[Paper Review] Joint segmentation and discontinuity-preserving deformable registration: Application to cardiac cine-MR images

Xiang Chen, Yan Xia+2 more|arXiv (Cornell University)|Nov 24, 2022

Medical Image Segmentation Techniques4 citations

TL;DR

This paper proposes SDDIR, a weakly-supervised deep learning framework that jointly performs cardiac MR image segmentation and discontinuity-preserving deformable registration. By integrating a co-attention module for segmentation and a discontinuity-aware regularization strategy in the registration branch, SDDIR achieves superior registration accuracy over state-of-the-art methods on cardiac cine-MRI datasets, particularly when high-quality segmentation masks are predicted.

ABSTRACT

Medical image registration is a challenging task involving the estimation of spatial transformations to establish anatomical correspondence between pairs or groups of images. Recently, deep learning-based image registration methods have been widely explored, and demonstrated to enable fast and accurate image registration in a variety of applications. However, most deep learning-based registration methods assume that the deformation fields are smooth and continuous everywhere in the image domain, which is not always true, especially when registering images whose fields of view contain discontinuities at tissue/organ boundaries. In such scenarios, enforcing smooth, globally continuous deformation fields leads to incorrect/implausible registration results. We propose a novel discontinuity-preserving image registration method to tackle this challenge, which ensures globally discontinuous and locally smooth deformation fields, leading to more accurate and realistic registration results. The proposed method leverages the complementary nature of image segmentation and registration and enables joint segmentation and pair-wise registration of images. A co-attention block is proposed in the segmentation component of the network to learn the structural correlations in the input images, while a discontinuity-preserving registration strategy is employed in the registration component of the network to ensure plausibility in the estimated deformation fields at tissue/organ interfaces. We evaluate our method on the task of intra-subject spatio-temporal image registration using large-scale cinematic cardiac magnetic resonance image sequences, and demonstrate that our method achieves significant improvements over the state-of-the-art for medical image registration, and produces high-quality segmentation masks for the regions of interest.

Motivation & Objective

• To address the limitation of existing deep learning-based registration methods that enforce globally smooth deformation fields, which can produce implausible results at tissue/organ boundaries.
• To develop a joint segmentation and registration framework that preserves discontinuities at organ interfaces while maintaining local smoothness in deformation fields.
• To reduce dependency on pre-existing segmentation masks by learning structural correlations directly from image pairs during training.
• To improve registration accuracy and anatomical plausibility in intra-subject cardiac cine-MRI registration, especially in regions with complex motion such as the heart.
• To enable robust, real-time quantification of cardiac motion and strain by producing accurate, discontinuity-aware deformation fields.

Proposed method

• A U-Net-based segmentation network with a co-attention block is used to learn structural correlations between input image pairs, enhancing segmentation accuracy.
• A discontinuity-preserving regularization strategy is introduced in the registration branch to allow locally smooth but globally discontinuous deformation fields.
• The network is trained in a weakly-supervised manner using image similarity losses (e.g., normalized cross-correlation) and segmentation supervision via Dice loss.
• The deformation field is estimated using a U-Net-like architecture with skip connections, and discontinuities are modeled via a boundary-aware regularizer that penalizes smoothness across tissue boundaries.
• The segmentation and registration branches are jointly optimized, with segmentation masks used to guide region-wise deformation estimation for improved anatomical consistency.
• The framework is end-to-end trainable and does not require ground-truth deformation fields or pre-segmented masks for inference.

Experimental results

Research questions

• RQ1 Can joint segmentation and registration improve the accuracy and plausibility of deformation fields in cardiac cine-MRI registration?
• RQ2 How does enforcing local smoothness with global discontinuities at tissue boundaries affect registration performance compared to globally smooth deformation fields?
• RQ3 To what extent does the quality of predicted segmentation masks influence the final registration accuracy in a joint learning framework?
• RQ4 Can a weakly-supervised deep learning model achieve state-of-the-art registration performance without requiring pre-segmented masks or ground-truth deformation fields?
• RQ5 How robust is the proposed method to domain shifts across different MRI acquisition protocols and scanners?

Key findings

• SDDIR significantly outperforms state-of-the-art deep learning and traditional registration methods on three cardiac cine-MRI datasets in terms of registration accuracy and anatomical plausibility.
• The method achieves superior registration performance when the segmentation masks for both fixed and moving images have a Dice score above 76% (ACDC) and 70% (MM), respectively.
• Registration accuracy of SDDIR surpasses the baseline network only when high-quality segmentation masks are predicted, highlighting the strong dependency on segmentation quality.
• The joint learning framework produces more anatomically plausible segmentation masks by leveraging both fixed and warped moving image predictions.
• SDDIR demonstrates improved performance on the UK Biobank (UKBB) dataset due to consistent image appearance, but performance degrades significantly on external datasets (e.g., ACDC, MM) due to domain shifts.
• The results indicate that future improvements should focus on domain generalization and reducing reliance on high-quality segmentation for accurate registration.