Predictive value of clinical radiomics model based on nnU-Net for prognosis of gallbladder carcinoma
10.3760/cma.j.cn115610-20220406-00185
- VernacularTitle:基于nnU-Net的临床影像组学模型对胆囊癌预后的预测价值
- Author:
Zhechuan JIN
1
;
Qi LI
;
Dong ZHANG
;
Chen CHEN
;
Jian ZHANG
;
Min YANG
;
Qiuping WANG
;
Zhimin GENG
Author Information
1. 西安交通大学一附属医院肝胆外科，西安 710061
- Keywords:
Gallbladder neoplasms;
Deep learning;
nnU-Net;
Radiomics;
Survival prediction model
- From:Chinese Journal of Digestive Surgery
2022;21(5):656-664
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To investigate the predictive value of clinical radiomics model based on nnU-Net for the prognosis of gallbladder carcinoma (GBC).Methods:The retrospective cohort study was conducted. The clinicopathological data of 168 patients who underwent curative-intent radical resection of GBC in the First Affiliated Hospital of Xi'an Jiaotong University from January 2012 to December 2020 were collected. There were 61 males and 107 females, aged (64±11)years. All the 168 patients who underwent preoperative enhanced computed tomography (CT) examina-tion were randomly divided into 126 cases in training set and 42 cases in test set according to the ratio of 3:1 based on random number table. For the portal venous phase images, 2 radiologists manually delineated the region of interest (ROI), and constructed a nnU-net model to automatically segment the images. The 5-fold cross-validation and Dice similarity coefficient were used to evaluate the generalization ability and predictive performance of the nnU-net model. The Python software (version 3.7.10) and Pyradiomics toolkit (version 3.0.1) were used to extract the radiomics features, the R software (version 4.1.1) was used to screen the radiomics features, and the variance method, Pearson correlation analysis, one-way COX analysis and random survival forest model were used to screen important radiomics features and calculate the Radiomics score (Radscore). X-tile software (version 3.6.1) was used to determine the best cut-off value of Radscore, and COX proportional hazard regression model was used to analyze the independent factors affecting the prognosis of patients. The training set data were imported into R software (version 4.1.1) to construct a clinical radiomics nomogram model of survival prediction for GBC. Based on the Radscore and the independent clinical factors affecting the prognosis of patients, the Radscore risk model and the clinical model for predicting the survival of GBC were constructed respectively. The C-index, calibration plot and decision curve analysis were used to evaluate the predictive ability of different survival prediction models for GBC. Observation indicators: (1) segmentation results of portal venous phase images in CT examination of GBC; (2) radiomic feature screening and Radscore calculation; (3) prognostic factors analysis of patients after curative-intent radical resection of GBC; (4) construction and evaluation of different survival prediction models for GBC. Measurement data with normal distribution were represented by Mean± SD. Count data were expressed as absolute numbers or percentages, and comparison between groups was analyzed using the chi-square test. Univariate and multivariate analyses were performed using the COX proportional hazard regression model. The postoperative overall survival rate was calculated by the life table method. Results:(1) Segmentation results of portal venous phase images in CT examination of GBC: the Dice similarity coefficient of the ROI based on the manual segmentation and nnU-Net segmentation models was 0.92±0.08 in the training set and 0.74±0.15 in the test set, respectively. (2) Radiomic feature screening and Radscore calculation: 1 502 radiomics features were finally extracted from 168 patients. A total of 13 radiomic features (3 shape features and 10 high-order features) were screened by the variance method, Pearson correlation analysis, one-way COX analysis and random survival forest model. Results of random survival forest model analysis and X-tile software analysis showed that the best cut-off values of the Radscore were 6.68 and 25.01. A total of 126 patients in the training set were divided into 41 cases of low-risk (≤6.68), 72 cases of intermediate-risk (>6.68 and <25.01), and 13 cases of high-risk (≥25.01). (3) Prognostic factors analysis of patients after curative-intent radical resection of GBC: the 1-, 2-, and 3-year overall survival rates of 168 patients were 75.8%, 54.9% and 45.7%, respectively. The results of univariate analysis showed that preopera-tive jaundice, serum CA19-9 level, Radscore risk (medium risk and high risk), extent of surgical resection, pathological T staging, pathological N staging, tumor differentiation degree (moderate differentiation and low differentiation) were related factors affecting prognosis of patients in the training set ( hazard ratio=3.28, 3.00, 3.78, 6.34, 4.48, 6.43, 3.35, 7.44, 15.11, 95% confidence interval as 1.91?5.63, 1.76?5.13, 1.76?8.09, 2.49?16.17, 2.30?8.70, 1.57?26.36, 1.96?5.73, 1.02?54.55, 2.04?112.05, P<0.05). Results of multivariate analysis showed that preoperative jaundice, serum CA19-9 level, Radscore risk as high risk and pathological N staging were independent influencing factors for prognosis of patients in the training set ( hazard ratio=2.22, 2.02, 2.89, 2.07, 95% confidence interval as 1.20?4.11, 1.11?3.68, 1.04?8.01, 1.15?3.73, P<0.05). (4) Construction and evaluation of different survival prediction models for GBC. Clinical radiomics model, Radscore risk model and clinical model were established based on the independent influencing factors for prognosis, the C-index of which was 0.775, 0.651 and 0.747 in the training set, and 0.759, 0.633, 0.739 in the test set, respectively. The calibration plots showed that the Radscore risk model, clinical model and clinical radiomics model had good predictive ability for prognosis of patients. The decision curve analysis showed that the prognostic predictive ability of the clinical radiomics model was better than that of the Radscore risk and clinical models. Conclusion:The clinical radiomics model based on the nnU-Net has a good predictive performance for prognosis of GBC.
