A Novel Fully Automated MRI-Based Deep-Learning Method for Segmentation of Nasopharyngeal Carcinoma Lymph Nodes

Purpose Nasopharyngeal carcinoma (NPC) is epidemic in south China, especially in Guangdong province. Radiotherapy is the main treatment, with the non-keratinizing type accounting for more than 95% of cases. Metastatic lymph nodes, which should be included in the radiotherapy target volume, are detected among approximately 70-80% of cases when the disease is first diagnosed. Accurate spatial modelling of metastatic lymph nodes is important for successful treatment. Methods We propose a coarse-to-fine deep supervision convolutional neural network (CF-Net) to perform metastatic lymph node segmentation using a 3D residual V-Net. Contrast-enhanced axial T1-weighted (T1C) magnetic resonance images of more than 6000 patients with NPC were enrolled in this study. We used the probability map predicted at a coarse scale as the weight map for training at a fine scale. This method draws attention to a fine scale within an area already detected at a coarse scale. Results CF-Net achieves a median Dice score of 81.0% in the segmentation of metastatic lymph nodes with a sensitivity and specificity of 79.1% and 99.2%, respectively. Conclusion The results show that our method can accurately identify, locate and segment NPC lymph nodes. We compared CF-Net with popular methods: V-Net, DeepLab-v3, HR-Net, and DenseNet. Our proposed method, across all variants, consistently and statistically outperformed the other models.