In the large, platform-based, independent validation study CATALINA, published by Dixon-Douglas et al in The Lancet Oncology, two artificial intelligence (AI)-derived computational tumor-infiltrating lymphocyte (cTIL) models, which were deployed without retraining or modification, provided statistically significant prognostic information and improved risk discrimination beyond clinicopathologic variables alone in patients with early-stage triple-negative breast cancer.
Tumor-infiltrating lymphocytes (TILs) are an established prognostic marker in triple-negative breast cancer. AI-based computational tools for TIL assessment could improve efficiency, but their prognostic value requires validation against clinical outcomes in independent patient cohorts.
“Although cTIL score did not incrementally improve prognostication compared with models combining clinicopathologic variables with sTIL [stromal TIL; scored by pathologists] score, [our] findings support the application of cTILs as a reproducible prognostic biomarker, particularly in settings where routine or widespread pathologist assessment is unavailable,” the investigators wrote.
Study Details
The investigators evaluated two separate cohorts:
- Analytical cohort: A total of 220 digitized hematoxylin-and-eosin (H&E) whole-slide images from patients with early-stage triple-negative or HER2-positive breast cancer that had previously been scored by trained pathologists.
- Clinical cohort: Digitized H&E whole-slide images and long-term outcome data from 1,759 patients (n = 1,356 with complete clinicopathologic, pathologist sTIL, and cTIL data) with early-stage triple-negative breast cancer pooled from seven prospective randomized adjuvant trials.
Two previously validated AI pipelines—AI-TIL and MuTILs—were independently deployed as locked models without retraining on the CATALINA slides, with algorithm developers masked to clinical data. Together, the pipelines generated five prespecified cTIL scores per slide, each quantifying lymphocytes relative to stromal or tumoral compartments.
The analytical cohort was used to assess correlations between cTIL scores and the mean pathologist sTIL score, whereas the clinical cohort was used to assess prognostic performance.
Key Findings
According to the investigators, the correlation between cTIL scores and the mean pathologist sTIL score was modest (Spearman’s correlation coefficient = 0.375–0.473).
Both sTIL and cTIL appeared to be independently associated with 5-year invasive disease–free survival, distant disease–free survival, and overall survival after adjustment for clinicopathologic factors, with hazard ratios of 0.73 (q < .0001), 0.70 (q < .0001), and 0.72 (q < .0001), respectively, for sTIL scores, and 0.80 (q < .0001), 0.77 (q < .0001), and 0.79 (q = .0002), respectively, for percentage_lymphocyte (an AI-TIL score representing lymphocytes as a proportion of all detected cells). In models adjusted for clinicopathologic variables and sTIL score, the prognostic association of cTIL score was no longer statistically significant, the investigators wrote.
Adding the sTIL score to standard clinicopathologic factors improved the model’s ability to distinguish between patients with better vs poorer 5-year invasive disease-free survival, distant disease-free survival, and overall survival; however, the investigators noted that adding cTIL score to clinicopathologic variables and sTIL score did not significantly further improve discrimination.
“This study affirms the prognostic value of TILs in patients with triple-negative breast cancer, whether scored by pathologists or computationally, using locked models employed entirely independently without modification or retraining on multisite data,” the investigators concluded.
They added, “The prognostic performance of five cTIL scores from two models…suggests the feasibility of widespread, multisite application of cTIL score….”
Sherene Loi, MD, PhD, of Peter MacCallum Cancer Centre, Melbourne, Australia, is the corresponding author of the article in The Lancet Oncology.
Disclosure: The study was funded by the Breast Cancer Research Foundation. For full disclosures of the study authors, visit thelancet.com.
