Neural Network PET Image Transformation for Cost Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High-end PET/CT systems face challenges in cost reduction while maintaining performance, as less expensive crystals compromise TOF resolution, and conventional software solutions are insufficient to bridge the gap between system performance and customer needs in low-end systems.
Innovation Solution
A deep learning approach that utilizes a neural network to transform images from low-end PET/CT systems into images comparable to those from high-end systems by training on datasets from both systems, minimizing difference metrics and improving image quality without hardware modifications.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If less expensive crystals are used in PET/CT systems, then manufacturing cost is reduced, but TOF resolution deteriorates
Solution Approach 1:
The patent creates a virtual copy of high-end system images through neural network transformation. The deep learning model learns the mapping between low-end and high-end system images, generating synthetic images that replicate the quality characteristics of expensive systems without requiring the actual expensive hardware components
Solution Approach 2:
The patent replaces the physical hardware improvement path (upgrading crystals) with a software-based neural network system. Instead of improving the physical detection mechanism, the invention uses deep learning algorithms to post-process and transform images, substituting computational processing for mechanical/hardware enhancement
2Measurement precision
If conventional reconstruction algorithms are used to improve low-end system performance, then image quality is partially improved, but the gap between system performance and customer needs remains insufficient
Solution Approach 1:
The patent fundamentally changes the approach parameter from traditional iterative reconstruction algorithms to a deep learning-based transformation model. This parameter change enables the system to achieve superior image quality improvement by learning complex non-linear mappings between low-end and high-end system images, rather than relying on conventional linear or iterative methods
3Measurement precision
If deep learning transformation is applied to low-end system images, then image quality comparable to high-end systems is achieved, but computational processing time increases
Solution Approach 1:
The patent performs preliminary training of the neural network model using paired datasets from low-end and high-end systems before actual clinical use. This preliminary action creates a pre-trained transformation model that can be rapidly applied to patient images, separating the computationally intensive learning phase from the clinical application phase
Data Source
Figure 1
Figure 2
AI summary
An imaging method (100) includes: acquiring first training images of one or more imaging subjects using a first image acquisition device (12); acquiring second training images of the same one or more imaging subjects as the first training images using a second image acquisition device (14) of the same imaging modality as the first imaging device; and training a neural network (NN) (16) to transform the first training images into transformed first training images having a minimized value of a difference metric comparing the transformed first training images and the second training images.