Next generation quantitative magnetic resonance imaging for individualized radiotherapy
To fully utilize the precision with which modern radiotherapy can deliver dose, images depicting tumor biology through for instance cell density and blood flow are needed. Magnetic resonance imaging (MRI) can be used to create these types of images, however, currently these images often have large and unknown uncertainties. New solutions to both reduce and quantify uncertainties are therefore required.
Current models used in quantifying tumor properties from MRI data typically do not consider information about spatial structures. By including this as prior information in the models, the data can be used more efficiently with reduced uncertainties as a consequence. In the project we will explore different methods and models to reduce the uncertainties and in parallel develop methods enabling estimation of the remaining uncertainties.
Project Description A great concern with current quantitative MRI is that the resulting images often have large and unknown uncertainties, in particular when the imaging procedures need to be short. An important contributing factor to the uncertainties is that the models used only consider one pixel at the time. This implies that much information in the images is ignored. In this project we will improve on this aspect by developing models that also consider the spatial properties of the images. In this way, we will decrease the effective number of parameters that needs estimation and by that reduce the uncertainties. However, when the models become more advanced, estimating the remaining uncertainties becomes more challenging and new methods will therefore be developed for evaluated.
Next Step The currently dominating method for parameter estimation in quantitative MRI is based on pixelwise curve fitting. By shifting perspective and make use of the ever-growing computing power one can instead imagine creating and fitting models for entire 3D volumes. For instance, one can then include in the models that neighboring pixels are similar and that the images have hierarchical structure. This can be modelled using for instance Bayesian models and that deep neural networks have preference for images with hierarchical structure. In this project we will study how different types of structured models can be used to reduce the uncertainties. Simultaneously we will develop algorithms based of e.g. Monte Carlo techniques to get distributions of possible values rather than just a single value, and thus enable quantification of uncertainty.
Applications The methods developed in the project will primarily be evaluated on MR based perfusion and diffusion imaging relevant for radiotherapy targeting, for instance, prostate cancer. Even though the methods are developed with a focus on applications in radiotherapy, they are not limited to this. Quantitative imaging with reduced and known uncertainties have many applications both within and outside medicine.