Dilated Capsule Network for Brain Tumor Type Classification Via MRI Segmented Tumor Region

Brain tumor recently is considered among the deadliest cancers according to research statistics and have several categories, based on the different characteristics of the tumor. Early detection of the tumor types help to devise treatment plans and achieve high survival rate. Human inspection is noted to be cost effective, error prone and time-consuming, which have led the interest in Convolutional Neural Networks (CNNs) to automatize the problem. However, CNNs fail to consider the precise location of the features as beneficial, which is harmful, because tumor location and its relationship with the surrounding tissue provide high influence on the brain tumor type. In addition, the CNNs require large amount of dataset for accurate training and prediction. CNNs increasingly reduce image resolution, which result to decrease in classification accuracy. In this work, we incorporated recently developed Capsule Networks (CapsNets) which overcome these drawbacks. The focused contribution is to enhance CapsNets with dilation to maintain the image resolution and improve classification accuracy. We proposed a less trainable CapsNet architecture for brain tumor classification, which takes the segmented tumor regions as inputs within the structure and has the capability of ensuring an increase focus of the CapsNets. While the baseline CapsNets consist of single convolutional layer, our proposed model introduced multiple convolutional layers which achieved an improved performance of 95.54% compared to the related works. Our results indicate that the proposed approach can improve brain tumor classification problem.