Volumes Of Interest Definition

1
Volumes Of Interest Definition

• Mario Quarantelli
• Biostructure and Bioimaging Institute CNR
• Naples - Italy
• HBM2004 - PVEOut Satellite Meeting
• Budapest, 12 June 2004

2
Background

• Manual delineation of VOIs is
• Operator-dependent, less reproducible?
• very time demanding (up to 8 hours for 37 VOIs
  per subject)
• prone to errors
• Ideally a method for VOI definition should be
• Fully automated
• Accurate (gold standard?) and reproducible
• Capable of working on multiple modalities
• PET (FDG, receptors)
• SPET (CBF, receptors)
• MRI (T1, EPI, segmented)

3
REQUIREMENTS FOR PVE-C

• VOIs must be brought in the single patient space
  (where resolution is defined)
• VOIs must cover the whole brain
• Possibly homogeneous VOIs should be defined
  (tracer distribution)
• Different VOI sets for different tracers

4

• The complete process of digitalized brain atlas
  based identification of anatomical structures
  requires three different tools
• A VOI database of 3D brain structures (atlas or
  template) in a standardized coordinate system
• A spatial normalization software for the
  definition of a correspondence between each
  individual 3D MRI data set and a standard space
  (Talairach, MNI, others). If we calculate a
  normalization matrix to move from the patient
  space to the standard space, this matrix will be
  used backward to superimpose the template onto
  the single subject study
• A software for applying the labeled VOI's to the
  functional images.

5
(No Transcript)
6
ATLAS - Talairach based
Andreasen, NC, Rajarethinam R, Cizadlo T, et al.
Automatic Atlas-Based volume estimation of human
brain regions from MR images. J Comput Assist
Tomogr 19962098-10 Quarantelli M , Larobina M,
Volpe U, Amati G, Tedeschi E, Ciarmiello A,
Brunetti A, Galderisi S, Alfano B.
Stereotaxy-based regional brain volumetry applied
to segmented MRI validation and results in
deficit and nondeficit schizophrenia. NeuroImage.
2002 Sep17373-384
7
Talairach stereotactic coordinate system is
widely used for inter-subject normalization and
localization of activation sites in nuclear
medicine functional studies. _____________________
__ Talairach J et al., 1952. Presse Med
28605-609 Talairach, J., and Tournoux, P. 1988.
Co-planar stereotaxic atlas of the human brain.
Thieme, New York

8

• Under the assumption of proportionality of normal
  brain structures, the proportional grid approach
  proposed by Talairach divides the supratentorial
  brain into
• 8 axial planes above the AC-PC line
• 4 axial planes below the AC-PC line
• 4 coronal planes anterior to the AC
• 3 coronal planes between AC and PC
• 4 coronal planes posterior to the PC
• 4 sagittal planes on each side of the midsagittal
  plane
• Defining 1056 small boxes

9
ATLAS - Talairach based

• Assignment of Talairach boxes was done
  preliminarily by visual inspection of the
  Talairach atlas Talairach, J.,1988, based on
  the labeling of cortical structures therein
  reported.
• The software then
• Allows for identification of the AC and PC on
  original axial images
• GM selection
• Segmentation is either
• performed binarily, i.e. each intracranial pixel
  is labeled as belonging univocally to GM, WM and
  CSF
• or segmented maps are binarized (for
  probabilistic segmentation, each voxel is zeroed
  if (pGMpWMpCSF) lt50, remaining voxels are
  assigned to the most probable tissue
• Rebinning of GM volume to take care of
  anisotropic voxels (e.g. 0.94x0.94x4mm).

10
ATLAS - Talairach based

• Re-alignment of the segmented GM volume to the
  AC-PC line
• Automated identification of the falx cerebri (FC)
  for correction of possible rotation around the Y
  axis (due to malpositioning of the head at the
  time of the MR scan).
• Identification of the boundaries of a box
  encompassing the supratentorial brain
• Application of the Talairach proportional grid to
  the segmented image set

11

• VALIDATION
• 10 MR studies have been analyzed twice using the
  manual technique and twice using the automated
  technique (one month apart)
• Volumetric accuracy
• Specificity
• Reproducibility

12
Difference in reproducibilities significant at
paired t-test after correction for multiple
comparisons. When pooling all structures
together, no differences in the reproducibilities
of the two methods emerged.
13
Representative slices from the segmented MRI
study of the validation set with the smallest
error (mean error per structure 3 ml).
14
... and with the largest error (mean error per
structure 11.2 ml).
15
ATLAS - MNI based

• Voxels of the MNI space belonging to cerebral
  lobes, cerebellum, PFC, Hyppocampus and Posterior
  cingulate have been labeled according to their
  MNI coordinates paralleling the Talairach Labels
  database served by the Talairach Daemon.
  http//ric.uthscsa.edu/projects/talairachdaemon.ht
  ml
• Lancaster JL, Rainey LH, Summerlin JL, Freitas
  CS, Fox PT, Evans AC, Toga AW, Mazziotta JC.
  Automated labeling of the human brainFA
  preliminary report on the development and
  evaluation of a forward-transformed method. Human
  Brain Mapping 19975238242

16
AtlasMNI based

• The MNI atlas module only works if SPM is
  installed on the same PC.
• PVELab will automatically invoke the SPM
  normalization tools, needed to measure the
  normalization matrix, which will be used to
  assign each GM voxel of the subject to the
  corresponding structure
• Currently it only uses affine transformation
  parameters
• Normalization is done using segmented GM and GM
  prior
• Template is made of binary volumes in analyze
  format, with a simple ascii file coupling each
  structure to a
• Validation is ongoing

17
(No Transcript)
18
Idea of proposed method

• Multiple sets of Regions of Interest (VOI's) is
  available in different template spaces.
• These have manually been delineated at high
  resolution MR scans (preregistered to the AC-PC
  line) for a number of template subjects and
  afterwards carefully been checked for errors
• Multiple template VOI sets is automatically
  transferred from template spaces to new
  subject space
• By combining multiple transferred VOI sets it is
  possible to limit some of the variation coming
  from delineation and identification of
  transformation parameters

19
Example of 4 template VOI sets
20 VOI sets (37 VOIs) have manually been
delineated at high resolution structural MR
images (2x2x2 mm voxels) for 10 healthy controls
and 10 MCI patients (Karine Madsen and Steen
Hasselbalch, NRU)
20
Transformations used between template and new
subject spaces
Affine (12 param.) transformation
Woods, JCAT, 1992
Warping (soft) transformation
Kjems, IEEE TMI, 1999
21
Transformation of three template MRs to new
subject space
Affine and warp transformation
22
Transformation of VOIs and generation of
probability maps for the VOIs

• Applying the identified transformation to the
  VOIs defined in template space multiple sets
  of VOIs are available in new subject space
• A probability map for voxels being included in
  the final VOI set is individually created for
  each VOI.
• Proposed method
• for each template VOI set transformed the
  probability being in the VOI is 1 for voxels
  inside the VOI and 0 outside
• create a common probability map by averaging the
  probability maps generated in new subject space
  
• threshold the probability map so the volume of
  the created VOIs are equal to the mean of the
  transformed template VOIs

23
Example of probability MAP for some VOIs

• Upper panel Probability map for cerebellum
• Lower panel Probability map for sensory motor
  cortex and parietal cortex
• As expected voxels in the middle of the VOIs
  have the highest probability while more exterior
  voxels have lower probabilities

24
Conclusion