Synthetic Electron Density Imaging From MRI Without CT Scans
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing radiotherapy planning methods using combined CT/MRI workflows are costly, time-consuming, and expose patients to radiation, while MRI-only techniques face challenges in accurately generating synthetic electron density images due to the lack of correspondence between MR image pixel intensity and tissue radiation attenuation properties.
Innovation Solution
A machine-learning model is trained to parameterize an image transfer function on MR images, generating synthetic electron density images without the need for segmentation, thus improving accuracy and detail by using a trained machine-learning model to generate synthetic electron density images from MR images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a combined CT/MRI workflow is used for radiotherapy planning, then soft tissue delineation accuracy is improved, but patient radiation exposure and treatment cost increase
Solution Approach 1:
The patent creates a synthetic CT image that copies the electron density information from a real CT scan, allowing the MRI to serve as the primary imaging modality while generating CT-like attenuation maps through machine learning. This eliminates the need for actual CT scans in MRI-only workflows, removing radiation exposure while preserving the ability to calculate radiation dose accurately.
Solution Approach 2:
The patent introduces a machine learning model as an intermediary that translates MRI signal intensities into CT electron density values. This mediator bridges the gap between MRI's superior soft tissue contrast and CT's radiation attenuation information, allowing accurate dose calculation without direct CT imaging.
2Object-affected harmful factors
If MRI-only workflow with synthetic CT generation is used, then patient radiation exposure is reduced, but electron density accuracy deteriorates due to lack of correspondence between MR pixel intensity and radiation attenuation
Solution Approach 1:
The patent transforms the relationship between MRI signal intensity and electron density by training a machine learning model on paired MRI-CT data. The model learns the complex, non-linear mapping between MRI parameters (T1, T2, proton density) and CT electron density values, enabling accurate conversion without direct correspondence in the raw data.
Solution Approach 2:
The patent moves from direct one-to-one pixel intensity mapping to a multi-dimensional approach using convolutional neural networks that analyze local neighborhoods and contextual information. This dimensional expansion allows the model to resolve ambiguities (e.g., bone vs. air both appearing dark in MRI) by considering surrounding tissue patterns and anatomical context.
3Measurement precision
If segmentation-based approach is used for sCT generation, then tissue-specific conversion models can be applied, but manual segmentation complexity and time consumption increase
Solution Approach 1:
The patent enables the system to automatically perform segmentation and tissue classification without human intervention. The machine learning model autonomously identifies different tissue types (bone, soft tissue, air, fat) and applies appropriate conversion parameters, eliminating the need for manual segmentation while maintaining tissue-specific accuracy.
Solution Approach 2:
The patent combines multiple functions into a single end-to-end machine learning model: segmentation, tissue classification, and electron density prediction are performed simultaneously rather than as separate manual steps. This integration maintains the benefits of tissue-specific conversion while eliminating manual segmentation complexity.
4Measurement precision
If atlas-based approach is used for sCT generation, then electron density information can be transferred from reference images, but registration time and computational complexity increase
Solution Approach 1:
The patent pre-trains the machine learning model on large datasets of paired MRI-CT images before actual sCT generation. This preliminary training phase captures the statistical relationships and anatomical variations, allowing rapid inference on new images without time-consuming registration or atlas matching during clinical use.
Solution Approach 2:
The patent replaces the mechanical registration and atlas-matching process with a data-driven machine learning approach. Instead of geometrically aligning images and transferring electron density values through complex registration algorithms, the neural network directly predicts electron density from MRI input, dramatically reducing computational time and complexity.
Data Source
Figure 1~3A
Figure 3B~3C
Figure 4A
AI summary
A conversion device (20) is operable to perform a learning-based method of generating a synthetic electron density image (sCT) of an anatomical portion based on one or more magnetic resonance, MR, images. The method is processing-efficient and capable of producing highly accurate sCT images irrespective of misalignment in the underlying training set. The conversion device (20) receives and installs a machine-learning model (22) trained to predict coefficients of an image transfer function (24). The conversion device (20) then receives a current set of MR images (MRI) of the anatomical portion, computes current coefficients ([C]) of the image transfer function (24) by operating the machine-learning model (22) on the current set of MR images (MRI), and computes a current sCT image of the anatomical portion by operating the current coefficients ([C]), in accordance with the image transfer function (24), on the current set of MR images (MRI).