Transcript Medical Image Classification using Mathematical Morphology

SP-ASC 2010
São Paulo Advanced School of Computing
Medical Image Classification by
Mathematical Morphology Operators
Dra. Mariela Azul Gonzalez
Director: Dra. Virginia Ballarin
Co-Director: Dr. Marcel Brun
Universidad Nacional de Mar del Plata
Mar del Plata, Buenos Aires
Argentina
Phd Thesis: Bone Marrow Biopsies Segmentation
Conventional Image Processing Techniques
Bone Marrow Image
Thereshold
Contour Tracing
Region Growing
Watershed
Transform
Watershed Transform
Flooding Algorithm
Watershed Lines
basins
Markers
Proposed algorithm for marker definition
Final results of other bone marrow biopsies.
Conclusion
•The classification by over-segmented regions has
proved to be advantageous.
• It is less sensitive to the noise present in the medical
images and reduces computational cost.
New Project
•The proper characterization and quantification of shape, size
and direction of 2D Medical Images Components.
•Future works oriented to process Medical Image 3D
2D Images Tissue Engineering Scaffolds and developing Neurons.
Granulometric Function
To obtain a Granulometric Function, first we applied
openings with increasing structuring elements, Then we
compute each area (or volume in gray level images).
Those values are normalize to obtain a probabilistic
distribution function. Finally we compute its moments to
compare them in order to analyze morphological
characteristics of objects of interest.
 ( A E )
G ( )  1 
A  E ( x)  ( AE )  E
 ( A)
G ( )  1 
 ( A E )   ( A E N )
( A)  ( A E N )
Granulometric Function
Proposed Method
1° - To obtain Granulometric Functions with
different structuring elements,
2° - To compute its moments and compare them,
3° - To analyze morphological characteristics
of objects of interest.
Area: NGF and its derivative
Area
Mean Value SE Lineal
0.96
0.94
0.92
0.9
Mean Value SE Round
Mean Value
0.88
0.86
0.84
0
100000
200000
300000
400000
Area
500000
600000
700000
800000
900000
Preliminary Results: Mean Value (round SE) vs. Diameter
Mean Value
Diameter (µm)
Preliminary Results: Mean Values (linear SE) vs. Orientation
Mean Value
SE orientation
Conclusion
The preliminary results shows there’s is an association
between the NGF moments and components
morphology (shape, size and orientation). Future studies
are oriented to process a higher number of images
Thanks to SP-ASC 2010, the organizate committee, speakers and students
Mar del Plata, Buenos Aires, Argentina
Proposed algorithm for marker definition
a) Over-segmented regions were obtained through the
application of the Watershed Transformation, using the
regional minima as markers
b) The region’s attributes were calculated. The average value
was determined, along with the standard deviation of the gray
level values from the pixels belonging to each region.
c) The values of the attributes from each region were
classified with several methods based on expert oriented
Clustering, Fuzzy Logic Inference Systems and
Compensatory Fuzzy Logic Systems. The selected regions
will be the Markers for a new application of the Watershed
Transform
e) Binarization.
f) Finally, openings with structuring elements of 3x3 pixels
was carried through by unifying the adjacent regions and
eliminating the noise and irrelevant objects.
SP-ASC 2010
São Paulo Advanced School of Computing
Medical Image Classification by
Mathematical Morphology Operators
Mariela Azul Gonzalez
[email protected]
Directora: Virginia Ballarin
Co-Director: Marcel Brun
Universidad Nacional de Mar del Plata
Buenos Aires
Argentina
Morphological operators for binary images
Erosion:
AB  x : B  x   A 
siendo B(x)  { b  x : b  B}
Dilation:
A  B   x B(x)  A  