A Breast Ultrasound Classification Method Based On High-level Semantic Features Mapping

Breast cancer is one of the most common cancers and the second leading causeof female cancer death, which has a serious impact on women’s physical and mentalhealth. Due to non-radioactive, non-invasive and highly accurate identification ofbenign and malignant tumors, etc., ultrasound imaging method has already becomean important method for breast cancer diagnosis. In order to assist doctors improvingthe accuracy and objectivity of ultrasonic detection of breast cancer and reducing themisdiagnosis rate of malignant tumors, many computer-aided diagnosis (CAD)methods have been widely used in breast cancer diagnostic process. But manyclassification results of CAD are hardly understood and accepted by doctors, sobreast ultrasound CAD system is limited in application. How to make doctorsaccepted and recognized the classification results given by breast ultrasound CADsystem easily is an urgent problem to be solved.In traditional CAD system for breast cancer detection and classification,computer mainly extract low-level features such as sonographic, texture, based onmodel etc. features. However, BI-RADS descriptors are critical in diagnosing benignand malignant for doctors. BI-RADS descriptors can be treated as high-levelsemantic features, because it combines the human visual characteristics andunderstanding of the image. There is a big ‘semantic gap’ between low-level featuresand high-level semantic features. In this paper, only choose three BI-RADSdescriptors highly predictive value for benign and malignant lesions: shape,orientation and margin. In order to narrow down the ‘semantic gap’, SVM isproposed for mapping from low-level features to high-level semantic features. Thenthe C4.5decision tree is applied to classify the mapped high-level semantic features.Experimental results illustrate that the classification based on high-level semanticfeatures performs much better than the one based on low-level features, furtherimproving the classification accuracy.Aimed at the situation that some samples do not show benign and malignantobviously, a minimum classification risk model is proposed and texture features are combined to aid to classify. This model considers the different cost of the wronglyclassified malignant and benign samples, which can ensure the high accuracy and thelowest wrongly classification cost. Experiments demonstrate that the proposedmethod is very effective and useful for classifying breast tumors, the classificationresults are more fit the human understanding and the classification accuracy has beensignificantly improved.