AI in Molecular Imaging: The Vast Potential

The potential is vast; by applying AI algorithms to molecular imaging studies and integrating other data, we might be able to improve diagnostic accuracy. It’s even on the radar that AI-aided imaging techniques may soon prove more accurate than standard biopsies.

But we’re still in the early stages. Only one presentation focused on AI during the recent Molecular Imaging World Summit in Lausanne, Switzerland – the one I gave in tandem with Sven Zuehlsdorff, PhD, Senior Director, Research for Molecular Imaging at Siemens Healthineers. Yet, when the next summit takes place a few years from now, I expect at least half of the talks will feature AI.

The clinical application of AI in molecular imaging is one of great potential. For example, AI could help physicians identify uptake patterns of glucose-analog tracers in examinations for non-small cell lung cancer. In this type of disease, the presence of tracer uptake is a potential indicator of a metastasis, but sometimes there’s also distinct uptake in small lymph nodes. Even with standard uptake value (SUV) thresholds, how do we confidently resolve the uncertainty around lymph node uptake? The use of AI could help differentiate a metastasis from other potential causes of uptake.

Another example is in the routine diagnosis of lymphoma. In a whole-body scan, there may be lesions in both the lymph nodes and organs. It would be useful to have algorithms that could help segment and quantify those lesions and compare imaging examinations with previous ones: the comparative data could also help determine the success of the treatment.

“Medicine is a science of uncertainty and an art of probability,” said Sir William Osler, one of the founders of modern medicine in the late-19th and early-20th century. Over a century earlier, a pioneering statistician named Thomas Bayes established a theorem in medicine: a successful diagnosis depends in large part on the known pre-examination conditions.

Since a diagnosis is a guess at the probability of a particular reality, the Bayes theorem suggests that relevant pre-test elements be factored into the equation as much as possible. Instead of just looking at images, physicians should consider demographic data, such as age, and the results of earlier examinations.

If you want to predict the mortality rates of patients with lung cancer, several factors are important: smoking, the FEV1 number (for lung capacity), and even the tumor size. But most examiners do not take such data into consideration when they analyze images. Algorithms could help prompt doctors to look more closely at lesions in patients who have red-flag pre-test characteristics, for example.

The nuclear medicine and molecular imaging field has a unique advantage over other imaging modalities as it can focus and quantify results in ways that allow physicians to determine targeted responses to disease.

It is also important to peruse what I call “convergence engagement” among disciplines. Professionals from different disciplines should learn how to combine their knowledge and expertise, which includes the combination of data from disparate tests and other sources. AI could prove crucial in our efforts to bring together these different parameters and optimally leverage their additional value.