A new study published in JAMA Network Open has found that artificial intelligence (AI)–assisted assessment of tumour-infiltrating lymphocytes (TILs) in melanoma significantly improves reproducibility over traditional pathologist scoring and offers stronger prognostic value for disease-specific survival.
In this multi-institutional, multi-operator study, researchers analysed 60 hematoxylin and eosin–stained primary melanoma slides. A total of 98 participants were recruited to assess the samples, including 40 board-certified pathologists using standard visual TIL scoring criteria, and 58 others using an AI-assisted platform. The AI-assisted cohort included both pathologists and non-pathologist users, supported by a deep learning algorithm that quantified lymphocyte infiltration as a continuous percentage.
The study reported that the AI-based evaluation of TILs yielded excellent interobserver reproducibility, with intraclass correlation coefficients (ICCs) exceeding 0.90. In contrast, manual scoring using Clark’s method demonstrated only moderate reliability, with Kendall’s coefficient of concordance of approximately 0.44. These findings suggest that AI-driven methods may help standardise immune biomarker assessment in melanoma, overcoming long-standing issues of subjectivity and variability in manual scoring.
To evaluate prognostic performance, the researchers applied the AI-generated TIL scores to an independent test set of 111 patients with stage I to IV melanoma. Using the median AI-derived TIL score as a threshold, they found that higher TIL scores were significantly associated with better melanoma-specific survival. In univariable analysis, the hazard ratio was 0.45 (95% CI, 0.26–0.80; p = 0.005), and remained significant in multivariable models adjusted for stage and other clinicopathologic factors (HR 0.53; 95% CI, 0.29–0.97; p = 0.04).
The prognostic value of AI-derived TILs held even when using alternative thresholds, such as the 16.6th percentile cut-off, which also demonstrated a statistically significant association with survival outcomes. In contrast, the traditional pathologist-read sTIL scoring showed only borderline significance after adjustment (HR 0.54; 95% CI, 0.28–1.01; p = 0.06).
According to the authors, these results suggest that deep learning tools can offer not only more reproducible assessments but also more clinically meaningful insights into tumour immune responses. They argue that AI-assisted histopathology could be a valuable addition to diagnostic workflows, especially for biomarkers like TILs that are known to predict response to immunotherapy.
Nonetheless, the study authors note that this was a retrospective analysis, and prospective studies will be necessary to confirm the clinical utility of AI-assisted TIL scoring. Implementation into routine practice will also require careful integration into pathology workflows and further validation across diverse clinical settings.
This study supports the growing role of AI in improving the consistency and prognostic performance of histopathologic biomarkers and provides a potential path forward for more objective, reproducible immune profiling in melanoma.
Paper: Aung TNLiu MSu 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 Access online here.
