AI outperforms pathologists in quantifying tumour-infiltrating lymphocytes in melanoma

A multinational study led by Yale School of Medicine and the Karolinska Institute in Sweden has found that artificial intelligence (AI)–driven assessment of tumour-infiltrating lymphocytes (TILs) in melanoma tissue significantly outperforms traditional pathology in both reproducibility and prognostic validity.

Published in JAMA Network Open, the study illustrates the potential of open-source AI tools to standardize and refine clinical workflow in dermatopathology, marking a substantial advance in melanoma management.

A critical biomarker in melanoma, TILs reflect the immune system’s response and correlate with improved patient outcomes. Accurate quantification can influence diagnosis, prognosis, and therapeutic decisions, particularly concerning immunotherapy. However, melanoma diagnosis has significant interobserver variability when pathologists visually assess TILs on hematoxylin and eosin–stained slides, impeding consistent clinical decision-making.

“Our findings suggest that an AI-driven lymphocyte quantification tool may provide consistent, reliable assessments with a strong potential for clinical use, offering a robust alternative to traditional methods,” lead author Thazin Nwe Aung, PhD, associate research scientist in pathology at Yale School of Medicine said in a press release.

The study enrolled 98 participants from 45 institutions globally. Forty pathologists assessed 60 melanoma tissue sections using conventional methods. The AI-assisted arm, comprising 11 pathologists and 47 non-pathologist scientists, utilized a machine learning algorithm to quantify TILs. The AI approach demonstrated intraclass correlation coefficients exceeding 0.90 for all TIL variables, substantially outperforming manual reading, which achieved an ICC of 0.61 for stromal TILs and a Kendall W value of 0.44 for manual Clark scoring.

Most critically, AI-based TIL scores also correlated with patient outcomes. Median cutoff–based stratification yielded a hazard ratio of 0.45 (95% CI, 0.26–0.80; p=0.005), while the 16.6 cutoff approach produced a hazard ratio of 0.56 (95% CI, 0.32–0.98; p=0.04), underscoring the potential prognostic utility of an AI system.

Despite its retrospective design, the availability of an open-source dataset and algorithm provides a platform for validation and future integration into clinical workflows, the researchers reported.

“I’m especially proud that 15 Yale School of Medicine faculty and staff contributed to this work. It’s a great example of how the Department of Pathology at Yale is leading the way in AI-driven pathology research,” Dr. Aung said.