Projection network with Spatio-temporal information: 2D+time DSA to 2D aorta segmentation

Aortic dissection (AD) is an acute cardiovascular disease with high mortality and disability rates. AD is commonly treated by the thoracic endovascular aortic repair(TEVAR) which relies on the aorta's position in the Digital Subtraction Angiography (DSA). However, patients and doctors need to be constantly exposed to the X-rays during the DSA process. Besides, the aorta is partially displayed in each frame with blurred boundaries and inhomogeneously distributed contrast agent. The accurate segmentation of AD in DSA is essential for stent placement. This paper proposes a projection network with spatio-temporal information (PNet-ST) for the aortic segmentation of DSA. We introduce a spatial encoder to learn the partial aortic structure information in each frame. Meanwhile, the max intensity projection (MIP) skip connections are used to fuse the temporal information preserved by the encoder to obtain the complete aortic structure. Furthermore, the dense biased connections integrate the multi-receptive field to enhance the network's sensitivity for the multi-resolution feature. The experiment results show that our PNet-ST with segmentation DSC of 0.897, Precision of 0.8757, Recall of 0.9202 and Acc of 0.9684, outperforming the previous image segmentation techniques, such as CE-Net, HRNET and U-Net. The segmentation results of our PNet-ST can offer assistance in endovascular surgery and help the doctors observe the entire aorta to place the stent more accurately.