MRI Reconstruction Using Auxiliary Data and Learned Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for restoring missing data in medical imaging, such as MRI, using deep neural networks are limited in accuracy and efficiency, particularly when dealing with undersampled k-space data, leading to suboptimal reconstruction images.
Innovation Solution
A medical data processing apparatus and method that employs a learned model with a multilayer network architecture, utilizing both target and auxiliary input data with different imaging conditions to restore missing portions in medical data, enhancing the accuracy of image reconstruction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep neural network is used to restore missing data in undersampled k-space data, then image reconstruction can be achieved, but restoration accuracy is limited
Solution Approach 1:
The patent introduces a multi-dimensional approach by incorporating auxiliary input data with different imaging conditions alongside target input data. This expands the input space from a single data dimension to multiple dimensions (target data + auxiliary data from different conditions), enabling the neural network to restore missing information more accurately by leveraging correlations across different imaging dimensions.
Solution Approach 2:
The patent uses auxiliary input data as an intermediary to bridge the information gap caused by undersampling. The auxiliary data, acquired under different imaging conditions, serves as a mediator that provides additional information to compensate for the missing portions in the target k-space data, enabling more accurate restoration without requiring complete sampling of the target data.
2Measurement precision
If complete sampling of k-space data is performed, then high quality reconstruction images are obtained, but examination time increases
Solution Approach 1:
The patent applies partial action by acquiring auxiliary input data under different imaging conditions rather than completely sampling the target k-space data. This partial sampling approach, combined with the neural network restoration using multiple inputs, achieves high-quality reconstruction without the time cost of complete sampling, effectively doing 'enough' sampling to enable accurate AI-based restoration.
Solution Approach 2:
The patent performs preliminary action by acquiring auxiliary input data with different imaging conditions before the actual reconstruction process. This preliminary data collection enables the neural network to have access to additional information that facilitates accurate restoration of the target data, reducing the need for extensive target data acquisition and thereby shortening examination time.
3Productivity
If undersampling rate of k-space data is increased, then examination time is reduced, but restoration accuracy deteriorates
Solution Approach 1:
The patent changes the parameters of the neural network by introducing multiple input channels that accept both target input data and auxiliary input data with different imaging conditions. This parameter change in the network architecture enables it to process highly undersampled data more effectively by leveraging the additional auxiliary information, thereby maintaining restoration accuracy even at higher undersampling rates.
Solution Approach 2:
The patent creates a universal neural network model that can handle various imaging conditions through its multi-input architecture. The network is designed to process both target data and auxiliary data from different imaging conditions, making it universally applicable to restore highly undersampled data while maintaining accuracy, thus enabling higher undersampling rates without sacrificing restoration quality.
Data Source
AI summary
A medical data processing apparatus includes a memory and processing circuitry. The memory stores a learned model including an input layer to which first MR data and second MR data having the same imaging target as the first MR data and an imaging parameter different from the first MR data are inputted, an output layer from which third MR data is output with a missing portion of the first MR data restored, and at least one intermediate layer arranged between the input layer and the output layer. The processing circuitry generates third MR data relating to the subject, from the first MR data serving as a process target and relating to the subject and the second MR data relating to the subject and acquired by an imaging parameter different from the first MR data serving as the process target, in accordance with the learned model.