March 2023
Song Xue is a biomedical engineer specialized in deep learning, conducting postdoctoral research at the AI for Translational Theranostics (AITT) group of the Department of Nuclear Medicine, Inselspital. Applying artificial intelligence to nuclear medicine, Song aims to reduce radiation in diagnostic PET imaging and to personalize dose prediction for radionuclide therapy. With new therapeutic fields just opening for nuclear medicine, Song feels that his research is at an exciting intersection of data science, clinical practice, and industry.
Song, what are you working on?
I am working on dose optimization in nuclear medicine for both diagnostic imaging and radioligand therapy with the help of artificial intelligence.
For the imaging part, the most widely used scanner is the PET/CT (positron emission tomography/computer tomography). I apply AI for image denoising or image quality recovery from low-dose imaging. A second part is attaining the same function as a CT via deep learning (attenuation and scatter correction) for CT-free PET imaging.
In radioligand therapy higher-dose injected tracer drugs destroy tumor cells. The problem is that today a standard dose is given for that which may be too high (harming other organs like the kidneys) or too low (not optimal results) for the patient. As this therapy form is newly accredited for prostate cancer, we have proposed a method for predicting the dose of radionuclide therapy based on pre-treatment information, enabling personalized treatment with greater precision.
Who will benefit from this?
With our research we aim to reduce adverse effects for the patients by minimizing the injected dose (as low as possible to the effective) and this way also try to make PET available for pediatric patients. But our AI-enhanced imaging also has advantages for hospitals as they need less time for image acquisition per patient, allowing them to treat more patients with the same resources.
Future work will focus on two aspects. On the one hand, we will develop a set of standardized guidelines for radionuclide therapy for prostate cancer and a software package for personalized treatment planning. On the other hand, we will develop a low-dose PET imaging system based on deep learning, which will enable high-performance noise reduction algorithms, CT-free PET imaging, and high-precision, high-sensitivity PET systems. These technologies will make PET imaging more convenient and feasible in scenarios such as routine medical screening.
Through domain knowledge we aim to improve robustness and generalizability of AI for nuclear medicine.
Do you see yourself as part of a bigger research community?
Definitely! I have previously worked on different imaging modalities as well as genomics data with the help of AI. So, I see myself as an AI developer for healthcare in general. To stay connected in this field, we have started a few collaborations here in Bern as this offers more inspiration from other groups using similar techniques but for other tasks. Maybe one group develops an innovative approach around AI that we could also use. It also helps to see that others face similar challenges in the algorithm development.
As we work in healthcare, it is very useful to understand how doctors think and how they would formulate a problem. This helps us to define our research tasks much better. In nuclear medicine, for example, we try to integrate the domain knowledge from physics into the design of the AI to improve its robustness and generalizability. We have the additional advantage that Switzerland is leading radionuclide prostate cancer therapy and thus has data from clinical trials available.
What motivated you to your field of research?
