Shaoting Zhang, Yiqiang Zhan, Yan Zhou, and Dimitris N. Metaxas

Efficient Sparse Shape Composition with its Applications in Biomedical Image Analysis:

An Overview

Shaoting Zhang, Yiqiang Zhan, Yan Zhou, andDimitris N. Metaxas

Outline• Introduction• Our methods• Experiments• Future work

2

IntroductionMotivations

• Segmentation (i.e., finding 2D/3D ROI) is a fundamental problem in medical image analysis.

• We use deformable models and shape prior. Shape representation is based on landmarks.

Chest X-ray

Lung CADLiver in low-dose whole-body CT

Rat brain structure in MR Microscopy

3

IntroductionMotivations

• Shape prior is important:– Automatic, accuracy,

efficiency, robustness.– Handle weak or misleading

appearance cues from image information.

– Discover or preserve complex shape details.

Segmentation system

4

IntroductionChallenges – shape prior modeling

• Good shape prior method needs to:– Handle gross errors of the input data.– Model complex shape variations.– Preserve local shape details.

5

IntroductionRelevant work – shape prior methods

• Handling gross errors. – RANSAC + ASM [M. Rogers, ECCV’02]– Robust Point Matching [J. Nahed, MICCAI’06]

• Complex shape variation (multimodal distribution).– Mixture of Gaussians [T. F. Cootes, IVC’97]– Patient-specific shape [Y. Zhu, MICCAI’09]

• Recovering local details.– Sparse PCA [K. Sjostrand, TMI’07]– Hierarchical ASM [D. Shen, TMI’03]

6

Our ApproachShape prior using sparse shape composition

• Our method is based on two observations:– An input shape can be approximately represented by a

sparse linear combination of training shapes.– The given shape information may contain gross errors, but

such errors are often sparse....

≈

7


• Formulation:–

• Sparse linear combination:–

...≈

...≈

Dense xSparse xAligned data matrix D

Weight x

Input y

Global transformationparameter

Number of nonzero elements

8

Global transformationoperator


• Non-Gaussian errors:–

9


• Why it works?– Robust: Explicitly modeling “e” with L0 norm constraint.

Thus it can detect gross (sparse) errors.– General: No assumption of a parametric distribution

model (e.g., a unimodal distribution assumption in ASM). Thus it can model complex shape statistics.

– Lossless: It uses all training shapes. Thus it is able to recover detail information even if the detail is not statistically significant in training data.

Zhang, Metaxas, et.al.: MedIA’11

10

Experiments2D lung localization in X-ray

• Setting:– Manually select landmarks for

training purpose.– 200 training and 167 testing.– To locate the lung, detect

landmarks, then predict a shape to fit them.

– Sensitivity (P), Specificity (Q), Dice Similarity Coefficient.

11

Experiments 2D lung localization in X-ray

• Handling gross errors

Detection PA ASM RASM NN TPS Sparse1 Sparse2

Procrustes analysisActive Shape ModelRobust ASMNearest NeighborsThin-plate-splineWithout modeling “e”

12

Proposed method


• Multimodal distribution

Detection PA ASM/RASM NN TPS Sparse1 Sparse2

13


• Recover local detail information

Detection PA ASM/RASM NN TPS Sparse1 Sparse2

14


• Sparse shape components

• ASM modes:

15

≈

0.5760

+ +

0.2156 0.0982


• Mean values (µ) and standard deviations (σ).

Left lung

Right lung

1)PA, 2)ASM, 3)RASM, 4)NN, 5)TPS, 6)Sparse1, 7)Sparse2

µσ

16

Future WorkDictionary Learning

17

... ≈... ...

Dictionary size = 128 #Coefficients = 800 #Input samples = 800

18

DInitialize D

Sparse CodingUse MP or BP

Dictionary UpdateColumn-by-Column by SVD computation

Use K-SVD to learn a compact dictionary Aharon, Elad, & Bruckstein (‘04)

YT

* The slides are from http://www.cs.technion.ac.il/~elad/talks/

Future WorkDictionary Learning

Future WorkOnline Update Dictionary

19

...≈

...

Future WorkMesh Partitioning for Local Shape Priors

20

Segmentation system

Lung Heart

Rib

Colon

Abdomen

Thanks!Questions and comments

21

Shaoting Zhang, Yiqiang Zhan, Yan Zhou, and Dimitris N. Metaxas

Documents

Transcript of Shaoting Zhang, Yiqiang Zhan, Yan Zhou, and Dimitris N. Metaxas