Super Resolution MR Image Processing for Radiotherapy Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current MR guided radiotherapy (MRgRT) systems face limitations in spatial resolution due to low-field MRI systems, leading to motion blurring and reduced signal-to-noise ratio (SNR), which affects target tracking and dose calculation accuracy.
Innovation Solution
A novel cascaded deep learning (DL) super-resolution (SR) framework is applied to enhance the spatial resolution of MR images without altering hardware or scanning parameters, utilizing low-resolution cine MRI images to generate high-resolution images through postprocessing techniques.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If low-field MRI systems are used for MR guided radiotherapy, then the system can provide onboard imaging with excellent soft-tissue contrast without additional radiation exposure, but the spatial resolution is limited leading to motion blurring and reduced signal-to-noise ratio
Solution Approach 1:
A deep learning-based super-resolution framework is introduced as an intermediary processing step between image acquisition and clinical use. The framework includes a downsampling model that creates training pairs and a super-resolution model that reconstructs high-resolution images from low-resolution inputs, effectively mediating the resolution limitation without hardware changes
Solution Approach 2:
The patent replaces the mechanical/hardware-based resolution limitation with a computational/software-based solution. Instead of upgrading MRI hardware to achieve higher resolution, a cascaded deep learning framework processes the existing low-resolution images to generate high-resolution outputs, substituting physical constraints with algorithmic enhancement
2Measurement precision
If image resolution is improved from 3.5 mm to 0.9 mm through postprocessing, then edge sharpness and signal-to-noise ratio are enhanced, but computational processing time and complexity increase
Solution Approach 1:
The framework performs preliminary downsampling of high-resolution images to create training pairs before the actual super-resolution processing. This preliminary action prepares the data in advance, enabling the main super-resolution model to learn the mapping from low to high resolution more efficiently during the training phase
Solution Approach 2:
The cascaded framework maintains continuous processing through multiple stages: downsampling model generation, super-resolution model training, and real-time inference. The continuous refinement through cascaded stages ensures that computational resources are efficiently utilized while maintaining high resolution output quality
Data Source
AI summary
A computer implemented method of treatment targeting includes receiving magnetic resonance (MR) images of a subject including a target region, generating at least one contour of at least one surrogate element apart from the target region in the MR images, and determining a location of the target region in each of the MR images based on a location of the at least one contour in the MR images.