Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FF-PNet: A Pyramid Network Based on Feature and Field for Brain Image Registration

Ying Zhang, Shuai Guo, Chenxi Sun, Yuchen Zhu, Jinhai Xiang

TL;DR

FF-PNet introduces a pyramid-based, unsupervised brain image registration framework that decouples coarse feature encoding from fine deformation refinement through two residual fusion modules. RFFM strengthens context-feature decoding, while RDFFM refines deformation fields via cross-layer fusion, enabling parallel coarse-to-fine alignment without relying on attention or transformers. Across LPBA and OASIS brain MRI datasets, FF-PNet demonstrates superior registration accuracy, evidenced by higher Dice similarity and lower HD95 and SDlogJ compared with both U-Net–based and pyramid-based baselines. The approach is lightweight, generalizable, and readily integrable into other pyramid architectures to enhance non-rigid registration performance.

Abstract

In recent years, deformable medical image registration techniques have made significant progress. However, existing models still lack efficiency in parallel extraction of coarse and fine-grained features. To address this, we construct a new pyramid registration network based on feature and deformation field (FF-PNet). For coarse-grained feature extraction, we design a Residual Feature Fusion Module (RFFM), for fine-grained image deformation, we propose a Residual Deformation Field Fusion Module (RDFFM). Through the parallel operation of these two modules, the model can effectively handle complex image deformations. It is worth emphasizing that the encoding stage of FF-PNet only employs traditional convolutional neural networks without any attention mechanisms or multilayer perceptrons, yet it still achieves remarkable improvements in registration accuracy, fully demonstrating the superior feature decoding capabilities of RFFM and RDFFM. We conducted extensive experiments on the LPBA and OASIS datasets. The results show our network consistently outperforms popular methods in metrics like the Dice Similarity Coefficient.

FF-PNet: A Pyramid Network Based on Feature and Field for Brain Image Registration

TL;DR

FF-PNet introduces a pyramid-based, unsupervised brain image registration framework that decouples coarse feature encoding from fine deformation refinement through two residual fusion modules. RFFM strengthens context-feature decoding, while RDFFM refines deformation fields via cross-layer fusion, enabling parallel coarse-to-fine alignment without relying on attention or transformers. Across LPBA and OASIS brain MRI datasets, FF-PNet demonstrates superior registration accuracy, evidenced by higher Dice similarity and lower HD95 and SDlogJ compared with both U-Net–based and pyramid-based baselines. The approach is lightweight, generalizable, and readily integrable into other pyramid architectures to enhance non-rigid registration performance.

Abstract

In recent years, deformable medical image registration techniques have made significant progress. However, existing models still lack efficiency in parallel extraction of coarse and fine-grained features. To address this, we construct a new pyramid registration network based on feature and deformation field (FF-PNet). For coarse-grained feature extraction, we design a Residual Feature Fusion Module (RFFM), for fine-grained image deformation, we propose a Residual Deformation Field Fusion Module (RDFFM). Through the parallel operation of these two modules, the model can effectively handle complex image deformations. It is worth emphasizing that the encoding stage of FF-PNet only employs traditional convolutional neural networks without any attention mechanisms or multilayer perceptrons, yet it still achieves remarkable improvements in registration accuracy, fully demonstrating the superior feature decoding capabilities of RFFM and RDFFM. We conducted extensive experiments on the LPBA and OASIS datasets. The results show our network consistently outperforms popular methods in metrics like the Dice Similarity Coefficient.
Paper Structure (15 sections, 7 equations, 7 figures, 5 tables)

This paper contains 15 sections, 7 equations, 7 figures, 5 tables.

Table of Contents

  1. Introduction
  2. related work
  3. U-Net-based Registration Method
  4. Pyramid-based Registration Method
  5. methods
  6. The Overview of FF-PNet
  7. Residual Feature Fusion Module
  8. Residual Deformation Field Fusion Module
  9. Loss Function
  10. experiments
  11. Datasets and Evaluation Metrics
  12. Experimental Settings
  13. Experimental Results and Analysis
  14. Ablation experiments
  15. conclusion

Figures (7)

  • Figure 1: One example to show the deformable image registration. a non-rigid deformation field (c) is estimated to warp the moving image (b) to match with the fixed image (a). (d) shows the warped moving image.
  • Figure 2: The overall architecture of our FF-PNet. It is a three-layer pyramid network composed of a hierarchical feature extraction encoder based on CNN and a coarse-to-fine registration decoder based on two fusion modules (including RFFM and RDFFM).
  • Figure 3: This figure illustrates the detailed implementation of RFFM. The first row in the figure shows the residual feature extraction operations, and the second row describes the composition of the cross-fusion module.
  • Figure 4: This figure details the process of RDFFM. It performs warping and fusion operations on the input deformation field, and after two iterations, outputs the deformation field via upsampling.
  • Figure 5: The visualization of the registration results from different methods on LPBA40 datasets. The two leftmost columns represent the original image and segmentation of the fixed and moving images, respectively. From top to bottom, the figure displays the warped images, warped segmentations, registration fields in grid format, and the deformed grids for different methods.
  • ...and 2 more figures