Tertiary lymphoid structures (TLS) are ectopic lymphoid formations which are composed predominantly of B cells, T cells and dendritic cells. The presence in the tumor compartment of mature TLS - characterized by mature follicles containing germinal centers - has been strongly associated with improved survival upon cancer immunotherapies for patients with solid tumors. For this reason, TLS could be used as a predictive factor biomarker to identify the patients more likely to benefit from immune checkpoint inhibitors. However, the pathological assessment of the TLS status remains time-consuming and usually requires additional analysis including CD3/CD20 staining. This study aims to show that artificial intelligence techniques applied to digital pathology could offer a fast and cheap mass patient screening solution to assess TLS from Hematoxylin/Eosin (HE) digital pathology slides used in clinical workflow.
We analyzed a pan-cancer cohort of 289 HE-stained Whole Slide Images (WSI) from Institut Bergonié (43% NSCLC, 12% sarcoma, 45% other solid tumors) on which TLS were manually contoured by expert pathologists. A Deep Learning (DL) model was trained on WSI to predict TLS status at the patient level (presence or absence of TLS). Models were evaluated using five-fold cross validation. The best performing architecture included two main components: a predicted score of TLS presence on small areas (112x112 micrometers) of the WSI followed by an aggregation at the patient level. To assess the transferability of the model, a validation cohort (PEMBROSARC) of 236 sarcoma WSI was used.
The best performing model achieved a ROC AUC score of 0.917 (std of 0.036) in five-fold cross validation on the pan-cancer cohort (specificity of 68% for a sensitivity of 90%). This model transferred very well on the PEMBROSARC study, the first clinical trial implementing TLS status as an inclusion criteria (Italiano et al Nature Medicine 2022) with a ROC AUC score of 0.89 (specificity of 64% for a sensitivity of 90%).
Our study demonstrates the predictive power of DL applied to predict the patient's TLS status from HE images. This model could be implemented in pathology labs as an efficient pre-screening tool.
Institut Bergonie.
Owkin, Inc.
L. Gillet, J. Le Douget, B. Schmauch, C. Maussion: Financial Interests, Institutional, Full or part-time Employment: Owkin, Inc. All other authors have declared no conflicts of interest.