Lightweight deep training network for lymph nodes segmentation from head and neck CT images

Accurate lymph node (LN) segmentation is highly beneficial for diagnosing and treating head and neck diseases. However, because of the varying sizes and complex shapes of LNs from the head and neck, as well as their blurred boundaries with surrounding tissues in computed tomography (CT) images, it is difficult for physicians to manually identify the region of interest (ROI). Although existing 3D-volumetric-convolution-based methods play an important role in LN boundary extraction, they suffer from high computational complexity.

To tackle these issues, we develop an efficient and lightweight volumetric convolutional neural network, named LNSNet, for the LN segmentation from the head and neck region.

Our LNSNet presented a 3D Volume Block, which mainly combines Volumetric Partial Convolution (VPConv) with point-wise convolution to decrease computational complexity and parameter count. In addition, both a Lightweight Boundary Enhancement Module (LBEM) and a depthwise separable convolution are added to the bottom of LNSNet to improve the accuracy of LN segmentation.

678 3D LNs extracted from 123 patients with head and neck cancer were used for evaluation. We trained the model using 5-fold cross-validation and tested it on an independent test set. Our model had fewer parameters and lower computational complexity than some state-of-the-art models, with a Dice Similarity Coefficient (DSC) of up to 73.81% and the Average Surface Distance (ASD) and 95th percentile' Hausdorff Distance (HD95) are only 0.92 and 2.52 mm, respectively.

LNSNet improves computational efficiency and robustness by reducing parameter count and complexity, making it more attractive in practical applications.