Continuous Spatio-Temporal Representation with Implicit Neural Networks for Fetal Brain MRI Atlas Construction

Download PDF
Open Webpage

Kai Zhang, Dan Hu, Kaibo Tang, Xiuyu Dong, Shijie Huang, Sheng-Che Hung, He Zhang, Zhengwang Wu, Weili Lin, Dinggang Shen, Gang Li

Published: 2025, Last Modified: 07 Nov 2025PIPPI@MICCAI 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Constructing a high-resolution spatiotemporal fetal brain atlas is essential for understanding early neurodevelopmental processes. Many existing atlas building methods assume a predefined temporal distribution, like a Gaussian distribution, and use Gaussian kernel regression for smoothing. However, this limits the ability to capture complex, non-linear developmental trajectories. Meanwhile, these methods suffer from limited resolution dependence. In this work, we leverage implicit neural representations to model continuous functions and jointly optimize both registration and atlas representation for learning a continuous, high-resolution spatiotemporal fetal brain atlas. A key innovation of our method is the explicit incorporation of gestational age as an additional input, allowing the network to learn the continuous deformation of brain structures over time rather than relying on a predefined distribution. We integrate tissue segmentation maps into the learning process to further enhance anatomical accuracy during the construction of atlas, guiding the network to capture structural details better. Additionally, we introduce a set of regularization constraints on the deformation field to ensure anatomical and physical plausibility. Experimental results demonstrate that our method outperforms existing methods, achieving superior anatomical alignment and continuity in both spatial and temporal domains. Compared to other methods, our approach improves anatomical alignment, achieving a higher Dice score (93.8%) and reducing surface distance errors. Specifically, the average symmetric surface distance is reduced to 0.584 mm, and the 95th percentile Hausdorff distance is decreased to 1.618 mm. These results highlight the effectiveness of our method in constructing a high-resolution, temporally consistent and continuous fetal brain atlas.