Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning accurate rigid registration for longitudinal brain MRI from synthetic data

Jingru Fu, Adrian V. Dalca, Bruce Fischl, Rodrigo Moreno, Malte Hoffmann

TL;DR

The paper tackles the problem of precise within-subject rigid brain registration for longitudinal MRI by adapting an anatomy-aware, acquisition-agnostic framework to rigid alignment. It introduces a Detector-based architecture that estimates rigid transforms from synthetic intra-subject data, using inverse-consistent, label-based losses and a velocity-field–driven nonlinear augmentation to mimic longitudinal changes. Across ADNI, MIRIAD, and QIN datasets, the method outperforms several baselines, with instance-specific optimization further improving skull-stripped performance and cross-contrast robustness. The work emphasizes task-specific data generation and architecture design for within-subject registration and outlines future directions like dual-input models and robust success/failure metrics.

Abstract

Rigid registration aims to determine the translations and rotations necessary to align features in a pair of images. While recent machine learning methods have become state-of-the-art for linear and deformable registration across subjects, they have demonstrated limitations when applied to longitudinal (within-subject) registration, where achieving precise alignment is critical. Building on an existing framework for anatomy-aware, acquisition-agnostic affine registration, we propose a model optimized for longitudinal, rigid brain registration. By training the model with synthetic within-subject pairs augmented with rigid and subtle nonlinear transforms, the model estimates more accurate rigid transforms than previous cross-subject networks and performs robustly on longitudinal registration pairs within and across magnetic resonance imaging (MRI) contrasts.

Learning accurate rigid registration for longitudinal brain MRI from synthetic data

TL;DR

The paper tackles the problem of precise within-subject rigid brain registration for longitudinal MRI by adapting an anatomy-aware, acquisition-agnostic framework to rigid alignment. It introduces a Detector-based architecture that estimates rigid transforms from synthetic intra-subject data, using inverse-consistent, label-based losses and a velocity-field–driven nonlinear augmentation to mimic longitudinal changes. Across ADNI, MIRIAD, and QIN datasets, the method outperforms several baselines, with instance-specific optimization further improving skull-stripped performance and cross-contrast robustness. The work emphasizes task-specific data generation and architecture design for within-subject registration and outlines future directions like dual-input models and robust success/failure metrics.

Abstract

Rigid registration aims to determine the translations and rotations necessary to align features in a pair of images. While recent machine learning methods have become state-of-the-art for linear and deformable registration across subjects, they have demonstrated limitations when applied to longitudinal (within-subject) registration, where achieving precise alignment is critical. Building on an existing framework for anatomy-aware, acquisition-agnostic affine registration, we propose a model optimized for longitudinal, rigid brain registration. By training the model with synthetic within-subject pairs augmented with rigid and subtle nonlinear transforms, the model estimates more accurate rigid transforms than previous cross-subject networks and performs robustly on longitudinal registration pairs within and across magnetic resonance imaging (MRI) contrasts.
Paper Structure (6 sections, 2 equations, 3 figures, 1 table)

This paper contains 6 sections, 2 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Experiments and results
  4. Discussion
  5. Compliance with ethical standards
  6. Acknowledgments

Figures (3)

  • Figure 1: Representative within-subject registration pairs. We overlay the image moved by each method with the absolute difference between fixed and moved brain masks in yellow. BrainMorph (BM) and SynthMorph (SM) use deep learning.
  • Figure 2: Rigid 3D registration accuracy on the test set, as mean Dice scores over left and right cortex, subcortex, and cerebellum. bbregister (BBR) uses FreeSurfer reconstructions, which include brain masks, and is thus classified as skull-stripped. The asterisk denotes instance-specific optimization after inference.
  • Figure 3: Effect of deformation and smoothing strengths of the non-linear deformation augmentation on registration accuracy in the validation set. Error bars indicate 95$\%$ confidence intervals (CI).