Fully-automated 3D volume segmentation in CT images for preoperatively predicting the overall survival of resectable pancreatic ductal adenocarcinoma: a cohort study / 中华胰腺病杂志

ABSTRACT

Objective:To verify the predictive value of fully-automated 3D volume segmentation of CT images for the overall survival prognosis of resectable pancreatic ductal adenocarcinoma (PDAC).Methods:From July 2018 to March 2019, the clinical data of 198 cases of resectable PDAC were continuously collected in the First Affiliated Hospital of Naval Medical University. According to the level of carbohydrate antigen(CA)19-9 and carcinoembryonic antigen(CEA), the patient were divided into low CA19-9 group(≤210 U/ml ), high CA19-9 group (>210 U/ml ), normal CEA group (<5 ng/ml ) and high CEA group (≥5 ng/ml). Using our fully-automated segmentation tool developed in the early stage, images at the plain phase and portal phase were matched to those at the late artery phase by taking the artery phase as the matching target to establish UNet model; and the PDAC tumor and pancreatic glands were three-dimensionally segmented to estimate the tumor 3D volume. Univariate and multivariate logistic regression analysis were performed to compare the tumor 3D volume with the common preoperative risk factors (tumor 2D long diameter, CA19-9 level, CEA level, etc.) in predicting the patients′ survival. C-index was used to estimate the accuracy for predicting the survival. Receiver operating characteristics curve (ROC) was drawn and AUC was calculated to evaluate the accuracy for predicting the 1-year and 2-year overall survival and the influence of CA19-9 and CEA level on the patients′ overall survival.Results:Univariate logistic analysis showed that age, tumor 3D volume, tumor location, CA19-9 and CEA level were correlated with the patients′ overall survival. Multivariate logistic analysis showed that tumor 3D volume, CA199 and CEA were correlated with the overall survival. Among them, tumor 3D volume was most strongly correlated with the overall survival ( HR=2.25, 95% CI1.49-3.39, P<0.0001). The prognostic C-index of automatic 3D tumor volume, tumor long diameter, serum CEA and CA19-9 was 0.667(95% CI0.617-0.717), 0.637(0.583-0.691), 0.593(0.527-0.659) and 0.585(0.526-0.644), respectively. The AUCs of 3D tumor volume, tumor location, tumor long diameter, serum CEA and CA19-9 for predicting 1-year and 2-year survival were 0.726 and 0.698, 0.562 and 0.562, 0.703 and 0.660, 0.583 and 0.624, 0.602 and 0.609 respectively. C-index and AUC of tumor 3D volume was significantly better than those of the other common preoperative risk factors, and the difference was statistically significant (all P value <0.05). The survival of patients with large tumor 3D volume was greatly poorer than that of patients with small tumor 3D volume in low CA19-9 group, high CA19-9 group, normal CEA group and high CEA group, and the differences were all statistically significant ( HR=2.27, 95% CI 1.39-3.72; HR=2.42, 95% CI1.23-4.74; HR=2.08, 95% CI1.07-4.06; HR=2.67, 95% CI1.63-4.38, all P value <0.01). And the automatic 3D volume was the strongest predictor for the survival in high CA19-9 group. Conclusions:The tumor 3D volume obtained by automatic CT segmentation was an objective and reliable prognostic biomarker, which can supplement the established preoperativel risk factors and was expected to guide the personalized choice of neoadjuvant therapy.