Combining Handcrafted and CNN Features for Robust Breast Cancer Detection Using Ultrasound Images

Breast cancer is a disease with an exponentially increasing risk of death and challenging to cure, at advanced stages. Therefore, early detection with adequate precision can considerably increase the probability of survival. Breast cancer diagnosis commonly involves the use of Breast Ultrasound (BUS) imaging. For interpretation of BUS images required expert radiologists. However, unavailability of expert radiologists is a major concern specifically for developing countries like India. Computer-Aided Diagnosis (CAD) systems play a crucial role in aiding clinicians with the early detection of breast cancer. Over the recent years, numerous research efforts have been directed towards devising solutions for detecting breast cancer, with a predominant focus on utilizing Convolutional Neural Networks (CNNs) The CNN based solutions have provided promising results due its automatic high level feature extraction capability; however incorporation of local handcrafted features can improve the classification performance because CNN extracts global features from an image which are not sufficient for accurate classification. Therefore, this paper proposes a breast cancer detection system named Hybrid Expert System for Breast Cancer (HESBC), which uses a combination of local hand crafted features and CNN features. For hand crafted feature Histogram Oriented Gradient (HOG) is utilized whereas the ResNet50 is employed for CNN feature extraction. Finally the fused feature vector is used to classify breast tumors using SVM. The HESBC is being tested using the BUS dataset, and achieving a mean classification accuracy of 85.76% with Precision, Recall, and F1-score of 0.8516, 0.8395, and 0.8428, respectively, using the 10-fold cross-validation method.