Deep Set Reconstruction for MRI Noise Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current medical imaging techniques, such as MR, CT, PET, and SPECT, often require repetitive scanning to reduce noise in images, which can be time-consuming and inefficient, especially when using deep learning methods that rely on unrolled iterative algorithms, as they require multiple iterations and significant computational resources.
Innovation Solution
A deep set-based approach is used to train a machine learning model for medical image reconstruction, where the loss function is based on the aggregation of images from multiple repetitions, allowing the model to learn invariance by permutation and improve reconstruction quality without increasing computational time or memory requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If unrolled iterative algorithms with multiple iterations are used for deep learning-based reconstruction, then reconstruction quality is improved, but computational time and memory requirements increase
Solution Approach 1:
The patent applies partial action by using a single iteration of the unrolled algorithm instead of multiple iterations, achieving sufficient reconstruction quality without the full computational burden. The deep set framework allows the model to learn from aggregated data across multiple repetitions in one pass, effectively performing partial iterations that balance quality and speed.
Solution Approach 2:
The patent performs preliminary action by aggregating data from multiple repetitions before reconstruction, using deep sets to learn permutation-invariant representations in advance. This pre-processing of aggregated information enables the single-iteration reconstruction to achieve quality comparable to multiple iterations, as the model has already processed combined information from all repetitions beforehand.
2Manufacturing precision
If unrolled iterative algorithms with multiple iterations are used for deep learning-based reconstruction, then reconstruction quality is improved, but memory requirements increase
Solution Approach 1:
The patent uses partial action by limiting the process to one iteration instead of multiple, thereby requiring less memory to store intermediate results from successive iterations. The deep set architecture processes aggregated data from multiple repetitions in a single pass, reducing the memory footprint compared to storing multiple iteration states.
Solution Approach 2:
The patent extracts only the essential aggregated information from multiple repetitions using deep set permutation-invariant functions, discarding redundant iterative processing steps. By taking out only the necessary aggregated features rather than maintaining full iterative computation states, memory requirements are significantly reduced while preserving reconstruction quality.
3Manufacturing precision
If repetition is used to acquire multiple contrasts or reduce noise, then image quality is improved, but scanning time increases
Solution Approach 1:
The patent merges information from multiple repetitions using deep set aggregation functions that are permutation-invariant, combining the benefits of multiple scans (noise reduction, contrast information) into a unified representation. This merging occurs in the learning phase rather than requiring sequential processing, effectively combining data from all repetitions simultaneously to improve image quality without proportionally increasing scanning time.
Solution Approach 2:
The patent changes the parameter of aggregation from simple averaging to deep set-based permutation-invariant aggregation, which extracts more meaningful features from the repeated measurements. This parameter change in the aggregation function allows the system to efficiently utilize multiple repetitions for improved image quality while minimizing the time penalty, as the deep set framework processes all repetitions in parallel during training.
Data Source
AI summary
For reconstruction in medical imaging using a scan protocol with repetition, a machine learning model is trained for reconstruction of an image for each repetition. Rather than using a loss for that repetition in training, the loss based on an aggregation of images reconstructed from multiple repetitions is used to train the machine learning model. This loss for reconstruction of one repetition based on aggregation of reconstructions for multiple repetitions is based on deep set-based deep learning. The resulting machine-learned model may better reconstruct an image from a given repetition and/or a combined image from multiple repetitions than a model learned from a loss per repetition.