OVERCOMING PARTIAL VOLUME EFFECTS IN MR TISSUE ANALYSIS

1
OVERCOMING PARTIAL VOLUME EFFECTS IN MR TISSUE
ANALYSIS

• M.S. Atkins and Z. Tauber
• School of Computing Science, Simon Fraser
  University, Burnaby, B.C., Canada

2
Introduction

• Simple intensity histogram analyses of brain
  tissue
• in dual-echo PD and T2-weighted MR brain images
• often fail because of partial volume effects.
• We overcome these effects for multiple
  sclerosis
• (MS) lesions in MR images with thick (5mm) slices
  
• in order to automatically identify them, noting
  that
• Conspicuous MS lesions have high intensities in
  both PD and T2 images 1.
• Lesion borders having partial volume effects are
  darker in both spectra.

3
Intensity Histograms
PD Histogram
T2 Histogram
4
Method Stage 1 Preprocessing

• Initially we enhance the contrast of the brain
• tissues in all the MR head images by
• performing the following operations
• Correct for RF-Inhomogeneity using Sleds method
  2.
• Isolate the brain tissue from the head skull
  using Atkins method 3.
• Scale the brain pixels to the range of 0-255.

5
PD-T2 signal analysis

• We present a typical slice of the brain MRI
• PD Image (slice 17) T2 Image (slice 17)
  PD/T2 Scatter Plot
• The scatter plot shown here represents all the
• tissues in the brain slice.

6
PD-T2 lesion signal analysis

• Manual segmentation of the MS lesions in the
• MR images has a characteristic PD-T2 plot
• PD Lesions (slice 17) T2 Lesions (slice 17)
  PD/T2 Scatter Plot
• Hence, we assume that lesion pixels form a wide
  line in PD-T2 feature space.

7
PD-T2 Scatter Plot Analysis

• Conspicuous MS lesions/lesion interiors have high
  intensity in both PD and T2 images.
• Subtle MS lesions/lesion boundaries have average
  intensity in both PD and T2 images.
• All lesion intensities however, in a single
  slice, lie on a line of specific width in the
  PD-T2 feature space.
• The subtle parts of the MS lesion pixels have the
  same intensity in both PD and T2 images as some
  of the other brain tissues (mostly gray matter).
• The conspicuous MS lesion pixels are the only
  pixels that have high intensity in both PD and T2
  images.

8
Conspicuous MS lesions

• To obtain only the conspicuous MS lesion pixels,
  a fairly simple intensity thresholding is applied
  to both PD and T2 images.
• Only points that pass the threshold in both PD
  and T2 images are considered conspicuous.

PD image after threshold
PD-T2 scatter plot after threshold
9
Hough Transform for Line Normal Parameterization
y

• ? xi cos ? yi sin ?
• 0 ? 2? , -N ? N
• ? r (xi / r cos ? yi / r sin ?)
• r cos (? - ?)
• Hough Transform
• - Create an accumulation array for the ?-?
  parameter space.
• - Every pixel in the image adds 1 to the
  accumulator array cells under the
  parameterization curve.

d
(xi,yi)
?
r
?
F
0
x
?
?
10
Line Detection

• Given the PD-T2 scatter plot of only the
  conspicuous pixels, a Hough transform for line
  detection (using normal parameterization) is
  applied to the scatter plot image.
• The peak point in the Hough transform space is
  selected using a
  Gaussian filter.
• The peak point corresponds to the
  detected line parameters.

Hough Transform
11
Line Detection (Continued)

• The Hough transform image represents many 1-pixel
  thin lines in many directions and positions
  constrained by the wide line that covers all the
  lesion pixels in PD-T2 space.
• The peak point in the Hough transform space
  represents the parameters of the average thin
  line inside the wide line, which is also the
  center line of the wide line.
• The thickness of the wide line is determined by
  summing the pixels along the peak line gradient
  (i.e. the same x coordinate) in Hough transform
  space until the sum equals nearly 100 of the
  number of pixels in the PD-T2 scatter plot of the
  conspicuous lesions.

12
Line Mask

• A line mask is automatically generated based on
  the allocated peak parameters.
• Only points with intensities in both PD and T2
  images that are covered by the line mask are
  allowed to be tested for being subtle lesions.
• Line Mask Image Allowable PD-T2 points
  Subtle tissues under mask

13
Conspicuous Lesions Postprocessing

• Thresholding doesnt eliminate all
  non-conspicuous pixels. However, typically most
  other pixels are isolated in the thresholded
  images.
• Very small pixel blobs are therefore eliminated
  from the conspicuous lesions images.
• Note The Hough transform is robust with respect
  to the extra non-conspicuous pixels.

14
Subtle Points Detection Criteria

• The following criteria are defined for the
• allowable pixels to be considered subtle lesions
• Shorter pixel distance to a conspicuous pixel is
  a better indication that its a subtle pixel.
• Larger number of conspicuous pixels in the
  neighborhood indicates higher likelihood of being
  a subtle pixel.
• Higher intensity of a conspicuous pixel in both
  PD and T2 images in the neighborhood indicates a
  subtle pixel.

15
Subtle Points Detection

• Applying a Gaussian smoothing filter with a
  cutoff of about 4 pixels to the conspicuous
  lesions images produces an image with subtle
  lesion pixel values appropriate to the criteria
  required.
• The resulting image is then further
  processed to remove all points that dont
  have the PD-T2 intensities allowable
  to represent a subtle lesion.

Possible Subtle Lesions
16
Lesion Detection

• The conspicuous lesions image is combined
  together with the subtle lesions image.
• The Radiologist can interactively threshold the
  result image to the required probability of the
  detected subtle lesion being a real MS
  lesion, and thereby obtain an image of
  only the lesions.
• Alternatively the threshold can be
  automatic.

Thresholded lesion pixels
17
Conclusions

• For brain MRI where we can assume the lesion
  pixels intensities lie under a line in PD-T2
  space, an automated thresholding method along
  with a Hough transform can be applied in order to
  automatically detect MS lesions and overcome the
  partial volume effect of the subtle lesions
  adjacent to conspicuous lesions.
• Spatial information can then be used to refine
  the tissue segmentation further, as in 4.

18
Future Work

• Investigate the sensitivity of our method to the
  assumption that there is low variability in
  lesion tissue material.
• Discover the exact reason why it is not always
  the case that the MS lesions pixels PD-T2
  intensities lie under a line.
• References
• 1 M.S. Atkins et al. ISMRM 97, 649.
• 2 J. Sled et al. IEEE Transaction on Medical
  Imaging, 17(1), 87-97, 1998.
• 3 M.S. Atkins and B. Mackiewich, IEEE
  Transaction on Medical Imaging, 17(1), 98-107,
  1998.
• 4 K.Krishnan and M.S. Atkins, Proceedings in
  SPIE 98, 33381106-1116.