CNN-Based Multi-Parametric 4D MRI Reconstruction for Real-Time Tumor Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing 4D MRI systems face challenges in achieving real-time ultra-high-quality imaging due to long acquisition times, offline reconstruction, respiratory motion artifacts, and insufficient image contrast, leading to poor image quality and low temporal resolution, which hinders effective tumor tracking in liver radiation therapy.
Innovation Solution
A deep learning-based joint MR image reconstruction and motion estimation model using a convolutional neural network (CNN) for generating ultra-quality multi-parametric 4D MR images, employing non-uniform fast Fourier transform (NUFFT) and inverse NUFFT to simulate real-time acquisition, and dual supervision with end-to-end point error (EPE) and normalized correlation coefficient (NCC) for improved tumor contrast and organ edge sharpness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If traditional iterative optimization deformation registration algorithm is used to extract deformation vector field, then image quality can be improved, but processing time increases from tens of minutes to several hours
Solution Approach 1:
The patent replaces the traditional iterative optimization deformation registration algorithm (mechanical/computational system) with a deep learning-based convolutional neural network. The CNN model is trained offline using ground truth deformation vector fields from traditional registration, then deployed for real-time inference. This substitution transforms the processing from minutes/hours to seconds/real-time while maintaining image quality through the learned deformation patterns.
Solution Approach 2:
The patent performs preliminary training of the convolutional neural network offline using comprehensive training data and ground truth deformation fields. This preliminary action prepares the model in advance so that during actual 4D MRI generation, the pre-trained CNN can rapidly predict deformation vector fields without requiring iterative optimization, thus achieving real-time performance while maintaining accuracy.
2Manufacturing precision
If 4D MRI acquisition is performed with sufficient sampling conditions, then image quality can be improved, but acquisition time becomes too long for real-time application
Solution Approach 1:
The patent introduces a convolutional neural network as an intermediary between the acquired MRI data and the final 4D MR images. The CNN learns to predict deformation vector fields from intermediate data, acting as a mediator that bridges the gap between limited sampling data and high-quality reconstructed images, enabling real-time generation without requiring extensive sampling.
Solution Approach 2:
The patent changes the parameter of image reconstruction from traditional iterative methods to deep learning-based prediction. By training the CNN on diverse training data with various sampling conditions, the model learns to generalize and produce high-quality images even with reduced sampling, thus improving acquisition speed while maintaining image quality.
3Measurement precision
If conventional registration methods are used for offline reconstruction, then deformation accuracy can be improved, but real-time generation of 4D MR images cannot be achieved
Solution Approach 1:
The patent substitutes conventional registration methods with a deep learning-based convolutional neural network for deformation field prediction. The CNN is trained offline using ground truth deformation fields from conventional methods, then deployed for rapid real-time inference, achieving both accuracy and speed requirements for real-time 4D MRI generation.
Solution Approach 2:
The patent creates a learned copy of the deformation patterns through training the CNN on ground truth deformation vector fields. Instead of performing computationally intensive conventional registration during real-time operation, the system uses the pre-learned deformation patterns from the trained network, which replicates the accuracy of conventional methods at real-time speeds.
4Productivity
If deep learning model is used for real-time image generation, then processing speed is improved, but model training complexity and computational resources increase
Solution Approach 1:
The patent performs the complex model training as a preliminary action offline, before actual 4D MRI generation. The convolutional neural network is trained using comprehensive training data and ground truth deformation fields in advance. Once trained, the model can be deployed for real-time inference with minimal computational resources, separating the complexity of training from the simplicity of real-time operation.
Data Source
AI summary
A computer-implemented method for training a convolutional neural network (CNN) using training data comprising a pair of original and downsampled 4D magnetic resonance imaging (MRI) data is provided. The CNN is used to generate multi-parametric 4D magnetic resonance (MR) images based on multi-parametric 3D MR images in real-time. The method includes receiving a 4D MR image formed by a plurality of fixed images of different frames; converting the plurality of fixed images into a plurality of k-space data by non-uniform fast Fourier transform (NUFFT); applying radial scan to the k-space data to simulate real-time MR image acquisition, and generating a plurality of downsampled fixed images by inverse NUFFT; training a CNN with training data comprising the 4D MR image and the corresponding downsampled 4D MR image; and estimating the multi-parametric 4D MR image in real-time by applying apply the predicted DVF to the multi-parametric 3D MR images.