Breast cancer diagnosis using modified Xception and stacked generalization ensemble classifier
Sagar Deep Deb, Aqhlaqur Rahman, Rajib Kumar Jha · Research on Biomedical Engineering · 2023
Cancer is caused when abnormal cells grow severely, go beyond their normal regions to invade adjoining parts of the body, and spread to other organs. It can start in any part of the body. According to the American Association for Cancer Research (AACR), breast cancer can seriously affect women’s physical and mental health. In some cases, it can even become a threat to life. Thankfully, early detection followed by proper medical treatment can save lives. Presently, mammography is one of the most common criteria for radiologists to diagnose breast cancer. However, the complexity in the mammogram images and lack of expertise lead to a lot of misclassification. In this study, we aim to propose a deep learning-based architecture for proper classification and detection of breast cancer. A modified Xception network is proposed for the accurate detection of breast cancer. The proposed network takes ROIs as input. It extracts features from different intermediate convolutional layers of the Xception network. This makes the final features extracted, a fusion of different low and high-level descriptors. Finally, a stacked generalization ensemble classifier is used for classification. We have successfully proved that employing such a technique can enhance the classification performance by about 8%. We have successfully proved that using the modified Xception network with the deep learning-based stacked generalization ensemble classifier enhances the detection performance. We have reported an accuracy of 84.3%, whereas a plain Xception network reports an accuracy of 78%. For reproducibility, the pre-processed dataset and the source code are made available at https://github.com/sagardeepdeb/rahman_xception_global . The quantitative and qualitative comparison of this work with various state-of-the-art research demonstrates that concatenating features from different intermediate layers of the proposed network makes the final features more discriminatory. It also shows that the deep learning-based stacked generalized ensemble classifier enhances the classification performance.