ConvCaps: Multi-input Capsule Network for Brain Tumor Classification

Brain tumor is considered one of the most deadly cancer. Determining the type of brain tumor has a significant effect on the choice of treatment and improves the survival rate of the patients. In general, manual diagnosis is time-consuming and error-prone. With the development of deep learning, especially the convolutional neural network (CNN), the process of automatic classification and diagnosis of medical images has been greatly promoted. However, CNN’s pooling operation decreases a lot of spatial relationships that play important roles in discriminating tumor types, resulting in inaccurate classification results. Capsule Network (CapsNet) is a novel architecture proposed in recent years to overcome shortcomings of CNN. However, it cannot handle inputs in large size. To tackle this problem, a Convolutional Capsule Network (ConvCaps) architecture for brain tumor classification is proposed which has the following properties: (1) It can accept large-size images as input without scaling them in advance. (2) Preserving the spatial relationships of components in the image. (3) Multiple convolutional layers are added in front of the primary capsule layer so that the network can extract low-level features. (4) Image of brain tumor region is fed into our model as extra input to improve the network’s attention to the region of interest. The experimental results on the brain MR dataset show that our model improves significantly compared with the previous work, the classification accuracy increases to 93.5%, and the training speed is also improved.