INFANiTE Implicit Neural Representation for High-Resolution Fetal Brain Spatio-Temporal Atlas Learning from Clinical Thick-Slice MRI

Jul 16, 2026

Upload by Patrick

Spatio-temporal fetal brain atlases enable normative neurodevelopment characterization and early detection of congenital anomalies, yet conventional atlas pipelines demand expensive slice-to-volume reconstruction (SVR) and iterative non-rigid registration, making large-cohort construction impractical. INFANiTE introduces an implicit neural representation framework that learns a high-resolution fetal brain spatio-temporal atlas from clinical thick-slice MRI while bypassing both SVR and iterative registration, substantially accelerating end-to-end processing. Experiments demonstrate top subject-consistency and reference-fidelity, competitive image and tissue-trajectory quality, robustness under sparse data, and reducing time from days to hours versus 3D volume pipelines.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

Abstract

Introduction

INFANiTE: Implicit Neural Representation for High-Resolution Fetal Brain Spatio-Temporal Atlas Learning from Clinical Thick-Slice MRI Xiaotian Hu 1 * , Mingxuan Liu2 * , Hongjia Yang2 * , Juncheng Zhu3 , Yijin Li 1 , Yifei Chen2 , Haoxiang Li2 , Tongxi Song2 , Zihan Li2 , Yingqi Hao2 , Ziyu Li4 , Yujin Zhang3 , Gang Ning3 , Yi Liao3 , Haibo Qu3 , Qiyuan Tian2 1Beihang University, Beijing, China 2Tsinghua University, Beijing, China 3 Sichuan University, Chengdu, China 4University of Oxford, Oxford, United Kingdom [qiyuantian@tsinghua.edu.cn](qiyuantian@tsinghua.edu.cn) arXiv :2605 .09977v2 [ cs .CV] 13 Jul 2026 Abstract Spatio-temporal fetal brain atlases are important for characterizing normative neurodevelopment and identifying congenital anomalies. However, existing atlas construction pipelines necessitate days for slice-to-volume reconstruction (SVR) to generate high-resolution 3D brain volumesand several additional days for iterative volume registration, thereby rendering atlas construction from large-scale cohorts prohibitively impractical. We address these limitations with INFANiTE, an Implicit Neural Representation (INR) framework for high-resolution Fetal brain spatio-temporal Atlas learNing from clinical Thick-slicE MRI scans, bypassing both the costly SVR and the iterative non-rigid registration steps entirely, thereby substantially accelerating atlas construction. Extensive experiments show that INFANiTE achieves the best reported mean subject-consistency and reference-fidelity scores among the evaluated methods, while providing competitive intrinsic image quality and tissuevolume trajectories that broadly agree with normative developmental models. These advantages remain consistent under challenging sparse-data settings. Additionally, INFANiTE reduces the end-to-end processing time (i.e., from raw scans to the final atlas) from days to hours compared to the traditional 3D volume-based pipeline (e.g., SyGN), facilitating large-scale population-level fetal brain analysis. Introduction Prenatal brain development is a critical phase with lasting implications for human neurodevelopment (Li et al. 2025; Hao et al. 2026) . Spatio-temporal atlases constructed from three-dimensional (3D) fetal brain MRI volumes support the identification of atypical brain patterns, providing indispensable insights into potential early manifestations of clinical conditions (Ciceri et al. 2024) . Nevertheless, persistent fetal motion and maternal respiration preclude the direct acquisition of artifact-free 3D MRI. To address the problem, traditional fetal brain atlas construction pipelines (Khan et al. 2019; Gholipour et al. 2017; Wu et al. 2021; Xu et al. 2022) *These authors contributed equally. Copyright © 2026, Association for the Advancement of Artificial Intelligence ([www.aaai.org](www.aaai.org)). All rights reserved. Figure 1: (a) Qualitative comparison of fetal brain atlases at GA 34 weeks (axial, coronal, sagittal), all built directly from clinical thick-slice stacks (with INFANiTE preprocessing) .(b) End-to-end processing time on the 175-subject MultiStack dataset. By replacing SVR with a lightweight sliceto-template registration, INFANiTE cuts total time to 5.8 h, an ≈30× speedup over CINeMA. employ fast 2D imaging sequences (e.g., turbo spin echo, TSE) to preserve high in-plane resolution while freezing intra-shot motion. The resulting multi-planar image stacks are subsequently reconstructed into a 3D brain volume using slice-to-volume reconstruction (SVR) algorithms (Xu et al. 2023; Ebner et al. 2020) . Then, atlases are constructed using SyGN (Avants et al. 2010), an iterative framework that warps individual subjects to an evolving template through non-rigid registration, with the template being refined by averaging the warped images until convergence to achieve astable population representation. However, the heavy computational cost associated with both SVR and iterative registration limits

FAQs

What problem does INFANiTE address in fetal spatio-temporal atlas construction?

INFANiTE targets the high computational burden of traditional pipelines that rely on slice-to-volume reconstruction and iterative non-rigid registration, which makes large-scale cohort atlas building impractical.

How does INFANiTE accelerate atlas construction compared with traditional pipelines?

INFANiTE learns the atlas directly from clinical thick-slice MRI using an implicit neural representation, bypassing both SVR and iterative non-rigid registration. Reported end-to-end time is reduced from days to hours (e.g., 5.8 h in the MultiStack setting).

What performance aspects does INFANiTE improve according to the experiments?

INFANiTE achieves the best reported mean subject-consistency and reference-fidelity scores among evaluated methods, while maintaining competitive intrinsic image quality and tissue-volume trajectories consistent with normative developmental models, including under sparse-data conditions.

Jul 16, 2026

Upload by Patrick