Research And Realization Of Mammogram Retrieval System Based On Content

Breast cancer has been considered as No.l killer of women. The mammogram is the most reliable clinical method to detect breast diseases. But the mammograms masses are usually difficult to detect as they often superimpose on complicated and dense structured background, and it is hard to detect and distinguish, besides, audit fatigue of doctors is another factor of error diagnose, so it is necessary to search for a similar picture compared with queried picture from mammogram database. The essay realizes the function of prototype image retrieval system based on image retrieval knowledge, the main work is as follows:A mammogram database is built and contains 750mammographic regions. Every region of interested (ROI) in the database has known pathology-benign and malignant. It is the base of data according to ROI queries.Due to the difficult problem that how to segment the mass region of the queried ROI, an interactive mass segmentation method is given. The size of mass by user provided is used. The improved growing algorithm is adopted to segment the suspicious region of mass, and evaluated by the method of overlapping percentage. For the effective feature selection problem,28features are extracted initially form each ROI as a feature set, then 22 features are selected from the feature set using 1-r-GA feature selecting method to form a 8 dimension vector.The dissertation adopts a kind of two layer classification structure to fulfill visual and pathology similarity between retrieval image and queried ROI. In the first layer, SVM classifier is used to fulfill pathological classification. In the second layer, K-nearest is adopted to get the most k visual similar images. At last, the first 15 ROIs are returned to doctors based on mutual information (MI) order.Mammogram retrieval system realizes and shows the retrieval process based on this system. The experiment shows the system offers an effective method to realize mammogram retrieval.