Neural Network MRI-to-CT Synthesis for Radiation Therapy Planning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating synthetic CT images from MRI images are inefficient, requiring significant computing resources and time, and often result in inaccurate representations due to the lack of a direct mathematical relationship between CT and MRI intensity values, particularly in distinguishing bone and air in traditional MRI images.
Innovation Solution
A computer-implemented method using a convolutional neural network (CNN) to convert MRI images into synthetic CT images, utilizing a training process that learns a direct mapping between MRI and CT images, reducing the need for inter-subject image registration and enabling faster generation of accurate synthetic CT images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional CT imaging is used to generate synthetic CT images, then accurate representation of patient geometry and electron densities is achieved, but additional radiation exposure to patients occurs
Solution Approach 1:
The patent creates a copy of the CT image appearance and characteristics from MRI data without using actual CT scanning. The neural network learns to generate synthetic CT images that replicate the Hounsfield unit values and tissue appearance of real CT images, allowing treatment planning to proceed with MRI-derived images that mimic CT characteristics without the associated radiation exposure
Solution Approach 2:
The patent replaces the physical CT imaging mechanism (X-ray generation and detection) with a computational approach using neural networks. Instead of using ionizing radiation to create images, the system uses learned mappings from MRI data to generate synthetic CT images, substituting a mechanical/physical imaging system with an information-processing system
2Object-affected harmful factors
If MRI images are used directly for radiation therapy planning, then radiation exposure is eliminated and soft tissue contrast is improved, but accurate electron density information and bone representation are lost
Solution Approach 1:
The patent introduces synthetic CT images as an intermediary between MRI and treatment planning systems. The neural network translates MRI intensity values into synthetic Hounsfield units that represent electron density, creating a bridge that allows MRI-derived data to be used in CT-based treatment planning workflows without losing the radiation-free advantage of MRI
Solution Approach 2:
The patent transforms the parameter space from MRI intensity values to CT Hounsfield units through the neural network. This parameter transformation allows the system to represent tissue electron density in the familiar CT scale while the underlying data comes from MRI, enabling accurate dose calculation without direct CT imaging
3Measurement precision
If atlas-based approaches are used to generate synthetic CT images, then tissue classification accuracy is improved, but computation time and processing resources are significantly increased
Solution Approach 1:
The patent performs the complex task of learning the mapping between MRI and CT characteristics in advance during a training phase. The neural network is pre-trained on paired MRI-CT datasets to learn tissue correspondence and intensity relationships, so that during actual synthetic CT generation, the pre-learned knowledge can be applied rapidly without repeating the complex analysis
Solution Approach 2:
The patent creates a computational model that copies the essential mapping relationships between MRI and CT data. Instead of performing complex atlas-based registration and tissue classification during image generation, the system uses a pre-trained neural network that has learned to directly translate MRI intensities to synthetic CT values, replicating the outcome of complex classification processes in a streamlined manner
Data Source
Figure 1A
Figure 1B
Figure 1C
AI summary
Systems, computer-implemented methods, and computer readable media for generating a synthetic image of an anatomical portion based on an origin image of the anatomical portion acquired by an imaging device using a first imaging modality are disclosed. These systems may be configured to receive the origin image of the anatomical portion acquired by the imaging device using the first imaging modality, receive a convolutional neural network model trained for predicting the synthetic image based on the origin image, and convert the origin image to the synthetic image through the convolutional neural network model. The synthetic image may resemble an imaging of the anatomical portion using a second imaging modality differing from the first imaging modality.