Image analysis methods for the detection and classification of mammographic masses

Mammography is well-established as a tool for early diagnosis of breast cancer. While the technique can be accurate, its performance is limited by various factors that lead to uncertainties in the diagnosis. Computer-aided image analysis methods could be helpful in prompting as well as for second opinion purposes. Breast cancer exhibits diverse characteristics on mammograms, of which masses constitute some of the primary signs of the disease. The present thesis proposes computer methods that address two major aspects concerning the analysis of masses in mammographic images.;Initially, methods are proposed to evaluate the efficacy of various radiographic notions of masses such as shape, gradient, and texture information, as apparent on mammograms, in the benign versus malignant classification of masses using manually segmented mass regions. The shape analysis methods proposed in the present thesis achieved improved benign versus malignant classification results as compared to other methods. Furthermore, the analysis of textural information using ribbons of pixels extracted across the boundaries of masses as proposed in the present thesis achieved the best benign versus malignant classification accuracy represented by an area (Az) of 0.82 under the receiver operating characteristics (ROC) curve.;The second major aspect of the study is related to the detection of masses and analysis of false-positives (FPs) in mammographic images. A new method based on pyramidal decomposition principles is developed for the detection of mass regions. Texture flow-field methods are introduced to compute features based on oriented textural information present in the margins of masses in order to classify the regions detected as true masses or FPs. The proposed features yielded a good mass versus normal tissue classification accuracy with Az = 0.87 with a dataset of 56 images. Malignant tumor versus normal tissue classification resulted in a higher A z value of 0.9. Furthermore, benign versus malignant classification of the segmented mass regions based on the texture information present in their margins resulted in Az = 0.79. The methods developed in the present thesis could lead to promising applications in computer-aided analysis of mammographic images for early diagnosis of breast cancer.