Nonrigid registration of medical image based on adaptive local structure tensor and normalized mutual information

Tiejun Yang, Qi Tang, Lei Li, Jikun Song, Chunhua Zhu, Lu TangJournal of Applied Clinical Medical Physics 2019 · DOI

TL;DR

This paper introduces a novel similarity measure for nonrigid medical image registration called Adaptive Local Structure Tensor-Normalized Mutual Information (ALST-NMI). By combining intensity-based normalized mutual information with spatial information extracted from an adaptive local structure tensor, the method achieves improved accuracy and robustness, particularly for large deformations, and reduces the risk of falling into local optima.

What Problem Does This Paper Address?

Medical image registration is crucial for tasks like comparing scans over time or fusing information from different imaging modalities. Nonrigid registration, which allows for complex, non-uniform deformations, is particularly important because anatomical structures can change significantly due to disease, patient movement, or the inherent flexibility of tissues. Traditional similarity measures like Mutual Information (MI) primarily focus on intensity statistics, ignoring valuable spatial and structural information. This can lead to inaccurate alignments, especially in areas with strong gradients or large deformations, and can cause registration algorithms to get stuck in suboptimal solutions (local extremes). The paper aims to address this by developing a registration method that better leverages both intensity and spatial information.

Methods

The proposed method, ALST-NMI, builds upon existing techniques by introducing several key innovations:

  • Adaptive Local Structure Tensor: Instead of using a fixed Gaussian filter for extracting local structure, the paper employs an adaptive Gaussian filter. The variance of the Gaussian kernel is adjusted based on the regional variance of the image. This allows the method to better preserve image details in consistent regions while effectively capturing structural information in edge and corner regions. The local structure tensor itself is used to extract spatial information.
  • Weighting Function: The spatial information extracted from the adaptive local structure tensor is used to define a weighting function. This function assigns higher weights to regions with strong spatial information (e.g., edges, strong gradients) and lower weights to smoother regions.
  • Combined Similarity Measure: The final similarity measure is a product of the discrete Normalized Mutual Information (NMI) and the derived weighting function. This ensures that regions with strong spatial features contribute more significantly to the overall similarity score, guiding the registration process more effectively.
  • Registration Framework: The paper utilizes a standard nonrigid registration framework involving a B-spline transformation model for representing deformations, the proposed ALST-NMI as the similarity measure, and a steepest gradient descent method for optimization.

Datasets

The paper uses real brain image datasets for validation. Specifically, it mentions:

  • Single-modality brain images: A dataset from "the brain web" with a size of 353x354 pixels.
  • Multi-modality brain images: MRI and CT brain images, both with a size of 256x256 pixels.

The paper does not introduce new benchmark datasets.

Key Insights and Ideas

  • Spatial Information is Key for Robustness: The core insight is that simply relying on intensity statistics (like NMI) is insufficient for accurate nonrigid registration, especially with large deformations. Incorporating spatial and structural information significantly improves performance.
  • Adaptivity Matters: Using an adaptive approach for feature extraction (Gaussian filtering in this case) is crucial. Fixed parameters can either smooth out important details or fail to capture sufficient structural information, whereas adaptive parameters allow for a more nuanced and effective representation of local image characteristics.
  • Prioritizing Strong Features: The weighting function effectively prioritizes regions with strong spatial information. This is intuitive, as these are often the most informative areas for aligning images and are less prone to ambiguity than smooth regions.
  • Reducing Local Optima: By considering spatial context and weighting important regions, the method is less likely to get trapped in suboptimal registration solutions, leading to more reliable results.

Key Results

The proposed ALST-NMI method demonstrated superior performance compared to several baseline methods (GD, SSD, NMI, BRMI, LST-NMI) in both single-modality and multi-modality brain image registration experiments.

  • Single-modality:
    • Achieved the lowest Mean Square Error (MSE) of 64.350.
  • Showed the highest Structural Similarity Index (SSIM) of 0.962.
  • Resulted in the highest Normalized Mutual Information (NMI) of 1.458.
  • Achieved the highest Correlation Coefficient (CC) of 0.990.
  • Significantly reduced displacement errors, with 94.7% of pixels effectively registered and errors larger than 0.2 cm being less than 0.4% of total pixels.
  • Multi-modality:
    • Achieved a lower MSE (2695.14) and higher SSIM (0.621), NMI (0.748), and CC (0.713) compared to other methods.
  • Visually, ALST-NMI produced smoother registration results closer to the reference image, with fewer artifacts than some other methods.
  • Demonstrated improved robustness against falling into local extremes, particularly in challenging multi-modality scenarios where intensity differences are large.

Limitations and Future Work

  • Computational Efficiency: The paper acknowledges that the NMI and local structure tensor computations require traversing each pixel, which can reduce registration efficiency.
  • Future Work: The authors suggest exploring methods to achieve higher registration precision and computational efficiency simultaneously. This implies a need for further optimization or algorithmic improvements to speed up the process without sacrificing accuracy.

Summary updated

Abstract

Nonrigid registration of medical images is especially critical in clinical treatment. Mutual information is a popular similarity measure for medical image registration; however, only the intensity statistical characteristics of the global consistency of image are considered in MI, and the spatial information is ignored. In this paper, a novel intensity-based similarity measure combining normalized mutual information with spatial information for nonrigid medical image registration is proposed. The different parameters of Gaussian filtering are defined according to the regional variance, the adaptive Gaussian filtering is introduced into the local structure tensor. Then, the obtained adaptive local structure tensor is used to extract the spatial information and define the weighting function. Finally, normalized mutual information is distributed to each pixel, and the discrete normalized mutual information is multiplied with a weighting term to obtain a new measure. The novel measure fully considers the spatial information of the image neighborhood, gives the location of the strong spatial information a larger weight, and the registration of the strong gradient regions has a priority over the small gradient regions. The simulated brain image with single-modality and multimodality are used for registration validation experiments. The results show that the new similarity measure improves the registration accuracy and robustness compared with the classical registration algorithm, reduces the risk of falling into local extremes during the registration process.
BibTeX
@article{yang2019nonrigid,
  author = {Tiejun Yang and Qi Tang and Lei Li and Jikun Song and Chunhua Zhu and Lu Tang},
  title = {{Nonrigid registration of medical image based on adaptive local structure tensor and normalized mutual information}},
  journal = {Journal of Applied Clinical Medical Physics},
  year = {2019},
  doi = {10.1002/acm2.12612},
}
Figure