AI tools improve detection of key immune structures in melanoma

Researchers from the ECOG-ACRIN Cancer Research Group (ECOG-ACRIN) have applied AI-driven processes for detecting tertiary lymphoid structures (TLS) in thousands of digital images of melanoma tumor tissue, significantly enhancing TLS identification and survival predictions for operable stage III/IV patients. The presence of TLS, a key biomarker for better prognosis and improved survival, is not yet a standard part of patients' pathology reports, and manual detection is labor-intensive and can be variable. Lead investigators Ahmad A. Tarhini, MD, PhD, and Xuefeng Wang, PhD, will present the new approach at the American Association for Cancer Research 2025 Annual Meeting in Chicago.

Our efforts reveal the potential of open-source AI tools to transform how we predict survival and immunotherapy benefits by detecting critical immune structures like TLS with unprecedented ease and accuracy."

Dr. Ahmad A. Tarhini, professor and senior member, cutaneous oncology and immunology, at the Moffitt Cancer Center and Research Institute in Tampa, Florida

The study retrospectively analyzed thousands of archived digital images coupled with corresponding RNA sequencing data from 376 patients with advanced, high-risk melanoma, linking TLS presence to significantly better overall survival. The cohort had participated in a landmark US cooperative group trial led by ECOG-ACRIN called E1609 that tested immune check point blockade and cytokine therapy in high-risk melanoma (Tarhini A. J Clin Oncol. February 2020). 

This analysis found TLS present in 55% of the E1609 cohort and predicted significantly better overall survival than those without TLS (36.23% vs. 29.59% at 5 years), especially in those with more than one TLS (38.04% in >1 TLS vs. 28.65%). TLS density was also significantly prognostic for overall survival (37.77% vs. 28.72% at 5 years for median cutoff). Survival also varied by AJCC stage group, age, sex, treatment type, and tumor ulceration, as shown in AACR Abstract 3358.

"These findings highlight the potential for AI-driven approaches to standardize TLS assessment using low-cost H&E-stained images, with the potential to improve prognostication and stratification within AJCC, and warrant further investigation," said Dr. Tarhini.

Researchers first applied HookNet-TLS, an open-source deep learning algorithm, to measure TLS and germinal centers (GC) within the E1609 digitized H&E-stained slides. After reviewing the initial results, they retrained the model for better accuracy. They evaluated the prognostic value of TLS scores by correlating the presence of TLS and GC found in the digitized images with normalized TLS counts. 

Next, the researchers applied Gigapth Whole-Slide Foundation Model for Digital Pathology feature extraction and investigated the potential in TLS detection in this cohort. Gigapth allowed for enhanced visualization of H&E image tiles through the generation of principal component analysis (PCA).

"Utilizing Gigapth Foundation Model, the generated PCA visualizations appear promising in enhancing TLS and GC detection. These are undergoing further fine-tuning, and the final results will be shared at a future meeting," said Dr. Wang, chair of biostatistics and bioinformatics at Moffitt Cancer Center.

This research was supported by a grant from the National Cancer Institute, one of the National Institutes of Health.

"The new survival prediction methods leverage low-cost, easily accessible technologies. They have the potential to speed up TLS testing adoption for high-risk melanoma patients, aiding discussions with physicians on potential immunotherapy benefits," added Dr. Tarhini.