| Summary: | In this thesis several global mammographic features were examined for their ability to classify the mammograms into (1) classes based on the proportion of dense tissue; (2) normal/abnormal groups. A set of 240 digitised mammograms was obtained from the Digital Database for Screening Mammography from the University of South Florida. The database was composed of mammograms that were digitized using one of three high resolution x-ray digitisers. It was necessary for the images to be corrected for three systematic differences between the x-ray digitisers: the resolution, the slope of the calibration curve and non-linearities in the calibration curve. A simple correction was also made for differences in the mammographic technique by adjusting the histogram of the breast shadow. The breast shadow was then segmented using a semi-automatic procedure and several mammographic properties were extracted: global moments of the histogram, the average local moments calculated for ~3 x 3 mm2 regions covering the breast shadow, subregions of the global histogram, multifractal dimensions and the texture energy, entropy and inertia calculated for the wavelet transform of the image. The classification accuracy, when considering the density grades, was consistently ~40% correct and independent of the properties used in the classifier. When classifying into normal/abnormal groups, the regional moments, histogram sub-regions and the multifractal dimensions all had approximately the same performance at ~60% correctly classified cases, While the global moments classified ~70% of the cases correctly. The texture energy, entropy and inertia also had approximately the same performance but at ~80-85% correct. In addition, the classifiers exhibited no significant change in classification performance for variations in age for any of the examined properties with ' p' = 0.001. The texture features resulted in the highest classification accuracy. The results may show some residual dependence on the x-ray digitiser but the small sample size precluded any definitive conclusions regarding the influence of the scanners. Overall, a classifier using six texture inertia features exhibited the best overall classification accuracy with minimal age dependence.
|
|---|
