Esophageal cancer (EC) remains a major global health problem characterized by high mortality due to frequent late‑stage diagnosis and the absence of widely implemented, noninvasive screening programmes. The disease includes two main histological subtypes—esophageal squamous cell carcinoma (ESCC) and esophageal adenocarcinoma (EAC)—which differ in risk factors, anatomic distribution and geography. Reported 5‑year overall survival rates cited in the source are 36.9% in China and 18.5% in the United States, underscoring the need for earlier detection.
Although population‑based endoscopic screening has contributed to earlier detection in some high‑risk regions, endoscopy is resource intensive. Noncontrast computed tomography (NC CT) is widely available but has been considered unsuitable historically for early EC detection because the esophagus is a hollow, collapsible tube subject to motion and other artefacts. These factors make small precancerous or early malignant lesions difficult to visualise and distinguish from normal tissue on standard CT.
To address the lack of accurate, scalable noninvasive screening methods, the authors developed the Esophageal AI‑Guided malignant Lesion Evaluation (EAGLE) model to detect precancerous lesions and cancers from chest NC CT. According to the source, EAGLE was trained on imaging from 6,813 patients collected at two centres. The provided excerpt does not include detailed information about the model architecture, training procedures, input preprocessing, or patient demographic distributions; those methodological details were not reported in the source excerpt.
EAGLE was validated across a large multicentre dataset spanning 12 centres in three countries, totalling 80,612 patients across opportunistic and population‑based settings. For opportunistic screening of existing chest CT scans, multicentre external test cohorts from eight centres (n = 11,466) produced a specificity of 98.5%. Sensitivity reported in these cohorts was 90.0% for cancer and 52.5% for precancerous lesions. These results indicate that the model prioritised a low false‑positive rate while maintaining high cancer detection sensitivity in opportunistic settings.
The authors evaluated EAGLE on low‑dose CT (LDCT) data from two centres (n = 1,607) and report comparable performance to standard NC CT validation. This suggests the potential for opportunistic EC detection when LDCT is performed for lung‑cancer screening programmes, although the excerpt does not provide granular LDCT performance metrics beyond the statement of comparable results.
Calibration of the model in a real‑world cohort from three centres (n = 35,402) substantially reduced false‑positive findings: the reported reduction was 72.7% while preserving sensitivity. The source does not elaborate in the excerpt on the specific calibration approach, thresholds adjusted, or the impact on downstream clinical workflows; those procedural details were not reported in the provided text.
In a prospective hospital validation cohort (n = 17,446), EAGLE achieved a positive predictive value (PPV) of 42.2%. In a separate real‑world LDCT screening cohort (n = 10,959), the model reached 99.94% specificity. These reported results combine to indicate strong specificity in large real‑world populations and a meaningful PPV in a prospective hospital setting. The excerpt does not provide additional metrics such as negative predictive value, per‑center variability, or confidence intervals.
EAGLE also demonstrated capacity to detect earlier disease stages. In paired CT–endoscopy cohorts from two centres (n = 702), operating at a higher‑sensitivity threshold, reported sensitivities were 65.0% for precancerous lesions and 78.4% for stage I EC. These findings indicate that the model can identify a proportion of lesions that would traditionally require endoscopic visualisation or targeted biopsy, although sensitivity for precancerous lesions remains lower than for invasive cancer.
Exploratory analyses in a prospectively enrolled cohort suggest that using EAGLE to flag high‑risk individuals for endoscopic referral could improve screening efficiency by concentrating invasive diagnostic resources on a smaller, higher‑risk group. The source excerpt does not provide protocol details, referral thresholds, or outcomes of such a pathway, and those operational specifics were not reported here.
The authors conclude that EAGLE has potential as a scalable tool for early EC screening, capable of operating on widely available NC CT and on LDCT used in lung screening. Key reported strengths include large multicentre validation (80,612 patients), high specificity (up to 98.5% and 99.94% in reported cohorts), high sensitivity for cancer (90.0% in opportunistic external cohorts), and measurable detection of precancerous and stage I disease in paired CT–endoscopy cohorts.
Limitations in the provided excerpt include lack of methodological detail about the AI model, absence of full demographic and per‑center performance breakdowns, and limited description of calibration methods and clinical implementation pathways. The authors registered the study on Chictr.org.cn (ChiCTR2300074806). Additional methodological and contextual detail would be needed for independent evaluation and clinical adoption.
EAGLE shows promise as an AI‑based approach to extract esophageal screening value from routine noncontrast CT and low‑dose CT scans, achieving very high specificity and substantial cancer sensitivity across large multicentre cohorts, with the potential to concentrate endoscopic resources on higher‑risk individuals. Full technical and implementation details were not available in the provided excerpt.