Deep learning-based fetal brain extraction method for in utero diffusion MRI

Motivation: As a critical step in diffusion MRI post-processing, fetal brain extraction remains challenging due to highly anisotropic resolution and age-related brain volume changes in in utero dMRI. Goal(s): We aimed to develop a deep learning-based model to accurately extract the fetal brain from in utero dMRI images. Approach: We introduced the Multi-Scale-Anisotropy Network (MSA-Net), which incorporates convolutional blocks of different dimensions to create an asymmetric receptive field and combines features from adjacent layers to capture multiscale information. MSA-Net was evaluated on two datasets with different resolutions. Results: Results show that MSA-Net achieves accurate fetal brain extraction (mean Dice > 0.95) with strong generalization. Impact: The proposed method can significantly streamline the tedious annotation process and improve segmentation accuracy, contributing to a fast and accurate post-processing pipeline.