Semi-supervised region-connectivity-based cerebrovascular segmentation for time-of-flight magnetic resonance angiography image

Deep-learning-based methods have achieved state-of-the-art results in cerebrovascular segmentation. However, it is costly and time-consuming to acquire labeled data because of the complex structure of cerebral vessels. In this paper, we propose a novel semi-supervised cerebrovascular segmentation with a region-connectivity -based mean teacher model (RC-MT) from time-of-flight magnetic resonance angiography (TOF-MRA), where unlabeled data is introduced into the training. Concretely, the RC-MT framework consists of a mean teachers (MT) model and a region-connectivity-based model. The region-connectivity-based model dynamically controls the balance between the supervised loss and unsupervised consistency loss by taking into account that the predicted vessel voxels should be continuous in the underlying anatomy of the brain. Meanwhile, we design a novel multi-scale channel attention fusion Unet (MSCAF-Unet) as a backbone for the student model and the teacher model. The MSCAF-Unet is a multi-scale channel attention fusion layer used to construct an image pyramid input and achieve multi-level receptive field fusion. The proposed method is evaluated on diverse TOF-MRA datasets (three clinical datasets and a public dataset). Experimental results show that the proposed method achieves high-performance gains by incorporating the unlabeled data and outperforms competing semi-supervised-based methods. The code will be openly available at https://github.com/IPIS-XieLei/RC-MT.

Automated summary

This paper proposes a novel deep-learning-based semi-supervised cerebrovascular segmentation method, which achieves high-performance gains by incorporating unlabeled data and utilizing a region-connectivity mean teacher model. The proposed method outperforms other semi-supervised methods.