Technical note: Minimizing CIED artifacts on a 0.35 T MRI‐Linac using deep learning

Austen Curcuru, Deshan Yang, Hongyu An, Phillip S. Cuculich, Clifford G. Robinson, H. Michael GachJournal of Applied Clinical Medical Physics 2024 · DOI

TL;DR

This paper tackles the challenge of severe image artifacts caused by implantable cardioverter-defibrillators (ICDs) in MRI-guided radiotherapy (MRgRT). They successfully employed a CycleGAN, a type of deep learning model, to significantly reduce these artifacts in MRI scans, leading to improved image quality and more accurate target tracking for radiation delivery.

What Problem Does This Paper Address?

Treatments like stereotactic arrhythmia radiotherapy (STAR) aim to precisely target cardiac lesions causing arrhythmias. However, these treatments often require expanded margins to account for patient motion and setup uncertainties, leading to unnecessary radiation exposure to healthy tissues. MRI-guided radiotherapy (MRgRT) offers superior soft tissue contrast and real-time target tracking, which could potentially reduce these margins. A major hurdle for MRgRT, especially in patients with ventricular tachycardia (VT), is the presence of implantable cardioverter-defibrillators (ICDs). ICDs generate significant artifacts in MRI scans, particularly in the balanced steady-state free precession (bSSFP) cine sequences commonly used for MRgRT. These artifacts obscure critical anatomical information, hindering accurate target tracking and potentially compromising treatment efficacy and safety. This paper aims to mitigate these ICD-induced artifacts to enable more effective MRgRT.

Methods

The core of this study is the application of a Cycle-Consistent Generative Adversarial Network (CycleGAN) for image-to-image translation. CycleGAN is an unsupervised deep learning architecture that learns to translate images from a source domain (e.g., images with ICD artifacts) to a target domain (e.g., artifact-free images) without requiring perfectly paired training data.

The CycleGAN architecture consists of two generators (one mapping from domain X to Y, and another from Y to X) and two discriminators (one for each domain). The training process involves three types of losses:

  • Adversarial Loss: Encourages the generated images to be indistinguishable from the target domain.
  • Cycle-Consistency Loss: Ensures that translating an image from domain X to Y and then back to X results in the original image (and vice-versa). This helps maintain image content and structure.
  • Identity Loss: Aims to preserve the color/content of the input image when it's already in the target domain, making the model more conservative.

The generators used a 9-block ResNet architecture, and the discriminators employed a 70x70 PatchGAN architecture. Training was performed using PyTorch and Python.

Datasets

The study utilized data from two groups of volunteers:

  • Group A: Fourteen healthy volunteers were scanned on a 0.35 T MRI-Linac. For these volunteers, MR conditional ICDs were taped to their pectoral region with leads running towards the ventricles to simulate implanted ICDs. Scans were repeated immediately after removing the ICD to create paired (with and without ICD) data.
  • Training Data: Data from 12 volunteers were used for training the CycleGAN.
  • Testing Data: Data from the remaining two volunteers were used for testing. Additionally, a Leave-One-Out scheme was employed, where one volunteer's data was used for testing, and the remaining twelve were used for training. This was repeated for three specific volunteers.
  • Group B: Three additional cine MRI datasets from 2018 were used for qualitative testing. These datasets were acquired from patients with implanted ICDs but did not have corresponding artifact-free reference scans, making them unsuitable for quantitative evaluation but useful for assessing the model's robustness.

Key Insights and Ideas

  • Unsupervised Learning is Key: The difficulty in acquiring perfectly paired MRI data (with and without ICD) for every patient makes unsupervised learning approaches like CycleGAN particularly valuable.
  • CycleGAN's Versatility: The study demonstrates CycleGAN's effectiveness not just for general image-to-image translation but specifically for a challenging medical imaging problem like artifact reduction in MRgRT.
  • Preserving Anatomy: A crucial aspect is that the CycleGAN not only removes artifacts but also preserves the underlying anatomical structures, which is vital for accurate treatment planning and delivery.
  • Impact on Tracking: The reduction in artifacts directly translates to improved tracking accuracy, as evidenced by the improved Dice Similarity Coefficient (DSC) and reduced Target Registration Error (TRE) and Hausdorff Distance (HD).

Key Results

The CycleGAN reconstruction demonstrated significant improvements in both tracking and image quality metrics:

Tracking Metrics (Whole-Heart Contour):

  • Dice Similarity Coefficient (DSC): Increased from a mean of 0.910 to 0.935.
  • Target Registration Error (TRE): Decreased from a mean of 4.488 mm to 2.877 mm.
  • 95th Percentile Hausdorff Distance (95% HD): Decreased from a mean of 10.236 mm to 7.700 mm.

Image Quality Metrics (Whole-Body Slice):

  • Normalized Root Mean Square Error (nRMSE): Decreased from a mean of 0.644 to 0.420.
  • Multiscale Structural Similarity (MS-SSIM): Increased from a mean of 0.779 to 0.819.
  • Peak Signal-to-Noise Ratio (PSNR): Increased from a mean of 18.744 to 22.368.

Qualitative evaluation on Group B datasets also showed a noticeable reduction in banding artifacts near the heart. All paired t-tests for these metrics showed p < 0.001, indicating statistically significant improvements.

Limitations and Future Work

  • Small Sample Size: The authors acknowledge that the study's primary limitation is the relatively small sample size of the training and testing datasets. They suggest that additional training data could further enhance the CycleGAN's artifact reduction capabilities.
  • Data Correlation: Due to the nature of the dataset acquisition (e.g., using data from a limited number of volunteers for training), there might be a high degree of correlation between training images, which could impact generalizability.
  • Single Model/Placement: The study used a single model and placement of the ICD for training. Future work could explore variations in ICD placement and models.
  • Imperfect Test Pairs: The test image pairs were not perfectly aligned, which is a common challenge in real-world scenarios. While CycleGAN handled this well, generating synthetic, anatomically correct MRI phantoms with realistic artifacts could provide a more robust data source for future studies.
  • Arm Position: The training was conducted with arms down, but cardiac patients often have their arms raised, which can shift the ICD. Training the model with arm positions relevant to treatment plans is suggested.
  • Future Directions:
    • Applying the method to 1.5 T MRI systems, where susceptibility artifacts are more severe.
  • Investigating alternative supervised learning networks.
  • Exploring application to other cine sequences used in MRgRT.
  • Optimizing for real-time reconstruction (< 30 ms) through hardware and software advancements.
  • Integrating CycleGAN into existing reconstruction pipelines, potentially through open-source frameworks like Gadgetron.

Summary updated

Abstract

CycleGAN-generated reconstructions significantly improved both tracking and image quality metrics when used to mitigate artifacts from ICDs.
BibTeX
@article{curcuru2024technical,
  author = {Austen Curcuru and Deshan Yang and Hongyu An and Phillip S. Cuculich and Clifford G. Robinson and H. Michael Gach},
  title = {{Technical note: Minimizing CIED artifacts on a 0.35 T MRI‐Linac using deep learning}},
  journal = {Journal of Applied Clinical Medical Physics},
  year = {2024},
  doi = {10.1002/acm2.14304},
}
Figure