PET Attenuation Map Generation with DCNNs for CT Artifact Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing PET imaging technologies face inaccuracies in attenuation estimation due to variations in effective X-ray energy and incorrect material classification, leading to artifacts and errors in image reconstruction.
Innovation Solution
Utilizing a trained deep convolutional neural network (DCNN) to translate CT images into accurate attenuation maps for 511 keV gamma rays, incorporating spectral CT data and material decomposition techniques to capture signatures of different materials and artifacts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional CT-based attenuation estimation is used, then the process is simple and fast, but the accuracy is limited due to beam-hardening artifacts and incorrect material classification
Solution Approach 1:
The patent replaces the traditional mechanical/mathematical CT-based attenuation estimation system with a deep learning-based system. Specifically, a U-Net convolutional neural network is trained to directly convert CT images into attenuation maps, substituting the conventional iterative reconstruction and material classification processes with a data-driven neural network that learns the complex mapping relationship between CT images and attenuation coefficients.
Solution Approach 2:
The patent changes the fundamental parameter used for attenuation estimation from CT numbers (which are affected by beam-hardening) to attenuation coefficients derived through deep learning. The neural network processes CT images with varying effective energies and outputs attenuation maps that are corrected for beam-hardening effects, effectively changing the parameter representation to achieve higher accuracy.
2Measurement precision
If spectral CT is used to improve material decomposition, then the accuracy of attenuation maps improves, but the scan time and complexity increase
Solution Approach 1:
The patent performs preliminary action by training the U-Net neural network in advance using paired data from spectral CT scans and ground truth attenuation maps. During actual use, the pre-trained network can rapidly convert routine CT images into accurate attenuation maps without requiring time-consuming spectral CT scans, effectively preparing the model beforehand to handle the transformation efficiently.
Solution Approach 2:
The patent creates a digital copy of the spectral CT processing capability through the neural network. Instead of physically performing spectral CT scans for every patient, the system creates a learned representation (copy) of the spectral information and material decomposition results through the trained U-Net model, which can then reproduce accurate attenuation maps from standard CT data much faster.
3Productivity
If simple material classification from CT images is used, then the processing is fast, but artifacts are introduced due to incorrect material identification
Solution Approach 1:
The patent replaces the simple, rule-based material classification system with a deep learning-based material decomposition system. The U-Net network learns complex patterns in CT images and automatically identifies different materials and their attenuation characteristics, substituting fast but inaccurate classification algorithms with a more reliable neural network that processes spatial and intensity information to achieve both speed and accuracy.
Data Source
Figure 1A
Figure 1B
Figure 2A
AI summary
An attenuation map generation method according to an embodiment comprises receiving computed tomography (CT) image data of a subject; and generating an attenuation map for Positron Emission Tomography (PET) image reconstruction by inputting the received CT image data into a trained deep convolutional neural network (DCNN) model that outputs the attenuation map, wherein the DCNN model was trained using training input image data and corresponding training attenuation map data generated from spectral CT image data.