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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Implantable cardioverter defibrillators (ICDs) create artifacts that challenge magnetic resonance imaging (MRI)-guided radiotherapy (MRgRT).
• Artifacts can compromise image quality and treatment accuracy.

Purpose of the Study:

• To evaluate an unsupervised generative adversarial network (CycleGAN) for mitigating ICD artifacts in balanced steady-state free precession (bSSFP) cine MRIs.
• To enhance image quality and tracking performance for MRgRT procedures.

Main Methods:

• A CycleGAN model was trained on bSSFP MRI data from 14 healthy volunteers with simulated ICDs.
• The model was tested using a Leave-One-Out cross-validation scheme.
• Image quality and tracking metrics (DSC, TRE, 95% HD, nRMSE, PSNR, MS-SSIM) were evaluated, alongside qualitative assessment on patient data.

Main Results:

• CycleGAN reconstruction significantly improved whole-heart contour tracking: Dice similarity coefficient (DSC) increased from 0.910 to 0.935, target registration error (TRE) decreased from 4.488 to 2.877 mm, and 95% Hausdorff distance (95% HD) decreased from 10.236 to 7.700 mm.
• Image quality metrics also improved: normalized root mean square error (nRMSE) decreased from 0.644 to 0.420, multiscale structural similarity (MS-SSIM) increased from 0.779 to 0.819, and peak signal-to-noise ratio (PSNR) increased from 18.744 to 22.368.
• Qualitative analysis of additional datasets confirmed a reduction in ICD artifacts.

Conclusions:

• CycleGAN effectively mitigates implantable cardioverter defibrillator (ICD) artifacts in MRI-guided radiotherapy (MRgRT).
• The AI-driven approach significantly enhances both image quality and tracking accuracy, crucial for precise radiotherapy delivery.