Deep Learning Bias Removal for fMRI Signal Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Functional MRI (fMRI) signals are contaminated by non-neuronal biases such as regional partial volume effects and cortical orientation effects, making it challenging to model and remove these biases due to their nonlinear relationships, which complicates the inference of neuronal activity and mapping across the brain.
Innovation Solution
A method using a computer system that accesses fMRI signal data and bias characterization data, inputting them into a trained neural network to generate bias-reduced fMRI signal data, employing convolutional neural networks (CNNs) to predict anatomical and physiological features and reduce variance in fMRI signals.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional linear models are used to model fMRI biases, then the model is simple and easy to implement, but it cannot accurately capture the nonlinear relationships between biases and fMRI signals
Solution Approach 1:
The patent replaces traditional linear statistical models with a neural network-based computational system. The neural network uses multiple layers of nonlinear transformations to model the complex relationships between bias characteristics and fMRI signals, substituting conventional mechanical/statistical approaches with a more powerful computational paradigm that can capture nonlinear patterns.
Solution Approach 2:
The patent employs a nested architecture where a first neural network predicts bias characteristics (such as cortical thickness, cortical orientation, and vascular density) from anatomical images, and a second neural network uses these predicted bias characteristics along with fMRI signals to remove biases. This nested structure allows the system to progressively process information through multiple levels of abstraction.
2Reliability
If multiple runs of fMRI scans are performed to improve statistical power, then the reliability of results improves, but the scan time and cost increase
Solution Approach 1:
The patent performs preliminary processing by predicting bias characteristics from anatomical images before analyzing fMRI signals. By pre-characterizing the biases using structural information and neural network predictions, the system prepares correction factors in advance, allowing for more efficient single-run or fewer-run fMRI analysis that achieves comparable statistical power.
Solution Approach 2:
The patent introduces predicted bias characteristics (cortical thickness, cortical orientation, vascular density) as intermediary variables that mediate between anatomical images and fMRI signals. These intermediaries capture the sources of bias, allowing the system to correct for confounding effects without requiring multiple fMRI runs, thereby improving reliability while reducing scan time.
3Measurement precision
If fMRI signals are corrected for bias, then the accuracy of functional activation maps improves, but the processing complexity and computational requirements increase
Solution Approach 1:
The patent segments the bias removal process into distinct modular components: (1) a first neural network that predicts anatomical bias characteristics from structural images, and (2) a second neural network that uses these characteristics to correct fMRI signals. This segmentation allows each component to be optimized independently and facilitates integration with existing fMRI processing pipelines.
Solution Approach 2:
The patent replaces traditional manual or algorithmic bias correction methods with neural network-based systems that automatically learn and apply corrections. The neural networks substitute conventional signal processing approaches with data-driven models that can adapt to individual subject characteristics, improving accuracy while the modular architecture manages computational complexity.
Data Source
AI summary
Anatomical, physiological, instrumental, and other related biases are removed from functional magnetic resonance imaging (“fMRI”) signal data using deep learning algorithms and/or models, such as a neural network. Bias characterization data are used as an auxiliary input to the neural network. The bias characterization data can be subject-specific bias characterization data (e.g., cortical thickness maps, cortical orientation angle maps, vasculature maps), hardware-specific bias characterization data (e.g., coil sensitivity maps, coil transmission profiles), or both. The subject-specific bias characterization data can be extracted from the fMRI signal data using a second neural network. The bias-reduced fMRI signal data can include time-series signals, functional activation maps, functional connectivity maps, or combinations thereof.