Applications of Image Processing to Medicine

paradepetAI and Robotics

Nov 5, 2013 (3 years and 7 months ago)

76 views

Applications of Image Processing to Medicine
(Beyond the limitations of the human visual system)
Maria Petrou
Informatics and Telematics Institute, CERTH, Greece
Informatics and Telematics Institute, CERTH, Greece
and
Electrical and Electronic Engineering,
Imperial College, London, UK
Limitations of the human vision
system
•People cannot see in volumes!
•People cannot perceive variation in volumes
(effectively a 4
th
dimension)
(effectively a 4
th
dimension)
•People cannot measure accurately
•People cannot see variations in high order
statistics
The role of image processing
•Go beyond the above limitations
•Measure objectively what people only
estimate subjectively
estimate subjectively
•Visualise things not perceivable or visible
•Discover new knowledge
3D texture characterisation
Use the 3D orientation histogram to characterise a volume
Quantify the shape of the orientation histogram.
Use these shape measurements as features.
Application to 3D MRI data of
schizophrenic patients
and normal controls
Statistically significant differences were observed
in the structure of the grey matter at the bottom
quarter of the brain between schizophrenics
and normal controls
MRI data: Schizophrenics and normal controls
Alzheimer’s
Patients: MRI data
Localising the places of difference
Generalised co-occurrence matrices:
They can be used to characterise surface
Shape.
Count the number of pairs of voxels that
Count the number of pairs of voxels that
are at a certain distance from each other
and have gradients with certain relative
orientation
and their gradients have certain values.
Application:
Can you tell
who suffers
from ulcerous
colitis and
who doesn’t?
Detecting Invisible Boundaries
Different mean (1
st
order): visible
Different variance (2
nd
order): visible
Different skewness (3
rd
order): invisible
Synthetic images
Skewness gradient maps for sphere and cube images
(skewness s=1.0 for background and s=2.0 for sphere & cube)
Two modes of applications:
We know that a boundary exists and we want to locate it.
We do not know that such boundary exists, but we want
We do not know that such boundary exists, but we want
to enhance our vision
When prior information is available
MRI brain data:The boundaries of malignant tumours
are diffuse and invisible.
Can we make them visible?
Do we have reasons to suspect that high order
gradients are present?
Tumours in real MRI images:
scanning along rays
Tumour border detection in real images:
MRI-T
1
patient-1patient-2patient-3
Tumour border detection in real images:
MRI-T
2
patient-4patient-5
In MRI-T1 the tumour has consistently higher skewness
In MRI-T2 the tumour has consistently lower skewness
Invisible Differences
When NO prior information is available
Invisible Differences
in Brain Images
of Schizophrenics and Controls
The method
222
ZYX
mmmm++=
gradient of mean

&

)
0

(

& )0( if




v
m
spherical
sliding window
s

+

cumulative
skewness gradient
map
222
ZYX
vvvv++=
gradient of variance
222
ZYX
ssss++=
gradient of skewness
Window Voxels
)(MAPMAP
then)0 (

&

)
0

(


+=
>


s
s
v
Real MR image data
40 subjects (21 schizophrenic patients + 19 controls), Maudsley Hospital
MRI-T
2
patients controls
MRI-PD
Results: skewness gradient maps
Results:statistical significance
There was no significant inter-group
differences found using MRI-PDimages;
In MRI-T
2
scans the percentage of
mapped brain volume was significantly
lower in patients (47.5% vs. 55.4% in
controls,
t=
4.57,
p
<0.0001)
controls,
t=
4.57,
p
<0.0001)
The percentage of non-cumulative
mapped brain volume was also lower in
patients (2.85% vs. 3.73%, t=3.34,
p<0.002).
No asymmetriesdetected when brain
hemispheres considered separately
Conclusions
The human vision system is limited
Image processing may be used to make explicit
the information that is only implicit in the data.
New knowledge may be discovered by doing that