Automatic Head-and-Neck Tumor Segmentation in MRI via an End-to-End Adversarial Network

Among head-and-neck tumors, nasopharyngeal carcinoma (NPC) is the most common type which accounts for high mortality. In the clinical treatment of NPC, magnetic resonance imaging (MRI) has been the primacy method to assess the local and intracranial infiltration of NPC. Due to the time-consuming and labor-intensive nature of NPC in MRI segmentation process, it is desirable to design an accurate and automatic tumor segmentation method. In light of this, we propose a novel end-to-end adversarial network, named as Dense-SegNet based Generative Adversarial Networks (DS-GANs), for NPC segmentation. First, to solve the problem of edge blurring of NPC in MRI images, we propose a hybrid U-shape architecture which integrates the advantages of SegNet and U-net, and the hybrid architecture is named as SU-net. Second, enlightened by the great success of densely connected convolutional networks, we introduce the dense blocks structure to replace the convolution and deconvolution blocks in the proposed SU-net, thus minimizing the number of training parameters while achieving higher performance. The improved composite architecture is referred as DSU-net and employed as the generator network. Third, the traditional adversarial network outputting a single true/false may not match our tumor segmentation task, we introduce a multiscale adversarial network to distinguish both global and local features between the segmented result and the ground truth. Finally, we propose to use a hybrid loss function that utilizes both the multiscale adversarial loss and the Dice loss to train the entire network for better segmentation performance. The effectiveness of the proposed method is evaluated through an in-house NPC dataset of MRI images and better segmentation results are obtained compared with the state-of-the-art segmentation methods.