Aim of the PhD Project:
Develop self-supervised AI for PET image reconstruction and analysis in order to optimise the clinically-important tasks of cancer diagnosis, treatment planning and treatment monitoring.
Project Description / Background:
Positron emission tomography (PET) is in widespread use for diagnosing and monitoring cancer, infection, neurological and cardiovascular diseases [1, 2]. PET images are used for assisting in disease diagnosis and lesion characterisation, cancer staging and treatment response. The images that are reported by clinicians use quantitative metrics, with particular reliance on standardised uptake values (SUVs) to provide reports that inform how patients are treated at various time points, given that more than one treatment pathway is often available. Importantly, PET scans are conducted during or after early stages of treatment to assess efficacy, and decide whether (or not) a treatment plan is effective. PET-guided therapy has become standard practice, continuing or even reducing treatment where effective or switching treatment if not. These are important decisions, as cancer therapies are often accompanied by side-effects which should be limited if at all possible, by avoiding unnecessary treatment, or avoiding inappropriate treatment escalation. On the other hand, discontinuation or pressing on with ineffective treatment can allow cancer to progress with detriment to patients who are sub-optimally treated. SUVs  are commonly used to characterise lesions, assess cancer aggressiveness and monitor treatment. SUVs are also used to measure uptake in reference regions such as the liver to establish whether patients have had a complete metabolic response. There is increasing interest in using SUV-derived metrics to measure metabolic tumour burden to develop risk-adapted (before treatment) plans. Most often the maximum SUV (SUVmax) or recently the ‘peak SUV’ is measured in a region of interest (ROI). However, PET images are reconstructed from relatively noisy, count-limited data, and it is necessary to compensate for noise by various empirical and heuristic methods. Given that the maximum value in a region will be very dependent on the noise level, or the amount of smoothing applied, it is without question that noisy images, or inappropriately smoothed images, can have a notable impact on these clinical metrics and therefore patient management. There is therefore a need for robust and objective ways to reconstruct PET images which compensate not only for noise, but also biases arising from the empirical or arbitrarily chosen levels of image smoothing, thereby mitigating against variability in SUV measures. We have recently proposed a new reconstruction methodology which exploits a simple statistical method known as bootstrapping . In previous work, bootstrapping has only been used to estimate the level of noise in PET images, which is a post hoc analysis of what has already happened in the reconstruction. The new method, in contrast, exploits bootstrapping during the iterative reconstruction process, seeking self-supervised optimal levels of noise reduction of the reconstruction in accordance with the noise level of the data. The outcome is a fixed reconstructed image with an appropriate level of detail, justified only by the measured data. Arbitrary smoothing is avoided, and robustness to differing levels of noise in the raw data has already been demonstrated. In effect, the newly proposed method is an early example of the rapidly growing area of self-supervised machine learning. We hypothesise that self-supervised regularised image reconstruction of clinical PET data will result in more robust SUV-metrics being available for improved clinical decision making for patient benefit. Beyond this, deep-learning enhanced versions of the method  will be explored by using self-supervised deep learning to optimise not just the strength, but also the type of prior information used in image reconstruction (drawing upon training data examples). Nonetheless, the version of the method without use of training data is potentially advantageous, with expected ready application across different clinical centres and different software platforms and scanners, through the method’s independence from training data (which tend to restrict utility to dataset sizes and types comparable to those used in the training corpus). Expected academic background of the candidate: mathematics, physics, and/or computational science.
Figure 1: An example of the impact of deep learning strategies on PET image reconstruction [5, 6]. OSEM and OSEM+PSF are conventional reconstruction methods compared to the two rightmost columns which embed deep learning into the reconstruction algorithm.
- Muehllehner, G. and Karp J.S., Positron emission tomography. Physics in Medicine and Biology, 2006. 51(13): p. R117-R137.
- Jones, T. and Rabiner, E. A., The development, past achievements, and future directions of brain PET. Journal of Cerebral Blood Flow and Metabolism, 2012. 32(7): p. 1426-1454.
- Kinahan, P.E. and Fletcher, J. W., Positron emission tomography-computed tomography standardized uptake values in clinical practice and assessing response to therapy. Semin Ultrasound CT MR, 2010. 31(6): p. 496-505.
- Reader, A.J. and Ellis, S., Bootstrap-Optimised Regularised Image Reconstruction for Emission Tomography. IEEE Trans Med Imaging, 2020.
- Reader, A. J., Corda, G., Mehranian, A., da Costa-Luis C. O., Ellis, S. and Schnabel, J. Deep Learning for PET Image Reconstruction, IEEE Trans Rad Plasma Medical Science, 2020
- Mehranian, A. and Reader, A. J. Model-Based Deep Learning PET Image Reconstruction Using Forward-Backward Splitting Expectation Maximisation, IEEE Trans Rad Plasma Medical Science, 2020