At the recent 2026 Society of Nuclear Medicine & Molecular Imaging (SNMMI) Annual Meeting, the promise of artificial intelligence (AI) in nuclear medicine was highlighted in multiple sessions, with some presentations showcasing applications that may be particularly well suited to gastroenteropancreatic neuroendocrine tumors (GEP-NETs). In the future, AI may accelerate discovery and facilitate a more individualized theranostic treatment paradigm.

Following these presentations, featured expert Simron Singh, MD, MPH, FRCPC, was interviewed by Conference Reporter Editor-in-Chief Tom Iarocci, MD. Clinical perspectives from Dr Singh on these findings are presented here.

There is clearly a lot of excitement around AI, as reflected by the number of AI-related topics at this meeting, and there is great potential. I think that there were several main themes presented. One is the acceleration of the path to new drug discovery, which is happening in all areas of oncology. Another theme was improved efficiency in terms of workflow, such as in segmenting organs and regions of interest. A third was predictive and prognostic modeling. All of these are of particular interest as we aim to further advance patient care in neuroendocrine cancers. Although multiple presentations touched on the promise of AI and explored specific potential applications, I think that we are still generally at the hypothetical stage, which reflects how fast AI is moving and how we, as a society, do not have a good hold on it yet.

During his presentation at the 2026 SNMMI Annual Meeting, Amir Iravani, MD, FRACP, spoke to the potential of AI in general and to why it may be a good option and how it may provide benefit in NETs, in particular. AI could facilitate the use of radiomics to help us understand the heterogeneity of GEP-NETs on a new level, perhaps even in ways that we have not yet been able to in the laboratory or by analyzing currently available biomarkers. Additionally, AI predictive models may help us reduce the number of time points needed for dose-response modeling and dosimetry optimization, which could provide a more resource-feasible model of dosimetry on a large scale. That is, we could still do personalized PRRT, but we would not necessarily need to devote a huge amount of time and resources to perform multiple scans, at multiple time points, for dosimetry.

With respect to immuno-oncology, there is an opportunity for AI to accelerate new drug discovery, from a process that once took years to a process that perhaps may take a matter of months. And so, it should be no different in the world of radiopharmaceuticals, where AI might also facilitate patient selection, dosimetry optimization, outcome prediction, and/or the prioritization of strategies involving alpha, beta, tandem, or combination therapy, or perhaps alternative systemic therapies.

AI may have huge potential in helping us predict responses in the future. First author Sebastian Ignacio Salgado-Maldonado and colleagues described the rationale for their feasibility study at the 2026 SNMMI Annual Meeting (abstract 261246). They noted that, presently, we have relatively basic measures, such as the Krenning score, Ki-67, and some blood markers, but really there is not a huge panel of sophisticated markers available for predicting response. If, however, we could use AI to predict those lesions that may be less likely to respond to PRRT, we could then deploy other methods (eg, surgery, stereotactic ablative radiotherapy, or other forms of local ablative therapy) to deal with those lesions either pre-, post-, or during PRRT. That could obviously also help us resolve sequencing questions in individual patients.

When it comes to patient-facing issues, we need to be really cautious about overly depending on AI, recognizing that it has the potential to produce hallucinations and mistakes. Interestingly, where I am in Toronto, the auditor general of the province recently produced a report on AI scribes in the use of medical documentation, highlighting the amount of incorrect information that even the scribes took down. If you do not monitor AI closely, mistakes can enter the patient’s medical record. So, I do not think that we are at a point yet where we can depend on AI without having additional safeguards in place.

The potential predictive and prognostic applications of AI are very important, and the added efficiency of AI is also very important, but we need validation. Validation of response models, for example, is crucial because we need to be confident when we take AI-generated insights back to the patient. My conclusion is that we are not quite ready for prime time. While these potential applications are interesting and exciting, we would need much more data and validation before we start using them.