Deep Learning MRI Reconstruction for B1+ Inhomogeneity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Ultrahigh field MRI systems face challenges with severe B1+ inhomogeneity when using conventional single-channel transmit RF coils, leading to flip angle variations and signal dropout, especially in lower brain regions, which conventional RF parallel transmission techniques struggle to address due to tedious workflows and the need for specialized and expensive hardware.
Innovation Solution
A method using a trained machine learning model to convert magnetic resonance images acquired with single-channel transmit hardware into pTx-like images, leveraging a deep-learning framework to enhance image quality and reduce B1+ artifacts without requiring pTx hardware, by accessing and processing single transmission data with a computer system.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If conventional single-channel transmit RF coil is used at ultrahigh field, then hardware complexity is reduced, but B1+ inhomogeneity increases causing flip angle variations and signal dropout
Solution Approach 1:
The patent creates a virtual copy of parallel transmission functionality through deep learning. A trained neural network model learns to map single-channel transmit data to multi-channel transmit-like images, effectively copying the beneficial effects of pTx (B1+ correction, flip angle uniformity) without requiring actual multi-channel hardware. This allows single-channel systems to achieve pTx-like image quality through software-based synthesis.
Solution Approach 2:
The patent replaces the physical hardware system (multi-channel RF transmit coils and transmitters) with a computational system. Instead of using multiple physical RF channels to correct B1+ inhomogeneity, the invention uses a deep learning model that processes single-channel data and synthesizes corrected images, substituting mechanical/electrical complexity with algorithmic processing.
2Manufacturing precision
If RF parallel transmission is implemented, then B1+ inhomogeneity is corrected improving image quality, but device complexity and cost increase due to specialized hardware requirements
Solution Approach 1:
The patent creates a virtual copy of parallel transmission functionality through deep learning. A trained neural network model learns to map single-channel transmit data to multi-channel transmit-like images, effectively copying the beneficial effects of pTx (B1+ correction, flip angle uniformity) without requiring actual multi-channel hardware. This allows single-channel systems to achieve pTx-like image quality through software-based synthesis.
Solution Approach 2:
The patent extracts the essential function of parallel transmission (B1+ correction and flip angle uniformity) from its complex hardware implementation. By using a deep learning model trained on paired single-channel and multi-channel data, the system separates the corrective function from the hardware complexity, applying only the necessary image processing to achieve pTx-like results with simpler equipment.
3Manufacturing precision
If RF parallel transmission is used, then flip angle uniformity is improved, but workflow complexity increases due to calibration scans and optimization requirements
Solution Approach 1:
The patent performs the complex calibration and optimization work in advance during the model training phase. The deep learning model is trained on datasets that include calibration scans and ground truth images from multi-channel systems. Once trained, the model encapsulates all the complex corrections and optimizations, allowing for simple one-step inference during actual imaging without requiring operators to perform tedious calibration procedures.
Solution Approach 2:
The patent enables the system to automatically correct B1+ inhomogeneity and optimize flip angle uniformity without human intervention. The deep learning model self-adjusts and applies corrections based on the input single-channel data, eliminating the need for operators to manually perform calibration scans or solve optimization problems, thereby simplifying the workflow significantly.
4Device complexity
If single-channel transmit hardware is used, then cost and hardware simplicity are maintained, but image quality deteriorates due to signal dropout in lower brain regions
Solution Approach 1:
The patent creates a virtual copy of parallel transmission functionality through deep learning. A trained neural network model learns to map single-channel transmit data to multi-channel transmit-like images, effectively copying the beneficial effects of pTx (B1+ correction, flip angle uniformity) without requiring actual multi-channel hardware. This allows single-channel systems to achieve pTx-like image quality through software-based synthesis.
Solution Approach 2:
The patent replaces the physical hardware system (multi-channel RF transmit coils and transmitters) with a computational system. Instead of using multiple physical RF channels to correct B1+ inhomogeneity, the invention uses a deep learning model that processes single-channel data and synthesizes corrected images, substituting mechanical/electrical complexity with algorithmic processing.
Data Source
AI summary
Magnetic resonance images with improved image quality consistent with those obtained using parallel radio frequency (“RF”) transmission (“pTx”) techniques are generated from data acquired using single transmission hardware (e.g., single channel RF transmission). A deep-learning framework is used to train a deep neural network to convert images obtained with single transmission into pTx-like images. The pTx-like images have reduced signal variations and dropouts that may otherwise be attributable to B1+ inhomogeneities.