Impact of Initialization on Intra-subject Pediatric Brain MR Image Registration: A Comparative Analysis between SyN ANTs and Deep Learning-Based Approaches

TL;DR

This paper addresses the challenge of intra-subject, pediatric brain MRI registration by comparing conventional SyN ANTs with a DL-based registration framework under three initialization schemes (NoReg, RR, RAR). The authors implement a Voxelmorph-inspired, unsupervised 3D U-Net registered via DeepReg and evaluate it against SyN ANTs on the Calgary Preschool longitudinal dataset, using 18 SynthSeg-defined regions plus WM/GM/CSF, with metrics including Dice scores and Jacobian determinants. Key findings show that DL-based registration with rigid or rigid+affine pre-alignments achieves Dice scores comparable to or exceeding ANTs (e.g., RR: $0.825\pm0.024$, RAR: $0.913\pm0.018$ vs ANTs $0.818\pm0.017$, $0.895\pm0.023$) and provides substantial inference speedups (22–74×) after training, despite requiring substantial training time ($\sim$127.5 units per dataset). However, performance declines with larger age intervals for all methods, and ANTs remains advantageous when no training is feasible. Overall, the study suggests a practical trade-off: DL-based methods offer fast, accurate registrations for population-specific pediatric datasets, while SyN ANTs provides robust, training-free performance and easier generalization to unseen data. The work highlights the importance of selecting registration methods based on the scope (global vs. local structures) and provides a public codebase for reproducibility.

Abstract

This study evaluates the performance of conventional SyN ANTs and learning-based registration methods in the context of pediatric neuroimaging, specifically focusing on intrasubject deformable registration. The comparison involves three approaches: without (NR), with rigid (RR), and with rigid and affine (RAR) initializations. In addition to initialization, performances are evaluated in terms of accuracy, speed, and the impact of age intervals and sex per pair. Data consists of the publicly available MRI scans from the Calgary Preschool dataset, which includes 63 children aged 2-7 years, allowing for 431 registration pairs. We implemented the unsupervised DL framework with a U-Net architecture using DeepReg and it was 5-fold cross-validated. Evaluation includes Dice scores for tissue segmentation from 18 smaller regions obtained by SynthSeg, analysis of log Jacobian determinants, and registration pro-rated training and inference times. Learning-based approaches, with or without linear initializations, exhibit slight superiority over SyN ANTs in terms of Dice scores. Indeed, DL-based implementations with RR and RAR initializations significantly outperform SyN ANTs. Both SyN ANTs and DL-based registration involve parameter optimization, but the choice between these methods depends on the scale of registration: network-based for broader coverage or SyN ANTs for specific structures. Both methods face challenges with larger age intervals due to greater growth changes. The main takeaway is that while DL-based methods show promise with faster and more accurate registrations, SyN ANTs remains robust and generalizable without the need for extensive training, highlighting the importance of method selection based on specific registration needs in the pediatric context. Our code is available at https://github.com/neuropoly/pediatric-DL-registration

Figures (18)

• Figure 1: Illustration of the three initialization strategies, NoReg (NR), RigidReg (RR) as well as RigidAffineReg (RAR), (blue) used for comparing deep learning (green) and conventional SyN ANTs (red) registration approaches.
• Figure 2: Schema of the training procedure to obtain a deformation field ($\phi$) with given moving (M) and fixed (F) 3D pair of images. The validation technique using WM, GM and CSF segmentations (WM depicted in this figure) to calculate Dice scores is also shown in the blackdashed region as well as the loss function in the upper right corner where $v$ indicates voxels for the L2 norm of the displacement gradient, $\nabla u$, which encourages a smooth deformation field. $\phi$ is calculated by adding the identity transform to the displacement field ($\phi = Id + u$). Image inspired by Balakrishnan2019-sr.
• Figure 3: Dice score results on the test sets represented as boxplots for each initialization approach (NoReg, RigidReg and RigidAffineReg) for DL-based methods compared to the initial Dice scores pre-conducting the registration steps. Each method is also compared to the SyN ANTs registration. The Dice scores are averaged over all 18 segmented regions. The table in the lower right corner shows the mean±SD Dice scores for all scenarios.
• Figure 4: Visual representation of the results obtained with all three DL-based approaches ( moved network) compared to SyN ANTs ( moved ANTs) results with an age interval between moving and fixed images of 0.116 years on the left and 3.37 years on the right. Red arrows highlight instances of misalignment, yellow arrows indicate blurriness or minor deviations from the fixed image, and green arrows denote successfully aligned areas.
• Figure 5: Dice scores against age intervals for all initialization methods compared to SyN ANTs. The three rows correspond to NoReg, RigidReg, and RigidAffineReg, and the three columns represent WM, GM, and CSF segmentations. SyN ANTs Dice scores are shown in red, while the results of DL-based approaches are in green, with corresponding trendlines in the same color. Dice scores in global regions are calculated by averaging SynthSeg sub-regions within these tissues from the total 18 regions available. Figure titles include coefficients of determination (R-squared) for reference. Note: the y-axis scale for NoReg differs due to a distinct range of Dice scores.