[Paper Review] Inverse-Consistent Deep Networks for Unsupervised Deformable Image Registration

Jun Zhang|arXiv (Cornell University)|Sep 10, 2018

Medical Image Segmentation Techniques58 references108 citations

TL;DR

ICNet introduces inverse-consistent and anti-folding constraints into an unsupervised deformable registration framework, achieving state-of-the-art performance on brain MRI without ground-truth registrations.

ABSTRACT

Deformable image registration is a fundamental task in medical image analysis, aiming to establish a dense and non-linear correspondence between a pair of images. Previous deep-learning studies usually employ supervised neural networks to directly learn the spatial transformation from one image to another, requiring task-specific ground-truth registration for model training. Due to the difficulty in collecting precise ground-truth registration, implementation of these supervised methods is practically challenging. Although several unsupervised networks have been recently developed, these methods usually ignore the inherent inverse-consistent property (essential for diffeomorphic mapping) of transformations between a pair of images. Also, existing approaches usually encourage the to-be-estimated transformation to be locally smooth via a smoothness constraint only, which could not completely avoid folding in the resulting transformation. To this end, we propose an Inverse-Consistent deep Network (ICNet) for unsupervised deformable image registration. Specifically, we develop an inverse-consistent constraint to encourage that a pair of images are symmetrically deformed toward one another, until both warped images are matched. Besides using the conventional smoothness constraint, we also propose an anti-folding constraint to further avoid folding in the transformation. The proposed method does not require any supervision information, while encouraging the diffeomoprhic property of the transformation via the proposed inverse-consistent and anti-folding constraints. We evaluate our method on T1-weighted brain magnetic resonance imaging (MRI) scans for tissue segmentation and anatomical landmark detection, with results demonstrating the superior performance of our ICNet over several state-of-the-art approaches for deformable image registration. Our code will be made publicly available.

Motivation & Objective

• Motivate unsupervised deformable registration due to scarcity of ground-truth registrations.
• Enforce inverse-consistent property to ensure bidirectional transforms are inverse mappings.
• Prevent folding in the learned transformation with an anti-folding constraint.
• Demonstrate improved tissue segmentation and landmark detection on brain MRI using ICNet.

Proposed method

• Use a dual-branch FCN (U-Net) to learn flows F_AB and F_BA between image pairs with shared weights.
• Incorporate an inverse-consistent loss that enforces F_AB to be close to its estimated inverse and F_BA to its inverse (L_inv).
• Introduce an anti-folding loss that penalizes folding by conditioning on gradient signs indicating folding (L_ant).
• Apply a standard smoothness loss to encourage local regularity (L_smo).
• Optimize a symmetric similarity loss based on MSD between aligned pairs (L_sim) and combine with regularizers (α,L_smo; β,L_inv; γ,L_ant).
• Implement with differentiable spatial transformer (STN) for warping and grid-based inverse flow estimation (G).
• Train end-to-end in PyTorch using Adam (lr = 5e-4) and validate on ADNI brain MRI data.

Experimental results

Research questions

• RQ1 Can an unsupervised Deep Network enforce true inverse-consistency between bidirectional deformation fields?
• RQ2 Does adding an anti-folding constraint reduce flow foldings and improve registration fidelity over smoothness alone?
• RQ3 How does ICNet perform on downstream tasks (tissue segmentation and anatomical landmark detection) compared with non-learning and other unsupervised methods?

Key findings

• ICNet achieves superior registration performance across brain tissues (CSF, GM, WM) versus Demons, SyN, and a baseline DL method.
• For CSF, ICNet achieves DSC 83.58 and SEN 82.55, outperforming Demons (DSC 77.16; SEN 71.99) and SyN (DSC 78.43; SEN 76.57).
• For CSF PPV, ICNet reaches 84.66, higher than Demons (83.16), SyN (80.41), and DL (82.34).
• For CSF ASD and HD, ICNet shows competitive errors with ASD 0.56 and HD 11.10, generally lower than competing methods.
• For GM and WM, ICNet yields the best DSC, SEN, and PPV (e.g., WM DSC 88.13, WM PPV 90.22) and the lowest HD (WM HD 12.71) among the tested methods.
• ICNet variants without inverse-consistent or anti-folding constraints perform worse, confirming the effectiveness of the proposed constraints.