Pseudo-CT Generation from MR Data Using Feature Regression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for generating pseudo-CT images from MR data in radiation therapy are limited by the lack of a simple mathematical relationship between MR intensity values and electron densities, leading to pseudo-CT images that lack anatomical details and are not standardized, resulting in inaccurate dose computation and increased radiation exposure.
Innovation Solution
A computer-implemented method using a learning-based approach with a training module and prediction module to construct a predictive model that predicts CT values from MR images by extracting multiple features and applying dimensionality reduction techniques, allowing for the generation of high-quality pseudo-CT images that can replace real CT images for radiation therapy planning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If traditional atlas-based methods are used to generate pseudo-CT images from MR data, then the process is simpler, but the pseudo-CT images lack anatomical details and have poor accuracy for dose computation
Solution Approach 1:
The patent applies preliminary action by constructing a predictive model in advance using training data from multiple patients. The model is trained offline to learn the mapping between MR image features and CT values, so that when generating pseudo-CT images for new patients, the pre-trained model can quickly and accurately predict CT values without requiring complex real-time processing or manual atlas registration.
Solution Approach 2:
The patent changes parameters by extracting multiple features from MR images (intensity, texture, shape) and using dimensionality reduction techniques to transform the feature space. The predictive model learns optimal parameter transformations to map MR features to accurate CT values, improving pseudo-CT accuracy while managing complexity through systematic feature engineering.
2Manufacturing precision
If multiple feature extraction and dimensionality reduction techniques are applied to improve pseudo-CT accuracy, then anatomical details are improved, but the processing time and computational complexity increase
Solution Approach 1:
The patent performs feature extraction, dimensionality reduction, and model training as preliminary actions during an offline phase using training datasets. This allows the system to pre-compute optimal feature transformations and store the learned model parameters, so that during actual pseudo-CT generation for new patients, only forward prediction is needed, significantly reducing processing time while maintaining high anatomical accuracy.
3Adaptability or versatility
If a learning-based predictive model is constructed to predict CT values from MR images, then pseudo-CT images can replace real CT images for radiation therapy planning, but the model requires extensive training data and computational resources
Solution Approach 1:
The patent creates a universal predictive model that can be applied to different patients, MR scanner types, and imaging sequences. The model is trained on diverse training data from multiple patients and is designed to generalize across different clinical scenarios, allowing a single model to replace the need for patient-specific or scanner-specific pseudo-CT generation methods, thereby achieving versatility while managing complexity through consolidation.
4Ease of manufacture
If atlas images are used as references for pseudo-CT generation, then the method is easier to implement, but the registration errors between atlas and patient images reduce accuracy
Solution Approach 1:
The patent uses copying by training the predictive model on paired MR and CT images from multiple patients to learn the statistical relationships and anatomical variations. Instead of copying from a single atlas image, the model learns from multiple examples and generalizes to new patients, eliminating registration errors while maintaining ease of implementation through automated prediction rather than manual atlas-based methods.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Systems and methods are provided for generating a pseudo-CT prediction model that can be used to generate pseudo-CT images. An exemplary system may include a processor configured to retrieve training data including at least one MR image and at least one CT image for each of a plurality of training subjects. For each training subject, the processor may extract a plurality of features from each image point of the at least one MR image, create a feature vector for each image point based on the extracted features, and extract a CT value from each image point of the at least one CT image. The processor may also generate the pseudo-CT prediction model based on the feature vectors and the CT values of the plurality of training subjects.