Train a 3D U-Net to segment cranial vasculature in CTA volume without manual annotation

Computed tomography angiography (CTA) is now applied as the gold standard in clinical diagnosis of cranial vascular diseases. Segmenting vasculature is a critical step of computer aided diagnosis. In this paper, we adopted a deep learning network architecture 3D U-Net to segment cranial vasculature from CTA images. Different from other traditional methods that require a large amount of manual annotation for network training, we adopted the incomplete vascular segmentation automatically obtained from time-of-flight magnetic resonance angiography (TOF-MRA) volume to train a segmentation network for CTA images. Our results showed that, by carefully tuning the network parameters, relatively complete cranial vascular segmentation can be achieved from CTA volume though the training truth is under-segmented. Our method does not require any human annotation.