CT-based deep learning model: a novel approach to the preoperative staging in patients with peritoneal metastasis

Abstract

AbstractPeritoneal metastasis (PM) is a frequent manifestation of advanced abdominal malignancies. Accurately assessing the extent of PM before surgery is essential for patients to receive optimal treatment. Therefore, we propose to construct a deep learning (DL) model based on enhanced computed tomography (CT) images to stage PM preoperatively in patients. All 168 patients with PM underwent contrast-enhanced abdominal CT before either open surgery or laparoscopic exploration, and peritoneal cancer index (PCI) was used to evaluate patients during the surgical procedure. DL features were extracted from portal venous-phase abdominal CT scans and subjected to feature selection using the Spearman correlation coefficient and LASSO. The performance of models for preoperative staging was assessed in the validation cohort and compared against models based on clinical and radiomics (Rad) signature. The DenseNet121-SVM model demonstrated strong patient discrimination in both the training and validation cohorts, achieving AUC was 0.996 in training and 0.951 validation cohort, which were both higher than those of the Clinic model and Rad model. Decision curve analysis (DCA) showed that patients could potentially benefit more from treatment using the DL-SVM model, and calibration curves demonstrated good agreement with actual outcomes. The DL model based on portal venous-phase abdominal CT accurately predicts the extent of PM in patients before surgery, which can help maximize the benefits of treatment and optimize the patient’s treatment plan. Graphical abstract