Skip to content

SMART: A Flexible, Interpretable, and Scalable Spatio-temporal Brain Atlas from High-Resolution Imaging Data

Home / Papers / SMART: A Flexible, Interpretable, and Scalable Spatio-temporal Brain Atlas from High-Resolution Imaging Data

SMART: A Flexible, Interpretable, and Scalable Spatio-temporal Brain Atlas from High-Resolution Imaging Data

Authors: John Kalkhof, Boris Gutman, Emile d’Angremont, Daniel C. Alexander, Marco Lorenzi Date: 2026-06-17 Paper ID: openalex:2606.18753

Summary

The SMART framework provides a novel approach for constructing spatio-temporal brain atlases from longitudinal high-resolution 3D medical images. By modeling global disease trajectories through region-specific differential equations and personalizing them via multi-scale Neural Cellular Automata, the method addresses existing limitations in interpretability and scalability found in black-box generative models. Evaluated across large longitudinal MRI cohorts including ADNI and OASIS-3, SMART outperforms traditional adversarial and diffusion baselines in both forecasting accuracy and temporal consistency.

Key Contributions

  • Introduced SMART, a spatio-temporal atlas construction framework that decouples global disease dynamics from patient-specific anatomical manifestations using region-specific differential equations.
  • Leveraged multi-scale Neural Cellular Automata for dense, scalable diffeomorphic deformation field modeling in high-dimensional medical image longitudinal data.
  • Demonstrated state-of-the-art disease progression forecasting accuracy and superior temporal consistency compared to diffusion and adversarial baselines on five large-scale MRI datasets.

Key Concepts

  • smart-framework: A spatio-temporal brain atlas framework that learns disease progression trajectories using region-specific differential equations and multi-scale Neural Cellular Automata.

Archivist Review

I approved the SMART framework as a distinct method for longitudinal medical image modeling that separates global disease progression from local patient anatomy. I rejected the Neural Cellular Automata subcomponent because it is an implementation detail of the overarching SMART framework rather than a primary conceptual contribution in its own right. No new open questions or datasets were approved to maintain strict quality and scarcity standards.

Approved Concepts

  • SMART Framework: The framework introduces a novel way to decouple global disease progression dynamics from patient-specific anatomical manifestations in 3D medical imaging.

Rejected Candidates

  • [concept] Multi-scale Neural Cellular Automata for Diffeomorphic Deformation (neural-cellular-automata-for-diffeomorphic-deformation) - subcomponent_of_broader_mechanism: This is a specific architectural subcomponent used to implement the broader SMART framework.

Metadata & Links

url
https://arxiv.org/abs/2606.18753
paper_id
2606.18753
paper_source
openalex
domain
computer-vision
tags
medical-imagingtime-series-forecastinglongitudinal-modelingspatiotemporal-modeling
architectures
datasets
concept_slugs
smart-framework
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-06-20T06:33:18Z