Residual-atrous attention network for lumbosacral plexus segmentation with MR image.

Accurate segmentation of the lumbosacral plexus is a crucial step for diagnosis and analysis of nerve damage in clinical. Due to the extremely low contrast and complicated structure around the lumbosacral plexus, it has been remaining a challenging task to effectively segment the lumbosacral plexus from spinal MR images. Even though several deep learning methods for spine segmentation have been developed, most of them only pay attention to the segmentation of vertebral bodies and intervertebral discs rather than nerves. To solve these problems, in this paper, we propose a residual-atrous attention network (RA-Net) for lumbosacral plexus segmentation with MR images. Specifically, the RA-Net consists of three main parts, (1) the atrous encoder module is employed to learn multi-scale contextual features from MR images in the encoder, (2) the residual skip connection operation is used to integrate the features with high-resolution spatial details in the encoder and the high-level contextual features in the decoder, and (3) the scale attention block is proposed for fusing the multi-scale high-level features in the decoder. We perform our proposed RA-Net for the lumbosacral plexus segmentation on the collected spinal MRI dataset with 10 patients (a total of 236 MRI scans). Extensive experiments demonstrate that our RA-Net achieves better performance in lumbosacral plexus segmentation with MR images when compared with several state-of-the-art methods.