A Lightweight Deep Network for 3D Medical Image Segmentation

Medical image segmentation is an important preprocessing task for computer-aided diagnosis and computer-aided surgery. As medical images such as computed tomography scans provide 3D volumetric data, 3D deep learning (DL) models (e.g., 3DU-Net) have been proposed for medical image segmentation. Compared with conventional 2D DL, 3D DL models have more network parameters, which may lead to overfitting when small training sets are available. In addition, 3D image processing has a high computational cost and consequently a long computation time. To reduce both the number of network parameters and computation time, we propose a lightweight formulation of 3DU-Net based on three-component decomposition. The proposed method can substantially reduce the number of parameters and computation time for 3D image segmentation while maintaining a comparable accuracy with the conventional 3DU-Net.