Design: Independent external validation using pooled data from 7 RCTs
Population: 1,356 women with early triple-negative breast cancer
Primary outcome: Association of AI and pathologist TIL scores with iDFS, DDFS and OS
Effect: iDFS HR 0.73 (pathologist) vs 0.80 (AI); AI not significant after adding pathologist score

CATALINA tested two locked, previously validated AI pipelines that score tumour-infiltrating lymphocytes on routine H&E slides, without retraining, in 1,356 women with early triple-negative breast cancer from seven randomised adjuvant trials.

AI scores correlated only modestly with pathologist scores (r 0.38–0.47). Both were independently prognostic: pathologist-scored stromal TILs had an HR for invasive disease-free survival of 0.73 (95% CI 0.66–0.82), and the best AI score 0.80 (0.73–0.89). Both improved 5-year discrimination over clinical factors alone, but once the pathologist score was in the model, the AI score added nothing significant.

The message is practical: where a trained pathologist reports TILs, AI does not improve prognosis. Where pathologist TIL scoring is not routinely available — common in many Indian centres — AI scoring could fill the gap.

Ask pathology to report stromal TILs on triple-negative breast cancer biopsies and resections.
Higher TILs indicate better prognosis in early triple-negative breast cancer.
Where pathologist TIL scoring is unavailable, AI-derived scores may be a reasonable substitute.
Do not use TIL scores alone to de-escalate treatment outside a trial.

Why it matters

It shows where AI pathology adds value — replacing a missing expert read — and where it does not.

The statistics, in plain English

A correlation of about 0.4 means AI and pathologist scores agree only moderately. A hazard ratio of 0.73 means each step up in TIL score was linked to lower risk of recurrence. The key test is whether AI adds information on top of the pathologist score — it did not.