Rapid 4D-MRI reconstruction using a deep radial convolutional neural network: Dracula.

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• 4D and midposition MRI are crucial for adaptive radiotherapy in lung and abdominal treatments.
• Current 4D-MRI reconstruction times are lengthy, hindering timely plan adaptation.

Purpose of the Study:

• To develop deep learning solutions for accelerated 4D-MRI reconstruction.
• To maintain high image quality and short scan times for MR-guided radiotherapy.

Main Methods:

• Two 3D U-net deep convolutional neural networks were trained for accelerated 4D joint MoCo-HDTV reconstruction.
• One network reconstructed 4D-MRI from gridded images; the second directly calculated midposition images.
• Deep learning MRI were validated against joint MoCo-HDTV reconstructions using SSIM and NIQE, with blinded observer contouring.

Main Results:

• High-quality 4D and midposition MRI were reconstructed in 28 seconds per subject.
• Excellent agreement (SSIM ≥ 0.96) was observed between deep learning and conventional methods.
• Deep learning 4D-MRI was clinically acceptable for delineation, with tumor positions agreeing within 0.7 mm.

Conclusions:

• Deep convolutional neural networks can effectively approximate joint MoCo-HDTV and midposition MRI algorithms.
• Rapid reconstruction of 4D and midposition MRI supports online treatment adaptation in thoracic and abdominal MR-guided radiotherapy.