Michał Obara receives Best Paper Presentation Award

PET (positron emission tomography) is an imaging technique used, among other applications, in cancer diagnosis. A patient receives a small dose of a substance involved in metabolism, such as glucose, labelled with a radioactive isotope. Radioactive decay in the patient’s body produces pairs of photons that travel in opposite directions. The scanner detects these pairs, and an algorithm reconstructs an image from millions of such detections, showing where the substance accumulates, for example in a tumour.

"Some of the recorded pairs, however, are ‘false’: the photons have scattered along the way or originate from two independent decays. This background reduces image contrast, making small lesions harder to detect. Our research aims to develop machine-learning algorithms that can distinguish useful photon pairs from background events before the image is even formed," explains Michał Obara of the WUT Faculty of Electronics and Information Technology.

Our researchers are developing a method that assigns each detected photon pair a probability of being a true coincidence. In the award-winning paper, the researchers demonstrated two things. First, rather than discarding pairs classified as background, it is better to include each one with a weight equal to the probability predicted by the model. Second, it is possible to generate a model-uncertainty map showing where in the image the model is confident in its output and where the result should be interpreted more cautiously. Determining the model’s actual error requires knowing which photon pairs are truly coincident—information that is never available in a real examination. The uncertainty map developed aligns closely with the true error distribution, allowing regions of high model accuracy to be identified from the scan alone. Ultimately, the method is intended to be incorporated into standard PET image-reconstruction algorithms and compared with methods used clinically.

"We build a computer model of the scanner and virtual patients—realistic models of the human body—and then simulate PET examinations. This tells us which photon pairs are true coincidences and which constitute background, something that cannot be established from a real scan. We use these data to train and compare machine-learning models, integrate their outputs into the image-reconstruction process, and assess whether image quality improves,” explains Michał Obara. “We perform the calculations on supercomputers. Importantly, the method not only improves the image but also indicates where its output is less certain. This is particularly important in medical applications."

Modern long axial field-of-view (LAFOV) PET scanners have a field of view approximately one metre long and cover all the major organs at once. They are highly sensitive, but they also record more false coincidences: in the researchers’ simulations, background events account for approximately half of all recorded pairs. Current correction methods estimate the background statistically for the entire image at once and are computationally expensive. None of them evaluates individual photon pairs. A second challenge is trust in artificial-intelligence methods in a field as sensitive as medicine, which is why the researchers want not only to improve the image but also to indicate how certain the result is.

"The solution will be used primarily in medical diagnostics, including oncology, cardiology and neurology. Nuclear-medicine physicians could receive clearer images that facilitate diagnosis, as well as more reliable quantitative measurements. In the longer term, better background correction could help shorten scans or reduce the patient’s radiation dose. The solution may also be of interest to scanner-software developers. The method is not limited to PET, however: it could also be applied to other imaging techniques that detect pairs or triplets of photons. As part of the IMPET project, devoted to multiphoton positron emission tomography for industrial imaging, we will test whether it can be transferred to industrial applications, including PEPT, the tracking of isotope-labelled particles in technological processes. At this stage, our work is based on simulated data," says our researcher.

Work on the subject began before 2024, and in 2024 it became the topic of a doctoral project at the WUT Doctoral School. The award-winning paper marks the next step following the earlier stages, with completion of the doctorate planned for 2028. The work is divided into four stages: simulation and data preparation; building and training machine-learning models, including photon-pair weighting and uncertainty maps; incorporating the results into standard image-reconstruction algorithms and comparing them with current methods; and determining whether the method can be transferred to industrial applications. The first two stages have been completed, and the researchers are now moving on to the third.

Simulations account for most of the workload. Digitally reproducing an entire scanner and realistic patients demands substantial computing power and generates enormous datasets. Preparing data for the models is also time-consuming. Every photon pair must be assigned the correct class, while information about it—such as the location, time and energy at which each photon was detected—must be represented in a form the model can use effectively. Selecting the model architecture and tuning it, by repeatedly testing which settings yield the best results, is equally labour-intensive. Finally, rigorous evaluation requires close attention: a model may perform well on global metrics such as accuracy yet still exhibit local inaccuracies that affect lesion visibility. The results are therefore always verified at the image level.

Cooperation with NCBJ

At WUT, the work involves two researchers from the Institute of Electronic Systems at the Faculty of Electronics and Information Technology: Michał Obara, a doctoral student at the WUT Doctoral School, and his doctoral supervisor, prof. Konrad Jędrzejewski, PhD. The research is being conducted in cooperation with the National Centre for Nuclear Research (NCBJ) under the IMPET project, “Multiphoton Positron Emission Tomography for Industrial Imaging” (No. FENG.02.02-IP.05-0152/23), funded by the Foundation for Polish Science under the European Funds for a Modern Economy programme. Michał Obara is also employed by NCBJ on the project.

Konrad Klimaszewski, PhD, of NCBJ, a co-author of the award-winning paper, has made a particularly significant contribution to the research. The other co-authors are members of the IMPET project team at NCBJ—Mateusz Bała, Lech Raczyński, Roman Y. Shopa, Wojciech Zdeb and Wojciech Krzemień—and Beatrix C. Hiesmayr of IT:U Interdisciplinary Transformation University in Linz and the University of Vienna. The computations are carried out at the Świerk Computing Centre (NCBJ) and using PLGrid infrastructure at the Academic Computer Centre CYFRONET AGH.