Inverse-Consistent Deep Networks for Unsupervised Deformable Image Registration

1 INTRODUCTION

• Deformable registration establishes dense, nonlinear correspondences that enable anatomical comparison across brain scans.The introduction connects this comparison to studying population variability and longitudinal changes in brain anatomy.Source: §1 · p. 1
• Supervised registration depends on task-specific ground truth, which is difficult to collect and can bias the learned mapping.These constraints limit the practical utility of supervised methods.Source: §1 · p. 1
• Unsupervised methods avoid predefined supervision but commonly estimate asymmetric transformations and omit inverse consistency –.The introduction defines inverse consistency as symmetric deformation of both images until the bidirectionally warped images match.Source: §1 · p. 1
• Increasing smoothness weighting can make registration inaccurate, while weaker weighting can leave foldings in the flow.The figure contrasts flows learned with large and small smoothness-constraint weights [17].Source: §1 · p. 2, Figure 1 · p. 2
• ICNet combines inverse-consistent and anti-folding constraints for unsupervised registration, evaluated on brain MRI tissue segmentation and landmark detection.The authors report superior performance over several state-of-the-art methods.Source: §1 · p. 2

2.1 Deformable Image Registration

• Registration optimizes alignment between a warped source and target while regularization favors desired flow properties such as local smoothness.Methods also differ in deformable models, similarity criteria, and optimization; examples include MSD, SSD, cross-correlation, and mutual information.Source: §2.1 · p. 2

2.2 Non-learning-based Registration Methods

• Non-learning methods use volume intensities or anatomical landmarks and typically require iterative parameter optimization.Their performance can degrade with large anatomical variation, motivating robust and tuning-free methods.Source: §2.2 · p. 2

2.3 Learning-based Registration Methods

• Learning-based methods predict transformations for unseen images from pretrained models and are categorized by whether they require supervision.The literature includes random forests, support vector regression, sparse representation, and deep neural networks.Source: §2.3 · pp. 2–3

2.3.1 Supervised Methods

• Supervised methods train on task-specific ground-truth registrations, whose collection and quality constrain practical utility and performance,,,, –.Deep-learning approaches combine feature extraction and model learning, but typically still require ground truth.Source: §2.3.1 · p. 3

2.3.2 Unsupervised Methods

• Unsupervised methods avoid predefined supervision, but existing deep methods may require predefined features or post-processing and often omit inverse consistency –.Earlier jointly estimated inverse mappings use human-engineered features,,; the paper proposes an unsupervised deep network with inverse-consistent and anti-folding constraints.Source: §2.3.2 · p. 3

3 PROPOSED METHOD

• The section first establishes the notation used in the paper.This prepares the notation for the proposed network and objective.Source: §3 · p. 3
• It presents the proposed network and a training objective that includes inverse-consistent and anti-folding constraints.Source: §3 · p. 3
• The section concludes with implementation details for the proposed method.Source: §3 · p. 3

3.1 Notations

• Images A and B are the two inputs, with F_AB deforming A toward B and F_BA deforming B toward A.Each flow is defined in the space of the image it deforms.Source: §3.1 · p. 3
• Each flow element is a three-dimensional displacement vector indicating a voxel's movement from its original location.The vector components correspond to the x, y, and z axes.Source: §3.1 · p. 3
• The deformed versions of A and B are denoted eA and eB; the illustrated ICNet pipeline takes a pair of images as input.Source: §3.1 · p. 3, Figure 2 · p. 4

3.2 Inverse-consistent Unsupervised Neural Network

• Two fully convolutional modules estimate flows in opposite directions, warping image A toward B and image B toward A.The modules share their structure and parameters.Source: §3.2 · p. 4
• The FCN follows a U-Net architecture with contracting and expanding paths to capture and combine global and local image structure.Its input has two image channels, and its three output channels correspond to the x, y, and z axes.Source: §3.2 · p. 4
• A spatial transformer uses each predicted flow to create a sampling grid and warp its input image with bilinear interpolation.The differentiable sampling process supports back propagation.Source: §3.2 · p. 4
• An inverse network estimates the inverse flow for each learned transformation, using grid sampling based on the flow and its negative.These estimated inverse flows are used by the inverse-consistency losses.Source: §3.2 · p. 4

3.3.1 Inverse-consistent Constraint

• The inverse-consistency constraint makes each of the two directional flows close to the estimated inverse of the other flow.An inverse network generates those estimated inverse mappings; minimizing both terms explicitly models inverse consistency.Source: §3.3.1 · pp. 4–5

3.3.2 Anti-folding Constraint

• The anti-folding constraint penalizes flow gradients at locations where the specified gradient condition indicates folding, while leaving non-folding locations unpenalized.The index function is nonzero when its input is at most zero; the penalty uses the magnitude of the flow gradient plus one.Source: §3.3.2 · p. 5

3.3.3 Smoothness Constraint

• The smoothness constraint regularizes the spatial gradients of both directional flows, approximated by differences between neighboring voxels.The paper uses this conventional constraint alongside its anti-folding constraint.Source: §3.3.3 · p. 5

3.3.4 Objective Function

• The objective uses symmetric mean squared distance to compare each warped image with its corresponding target image.The full loss balances similarity with smoothness, inverse-consistency, and anti-folding regularizers using α, β, and γ, respectively.Source: §3.3.4 · p. 5

3.4 Implementation

• The network is implemented in PyTorch and optimized with Adam and back-propagation, using a learning rate of 5 × 10−4.The validation set contains a random 10% of ADNI-1 subjects, with the remainder used for training.Source: §3.4 · p. 5
• On ADNI-1, the authors report that training and validation losses indicate almost no over-fitting.The figure shows training and validation loss curves.Source: §3.4 · p. 5
• The authors report that the objective converges quickly within 40,000 iterations.Source: §3.4 · p. 5

4.1 Materials and Image Pre-processing

• Experiments use 860 ADNI subjects: 805 from ADNI-1 and 55 from ADNI-2, with 1.5 T scans in ADNI-1 and 3.0 T scans in ADNI-2.Tissue references comprise manually corrected CSF, GM, and WM segmentations; landmark references are five radiologist-annotated points per image.Source: §4.1 · p. 6, Figure 4 · p. 6

4.2 Competing Methods

• ICNet is compared with Demons, SyN, an unsupervised deep-learning method (DL), and variants ICNet-1 and ICNet-2, each omitting one proposed constraint.DL uses an MMSE loss and smoothness constraint; ICNet-1 omits inverse consistency, while ICNet-2 omits anti-folding.Source: §4.2 · p. 6

4.3 Experimental Settings

• Learning-based methods train on ADNI-1 and test on ADNI-2; testing uses five atlas images and 50 testing images, with atlas-based segmentation and landmark detection.The study uses grid search or cross-validation for parameters and evaluates segmentation with five metrics and landmarks by Euclidean distance.Source: §4.3 · p. 7

4.4 Results of Brain Tissue Segmentation

• 83.58% DSC for ICNet versus 79.76% for DL on CSF segmentation; the proposed methods usually achieve the best overall results across tissues and metrics.Visual comparisons also report more accurate registration of the left and right planum temporale with ICNet.Source: Table 1 · p. 7, Figure 5 · p. 7, §4.4 · p. 8

4.5 Results of Anatomical Landmark Detection

• On ADNI landmark detection using deformed atlas images, ICNet achieved an average error of 2.61 mm versus Demons’ 3.16 mm and had superior overall performance to five competing methods.The proposed methods (ICNet-1, ICNet-2, and ICNet) are unsupervised and do not require generating ground-truth flow for each image during network training.Source: §4.5 · p. 8, Table 2 · p. 8

4.6 Computational Cost

• About 0.25 second per MRI for the deep-learning methods, versus about 1 minute for Demons and 2 hours for SyN in application-stage registration.The reported implementations use a GPU for deep-learning methods and a CPU for Demons and SyN.Source: §4.6 · p. 8, Table 3 · p. 8

5.1 Influence of Inverse-consistent Constraint

• With α = 1 and β = 0.1, ICNet produces reasonably smooth, inverse-consistent flows, unlike ICNet-1 settings that yield foldings or over-smoothed flows.The validation losses also indicate that ICNet balances image similarity and inverse consistency better than either ICNet-1 setting.Source: §5.1 · pp. 9–10

5.2 Influence of Anti-folding Constraint

• Without the anti-folding constraint, ICNet-2 produces flow foldings; with it, ICNet effectively avoids them.Source: Figure 8 · p. 9, §5.2 · p. 10

5.3 Influence of Network Architecture

• ICNet 16 with 16 starting filter channels slightly improves tissue segmentation over ICNet with 8, but requires about 20GB rather than 10GB of training memory and doubles training time.The authors characterize the performance improvement as marginal and suggest choosing channel count based on available memory.Source: §5.3 · p. 10, Figure 9 · p. 10

5.4 Influence of Network Refining Strategy

• ICNet R is obtained by refining the trained network on each new testing-image pair, using a learning rate of 1 × 10^-5 for 100 iterations.The learned parameters from training initialize this refinement.Source: §5.4 · p. 10
• ICNet R consistently outperforms ICNet for segmenting three tissue types across five evaluation metrics.For CSF, PPV is 0.87 with ICNet R versus 0.84 with ICNet.Source: §5.4 · p. 10

• ICNet combines an inverse-consistent constraint for symmetric image deformation with an anti-folding constraint to minimize foldings in the estimated flow.Source: §6 · p. 10
• On T1-weighted brain MR registration for tissue segmentation and anatomical landmark detection, ICNet outperforms several state-of-the-art deformable registration algorithms.Source: §6 · p. 10
• The current work measures similarity using MSD and proposes only a simple network-refining strategy for data heterogeneity, while noting other similarity measures and more advanced adaptation algorithms as alternatives.Source: §6 · p. 11

• For neighboring points m and m + 1 along axis i, avoiding folding requires their displaced locations to remain in order.Their displacements are denoted Fi_AB(m) and Fi_AB(m + 1).Source: Figure 11 · p. 11, APPENDIX · p. 11
• At m, the discrete mapping is non-folding when ∇Fi_AB(m) + 1 > 0; a value at or below zero indicates folding.Source: APPENDIX · p. 11
• The anti-folding constraint penalizes the flow gradient only at locations that violate the non-folding rule, using δ(Q) = |Q| for Q ≤ 0 and 0 otherwise.Here, Q = ∇Fi_AB(m) + 1.Source: APPENDIX · p. 11