Brain feature prediction using geometric deep learning on graph representations of medical image data
Inventors
Besson, Pierre Alain • Bandt, Sarah Kathleen
Assignees
Publication Number
US-12354256-B2
Publication Date
2025-07-08
Expiration Date
2041-10-19
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Abstract
Described here are systems and method for predicting clinically relevant brain features using geometric deep learning techniques, such as may be implemented with graph convolutional neural networks or autoencoder networks that are applied to graph representations of brain surface morphology derived from medical images. As an example, graph convolutional neural networks can be applied to brain surface morphology data derived from magnetic resonance images (e.g., T1-weighted) using surface extraction techniques in order to predict brain feature data.
Core Innovation
The invention provides systems and methods for predicting clinically relevant brain features using geometric deep learning techniques, particularly graph convolutional neural networks and autoencoder networks applied to graph representations of brain surface morphology derived from medical images such as magnetic resonance images.
The method involves accessing medical image data depicting a subject's brain, generating brain surface data representing cortical and/or subcortical surfaces, converting these surfaces into graph representations composed of connected nodes and edges, inputting this graph data into a trained neural network, and generating brain feature data that indicate estimated brain characteristics such as age, sex, or disease state.
The problem addressed is the limitation of traditional convolutional neural networks, which require data on regular Euclidean grids and thereby cannot be directly applied to non-Euclidean data like surface meshes and connectivity graphs that are important in neuroimaging. This invention enables the use of CNN-like architectures on such non-Euclidean neuroimaging data to generate predictive brain feature data.
Claims Coverage
The patent includes multiple inventive features spanning methods of generating brain feature data from medical images using neural networks with graph representations.
Method for brain feature data generation using graph representations
A method comprising accessing medical images of a brain, generating brain surface data (cortical or subcortical), converting to graph representations of connected nodes, inputting this graph data to a trained neural network, and outputting brain feature data to a user.
Use of graph convolutional and autoencoder neural networks
Employing neural networks such as graph convolutional neural networks (including spectral variants) or graph-based autoencoder networks for processing the graph representation data to predict brain features.
Graph representation with three-dimensional mesh and spiral sequences
Utilizing three-dimensional mesh graph representations with spiral sequences for convolutional encoder modules, including spiral residual blocks.
Inclusion of various cortical and subcortical surfaces as brain surface data
Using inner cortical surfaces, outer cortical surfaces, subcortical surfaces, or combinations thereof as input brain surface data.
Generation and use of class activation maps for visualization
Generating class activation maps via gradient extraction at last convolutional layers, computing neuron importance weights, and overlaying activation maps on brain surfaces to visualize influential brain regions.
Prediction of neurological and cognitive features
Generating brain feature data indicative of Alzheimer’s disease, mild cognitive impairment, brain age (including accelerated brain age due to disease or radiotherapy), and fluid intelligence scores.
Definition of vertex features in the graph representation as registered native space coordinates
Using vertex features defined as Cartesian coordinates registered to native space in the graph representations.
The claims cover methods of using trained neural networks on graph representations derived from medical brain images to predict brain features, with inventive aspects including specific types of neural networks, graph representations, brain surface data types, visualization techniques, and predictive clinical and cognitive feature outputs.
Stated Advantages
The system provides an unbiased, data-driven approach that avoids manual feature selection.
It enables integration of cortical topology and morphometry through explicit modeling of brain surfaces, reducing dimensionality compared to volumetric approaches.
The method allows mapping and visualization of brain regions influential in predictions enhancing interpretability.
The techniques work with commonly available T1-weighted MRI data and can integrate connectivity analysis.
The approach brings clinical applications such as Alzheimer’s disease and stroke risk prediction to greater prominence.
Documented Applications
Biomedical research tool across a wide range of neurologic and psychiatric disorders including Alzheimer's disease, epilepsy, Parkinson's disease, multiple sclerosis, stroke, autism, ADHD, depression, anxiety, PTSD, brain tumor diagnosis and management.
Investigation and monitoring of remote brain region relationships in individuals.
Clinical applications including epilepsy patient management, Parkinson’s disease management, multiple sclerosis management, stroke risk and outcome prediction, traumatic brain injury outcome prediction.
Generation of brain feature data such as brain age scores for risk prediction and lifestyle modification recommendations.
Prediction of cognitive features such as fluid intelligence in individuals across the aging spectrum.
Application in whole brain radiotherapy patients to estimate brain age and assess neurocognitive toxicity.
Identification of electrophysiologically abnormal brain regions in epilepsy using single-subject analysis.
Evaluation of accelerated brain aging in diseases such as HIV infection.
Interested in licensing this patent?