Self-supervised 3D deep learning on preoperative contrast-enhanced computed tomography for predicting high pathologic nodal burden in esophageal squamous cell carcinoma: temporal and external multicohort validation.

BACKGROUND

Accurate preoperative identification of high nodal burden (pathologic N2 or N3; hereafter N2+) is important in esophageal squamous cell carcinoma, but contrast-enhanced computed tomography criteria based mainly on nodal size and morphology have limited sensitivity.

METHODS

In this retrospective multicohort study, 1,060 consecutive patients with esophageal squamous cell carcinoma who underwent preoperative contrast-enhanced computed tomography and curative-intent esophagectomy with lymphadenectomy were enrolled from two centers. Center A contributed a development cohort (n = 612; train/validation/internal test, 428/92/92) and a temporally held-out cohort (n = 238), and Center B contributed an external test cohort (n = 210). A three-dimensional residual convolutional neural network encoder was pretrained on 3,200 unlabeled chest computed tomography examinations using masked-volume reconstruction and then fine-tuned on tumor-centered volumes comprising the primary tumor plus a 5-mm margin.

RESULTS

The self-supervised model achieved area under the receiver operating characteristic curve values of 0.881 (95% confidence interval, 0.793-0.955) in the internal test cohort, 0.860 (95% confidence interval, 0.810-0.903) in the temporal cohort, and 0.860 (95% confidence interval, 0.810-0.906) in the external cohort. In the external cohort, sensitivity and specificity at the main operating point were 0.581 and 0.845 for the self-supervised model versus 0.339 and 0.784 for the guideline-inspired comparator. Calibration also improved with self-supervised pretraining (Brier score, 0.148; expected calibration error, 0.053).

CONCLUSION

A contrast-enhanced computed tomography-only self-supervised three-dimensional model predicted high pathologic nodal burden in esophageal squamous cell carcinoma with robust temporal and external validation and showed numerically higher performance than a transparent computed tomography-only comparator. Calibrated risk estimates may help prioritize additional nodal workup when staging resources are limited or routine computed tomography findings are equivocal.