摘要
Objective:To predict pathological nodal stage of locally advanced rectal cancer by a radiomic method that uses collective features of multiple lymph nodes(LNs)in magnetic resonance images before and after neoadjuvant chemoradiotherapy(NCRT).Methods:A total of 215 patients were included in this study and chronologically divided into the discovery cohort(n=143)and validation cohort(n=72).In total,2,931 pre-NCRT LNs and 1,520 post-NCRT LNs were delineated from all visible rectal LNs in magnetic resonance images.Geometric,first-order and texture features were extracted from each LN before and after NCRT.Collective features are defined as the maximum,minimum,mean,median value and standard deviation of each feature from all delineated LNs of each participant.LN-model is constructed from collective LN features by logistic regression model with L1 regularization to predict pathological nodal stage(ypN0 or ypN+).Tumor-model is constructed from tumor features for comparison by using DeLong test.Results:The LN-model selects 7 features from 412 LN features,and the tumor-model selects 7 features from 82 tumor features.The area under the receiver operating characteristic curve(AUC)of LN-model in the discovery cohort is 0.818[95%confidence interval(95%CI):0.745-0.878],significantly(Z=2.09,P=0.037)larger than 0.685(95%CI:0.602-0.760)of the tumor-model.The AUC of LN-model in validation cohort is 0.812(95%CI:0.703-0.895),significantly(Z=3.106,P=0.002)larger than 0.517(95%CI:0.396-0.636)of the tumor-model.Conclusions:The usage of collective features from all visible rectal LNs performs better than the usage of tumor features for the prediction of pathological nodal stage of locally advanced rectal cancer.
Objective: To predict pathological nodal stage of locally advanced rectal cancer by a radiomic method that uses collective features of multiple lymph nodes(LNs) in magnetic resonance images before and after neoadjuvant chemoradiotherapy(NCRT).Methods: A total of 215 patients were included in this study and chronologically divided into the discovery cohort(n=143) and validation cohort(n=72). In total, 2,931 pre-NCRT LNs and 1,520 post-NCRT LNs were delineated from all visible rectal LNs in magnetic resonance images. Geometric, first-order and texture features were extracted from each LN before and after NCRT. Collective features are defined as the maximum, minimum,mean, median value and standard deviation of each feature from all delineated LNs of each participant. LN-model is constructed from collective LN features by logistic regression model with L1 regularization to predict pathological nodal stage(ypN0 or ypN+). Tumor-model is constructed from tumor features for comparison by using DeLong test.Results: The LN-model selects 7 features from 412 LN features, and the tumor-model selects 7 features from 82 tumor features. The area under the receiver operating characteristic curve(AUC) of LN-model in the discovery cohort is 0.818 [95% confidence interval(95% CI): 0.745-0.878], significantly(Z=2.09, P=0.037) larger than 0.685(95% CI: 0.602-0.760) of the tumor-model. The AUC of LN-model in validation cohort is 0.812(95% CI:0.703-0.895), significantly(Z=3.106, P=0.002) larger than 0.517(95% CI: 0.396-0.636) of the tumor-model.Conclusions: The usage of collective features from all visible rectal LNs performs better than the usage of tumor features for the prediction of pathological nodal stage of locally advanced rectal cancer.
基金
supported by Beijing Municipal Administration of Hospitals Clinical Medicine Development of Special Funding Support (No. ZYLX201803)
Beijing Hospitals Authority’ Ascent Plan (No. DFL20191103)
National Key R&D Program of China (No. 2017YFC1309101, 2017YFC1309104)
