Diagnostic accuracy of an AI-based pathologic response assessment in locally advanced non-small cell lung cancer after neoadjuvant chemo-immunotherapy

Abstract Background Neoadjuvant chemo-immunotherapy has shown promise in improving survival outcomes for non-small cell lung cancer (NSCLC) patients, with pathologic response serving as a critical predictor of long-term outcomes. However, manual assessment of pathologic response is labor-intensive and subject to inter-observer variability. This study aimed to develop an automated AI-based solution to address these limitations. Methods We developed an AI-powered patch-based image analysis model to quantify residual viable tumor (RVT) in hematoxylin and eosin (H&E)-stained whole slide images. The model was evaluated on resected specimens from 47 NSCLC patients treated with neoadjuvant chemo-immunotherapy. The AI-derived estimates of RVT were compared with visual assessments by a board-certified pathologist. Statistical analysis included Pearson’s correlation for continuous tumor estimation and Cohen’s Kappa for concordance in major pathologic response (MPR) and pathologic complete response (pCR) classification. Results The AI model demonstrated a strong correlation with the pathologist’s continuous estimation of RVT (r = 0.77, p < 0.001, confidence interval [CI]: 0.73–0.81). In the assessment of clinical endpoints, the model achieved an 89.36% concordance rate for MPR (Kappa = 0.79, p < 0.001, CI: 0.61–0.96) and 89.36% concordance rate for pCR (Kappa = 0.56, p < 0.001, CI: 0.24–0.89) when compared with the board-certified pathologist. Conclusions Our AI-powered model demonstrates potential as a decision-support tool for pathologic response assessments in NSCLC patients treated with neoadjuvant chemo-immunotherapy.