The 2026 ASCO Gastrointestinal (GI) Cancers Symposium spotlighted the current and potential future use of artificial intelligence (AI) in the diagnosis, prognostication, and treatment of GI cancers. Although AI has limitations, its role in supporting diagnosis and treatment continues to evolve, with opportunities for innovation in the management of gastroenteropancreatic neuroendocrine tumors (GEP-NETs).

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

There have been a few recent publications in the GEP-NET space looking at how we can potentially incorporate the use of AI. On the diagnostic side, studies have examined how we can improve the accuracy of pathological diagnosis, such as distinguishing between well-differentiated grade 3 NETs and poorly differentiated grade 3 neuroendocrine carcinomas. Even an experienced pathologist can struggle with this. The question that we have to answer when considering the use of AI in this space is: Who do you compare it against? Not all pathologists are equal in their diagnostic capabilities, but if we can standardize this with the incorporation of AI, I could see it being an incredibly helpful future tool.

Similarly, machine learning AI has been used to look at large data sets of imaging, including radiomics, to predict tumor grade and outcomes of therapy. Overall, I can see a lot of exciting possibilities for AI in this space, but I do wonder whether we are putting too much value on this at the expense of performing a classic histological assessment, which is also an important skill for clinicians to have.

One challenge with using AI in NETs is that some patients, especially those with well-differentiated NETs, do not have a lot of detectable mutations. Small bowel NETs also often have very low amounts of detectable molecular alterations, which make a lot of the current tests uninformative. For example, when we do germline genomic testing on patients with small bowel NETs, the test results almost always come back normal. These are also low-proliferative tumors and may not shed DNA into circulation. However, that is not to say that future methods will not be able to better detect low amounts of molecular alterations.

At the 2026 ASCO GI Cancers Symposium, Caroline Chung, MD, FRCPC, discussed the importance of context when considering the adoption of AI models in the clinic. And, to her point, computed tomography scans of the abdomen and pelvis can be drastically different from one institution to another, depending on the timing of the contrast administration and the acquisition of the images. Most pancreatic NETs also tend to be arterially hyperenhancing, so you may miss or undercall liver metastases if you do not do an arterial phase computed tomography scan. Moreover, if you are feeding these highly variable images into an AI model, that is where it can break down. I am also concerned about AI’s generalizability. Perhaps your AI model works great on your scanner at your institution, but can it be used at another institution or with other techniques?

Regarding the role of AI in imaging analysis, Yuichi Mori, MD, PhD, talked about AI applications in endoscopy, specifically the concerning issue of deskilling among endoscopists, during a presentation at this year’s ASCO symposium. To me, there is no question that AI will improve or enhance our diagnostic capabilities, but, to Dr Mori’s point, if you are at a center that does not have the equipment or your equipment does not work that day, are you still able to do the diagnostic procedure? I am personally approaching this with both caution and enthusiasm. I think that these issues are something we can work around.

I also think that most clinicians are willing to adopt AI, and, in many ways, non-NET oncology is much further along with this. I do some work with cancers of unknown primary, and we are starting to apply these tumor-of-origin classifiers, where we use the genomic signature of a tumor to map against the genomic signatures of known tumors, such as colon, breast, and lung cancers. It is not perfect, but, with this, we can predict the origin of the tumors with surprising precision.

Also at the 2026 ASCO GI Cancers Symposium, William A. Hall, MD, discussed the importance of embracing commercially developed genomic signatures in clinical trials. I am an enthusiastic supporter of extensive genomic testing. Unfortunately, so far in GEP-NETs, the utility of genomic testing seems to be mostly limited to the high-grade tumors and, especially, the neuroendocrine carcinomas.

Overall, I think that AI needs to be embraced in neuroendocrinology. We have a lot of imaging and genomics data from large clinical trials that we could explore in terms of treatment outcomes, toxicity, and things of that nature. I think that we are at a point where we can start combining all this information to try to improve therapy selection and even to select surveillance strategies. There are a few tests that are looking at this (eg, the NETest), but, so far, I do not think that any of them have lived up to our expectations. We just do not have clinical data supporting their use.