MIAUnet++: Multi-inception attention network for MR spine image segmentation

Abstract

Accurate segmentation of MR spine images is of great important for the evaluation of spinal diseases. With its unique encoder-decoder symmetric architecture and jump connection design, Unet has become a benchmark model in medical image segmentation since its proposal in 2015. However, the traditional Unet encoder-decoder structure suffers from the problem of semantic information loss during deep feature extraction, and the single-hop connection method is difficult to effectively fuse multi-scale features, resulting in limited segmentation accuracy for complex structures. Responding to these challenges, this study proposes a multi-Inception attention Unet++ (MIAUnet++) model. The model uses different Inception modules to replace the standard convolutional layers of the Unet++, which significantly enhances the multi-scale feature extraction capability by extending the network width and depth. At the same time, multiple attention mechanisms are introduced to further enhance the sensitivity of the network to the boundary information, so that the model can more accurately capture the subtle anatomical structural differences between spine-soft tissues, thus improving the segmentation performance of the network. The experimental results show that the proposed model performs well in the spine image segmentation task with IoU, DSC, TPR and PPV reaching 0.8327, 0.9041, 0.9060 and 0.9068 respectively, outperforming the benchmark method in all metrics. It shows that the proposed method has good performance in MR Spine image segmentation.