nnU-Net-based CT segmentation of perigastric varices in sinistral portal hypertension: a multicenter study.

PURPOSE: To develop a deep learning model based on nnU-Net for automated segmentation of all perigastric veins on contrast-enhanced CT images in patients with sinistral portal hypertension (SPH).

METHODS

Retrospectively including contrast-enhanced computed tomography (CT) portal venous phase images from 172 pancreatic cancer patients with SPH across three hospitals in Nantong. The patients were divided into the training dataset (n = 99), model optimization dataset (n = 22), internal testing dataset (n = 22), and external testing dataset (n = 29). The self-configuring nnU-Net model was trained on the manually segmented training dataset. Evaluation metrics on the testing datasets included Dice similarity coefficient (DSC), recall, precision and Hausdorff distance (HD). Pearson correlation and Bland-Altman analyses were conducted between the reference standard and predicted diameters. Intraclass correlation coefficients (ICCs) were used to assess inter-rater agreement and reproducibility of radiomic features.

RESULTS

The nnU-Net model achieved a DSC of 0.784 (95% CI 0.710, 0.857) on the internal testing dataset and 0.780 (95% CI 0.705, 0.856) on the external testing dataset, outperforming the comparative models. Predicted diameters correlated strongly with the reference standard in both testing datasets, notably for the gastric coronary vein (r = 0.967), with all perigastric varices achieving correlations > 0.770 (p < 0.001). High intra-rater (ICC = 0.852; 95% CI 0.806, 0.897) and inter-rater agreement (ICC = 0.839; 95% CI 0.793, 0.885) were observed, and feature reproducibility remained robust in both the internal (ICC = 0.853; 95% CI 0.827, 0.879) and external testing datasets (ICC = 0.870; 95% CI 0.833, 0.906).

CONCLUSION

The nnU-Net model provides promising segmentation performance for perigastric varices in SPH.