A Study On The Similarity Metric Algorithm For Mammogram Retrieval

Breast cancer is one of the most common malignant tumors that hazard the middle-aged women's health. In China, the incidence of breast cancer in women presents persistent high growth. Currently, the annual growth rate sharply rises to 3% ~ 4%. Due to the easy, low-cost, non-invasive and good detection results of early asymptomatic of occult breast cancer, mammography becomes the primary method for breast cancer detection in the current clinical environment.The studies show that it will be very helpful for radiologist to make the judgment in the diagnosis process if he can take the advantage of content-based image retrieval technology to retrieve the similar cases with biopsy-proven results over the years.The thesis aims at the study of similarity metric for retrieval of the mammographic image with masses. A two-stage learning hierarchical framework was proposed for similarity metric learning. In addition, an experimental study was taken to make a comparison of three typical CBIR methods for computer-aided diagnosis in mammography. The results presented in this thesis could be used as a valuable reference for other researchers. The thesis consists of three parts.In the first part, the database of mass regions-of-interest (ROIs) that will be used in this study is presented, which is followed by a description of feature extraction and normalization methods.The second part presents a new method for similarity metric learning which is machine-learning based two-stage framework. With the features extracted in the first part, we first use 5 different classifiers separately on the two ROIs to judge if they are similar in semantics of malignancy, then voting method is used to get the final classification results. After the classification, the ROIs that are considered to be"similar"in semantics (either benign or malignant) are entered into the next phase for the visual similarity assesment. In this step, we use the strategy of combining the k nearest neighbor classifier and mutual information to judge if the two ROIs are visual similar. Through these two stages of similarity evaluation, the selected ROIs are expected to be"similar"to the query ROI both in"semantic"and"visual"meaning.In the third part, we propose two indexes, association degree and weight association degree, to evaluate the relationship of retrieval results for different CBIR methods. Extensive experiments were taken in the comparison study of different CBIR methods for CAD in mammograms. The results are of scientific significance and reference value for peer research.Finally, give the summaries and prospects of the full text. In summing up the work of this thesis and predicting the direction of future research.