Training Cross-Modality Cerebrovascular Segmentation Networks with Paired Images

Automatic cerebral vessel segmentation from Computed Tomography Angiography (CTA) and Magnetic Resonance Angiography (MRA) is an important task in clinical diagnosis of cranial vascular diseases. With the success of deep convolution neural network, cerebrovascular segmentation has made great progress. But in most of previous work, training the segmentation network in different modalities usually requires a large amount of data to be labeled for each modality, which is expensive and time-consuming. In this paper, we proposed a framework to realize cross-modality cerebrovascular segmentation between CTA and MRA with paired data. We first train the source domain segmentation network, through which the initial segmentation results of the target domain images can be obtained. Then pseudo-label of the target domain is generated by registering the initial segmentation of the target domain with the paired source domain label, and it is used to train the target domain segmentation network. Experiments demonstrate the effectiveness of our proposed framework in cross-modality segmentation.