2. –
™ Preprocessing
–  Grayscaling
–  Region of interest (ROI)
–  Separation of jaws
–  Adjust the image
–  Separation of teeth
™ Teeth segmentation
–  Water shedding
–  Mean shift filtering
™ Teeth classification
Content
5. –
™ Image contains a lot of extra information (jaw, nose,
etc.)
âž crop to a region of interest (ROI)
™ Characteristics of the radiographs:
–  Variation in scale and position of head and teeth is
limited
–  Head is always centered horizontally
™ Define the ROI based on the mean of the Gaussian
distribution of the ROI of each image
Region of interest (ROI)
7. –
™ Teeth have a higher grey level intensity than jaws
and other (soft) tissue, because of their higher tissue
density
âž gap between jaws forms a valley in the y-axis
projection histogram
™ How?
–  Determine a set of points that have minimal intensity
–  Use interpolation to estimate the gap valley
–  Determine a split line (parallel to the x-axis)
Separation of jaws
16. –
™ Segment image based on similar intensity
™ No proper segmentation : toothwas not sufficiently
delineated
âž exterior is also flooded
Water shedding
17. –
= (Partial) solution to the previous problem:
–  Remove part of the noise (upper part of the image)
Mean shift filtering
18. –
= (Partial) solution to the previous problem:
–  Remove part of the noise (upper part of the image)
–  Apply a Gaussian blurring (again)
Mean shift filtering
19. –
= (Partial) solution to the previous problem:
–  Remove part of the noise (upper part of the image)
–  Apply a Gaussian blurring (again)
–  Apply mean shift filtering to smoothen the image
Mean shift filtering
20. –
= (Partial) solution to the previous problem:
–  Remove part of the noise (upper part of the image)
–  Apply a Gaussian blurring (again)
–  Apply mean shift filtering to smoothen the image
–  Reapply the water shedding algorithm (and adjust the
image)
Mean shift filtering
22. –
™ Both methods do not work really well
™ The delineation is not good enough to perform a
good classification
–  Not implemented
™ Possible methods to consider:
–  Hamming distance
–  Eigenfaces
–  Principal Component Analysis (PCA)
Results from
segmentation
23. –
™  Use segmented image as a mask to compare with the
retrieved segmentation
= compare
™  Determine common scale
–  E.g. Smallest box arround both teeth
™  Determine number of not-matching pixels
Hamming distance
24. –
™ Created bitmap that contains most characteristics of
an incisor
™ If a segmented tooth can be described as a weighted
sum of a number of the bitmap images, it is classified
as an incisor
Eigenfaces
25. –
™ Determine the principal components in both the
segmented image and the retrieved segmentation
™ Determine the distance between the principal
components
Principal Component
Analysis (PCA)