Pseudo CT Image Generation via Multi-Task Neural Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional radiation therapy treatment planning and PET imaging rely on CT images for accurate electron density information, but MR-only workflows are sought to reduce patient CT dose exposure, with previous methods failing to accurately generate pseudo CT images due to biased training towards soft tissue and background regions, leading to errors in bone value assignment and dose calculation.
Innovation Solution
A deep multi-task neural network is employed to generate pseudo CT images by translating MR images, with specific tasks of whole image translation, accurate segmentation, and image value estimation, using a U-Net architecture with multiple output layers to improve bone value accuracy and reduce errors in bone regions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional CT imaging is used to obtain electron density information, then accurate dose calculation and attenuation correction are achieved, but patients receive unnecessary CT radiation dose exposure
Solution Approach 1:
The patent creates a pseudo-CT image as a copy of the MR image, translating MR intensity values to CT-like Hounsfield units through neural network processing. This pseudo-CT copy provides the necessary electron density information for dose calculation without requiring actual CT scanning, thus eliminating radiation exposure while maintaining accuracy
Solution Approach 2:
The patent replaces the physical CT scanning mechanism with a computational translation process using deep learning neural networks. Instead of using X-rays to generate density information, the system uses MR images processed through trained neural networks to generate equivalent electron density data, substituting mechanical/physical measurement with computational modeling
2Object-affected harmful factors
If MR-only workflow is used to eliminate CT dose exposure, then patient radiation safety is improved, but accuracy of pseudo CT image generation deteriorates due to biased training towards soft tissue and background regions
Solution Approach 1:
The patent applies local quality by creating a focused loss function that specifically targets bone regions during neural network training. Instead of uniform training across all image regions, the system identifies bone areas and applies weighted loss calculations that prioritize accurate prediction of bone Hounsfield units, ensuring high precision in critical regions while maintaining overall image quality
Solution Approach 2:
The patent extends the training objective from standard pixel-wise loss to include regional and structural dimensions. By incorporating bone segmentation masks and applying loss functions across multiple spatial scales and anatomical regions, the system optimizes performance in the critical bone dimension without compromising soft tissue representation
3Ease of manufacture
If standard neural network training is used for pseudo CT generation, then training simplicity is maintained, but bone region accuracy deteriorates due to large dynamic range and spatial sparsity of bone values
Solution Approach 1:
The patent segments the training process into distinct components: bone region identification through segmentation masks, focused loss calculation for bone areas, and separate optimization targets for different tissue types. This segmentation allows the network to learn bone characteristics independently while maintaining overall image translation accuracy
Solution Approach 2:
The patent introduces bone segmentation masks as an intermediary element that guides the training process. These masks act as mediators between the input MR image and the loss function, highlighting regions that require special attention and enabling the network to focus computational resources on accurately predicting bone Hounsfield units
Data Source
AI summary
Various methods and systems are provided for translating magnetic resonance (MR) images to pseudo computed tomography (CT) images. In one embodiment, a method comprises acquiring an MR image, generating, with a multi-task neural network, a pseudo CT image corresponding to the MR image, and outputting the MR image and the pseudo CT image. In this way, the benefits of CT imaging with respect to accurate density information, especially in sparse regions of bone which exhibit with high dynamic range, may be obtained in an MR-only workflow, thereby achieving the benefits of enhanced soft-tissue contrast in MR images while eliminating CT dose exposure for a patient.