A 3D Dual Path U-Net of Cancer Segmentation Based on MRI

Nasopharyngeal Carcinoma (NPC) is one of the most common malignant tumors in China. However, the cancer's region is subtle, variability and irregular. In the traditional diagnostic way, clinicians' diagnosis relies on manual delineations which are time consuming and require rich prior experience. Recently, the deep learning architecture of U-Net and Dual Path Network (DPN) apply well in the biomedical segmentation and nature scene respectively. However, U-Net cannot extract abundance texture information from the data and DPN cannot utilize the information of shallow layer and deep layer closely. Moreover, both of them are applied on the slices of images instead of 3D images directly, which discard the anatomic context in 3D spatial domain. Consequently, this paper proposed a novel 3D convolutional network-Dual Path U-Network (DPU) which integrates U-Net and DPN to segment the cancer's region of NPC automatically. The experiment on the MRI dataset of NPC patients has shown that the DPU is more successful than the corresponding 3D version of U-Net and DPN in the field of 3D biomedical image segmentation automatically.