Evaluation of correspondence
between left and right mammograms
work was done in Department
of Mathematics and Statistics, Flinders University, Adelaide, Australia
the aim: to find an appropriate
type of mammogram description, where the dissimilarities between corresponding
left and right breasts for one pair are apparent (for example malignant
masses are present). This description should also capture the relation
between corresponding sections of the left and right breasts of one person
(their description should be similar in case of the healthy person).
Co-occurence matrices
two-dimensional histograms
of grey level distribution in pairs of pixels with same relative displacement.
given the shape operator SO
and the number of colours N of the image F
the value at position [x,y]
in the co-occurence matrix is given by the number of pixels [j, k] such
that
F(j, k) = x and F(j + SO_x,
k + SO_y) = y
12 various features of coocurence
matrices were computred from parts of mammograms
feature vectors were computed
for co-occurence matrices, computed from
the original mammogram image
A
the edge image E of A
the magnitude image M of
the gradient of A
Pictures of (a) the original
mammogram image, (b) the edge image and (c) the magnitude image
Following figure depicts several
examples of fragments of original mammograms and their features. In the
two left columns there are different parts of the left breast mammograms
and their features and in the right columns their corresponding right pairs.
It is possible to find out that the similarity between the left and right
subregions from one pair can be detected in features (while the features
of the different fragments have not this property). But the disimilarities
are not apparent.
The conclusion from these studies
was negative. It appeared that the only detected trends in the features
were caused by the general darkening of images towards the surface (edge)
of the breasts. The next step therefore aimed to remove this anomaly in
the images. This lead to an independent part of the project.
