Application of Wavelets and Principal Component Analysis in Image Query and Mammography

Breast cancer is currently one of the major causes of death for women in the U.S. Mammography is currently the most effective method for detection of breast cancer and early detection has proven to be an efficient tool to reduce the number of deaths. Mammography is the most demanding of all clinical imaging applications as it requires high contrast, high signal to noise ratio and resolution with minimal x-radiation. According to studies [36], 10% to 30% of women having breast cancer and undergoing mammography, have negative mammograms, i.e. are misdiagnosed. Furthermore, only 20%-40% of the women who undergo biopsy, have cancer. Biopsies are expensive, invasive and traumatic to the patient. The high rate of false positives is partly because of the difficulties in the diagnosis process and partly due to the fear of missing a cancer. These facts motivate research aimed to enhance the mammogram images (e.g. by enhancement of features such as clustered calcification regions which were found to be associated with breast cancer) , to provide CAD (Computer Aided Diagnostics) tools that can alert the radiologist to potentially malignant regions in the mammograms and to develope tools for automated classification of mammograms into benign and malignant classes. In this paper we apply wavelet and Principal Component analysis, including the approximate Karhunen Loeve aransform to mammographic images, to derive feature vectors used for classification of mammographic images from an early stage of malignancy. Another area where wavelet analysis was found useful, is the area of image query. Image query of large data bases must provide a fast and efficient search of the query image. Lately, a group of researchers developed an algorithm based on wavelet analysis that was found to provide fast and efficient search in large data bases. Their method overcomes some of the difficulties associated with previous approaches, but the search algorithm is sensitive to displacement and rotation of the query image due to the fact that wavelet analysis is not invariant under displacement and rotation. In this study we propose the integration of the Hotelling transform to improve on this sensitivity and provide some experimental results in the context of the standard alphabetic characters.