V2C-Long: Longitudinal Cortex Reconstruction with Spatiotemporal Correspondence

TL;DR

V2C-Long addresses the core challenge of comparing cortical morphology over time by enforcing inherent spatiotemporal correspondence during reconstruction. The method couples two sequential template-deformation networks to create robust within-subject templates in mesh space and then deform them across visits, yielding directly comparable vertex-wise surfaces without post-processing. Across ADNI, OASIS, and TRT data, V2C-Long delivers superior longitudinal consistency and surface accuracy compared with Strong baselines, and demonstrates stronger evidence of Alzheimer's disease–related cortical atrophy, along with higher diagnostic AUC. This approach enables more sensitive, scalable, and clinically applicable longitudinal brain analyses with reduced processing time and fewer compatibility issues across pipelines.

Abstract

Reconstructing the cortex from longitudinal magnetic resonance imaging (MRI) is indispensable for analyzing morphological alterations in the human brain. Despite the recent advancement of cortical surface reconstruction with deep learning, challenges arising from longitudinal data are still persistent. Especially the lack of strong spatiotemporal point correspondence between highly convoluted brain surfaces hinders downstream analyses, as local morphology is not directly comparable if the anatomical location is not matched precisely. To address this issue, we present V2C-Long, the first dedicated deep learning-based cortex reconstruction method for longitudinal MRI. V2C-Long exhibits strong inherent spatiotemporal correspondence across subjects and visits, thereby reducing the need for surface-based post-processing. We establish this correspondence directly during the reconstruction via the composition of two deep template-deformation networks and innovative aggregation of within-subject templates in mesh space. We validate V2C-Long on two large neuroimaging studies, focusing on surface accuracy, consistency, generalization, test-retest reliability, and sensitivity. The results reveal a substantial improvement in longitudinal consistency and accuracy compared to existing methods. In addition, we demonstrate stronger evidence for longitudinal cortical atrophy in Alzheimer's disease than longitudinal FreeSurfer.

Figures (13)

• Figure 1: V2C-Long inherently establishes spatiotemporal correspondence of cortical surfaces during the reconstruction. This renders all reconstructed surfaces, i.e., within a certain subject and across subjects, directly comparable on the vertex level.
• Figure 1: Medial view on differences in longitudinal cortical thickness between stable AD subjects ($n=71$) and healthy controls ($n=193$). We show uncorrected negative log10(p-value)-maps on FsAverage --- based on meshes from FS-Long (v7.2), V2CC, and V2C-Long.
• Figure 2: Architecture of V2C-Long. From a sequence of 3D brain MRI scans, V2C-Long computes a within-subject template, enriches it with vertex features from the template-creation model, and deforms it to cortical surfaces with strong spatiotemporal point correspondence.
• Figure 2: Conceptual comparison of V2C-Long with the longitudinal FreeSurfer pipeline.
• Figure 3: Region-wise consistency of reconstructed surfaces based on the Destrieux atlas (mean of WM and pial surfaces, ADNI test set). a) Lateral view of per-region ParcF1 scores (higher is better) from V2C-Long. b) Comparison of ParcF1 scores for all 75 regions from different methods. c) Medial view of per-region ParcF1 scores from V2C-Long. We also indicate the location and name of three regions in the atlas and the polar chart. A list of all Destrieux regions in the order of the plot (counter-clockwise) is in the Supplementary Material. For plotting, we used Pyvista (v0.35.2) and Plotly (v5.22.0).

Theorems & Definitions (1)

• Theorem 1