← Back to Home

A Closer Look at Chest CT: Multinational Validation of AI for Identifying Esophageal Cancer Risk

Chest scans originally acquired for other purposes may offer another opportunity to detect esophageal lesions. EAGLE demonstrated detection capabilities in large-scale validation, but early lesions may still be missed, and whether it can improve patient survival remains unproven.

By SURL BioNews

Could an existing chest computed tomography scan offer another opportunity to detect esophageal cancer? A study published on September 22 in *Nature Medicine* provided large-scale validation evidence for this approach. An artificial intelligence model named EAGLE analyzes noncontrast chest CT scans to identify people who may need further endoscopic evaluation, adding a risk alert for esophageal lesions to an existing examination.

This is no easy task. The esophagus is a hollow, tubular organ, and its images are readily affected by collapse of the lumen and motion artifacts. Small lesions are also difficult to distinguish from normal tissue. According to development details from Alibaba DAMO Academy, the team trained the model by pairing lesion locations provided by endoscopy reports and contrast-enhanced CT with the corresponding noncontrast images. Sichuan Cancer Hospital noted that, in addition to providing a positive or negative classification, the model marks suspicious regions and provides heatmaps for physicians to review.

The study trained the model using data from 6,813 patients at two centers, then validated it in a total of 80,612 patients across 12 centers. The development team said the validation covered China, the Czech Republic, and Australia. In external testing involving 11,466 people at eight of these centers, sensitivity for esophageal cancer was 90.0% and specificity was 98.5%. The former means the model could identify approximately nine out of ten cancer cases, while the latter reflects its ability to rule out people without the disease. Neither figure can be directly interpreted as meaning that someone who receives an alert has a 90% chance of having cancer.

The results for early lesions require separate consideration. Sensitivity for precancerous lesions in external testing was 52.5%. In another dataset of 702 people who underwent both CT and endoscopy, sensitivity for high-grade intraepithelial neoplasia and stage I esophageal cancer was 65.0% and 78.4%, respectively, when a threshold favoring higher sensitivity was used. These results show that the model can capture some early signs, while also reminding readers that figures from different populations and thresholds cannot be directly compared, and that negative results may still miss lesions.

How many follow-up examinations an alert generates is another challenge for clinical implementation. A study abstract included by the French National Cancer Institute stated that the team calibrated the model using real-world data from 35,402 people, reducing false positives by 72.7% while retaining sensitivity. However, in prospective validation within a hospital involving 17,446 people, the positive predictive value was still 42.2%. This means that more than half of positive alerts were not confirmed as positive cases under the study's definition, making the subsequent confirmation process essential.

The study also found similar performance in low-dose CT validation involving 1,607 people, suggesting the possibility of integration with existing lung cancer screening workflows. If existing images are analyzed, another CT scan would not be needed for this AI assessment. The development team also said that a clinical simulation using EAGLE for initial risk stratification could increase the endoscopic lesion detection rate approximately threefold. This reflects the efficiency of selecting people for examination; it does not yet support an inference that cancer detection rates or survival in the overall population would improve in parallel.

EAGLE currently offers a screening pathway that merits further evaluation: identifying clues in existing images, followed by confirmation through clinical examinations. Validation across centers adds evidence of its applicability, but the study has not established a benefit in reducing mortality. Before it enters routine care, further clarification is needed on its performance in local populations, the referral burden following alerts, and how to balance missed diagnoses against additional examinations.

References

  1. Nature Medicine via PubMed
  2. Sichuan Cancer Hospital
  3. Alibaba Cloud Community
  4. Institut national du cancer