Cycle-consistent Learning for Fetal Cortical Surface Reconstruction.

Fetal cortical surface reconstruction is crucial for quantitative analysis of normal and abnormal prenatal brain development. While there are many cortical surface reconstruction methods available for adults and infants, there remains a notable scarcity of dedicated techniques for fetal cortical surface reconstruction. Of note, fetal brain MR images present unique challenges, characterized by nonuniform low tissue contrast associated with extremely rapid brain development and folding during the prenatal stages and low imaging resolution, as well as susceptibility to severe motion artifacts. Moreover, the smaller size of fetal brains results in much narrower cortical ribbons and sulci. Consequently, the fetal cortical surfaces are more prone to be influenced by partial volume effects and tissue boundary ambiguities. In this work, we develop a multi-task, priori-knowledge supervised fetal cortical surface reconstruction method based on deep learning. Our method incorporates a cycle-consistent strategy, utilizing prior knowledge and multiple stationary velocity fields to enhance its representation capabilities, enabling effective learning of diffeomorphic deformations from the template surface mesh to the inner and outer surfaces. Specifically, our framework involves iteratively refining both inner and outer surfaces in a cyclical manner by mutually guiding each other, thus improving accuracy especially for ambiguous and challenging cortical regions. Evaluation on a fetal MRI dataset with 83 subjects shows the superiority of our method with a geometric error of 0.229 ± 0.047 mm and 0.023 ± 0.058% self-intersecting faces, indicating promising surface geometric and topological accuracy. These results demonstrate a great advancement over state-of-the-art deep learning methods, while maintaining high computational efficiency.