Self-supervised 3D CT predicts high nodal burden in esophageal SCC with external validation AUC 0.86

A self-supervised 3D deep learning model using preoperative contrast-enhanced CT predicted high pathologic nodal burden (pN2+) in esophageal squamous cell carcinoma. In external validation, AUC was 0.86, sensitivity 58% vs 34% for conventional CT criteria.

• Retrospective multicohort study: 1,060 patients from two centers; development cohort n=612, internal test n=92, temporal validation n=238, external test n=210.
• Model pretrained on 3,200 unlabeled chest CTs via masked-volume reconstruction, fine-tuned on tumor volumes. Self-supervised pretraining improved calibration (Brier score 0.148, expected calibration error 0.053).
• Limitation: retrospective design, moderate sensitivity (58%) in external cohort, model limited to primary tumor region, and single-center development.

RadPigeon summaries are original and for information only. They are not clinical advice.