Deep Learning MRI Reconstruction Suppressing Fine-Line Artifacts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep-learning MRI reconstruction networks applied to fast-spin-echo data often produce images with fine-line artifacts, which can be misinterpreted as pathology or degrade image quality, and existing methods to address this either increase imaging time or compromise image resolution.
Innovation Solution
A deep-learning network is trained using fully sampled NEX=2 data to generate ground-truth images without fine-line artifacts, with one excitation being retrospectively undersampled for input during training, allowing the network to learn and suppress both undersampling and fine-line artifacts, enabling reconstruction of NEX=1 images with suppressed artifacts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep-learning networks are applied to fast-spin-echo data reconstruction, then image reconstruction quality is improved, but fine-line artifacts are introduced that degrade diagnostic accuracy
Solution Approach 1:
The patent converts the harmful fine-line artifacts into a training target by using NEX=2 data (which has suppressed artifacts) as ground truth to train the deep-learning network. The network learns to recognize and eliminate the artifacts it would normally generate when reconstructing NEX=1 data, transforming the artifact problem into a learnable pattern recognition task.
Solution Approach 2:
The patent performs preliminary artifact suppression by acquiring NEX=2 data during the training phase. This preliminary action creates artifact-free ground truth images that guide the network learning process, enabling the network to learn artifact suppression before actual reconstruction is performed.
2Measurement precision
If NEX=2 acquisition is used to suppress fine-line artifacts, then image quality is improved, but scanning time is doubled
Solution Approach 1:
The patent creates a copy of the beneficial artifact-suppression effect achieved by NEX=2 acquisition through training data. Instead of requiring actual NEX=2 acquisition for every scan, the network learns from copied ground truth images and applies the learned artifact suppression to rapid NEX=1 acquisitions, preserving image quality while reducing scan time.
Solution Approach 2:
The patent changes the parameter from NEX=2 (two excitations) to NEX=1 (one excitation) for the actual acquisition, while using the NEX=2 data only for training purposes. This parameter change reduces scanning time by half while the trained network maintains the image quality benefits of NEX=2 through learned artifact suppression.
3Productivity
If undersampled NEX=1 data is used for fast imaging, then scanning speed is increased, but fine-line artifacts appear and image quality degrades
Solution Approach 1:
The patent implements feedback by comparing the network's reconstructed NEX=1 images against ground truth images derived from NEX=2 data during training. This feedback mechanism allows the network to learn the correct reconstruction patterns that eliminate artifacts while maintaining the speed benefits of undersampled NEX=1 acquisition.
Data Source
AI summary
The subject matter discussed herein relates to a fast magnetic resonance imaging (MRI) method to suppress fine-line artifact in Fast-Spin-Echo (FSE) images reconstructed with a deep-learning network. The network is trained using fully sampled NEX=2 (Number of Excitations equals to 2) data. In each case, the two excitations are combined to generate fully sampled ground-truth images with no fine-line artifact, which are used for comparison with the network generated image in the loss function. However, only one of the excitations is retrospectively undersampled and inputted into the network during training. In this way, the network learns to remove both undersampling and fine-line artifacts. At inferencing, only NEX=1 undersampled data are acquired and reconstructed.