Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Attention-Driven Framework for Non-Rigid Medical Image Registration

Muhammad Zafar Iqbal, Ghazanfar Farooq Siddiqui, Anwar Ul Haq, Imran Razzak

TL;DR

Deformable medical image registration faces large deformations and anatomical plausibility challenges. The paper introduces AD-RegNet, an attention-driven framework that combines a 3D U-Net backbone, a bidirectional cross-attention module, regional adaptive attention, and multi-resolution deformation field synthesis to estimate a dense displacement field $\ Phi \in \mathbb{R}^{3 \times D \times H \times W}$ with $I_R = \mathcal{T}(I_M, \Phi)$. Training optimizes a multi-term loss $\mathcal{L} = \lambda_{sim} \mathcal{L}_{sim} + \lambda_{reg} \mathcal{L}_{reg} + \lambda_{landmark} \mathcal{L}_{landmark}$, and evaluations on DIRLab and IXI show competitive TRE (1.51 mm) and DSC (0.759) with highly plausible deformations (mean Jacobian near 1 and minimal negative Jacobians). The approach demonstrates cross-modality robustness (lung CT and brain MRI) and positions attention-guided registration as ready for clinical contexts such as disease diagnosis and image-guided interventions.

Abstract

Deformable medical image registration is a fundamental task in medical image analysis with applications in disease diagnosis, treatment planning, and image-guided interventions. Despite significant advances in deep learning based registration methods, accurately aligning images with large deformations while preserving anatomical plausibility remains a challenging task. In this paper, we propose a novel Attention-Driven Framework for Non-Rigid Medical Image Registration (AD-RegNet) that employs attention mechanisms to guide the registration process. Our approach combines a 3D UNet backbone with bidirectional cross-attention, which establishes correspondences between moving and fixed images at multiple scales. We introduce a regional adaptive attention mechanism that focuses on anatomically relevant structures, along with a multi-resolution deformation field synthesis approach for accurate alignment. The method is evaluated on two distinct datasets: DIRLab for thoracic 4D CT scans and IXI for brain MRI scans, demonstrating its versatility across different anatomical structures and imaging modalities. Experimental results demonstrate that our approach achieves performance competitive with state-of-the-art methods on the IXI and DIRLab datasets. The proposed method maintains a favorable balance between registration accuracy and computational efficiency, making it suitable for clinical applications. A comprehensive evaluation using normalized cross-correlation (NCC), mean squared error (MSE), structural similarity (SSIM), Jacobian determinant, and target registration error (TRE) indicates that attention-guided registration improves alignment accuracy while ensuring anatomically plausible deformations.

Attention-Driven Framework for Non-Rigid Medical Image Registration

TL;DR

Deformable medical image registration faces large deformations and anatomical plausibility challenges. The paper introduces AD-RegNet, an attention-driven framework that combines a 3D U-Net backbone, a bidirectional cross-attention module, regional adaptive attention, and multi-resolution deformation field synthesis to estimate a dense displacement field with . Training optimizes a multi-term loss , and evaluations on DIRLab and IXI show competitive TRE (1.51 mm) and DSC (0.759) with highly plausible deformations (mean Jacobian near 1 and minimal negative Jacobians). The approach demonstrates cross-modality robustness (lung CT and brain MRI) and positions attention-guided registration as ready for clinical contexts such as disease diagnosis and image-guided interventions.

Abstract

Deformable medical image registration is a fundamental task in medical image analysis with applications in disease diagnosis, treatment planning, and image-guided interventions. Despite significant advances in deep learning based registration methods, accurately aligning images with large deformations while preserving anatomical plausibility remains a challenging task. In this paper, we propose a novel Attention-Driven Framework for Non-Rigid Medical Image Registration (AD-RegNet) that employs attention mechanisms to guide the registration process. Our approach combines a 3D UNet backbone with bidirectional cross-attention, which establishes correspondences between moving and fixed images at multiple scales. We introduce a regional adaptive attention mechanism that focuses on anatomically relevant structures, along with a multi-resolution deformation field synthesis approach for accurate alignment. The method is evaluated on two distinct datasets: DIRLab for thoracic 4D CT scans and IXI for brain MRI scans, demonstrating its versatility across different anatomical structures and imaging modalities. Experimental results demonstrate that our approach achieves performance competitive with state-of-the-art methods on the IXI and DIRLab datasets. The proposed method maintains a favorable balance between registration accuracy and computational efficiency, making it suitable for clinical applications. A comprehensive evaluation using normalized cross-correlation (NCC), mean squared error (MSE), structural similarity (SSIM), Jacobian determinant, and target registration error (TRE) indicates that attention-guided registration improves alignment accuracy while ensuring anatomically plausible deformations.
Paper Structure (33 sections, 17 equations, 5 figures, 5 tables)

This paper contains 33 sections, 17 equations, 5 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Traditional Deformable Image Registration
  4. Learning-Based Deformable Registration
  5. Hybrid, Multistage, and Groupwise Registration
  6. Attention-Driven and Transformer-Based Registration
  7. Methodology
  8. Problem Formulation
  9. Network Architecture
  10. 3D UNet Backbone
  11. Bidirectional Cross-Attention Module
  12. Regional Adaptive Attention (RAA)
  13. Multi-Resolution Deformation Field Synthesis
  14. Loss Function
  15. Similarity Loss
  16. ...and 18 more sections

Figures (5)

  • Figure 1: Panel A presents representative central axial slices (102 to 106) from the fixed images of Case 1 in the DIRLab dataset, whereas Panel B illustrates the corresponding slices from the moving images of the same case.
  • Figure 2: Panel A shows coronal slices (71–75) from the fixed volume of Case 1 (DIRLab), while Panel B presents the corresponding slices from the moving volume.
  • Figure 3: Panel A shows sagittal slices (71–75) from the fixed volume of Case 1 in the DIRLab dataset, while Panel B presents the corresponding slices from the moving volume.
  • Figure 4: Overview of the proposed Attention-Driven Framework for Non-Rigid Medical Image Registration (AD-RegNet). The network comprises a 3D UNet backbone for feature extraction, a bidirectional cross-attention module for establishing correspondence, a regional adaptive attention mechanism, and a multi-resolution deformation field synthesis module.
  • Figure 5: Qualitative registration results on the DIRLab 4D CT lung and IXI MRI brain datasets. Axial, coronal, and sagittal slices are shown. From left to right: fixed image, moving image, warped moving image, appearance uncertainty maps (coarse-to-fine) visualized as heatmaps, and Jacobian determinant maps (coarse-to-fine), color-coded to highlight anatomical plausibility and consistency of the deformations.