Abstract PS3-04-26: Reproducibility and Robustness of an AI-based Stromal Tumor-Infiltrating Lymphocyte Pipeline in Triple-Negative Breast Cancer

Abstract Introduction: Triple-negative breast cancer (TNBC) has limited treatment options and significant treatment-related toxicity. Recent studies have shown that quantifying stromal tumor-infiltrating lymphocytes (sTILs) can aid in identifying early-stage TNBC patients likely to achieve a pathological complete response to neoadjuvant chemotherapy (Abuhadra et al., 2023), suggesting its utility in treatment stratification. However, manual assessment of sTILs is prone to interobserver variability (Van Bockstal et al., 2021). Artificial intelligence (AI) promises a more consistent alternative, but its reproducibility and robustness to input variability remains insufficiently evaluated, limiting its clinical adoption (Park et al., 2020). Methods: We aimed to validate the reproducibility and robustness of an in-house AI-sTIL pipeline (AbdulJabbar et al., 2020; Pan et al., 2025) developed using H&E-stained slides from the ARTEMIS trial (NCT02276443), scanned between 2016 and 2024 on a Leica Aperio AT2 scanner at 20x magnification. We first compared AI-generated sTIL scores with two independent sets of manual sTILs (n=201) provided by pathologists according to the International TIL Working Group guidelines. Moreover, to establish deployment readiness, we examined the effects of input variability, including scanner type, scanning magnification, operator, and slide aging by rescanning up to 150 slides and conducting controlled, pairwise comparisons. Rescanned images were generated at 20x and 40x magnifications using Aperio AT2 and Hamamatsu NanoZoomer scanners, and at 40x using Aperio GT450. To assess operator viability, the scanning was run independently by different operators with AT2 (20x and 40x) and GT450 (40x). Results: Manual sTILs showed strong concordance between pathologists (ρ=0.884, R2=0.855). Comparing this agreement to the inter-operator reproducibility of the pipeline (ρ=0.980-0.994, R2=0.978-0.984) suggests AI-sTIL is robust to typical user handling variations in clinical workflows and offers a more consistent assessment. The pipeline correlated well with individual manual sTILs (ρ=0.768, R2=0.502 for pathologist1, and ρ=0.765, R2=0.579 for pathologist2) and even more strongly with the average of the two sets of manual sTILs (ρ=0.793, R2=0.583), indicating the pipeline may approximate a consensus interpretation while minimizing individual variability. Scores remained highly consistent across scanner models (AT2 vs. GT450: ρ=0.972, R2= 0.941; AT2 vs. Hamamatsu: ρ=0.987, R2=0.975), though a systematic bias toward higher scores for AT2 (regression slopes of 0.853 and 0.885, when compared with GT450 and Hamamatsu scores) suggests calibration may be necessary to enable accurate cross-platform comparisons. Similarly, scores were consistent across scanning magnification (20x vs. 40x: AT2: ρ=0.984, R2=0.964; Hamamatsu: ρ=0.992, R2=0.979). Interestingly, we observed lower scores for rescanned images of older slides (median slide age: 6.3 years), likely due to stain fading. However, this is unlikely to impact clinical deployment, as such delays between tissue sectioning and scanning are rare in routine clinical practice. The pipeline demonstrated excellent inter-run consistency with 100% concordance (ρ=0.991, R2=0.988), confirming the repeatability and stability of the pipeline. Conclusion: This study highlights the reproducibility and robustness of our in-house developed AI-sTIL pipeline across common variations in pathology workflows. Through controlled data acquisition experiments, we identified the appropriate data acquisition setup to pair with the pipeline. Our approach provides a practical framework for institutions seeking to assess deployment readiness of pathology AI tools before formal validation and clinical integration.</abstract> Citation Format: S. Ranjbar, A. San Lucas, P. Chen, K. Shah, M. Suchko, B. Guerrouahen, C. Ercan, E. Barrientos Toro, K. Sweeney, X. Pan, J. DeLa Cruz, B. Rodriguez, N. Reddy, C. Yam, G. Mantha, M. Riben, L. Huo, Y. Yuan. Reproducibility and Robustness of an AI-based Stromal Tumor-Infiltrating Lymphocyte Pipeline in Triple-Negative Breast Cancer [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS3-04-26.