DeepSENSE CNN for MRI Sensitivity Profile Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current MRI techniques face challenges in reducing scan time while maintaining high image quality and spatial resolution, as they require lengthy calibration processes and are hindered by the computational expense of accurately estimating sensitivity profiles, especially in multi-dimensional acquisitions with high coil counts.
Innovation Solution
A deep convolutional neural network (CNN) called DeepSENSE is used to estimate sensitivity profiles unsupervisedly, transforming multi-channel data into a latent space that enables more accurate and efficient image reconstruction, allowing for higher subsampling factors and reducing the need for separate calibration scans.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If separate calibration scan is performed to measure sensitivity maps, then sensitivity map accuracy is improved, but exam time increases
Solution Approach 1:
The patent combines the calibration scan with the actual imaging scan by acquiring calibration data from a calibration region within the imaging data itself. This merging eliminates the need for a separate calibration scan, thereby maintaining sensitivity map accuracy while reducing total exam time.
Solution Approach 2:
The imaging scan serves dual purposes: it acquires both the calibration data (from the calibration region) and the actual imaging data. This multi-functionality allows the same scan to fulfill both calibration and imaging requirements, reducing time loss.
2Measurement precision
If calibration region is included in highly subsampled acquisition, then sensitivity map estimation is enabled, but scan time cost increases proportionally
Solution Approach 1:
Instead of requiring a full separate calibration scan, the patent uses only a calibration region (partial action) extracted from the imaging data. This partial approach provides sufficient calibration information while minimizing the time cost proportionally.
3Measurement precision
If conventional sensitivity profile estimation methods are used, then accuracy is maintained, but computational time increases to several minutes
Solution Approach 1:
The patent replaces the conventional iterative mathematical estimation process (mechanical/computational system) with a deep learning neural network model. This substitution maintains sensitivity profile accuracy while reducing computational time from minutes to milliseconds.
Solution Approach 2:
The deep learning model is pre-trained on calibration data to learn the mapping from k-space data to sensitivity maps. During actual scanning, the pre-trained model can rapidly estimate sensitivity profiles without performing heavy computational iterations, thus reducing real-time computational time.
Data Source
AI summary
An MRI apparatus performs multi-channel calibration acquisitions using a multi-channel receiver array and uses a convolutional neural network (CNN) to compute an estimated profile map that characterizes properties of the multi-channel receiver array. The profile map is composed of orthogonal vectors and transforms single-channel image space data to multi-channel image space data. The MRI apparatus performs a prospectively subsampled imaging acquisition and processes the resulting k-space data using the estimated profile map to reconstruct a final image. The CNN may be pretrained in an unsupervised manner using subsampled simulated multi-channel calibration acquisitions and using a regularization function included in a training loss function.