|
Abstract:
|
Breast cancer detection is an important social requisite as it is the
leading cause of death due to cancer among women. The mortality rate of
breast cancer is second among all cancers. The cause for breast cancer is
not known to date and early detection & treatment are the only means to
reduce breast cancer related deaths. Mammography is the main radiological
tool that is employed for identifying breast cancer at the earliest stage.
Computer aided techniques have great relevance in detection of
abnormalities from mammographic images, as often the features associated
with various abnormalities are difficult to detect and might be missed by
even trained radiologists. In addition, when screening mammography is
employed, a large number of mammographic images need to be checked for
signs of abnormality, justifying the use of computer aided diagnosis.
Three problems are addressed in this thesis: delineation of the
pectoral muscle region by properly identifying the pectoral muscle
boundary, detection of architectural distortion and enhancement of
microcalcification features in the mammographic images. Two novel
methods were developed for identifying the pectoral muscle boundary from
mediolateral oblique view mammograms that employed multiscale
decomposition and local segmentation. The breast area is extracted after
this step following the removal of the Pectoral muscle region. The breast
abnormalities are searched for in this region. Architectural distortion is the
most commonly missed abnormality in mammograms. A novel method for
detecting architectural distortion is proposed in this thesis that employs
geometrical features obtained from selected edge structures in the
mammographic image. These features are used to train a feedforward neural
network classifier initialized using metaheuristic algorithms for better
classification. Microcalcification is another breast cancer symptom which is
ii
said to be the most commonly occurring. However the visibility of the
microcalcification structures is often poor, especially when they are located
in dense parenchymal tissues. Therefore an algorithm is proposed to
enhance such features, employing the singularities, viz. zero-crossings and
modulus maxima of coefficients obtained after computing the contourlet
transform of the mammographic image. Contourlet transform is employed
for the directional information it provides. |