AI detects esophageal cancers radiologists miss on chest CT

Clinical takeaway: Esophageal cancer has no US screening pathway. Performance in this study was weakest for adenocarcinoma and junctional tumors, the US-dominant forms of the disease. Watch for opportunistic AI screening to reach existing lung cancer screening programs first. 

Esophageal cancer is usually caught only when swallowing becomes difficult. By then most tumors have outgrown effective treatment with five-year survival in the US at just 18.5%. No screening test currently exists to move diagnosis earlier. Endoscopy finds early disease but is too invasive and too scarcely used to cover asymptomatic populations, and less invasive candidates like blood tests and sponge cytology have fallen short on early-stage sensitivity, cost, or both. 

Chest CT is the obvious place to look: noncontrast scans account for roughly 40% of all CT examinations and already cover the full esophagus. The organ itself is the obstacle, a collapsed, motion-blurred tube in which early tumors blend into normal tissue. According to one analysis, the majority of patients who died of esophageal cancer had no suspicious esophageal findings documented on chest CTs from the preceding five years. Researchers trained a deep learning model, EAGLE, to find malignant and premalignant lesions on those same scans, then validated it across settings running from everyday hospital imaging to organized screening, including prospective deployment. The results make opportunistic detection look workable rather than theoretical. 

EAGLE detected 90.0% of esophageal cancers at 98.5% specificity in external testing. The soft spots showed early: sensitivity ran 52.5% for high-grade precancerous lesions and 60.1% for stage I cancers, and detection was weaker for adenocarcinoma and for tumors at the esophagogastric junction than for the squamous cell disease that dominated every cohort. 

The model outperformed every radiologist reading the same scans, and as an assist it raised their pooled sensitivity from 71.9% to 85.7% and their specificity from 79.6% to 91.7%, lifting residents toward specialist-level accuracy. In real-world hospital imaging the model found three malignancies the initial standard of care had missed, one of them 21 months before clinical diagnosis. Recalibration on those real-world scans cut false positives by 72.7% and more than doubled positive predictive value, from 25.3% to 55.0%, without losing sensitivity. 

Deployed prospectively in hospital imaging, the recalibrated model flagged 90 individuals, 38 of them confirmed malignancies, for a positive predictive value of 42.2% and an AI-triggered referral rate of 0.09%; one flagged cancer had escaped standard care entirely. On low-dose CT, the scan used in lung cancer screening, sensitivity held at 88.4%, and in a real-world low-dose screening cohort specificity reached 99.94%. 

EAGLE was trained on noncontrast chest CTs from 6,813 patients at two Chinese cancer centers and validated across 12 centers in three countries involving 80,612 patients, including 17 radiologists reading 300 scans and a prospective hospital deployment of 17,446 patients. Malignancy was confirmed by surgical or biopsy histopathology. 

Smoking raises risk for both lung and esophageal cancer, so the low-dose CT population the USPSTF already screens annually doubles as the esophageal high-risk group, and every scan images the whole esophagus. A model reading them adds a second cancer screen without new imaging, new visits, or new radiation. But the adenocarcinoma gap needs closing, and Western populations need larger prospective validation before any US deployment case can be made. 

"As an accurate, noninvasive and rapid tool, EAGLE demonstrates the feasibility of identifying esophageal malignant and precancerous lesions from NC CT [noncontrast CT] and LDCT [low-dose CT]," the authors conclude, adding that the model "has the potential to extend the clinical applicability of AI-based CT analysis beyond opportunistic screening settings to risk stratification in population-based prevention." 

Source: Zhou J, et al. (2026 Sep 22) Nat Med. Large-scale esophageal cancer screening through noncontrast computed tomography and artificial intelligence