MRI Image Segmentation of Nasopharyngeal Carcinoma Using Multi-Scale Cascaded Fully Convolutional Network

Nasopharyngeal carcinoma (NPC) is one of the most common malignant tumors head and neck, its incidence highest all around world. Intensive radiotherapy using computer-aided diagnosis best technique for treatment NPC. The key step delineation target areas organs at risk, that is, tumor images segmentation. We proposed segmentation method NPC image based on multi-scale cascaded fully convolutional network. It used network feature a coarse-to-fine to improve effect. In coarse segmentation, blocks data augmentation were compensate shortage training samples. fine Atrous Spatial Pyramid Pooling (ASPP) was increase receptive field transfer, which added in Dense block DenseNet. process up-sampling, features multiple views fused reduce false positive Additionally, order class imbalance problem, Focal Loss weight loss function voxel distance because it could background category can alleviate problem gradient disappearance obtain smoother boundary. experimental results quantitatively analyzed by DSC, ASSD F1_score values, showed effective nasopharyngeal compared with other methods this paper.