Title: Segmentation of Brain Tissues on MR Images Using Data Driven Techniques
1Segmentation of Brain Tissues on MR ImagesUsing
Data Driven Techniques
- Ali Hojjat1, Judit Sovago2, Alan Colchester1
- 1University of Kent, Canterbury
- 2Karolinska Institute, Stockholm
- PVEO Project Meeting
- Budapest, 11-12 June 2004
2Classification of segmentation techniques
- Segmentation techniques
- Design based on different principles, eg.
- Edge based- Use local changes (gradient)
Problem Very sensitive to noise - Region based- Use homogeneity to segment the
image - Region growing techniques Problem of
overgrowing - Statistical analysis in feature space like
histogram Problem lack of spatial information - Combined region and edge based techniques
- Atlas Based- Geometric model of the image
Problem Relies on the performance of the
co-registration step. - Hybrid techniques- Combination of two or more of
the segmentation techniques.
3MR acquisition problems affecting segmentation
- Noise
- Spatial resolution and partial volume effect
- Low resolution
- Anisotropic sampling
- Intensity inhomogeneity
- Motion artefacts
4Evaluation of segmentation techniques
- Evaluating complete systems in an application
domain requires proper definition of application
goals and of criteria by which performance is to
be judged - Evaluation of algorithms in isolation is
difficult - Gold standard usually not available
- For specific task we might need more than one
segmentation algorithm to cover the wide
categories of images.
5Segmentation techniquesused
6SPM
- Co-register the image with an atlas using an
affine transformation. - Perform Cluster Analysis with a modified Mixture
Model and a-priori information about the
likelihoods of each voxel being one of a number
of different tissue types. - Do a "cleanup" of the partitions
- Write the segmented images. The names of these
images have "_seg1", "_seg2" "_seg3" appended
to the name of the first image passed.
7Growing process
- The grey values of successive
- points joining the region are
- shown bellow
8Some definitions
- A region containing 20 pixels (shown in black and
green).
External boundary, EB
- External boundary (EB) regions boundary
- Internal boundary (IB) outermost pixels
inside the region - Peripheral contrast (Mean IB) / (Mean EB)
9- Grey level (top) and PC (bottom) mapping when
region growing started inside the scalp.
10Segmentation of scalp and skull base
11Flowchart Segmentation of scalp
Original T1 MRI
Seed point In scalp
Region growing
Find maximum Peripheral contrast
Segmented scalp
12- Grey level (top) and peripheral contrast (bottom)
- during the growing process for two conditions,
with - and without the mask.
13Sample MR image
- MR image of an elderly patient with cerebral
atrophy. - Segmented WM, GM and CSF are shown in
- Grey, dark grey and white, respectively.
14Sample MR image
153-D Visualisation of the brain
Right
16BrainSeg Summary
- It needs manually selected seed points in the
scalp and in the brain - Works well on abnormal as well as normal images
- Speed of segmentation of every structure is about
3-4 minutes - Total speed would be about 10 minutes. The speed
may vary according to the resolution of the input
image and the level of artefacts.
17Evaluation of the segmentation techniques
18Practical issues related to manual segmentation
- Questions
- Should we use anatomical knowledge to segment a
region or only rely on intensity values? - What is the best decision when the distance
between two parts of a sulcus is very close to
zero (touching each other, or less than one
pixel)? Crack between pixels might be a good
idea, but we should go to subpixel level. - Should we rely mainly on our anatomical knowledge
when partial volume affects the intensity?
19 20RG
BrainSeg
SPM
21Performance of automatic segmentation technique
for brain (WM GM) tissue
BrainSeg
22Rate of falsely detected voxels in every slice
for the two segmentation techniques
BrainSeg
23SPM
Voxels which are missed by both techniques are
shown in green
BrainSeg
24Voxels which are falsely segmented by the two
techniques are shown in red
SPM
RG
BrainSeg
25Issues related to segmentation of WM
- WM segmentation is more difficult.
- Boundary between WM and GM is not clear.
- Segmenting tissues changed from their original
form,like WM lesions, is difficult. Should we
segment the WM lesions as WM, GM or
separate tissue?
26Performance of automatic segmentation technique
for WM tissue
BrainSeg
27Rate of falsely detected voxels in every slice
for the two segmentation technique
BrainSeg
28False positive points (in orange) in the WM
overlaid on manual segmentation
Original image
Manual
SPM
29False positive (in green) and missed points (in
purple) in whole brain
Original image
Manual
30Summary
- Manual segmentation of images is very difficult.
- Performance of SPM and BrainSeg techniques are
very similar. - SPM can segment GM structures (in basal ganglia)
better than BrainSeg. - BrainSeg performs better around the cortex.
- BrainSeg is three times faster than SPM.
- Evaluation should be extended to a larger number
of subjects. - We plan to compare the results of PVE correction
using different techniques against manual
segmentation.
31(No Transcript)