The efficacy of combining Deep and Handcrafted Features for Breast Cancer Classification using Ultrasound Images

Breast cancer is a potentially fatal condition, and timely detection plays a vital role in enhancing survival rates. To address this issue and aid clinicians in early detection, Computer Aided systems have been explored. Many researchers have proposed solutions for breast cancer detection, and a popular approach involves using Convolutional Neural Networks (CNNs). CNN-based approaches have displayed encouraging outcomes owing to their capacity to autonomously capture advanced features from medical images. But relying solely on CNN-based global features might lead to suboptimal classification results, as local image details may be overlooked. To improve the classification performance, this paper introduces a breast cancer detection system called the Multi Featured Breast Cancer Detection System (MFBCDS). This system takes advantage of both CNN features and local handcrafted features. Histogram Oriented Gradient (HOG), and Local Binary Pattern (LBP) are utilized for extracting local handcrafted features, while global features are obtained using ResNet50. The MFBCDS model integrates both global and local features to create a comprehensive feature vector. This vector captures essential information from localized regions, complementing the global features extracted by the CNN. Therefore, this combination of features significantly enhances the performance of the model. The combined feature vector is classified using Support Vector Machine (SVM). The evaluation of the proposed MFBCDS model is carried out on the widely used BUSI dataset using a 5-fold cross-validation approach where the proposed model has achieved satisfactory performance on various evaluation matrices with an average accuracy of 88.87%.