The value of multi-sequence MRI radiomics in predicting the early response to concurrent chemoradiotherapy for locally advanced cervical squamous cell carcinoma
10.3760/cma.j.cn112149-20211117-01024
- VernacularTitle:多序列MRI影像组学预测局部晚期宫颈鳞癌同步放化疗早期治疗反应的价值
- Author:
Jing TONG
1
;
Guiguang BEI
;
Yangyang KAN
;
Shuai HE
;
Lijun WEI
;
Yahong LUO
Author Information
1. 中国医科大学肿瘤医院 辽宁省肿瘤医院医学影像科，沈阳 110042
- Keywords:
Uterine cervical neoplasms;
Carcinoma, squamous cell;
Magnetic resonance imaging;
Radiomics;
Concurrent chemoradiotherapy
- From:
Chinese Journal of Radiology
2022;56(11):1223-1229
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To investigate the predictive value of MRI radiomics model in assessing the early response to concurrent chemoradiotherapy for locally advanced cervical squamous cell carcinoma.Methods:A total of 367 patients with pathologically proven locally advanced cervical squamous cell carcinoma (International Federation of Gynecology and Obstetrics stage ⅡB-ⅣA) in Liaoning Cancer Hospital & Institute from January 2013 to June 2019 were retrospectively collected. The patients were unable to undergo surgery and received complete concurrent chemoradiotherapy. Pelvic plain MRI, DWI and dynamic contrast-enhanced MRI were performed within 2 weeks before treatment and at the end of the 4th week of treatment. Patients were divided into complete response (CR) group ( n=247) and non-CR group ( n=120) according to response evaluation criteria in solid tumors 1.1. The patients were divided into a training set ( n=256) and a validation set (n=111) via a randomized split method at a ratio of 7∶3. Two radiologists drew the region of interest on the DWI, T 2WI and contrast-enhanced T 1WI (delayed phase) images before treatment to form the volume of interest finally. Totally 1 906 radiomics features were extracted from 3 single sequence images, respectively. Feature correlation analysis and tree model were used for feature selection. Three classifier learning algorithms, namely logistic regression (LR), support vector machine and random forest, were used for machine learning and the best classifier was selected. Based on the best classifier, 3 single sequence radiomics models were built, and a multi-sequence combined model was obtained by multivariate LR analysis. The differences in the area under the receiver operating characteristic curve (AUC) of the 3 single sequence models and the multi-sequence combined model were compared by DeLong test. The clinical application value of the multi-sequence combined model was evaluated by decision analysis curve. Results:In the training set and validation set, the LR classifier model had the best performance. Based on the LR classifier, AUCs of DWI, T 2WI, contrast-enhanced T 1WI and combined sequences in the training set were 0.77, 0.74, 0.79 and 0.86, respectively, and AUCs in the validation set were 0.71, 0.66, 0.75 and 0.77, respectively. In the training set, the AUC of multi-sequence combined model was higher than those of DWI, T 2WI and contrast-enhanced T 1WI sequence models, and the differences were statistically significant ( Z=3.01, 3.56, 2.83; P=0.003, 0.001, 0.005). In the validation set, the AUC of multi-sequence combined model and T 2WI model had significant difference ( Z=2.46, P=0.015). The decision analysis curve showed that when the threshold probability was in the range of 0.44 to 0.88, the multi-sequence combined model yielded a net benefit. Conclusion:Based on the LR classifier, the combined model built by radiomics features of multi-sequence MRI images has predictive value for assessing the early response of concurrent chemoradiotherapy for locally advanced cervical squamous cell carcinoma.
