Developing deep learning-based AI for quantitative PET imaging

The problem at hand

Positron emission tomography (PET) imaging plays a vital role in detection, staging, and treatment response assessment of cancer.

Dynamic PET, in particular, visualizes the time-dependent uptake of an injected radiotracer. By application of a kinetic model, it allows quantification of the underlying biological process. Kinetic modelling requires knowledge of the time-dependent tracer concentration in blood, the so-called arterial input-function (AIF).

The gold-standard AIF is obtained by arterial blood sampling, which is invasive, laborious, time-consuming, and potentially painful, with risk for complications. Existing approximations that aim to overcome blood sampling suffer drawbacks that limit their usability.

Groups joining forces

Over the past years, in the Nuclear Medicine and Radiation Biology research group at UiT The Arctic University of Norway, researchers have gained significant knowledge and state-of-the-art infrastructure for performing cutting-edge PET research in humans and small-animals in the PET Imaging Center at the University Hospital of North Norway (UNN).

These possibilities were realized through The Norwegian Nuclear Medicine Consortium 180°N funded by Tromsø Research Foundation.

Head of the PET Imaging Center at UNN, Rune Sundset explains:

- The PET Imaging Center in Tromsø is the northern most PET imaging center in the world. It is physically connected to the hospital and physically connected to the university. This allows us to perform research from basic organic chemistry, all the way to research in human beings.

- In the basement we have a cyclotron and equipment for producing radiopharmaceuticals. On the ground floor we have clinical scanners for patient examinations and at the first floor we have equipment for preclinical studies.

Likewise, the UiT Machine Learning Group has over many years built a unique team of top competent researchers within artificial intelligence (AI) and deep learning for image analysis.

The group was recently acknowledged national status as Center for Research-based Innovation (SFI) by the Research Council of Norway, with the name SFI Visual Intelligence.

-As a machine learning group, our main competence is to develop the next generation AI methodology to solve real world problems of great value to society. Our collaboration with the PET Imaging Center gives unprecedented opportunities and is of strategic importance for us, explains Michael Kampffmeyer.

Jenssen follows up:

As a Norwegian centre for research-based innovation (senter for forskningsdrevet innovasjon – SFI), Visual Intelligence shall advance deep learning research for value creation by developing innovative solutions that impact people. I am very pleased with the innovation path we are pursuing for quantitative PET imaging, the competence we are building, and the results we are obtaining.  

In this favourable environment, researchers from both 180°N and SFI Visual Intelligence joint their efforts and invented a novel AI-based method that predicts the input-function from dynamic PET images, and which does not suffer the limitations of existing methods.

This deep learning-based input-function (DLIF) has shown promising results for a few tracers in a few animal models and in a human cohort, based on data obtained from collaborators.

These achievements were acknowledged at the 180°N Conference held in Tromsø in March 2022, when the abstract presented by Nils Thomas Doherty Midtbø, Ms student involved in the project, won the 10k NOK price for the Best Abstract Award:

- I am honored to be part of a project such as this, in which my thesis has strong tides with cutting-edge research which, moreover, can soon have a clear impact in real-world applications.

The method so far has been developed using limited available data (several tens of scans without blood samples). Larger amounts of data (several hundreds of scans with blood samples) are still necessary to improve the method further and to validate it in several settings and scenarios, however there is no available database to meet such requirements.

For this reason, UiT recently granted funding for an innovation project whose goal is to collect such data and, consequently, develop and eventually commercialize this novel AI-product that will significantly advance tumor imaging with PET.

The innovation project

This innovation project aims at collecting a large database of dynamic small-animal PET images with arterial blood sampling. This will allow to build commercial DLIF models based on AI which will significantly simplify quantitative tumor imaging with dynamic PET by evading of blood sampling.

With future translational research, the DLIF models could also be adopted for human PET imaging.

The innovation project is a unique opportunity. It requires top expertise in small-animal PET imaging for data collection, as well as for data analysis and machine learning, all available through the 180°N consortium.

In particular, the researcher hired for this project, Rodrigo Berzaghi explains:

- This opportunity means a lot to me, both personally and professionally. As a biologist, I have worked a lot with small animals, but carrying out arterial cannulation in them is a major challenge. Learning such technique will be a great achievement, which, together with PET imaging, will grant great visibility to our group, both nationally and internationally

Collecting such an unprecedented database for AIF prediction represents another great added value, as it may attract collaborators keen to perform state-of-the-art research which cannot be conducted elsewhere. Samuel Kuttner, medical physicist and researcher at the PET Imaging Center and the UiT Machine Learning Group/SFI Visual Intelligence, summarises:

- The task of performing simultaneous PET scanning and arterial blood sampling is very challenging, and very few centers in the world are able to perform this. We are actively collaborating with international experts in the field, for instance at Sherbrooke Molecular Imaging Center, University de Sherbrooke, Canada, where we will have exchange methodology and techniques for small-animal arterial blood sampling with simultaneous dynamic PET imaging.

Finally, an innovative product based on state-of-the-art AI methodology, developed at the SFI Visual Intelligence using the above mentioned data, is under way of being commersialized. Luigi Tommaso Luppino and Kristoffer Knutsen Wickstrøm, postdoctoral research fellow and Ph.D. student with the UiT Machine Learning Group/SFI Visual Intelligence, identify several challenges also from their point of view:

- The model we aim to come up with not only must perform adequately well, but it also must provide quantitative measures of its uncertainty and clear interpretations of its behaviour, namely by visualising which parts of the input lead the model to a certain output. These are the keys to gain the trust of doctors and medical practitioners in general, the final users of our proposed method. Thus, the development of new methodologies for interpretable AI (XAI), prominent topic in today’s AI-related research, is fundamental.