AI model combines scans and tissue clues to assess survival risk in esophageal cancer

An artificial intelligence framework combines CT scans, digital tissue images and clues from the tumor’s surroundings to estimate survival risk after surgery for esophageal squamous cell carcinoma. In a retrospective study of 450 patients, the system distinguished groups with different survival outcomes and performed better than models relying on any single information source. The work was published in iMetaMed on November 3, 2026.

Esophageal squamous cell carcinoma is a major form of esophageal cancer, particularly in East Asia. Clinicians commonly use tumor stage to judge prognosis, but stage alone does not capture all the biological differences between patients. The research team set out to test whether combining several views of the same tumor could provide a more informative picture of the risk patients face after surgery.

The framework, called MENISA, brings together routine clinical information, CT imaging and whole-slide pathology images with seven markers describing the tumor microenvironment. These markers include cancer invasion around nerves and into blood or lymphatic vessels. Such tissue features sit between the larger view provided by a scan and the fine detail visible in a pathology image, potentially adding information that either imaging approach alone may miss.

To develop and evaluate the model, the researchers analyzed records from patients who had undergone surgical resection. Data from 273 patients at Qilu Hospital of Shandong University were used for training, and another 63 patients from the same hospital formed the internal validation group. The external validation group included 114 patients treated at Liaoning Cancer Hospital. Patients who had received preoperative neoadjuvant treatment were excluded.

In internal validation, MENISA achieved a concordance index of 0.778; in the external hospital cohort, the score was 0.675. This measure assesses how well a model ranks patients by survival risk, with higher scores indicating better agreement with observed outcomes. The model also separated patients into groups with different survival outcomes. Analysis of its predictions highlighted nerve and vessel invasion as important contributors, alongside information from pathology and CT images.

Another practical challenge is that hospitals do not always collect or annotate images in the same way. The team therefore tested a feature-imputation approach to bridge differences between two-dimensional and three-dimensional CT annotations. In the external cohort, this approach improved the model’s concordance index from 0.611 to 0.675, suggesting a possible way to accommodate differences in imaging workflows.

The findings remain a step toward clinical translation. The study used existing patient records, the external validation group was relatively small, and differences between hospitals could affect how well the framework works elsewhere. Limited follow-up and relatively few events beyond five years also call for caution in interpreting long-term predictions. Larger prospective studies with longer follow-up are needed to assess reproducibility and clinical usefulness. The study did not test whether using MENISA to guide treatment improves survival; that remains a question for future research.