Refining Artificial Intelligence Tools in Cancer Pathology and Research

A study assessing a machine-learning algorithm for quantifying tumor-infiltrating lymphocytes in melanoma was recently published in JAMA Network Open. Findings demonstrated that AI may overcome the subjectivity associated with traditional, pathologist-based assessments, offering a more reproducible and objective method for a critical prognostic biomarker.

Based on these results, David Rimm, MD, PhD, provided his perspective on the trajectory of AI in oncology, discussing other foreseeable roles for AI in cancer research and management. He outlined how AI may have an impact on analyzing cancer tissue and determining genomic alterations, although some tools may be more accurate than others. He also touched upon the use of AI as an evidence-based clinical decision support tool, noting how these technologies are already beginning to appear in daily practice.

Rimm is the Anthony N. Brady Professor of Pathology and a professor of Medicine (Medical Oncology) at Yale University School of Medicine.

Transcript:

Everyone’s enthusiastic about AI being able to do everything; I think that’s an overstatement. However, AI can do some things well, and those things will save pathologists effort. There’s no reason for a pathologist to look at normal tissue and say that it’s normal if a machine can do that as well and show anything that might not be normal to the pathologist. Then, the pathologist makes the diagnosis of abnormal. That’s an application of AI that has already been FDA-approved and is likely to see increased usage because it saves pathologists work, money, and time. All those things are [quite] valuable. There are other things that people are working on AI for; for example, using AI-based image assessment to determine genomic alterations. Some of those might work, and some do not. That’s an area that is much less likely to make it into the clinic in the near term because not all of them work. It might be that you can tell a BRAF mutation with 95% sensitivity and specificity, but a P53 mutation can only be assessed with 60% sensitivity and specificity. You are not going to replace sequencing with AI image analysis if it only can do 60%, and even 95% might not be good enough in some cases, although 95% to 98% is usually the threshold for making it into the clinic. Time will tell if that’s an AI application.

AI has also been applied in many other places, and we will probably have other roles in medicine and radiology in transcribing notes for clinicians. I know when I went to visit my doctor, we were talking about some symptom, and he said, "I don’t think those 2 are related. Let’s check." He went to this evidence-based AI machine that was right there, and he asked, "Is there any association between this symptom and this drug?" It could check all the literature using AI quickly, and it found none. That’s the use of an AI by a PCP, a personal physician. That kind of AI will be common if it is not already.

Reference

Aung TN, Liu M, Su D, et al. Pathologist-read vs AI-driven assessment of tumor-infiltrating lymphocytes in melanoma. JAMA Netw Open. 2025;8(7):e2518906. doi:10.1001/jamanetworkopen.2025.18906