A ConvMixEst and Multi-attention UNet for Intervertebral Disc Segmentation in Multi-modal MRI.

Accurate segmentation of spinal magnetic resonance imaging (MRI) plays a critical role in the diagnosis and evaluation of intervertebral discs. However, accurate disc segmentation is not easy due to the overly trivial and high similarity of disc tissues, as well as the huge variability between slices, and multi-modal MRI disc segmentation is even more challenging. For this reason, we propose a model called ConvMixEst and Muti-Attention Unet (CAM-Unet), which combines MLP with the attentional mechanisms of Inverted Variational Attention (IVA) and Dilated Gated Attention (DGA). Specifically, in this work, we propose the IVA module for detailing the overall feature information of the dataset and capturing feature information at different scales, and also design a ConvMixEst for enhancing the global context information. After doing trials on the MICCAI-2018 IVD challenge dataset, we obtain Dice similarity coefficient equal to 92.53(%) and Jaccard coefficient equal to 86.10(%) and Precision equal to 94.09(%).