Dynamic abdominal MRI image generation using cGANs: A generalized model for various breathing patterns with extensive...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Organ motion during abdominal interventions limits treatment accuracy.
• Intraoperative imaging for guidance increases operative time and radiation exposure.
• Developing methods to reduce intraoperative imaging is crucial for patient safety and treatment efficiency.

Purpose of the Study:

• To implement conditional generative adversarial networks (cGANs) for generating dynamic magnetic resonance imaging (dMRI) using external abdominal motion.
• To investigate different cGAN models for improved motion quality in generated dMRI.
• To objectively and subjectively assess the quality of generated dMRI sequences.

Main Methods:

• Conditional generative adversarial networks (cGANs) were trained using external abdominal motion as a surrogate signal for breathing variability.
• Evaluated multiple generator models to enhance motion representation in synthetic dMRI.
• Conducted objective assessments using Structural Similarity Index Measure (SSIM) and Mean Absolute Error (MAE).
• Performed subjective assessments with 32 clinical experts evaluating realism of static images and dynamic sequences.

Main Results:

• The best-performing cGAN model achieved an SSIM of 0.73 ± 0.13.
• Mean Absolute Error (MAE) for motion prediction was below 4.5 mm (superior-inferior) and 1.8 mm (anterior-posterior).
• Over 50% of generated synthetic images and dynamic sequences were classified as real by clinical experts.

Conclusions:

• Synthetic dynamic MRI generated using cGANs shows potential to reduce intraoperative imaging needs.
• This approach can decrease operative time and radiation exposure during abdominal tumor interventions.
• The developed framework offers a promising alternative for real-time motion management in radiotherapy.