A Novel 3D U-Net Deep Network with Paralleling Structure for Stroke Lesion Image Segmentation

Image segmentation technologies play a crucial role in medical diagnosis. This paper proposed a novel paralleling structure based on conventional 3D U-net deep network for improving the performance of CT image segmentation. In our model architecture, a new connection channel from analysis path to synthesis path was constructed for exploiting feature maps from deep spatial dimensions. 60 CT scan images of stroke patients were collected for lesion location. Finally, there were 36 valid data were selected for further analysis. The improved method led to better achievement for this task, which segment stroke CT scan images into healthy parts and injury parts. The performance on the test set obtained by our method was compared with other state-of-art U-net models, to demonstrate the effectiveness of our architecture. Furthermore, the result verified that paralleling structure was useful for the convergence of loss curve.