Simultaneous PET-MR Modelling and Reconstruction for Imaging Brain Disorders

Lead Research Organisation: King's College London
Department Name: Imaging & Biomedical Engineering


Positron emission tomography (PET) combined with computed tomography (PET-CT) is now a very successful medical imaging technique. It has widespread application to the detection and monitoring of cancer throughout the human body. However, beyond this crucial application, PET is also a powerful method for investigating problems with the human brain. It allows us to understand what is going on in dementia, epilepsy and other distressing brain disorders which afflict in excess of one million people in the UK. One specific example is PET's ability to look at the brain's inflammatory response. This inflammatory response is now recognized as linked not only with dementia and many other disorders of the brain, but is even implicated in mental health problems such as depression. Only PET can allow us to look at this inflammation at the molecular level, which means that PET has a unique role both in diagnosing and seeking to remedy these often debilitating conditions. But there are limits on just how much information we can currently gain from PET, mainly due to the noise levels in the data which limit our ability to accurately measure quantities of interest within small regions of the brain.

Recently, with a view towards the future of medical imaging, PET scanning has been made available with magnetic resonance (MR) imaging in one single scanner, PET-MR. This means we can see the structure of the human body at the same time as seeing how it functions. These images come without the radiation exposure of a CT scan, possess the amazing details of MR, and in contrast to PET-CT all the medical information is obtained at the same time with the patient in the same position in one single and convenient scan. The potential of PET-MR is exciting, not only in terms of its benefits to patients and to research, but also for opening new possibilities for processing the medical imaging data. This is very relevant to PET, where problems with image noise have led many to invest considerable effort into reducing noise in the images, so that more detailed medical information can be gained. This current project seeks to go further by joining together the PET and the MR image reconstruction and processing. We will model what PET does, as well as what MR does, together, at the same time. Also, we will build high quality reference brain images which reflect a number of different functional and anatomical attributes of the human brain, based on PET and MR scans of different people. This will allow future PET-MR scans to benefit from previous knowledge, and so be more intelligent when scanning a new patient. The project will apply these special techniques to three main areas. First, to look at brain inflammation with PET, to gain insight into what is happening in people suffering from depression, especially when their medication is having no effect. Second, we will look at what is happening in the brain for people suffering from epilepsy. Finally we will also look at detecting the presence of a particular type of protein associated with Alzheimer's disease, which is highly important for both understanding and possibly resolving this urgently growing issue (the numbers of sufferers from Alzheimer's is predicted to double in the next 40 years).

The proposed developments for imaging brain disorders is very timely. Firstly, a brand new PET-MR scanner has just arrived at St. Thomas' Hospital, one of the first sites in the UK to obtain such a technologically advanced system. Secondly, brain imaging studies with this scanner will soon get underway during 2015, with a funded study already scheduled to scan 60 participants for research into mood disorders. The timing is therefore perfect for the parallel development of novel methodology suited to the unique capabilities of PET-MR. Though many have seen the potential for joining together the PET and MR data processing, none of the methods so far are as unified and advanced as put forward in this proposal.

Planned Impact

A) Three patient groups are poised to benefit:

1) Patients suffering from neurodegenerative disorders
Alzheimer's disease has a hidden but significant impact upon the UK. Advancing brain imaging capabilities, as proposed here by the powerful combination of PET and MR imaging hardware with innovative algorithms, will help understand progression of the disease from an early stage, providing scope for early remedial actions to be taken. Furthermore, given the key role of brain imaging in drug development, the improved functional and anatomical imaging capabilities should also help in the search for treatments for this inevitably increasing patient group, currently arising from an aging UK population.

2) Patients suffering from neuropsychiatric disorders
The proposed work will look into enhancing the image quality for a Wellcome Trust funded study into treatment-resistant depression, which will use the newly installed simultaneous PET-MR scanner at St. Thomas' hospital. Since the proposed methodology is expected to deliver improved images for analysis, this in turn will allow more information to be gathered from this study, ultimately assisting in finding answers and treatments for these more challenging depressive disorders.

3) Patients suffering from epilepsy
A further area of emphasis of this project is to generate improved images for understanding the role of the dopaminergic system in epilepsy. Epilepsy can seriously degrade quality of life and restrict normal human freedoms, and the methods to be developed in this work will ultimately also contribute to the better understanding of this disorder, and in turn, potential new ways of treating and resolving epilepsy.

B) Wider impact:

1) The pharmaceutical and medical imaging industries
The manufacturers of the new generation of medical PET-MR imaging systems (Siemens, GE and Philips), as well as pharmaceutical companies (e.g. Lundbeck UK, Johnson & Johnson) are poised to benefit from the proposed work. Pharmaceutical companies can exploit improved brain imaging hardware with the software developments of this present research proposal to deliver more accurate and precise analyses during trials of new drugs. The medical imaging industry will also benefit from the developments which can both justify and showcase the considerable investment into simultaneous PET-MR imaging technology. Recently the Division of Imaging Sciences and Biomedical Engineering was awarded EPSRC funding for a Centre of Doctoral Training (CDT), and some of these studentships have collaborative funding from Siemens, including projects fitting within the auspices of this present project proposal.

2) The UK population as a whole
The advanced imaging methods to be developed in this proposal should allow new insights into the function of the human brain as well as its disorders. Given that a significant percentage of the national population will likely suffer some form of neuropsychiatric problem at some stage in life, the outcomes of this research are pertinent to literally everyone. Ultimately, better understanding of neurological and psychiatric disorders contributes to enhancing quality of life, health and the therefore also the ultimate creative output of the UK population. These impacts cannot be underestimated.

3) Cultural impact and public awareness of science
Dr. Reader retains strong links with the Montreal Neurological Institute, who initiated projects such as "NEURO PORTRAIT: The Brain Inspires Us", an exhibit which brings neuroscience to the public through art. In a similar fashion, since this research concerns creation of a new generation of multi-parametric medical images, these could be exploited in a highly visual and artistic manner, seizing the imagination and inspiring the next generation of researchers. Prof. Hammers also has extensive experience in the area of the public awareness of science, which will be brought to bear on the results of this research project.


10 25 50
publication icon
Ovtchinnikov E (2020) SIRF: Synergistic Image Reconstruction Framework in Computer Physics Communications

publication icon
Mehranian A (2018) Synergistic PET and SENSE MR Image Reconstruction Using Joint Sparsity Regularization. in IEEE transactions on medical imaging

publication icon
Reader AJ (2020) Bootstrap-Optimised Regularised Image Reconstruction for Emission Tomography. in IEEE transactions on medical imaging

publication icon
Mehranian A (2020) Model-Based Deep Learning PET Image Reconstruction Using Forward-Backward Splitting Expectation-Maximization. in IEEE transactions on radiation and plasma medical sciences

publication icon
Bland J (2018) MR-Guided Kernel EM Reconstruction for Reduced Dose PET Imaging. in IEEE transactions on radiation and plasma medical sciences

publication icon
Ellis S (2018) Multi-Tracer Guided PET Image Reconstruction. in IEEE transactions on radiation and plasma medical sciences

publication icon
Belzunce MA (2019) Enhancement of Partial Volume Correction in MR-Guided PET Image Reconstruction by Using MRI Voxel Sizes. in IEEE transactions on radiation and plasma medical sciences

publication icon
Belzunce M (2018) Simulation and Design Considerations of a Dual Layer Plastic Scintillator Intraoperative Probe for Radiolabeled Tumours in IEEE Transactions on Radiation and Plasma Medical Sciences

publication icon
Da Costa-Luis CO (2021) Micro-Networks for Robust MR-Guided Low Count PET Imaging. in IEEE transactions on radiation and plasma medical sciences

publication icon
Bland J (2018) Spatially-Compact MR-Guided Kernel EM for PET Image Reconstruction. in IEEE transactions on radiation and plasma medical sciences

Description We have found that there is clear potential with simultaneous PET-MR imaging to improve positron emission tomography (PET) image quality by using magnetic resonance (MR) data, and furthermore also to improve MR images (when using highly undersampled data) using PET data. We have found that synergistic PET-MR reconstruction is therefore feasible, giving imaging results which would not be achievable if each modality were used in isolation.

Furthermore, use of MR data during PET image reconstruction can significantly improve image quality. The level of improvement is such that it can potentially be exploited to lower the amount of injected radioactivity (down to just 5%, or lower, of the normal dose), while still achieving standard diagnostic image quality at a much lower radiation dose to the patient or research participant.

The methods developed in the context of PET-MR reconstruction have also been found to be highly effective for multi-contrast or multi-parametric MR data. For example, high quality anatomical MR images can be used to notably reduce noise and improve spatial resolution for functional MR images, such as arterial spin labelling (ASL).

Furthermore, we have through this award discovered a novel solution to a decades-old problem of deciding how to compensate for noise in PET images in an objective way. The method is automatic and robust, delivering precisely-optimised image quality for each PET scan, without user intervention.

Finally, this award opened the way to our subsequent successful demonstration of the power of AI / deep learning to push PET image quality still further compared to the prior state of the art.
Exploitation Route These findings may open the way for reduced radiation dose in PET imaging, or possibly faster MRI data acquisition. In light of this, the case for using simultaneous PET-MR scanners, rather than PET-CT, becomes stronger.

These findings may also be put to use to enhance image quality for other types of MR parametric / quantitative images, which can be characterised by high noise and lower spatial resolution (e.g. fMRI).

The findings for automatic, robust and precise compensation for noise in PET images might find application for routine clinical imaging, to avoid empirically chosen or subjective selections of noise reduction levels.

Overall, with the launch into the use of deep learning and AI which opened up via this funding, the outcomes of this work will provide still further potential to reduce radiation doses in PET, and/or potentially shorten data acquisition times (for higher patient throughput and comfort).
Sectors Healthcare

Description Deep-learning PET-MR longitudinal reconstruction for lower-dose antibody-imaging in the understanding and treatment of cancer
Amount £25,413 (GBP)
Organisation GlaxoSmithKline (GSK) 
Sector Private
Country Global
Start 10/2020 
End 09/2022
Description Joint PET-MR image reconstruction with machine learning
Amount £81,000 (GBP)
Organisation Siemens Healthcare 
Sector Private
Country Germany
Start 04/2020 
End 03/2023
Title BrainWeb-based multimodal models of 20 normal brains 
Description A set of tools for working with the pre-existing "BrainWeb" dataset. Tools include downloading, image processing (such as re-segmentation based on alternative medical imaging modalities and noise levels, and reshaping based on scanner geometry, as well as image coregisration) & display functions (interactive comparison of multiple 3D volumes). 
Type Of Material Data handling & control 
Year Produced 2019 
Provided To Others? Yes  
Impact Data used in publications and workshops. 
Title High-Resolution Heterogeneous Digital PET Brain Phantom Based on the BigBrain Atlas 
Description In positron emission tomography (PET), the evaluation of image reconstruction algorithms needs realistic simulated data sets where the ground truth is known and the image quality and the quantification errors can be evaluated. In the context of brain imaging, qualitative and quantitative assessments of the radiotracer uptake in anatomical regions, such as the striatum or the cortical grey matter, are important to study brain disorders. Therefore, brain phantoms that emulate brain scans are then needed to assess the accuracy of reconstruction and post-processing algorithms. However, most of the available digital brain phantoms are usually of limited spatial resolution making them not ideal to evaluate quantification errors due to the partial volume effect (PVE). In addition, they are piece-wise constant and usually generated from segmented MRI images of the brain. As a result, quantitative errors can be underestimated when doing regularized MR-guided reconstructions. A realistic [18F]FDG digital phantom that overcomes the problems of the current PET digital brain phantoms, particularly for the simulation of simultaneous PET-MRI data sets, was created. The phantom data set consists of the PET phantom, a T1-weighted MRI image, a CT image and an attenuation map for 511 keV. They are available in three different resolutions. 
Type Of Material Database/Collection of data 
Year Produced 2018 
Provided To Others? Yes  
Impact This phantom was presented to the PET-MR community at the PSMR conference in Elba, Italy, in 2018. 
Description Aspects and embodiments relate to a method of creating an image representative of a measured dataset by iteratively updating a base image. The method comprises: generating a principal dataset from the measured dataset, the principal dataset comprising a dataset having noise at substantially the same level as the measured dataset. The method comprises: generating at least one additional dataset from the measured dataset, the additional dataset comprising: a dataset generated from the measured dataset such that each additional dataset has noise at substantially the same level as the measured dataset and such that each additional dataset is not identical to the principal dataset. The method comprises: processing the base image without noise compensation using the principal dataset and each additional dataset to obtain a principal interim image and at least one additional interim image respectively; comparing the principal and the at least one additional interim image to determine an indication of a level of noise present; and using the determined indication of noise present to select noise compensation to apply when processing the base image using the measured dataset to create a new base image representative of the measured dataset. Approaches described herein may be implemented for any standard iterative image reconstruction method and provide a method in which use of, for example, just one bootstrapped resampled dataset allows powerful state-of-the-art denoising algorithms to be directly embedded into image reconstruction with relative simplicity. Approaches provide a mechanism for data-dependent automatic and precise optimisation of the strength of any denoising at every update. 
IP Reference WO2020201755 
Protection Patent granted
Year Protection Granted 2020
Licensed No
Impact Options are currently being explored, no licensing has yet occurred.
Description Imaging Sciences Patient-Public Involvement (PPI) group 
Form Of Engagement Activity A formal working group, expert panel or dialogue
Part Of Official Scheme? No
Geographic Reach Local
Primary Audience Patients, carers and/or patient groups
Results and Impact Invitation to present research outcomes and plans to the Imaging Sciences PPI group in September 2018. The group consists of 7 regular members, who meet every two months, to review research plans for provision of feedback at key stages.
Year(s) Of Engagement Activity 2018