Real-time 3D motion estimation from undersampled MRI using multi-resolution neural networks.

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Real-time adaptive radiotherapy requires precise tracking of internal organ motion.
• Magnetic resonance imaging-guided radiotherapy (MRIgRT) offers superior soft-tissue contrast for motion visualization.
• Current methods for obtaining 3D deformation vector fields (DVFs) lack the necessary spatiotemporal resolution and speed for real-time adaptation.

Purpose of the Study:

• To develop and validate a deep learning model for generating time-resolved 3D DVFs from undersampled 4D-MRI data.
• To achieve high spatiotemporal resolution and low latency (< 200 ms) for DVFs essential for real-time MRIgRT.
• To enable adaptive radiotherapy by providing accurate, real-time motion information.

Main Methods:

• A convolutional neural network (CNN) named TEMPEST was trained on retrospectively undersampled 4D-MRI data from lung cancer patients.
• The model learned to predict optical flow DVFs using a multiresolution architecture and nonuniform fast Fourier transform reconstruction.
• Validation was performed using 4D-MRI, digital phantoms, physical motion phantoms, and in-vivo volunteer scans on a hybrid MR-scanner.

Main Results:

• TEMPEST achieved accurate DVFs on respiratory-resolved MRI with 20-fold acceleration (average endpoint error $$ mm).
• Accurate time-resolved DVFs were estimated on a motion phantom with $$ mm error at 28x undersampling.
• In-vivo scans demonstrated accurate motion estimation within 200 ms, including MRI acquisition, at 366x undersampling.

Conclusions:

• A CNN trained on undersampled MRI can generate accurate 3D DVFs with high spatiotemporal resolution.
• The TEMPEST model shows significant promise for enabling real-time adaptive MRIgRT.
• This approach facilitates precise, motion-adaptive radiotherapy delivery.