Proyectos en procesamiento de imágenes

Proyectos en
procesamiento de
imágenes
Grupo Imagine
Grupo de Ingeniería Biomédica
Ingeniería de Sistemas y Computación
UNIVERSIDAD DE LOS ANDES
Motivación y objetivo de la
presentación
 Presentar algunos de los trabajos de
investigación en curso en procesamiento
de imágenes biomédicas
 Mostrar las dificultades de procesamiento
inherentes a las modalidades de
adquisición
 Descubrir los usos potenciales del
MEDIPIX en estos tipos de proyectos
Contenido
 Proyectos en curso
Motivación
Objetivos
Método
Limitaciones
 Retos con Medipix
Contenido
 Proyectos en curso
Motivación
Objetivos
Método
Limitaciones
 Retos con Medipix
•
•
•
Segmentación de canales
internos de corales
Caracterización de la placa
aterosclerótica en la arteria
carótida
Clasificación de tejido óseo
en el maxilar inferior
Extraction of axes from
coral
channels
Context

Marine enviroment changes




Climatic changes
Environmental pollution
Ocean acidification

Corals are more affected
Why are corals important?

Habitat for several animal species

Great part of food chain
Muricea Muricata
6
Image taken from: http://gorgonien.npage.de
Context
• Corals: Colonial animals called polyps
Polyp
Channel
Gorgonian
Axis
Channel
Muricea Muricata
Sclerit
7
Images taken from: Illustrated Trilingual Glossary of Morphological and Anatomical Terms Applied to Octocorallia
Objectives
•
Extract the axes of internal channels of corals
on µCT images
–
–
–
Visualize the internal structure
connections
Count the number of channels and
determine their sizes
Avoid manual channel extraction
8
Images
•
–
µCT
Resolution: 4.5, 10, 14 and 25 microns
Channel
Sclerit
Polyp
9
Gorgonian Axis
Method
•
Segmentation of
Gorgonian axis
3 stages
Initialization of
seed points
for channels'
axes
Extraction of
axes channels
10
Method
•
Segmentation of
Gorgonian axis
3 stages
Initialization of
seed points
for channels'
axes
Extraction of
axes channels
11
Method
Contours detection
Axis-channels
separation
Axis extraction
3D image gradient
12
Method
Contours detection
Axis-channels
separation
Axis extraction
Dilation with size mask equals to
average channel radius
13
Method
Contours detection
Axis-channels
separation
Axis extraction
Region growing. Seed
inside the axis.
14
Method
Contours detection
Axis-channels
separation
Axis extraction
Dilation to
compensate
15
Method
•
Segmentation of channels' region
Dilation with mask size equals
to channels diameter
XOR
16
Method
•
Segmentation of
Gorgonian axis
3 stages
Initialization of
seed points
for channels'
axes
Extraction of
axes channels
17
Method
Extract
center line
Obtain extern
contour of channels
Extract the
channels' signal
Detect seed
points
Gaussian
smooth
Iso contour
18
Method
Extract
center line
Obtain extern
contour of channels
Extract the
channels' signal
Detect seed
points
Gaussian
smooth
Iso-contour
19
Method
Extract
center line
Obtain extern
contour of channels
Extract the
channels' signal
Detect seed
points
Contour:
Contour normals:
Unwinding signal:
16
20
slices
Method
Extract
center line
Obtain extern
contour of channels
Extract the
channels' signal
Detect seed
points
Channels signal:
Integrate
21
Method
Extract
center line
Obtain extern
contour of channels
Extract the
channels' signal
Detect seed
points
Fourier transform, filter and reconstruction
Filter:
22
Method
Extract
center line
Obtain extern
contour of channels
Extract the
channels' signal
Detect seed
points
Problem: Sclerit near channels
23
Method
Extract
center line
Obtain extern
contour of channels
Extract the
channels' signal
Detect seed
points
Solution: Weighted integral
max
ai
∆
min
24
Method
Extract
center line
Obtain extern
contour of channels
Extract the
channels' signal
Detect seed
points
Sclerit presence is reduced
Sclerit
25
Method
Extract
center line
Obtain extern
contour of channels
Extract the
channels' signal
Detect seed
points
Initial channel points are the valleys of the signal
26
Method
•
Segmentation of
Gorgonian axis
3 stages
Initialization of
seed points
for channels'
axes
Extraction of
axes channels
27
Method
Calculate local orientation
•
Place the new point
Local orientation computation using 2 cylinders (inspired filter HD)
Rout
Cin Cout
h
Seed point
Cout
Cin
28
h = 2*Rout
Method
Calculate local orientation
•
Place the new point
Measurement of local orientations
µIn : Mean intensity inside Cin
Cin
Cout
µOut: Mean intensity inside Cout
σIn: Standard deviation inside Cin
29
Method
Calculate local orientation
•
Place the new point
Next point position
Forward direction
30
Results

4 Images
31
Results
Bifurcation region
Apex
32
Limitations
•
–
µCT
Acquisition constrains (size and temperature)
– Low contrast
µCT. Laboratory MATEIS
33
Connections between
channels
Carotid artery segmentation and
characterization in 3D computed
tomography (CT) images
Motivation
 Medical problem:
The type of atherosclerotic plaque,
i.e. the composition and structure of
the pathological vascular wall, is an
important indicator of the risk of
thromboembolic vascular events.
It determines the plaque vulnerability
to disruption.
Courtesy of: Dr
E. J. Tweedie,
University of
Western
Ontario, Canada
Objective
Lumen
Hypodense
area
 In the long term:
 Study of morphological features of
atherosclerotic plaques and the
distinction between their
components (lipids, fibers,
calcifications, thrombus).
 At present:
 Attempt to extract lumen, wall
thickness and calcifications from CT
images.
Calcification
Due to current
limitations of spatial
resolution and of
contrast in the vicinity
of the vessel wall
Method
Step :
Extraction of
the vessel
centerline
Method
Step :
Extraction of
the vessel
contours in
cross-sections
locally
perpendicular
to the
centerline
Lumen
Calcification
Outer boundary
Method: Step 1
Step : Extraction of
the vessel centerline
Image
preprocessing
Axis
extraction
Seeded bithresholding:
• Upper
threshold to
remove
calcifications.
• Lower
threshold to
remove
noncalcified
(hypodense)
plaque
components.
Method: Step 1
Step : Extraction of
the vessel centerline
Image
preprocessing
Axis
extraction
Extensible
skeleton:
• Tracking
strategy based
on a multi-scale
analysis of the
image
moments.
[Hernández Hoyos, IJCARS
2006]
Method: Step 2
Step : Contour
segmentation
Discontinuity
detection
Search ray
scheme:
• Based on
gradient
analysis.
[Lorenz, CARS 2005]
Contour extraction:
outer, lumen and
calcifications
Search ray scheme
From a point belonging to the lumen a number of rays
is cast. Directional derivative fr is computed along each
ray.
f 
f r   r  f
r

r
Search ray scheme
When a local minimum with fr < −t is found, the
search is stopped: a discontinuity point is detected.
Search ray scheme
Otherwise … when a local maximum with fr > t is
detected, it indicates a lumen-calcification transition.
This discontinuity is marked and the search is
continued.
Search ray scheme
Again … when a local minimum with fr < −t is found,
the search is stopped: a discontinuity point is detected.
Search ray scheme
At the end of this phase:
• Each detected region is
represented by a set of points
defining a polygon Pi.
• The index i = 0 corresponds
to the lumen.
• The remaining indices i > 0
correspond to the
calcifications (if any).
P0
P2
P1
Search ray scheme
At the end of this phase:
• Approximate boundaries thus
obtained are to be refined in
order to improve continuity
and to recover possibly
missing parts of the lumen
“hidden” behind the
calcifications.
• We also need to detect the
outer boundary of the wall.
P0
P2
P1
Method: Step 2
Step : Contour
segmentation
Discontinuity
detection
Contour extraction:
outer, lumen and
calcifications
Iso-contours:
• Adaptive isovalues based
on local
maximum of
intensity.
Outer boundary and
calcifications
 For each polygon Pi we compute:
 fi: maximum intensity
xi: gravity center of Pi
P0
P2
P1
Outer boundary and
calcifications
 For each Pi an iso-contour is calculated using
an adaptive iso-value i (% of the local
maximum of intensity) :
For the outer boundary:
P0
0 = out * f0
For calcifications:
P2
i = calc * fi
The iso-contour the closest to xi is automatically
selected
P1
Lumen
PP00
 A binary image is constructed: white (255)
within P0 and black (0) outside of it.
Lumen
PP00
 This image is eroded and low-pass filtered
to obtain a smooth transition.
Lumen
PP00
 An iso-contour is calculated with the isovalue initially fixed at 128.
Lumen
PP00
 If the lumen contour intersects one of the
other contours the iso-value is iteratively
incremented by a unit ...
Lumen
PP00
… until the intersection disappears.
Limitations
 Precise detection of outer wall and lumen contours
 Detection of hypondense zones
Inferior Maxillary Bone Tissue
Classification in 3D CT Images
Motivation
 Bone regeneration research project:
 Dental implants are used to replace missing
teeth.
 Atrophy of the jaw bone may occur. This would
make the placement of dental implants difficult.
 Bone grafts have been used to achieve bone
regeneration.
Motivation
 Bone regeneration research project:
 It is interesting to study the regeneration process
through Computed Tomography (CT) images.
 One objective of this study is to compare the
density of the regenerated bone to normal values.
 Little is known about the total 3D density, or the
bone tissues proportion in the anatomical zones of
the mandible.
Objective
Bone tissue classification
in 3D CT images to
determine normal density
values in 7 anatomical
zones of the mandible.
Cortical
Bone
Cancellous
Bone
Medical Context
 Anatomical zones of the mandible
Method: Step 1
Step :
Inferior
Maxillary Bone
Segmentation
Method: Step 2
Step :
Bone Tissue
Classification
Cancellous
bone
Cortical
bone
Method: Step 1
Step : Inferior
Maxillary Bone
Segmentation
Bone outer shell
segmentation
Bone internal
tissue
segmentation
3D region
growing
technique
• Lower
threshold:
Removes
other tissues
Method: Step 1
Step : Inferior
Maxillary Bone
Segmentation
Bone outer shell
segmentation
Morphological
operations
• Closing:
Fills holes
Bone internal
tissue
segmentation
Method: Step 1
Step : Inferior
Maxillary Bone
Segmentation
Bone outer shell
segmentation
Bone internal
tissue
segmentation
3D Ray
Casting:
• Fills the
segmented
surface
3D Ray Casting Scheme
 Six rays are casted from every voxel in the main
directions of the x, y and z axes.
 If all the rays intersect the external shell of the
bone, the point is included in the region.
Method: Step 1
Step : Inferior
Maxillary Bone
Segmentation
Bone outer shell
segmentation
Bone internal
tissue
segmentation
Logical
Operations
• AND:
Obtains the
original gray
levels
Method: Step 2
Step : Bone Tissue
Classification
Classification:
cortical,
cancellous,
undetermined.
Refining: cortical,
cancellous.
Clustering :
• Fuzzy C
Means.
Method: Step 2
Fuzzy c-means

Each point of the image has membership grade to a
class or cluster.

The membership grade is inversely related to the
distance between the intensity value of the point and
the gray level of the centroid of the class.
Method: Step 2
Step : Bone Tissue
Classification
Classification:
cortical,
cancellous,
undetermined.
.
Refining: cortical,
cancellous.
2D Ray
casting:
• Determines the final
types of bone tissue
2D Ray Casting Scheme
 From
every point of undetermined tissue eight rays
are casted.
 The ray analyzer searches for the first pixel with a
different intensity level.
 If a ray intercepts the background, the point is
classified as cortical bone.
Limitations
 It
was observed that narrow areas of dense
cancellous bone may be mistakenly classified as
cortical bone.
 Elimination of teeth from the image to take
measurements of the alveolar zone.
Retos con Medipix
Retos con Medipix
1. Obtener imágenes con alta resolución espacial:



De segmentos de corales, reemplazando el microTAC
De especímenes de arterias carótida provenientes de donantes
cadavéricos.
De fragmentos de mandíbula de animales pequeños.
2. Distinguir diferentes componentes (tejidos) no visibles en las
imágenes tradicionales, a partir de información de absorción
para diversos rangos de energía, aprovechando los umbrales
variables y ventanas de energía en rangos variables.
Retos con Medipix
1. Distinguir diferentes componentes (tejidos) no visibles en las imágenes
tradicionales, a partir de información de absorción para diversos rangos
de energía, aprovechando los umbrales variables y ventanas de energía
en rangos variables.
En corales:
Visualizar los tejidos que
separan los canales
entre sí y los canales del
eje córneo
Retos con Medipix
1. Distinguir diferentes componentes (tejidos) no visibles en las imágenes
tradicionales, a partir de información de absorción para diversos rangos
de energía, aprovechando los umbrales variables y ventanas de energía
en rangos variables.
En arterias:
Lumen
Estudiar las
características
morfológicas de las
placas ateroscleróticas y
distinguir sus
componentes (lípidos,
fibra, calcificaciones,
trombos)
Fibra
Grasa
Retos con Medipix
1. Distinguir diferentes componentes (tejidos) no visibles en las imágenes
tradicionales, a partir de información de absorción para diversos rangos
de energía, aprovechando los umbrales variables y ventanas de energía
en rangos variables.
En tejido óseo:
Diferenciar tejido óseo
cortical de los dientes