Multi-scale multi-hierarchy attention convolutional neural network for fetal brain extraction

Fetal brain extraction from in utero magnetic resonance imaging (MRI) scans is a key step for fetal brain development analysis. As the unpredicted fetal motion and maternal breathing generally result in blurring and ghosting in the slices of phase encoding direction, using the conventional 3D convolutional neural networks for fetal brain extraction with pseudo 3D fetal brain MR scans will lead to sub-optimal extraction performance. To address this issue, in this paper, we propose a novel multi-scale multi-hierarchy attention convolutional neural network (MSMHA-CNN) for fetal brain extraction in MR images. Specifically, to effectively utilize the 3D contextual information of the in utero MR image for fetal brain extraction, we employ multiple convolutional operations with different local receptive fields (i.e., with different kernel sizes) in each layer to learn the multi-scale feature representation for fetal brain extraction. To effectively use the learned multi-scale feature maps, we introduce a channel-wise spatial attention architecture to adaptively fuse those multi-scale feature maps derived from convolutional operations with different kernel sizes. In this way, the learned multi-scale features can be explicitly used to fetal brain extraction process. Besides, to take advantage of high-level feature maps at all spatial resolutions, we adopt the feature pyramid architecture to learn multi-hierarchy features for boosting the performance. We compare our proposed method with several state-of-the-art methods on two in utero MRI scan datasets (a total of 180 scans) for fetal brain extraction. The experimental results suggest the superior performance of the proposed MSMHA-CNN in comparison with its competitors.