MFR-Net: Multi-Scale Feature Representation Module for 3D Cerebrovascular Segmentation

Cerebrovascular segmentation of Time-of-Flight magnetic resonance angiography (TOF-MRA) is a necessary step for computer-aided diagnosis. At present 3D U-Net is the most popular 3D medical image segmentation framework, but it can only capture the vascular features of single-size receptive field, and cannot distinguish different structural information of large, medium and small vessels. CNN-Transformer hybrid model requires more labelled datasets to learn effective segmentation, while 3D cerebrovascular annotation is difficult to obtain. In this work, we propose MFR-Net, novelly designed a Multi-scale Feature Representation module to make up for the defect that traditional convolution units only extract single scale features. At the same time, we introduce residual extraction path in skip connection to reduce the encoder-decoder semantic gap. In addition, due to the lack of public 3D cerebrovascular segmentation annotation dataset, we publish the 3D cerebrovascular annotation ground truth of public dataset TubeTK and official data annotation algorithm. Compared with numerous advanced 2D/3D segmentation models and the most advanced deep learning medical image segmentation benchmark nnU-Net , the proposed approach shows better performance. Code and 3D cerebrovascular annotation ground truth of public dataset TubeTK are available at: https://github.com/EllisLyu/TubeTK-Dateset-Annotation.