNC-Chapel Hill
LINKS: Learning-based multi-source IntegratioN frameworK for Segmentation of infant brain images
Li Wang, Yaozong Gao, Feng Shi, Gang Li, Dinggang ShenPresented by Li Wang
09-18-2014
Department of Radiology and BRIC, UNC-Chapel Hill
Content
Motivation Proposed method Experimental results Conclusion
Department of Radiology and BRIC, UNC-Chapel Hill
Motivation
Limitations of multi-atlas label fusion1. nonlinear registrations2. simple intensity patch3. equal weight for different modality
Fractional anisotropy (FA) was calculated from Diffusion MRIs.
Our proposed work will 1. linear registrations2. appearance features and context features3. adaptive weights for different modality
2-weeks
6-months
12-months
T1 T2 FA Manual segmentation
Department of Radiology and BRIC, UNC-Chapel Hill
Flowchart of our proposed work
Context features
Appearance features
Classifier 2
Ground truth
T1 T2 FAAppearance
features
Probability mapsSequence classifier
Feature vectors
Context features
Appearance features
Classifier τ
Haar-like featuresClassifier 1
Random forests
Department of Radiology and BRIC, UNC-Chapel Hill
Result of an unseen target subject
T1 T2 FA
Original images
Iteration 1
Iteration 2
Iteration 10
Ground truth
Department of Radiology and BRIC, UNC-Chapel Hill
Probabilities of training image by the random forest
Post-processing: Anatomical constraint
To deal with the possible artifacts due to independent voxel-wise classification, we use patch-based sparse representation to impose an anatomical constraint [1] into the segmentation.
1. Wang, L., Shi, F., Gao, Y., Li, G., Gilmore, J.H., Lin, W., Shen, D., 2014. Integration of sparse multi-modality representation and anatomical constraint for isointense infant brain MR image segmentation. NeuroImage 89, 152-164.
Ground truth of training images
Probabilities of target image by the random forest
𝛼1 𝛼 𝑖
Without anatomical With anatomical Ground truth
Department of Radiology and BRIC, UNC-Chapel Hill
Dataset
Dataset 1: UNC 119 infants consisting of 26, 22, 22, 23, and 26 subjects at 0-, 3-, 6-, 9- and 12-months of age, respectively.
Dataset 2: NeoBrainS12 MICCAI2012 Challenge. Dataset 3: SATA MICCAI2013 Challenge.
Department of Radiology and BRIC, UNC-Chapel Hill
Importance of the context features
Iterations Iterations Iterations
Department of Radiology and BRIC, UNC-Chapel Hill
Importance of the multi-source
Department of Radiology and BRIC, UNC-Chapel Hill
Dataset 1: UNC 119 infants
(a) Majority voting (MV)(b) Nonlocal label fusion [1](c) Atlas forest [2](d) Patch-based sparse labeling [3](e) Proposed1 (Random forest)(f) Proposed2 (Random forest + Anatomical constraint)
1. Coupé, P., Manjón, J., Fonov, V., Pruessner, J., Robles, M., Collins, D.L., 2011. Patch-based segmentation using expert priors: Application to hippocampus and ventricle segmentation. NeuroImage 54, 940-954.
2. Zikic, D., Glocker, B., Criminisi, A., 2013. Atlas Encoding by Randomized Forests for Efficient Label Propagation. MICCAI 2013, pp. 66-73.3. Wang, L., Shi, F., Gao, Y., Li, G., Gilmore, J.H., Lin, W., Shen, D., 2014. Integration of sparse multi-modality representation and anatomical
constraint for isointense infant brain MR image segmentation. NeuroImage 89, 152-164.
Department of Radiology and BRIC, UNC-Chapel Hill
T1 T2 FA
(a) Majority voting(b) Nonlocal label fusion
(c) Atlas forest(d) Patch-based sparse labeling
Ground truth
(e) Proposed1 (f) Proposed2
Slice comparisons
Segmentation
Difference maps with the ground truth
Department of Radiology and BRIC, UNC-Chapel Hill
Inner surface comparisons
(a) Majority voting(b) Nonlocal label fusion
(c) Atlas forest(d) Patch-based sparse labeling
(e) Proposed1 (f) Proposed2 (g) Ground truth
Department of Radiology and BRIC, UNC-Chapel Hill
Quantitative measurement
Methods MV Nonlocal
label fusion Atlas Forest
Patch-based Sparse labeling
Proposed1 Proposed2
Time cost 1h 1.2h 12m 2h 5m 1.8h
WM
0 81.6±0.28 89.0±0.74 88.9±0.60 89.7±0.59 91.7±0.64 92.1±0.62 3 76.6±1.48 85.0±1.21 85.1±1.33 85.3±1.71 88.8±1.09 89.1±0.95 6 80.1±0.83 83.6±0.80 82.1±0.91 84.2±0.78 86.4±0.79 87.9±0.68 9 79.2±0.98 86.1±2.00 84.2±1.34 87.1±1.89 89.0±0.78 89.4±0.56 12 82.5±1.05 88.6±1.22 87.2±1.29 90.3±1.42 90.7±0.74 91.8±0.65
GM
0 78.6±1.02 85.1±0.78 87.1±0.76 86.7±0.81 89.6±0.66 90.8±0.42 3 77.3±1.42 83.4±0.78 85.5±1.12 85.3±0.51 88.1±1.00 88.3±0.90 6 79.9±1.04 83.9±0.83 83.1±0.93 84.8±0.77 88.2±0.77 89.7±0.59 9 83.6±0.69 88.1±0.75 87.4±0.66 87.4±0.54 90.0±0.49 90.3±0.54 12 84.9±1.01 89.3±0.90 88.8±1.02 88.9±0.57 90.3±0.74 90.4±0.68
CSF
0 76.6±1.57 80.2±1.87 77.7±4.52 76.1±2.59 83.9±2.20 84.2±2.02 3 80.6±1.55 84.1±1.88 82.4±2.17 80.1±1.10 83.7±1.52 85.4±1.49 6 71.2±0.71 79.2±1.69 86.7±1.16 83.0±0.77 92.7±0.63 93.1±0.55 9 68.7±1.27 80.6±2.40 84.1±1.57 81.0±2.27 85.8±1.53 86.7±1.09 12 65.2±3.69 81.5±1.66 83.6±1.83 81.7±2.59 84.1±1.90 85.2±1.69
Department of Radiology and BRIC, UNC-Chapel Hill
Dataset 2: NeobrainS12 MICCAI Challenge
2 training images with the manual segmentations. 3 target images for testing.
Department of Radiology and BRIC, UNC-Chapel Hill
Our results of 3 target images
Department of Radiology and BRIC, UNC-Chapel Hill
Quantitative measurement
Table 1. Dice ratios (DC) and modified Hausdorff distance (MHD) of different methods on NeoBrainS12 MICCAI Challenge data. (Bold indicates the best performance)
WM CGM BGT BS CB CSF
Placed Team Name DC MHD DC MHD DC MHD DC MHD DC MHD DC MHD UNC-IDEA 0.92 0.35 0.86 0.47 0.92 0.47 0.83 0.9 0.92 0.5 0.79 1.18 1
Imperial 0.89 0.70 0.84 0.73 0.91 0.8 0.84 1.04 0.91 0.7 0.77 1.55 2 Oxford 0.88 0.76 0.83 0.61 0.87 1.32 0.8 1.24 0.92 0.63 0.74 1.82 3 UCL 0.87 1.03 0.83 0.73 0.89 1.29 0.82 1.3 0.9 0.92 0.73 2.06 4
UPenn 0.84 1.79 0.80 1.01 0.8 4.18 0.74 1.96 0.91 0.85 0.64 2.46 5
http://neobrains12.isi.uu.nl/mainResults_Set1.php
Department of Radiology and BRIC, UNC-Chapel Hill
Dataset 3: SATA MICCAI2013 Challenge
35 training images with the 14 ROIs in subcortical regions. 12 target images for testing.
Department of Radiology and BRIC, UNC-Chapel Hill
Our results on one target image
Department of Radiology and BRIC, UNC-Chapel Hill
Quantitative measurement
Team Name Submission Date/Time
Mean (Median) DSC
Mean (Median) Hausdorff Distance (mm)
UPENN_SBIA_MAM 12-Jul-2013 0.8686 (0.8772) 3.3043 (3.1006)
PICSL 02-Jul-2013 0.8663 (0.8786) 3.5381 (3.2369)
LINKS 04-May-2014 0.8613 (0.8722) 3.6453 (3.3637)
deedsMIND 12-Jul-2013 0.8402 (0.8573) 4.1027 (3.8983)
MSRC_AF_NEW 18-Feb-2014 0.8247 (0.8392) 3.8437 (3.6799)
MSRC_AF_NEW_STAPLE 18-Feb-2014 0.8063 (0.8169) 4.6494 (4.3760)
deedsMIND no marginals 15-Jul-2013 0.7216 (0.7539) 6.1614 (5.5120)
Table 2. Dice ratios (DC) and Hausdorff distance (HD) of different methods on SATA MICCAI Challenge data.
http://masi.vuse.vanderbilt.edu/submission/leaderboard.html
Department of Radiology and BRIC, UNC-Chapel Hill
Conclusion
We have presented a learning-based method (LINKS) to effectively integrate multi-source images and the tentatively estimated tissue probability maps for infant brain image segmentation.
Experimental results on 119 infant subjects and MICCAI grand challenge show that the proposed method achieves better performance than other state-of-the-art automated segmentation methods.
Department of Radiology and BRIC, UNC-Chapel Hill
Thanks for your attention!
http://www.unc.edu/~liwa/
Google scholar