Cooccurrence Statistical Medical Image Processing

1
Co-occurrence Statistical Medical Image
Processing

• Bradley D. Clymer
• Dept of Electrical Computer Engineering
• Dept of Biomedical Informatics
• Participating Faculty in Biomedical Engineering

2
Outline

• Brief overview of medical image research of
  Clymers group
• Using in vivo resolution to infer hidden data in
  spatial statistics -- osteoporosis, microvessels
• Using statistical co-occurrence methods for
  multispectral medical images -- dynamic contrast
  MRI
• Using 4-D co-occurrence techniques to process raw
  dynamic contrast data for tissue characterization

3
Clymer Group Medical Image Research Activities

• Image acquisition reconstruction
• Parallel channel MRI (SMASH) at ultrahigh Fields
  Dr. P. Schmalbrock (MRI), P. Wassenaar (MS EE)
• Electron Paramagnetic Resonance Imaging (EPRI)
  Dr. J. Zweier (DHLRI), Dr. Y. Deng (DHLRI), Dr.
  P. Kuppusamy (DHLRI), R. Ahmad (EE PhD Student)
• Multimodal Data Fusion Display
• Temporal Bone Surgery Simulator Dr. G. Wiet,
  (Childrens Hosp.), D Stredney (OSC), Dr. P.
  Schmalbrock (MRI), S Raghunathan (BME PhD
  Student)
• Using audio/haptics for multimodal image data
  fusion and perception Dr. M. Knopp, Hee Chun
  (ECE PhD Student).

4
Clymer Group Medical Image Research
Activities(cont.)

• Directionally sensitive co-occurrence textures
  for tissue characterization
• Osteoporosis assessment Dr. K. Powell (CCF),
  Chad Showalter (ECE PhD Student), plus pending
  NIH grant with 4 Colleges 6 Depts.
• Microvessel volume density from ultrahigh field
  MRI Dr. G. Christoferidis (MRI), Dr. P.
  Schmalbrock (MRI), Dr. D. Chakeres (MRI), P.
  Barnes (BME PhD Student)
• Dynamic Contrast MRI
• Parameter pseudo-image spatial co-occurrence
  statistics Dr. M. Knopp (MRI), Dr. J. Heverhagen
  (MRI), M. Kale, (ECE PhD Student)
• 4-D spatial-temporal co-occurrence statistics on
  raw data Dr. M. Knopp (MRI), Dr. J. Heverhagen
  (MRI), Dr. T. Kurc (BMI), B Woods (ECE MS
  Student)

5
Statistical Co-occurrence Image Textures
(Haralick et al.)

• Even when fine structures are not resolved
  completely, intravoxel mixing effects create
  statistical spatial patterns
• Haralicks methods can be directionally sensitive
  can sense isotropy/anisotropy of statistics in
  image data
• Microvessels long tubes with diameters smaller
  than voxel (pixel) size.
• Trabeculae sheets and rods of calcified tissue
  with cross sections smaller than voxel (pixel)
  size.

6
Statistical Co-occurrence Image Textures (cont.)

• Haralicks general approach
• Build a co-occurrence matrix (joint histogram) of
  nearest neighbor values in a specific direction
• Generate a group of moments and entropies to
  characterize local joint statistics in the given
  neighbor directions
• Use combination of calculated moments and
  entropies to classify local image content
• Have been used since 1973 on 2-D imagery, more
  recently on 3-D, we are extending to 4-D with
  parallel cluster computing approaches

7
Example from bone imaging

• We used mCT images of bone core samples
  (calcaneus)
• Simulated blurring to in vivo CT resolution by
  local averaging of mCT data
• Calculated directionally sensitive co-occurrence
  textures on blurred images
• Compared texture outcomes as predictors for
  morphometric parameters (accepted standard) using
  linear regression

8
Texture parameters vs Morphometric parameters at
different resolutions
9
Next phase of bone imaging (pending NIH R21/R33
proposal)

• Obtain in situ/vivo images from human cadavers
  and live thoroughbred horse remodeling with
  exercise and atrophy
• Establish translational model to bone mechanics
• Compare in situ/vivo CT MRI via directional
  textures with
• Morphometric measures from high resolution images
  after excision
• Mechanical testing, localization of fracture
  sites
• Participants ECE/BME/BMI Clymer (PI)
  Radiology M Knopp, P Schmalbrock BME/Orth Surg
  A. Litsky Veterinary Science A. Bertone Mech
  Eng M. Dapino, OSC D. Stredney

10
Co-occurrence Processing of Dynamic Contrast MRI
4 approaches

• Spatial pattern textures of diffusion model
  parameter pseudo-images (k21, kel, Amp)
• Use co-occurrence technique to assess parameter
  combination statistics and moments and entropies
  similar to textures at matched points in
  k21-image, kel -image, Amp-image
• Combine methods 1 2
• Use raw 4-D DCE MRI data and calculate
  co-occurrence textures in specific directions
  through hyperspace, build classifier without
  using diffusion model parameters

11
Approach 1 (spatial textures of parameter
pseudoimages)

• John David Fleig (MSEE, MSCIS 2003)
• Joel Saltz, BMI
• Tahsin Kurc, BMI
• Bradley Clymer, ECE, BMI

12
Approach 2 (pixelwise co-occurrence across
different parameter pseudo-images)

• Mehmet Kale (MSEE 2004, ECE Phd Student)
• Johannes Heverhagen, Radiology
• Bradley Clymer, ECE, BMI

13
Approach 4 (4-D Image texture classification of
raw DCE image sequences)

• Brent Woods (ECE MS Student)
• Tahsin Kurc, BMI
• Johannes Heverhagen, Radiology
• Bradley Clymer, ECE, BMI