A polit study of using CT-radiomics based machine learning model in predicting immune cells infiltrating and prognosis of pancreatic cancer
10.3760/cma.j.cn112149-20211214-01104
- VernacularTitle:基于CT图像影像组学的机器学习模型预测胰腺癌免疫细胞浸润及预后的初步研究
- Author:
Tiansong XIE
1
;
Weiwei WENG
;
Wei LIU
;
Kefu LIU
;
Weiqi SHENG
;
Zhengrong ZHOU
Author Information
1. 复旦大学附属肿瘤医院放射诊断科 复旦大学上海医学院肿瘤学系，上海 200032
- Keywords:
Pancreatic neoplasms;
Tomography, X-ray computed;
Radiomics;
Machine learning
- From:
Chinese Journal of Radiology
2022;56(4):425-430
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To investigate the value of CT-radiomics based machine learning model in predicting the abundance of tumor infiltrating CD8 +T cells and the prognosis of pancreatic cancer patients. Methods:A total of 150 pancreatic cancer patients who underwent surgical excision and confirmed by pathology from Fudan University Shanghai Cancer Center between December 2011 and January 2017 were retrospectively enrolled. The patients were randomly divided into the training set ( n=105) and the validation set ( n=45) in a 7∶3 ratio with simple random sampling. The immunohistochemical method was used to assess the abundance of tumor infiltrating CD8 +T cells, and the patients were then divided into high infiltrating group ( n=75) and low infiltrating group ( n=75) according to the median. The prognosis between the 2 groups was evaluated using Kaplan-Meier method and log-rank test. Radiomic features were extracted from preoperative venous-phase enhanced CT images in the training set. The Wilcoxon test, the max-relevance and min-redundancy algorithm were used to select the optimal feature set. Three supervised machine learning models (decision tree, random forest and extra tree) were established based on the optimal feature set to predict the abundance of tumor infiltrating CD8 +T cells. Performance of above-mentioned models to predict the abundance of tumor infiltrating CD8 +T cells in pancreatic cancer was tested in the validation set. The evaluation parameters included area under the receiver operating characteristic curve (AUC), F1-score, accuracy, precision and recall. Results:The median overall survival time of patients in high infiltrating group and low infiltrating group were 875 days and 529 days, respectively (χ2=11.53, P<0.001). The optimal feature set consisted of 10 radiomic features in training set. In the validation set, the decision tree, random forest and extra tree model showed the AUC of 0.620, 0.704 and 0.745, respectively; corresponding to a F1-score of 0.457, 0.667 and 0.744, the accuracy of 57.8%, 68.9% and 75.6%, the precision of 66.7%, 73.7% and 80.0%, the recall of 34.8%, 60.9% and 69.6%. Conclusions:Pancreatic cancer patients with high tumor infiltrating CD8 +T cells have better prognosis than those with low tumor infiltrating CD8 +T cells. The radiomics-based extra tree model is valuable in predicting the CD8 +T cells infiltrating level in pancreatic cancer.
