Cerebral Arterial Stenosis Detection Based on a Retained Two-Stage Detection Algorithm

Figure 2

Flowchart of the retrained algorithm in this study. CTA images are fed into the backbone, consisting of Resnet50, deformable convolutional networks, and nonlocal neural networks incorporated in the third, fourth, and fifth stages of Resnet50. Feature maps from feature pyramid network and region proposal network with multiscale and proposal region information are integrated and then input to the cascade detector to refine the classification and regression by increasing the IoU thresholds in stages.