CNN Encoder-Decoder for Perfusion MRI Quantitative Map Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current perfusion MRI post-processing methods are slow, prone to noise, require high doses of contrast agents, and are user-dependent, leading to variable results.
Innovation Solution
A convolutional neural network (CNN) with an encoder-decoder architecture and skip connections is used to process perfusion sequences, generating quantitative maps of perfusion parameters automatically, quickly, and robustly to noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If classical deconvolution methods are used for perfusion sequence processing, then quantitative perfusion maps can be generated, but the processing is very slow and requires up to several hours
Solution Approach 1:
The patent replaces the mechanical deconvolution computation system with a neural network-based system. The CNN is trained on perfusion sequences and their corresponding quantitative maps, then uses the learned patterns to directly generate quantitative maps from new perfusion sequences, eliminating the need for slow iterative deconvolution calculations.
Solution Approach 2:
The patent performs preliminary training of the neural network using a large dataset of perfusion sequences and their corresponding quantitative maps before actual processing. This preliminary action allows the network to learn the complex relationships between perfusion data and quantitative parameters, enabling fast processing without repeated slow calculations.
2Measurement precision
If deconvolution methods are used to generate quantitative perfusion maps, then perfusion parameters can be estimated, but the results are highly sensitive to noise and signal-to-noise ratio
Solution Approach 1:
The patent replaces the noise-sensitive deconvolution mathematical system with a neural network system that learns robust patterns from training data. The CNN processes the entire perfusion sequence and extracts features that are inherently more resistant to noise, providing reliable quantitative maps even in low signal-to-noise conditions.
Solution Approach 2:
The patent uses skip connections in the U-Net architecture that provide feedback pathways, allowing the network to refine its predictions by comparing intermediate features with target information. This feedback mechanism helps correct errors and maintain accuracy even when input data contains noise.
3Illumination intensity
If higher doses of gadolinium-based contrast agents are used to improve signal quality, then perfusion imaging quality improves, but patient safety concerns increase and contrast agent usage should be minimized
Solution Approach 1:
The patent changes the processing approach parameter from traditional deconvolution to neural network-based processing. This parameter change allows the system to extract meaningful perfusion information from low-dose contrast agent data, as the neural network learns to identify patterns even in noisier, lower-signal conditions that result from reduced contrast usage.
Solution Approach 2:
The patent introduces the neural network as an intermediary between the raw perfusion sequences and the quantitative maps. This intermediary learns to handle the noise and low signal quality inherent in low-dose imaging, translating the degraded input into accurate quantitative outputs without requiring high contrast agent doses.
4Ease of operation
If manual or semi-automatic delineation of large arteries is performed to estimate arterial input function, then perfusion processing can be initiated, but user input is cumbersome and results vary when users change
Solution Approach 1:
The patent enables the system to perform the arterial input function estimation automatically without user intervention. The neural network processes the entire perfusion sequence and self-determines the necessary parameters, eliminating the need for manual artery delineation and ensuring consistent results regardless of which user operates the system.
Solution Approach 2:
The patent replaces the manual mechanical process of artery delineation and AIF estimation with an automated neural network process. The CNN automatically identifies relevant features and computes quantitative maps without requiring user interaction, thereby eliminating inter-user variability and simplifying operation.
Data Source
Figure 1
Figure 2
Figure 3a
AI summary
The present invention relates to a method for processing an inputted perfusion sequence, by means of a convolutional neural network, CNN, the method being characterized in that it comprises the implementation, by a data processor (11b) of a second server (1b), of steps of: (b) extracting, using an encoder branch of the CNN, a plurality of initial n+1-dimensional features maps representative of the inputted perfusion sequence at different scales, n≥3, said CNN further comprising a decoder branch and skip connections between the encoder branch and the decoder branch, each skip connection projecting the initial n+1-dimensional features maps into initial n- dimensional feature maps; (c) generating, using said decoder branch of the CNN, a plurality of enriched n-dimensional feature maps also representative of representative of the inputted perfusion sequence at different scales, each enriched n-dimensional feature map incorporating the information from the initial n-dimensional feature maps of smaller or equal scale; (d) generating at least one quantitative map of the inputted perfusion sequence from the largest-scale enriched n-dimensional feature maps.