Anatomical and Physical Supervision for CT-less PET Attenuation Correction: BIC-MAC 2026 Challenge
This paper presents a submission to the BIC-MAC 2026 Challenge that enhances CT-less PET attenuation correction by combining anatomical supervision via a frozen TotalSegmentator extractor and physical supervision through differentiable attenuation projection losses within a pretrained nnU-Net framework.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of medical imaging, a machine called a PET scanner acts like a highly sensitive camera, capturing the glow of tiny radioactive tracers as they move through the body to reveal how organs are working. To turn these glowing signals into a clear, accurate map, the computer needs to know exactly how much the body's tissues block or weaken that light as it travels. Traditionally, doctors have used a quick X-ray scan, known as a CT, to create a detailed map of the body's density, which the computer then uses to correct the PET image. However, in situations where patients are already receiving high doses of radiation, or when the scanner is a combined PET and MRI machine that does not have an X-ray component, taking that extra X-ray is either impossible or undesirable. This creates a difficult puzzle: how can doctors get the necessary correction data without exposing the patient to more radiation or using a machine that isn't there?
Researchers have long tried to solve this by teaching computers to imagine what the missing X-ray map would look like, based only on the images the machine can already take, such as MRI scans. This process is called synthesizing a "pseudo-CT," a fake X-ray map that is good enough for the computer to do its job. A new study from a team at the National Technical University of Athens tackles this challenge by refining how these computers learn to imagine the missing map. Instead of just trying to make the fake image look like a real one, the researchers taught the computer to understand the actual physics of how the body blocks radiation, ensuring that the final medical images are not just visually convincing but scientifically accurate for diagnosis.
The team entered their work into a global competition called the BIC-MAC 2026 Challenge, which tested methods for creating these pseudo-CT maps using a mix of MRI, a special type of PET scan, and a 2D projection image known as a topogram. Their approach relied on a standard, robust computer architecture that had already been trained on a massive dataset of MRI-to-CT conversions. Rather than redesigning the entire computer brain, they focused on how they guided its learning. They introduced two distinct types of supervision to the training process. The first was anatomical supervision, where the computer was shown a guide that identified specific body parts like the skull, organs, and soft tissues. This helped the computer focus on the structures that matter most, ensuring it paid extra attention to the skull, which is critical for accurate brain imaging, and the boundaries between different tissues.
The second, and perhaps more innovative, layer was physical supervision. Here, the researchers did not just ask the computer to match pixel colors; they asked it to prove that its fake map would work in the real world. They took the computer's generated map and ran a mathematical simulation that mimicked how radiation travels through the body from many different angles. The computer was then penalized if the results of this simulation did not match the expected behavior of a real patient. This forced the model to learn the actual physical rules of attenuation—how much the body stops the signal—rather than just memorizing what an image should look like. By combining these anatomical guides with the laws of physics, the team created a system that learned to generate pseudo-CT maps that were both structurally sound and physically consistent.
The results of this method were tested against a set of real patient data to see how well the generated maps performed in practice. The computer produced maps where the density values were extremely close to the real X-ray scans, with a very small average error. More importantly, when these maps were used to reconstruct the final PET images, the measurements of how much tracer the body absorbed were highly accurate. The study found that the method worked particularly well for the brain, a region where even tiny errors in the skull map can lead to significant mistakes in diagnosis. The team reported a very low error rate for brain imaging, suggesting that their approach successfully captured the complex details of the skull without needing a real X-ray.
This work suggests that the best way to solve the problem of missing X-rays is not just to make better pictures, but to teach the computer the underlying physics of the body. By blending a deep understanding of anatomy with a strict adherence to physical laws, the researchers demonstrated that it is possible to create reliable, radiation-free corrections for PET scans. While the final test on a hidden set of data is still pending, the results so far indicate that this method offers a simple, reproducible, and highly effective path forward for medical imaging, allowing doctors to get precise diagnostic information without the need for additional radiation exposure.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.