Comparative Analysis of Pathology Foundation Models for Automated Detection of Tertiary Lymphoid Structures in H&E-Stained Digital Pathology Images

Abstract

Tertiary lymphoid structures (TLS) have been observed in solid tumors and have been associated with better outcomes in patients treated with immunotherapy, but their dynamic nature makes identifying TLS in clinical samples challenging. Recently, pathology foundation models have emerged as powerful tools in computational pathology. In this study, we aimed to develop a computational tool capable of identifying TLS across different cancer types. To this end, we utilized multiple pathology foundation models, along with the ImageNet-pretrained ResNet50 as a baseline, to identify TLS in pancreatic ductal adenocarcinoma (PDAC) and head and neck squamous cell carcinoma (HNSCC) using hematoxylin and eosin (H&E) stained images from The Cancer Genome Atlas (TCGA) and a licensed Real-World Evidence (RWE) cohort. Both pathologist-annotated and transcriptomic signature-based TLS outcomes were employed for performance assessment. Pathologist-identified TLS-positive tumors showed higher expression of TLS signatures in both diseases. Among the models tested, PLIP and CTransPath exhibited strong performance in identifying TLS within PDAC samples (AUC = 0.94 and 0.89, respectively). However, all models struggled in the analysis of HNSCC, likely due to the increased heterogeneity of the tumor microenvironment (TME). Despite their overall utility in detecting TLS in PDAC, all foundation models demonstrated poor performance in predicting transcriptomic signature-based outcomes in both PDAC and HNSCC. This suggests that TLS signatures may reflect broader or more transient aspects of TLS biology, whereas pathology-based assessment anchors on visible morphological features, more closely aligned with the focus of foundation model training. These findings highlight the potential of advanced pathology foundation models for TLS detection and broader tumor immune profiling tasks. These models can also be utilized on routinely collected patient biosamples with nominal costs. However, further refinement is needed to enhance their utility in tumors with more complex TME. Additionally, the identification of transcriptomics-based biomarkers from H&E images remains a significant challenge, despite advancements in digital pathology.