Label-efficient computational tumour infiltrating lymphocyte assessment in breast cancer (ECTIL): multicentre...

The density of stromal tumour-infiltrating lymphocytes (TILs) is a prognostic factor for patients with triple-negative breast cancer and reflects their immune response. Computational TIL assessment has the potential to assist pathologists in this labour-intensive task, because it can be quick and reproducible. However, computational TIL assessment models heavily rely on detailed annotations and use complex deep learning pipelines that pose challenges for model iterations and clinical deployment. Here, we propose and validate a fundamentally simpler deep learning-based model that is trained in only 10 min on 100 times fewer pathologist annotations.

We collected whole slide images (WSIs) with TIL scores and clinical data of 2340 patients with breast cancer, including 790 patients with triple-negative breast cancer, from three cohorts in three countries (one each in the USA, UK, and Netherlands) and three randomised clinical trials in the Netherlands. Morphological features were extracted from WSIs using a pathology foundation model. Our model, label-efficient computational stromal TIL assessment (ECTIL), directly regresses the WSI TIL score from these features. We trained ECTIL on a single cohort from The Cancer Genome Atlas (n=356, ECTIL-TCGA), on only triple-negative breast cancer samples from four cohorts (n=400, ECTIL-TNBC), and on all molecular subtypes of five cohorts (n=1964, ECTIL-combined). We computed the concordance between ECTIL and the pathologist using the Pearson's correlation coefficient (r) and computed the area under the receiver operating characteristic curve (AUROC) using the pathologist TIL scores split into the clinically relevant TILs-high (≥30%) and TILs-low (<30%) groups. We also performed multivariate Cox regression analyses on the PARADIGM cohort with complete clinicopathological variables (n=384) to assess hazard ratios for overall survival, independent of clinicopathological factors.

ECTIL-TCGA showed concordance with the pathologist over five heterogeneous external cohorts (r=0·54-0·74, AUROC 0·80-0·94). ECTIL-TNBC showed a higher performance than ECTIL-TCGA on the PARADIGM cohort (r 0·64, AUROC 0·83 vs r 0·58, AUROC 0·80), and ECTIL-combined attained the highest concordance on an external test set (r 0·69, AUROC 0·85). Multivariate cox regression analyses indicated that every 10% increase of ECTIL-combined TIL scores was associated with improved overall survival (hazard ratio 0·85, 95% CI 0·77-0·93; p=0·0007), which was independent of clinicopathological variables and similar to the pathologist score (0·86, 0·81-0·92; p<0·0001).

In conclusion, our study showed that ECTIL could score TILs on haematoxylin and eosin-stained, formalin-fixed, paraffin-embedded WSIs in a single step, attaining high concordance with an expert pathologist. Without using deep learning-based segmentation and detection pipelines, ECTIL attained similar hazard ratios to the pathologist's score in an overall survival analysis independent of clinicopathological variables. In the future, such a computational TIL assessment model could be used to pre-screen patients for prospective de-escalation trials in patients with triple-negative breast cancer or as a tool to assist pathologists and clinicians in the diagnostic investigation of patients with breast cancer. Furthermore, our model is available online under an open-source licence, allowing translational researchers to validate and use ECTIL in future studies in breast or other cancers.

Dutch Cancer Society; Dutch Ministry of Health, Welfare and Sport; and Health∼Holland, Top Sector Life Sciences & Health.