Automatic Generation of Training Data for Brain Tissue Classification from MRI

Transcription

1 Automatic Generation of Training Data for Brain Tissue Classification from MRI Chris A. COCOSCO, Alex P. ZIJDENBOS, and Alan C. EVANS McConnell Brain Imaging Centre Montreal Neurological Institute McGill University COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.1/22

2 Brain tissue classification: the procedure of labeling each image voxel as a tissue class. T1 MRI T2 MRI PD MRI 4 classes: CSF, Grey-matter, White-matter, background COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.2/22

3 Preview: Main problem: subject different than anatomical model. COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.3/22

4 Preview: Main problem: subject different than anatomical model. COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.3/22

5 Outline: Requirements Existing methods Our method Validation and results COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.4/22

6 Target application: Quantitative measurements, such as: normalized tissue/structure volume atrophy measures voxel-based morphometry cortical surface cortical thickness... Studies on a large number of subjects ( ); data acquired at many different sites. COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.5/22

7 Requirements: Tissue classification method should be: ACCURATE application: quantitative measurements COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.6/22

8 Requirements: Tissue classification method should be: ACCURATE application: quantitative measurements FULLY AUTOMATIC reproducibility; many datasets to process COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.6/22

9 Requirements: Tissue classification method should be: ACCURATE application: quantitative measurements FULLY AUTOMATIC reproducibility; many datasets to process ROBUST against variability in: subject brain s morphology MRI data: image contrast, artifacts,... COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.6/22

10 Existing methods: Kamber et al. (IEEE TMI 95) Van Leemput (IEEE TMI 99), Ashburner ( SPM-99 ) all use a probabilistic brain anatomy atlas problems with brain anatomies significantly different than atlas. COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.7/22

11 Existing methods: EM EM-style schemes by Van Leemput (IEEE TMI 99), Ashburner ( SPM-99 ). Drawback: assume multi-variate Normal ( Gaussian ) intensity distributions. poor assumption for multi-spectral brain amri? biology, acquisition artifacts... COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.8/22

12 Existing methods: Kamber Kamber (IEEE TMI 95) used Tissue Probability Maps (TPM), defined in a stereotaxic space. 1. subject MRI spatially registered to stereotaxic space (linear registration) 2. select MRI intensity samples from spatial locations very likely to contain a given tissue type 3. use these samples to train a supervised classifier (such as: Bayes, neural network, knn,... ). COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.9/22

13 Stereotaxic space TPM: Subject MRI: TPM: CSF = 0% GM = 4% WM = 96% 0% 100% COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.10/22

14 Stereotaxic space TPM: young normal elderly, Alzheimer s Disease young normal population = CSF = 0% GM = 4% WM = 96% 0% 100% COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.10/22

15 Training samples selection: choose spatial locations with TPM value τ τ = CSF Grey matter White matter COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.11/22

16 Training samples selection: choose spatial locations with TPM value τ τ = CSF Grey matter White matter lower τ desired for more spatial coverage robust estimation of tissue intensity distributions. higher τ desired for reducing wrong class guesses. COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.11/22

17 Our novel method: TPM s MRI(s) "raw" samples PRUNING pruned samples knn supervised classifier Classification Pruning : removal of samples with incorrect class labels. COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.12/22

18 Our novel method: accommodates subject anatomies significantly different than model non-parametric: no assumptions about feature space (intensity) distributions allows for a lower TPM τ better estimation of intensity distributions accuracy, robustness COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.13/22

19 Pruning method: raw samples in feature space: PD T1 COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.14/22

20 Pruning method: [Step 1] PD Minimum Spanning Tree T1 COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.14/22

21 Pruning method: [Step 2] PD edge (i, j) is removed if length(i, j) > T A(i) or if length(i, j) > T A(j), where A(i) = average length of all other edges incident on node i T1 COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.14/22

22 Pruning method: [Step 2] PD edge (i, j) is removed if length(i, j) > T A(i) or if length(i, j) > T A(j), where A(i) = average length of all other edges incident on node i T = 9.2 T1 COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.14/22

23 Pruning method: [Step 2] PD edge (i, j) is removed if length(i, j) > T A(i) or if length(i, j) > T A(j), where A(i) = average length of all other edges incident on node i T = 9.2 T1 COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.14/22

24 Pruning method: [Step 2] PD edge (i, j) is removed if length(i, j) > T A(i) or if length(i, j) > T A(j), where A(i) = average length of all other edges incident on node i T = 5.3 T1 COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.14/22

25 Pruning method: [Step 2] PD edge (i, j) is removed if length(i, j) > T A(i) or if length(i, j) > T A(j), where A(i) = average length of all other edges incident on node i T = 5.3 T1 COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.14/22

26 Pruning method: [Step 2] PD cluster = a connected component of graph. CSF cluster = cluster with most CSF samples.... T = 5.3 Stop when BG, CSF, GM, WM clusters are distinct. T1 COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.14/22

27 Pruning method: [Step 2] PD cluster = a connected component of graph. CSF cluster = cluster with most CSF samples.... T = 4.9 Stop when BG, CSF, GM, WM clusters are distinct. T1 COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.14/22

28 Pruning method: [Step 2] PD cluster = a connected component of graph. CSF cluster = cluster with most CSF samples.... T = 4.9 Stop when BG, CSF, GM, WM clusters are distinct. T1 COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.14/22

29 Pruning method: [Step 3] PD discard samples that are not found in correct cluster. T = 4.9 T1 COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.14/22

30 Validation: simulated MRI T1-T2-PD multi-spectral simulated MRI, 10 different elderly brain phantoms young-normal model (TPM), N=53 quantitative evaluation: Kappa = chance-corrected similarity measure between two image labelings (classifications). classified image gold standard (phantom) COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.15/22

31 Validation: simulated MRI PRUNED 0.85 Kappa NOT PRUNED (RAW) TPM τ (elderly brain simulated MRI, young-normal model) COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.15/22

32 Validation: real MRI 1. young & normal individual (T1+T2+PD, and also T1 only), against full-brain manual segmentation Ischemia patients (T1+T2+PD) Alzheimer s Disease (A.D.) elderly patients (T1+T2+PD). COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.16/22

33 Results: Ischemia T1 MRI T2 MRI PD MRI raw (not pruned): pruned: COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.17/22

34 Results: Ischemia T1 MRI T2 MRI PD MRI raw (not pruned): pruned: COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.17/22

35 Results: Ischemia T1 MRI T2 MRI PD MRI raw (not pruned): pruned: COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.17/22

36 Results: A.D. elderly T1 MRI T2 MRI PD MRI raw (not pruned): pruned: COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.18/22

37 Results: A.D. elderly T1 MRI T2 MRI PD MRI raw (not pruned): pruned: COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.18/22

38 Future work: Current limitations: inherent to intensity-only, discrete classification. due to overlap of tissue intensity distributions (brain biology, MRI partial volume). COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.19/22

39 Future work: Current limitations: inherent to intensity-only, discrete classification. due to overlap of tissue intensity distributions (brain biology, MRI partial volume). also use voxel neighbourhood information (e.g. image gradient),... COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.19/22

40 Summary: fully automatic brain tissue classification procedure. robust against anatomical variability. non-parametric: no assumptions about tissue intensity distributions ( robust against imaging artifacts). validated qualitatively and quantitatively on simulated and on real MRI data. COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.20/22

41 Acknowledgements: John Sled, Steve Robbins, Peter Neelin, Godfried Toussaint, Noor Kabani, Louis Collins, Jean-François Mangin, Jason Lerch, Jennifer Campbell, Ives Levesque, Najma Khalili. the anonymous MICCAI-2002 reviewers. Alma Mater Student Travel Fund, Faculty of Graduate Studies and Research, McGill University, Montreal. COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.21/22

42 Automatic Generation of Training Data for Brain Tissue Classification from MRI Chris A. COCOSCO, Alex P. ZIJDENBOS, and Alan C. EVANS McConnell Brain Imaging Centre Montreal Neurological Institute McGill University COCOSCO, ZIJDENBOS, EVANS: Automatic Generation of Training Data for Brain Tissue Classification from MRI p.22/22