BP-CapsNet: An image-based Deep Learning method for medical diagnosis

Abstract

Deep Learning is one of the promising branches of artificial intelligence and has made phased progress in medical diagnosis. Widely used Convolution Neural Network (CNN) methods require considerable data to train the models to be comparable to professionals. However, it is difficult to acquire medical data in real life to support the application of the CNN methods due to its industrial particularity. Besides, CNN is easy to lose the spatial relationships among the features, which is caused by pooling. The emergence of Capsule Network (CapsNet) effectively alleviates these problems. Nonetheless, CapsNet has an obvious drawback of overlearning: it is prone to learning the noise latent in the image. The named dynamic routing algorithm may lead to the degeneration of capsules on some datasets as the routing iteration increases, making inaccurate diagnoses. This paper proposes a Bayes–Pearson routing CapsNet (BP-CapsNet) with a Singular Value Decomposition (SVD) module to solve the above problems and process medical images to achieve an adequate diagnosis. The proposed method’s effectiveness was evaluated by training and testing the entire model on seven distinct medical image datasets. It achieves state-of-the-art performance on two datasets and the greatest accuracy on another two datasets. The simulation results show that the method can significantly reduce the impact of noise, achieving effective diagnosis and potent generality.