Spectral Imaging via Deep Learning on Non-Spectral CT Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional computed tomography (CT) scanners without spectral imaging capabilities require expensive specialized hardware and cannot utilize existing protocols, leading to degraded image quality and increased data storage needs when attempting to produce spectral images.
Innovation Solution
A deep learning-based system processes non-spectral volumetric image data to estimate spectral data, including basis components and images, without the need for multiple X-ray tubes or spectral detectors, using a spectral results module with a denoiser, spectral basis estimator, and spectral image estimator, allowing for the use of existing acquisition and reconstruction protocols.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If specialized spectral imaging hardware (multiple X-ray tubes, kVp switching, spectral detectors) is used, then spectral imaging capability is improved, but system cost and device complexity increase
Solution Approach 1:
The patent uses a neural network to learn the mapping between non-spectral and spectral images by training on paired data. The network copies the spectral imaging function through software rather than hardware, enabling spectral imaging capability in conventional scanners without physical spectral detectors
Solution Approach 2:
The patent replaces the mechanical/optical spectral imaging system (multiple X-ray tubes, kVp switching hardware, spectral detectors) with a computational approach using deep learning. The neural network processes conventional non-spectral images to generate spectral images, substituting physical spectral measurement mechanisms with algorithmic transformation
2Adaptability or versatility
If specialized spectral imaging hardware is used, then spectral imaging capability is improved, but overall scanner cost increases
Solution Approach 1:
The patent copies the spectral imaging function through software (neural network) rather than requiring expensive spectral imaging hardware. This allows conventional scanners to provide spectral imaging capability through a software add-on, avoiding the need to manufacture and deploy costly specialized hardware systems
Solution Approach 2:
The patent makes conventional non-spectral scanners multi-functional by adding spectral imaging capability through software. The same hardware can perform both conventional non-spectral imaging and spectral imaging by applying the trained neural network, eliminating the need for separate specialized spectral imaging systems
3Adaptability or versatility
If spectral imaging protocols are implemented on non-spectral scanners, then spectral images can be produced, but image quality degrades and data storage requirements increase
Solution Approach 1:
The patent performs preliminary training of the neural network using paired spectral and non-spectral images before deployment. This pre-training phase allows the network to learn optimal transformation parameters and characteristics, ensuring high-quality spectral image generation when the trained network is applied to new non-spectral images
Solution Approach 2:
The patent transforms the non-spectral image data into spectral image space by applying learned parameter transformations through the neural network. The network learns to map intensity values and spectral characteristics from the non-spectral domain to the spectral domain, effectively changing the parameter space to generate spectral images
Data Source
Figure 1
Figure 2~3
Figure 4~5
AI summary
An imaging system (102) includes a radiation source (112) configured to emit X-ray radiation, a detector array (114) configured to detect X-ray radiation and generate a signal indicative thereof, an a reconstructor (116) configured to reconstruct the signal and generate non-spectral image data. The imaging system further includes a processor (124) configured to process the non-spectral image data using a deep learning regression algorithm to estimate spectral data from a group consisting of spectral basis components and a spectral image.