Feature selection for breast cancer detection from ultrasound images

Breast cancer is the most lethal form of cancer in women after lung cancer. Early detection of cancer is likely to improve the patient's ability to deal with the disease and live further. Ultrasound imaging technique is one of the available tools for cancer diagnosis. Over the years, several high precision features were suggested by researchers to distinguish between malignant and benign lesions. This work employs more than fifty of these features which may serve as a reference feature pool to the researchers. Eventually, we seek to select an optimized subset of this feature set by using three different feature selection methods. In this work, we have successfully employed Multi-Cluster Feature Selection, a recently developed feature selection method, to find a feature set that best describes breast cancer. Thus, we propose a Computer Aided Diagnosis tool with an optimum combination of 25 different features to differentiate between malignant and benign tumors. These features were fed into Sparse Representation Classifier to classify tumors. The proposed technique was examined on ultrasound scans of 504 pathologically diagnosed breast tumors including 454 benign and 50 malignant tumors. The resulting Area Under the Receiver Operating Characteristic Curve was found to be 93.31%.