Semi-supervised MRI Synthesis with Physics Guidance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for synthesizing contrast-weighted images in time-resolved MRI rely on fully-sampled ground truths for model training, which is impractical due to lengthy scan times, limiting the compilation of diverse clinical training datasets and hindering the development of robust models.
Innovation Solution
A semi-supervised learning approach using heavily under-sampled acquisitions with a physics-guidance module and complementary sampling masks across subjects and contrasts, enabling the synthesis of contrast-weighted images from time-resolved data, reducing scan time requirements and increasing the diversity of training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If fully-sampled ground truth images are used for model training, then image synthesis quality is improved, but scan time increases to 20-40 minutes making data collection impractical
Solution Approach 1:
The patent applies partial action by using heavily under-sampled acquisitions (e.g., 8-fold or 16-fold acceleration) instead of fully-sampled ground truths for training. The model learns to synthesize contrast-weighted images from time-resolved data using a semi-supervised loss function that operates on the accelerated, under-sampled data, thereby reducing scan time from 20-40 minutes to 30 seconds-3 minutes while maintaining synthesis quality
Solution Approach 2:
The patent introduces an intermediary approach by using a physics-guidance module that generates under-sampled multi-coil counterparts as a bridge between the accelerated acquisitions and the synthesis target. This intermediary enables the model to learn from accelerated data without requiring fully-sampled ground truths, resolving the contradiction between training quality and acquisition time
2Measurement precision
If contrast-weighted images are acquired using conventional methods, then diagnostic quality is improved, but examination time increases to 20-40 minutes
Solution Approach 1:
The patent applies preliminary action by acquiring time-resolved data with quantitative tissue parameter maps first, which serve as the foundation for subsequent contrast-weighted image synthesis. The model uses these preliminary quantitative maps to generate the final diagnostic images, eliminating the need for separate lengthy contrast-weighted acquisitions
Solution Approach 2:
The patent replaces the mechanical acquisition process (actual contrast-weighted scans taking 20-40 minutes) with a computational synthesis process using deep learning models trained on accelerated data. This substitution transforms the time-consuming physical imaging process into a rapid computational task, reducing examination time while maintaining diagnostic quality
3Adaptability or versatility
If training data is collected from diverse clinical populations, then model robustness is improved, but data collection becomes impractical due to long scan times
Solution Approach 1:
The patent applies parameter changes by modifying the acquisition parameters to use heavy acceleration factors (8-fold or 16-fold) in the k-space sampling pattern. This parameter change reduces the scan time from 20-40 minutes to 30 seconds-3 minutes, making it feasible to collect data from diverse clinical populations and train robust models
Data Source
AI summary
A method for magnetic resonance imaging acquires time-resolved k-space data by a magnetic resonance imaging apparatus, and generates contrast-weighted images by a multi-task generator G from the time-resolved k-space data. The multi-task generator G comprises a deep learning neural network trained using prospectively under-sampled ground truth images acquired using an acceleration factor of at least 8 without any fully-sampled ground truth images. The multi-task generator G is also trained using a physics guidance model and a semi-supervised loss function.