Comparative Analysis of Pathology Foundation Models for Automated Detection of Tertiary Lymphoid Structures in...

Tertiary lymphoid structures (TLSs) are observed in solid tumors and are associated with improved immunotherapy outcomes, yet their dynamic nature makes TLS identification in clinical samples challenging. Pathology foundation models have recently emerged as powerful tools in computational pathology. In this study, we developed a computational approach to identifying TLSs across cancer types using pathology foundation models, with ImageNet-pretrained ResNet50 as a baseline. Models were applied to hematoxylin and eosin images from pancreatic ductal adenocarcinoma (PDAC) and head and neck squamous cell carcinoma (HNSCC) from The Cancer Genome Atlas and a licensed Real-World Evidence cohort. TLS status was evaluated using both pathologist annotations and transcriptomic-signature-based labels. Pathologist-identified TLS-positive tumors showed higher TLS signature expression in both diseases. Among the models assessed, PLIP and CTransPath demonstrated strong performance in PDAC (area under the curve = 0.94 and 0.89), whereas all models struggled in HNSCC, likely reflecting greater tumor microenvironment heterogeneity. Despite effectively detecting TLSs in PDAC, foundation models performed poorly when predicting transcriptomic-signature-based outcomes in both PDAC and HNSCC. This discrepancy suggests that transcriptomic TLS signatures may capture broader or more transient biological processes, while pathology-based assessment reflects visible TLS morphology, more closely aligned with the features foundation models learn. Overall, these findings highlight the potential of pathology foundation models for TLS detection and immune microenvironment profiling using routinely collected biosamples. However, further refinement is needed to improve performance in tumors with complex tumor microenvironment and to enable reliable prediction of transcriptomic biomarkers directly from hematoxylin and eosin slides.