FetalCSR: Multi-input Attention Fusion Network for Neural ODE-based Fetal Cortical Surface Reconstruction

Cortical surface reconstruction (CSR) aims to quantitatively represent, visualize and analyze the 3D structure of the cerebral cortex by generating its inner and outer surface meshes from brain MRI scans. Various learning-based methods have been proposed to address the long runtimes required by traditional image processing-based CSR techniques like FreeSurfer, mostly for adults and infants. Fetal brain CSR is critical for assessing prenatal brain development but faces unique challenges, such as low imaging resolution, poor tissue contrast, severe motion artifacts, and smaller brain size with narrower cortical ribbons and sulci. Additionally, fetal brains undergo rapid changes during development, imposing another major challenge for existing learning-based methods. To address these challenges, we propose FetalCSR, a multi-input attention fusion deep learning framework specifically designed for CSR from fetal brain MRI data. FetalCSR is based on neural ordinary differential equations (ODEs), learns a diffeomorphic flow to deform an input surface into a target shape. Specifically, FetalCSR integrates positional encoded coordinates, sampling of the MRI data along the surface normal and from the surrounding volume, and their tissue segmentation masks. Experiments on the Developing Human Connectome Project (dHCP) dataset and a private fetal brain MRI dataset with 40 subjects demonstrate that FetalCSR achieves improvements over existing approaches in both average symmetric surface distance (ASSD) and Hausdorff distance (HD). Comprehensive ablation studies further validate the benefits of the multi-input and cross-attention mechanisms for feature fusion. Pretrained models will be released upon paper acceptance.