Quantitative Brain Structure Analysis on MR Images

1
Quantitative Brain Structure Analysis on MR
Images

• Zuyao Shan, Ph.D.
• Division of Translational Imaging Research
• Department of Radiological Sciences

2
Outline

• Introduction
• Cerebellum segmentation (Preliminary study)
• Cortical structure segmentation

3
Brain Segmentation

• With the ability to identify brain structures on
  MR images and to detect anatomic changes, the new
  volumetric tools aid in the diagnosis, treatment,
  and elucidation of changes associated with
  disease or abnormality.
• Registration based approaches
• Pros Straightforward tenet, robustness
• Cons Accuracy limited by match quality, mismatch
  leading to significant errors, relying on image
  only. One-one mapping may not existed, Speed
• Deformable model based approaches
• Pros Prior knowledge incorporated, high
  accuracy.
• Cons Good initialization needed, identification
  of landmarks

4
Brain Segmentation inter-personal variability

• More challenges in pediatric patients with brain
  tumors
• Removal of tissues
• Different stages of development
• An adequate method should cope with high
  inter-subject variability with high accuracy

5
Brain Segmentation Cerebellum

• Knowledge guided active contour
• Rigid-body registration good initialization
• Prior defined template Knowledge incorporated
• Active contour adjustment high accuracy,
  robustness

6
Brain Segmentation Cerebellum
Active contour (Snake) energy-minimizing spline
7
Brain Segmentation Cerebellum
Active contour (Cont.) Internal energy
8
Brain Segmentation Cerebellum
Active contour (Cont.) External energy
Distance
Sobel edge
detection
transform
9
Brain Segmentation Cerebellum
Visual inspection
10
Brain Segmentation Cerebellum
Visual inspection
11
Brain Segmentation Cerebellum

• Similarity evaluation
• Kappa index
• A vs. M1 0.94 A vs. M2 0.93 M1 vs. M2
  0.97
• Compared with 0.770.84 for pediatric brain tumor
  patient in recent report1

1. DHaese P et al. Int J Radiat Oncol Biol Phys
   2003 57 (2 Suppl) S205

12
Brain Segmentation Cortical Structures
KAM, Knowledge-guided Active Model

• New object functions

In contrast, Registration based approaches
maximize S deformable model based approaches
minimize H

• Pediatric brain atlas

• Affine registration (H)

• 3D active mesh (S)

13
Brain Segmentation Pediatric Brain Atlas
14
Brain Segmentation Pediatric Brain Atlas
15
Brain Segmentation Pediatric Brain Atlas
16
Brain Segmentation Affine Registration
12 DOF 3 translations, 3 rotations, 3 scaling,
and 3 shearing
17
Brain Segmentation Active Models
External Energy attract triangle vertex to the
edge of the image
18
Brain Segmentation Active Models
Internal Energy control the behavior of triangle
mesh models
19
Brain Segmentation Cortical Structures
Segmentation results
20
Brain Segmentation Cortical Structures
Segmentation results
21
Brain Segmentation Cortical Structures
Segmentation results compared with SPM2

• Volumetric agreement
• KAM 95.4 3.7
• SPM2 90.4 7.4
• Image similarities
• KAM 0.95 SPM2 0.86

22
Brain Segmentation Summary

• Pediatric brain atlas
• www.stjude.org/brainatlas
• KAM, Knowledge-guided Active Model
• preliminary results indicate that when segmenting
  cortical structures, the KAM was in significantly
  better agreement with manually delineated
  structures than the nonlinear registration
  algorithm provided by SPM2.

23
Brain Segmentation Future Studies

• Validation of KAM
• Application of KAM
• Incorporating KAM into radiation therapy
  planning
• Quantitative evaluation of cortical structure
  changes
• Further development of KAM
• Subcortical Structures
• Brain Tumors

24
Acknowledgements
Mentor Dr. Wilburn E Reddick Colleagues Dr.
Robert J Ogg Dr. Fred H. Laningham Dr.
Claudia M. Hillenbrand Carlos Parra, John
Stagich, Dr. Qing Ji, John Glass, Jinesh Jain,
Travis Miller, Rhonda Simmons