Accelerated MRI Reconstruction with Shifted Inputs for Aliasing Suppression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current accelerated MRI reconstruction methods using convolutional neural networks (CNNs) are inefficient in processing non-local aliasing artifacts due to their limited receptive field of view and locality assumptions, and require extensive data training which is not feasible with limited medical imaging datasets.
Innovation Solution
A de-aliasing CNN architecture is introduced that incorporates information about the periodicity of parallel imaging aliasing artifacts, allowing for non-local pixel dependencies by shifting input images based on the acceleration factor, and using a machine-learned model for regularization to suppress both local and non-local artifacts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If standard CNNs are used for accelerated MRI reconstruction, then local aliasing artifacts are suppressed, but non-local aliasing artifacts cannot be effectively reduced due to limited receptive field
Solution Approach 1:
The patent extends the CNN receptive field by incorporating temporal dimension through recurrent connections. The hidden state h_t carries information across time steps, allowing the network to access non-local spatial information without increasing spatial receptive field dimensions. This resolves the contradiction by adding a temporal dimension to capture non-local correlations.
Solution Approach 2:
The patent nests multiple processing levels within the UNet architecture, including spatial convolutional layers, temporal recurrent layers, and skip connections that aggregate information from different scales. This nested structure allows the model to capture both local and non-local dependencies within a unified framework, overcoming the limited receptive field of standard CNNs.
2Area of stationary object
If deeper CNN architectures with increased capacity are used to improve receptive FOV, then non-local pixel correlations are better captured, but computational cost increases significantly
Solution Approach 1:
The recurrent neural network maintains continuous temporal processing where hidden states are updated at each time step and fed back into the network. This continuous action allows the model to iteratively refine predictions and capture non-local correlations through temporal dynamics rather than requiring deeper spatial architectures, reducing computational cost.
Solution Approach 2:
The patent changes the parameters of the network by introducing temporal dynamics through recurrent connections and hidden states. Instead of increasing spatial capacity with deeper layers, the model uses temporal parameters (hidden state dimensions, recurrence rates) to capture non-local correlations, achieving the same effect with lower computational cost.
3Adaptability or versatility
If vision transformers are used to reduce locality assumptions, then non-local pixel correlations are improved, but extensive data training is required which is not feasible with limited medical imaging datasets
Solution Approach 1:
The UNet architecture performs preliminary feature extraction and spatial encoding in its encoder path before temporal processing occurs in the decoder. This preliminary action prepares the data in a form that requires less training, allowing the recurrent temporal model to focus specifically on capturing non-local correlations without needing extensive training data for basic feature learning.
Solution Approach 2:
The hidden state h_t acts as an intermediary that carries temporal and spatial information between network layers. This intermediary structure allows the model to capture non-local correlations in a parameter-efficient manner, requiring less training data compared to vision transformers that directly model all pairwise interactions.
4Loss of time
If under-sampling with high acceleration factor R is used, then scan time is reduced, but aliasing artifacts increase and reconstruction quality deteriorates
Solution Approach 1:
The recurrent neural network implements feedback through hidden states that are updated at each time step and fed back into the network. This feedback mechanism allows the model to iteratively refine the reconstruction, suppressing aliasing artifacts that result from high acceleration factor under-sampling and improving reconstruction quality over time.
Solution Approach 2:
The patent introduces dynamic temporal processing to the reconstruction model, where the network adapts its processing at each time step based on the current hidden state. This dynamic approach allows the model to effectively handle the varying aliasing patterns introduced by high acceleration factors, improving reconstruction quality while maintaining fast scan times.
Data Source
AI summary
For reconstruction in medical imaging, such as reconstruction in MR imaging, scanning is accelerated by under-sampling. In iterative reconstruction, the input to the regularizer is altered provide for correlation of non-local aliasing artifacts. Duplicates of the input image are shifted by different amounts based on the level of acceleration. The resulting shifted images are used to form the input to the regularizer. Providing an input based on shifts allows the regularization to suppress non-local as well as local aliasing artifacts.