A Marker-Based Watershed Algorithm For Segmentation Of Microcalcifications In Mammograms

Breast cancer is one of the major malignant diseases that threaten women’s life and health.The early detection of breast cancer is the key to cure this disease.So far,mammography has been used most commonly and effectively for the early detection.The early symptoms of breast cancer are mainly presented as microcalcifications,with a size ranging from 0.1mm to 1mm,and usually appear as small bright spots in mammograms.Due to the small size of some microcalcifications and the low brightness,visual identification is sometimes difficult and error-prone.Misdiagnosis can make the patient miss the best opportunity for treatment,so that the patient may not survive.CAD(Computer Aided Detection)is a system of detecting microcalcifications by computer-assisted computation.In fact,CAD systems makes up for the inadequacies of manual detection and improves the accuracy of detection.A CAD system usually includes several major steps such as preprocessing,enhancement,segmentation,detection,classification,etc.This article proposes a new method for some key steps in CAD system.The MIAS database has been used in our experiments.The gamma conversion has been used in the preprocessing step of the image for enhancing the contrast between calcifications and the background tissue.At last,the experiment used wavelet transform and contourlet transform to specifically enhance microcalcification features.The results showed that calcifications have been highlight effectively,which laid the foundation of subsequent process.The experiment then used a marker-based watershed segmentation algorithm to segment the microcalcifications.This method can clearly segment each individual microcalcification,especially in areas where microcalcifications are dense,it can also ensure the effectiveness of segmentation.At the same time,the segmentation results generated some non-microcalcification areas which can be removed by classification.The segmented results were classified by SVM(Support Vector Machine)to determine the final consequence.First,we obtained a vector of feature values corresponding to each area,including area,perimeter,average value,standard deviation,etc.After getting a SVM model on a training set,we implement the final classification on testing set.At last,we got a resulting image with only real microcalcifications.