Efficient segmentation of medical images using dilated residual networks

Abstract

Medical image segmentation is an essential part in many medical applications such as automatic measurement of tumour size, volume of organs and content-based image retrieval, etc. Various fully convolutional architectures have been proposed in the literature to tackle this problem. Lack of generalization is one of the main challenge in the field of medical imaging and all existing fully convolutional architectures involve huge number of parameters which make them prone to overfit the data. In this study, we proposed an efficient convolutional architecture called Dilated Residual Network (DRN) for medical image segmentation. By the design of DRN architecture, we have reduced number of parameters involved drastically, making the architecture less prone to overfitting hence by improving the generalization ability. We demonstrate that DRN outperforms the previous state of the art architecture called U-Net in medical image segmentation on various datasets in terms of training time, Dice score and Jaccard score. The source code (based on Keras with Tensorflow as the background) of the DRN and sample train and test image results are available at https://github.com/LokeshBonta/Dilated-Residual-Networks.