Deep Learning-Based Motion-Compensated Reconstruction for Accelerating 4-Dimensional Magnetic Resonance...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiotherapy Planning

Background:

• Conventional 4D-MRF reconstruction is time-consuming, limiting its clinical utility.
• Motion compensation is crucial for accurate 4D-MRF in abdominal imaging.
• Accelerated reconstruction methods are needed for efficient treatment planning.

Purpose of the Study:

• To develop and validate DeepMocor, a deep learning-based method for motion-compensated 4D-MRF.
• To accelerate conventional 4D-MRF reconstruction for improved clinical workflow.
• To enable more efficient clinical treatment planning, particularly for liver cancer.

Main Methods:

• Prospective study involving 19 hepatocellular carcinoma patients using a 3T MRI scanner.
• DeepMocor employs motion field initialization, refinement, and 4D-MRF reconstruction.
• Performance evaluated against alternative methods using metrics like PSNR, SSIM, MAPE, CNR, AMD, and PCC.

Main Results:

• DeepMocor achieved high image quality (PSNR: ~25.5, SSIM: ~0.86) and tissue property accuracy (MAPE: 3.1%-15.8%).
• Accurate tumor motion tracking was demonstrated with low average motion difference (AMD: 0.32-0.62 mm) and high correlation (PCC: 0.94-0.96).
• DeepMocor significantly outperformed alternative methods across most evaluated metrics.

Conclusions:

• DeepMocor enables a 24-fold acceleration in 4D-MRF reconstruction compared to conventional methods.
• The method demonstrates potential for significantly improving the efficiency of liver radiation therapy planning.
• DeepMocor represents a promising advancement in medical imaging for cancer treatment.