Research On Mammography Based Breast Tumor Detection

Breast cancer is one of the deadly tumor among older women. Early detection andearly treatment is the key method to reduce breast cancer mortality. Mammography isrecognized for early diagnosis method of breast cancer, but interpreting screeningmammograms is a time-consuming and boring work, so a kind of computer aideddetection and diagnosis(CAD) based on computer image processing and patternrecognition is presented, which can help doctor improve the diagnostic accuracy.CAD has many problems, for example, false-positive is higher, features optimizationand classification accuracy is not high. Some key methods of CAD in mammography wereresearched, including breast region segmentation, breast lumps features optimization andclassification of breast lumps.According to problems of over-segmentation of the watershed and breast region grayvalue close to background, a method of breast region segmentation based on watershed isproposed. The algorithm is to get the initial contour by the two times Otsu algorithmfirstly. Then it calculates the mark of breast region and background region, and sets thegray value of maker areas of morphological gradient image to0. Finally it uses watershedmethod to get the breast contour.According to problems of the redundancy of features and the influence of featurevalue range on classification, a weighted feature optimization method based on geneticalgorithm was designed. The algorithm uses the score value of classification as fitnessvalue. It will select features at the same time when it calculates feature weight. If thefeature weight lower than a certain threshold, we will remove this feature. Differentfeatures provide different classification contribution by feature weight.According to problems of small sample size and high sample dimension, this paperimplemented SVM classifier. It uses sequential minimum optimization(SMO) algorithmwhich proposed by Platt to implement.The experiment uses368images which is selected from DDSM database. Using tenfolds cross validation and receiver operating characteristic(ROC) curve to evaluate theperformance of CAD system, the Az value is0.92, which indicates that this research has ahigh performance.