Advancing AI in Pathology, Precision Medicine, and the Melanoma Landscape

A potential prognostic indicator in melanoma management is the presence of tumor-infiltrating lymphocytes (TILs). A higher density of TILs generally correlates with a better prognosis and response to certain therapies. However, a traditional method of quantifying these cells relies on a pathologist’s subjective, semi-quantitative assessment, which may have inter-observer variability that can impact clinical decision-making.

Seeking to bridge this gap, David Rimm, MD, PhD, spoke with CancerNetwork® about a novel machine-learning algorithm that may provide a more objective and precise method for TIL quantification. In a study that he and coauthors published in JAMA Network Open, investigators aimed to validate this AI-based approach against conventional methods. Rimm, Anthony N. Brady Professor of Pathology and professor of Medicine (Medical Oncology) at Yale University School of Medicine, elaborated on the rationale, key takeaways, and next steps of this research in melanoma.

Transcript:

The rationale for developing the algorithm was that it’s important to know whether a patient with melanoma is going to progress or what therapy they need. We have already known in the literature that when you estimate the number of lymphocytes that infiltrate the tumor, the more lymphocytes, the more likely the patient is to be cured by the current therapeutic approaches. The way we do that now is have a pathologist look at a slide and estimate by judgment how many lymphocytes there are and give a semi-quantitative approach to that diagnosis. That depends on the pathologist; it does not have [much] precision.

Maybe some pathologists are accurate, and other pathologists are less accurate. But the fact is that if you have multiple pathologists do the same section, you will not get the same answer. That’s the driving force behind it: to increase the precision in precision medicine. To increase the precision of the pathologist making that diagnosis. Now that we have tools that can find these cells and count them in a more accurate way than we did because we have AI and image analysis, we have the tools to do something that has not been done in the past. That’s what we did.

The reason [the algorithm] got into JAMA Network Open is because we did a lot of it. We had [many] pathologists, we looked at a lot of reproducibility, and we numerous non-pathologists operating a machine. [The study] showed that they had better precision than the pathologists who were reading the slide. Making the distinction of reading is what pathologists do [while] giving an expert opinion, and measurement is what machines do, given a numeric answer that is limited only by the accuracy and reproducibility of the machine. That was behind the effort.

The main conclusion moving forward is that the machines are more precise than the humans, but both the humans and the machines could be accurate. We did not try to compare accuracy because we did not have a big enough cohort for that. [The study] was retrospective; the tissues that we used were all retrospectively collected. To compare accuracy, we need to do a prospective trial, but we would not want to begin a prospective trial if we did not know if our machine worked or not, so we had to check our precision before we approached accuracy, which is our next step.

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