Fig. 1

6 regions of intracranial arteries: internal carotid arteries (ICA, blue), basilar artery (BA, green), vertebral artery (VA, purple), middle cerebral artery (MCA, yellow), anterior cerebral artery (ACA, red) and posterior cerebral artery (PCA, cyan)

Table 1

Image acquisition parameters

Fig. 2

Experimental process of this research

Fig. 3

The architecture diagram of DBCNet network. Where Dec is the decoding block of the network, BiA and SC are the branch decoupling module and deep feature extraction module proposed in this study. The final output feature maps $f_{i}^{\text{C}}$ and $f_{i}^{\text{D}}$ of the BiA module are obtained, where C and D represent the localization branch and the segmentation branch, respectively, and i takes 1, 2 and 3 to represent different BiA modules

Fig. 4

Structures of BiA (left) and SC (right) modules

Fig. 5

Training loss and validation loss curves

Fig. 6

3D reconstruction for DBCNet intracranial arterial tree region segmentation prediction results. Row A shows the arterial tree 3D reconstruction (threshold segmentation result of TOF-MRA), labeled real values and segmentation results of a healthy person; row B shows the arterial tree 3D reconstruction, labeled real values and segmentation results of a patient with intracranial aneurysm, the patient’s aneurysm is in the anterior cerebral artery (ACA) region. The visual effect of the real values and model segmentation results is different because the labeled real values are drawn manually using solid spheres, while the model segmentation results are obtained by up-sampling back to the original image size after voxel-level segmentation. Internal carotid arteries (ICA, blue), basilar artery (BA, green), vertebral artery (VA, purple), middle cerebral artery (MCA, yellow), anterior (ACA, red) and posterior cerebral artery (PCA, cyan)

Table 2

Segmentation performance evaluation of each region of the intracranial arterial tree in the testing data set

Fig. 7

Box plots of Dice coefficients and HD95 for each region segmentation in the testing data set

Table 3

Segmentation performance evaluation of each region of the intracranial arterial tree in the testing data set using DBCNet and other common deep learning networks

Fig. 8

Arterial tree segmentations using different deep learning networks for 3D TOF-MRA images of a healthy people subject on MIP

Fig. 9

Arterial tree segmentations using different deep learning networks for 3D TOF-MRA images of intracranial aneurysm patients on MIP

Table 4

Ablation experiment of DBCNet segmentation of cerebral artery region in validation set

Fig. 10

Arterial tree segmentations of ablation experimental for 3D TOF-MRA images of intracranial aneurysm patients on MIP. The red circles in the ground truth image indicate the apparent differences in segmentation results generated by different strategies