Retrospective MRI Tissue Contrast Tuning via Deep Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Standard clinical MRI techniques often struggle to distinguish pathology from surrounding normal tissues due to fixed tissue contrast, requiring repeated scans with different protocols for optimal visualization.
Innovation Solution
Deep learning-based image processing methods using self-attention convolutional neural networks to retrospectively adjust soft tissue contrast in MRI images by deriving tissue relaxation parametric maps and field maps, allowing for the generation of images with alternative imaging protocols without additional data acquisition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If standard clinical MRI with fixed imaging protocol is used, then data acquisition is simple and fast, but tissue contrast cannot be optimized for different pathologies
Solution Approach 1:
The patent performs preliminary action by acquiring multiple MRI images with different imaging protocols (different T1, T2, proton density weightings) during the initial scan. These pre-acquired images serve as input for the deep learning model, which then generates additional contrast-optimized images without requiring additional patient scanning time.
Solution Approach 2:
The patent creates synthetic copies of MRI images with desired tissue contrasts by training a deep learning model to map from acquired images to parametric maps and then generate images corresponding to alternative imaging protocols. This copying approach avoids the need to physically rescan the patient with different protocols.
2Measurement precision
If repeated scans with different imaging protocols are performed to optimize tissue contrast, then optimal pathology visualization is achieved, but scanning time and patient burden increase
Solution Approach 1:
Instead of performing repeated physical scans, the patent uses a deep learning model to copy and synthesize images that would have been obtained from additional scans. The model learns the relationship between different imaging protocols from training data and generates corresponding images from the initially acquired set, achieving optimal pathology visualization without additional scanning time.
Solution Approach 2:
The patent replaces the mechanical scanning process with a computational approach. Rather than physically rescan the patient with different imaging protocols (mechanical system), a deep learning model computationally generates images with desired contrasts from the already acquired data, substituting the mechanical scanning process with an information-processing system.
3Measurement precision
If quantitative parametric maps are derived using conventional methods, then accurate tissue characterization is achieved, but additional data acquisition and processing time are required
Solution Approach 1:
The patent merges the tasks of image acquisition and parametric map generation into a single integrated process. The deep learning model simultaneously performs multiple functions: it derives T1, T2, and proton density parametric maps while also generating images corresponding to alternative imaging protocols, all from the initially acquired images without requiring separate data acquisition or processing steps.
Solution Approach 2:
The patent uses copying to generate parametric maps and alternative-protocol images from the acquired images through learned mappings. This approach avoids the need for separate quantitative mapping acquisitions and processing, achieving accurate tissue characterization efficiently through computational synthesis.
4Adaptability or versatility
If images are generated with alternative imaging protocols, then versatile tissue contrast is achieved, but additional data acquisition is traditionally required
Solution Approach 1:
The patent generates images corresponding to alternative imaging protocols by copying and transforming the information from the initially acquired images through a deep learning model. The model learns the relationships between different protocols and synthesizes appropriate images without requiring additional patient scans or data acquisition, achieving versatile tissue contrast from a single data set.
Solution Approach 2:
The deep learning model serves multiple functions: it derives quantitative parametric maps, generates images with alternative tissue contrasts, and provides versatile imaging protocol flexibility. This multi-functional approach achieves diverse imaging capabilities from a single data acquisition, eliminating the need for separate scans for different imaging needs.
Data Source
AI summary
Retrospective magnetic resonance imaging (MRI) uses a deep neural network framework [102] to generate from MRI imaging data [100] acquired by an MRI apparatus using a predetermined imaging protocol tissue relaxation parametric maps and magnetic/radiofrequency field maps [104] which are then used to generate using the Bloch equations [106] predicted MRI images [108] corresponding to imaging protocols distinct from the predetermined imaging protocol. This allows obtaining a wide spectrum of tissue contrasts distinct from those of the acquired MRI imaging data.