Automatic Segmentation of Multimodal Brain Tumor

0 downloads 0 Views 1MB Size Report
Boosted Multi-Scale Dictionaries for Image Compression. 3/12. Introduction (Cont.) Segmentation of glioma tumors: ✓ BraTS 2012: a dataset for glioma tumors.
Missing:
Automatic Segmentation of Multimodal Brain Tumor Images Based on Classification of Super-Voxels M. Kadkhodaei1, S. Samavi2, N. Karimi1, H.Mohaghegh1 S. M. R. Soroushmehr2, A. All2, K. Ward2, K. Najarian2 1Isfahan

University of Technology, Iran 2University of Michigan, Ann Arbor, USA

EMBC 2016 August 2016

Outline Introduction Proposed Method Experimental Results Conclusion

Boosted Multi-Scale Dictionaries for Image Compression Automatic Segmentation of Multimodal Brain Tumor Images Based on Classification of Super-Voxels

1/12 1/13

Introduction Glioma Tumor: the most frequent brain tumor in adults  30% of all brain and central nervous system tumors  80% of all malignant brain tumors

Low grade glioma:  Aggressive treatment is delayed.  Median survival rate.

High grade glioma:  Worse prognosis and aggressive treatments.  Survival rate of two years or less. Boosted Multi-Scale Dictionaries for Image Compression Automatic Segmentation of Multimodal Brain Tumor Images Based on Classification of Super-Voxels

2/12 2/13

Introduction (Cont.) Segmentation of glioma tumors:  BraTS 2012: a dataset for glioma tumors.  Different MRI modalities, i.e. Flair, T1, T1c, T2.

Discriminative methods:  Learn from a train image set.  Classify based on extracted features from images.

Generative methods:  Make use of prior information or models.  Generalize the model to the unseen images. Boosted Multi-Scale Dictionaries for Image Compression Automatic Segmentation of Multimodal Brain Tumor Images Based on Classification of Super-Voxels

3/12 3/13

Our Goal Our goal:  Segmenting glioma tumors from multimodal MR images.

Challenge:  Size, location, shape and appearance of glioma tumors are highly variable. Flair

T1c

T2

Our Solution: Segmentation Method  Automatic segmentation of glioma tumors using classification of super-voxels.

Boosted Multi-Scale Dictionaries for Image Compression Automatic Segmentation of Multimodal Brain Tumor Images Based on Classification of Super-Voxels

4/12 4/13

Proposed Method Our strategy uses the merits of:  Super-voxel segmentation.  Saliency detection algorithm. Flair Channel

T2 Channel

T1c Channel

Preprocessing

Preprocessing

Preprocessing

Image Channel Fusion

Super-voxel Segmentation Feature Extraction

Super -Voxel Labels

Saliency Feature

Texture Features

Classification

Tumor Region Mask

Boosted Multi-Scale Dictionaries for Image Compression Automatic Segmentation of Multimodal Brain Tumor Images Based on Classification of Super-Voxels

5/12 5/13

Proposed Method (Cont.) preprocessing:  Contrast enhancement.  Intensity normalization.

Image Channel Fusion:  A three stacked color image similar to RGB format.

pseudo RGB image

Boosted Multi-Scale Dictionaries for Image Compression Automatic Segmentation of Multimodal Brain Tumor Images Based on Classification of Super-Voxels

6/12 6/13

Proposed Method (Cont.) Super-voxel Segmentation:  A meaningful grouping of voxels k-means clustering.  Initial cluster centers: sampled in regular cubes of side-size S in three dimensions.  Update a distance function .  Iterates the algorithm n-times.  A super-voxel: final members of a cluster.

Boosted Multi-Scale Dictionaries for Image Compression Automatic Segmentation of Multimodal Brain Tumor Images Based on Classification of Super-Voxels

7/12 7/13

Proposed Method (Cont.) Saliency Features Extraction:  Salient object: an object which stands out from its neighbors.  In MR images: salient object = TUMOR!  The fused images fed into saliency detection algorithm.

Saliency map of images Boosted Multi-Scale Dictionaries for Image Compression Automatic Segmentation of Multimodal Brain Tumor Images Based on Classification of Super-Voxels

8/12 8/13

Proposed Method (Cont.) Texture Features Extraction :  Mean, variance, skewness, kurtosis, range, and entropy, of histogram of every super-voxel.  local binary pattern (LBP)  Laws’ masks and oriented edge detectors

Some of the features used in the feature extraction procedure Boosted Multi-Scale Dictionaries for Image Compression Automatic Segmentation of Multimodal Brain Tumor Images Based on Classification of Super-Voxels

9/12 9/13

Proposed Method (Cont.) Classification:  The pattern recognition neural network.  training function : resilient back propagation.  Two classes, i.e. tumor core and everything else.

 Last step: Applying Weighted Median Filter those of the saliency map.

aligns tumor edges to

Boosted Multi-Scale Dictionaries for Image Compression Automatic Segmentation of Multimodal Brain Tumor Images Based on Classification of Super-Voxels

10/12 10/13

Evaluation of Our Method 30 real MR images in the dataset of BraTS 2012.  High grade: 20  low grad: 10.  Each sequence : 130 to 150 slices.

red color: our segmentation, green color ground truth, yellow: overlap between our segmentation and ground truth

Boosted Multi-Scale Dictionaries for Image Compression Automatic Segmentation of Multimodal Brain Tumor Images Based on Classification of Super-Voxels

11/12 11/13

Experimental Results Evaluation of the proposed method in a three-fold crossvalidation manner. 𝐷𝑖𝑐𝑒 𝑠𝑐𝑜𝑟𝑒 =

2𝑇𝑃 𝑇𝑃 + 𝐹𝑃 + 𝑇𝑃 + 𝐹𝑁

AVERAGE DICE SCORE AND STANDARD DEVIATION (%) OF TUMOR DETECTION IN COMPARISON TO [E. GEREMIA ET AL.]

Methods Mean Proposed Method Std. Proposed Method Mean [E. Geremia et al.] Std. [E. Geremia et al.]

Low Grade 68 12 20 27

High Grade 59 25 58 27

E. Geremia, B. H. Menze, N. Ayache, “Spatial Decision Forests for Glioma Segmentation in Multi-Channel MR Images” in MICCAI Challenge on Multimodal Brain Tumor Segmentation. 2012 Oct.

Boosted Multi-Scale Dictionaries for Image Compression Automatic Segmentation of Multimodal Brain Tumor Images Based on Classification of Super-Voxels

12/12 12/13

Conclusion Automatic segmentation of brain tumors has a critical role in diagnosis procedure. We developed an algorithm using 3D super-voxels and saliency detection for the segmentation purposes. Assessment via dice score rate confirmed the effectiveness of the proposed classification approach.

Boosted Multi-Scale Dictionaries for Image Compression Automatic Segmentation of Multimodal Brain Tumor Images Based on Classification of Super-Voxels

13/12 13/13

Thanks for your attention

Boosted Multi-Scale Dictionaries for Image Compression

14/12