Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

NCF: Neural Correspondence Field for Medical Image Registration

Lei Zhou, Nimu Yuan, Katjana Ehrlich, Jinyi Qi

TL;DR

Deformable image registration in medical imaging is challenged by limited training data. The authors propose Neural Correspondence Field (NCF), a training-data-free method that learns a per-image-pair correspondence field using a lightweight network with a coarse correspondence module and a smoothing module. The approach is trained in a self-supervised manner by warping the moving image to the fixed image through grid sampling and optimizing a composite loss including photometric, SSIM, and occupancy terms. It achieves superior or competitive performance on Lung CT and head-and-neck datasets while using only about 0.06 million parameters, demonstrating strong generalization and efficiency suitable for data-scarce clinical scenarios.

Abstract

Deformable image registration is a fundamental task in medical image processing. Traditional optimization-based methods often struggle with accuracy in dealing with complex deformation. Recently, learning-based methods have achieved good performance on public datasets, but the scarcity of medical image data makes it challenging to build a generalizable model to handle diverse real-world scenarios. To address this, we propose a training-data-free learning-based method, Neural Correspondence Field (NCF), which can learn from just one data pair. Our approach employs a compact neural network to model the correspondence field and optimize model parameters for each individual image pair. Consequently, each pair has a unique set of network weights. Notably, our model is highly efficient, utilizing only 0.06 million parameters. Evaluation results showed that the proposed method achieved superior performance on a public Lung CT dataset and outperformed a traditional method on a head and neck dataset, demonstrating both its effectiveness and efficiency.

NCF: Neural Correspondence Field for Medical Image Registration

TL;DR

Deformable image registration in medical imaging is challenged by limited training data. The authors propose Neural Correspondence Field (NCF), a training-data-free method that learns a per-image-pair correspondence field using a lightweight network with a coarse correspondence module and a smoothing module. The approach is trained in a self-supervised manner by warping the moving image to the fixed image through grid sampling and optimizing a composite loss including photometric, SSIM, and occupancy terms. It achieves superior or competitive performance on Lung CT and head-and-neck datasets while using only about 0.06 million parameters, demonstrating strong generalization and efficiency suitable for data-scarce clinical scenarios.

Abstract

Deformable image registration is a fundamental task in medical image processing. Traditional optimization-based methods often struggle with accuracy in dealing with complex deformation. Recently, learning-based methods have achieved good performance on public datasets, but the scarcity of medical image data makes it challenging to build a generalizable model to handle diverse real-world scenarios. To address this, we propose a training-data-free learning-based method, Neural Correspondence Field (NCF), which can learn from just one data pair. Our approach employs a compact neural network to model the correspondence field and optimize model parameters for each individual image pair. Consequently, each pair has a unique set of network weights. Notably, our model is highly efficient, utilizing only 0.06 million parameters. Evaluation results showed that the proposed method achieved superior performance on a public Lung CT dataset and outperformed a traditional method on a head and neck dataset, demonstrating both its effectiveness and efficiency.

Paper Structure

This paper contains 14 sections, 1 equation, 4 figures, 1 table.

Table of Contents

  1. INTRODUCTION
  2. Related Work
  3. Method
  4. Network Design
  5. Loss Design
  6. Experiments
  7. Implementation Details
  8. Dataset
  9. Unpaired/paired Lung CT dataset
  10. Private Head and Neck data
  11. Results
  12. Result on the Lung CT dataset
  13. Result on Head and Neck dataset
  14. Conclusion

Figures (4)

  • Figure 1: Features and advantages of the proposed method. (a) Comparison with learning-based methods in terms of the need for a training dataset. (b) Comparison of the registration performance (measured by Dice Similarity Coefficient (DCS) of segmentation masks) vs. the number of network parameters.
  • Figure 2: Architecture. The proposed method, NCF, consists of two parts: the CCM, represented in orange, is a 5-layer MLP network, and the SM, represented in green, is a 2-layer 3D CNN network. The loss function $\mathcal{L}$ evaluates the difference between the warped image $I_m \circ \Phi$ and the fixed image $I_f$.
  • Figure 3: A 2D illustration of generating a warped image through grid sampling of the moving image based on the warped grid.
  • Figure 4: Comparing the results of our NCF method and the 3D Slicer on the head and neck dataset. Red rectangles mark the areas where our method outperformed the 3D Slicer.