Source-linked AI summary
FAIM -- A ConvNet Method for Unsupervised 3D Medical Image Registration
Dongyang Kuang, Tanya Schmah
TL;DR
Medical image registration needs complex deformations, while unsupervised neural methods can produce non-invertible foldings. FAIM combines a displacement-field network with explicit anti-folding regularization; on MindBoggle101, it maintains accuracy and folding advantages over VoxelMorph across tested penalty strengths.
Problem
Unsupervised neural registration is fast but can produce non-invertible foldings, which are registration errors and have limited its adoption in clinical applications.
Source: §1 · p. 2Method
FAIM predicts displacement fields with a spatial deformation module and combines image similarity and smoothness losses with an explicit penalty on negative Jacobian determinants.
Source: §2 · p. 3Results
Across tested β values, FAIM has higher registration accuracy and fewer foldings than VoxelMorph, with β = 10^-2 yielding an average of 2 folding voxels per brain and better accuracy than ANTs SyNQuick.
Source: §3.3 · p. 5, §4 · p. 8Takeaways & Limitations
FAIM’s accuracy advantage is consistent across tested β values, and it sacrifices less accuracy than VoxelMorph when reducing foldings.
Source: §3.3 · p. 6
Abstract
from arXiv · showhide
We present a new unsupervised learning algorithm, "FAIM", for 3D medical image registration. With a different architecture than the popular "U-net", the network takes a pair of full image volumes and predicts the displacement fields needed to register source to target. Compared with "U-net" based registration networks such as VoxelMorph, FAIM has fewer trainable parameters but can achieve higher registration accuracy as judged by Dice score on region labels in the Mindboggle-101 dataset. Moreover, with the proposed penalty loss on negative Jacobian determinants, FAIM produces deformations with many fewer "foldings", i.e. regions of non-invertibility where the surface folds over itself. In our experiment, we varied the strength of this penalty and investigated changes in registration accuracy and non-invertibility in terms of number of "folding" locations. We found that FAIM is able to maintain both the advantages of higher accuracy and fewer "folding" locations over VoxelMorph, over a range of hyper-parameters (with the same values used for both networks). Further, when trading off registration accuracy for better invertibility, FAIM required less sacrifice of registration accuracy. Codes for this paper will be released upon publication.
1 Introduction
Medical image registration requires highly nonlinear deformations, but conventional geometric methods are computationally intensive and may not find optimal deformations. Unsupervised neural methods are fast but can produce non-invertible foldings, motivating FAIM’s explicit anti-folding regularization and comparison with VoxelMorph.
1 Introduction
- Geometric methods such as ANTs are computationally intensive, may not find optimal deformations, and optimize each image pair de novo without learning.The passage attributes these challenges partly to highly nonconvex optimization problems.Source: §1 · p. 1
- Unsupervised neural registration can produce non-invertible foldings that constitute clinical registration errors and have limited adoption of these methods.Foldings correspond to negative determinants of the deformation Jacobian; Figure 1 marks them in red.Source: §1 · p. 2, Figure 1 · p. 2
- FAIM is proposed as an unsupervised registration algorithm with explicit anti-folding regularization, compared with U-net-based VoxelMorph on accuracy and folding trade-offs.The experiments use the MindBoggle101 dataset.Source: §1 · p. 2
2 Methods
FAIM uses a spatial deformation module to predict a displacement field that warps a source volume toward a target. Its training objective combines image similarity with smoothness and anti-folding regularization.
2 Methods
- FAIM parameterizes the deformation with displacement field u(x), so the spatial deformation module warps source S to target T as S(x + u(x)).The architecture is inspired by spatial transformer networks; Figure 2 depicts the FAIM architecture.Source: §2 · p. 3, Figure 2 · p. 3
- The spatial deformation module takes stacked moving and target volumes and predicts displacement fields that a sampling module uses to generate the warped source image.Its network uses multiscale comparison, strided convolutions, transposed-convolution upsampling, and three skip connections.Source: §2 · p. 3
- The training loss combines image dissimilarity with αR1(u) for displacement smoothness and βR2(u) to penalize transformations with negative Jacobian determinants.Transformations with only non-negative Jacobian determinants are not penalized by R2; Table 1 presents the loss and regularization functions.Source: §2 · p. 3, Table 1 · p. 4
3 Experiments
The experiments evaluate FAIM and VoxelMorph on Mindboggle101 brain-volume registration, measuring label overlap and folding locations. Across tested regularization strengths, FAIM has higher accuracy and fewer foldings, with a more favorable accuracy–folding trade-off.
3.1 Mindboggle101 dataset
- The study uses 62 Mindboggle101 volumes from three subsets, with images truncated to 144×180×144 voxels; FAIM has about 70% of VoxelMorph’s trainable parameters.The images are already warped to MNI152 space, and the region labels follow the DKT25 variant.Source: §3.1 · p. 4
- The cortical parcellation’s geometrical complexity makes these registration tasks challenging.Source: §3.1 · p. 4
3.2 Evaluation Methods
- Training uses 1,722 ordered pairs from NKI-RS-22 and NKI-TRT-20, while testing uses 380 OASIS-TRT-20 pairs; accuracy is primarily measured by Dice score.Predicted deformations warp source ROI labels to target labels; the section also describes assessing deformation quality by counting locations with negative Jacobian determinants.Source: §3.2 · pp. 4–5
3.3 Results
- Figure 4 shows similar-looking warped labels from all three methods, although their underlying transformations differ.The example depicts one left superior parietal label; the described deformations include thinning and straightening regions.Source: §3.3 · p. 5, Figure 4 · p. 6
- FAIM has higher registration accuracy across the tested β values, while increasing β reduces its predicted folding locations.The experiments compare both networks using the same hyperparameters; Table 2 reports mean Dice and mean folding counts across β.Source: §3.3 · p. 5, Table 2 · p. 6
- FAIM’s flatter Dice-versus-folding curve indicates less accuracy sacrifice than VoxelMorph for a comparable reduction in folding locations.The summary plots also compare Dice across β and show the ANTs SyNQuick mean Dice as a horizontal reference in panels (b) and (c).Source: §3.3 · p. 6, Figure 5 · p. 7
4 Discussion
The discussion presents FAIM as an unsupervised 3D registration method with an optional direct penalty on non-invertible regions. On Mindboggle101, it reports fewer parameters, higher Dice accuracy, and less accuracy sacrifice when reducing foldings than VoxelMorph.
4 Discussion
- FAIM’s advantages over VoxelMorph include fewer trainable parameters, higher Dice accuracy, and less accuracy sacrifice when trading accuracy for fewer foldings.Source: §4 · p. 7
- The authors note that ANTs can produce more accurate registrations with well-chosen hyperparameters, so the reported comparison does not establish a universal accuracy advantage over ANTs.Source: §4 · p. 8