The number of stromal tumor-infiltrating lymphocytes (TILs) is a significant prognostic indicator in triple-negative breast cancer and reflects the immune system's response to the tumor. Automated computational TIL evaluation helps pathologists with a tedious task and delivers fast, consistent results. Yet most existing computational models for TILs depend on extensive, detailed labeling and rely on elaborate deep learning systems, which hinder iterative improvements and real-world clinical application. We present and test a much simpler deep learning solution trained in merely 10 minutes on far fewer — specifically 100 times fewer — pathologist labels. Whole slide images (WSIs) paired with TIL scores and clinical details were obtained for 2340 breast cancer patients, 790 of whom had triple-negative disease. These came from three separate cohorts in three nations (USA, UK, Netherlands) plus three randomized clinical trials from the Netherlands. Feature extraction on the WSIs used a pathology foundation model to capture morphological information. The new model, termed label-efficient computational stromal TIL assessment (ECTIL), predicts the overall WSI TIL score straight from these features. Training variants included one cohort from The Cancer Genome Atlas (n = 356, ECTIL-TCGA), only triple-negative samples from four cohorts (n = 400, ECTIL-TNBC), and all molecular subtypes across five cohorts (n = 196464, ECTIL-combined). We quantified concordance with pathologist scores using Pearson's correlation coefficient (r). AUROC values were computed after splitting pathologist TIL scores into clinically relevant TILs-high (≥ 30%) versus TILs-low (< 30%) categories. Multivariate Cox regression in the PARADIGM cohort with full clinicopathological information (n = 384) estimated hazard ratios for overall survival, adjusted for other clinicopathological factors. ECTIL-TCGA-matched pathologist evaluations were well correlated with five heterogeneous external groups (r = 0.54–0.74; AUROC = 0.80–0.94). ECTIL-TNBC outperformed ECTIL-TCGA on the PARADIGM set (r = 0.64, AUROC = 0.83 compared with r = 0.58, AUROC = 0.80). ECTIL-combined delivered the strongest results on an independent external test collection (r = 0.69, AUROC = 0.85). In multivariate Cox models, each 10% rise in ECTIL-combined TIL scores was linked to better overall survival (hazard ratio 0.85, 95% CI: 0.77–0.93; P = 0.0007), remaining independent of clinicopathological factors and closely matching the pathologist-derived score (0.86, 0.81–0.92; P < 0.0001). This investigation shows that ECTIL successfully assigns TIL scores to hematoxylin and eosin-stained, formalin-fixed, paraffin-embedded WSIs in a single step and achieves high agreement with experienced pathologists. By avoiding deep learning segmentation or detection pipelines entirely, ECTIL still yielded survival hazard ratios equivalent to manual pathologist scoring in analyses independent of clinicopathological variables. Such computational TIL tools may eventually support patient pre-screening for future de-escalation trials in triple-negative breast cancer or help pathologists and doctors during breast cancer diagnostic assessments. The model is openly accessible online under an open-source license, allowing researchers to test and apply it in future work on breast cancer or other tumor types.
