The aneurysm and vessel segmentation for 3D digital subtraction angiography images

Project overview

For this project, the No New U-Net (nnUNet) was used for accurate segmentation of 3D digital subtraction angiography images. The dice coefficient was taken as the evaluation standard, and the accuracy had reached above 90 percent.

Model selection

In this experiment, we chose 3D nnUNet for aneurysm segmentation. Compared with the common 3D UNet model, nnUNet has made great innovations in data set preprocessing. And achieved good results in various medical imaging competitions.


The segmentation dice accuracy

ModelVessel dice on validation set (mean/std)Aneurysm dice on validation set (mean/std)
nnUNet (patch size = 192)0.933/0.0510.841/0.206
nnUNet (patch size = 96)0.934/0.0570.905/0.093

Visualize the segmentation

We visualize the segmentation with the Visualization Toolkit (VTK).

The left image is the model predict and the right image is the ground truth.

Example image
Example image