Ultra-fast multi-parametric 4D-MRI image reconstruction for real-time applications using a downsampling-invariant...

Area of Science:

• Medical imaging
• Artificial intelligence in medicine
• Radiotherapy

Background:

• Accurate motion estimation from 4D-MRI is crucial for real-time imaging and tumor tracking during radiotherapy.
• Severely downsampled 4D-MRI data presents challenges for traditional motion estimation techniques.

Purpose of the Study:

• To develop a novel deep learning model for simultaneous MR image reconstruction and motion estimation from downsampled 4D-MRI.
• To evaluate the performance of the proposed model against existing methods.

Main Methods:

• A deep learning model, Downsampling-Invariant Deformable Registration (D2R), was developed.
• The D2R model was trained and validated using 4D-MRI data from 43 liver tumor patients undergoing radiotherapy.
• Performance was compared against Demons, Elastix, pTV, and VoxelMorph algorithms.
• High-quality 4D-MR images were reconstructed for real-time imaging feasibility.

Main Results:

• The D2R model demonstrated superior and robust registration performance compared to all baseline methods, even at high downsampling factors (up to 500).
• Successfully reconstructed high-quality T1-weighted and T2-weighted 4D-MR images with improved image quality and sub-voxel motion accuracy.
• External validation confirmed the model's robustness and generalizability.

Conclusions:

• The developed D2R model effectively estimates deformation from downsampled 4D-MR images and reconstructs high-quality 4D-MR images.
• This model holds potential for enhancing clinical implementation of 4D-MRI for real-time motion management in liver cancer treatment.