Synthetic-Data Neural Networks for 4D CT Hemodynamic Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for generating hemodynamic parametric maps from 4D computed tomography perfusion data suffer from low signal-to-noise ratio and errors due to bad registration and bolus superposition, leading to inaccurate hemodynamic parameter estimation.
Innovation Solution
Utilize deep learning algorithms trained on synthetic data to estimate hemodynamic parameters by correcting image non-idealities in 4D computed tomography perfusion data, using neural networks to determine residual impulse functions and hemodynamic parameters such as blood flow, blood volume, and mean transit time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deconvolution algorithm is used to retrieve hemodynamic features, then hemodynamic parameters can be obtained, but signal-to-noise ratio is low and measurement precision deteriorates
Solution Approach 1:
The patent uses synthetic data that copies the structural characteristics of real CT perfusion data while eliminating noise and registration errors. The neural network is trained on these synthetic copies to learn the mapping from input signals to hemodynamic parameters, achieving high measurement precision without the reliability issues of direct deconvolution methods.
Solution Approach 2:
The patent replaces the traditional mechanical deconvolution algorithm with a neural network-based computational model. The neural network learns non-linear relationships from synthetic data and can handle complex patterns that linear deconvolution methods cannot capture, thereby improving both measurement precision and signal-to-noise ratio.
2Measurement precision
If traditional least squares regression is used, then hemodynamic maps can be generated, but measurement precision deteriorates due to registration errors and bolus superposition
Solution Approach 1:
The patent converts the harmful effects of registration errors and bolus superposition into beneficial training data characteristics. By deliberately introducing these artifacts in the synthetic data generation process, the neural network learns to recognize and correct them, transforming what were previously sources of error into training opportunities that improve robustness.
Solution Approach 2:
The patent changes the approach from using fixed mathematical models (least squares regression) to using learned parameters from synthetic data. The neural network adapts its internal parameters during training to capture the relationships between input signals and hemodynamic parameters, allowing it to handle variations in registration and bolus conditions that rigid models cannot accommodate.
3Reliability
If deep learning algorithms are trained on synthetic data, then image non-idealities can be mitigated, but device complexity increases
Solution Approach 1:
The patent performs preliminary action by pre-training the neural network on synthetic data that mimics real CT perfusion conditions. This pre-training phase captures the essential relationships and robustness to artifacts before the network is applied to actual patient data, thereby improving reliability without requiring complex real-time computations during diagnosis.
Solution Approach 2:
The patent creates synthetic copies of real CT perfusion data that preserve the structural information needed for hemodynamic analysis while eliminating noise and artifacts. These synthetic copies serve as training data that simplifies the learning process and reduces the computational complexity required for accurate parameter estimation compared to processing real noisy data directly.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Methods and systems are described herein for hemodynamic parameter estimation. In certain embodiments, a set of perfusion data is acquired for a region of interest using an imaging system. An artery signal is obtained from the set of perfusion data. A tissue signal is obtained from the set of perfusion data. The artery signal and the tissue signal are provided as inputs to one or more neural networks to determine one or more hemodynamic parameters for the region of interest. The one or more neural networks are trained using one or more synthetic data.