A new AI algorithm is showing real promise in one of oncology's trickier tasks: figuring out how much viable tumor remains after pre-surgery chemoimmunotherapy in non-small cell lung cancer (NSCLC) patients. In a retrospective study spanning 135 patients and over 1,300 tissue slides across four international cohorts (Spain, China, and Germany), the AI matched expert pathologist assessments with 91% accuracy — a result that held up consistently across sites using different scanners and imaging setups.

The AI tool works by analyzing stained tissue slides to quantify tumor, necrosis, and stroma, generating a weighted score for residual viable tumor (%RVT). This score is used to classify whether a patient achieved a "major pathologic response" (MPR) — a key benchmark tied to treatment success. Discordant calls between AI and human experts occurred in fewer than 9% of cases, with most errors being false negatives.

By the Numbers:

• 91% pooled accuracy (95% CI, 85%–95%) for MPR classification
• 90% sensitivity and 93% specificity across all cohorts
• AUC ≥ 0.95 in all four international cohorts
• Only 12 of 135 cases had discordant classifications between AI and pathologists

Why it matters: Pathologic response assessment is increasingly critical as a surrogate endpoint in NSCLC treatment trials, but manual scoring is time-consuming and prone to inter-observer variability. An AI tool that's accurate, explainable, and generalizable across international settings could meaningfully standardize this process — though pathologist oversight remains essential for edge cases like complete response.