Automated CT segmentation of diffuse pseudomyxoma peritonei lesions using a hybrid deep learning model

Abstract Purpose To develop and externally validate DATUnet—a hybrid deep learning framework integrating DenseBlock encoding, atrous spatial pyramid pooling (ASPP), and a lightweight Transformer bottleneck—for automated segmentation of diffuse pseudomyxoma peritonei (PMP) lesions on contrast-enhanced CT. Materials and Methods This retrospective, dual-centre study enrolled 71 patients (59 primary centre, 12 external centre). Primary centre cases were allocated to training (n = 46, 6-fold cross-validation), independent held-out test (n = 7), and internal prospective validation (n = 6) cohorts; 12 external cases formed an independent external validation cohort. Lesions were independently annotated by two senior radiologists using intraoperative records as reference. DATUnet was evaluated using DSC, IoU, HD95, precision, and recall, and benchmarked against SwinUNet and nnUNetv2 under identical conditions. Model interpretability was assessed with Grad-CAM. Results On the independent test set, DATUnet achieved DSC 92.7%, IoU 87.2%, and HD95 24.098 mm. Internal prospective validation yielded DSC 92.4% and HD95 20.192 mm; external validation yielded DSC 84.9% and HD95 24.901 mm, confirming preserved generalization across institutions, scanners, and bowel-preparation protocols. DATUnet outperformed SwinUNet (DSC 81.3%) and nnUNetv2 (DSC 55.3%) across all overlap and boundary metrics. Grad-CAM visualizations demonstrated strong spatial correspondence between model attention and radiologist-annotated mucinous deposits, including regions with infiltrative growth patterns. Conclusion DATUnet achieves robust, reproducible automated segmentation of diffuse PMP lesions across diverse imaging environments, offering a quantitative foundation for preoperative lesion mapping, anatomical characterization, and cytoreductive surgery decision-making in clinical practice.