Physics informed neural networks for medical imaging using diffusion MRI

Overview

Neural networks are a modern method for machines to learn informative mappings from input data to output data. Physics-informed neural networks supplement the traditional neural network architecture with structure or constraints derived from physical principles. This is known to greatly improve the rate and quality of learning when such physical knowledge is available. In the case of diffusion MRI, the physics knowledge comes in the form of model equations describing the physical process of diffusion. The objective of this project is to develop improved methods of deriving clinically-relevant parameters such as diffusivity from MRI data using physics-informed neural networks.

Research engagement

Literature review

Mathematical analysis

Computation and simulation

Analysing experimental data

Machine learning

Research activities

Review the mathematical and physical background of diffusion MRI

Review the principles of physics-informed neural networks

Generate simulated MRI data for training and testing purposes

Design and train physics informed neural networks using real and simulated data

Analyse the efficiency and accuracy of physics informed neural networks at extracting diffusion information from MRI data

Outcomes

The outcomes will be new physics informed neural network approaches for extracting diffusion data from MRI scans

Skills and experience

Coding experience in MATLAB, Julia, Python, or similar is required

A background in mathematics covering first-year calculus and linear algebra

No prior experience with neural networks or machine learning is required

With higher mathematical background, especially partial differential equations, the project can be tailored to use this.

Start date

1 November, 2024

End date

1 February, 2025

Location

QUT Gardens Point

Additional information

Extensive resources and assistance with the mathematical, simulation, and coding aspects will be provided.

Keywords

• diffusion MRI
• physics informed neural networks
• machine learning
• medical imaging
• PINN

Contact

Tim Moroney  t.moroney@qut.edu.au