Neural Network Tissue Map for Accurate SAR Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current magnetic resonance (MR) scanners estimate Specific Absorption Rate (SAR) roughly and without spatial accuracy, leading to safety margin limitations, especially at high magnetic field strengths and with parallel excitation, due to the use of generic subject models that fail to accurately differentiate between bone and air in MR images.
Innovation Solution
A computer-implemented method using a trained neural network to generate a subject-specific map of tissue properties by improving contrast between bone and air in MR images, segmenting them into distinct tissue types, and assigning predetermined values for density and electromagnetic properties, allowing for accurate SAR estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a generic model of the subject's body is used for SAR estimation, then the estimation can be performed quickly and simply, but the spatial accuracy and reliability of the SAR estimation deteriorates
Solution Approach 1:
The patent performs preliminary segmentation of the subject's anatomy into tissue types (bone, air, soft tissue) before SAR estimation. This preliminary action creates a subject-specific anatomical model that can be reused for accurate SAR calculations, resolving the contradiction by preparing accurate tissue maps in advance rather than during the SAR estimation process itself.
Solution Approach 2:
The patent assigns different electromagnetic properties to different tissue types (bone, air, soft tissue) based on their local characteristics. This local quality approach allows the SAR estimation to account for spatial variations in tissue properties, improving accuracy while maintaining computational efficiency through targeted detailed modeling only where needed.
2Measurement precision
If a subject-specific model with detailed tissue differentiation is created, then the SAR estimation accuracy improves, but the complexity of image processing and model generation increases
Solution Approach 1:
The patent segments the MR image into distinct tissue types (bone, air, soft tissue) based on signal intensity characteristics. This segmentation simplifies the complex task of creating subject-specific models by dividing it into manageable classification steps, reducing processing complexity while maintaining high SAR estimation accuracy through precise tissue differentiation.
Solution Approach 2:
The patent uses parameter changes in the neural network processing to differentiate tissue types. By transforming the MR image data through learned parameter transformations, the system automatically identifies tissue boundaries and characteristics without requiring complex manual processing, thus improving accuracy while controlling complexity.
3Ease of manufacture
If standardised patient information from MR-CT Atlas dictionary is used for tissue identification, then the processing is simplified, but the patient-specific accuracy of SAR calculations deteriorates
Solution Approach 1:
The patent employs a neural network that automatically identifies and segments tissue types directly from the patient's own MR image without requiring external atlas data or manual intervention. This self-service approach allows the system to generate patient-specific anatomical models autonomously, improving reliability while maintaining ease of use through automated processing.
Solution Approach 2:
The patent replaces the mechanical/atlas-based tissue identification system with a neural network-based automated recognition system. This substitution eliminates the need for standardized atlas dictionaries and manual matching processes, providing patient-specific accuracy through learned patterns while simplifying the overall process through automation.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
The invention relates to a method for generating a subject-specific map (20) of tissue properties of a region of interest within a subject, the method comprising the steps of receiving an MR image (2) of the region of interest; feeding the MR image (2) as input into at least one trained neural network (4, 6), wherein the output of the neural network (4, 6) is an output image (8, 10) having improved contrast between bone and air; segmenting (12) the output image (8, 10) into air, bone and at least one further type of tissue to obtain a segmented image (13); and assigning (14) predetermined values for the density and for at least one electromagnetic property to air, bone and each other type of tissue in the segmented image (13), to obtain a subject-specific map (20) of tissue properties. The invention is also directed to a method of estimating SAR from the map of tissue properties.