Deep-Learning SPECT Attenuation Correction Without CT
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current SPECT scanners, especially those without integrated CT, face challenges in accurate attenuation correction due to increased radiation exposure and motion artifacts, and existing deep-learning methods struggle to generate attenuation maps efficiently for dedicated cardiac scanners with small field-of-view.
Innovation Solution
A 3D Dual Squeeze-and-Excitation Residual Dense Network (DuRDN) is employed to directly generate attenuation-corrected SPECT images from non-attenuation-corrected SPECT images using photopeak and scatter windows, incorporating patient-specific information like BMI and gender, without intermediate attenuation map estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If CT scanning is used for attenuation correction in SPECT, then attenuation correction accuracy is improved, but radiation exposure to patients increases
Solution Approach 1:
The patent uses a deep learning network as an intermediary to generate attenuation maps from scatter window SPECT images, replacing the direct CT scanning approach. The network learns the mapping between scatter images and attenuation maps, enabling accurate attenuation correction without requiring additional CT radiation exposure to the patient
Solution Approach 2:
The invention creates a synthetic copy of the attenuation map information by training the deep learning network on paired data of CT-based attenuation maps and corresponding scatter window images. The network learns to reproduce attenuation map characteristics from scatter images alone, eliminating the need for actual CT scanning while maintaining correction accuracy
2Measurement precision
If CT scanning is used for attenuation correction in SPECT, then attenuation correction accuracy is improved, but device complexity and cost increase
Solution Approach 1:
The patent makes the SPECT scanner multi-functional by enabling it to perform both primary imaging and attenuation map generation using the same hardware components. The scatter window data, originally intended only for correction purposes, is repurposed as input for the deep learning network to generate attenuation maps, eliminating the need for separate CT hardware
Solution Approach 2:
The invention extracts the attenuation map generation function from the CT scanner and relocates it to the SPECT system's reconstruction pipeline. By separating the attenuation correction capability from the CT hardware dependency, the system achieves accurate attenuation correction using only SPECT scanner components and software-based deep learning
3Measurement precision
If iterative optimization methods are used to estimate attenuation maps from SPECT data, then attenuation correction is achieved, but processing time increases
Solution Approach 1:
The patent performs preliminary action by pre-training the deep learning network on a large dataset of paired scatter images and attenuation maps before actual clinical use. This offline training phase captures the complex mapping relationships, enabling rapid inference during reconstruction without requiring time-consuming iterative optimization during patient scanning
Solution Approach 2:
The invention replaces the mechanical iterative optimization process with a software-based deep learning inference system. The trained network directly predicts attenuation maps from scatter images in a single forward pass, substituting the gradual convergence of iterative algorithms with immediate computational prediction, thereby dramatically reducing reconstruction time
4Measurement precision
If iterative optimization methods are used to estimate attenuation maps, then attenuation correction is achieved, but noise level increases
Solution Approach 1:
The patent incorporates feedback mechanisms through the deep learning network's training process, where the network learns from labeled data with known ground truth attenuation maps. During training, the network adjusts its parameters based on the error between predicted and actual attenuation maps, enabling it to generalize accurately even in low-activity regions where iterative methods struggle with noise amplification
Solution Approach 2:
The invention changes the fundamental parameter of how attenuation maps are computed - from solving an ill-posed inverse problem through iterative optimization to direct prediction through a trained neural network. This parameter change in the computational approach transforms a noise-sensitive iterative process into a noise-robust prediction process that leverages learned patterns from training data
Data Source
AI summary
A system based upon artificial neural networks generates attenuation-corrected SPECT from non-attenuation-corrected SPECT (single photon emission computed tomography) without or with an intermediate step of attenuation map estimation. The system includes a SPECT scanner with CZT cameras for dynamic SPECT imaging. The system also includes a machine learning system including a 3D Dual Squeeze-and-Excitation Residual Dense Network for generating attenuation-corrected SPECT or attenuation maps from non-attenuation-corrected SPECT. The machine learning system reconstructs images from photopeak window and one or more scatter windows of the SPECT scanner are fed to the 3D Dual Squeeze-and-Excitation Residual Dense Network to generate attenuation-corrected SPECT or attenuation maps.