3D Segmentation of Rodent Brain Structures Using Active Volume Model With Shape Priors

Shaoting Zhang1, Junzhou Huang1, Mustafa Uzunbas1, Tian Shen2, Foteini Delis3, Xiaolei Huang2, Nora Volkow3, Panayotis Thanos3, Dimitris Metaxas1

1 CBIM, Rutgers, The State University of New Jersey, Piscataway, NJ, USA2 Computer Science and Engineering Department, Lehigh University, PA, USA

3 Brookhaven National Laboratory, NY, USA

Motivations

• Rodents are often used as models of human disease.

• Use Magnetic Resonance Microscopy (MRM) to get 3D image for rodent brain.

• 3D segmentation of brain regions based on MR images of the rodent brain.

• Deformable model based segmentation.

Motivations

• Three challenges: 1) unclear boundary, 2) complex textures, 3) complex shape.

Relevant work

• Deformable model based segmentation– Deformable Models with Smoothness Constraints• Active contour [M. Kass, IJCV’88]• Gradient Vector Flow [C. Xu, TIP’98]• Deformable Superquadrics and Metamorphs [Metaxas

91,92; Huang, 08]

– Priors from training data• ASM [T.F. Cootes, CVIU’95]• 3D ASM [Y. Zheng, TMI’08]

4

Proposed method-Framework

Offline Learning

Geometry Processing

Shape Registration

Training Shapes PCA Shape

Statistics

Runtime Segmentation System

Input Image

Image Alignment

Volumetric Deformation

Shape Constraint Result

Proposed method-Build Shape Statistics

• Geometry processing (decimation, detail-preserved smoothing)

Nealen, et.al.: LMO, GRAPHITE’06

Proposed method-Build Shape Statistics

• Shape registration using AFDM

…

Shen, et.al.: AFDM, TMI’01

Proposed method-Build Shape Statistics

• PCA analysis (mean and variance)

Cootes, et.al.: ASM, CVIU’95

Proposed method-Deformation module

• Evolution of probability density function computed from region information

Huang, et.al.: Metamorphs, PAMI’08

Proposed method-Deformation module

• 3D Finite Element Method (A3D·V=LV)

Metaxas 92, Shen, et.al.: Active Volume Model, CVPR’09

Proposed method-Deformation module

A3D (smoothness)

Sorkine, et.al.: Laplacian Mesh Processing, EG’05

Proposed method-Framework, revisit

Input Image Image

Alignment

3D Metamorphs (AVM)

ASM Shape Refinement Result

Mean Mesh

Initialization

Shape Statistics

Reference Image

Initialization

Deformation

Experiments

• Settings– Adult male Sprague-Dawley rats– 21.1T Bruker Biospin Avance scanner– FOV of 3.4 × 3.2 × 3.0mm, voxel size 0.08mm– Data: 2/3 training and 1/3 for testing– All normal cases– Segment the cerebellum, the left and right striatum.– C++ and Python2.6 and tested on a 2.40 GHz Intel

Core2 Quad computer with 8G RAM.

Experiments

• Cerebellum (complex texture and shape details)

Our method

No prior

Experiments

• Striatum (unobvious boundaries)

Our method

No prior

Experiments

• p: sensitivity; q: specificity; DSC: dice similarity coefficient; RE-V: relative error of volume magnitude.

TP/(TP+FN)TN/(TN+FP)2TP/(2TP+FP+FN)

Conclusions

• Proposed a segmentation framework using 3D Metamorphs based deformation module and ASM based shape prior module.

• It is particularly useful when there are a limited number of training samples.

• In the future, we will test this algorithm on a larger dataset and also investigate how to segment multiple structures simultaneously and effectively.

Thanks