TOPLINE

An AI algorithm achieved 91% accuracy in assessing pathologic response (PR) to neoadjuvant chemoimmunotherapy in patients with non-small cell lung cancer (NSCLC), with performance consistent across four international cohorts. The AI-based percentage of residual viable tumour (%RVT) scoring showed at least moderate concordance with expert pathologist consensus, and discordant classifications occurred in fewer than 9% of cases.

METHODOLOGY

• Researchers conducted a retrospective study of four independent NSCLC cohorts treated with neoadjuvant chemoimmunotherapy followed by surgical resection; an internal cohort of 38 patients from Spain and three external cohorts from China, Germany, and Spain were included, with a total of 135 patients and 1344 haematoxylin and eosin-stained slides from primary tumours.
• For the internal cohort, two blinded pathologists independently scored %RVT, with a third pathologist adjudicating differences greater than 5 percentage points; a consensus meeting then adopted the more experienced reader's scores as the reference standard. External cohorts were each scored by a single pathologist.
• Slides were digitised using different scanners and magnifications across sites; the AI algorithm quantified tumour, necrosis, and stroma areas to generate weighted average percentages of RVT.
• The researchers analysed the concordance between AI-based and manual assessment of major PR (MPR), defined as less than or equal to 10% RVT in the primary tumour; performance was evaluated using accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC).

TAKEAWAY

• Concordance between AI-based and manual %RVT assessment was at least moderate across all cohorts, and the AI algorithm achieved a pooled overall accuracy of 91% (95% CI, 85%-95%) for MPR classification.
• The pooled sensitivity and specificity were 90% (95% CI, 82%-95%) and 93% (95% CI, 81%-99%), respectively, with an AUC of 0.95 or higher in all four cohorts.
• Interobserver agreement for manual %RVT scoring between pathologists of the internal cohort was excellent for both unweighted (intraclass correlation coefficient [ICC], 0.95; 95% CI, 0.91-0.99) and weighted (ICC, 0.96; 95% CI, 0.92-0.99) approaches.
• Discordant MPR classifications between manual and AI-based assessment occurred in 12 of 135 cases (8.9%), predominantly false negatives (nine cases) vs false positives (three cases), with main sources of error being stromal underrepresentation and misclassification of reactive pneumocytes and multinucleated giant cells.

IN PRACTICE

"We demonstrate that AI-based assessment of PR after neoadjuvant chemoimmunotherapy in NSCLC is consistent, explainable, and generalizable across independent international cohorts," the authors wrote. "As perioperative treatment strategies continue to evolve and the demand for reliable surrogate endpoints increases, AI-assisted quantification of %RVT offers a practical pathway toward greater standardization of PR assessment," they added.

SOURCE

The study was led by Daniel Curto, MD, and Susana Hernandez, PhD, Hospital Universitario 12 de Octubre, Molecular and Computational Pathology Group, Research Institute Hospital 12 de Octubre, Madrid, Spain. It was published online on September 05, 2026, in Lung Cancer.

LIMITATIONS

The study lacked definitive prognostic information due to short follow-up and small cohort sizes. The AI algorithm classified no case as complete PR (0% RVT), indicating that it cannot autonomously adjudicate this clinically important category and requires pathologist review for cases with low %RVT.

DISCLOSURES

This study was supported by grants from Fundación Mutua Madrileña, Instituto de Salud Carlos III, and Comunidad de Madrid iLUNG Programme. Hernandez reported receiving institutional grants from AstraZeneca, Roche, and Thermo Fisher; personal fees for lectures from Roche, AstraZeneca, Thermo Fisher, Lilly, Johnson & Johnson, and Pfizer; and support for travel from Roche, AstraZeneca, and Thermo Fisher, outside the submitted work. Multiple authors reported receiving grants, consulting fees, or honoraria or having other ties with various pharmaceutical companies. Full disclosures are reported in the original article.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.