Artificial Intelligence-Assisted Endoscopic Assessment of Response to Neoadjuvant Chemoradiotherapy in Esophageal Cancer: From Residual Disease Detection to Active Surveillance and Surgical Decision-Making

Neoadjuvant chemoradiotherapy (NCRT) followed by esophagectomy remains standard for locally advanced resectable esophageal cancer. Active surveillance defers or avoids surgery in patients who meet clinical complete response (cCR) criteria. What that judgment rests on is fragile: cCR is assembled from whichever modality was used, at whatever time point, under whatever follow-up intensity. cCR is a clinical classification, whereas pathologic complete response (pCR) is established on a resection specimen. Endoscopy carries most of the weight, and observer-dependent interpretation remains a concern. We searched PubMed, Embase, and the Cochrane Library from inception to 15 July 2026 and synthesized the findings narratively. Treatment settings differ enough to distort any pooled reading, so we stratified the evidence into four tiers—endoscopic AI in NCRT-treated cohorts, separated by pretreatment versus post-treatment image acquisition, conventional endoscopic and pathological response assessment, indirect AI evidence from neoadjuvant chemotherapy or immunotherapy, and other clinical and methodological context—rather than pooling it. Prediction model and diagnostic studies were independently appraised by two physicians using design-specific tools. Active surveillance achieved non-inferior two-year overall survival compared with immediate surgery in the SANO trial. preSINO measured how partial that reassurance is: bite-on-bite biopsy with endoscopic ultrasound-guided fine-needle aspiration yielded a negative predictive value of 68.7% for residual disease. Among the seven endoscopy-based modeling reports summarized, one used pretreatment images in an NCRT cohort, and six were indirect studies from non-radiotherapy or predominantly non-radiotherapy cohorts. None separately validated AI on post-NCRT images for current residual disease detection. Most are retrospective, single-center, static image studies conducted in chemotherapy rather than chemoradiotherapy cohorts; the largest enrolled 1041 patients across 46 centers and reported 80% sensitivity and 90% specificity for primary site pCR, not patient-level ypT0N0. Patient-level sensitivity, negative predictive value, calibration, and decision curve analysis are seldom reported. Endoscopic AI belongs inside composite response assessment, as a second reader—not as the thing that decides whether surgery can be omitted or surveillance relaxed. The most informative study still unperformed is not another AUC comparison; it is a prospective trial adding AI-targeted sampling to systematic bite-on-bite biopsy, with the patient-level false-negative rate for residual disease as the primary endpoint. AI has the potential to standardize image interpretation; prospective reader studies are needed to establish reproducibility in practice. Whether it makes active surveillance safer is not yet known.