MSAF-Net: Multi-scale attention fusion network for fetal brain tissue segmentation

Tissue segmentation of fetal brain Magnetic Resonance (MR) images is critically important for characterizing prenatal brain development. At present, automatic segmentation methods for fetal brain tissue perform unsatisfactorily. In this paper, we propose a fetal brain segmentation network that combines multi-scale feature fusion and attention mechanisms into the encoder–decoder structure. Specifically, we enhance the Residual and Squeeze-and-Excitation (RaSE) module by combining the channel gating with residual convolution and propose a spatial-attention-based Multi-scale Fusion (MSF) module. The two modules are incorporated into the encoder–decoder architecture, enhancing the model's capacity for feature extraction and feature fusion, thereby effectively improving segmentation performance. Experimental results on the Fetal Tissue Annotation (FeTA) dataset, the Computational Radiology Laboratory (CRL) atlas, and the private dataset show that, compared with state-of-the-art networks for fetal brain segmentation, our method improves segmentation performance, both quantitatively and qualitatively.