Deep Learning MR Image Standardization Across Vendors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Radiologists face challenges in comparing sequential imaging studies acquired on different magnetic resonance (MR) hardware systems due to varying contrast and distortions, making clinical imaging trials more complex when multiple vendor scanners are involved.
Innovation Solution
The development of methods and systems that use deep learning to transform MR images from one vendor's style to another or to a standardized MR style, preserving anatomical information while adjusting contrast, resolution, and image distortion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If MR images are acquired using different vendor scanners, then more imaging options and hardware availability are provided, but image contrast and distortion vary making comparison challenging
Solution Approach 1:
The patent introduces a deep learning-based image transformation system as an intermediary that converts images from different vendor scanners into a standardized target style. This mediator processes the varying contrast and distortion characteristics of different vendors' images, transforming them into a unified representation that enables accurate comparison while preserving the ability to use diverse hardware platforms
Solution Approach 2:
The system changes the stylistic parameters of MR images (contrast, distortion, intensity distribution) through deep learning transformation. By learning the mapping between different vendor-specific image styles and a standardized target style, the system transforms images to have consistent parameters across vendors, enabling accurate comparison while maintaining hardware diversity
2Productivity
If multiple vendor scanners are used in clinical imaging trials, then scanner availability and patient access improve, but image standardization and quantifiable comparison become more difficult
Solution Approach 1:
The system performs preliminary image style transformation before clinical analysis and comparison. By pre-processing images from multiple vendors to a standardized target style using deep learning, the system establishes uniformity in advance, allowing subsequent clinical workflows to proceed with images that are already standardized, thus maintaining both high patient access and image standardization
3Measurement precision
If deep learning transformation is applied to MR images, then image style standardization and diagnostic accuracy improve, but processing time and computational complexity increase
Solution Approach 1:
The deep learning transformation model is trained in advance on large datasets of multi-vendor images. Once trained, the model can rapidly transform new images to the target style. This preliminary training phase separates the computationally intensive learning process from the actual clinical application, allowing fast inference during patient workflows while still achieving high diagnostic accuracy through the learned transformations
Data Source
AI summary
A computer-implemented method for transforming magnetic resonance (MR) imaging across multiple vendors is provided. The method comprises: obtaining a training dataset, wherein the training dataset comprises a paired dataset and an un-paired dataset, and wherein the training dataset comprises image data acquired using two or more MR imaging devices; training a deep network model using the training dataset; obtaining an input MR image; and transforming the input MR image to a target image style using the deep network model.