Automated Fetal Brain Volume Reconstruction from Motion-Corrupted Stacks with Deep Learning

The fetal brain MRI 3D volume is critical for malformation diagnosis and development assessment. However, the inevitable fetal motion during MRI acquisition can result in loss of structural continuity and corrupted volumetric information, making it challenging to reconstruct a high-quality 3D volume of the fetal brain from multiple motion-corrupted stacks. Conventional 3D volume reconstruction methods based on slice-to-volume registration are limited to capturing relatively small motion, time-consuming, and sensitive to volume initialization. To address these challenges, we propose a novel end-to-end deep learning method for automated fetal brain MRI 3D volume reconstruction. Firstly, a multi-scale feature fusion model is proposed to solve arbitrary motion correction of 2D slices. It can capture both fine-grained details and global context information at different scales. By fusing features of different scales, the model can accurately estimate fetal motion. Secondly, an initial 3D volume is estimated by point spread function, and an outlier removal strategy is used to reject failed slices. Next, we design a residual-based encoder-decoder model, which can learn an enriched set of features, to improve the quality of the initial 3D volume and reconstruct high-resolution fetal brain MRI 3D volume. We specifically design a perceptual loss, which measures differences based on perceptual features to generate texture and structure information, and introduce an adversarial loss for encouraging anatomically more realistic details. We performed experiments on both simulated and real fetal MR images. The results demonstrate the proposed method is able to solve arbitrary motion correction of 2D slices and reconstruct high-resolution 3D volumes with high efficiency and accuracy, compared to state-of-the-art methods.