Ultrasound Image Based Tumor Classification via Deep Polynomial Network and Multiple Kernel Learning

Background: Ultrasound imaging is widely used for tumor detection and diagnosis. Feature extraction plays a critical role in the ultrasound-based computer-aided diagnosis system. Deep Polynomial Network (DPN) is a newly proposed deep learning algorithm, which also has the potential to learn for excellent representation from small dataset. Discussion: However, the final feature representation of DPN is the simple concatenation of the learned hierarchical features from different network layers, which essentially loses some properties exhibited by different network layers, and depresses the representative performance. Since the hierarchical features in DPN can be regarded as heterogeneous multi-view features, they can be effectively integrated by Multiple Kernel Learning (MKL) methods. Conclusion: In this work, we proposed a DPN and MKL based feature learning and classification framework (DPN-MKL) for ultrasound image based tumor diagnosis. The experimental results on breast ultrasound image dataset and prostate ultrasound image dataset show that DPN algorithm has superior performance to the commonly used deep learning algorithms, while the proposed DPNMKL framework outperforms all the single-view feature based algorithms.