PhD offer: Artificial intelligence based 3-gamma PET image reconstruction

3-gamma medical imaging is a novel nuclear imaging modality that relies on the acquisition of 3 gamma photons

coming from a b+ and gamma emitter such as Scandium 44 (Sc-44). The development of this approach is

motivated by the possibility of utilising the information of the third gamma photon, which can help localise the

annihilation site; this information can improve overall image quality but also allow for potential dose reduction

and high speed dynamic imaging. Within the context of this imaging paradigm the development of new fast

image reconstruction techniques is crucial.

The objective of this thesis is to develop new image reconstruction techniques for 3-gamma PET imaging. This

imaging is based on the utilisation of the coincidence photons to define a line of response (LOR) and on the

detection of the emitted third photon to derive a Compton cone that can help in localising the annihilation

position on the LOR. Over the last few years artificial intelligence (AI) has been massively employed in medical

image analysis [1] and has more recently been also considered in the field of image reconstruction [2]. A potential

advantage of these methods include the incorporation of physical processes without deriving a precise physical

model and a reduced computational burden once the algorithm is trained. The hypothesis of this work is that AIbased

localisation of the third gamma will enhance both qualitative and quantitative accuracy of static and

dynamic PET imaging using the proposed approach.

Two distinct stages can be identified in the proposed thesis work. The first will focus on the implementation of a

reconstruction algorithm for 3-gamma imaging based on AI. Within this context two scenarios will be considered.

The first will be based on the use of AI approaches for estimating the coordinates of the cone-LOR intersection

within the "pseudo-TOF" approach we have previously proposed for 3-gamma reconstruction [3]. The alternative

will be the application of approaches such as classical convolutional neural networks (CNNs) and generative

adversarial networks (GANs) for direct 3-gamma reconstruction (direct transfer from raw data to 3D images).

The AI based approaches will be trained using simulated datasets based on validated scanner models already

available but also on future imaging system designs. The second part of the thesis will consist of a performance

comparison between the algorithms developed in the first step with those of (i). a "pseudo-TOF" algorithm as

well as with (ii). a standard iterative PET image reconstruction using the CASToR platform [4]. The evaluation

performed on acquired datasets will concern both performance in terms of sensitivity (according to the level of

activity injected) but also in terms of image quality (signal-to-noise ratio, spatial resolution) and computation

times. Finally, we will focus on the performance of the developed reconstruction approaches in the context of

dynamic imaging.

References:

[1] Geert Litjens et al ; A survey on deep learning in medical image analysis; Medical Image 42, 60-88, 2017

[2] Reader A et al; Deep Learning for PET image reconstruction; IEEE Transactions on Radiation and Plasma Medical Sciences, 5(1), 1-25, 2021

[3] Giovagnoli D et al ; A Pseudo-TOF Image Reconstruction Approach for Three-Gamma Small Animal Imaging; IEEE Transactions on Radiation

and Plasma Medical Sciences, 2020, doi: 10.1109/TRPMS.2020.3046409

[4] https://castor-project.org

Education: Master degree in physics, computer science, applied mathematics or equivalent and have a background / experience in deep learning / machine learning and image analysis, Machine learning/deep learning, image reconstruction, medical imaging

Skills: Python, R, C/C++

Languages: English, French optional

Application: Send before the 10th of May (at the latest) your CV, cover letter, grades/marks (Master, License/Bachelor) and a reference letter by e-mail to: visvikis@univ-brest.fr