3D Partial U-Net: A Lightweight ConvNet for Head and Neck Lymph Node Segmentation

Accurate lymph node (LN) segmentation plays a crucial role in tumor diagnosis and treatment for head and neck cancer patients. Automatic LN segmentation remains challenging due to large size variation, extracapsular extension, and similar appearance to surrounding vessels on computed tomography (CT). The morphology and volume of LNs are crucial for diagnosis, while the current 3D convolution-based methods are well used for boundary extraction. However, existing volumetric convolutional neural networks (ConvNets) are parameter-heavy, computationally complex and slow to use when dealing with 3D medical images. To this end, we introduce a lightweight volumetric ConvNet termed 3D Partial U-Net to facilitate LN segmentation applications on head and neck cancer patients. Our approach introduces a combination of volumetric partial convolution and pointwise convolution, which significantly reduces the parameter count and computational complexity. A light boundary refinement output module with a large kernel depth-wise separable convolution placed at the end of our model is proposed to enhance the precision of LN segmentation. Compared to the symmetric U-Net, our model features a larger encoder and a smaller decoder, empowering it with enhanced capability to learn complex LN features. We evaluate the model performances on 678 LNs from 123 head and neck cancer patients. The results of experiments showcase that 3D Partial U-Net surpasses existing approaches with the fewest floating-point operations (FLOPs) and parameters, achieving state-of-the-art. The source code with our proposed 3D Partial U-Net is available at https://github.com/cii030/3D-Partial-U-Net .