Photoacoustic-MR Image Registration Based on a Co-Sparse Analysis Model to Compensate for Brain Shift

Parastoo Farnia, Bahador Makkiabadi, Maysam Alimohamadi, Ebrahim Najafzadeh, Maryam Basij, Yan Yan, Mohammad Mehrmohammadi, Alireza AhmadianbioRxiv 2021 · DOI

TL;DR

This paper introduces a novel photoacoustic-MR (PA-MR) image registration method using a co-sparse analysis model to compensate for brain shift during neurosurgery. By effectively capturing the interdependencies between PA and MR images, this approach offers a promising solution for real-time image guidance when pre-operative MR data becomes unreliable due to tissue deformation.

What Problem Does This Paper Address?

Neurosurgical interventions, especially those targeting tumors in critical brain regions, rely heavily on image-guided surgery systems. However, a significant challenge is "brain shift"—the deformation of brain tissue that occurs after the skull is opened. This shift renders pre-operative MRI scans inaccurate, compromising the precision of surgical navigation. While intra-operative imaging modalities like ultrasound (US) and MRI exist, they have limitations such as radiation exposure, cost, and limited imaging depth. The paper aims to address this by proposing the use of photoacoustic (PA) imaging, a non-ionizing hybrid modality, in conjunction with MR imaging for accurate and real-time brain shift compensation. The core challenge lies in developing a robust image registration method that can handle the complex and unpredictable nature of brain deformation between these different imaging modalities.

Methods

The paper proposes a co-sparse analysis model for PA-MR image registration. This model leverages the idea that different imaging modalities of the same scene can be represented by a common sparse component. Specifically, they introduce a Joint Co-Sparse Analysis Model (JACSM). This approach learns a pair of analysis operators, one for PA images and one for MR images, that capture the interdependencies between the two modalities. The learning of these operators and the optimization of the transformation between the images are achieved using the Alternating Direction Method of Multipliers (ADMM). This method breaks down the complex optimization problem into smaller, manageable sub-problems, making it computationally efficient. The co-sparse analysis model is favored over synthesis sparse models due to its ability to yield richer feature representations and potentially faster real-time performance, which is crucial for intra-operative applications.

Datasets

The proposed method was evaluated using two types of datasets:

  • Brain-mimicking phantom data: This phantom was constructed from Polyvinyl Alcohol Cryogel (PVA-C) and featured simulated blood vessels filled with either copper sulfate or human blood. This allowed for controlled experiments with known deformations.
  • Ex-vivo murine brain data: This dataset consisted of MR and PA images of mouse brains, provided by a previous study. Anatomical landmarks were used for evaluation.

The paper does not introduce a new benchmark dataset but rather uses existing data types to validate their novel registration approach.

Key Insights and Ideas

  • Photoacoustic Imaging as a viable intra-operative tool: The study highlights PA imaging's potential as a non-ionizing, high-contrast, and spatially/temporally resolved modality for intra-operative guidance, particularly for visualizing vascular structures which serve as important landmarks.
  • Co-sparse analysis for multimodal registration: The core innovation is the application of co-sparse analysis to capture the inherent relationships between PA and MR images, enabling more accurate registration than traditional methods. This approach moves beyond simple intensity-based similarity measures that often fail with disparate modalities.
  • ADMM for efficient operator learning: The use of ADMM is key to efficiently learning the analysis operators and optimizing the transformation, making the method suitable for real-time applications where computational speed is critical.
  • Brain deformation is a complex, non-linear challenge: The paper emphasizes that accurately modeling and compensating for brain shift is inherently difficult due to its unpredictable nature, necessitating advanced registration techniques.

Key Results

The proposed JACSM-based PA-MR registration method demonstrated significant improvements:

  • Phantom Data: Achieved approximately 63% improvement in target registration error (TRE) compared to the commonly used normalized mutual information (NMI) method. The RMSE was also significantly lower than US-MR registration.
  • Mouse Brain Data: The method showed acceptable results, with calculated mean RMSE, TRE, and Hausdorff Distance (HD) of 1.13, 0.98, and 0.85 mm, respectively. These results were comparable to MR-MR registration (considered a gold standard) and showed a notable improvement over US-MR registration.
  • Performance Comparison: The study found that PA-MR registration performed better than US-MR registration, and MR-MR registration was the most accurate, as expected. However, PA-MR offered a good balance of accuracy and the advantages of PA imaging.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the proposed method shows improvement, MR-MR registration remains the most accurate.
  • The study used ex-vivo data and simulated deformations. Real-time intra-operative testing on human patients with actual brain shift would be the ultimate validation.
  • The computational cost, while improved by co-sparse analysis and ADMM, might still be a factor for extremely time-sensitive procedures.
  • Future Work:
    • The authors suggest that intra-operative PA imaging, coupled with their robust registration method, could become a valuable tool for neurosurgeons when pre-operative MRI is invalidated by brain shift.
  • Further research could focus on optimizing the real-time performance and exploring its application in more complex surgical scenarios.
  • Investigating the use of other advanced sparse representation techniques or deep learning approaches in conjunction with co-sparse models could also be beneficial.

Summary updated

Abstract

Abstract Brain shift is an important obstacle to the application of image guidance during neurosurgical interventions. There has been a growing interest in intra-operative imaging to update the image-guided surgery systems. However, due to the innate limitations of the current imaging modalities, accurate brain shift compensation continues to be a challenging task. In this study, the application of intra-operative photoacoustic imaging and registration of the intra-operative photoacoustic with pre-operative MR images is proposed to compensate for brain deformation. Finding a satisfactory registration method is challenging due to the unpredictable nature of brain deformation. In this study, the co-sparse analysis model is proposed for photoacoustic -MR image registration, which can capture the interdependency of the two modalities. The proposed algorithm works based on the minimization of the mapping transform via a pair of analysis operators that are learned by the alternating direction method of multipliers. The method was evaluated using experimental phantom and ex-vivo data obtained from the mouse brain. The results of phantom data show about 63% improvement in target registration error in comparison with the commonly used normalized mutual information method. Results proved that intra-operative photoacoustic images could become a promising tool when the brain shift invalidated pre-operative MRI.
BibTeX
@misc{farnia2021photoacoustic,
  author = {Parastoo Farnia and Bahador Makkiabadi and Maysam Alimohamadi and Ebrahim Najafzadeh and Maryam Basij and Yan Yan and Mohammad Mehrmohammadi and Alireza Ahmadian},
  title = {{Photoacoustic-MR Image Registration Based on a Co-Sparse Analysis Model to Compensate for Brain Shift}},
  year = {2021},
  doi = {10.1101/2021.07.28.453095},
  howpublished = {bioRxiv},
}
Figure