2026-10-01
EAGLE reads routine chest CT to flag esophageal cancer and precancerous lesions: 98.5% specificity across 80,612 patients, 42.2% PPV in prospective hospital deployment.
Esophageal cancer caused an estimated 445,000 deaths worldwide in 2022, and five-year survival sits at 36.9% in China and 18.5% in the US, mostly because tumors are found late. Endoscopy, the reference standard for early detection, is invasive and compliance is poor, so it cannot scale beyond high-incidence regions; liquid biopsy and sponge cytology have only moderate sensitivity for early lesions and require repeat visits. A cheap substitute already exists: chest noncontrast CT accounts for 40% of all CT examinations and covers the entire esophagus. The catch is that early esophageal lesions are invisible on these scans. The esophagus is a long hollow tube that collapses, sits against the pulsing heart and great vessels, and early malignancy is a thin mucosal change that experienced radiologists cannot see. A study cited in the paper found that most patients who died of esophageal cancer had no suspicious esophageal findings in their chest CT reports during the previous five years. This Nature Medicine paper takes on exactly that reading task.
EAGLE (Esophageal AI-Guided malignant Lesion Evaluation) is a two-stage 3D pipeline built on nnU-Net V2. Stage one localizes the esophagus with a low-resolution 3D UNet (ResEnc backbone), cropping away irrelevant anatomy. Stage two runs joint segmentation and classification on that crop: multiscale decoder features are aligned by trilinear interpolation, concatenated along depth, and fed to a classification head of one 3×3×3 convolution, global average pooling and a fully connected layer, while a segmentation head outputs the lesion mask. Total loss is segmentation loss plus α times cross-entropy, and a class activation map highlights the regions driving each decision.
The architecture is standard; the contribution is the annotation protocol. Early lesions are nearly invisible even on contrast CT, so ground truth was synthesized from other stops in the clinical workflow: surgical and endoscopic pathology reports gave centimeter-level lesion location, two radiologists with three or more years of esophageal imaging experience independently delineated lesions on paired contrast-enhanced CT, masks with Dice above 0.6 became the gold standard by intersection, disagreements went to a senior reviewer with more than eight years of expertise or the case was dropped, and deformable registration transferred the labels onto noncontrast CT for voxel-level supervision. Training used 6,813 patients from two Chinese cancer centers, including 264 cases of HGIN and 548 stage I cancers.
Two engineering choices determine deployability. Thresholds are set per scenario: 99% specificity in cross-validation (threshold 0.93) for opportunistic screening, 98% sensitivity for population screening, one model with two operating points. And after launch, hard cases from the real world (all false positives, plus false negatives and borderline positives at roughly 1:1) were merged back for 200 more epochs of fine-tuning, producing EAGLE-Plus.
Validation spanned 12 centers in three countries and 80,612 patients, along four tracks:
| Setting | Cohort | Key result |
| External multicenter test | 8 centers, n=11,466 | AUC 0.976, sensitivity 89.5%, specificity 98.5% |
| Reader study | 17 radiologists, 300 scans | EAGLE beat all 17 readers; readers with AI: sensitivity 71.9%→85.7%, specificity 79.6%→91.7% |
| Real-world calibration | 3 centers, n=35,402 | EAGLE-Plus cut false positives by 72.7%, PPV 25.3%→55.0%, sensitivity unchanged (88.3% vs 89.8%) |
| Prospective + LDCT program | 17,446 + 10,959 | Hospital PPV 42.2%; LDCT screening specificity 99.94% |
The early-disease numbers carry the most information: external sensitivity was 52.5% for HGIN and 60.1% for stage I cancer, 90.0% for confirmed cancer overall, rising with stage; at a specificity of 91.0%, matched to liquid biopsy, stage I sensitivity reaches 73.4%. In the reader study the 17 radiologists (3 esophageal specialists, 6 general radiologists, 8 residents) averaged 36.3% sensitivity on HGIN and 59.8% on stage I, against 65.7% and 87.1% for EAGLE alone; with AI assistance readers reached 82.5% on stage I, still below the unaided model, so human-AI collaboration did not capture the full model capability. Residents with AI approached specialist-level balanced accuracy.
On the population-screening track, a 702-patient paired CT-endoscopy cohort (98%-sensitivity operating point) showed 65.0% sensitivity for HGIN and 78.4% for stage I; 18 of 28 patients (64.3%) were flagged on CT scans taken before their diagnosis, four of them at least 9 months early. At Suining, 530 prospectively enrolled participants were combined with 3 resampled cancer cases into a 533-person hybrid cohort, and simulation of EAGLE-first triage lifted the endoscopic detection rate from 1.7% to 5.2%, cut the number needed to screen from 59.2 to 19.2, avoided 67.6% of endoscopies, with a negative predictive value of 99.46%. On the 1,607-scan LDCT test set, the version trained with low-dose simulation reached 88.4% sensitivity versus 83.0% without it (P=0.024). Among backbones compared under the same framework, ResEnc beat Swin-UNETR, Mask2Former and two Mamba variants; the nnDetection framework scored AUC 0.954 versus 0.976. Lesion segmentation Dice was 0.744.
Concrete deployment cases anchor the claims: real-world review caught 3 malignancies missed by the initial care pathway, one 21 months early; in the prospective hospital run, 16 AI-triggered referrals produced 3 completed endoscopies and 1 T3N0M0 cancer that radiology had missed; in the LDCT program, 1 of 8 AI-positive participants with a clean initial report was confirmed with cancer 8 days later, and the run also surfaced an esophagotracheal fistula needing immediate care.
This is the newest tile in the multi-cancer screening mosaic on noncontrast CT, a line that already produced models for pancreatic cancer (Nat Med 2023), gastric cancer (Nat Med 2025) and colorectal cancer (Ann Oncol 2026), with the esophagus long treated as a blind spot. The transferable insight is about labels: when the target is invisible to human eyes on the modality you have, supervision can be synthesized from other parts of the clinical workflow through pathology-based localization, contrast-enhanced delineation and registration transfer. That protocol ports directly to other early-detection targets that are invisible on imaging but localized in medical records. Deployment is pure software over existing scans, adding no radiation and no new procedure, and LDCT lung programs become dual-cancer programs for free. Alibaba DAMO built the model and the deployment system; 8 authors are Alibaba employees. Code stays closed behind patents (CN117853490B, US 18046385), data is not public per IRB rules, and a sample dataset plus an interactive portal are available at eagle.damomed.com.
The authors list four. Population diversity is thin: only one center each in the Czech Republic and Australia, distal and adenocarcinoma-heavy populations untested, and adding esophagogastric-junction and adenocarcinoma cases to training lifted EAC sensitivity from 86.6% to 91.2%, which shows the original model was weak there. Sensitivity is lower in women, likely from training-set imbalance. The pre-endoscopy triage evidence rests on retrospective paired cohorts plus a single-site simulation with few positive cases. Both prospective validations run under two years of follow-up with suboptimal adherence to recommended endoscopy, so true sensitivity may be understated, and the reader study lacked a full cross-over design, leaving recall bias unexcluded.
Closer reading adds caveats. The Suining hybrid cohort held 9 malignancies in 533 people, so the confidence interval around the tripling-of-efficiency claim is wide. In the hospital deployment, 52 of 90 AI-positive cases were not esophageal malignancy and 5 were lost to follow-up, so the 42.2% PPV depends on follow-up quality. Specificity at the population operating point is 84.7%, and false-positive load would grow in lower-prevalence settings. Over 90% of missed stage I lesions were Paris type 0-II, the flat superficial morphology that is hardest to find even under endoscopy, an honest disclosure of where the model fails.