Program Information
Radiomics Prediction Model for Locally Advanced Rectal Cancer
J Wang*, H Zhong , L Shen , P Hu , J Gan , R Luo , z zhou , W Hu , Z Zhang , Fudan University Shanghai Cancer Center, Shanghai, shanghai
Presentations
TU-H-FS4-5 (Tuesday, August 1, 2017) 4:30 PM - 6:00 PM Room: Four Seasons 4
Purpose: The purpose of this study was using radiomics to develop robust models to predict prognosis for locally advanced rectal cancer patients treated with neoadjuvant chemoradiation followed by surgery.
Methods: A total of 411 with locally advanced rectal cancer patients treated with neoadjuvant chemoradiation were enrolled in this study. All patients’ radiotherapy treatment planning CT were collected. After segmentation, 271 radiomics features were calculated by an in-house radiomcis software. According to the results of test-retest and contour-recontour studies, unstable radimomics (spearman correlation coefficient < 0.7) was filtered. 21 radiomics features was left for final modeling. To get a reliable result, two statistic method were implemented including cluster analysis and cross validation based multivariable modeling. The prediction value of the radiomics features and clinical features were evaluated by comparing 3 models, which including model with radiomics only, model with clinical features only and model with radiomics and clinical features. The outcome includes local control, distant control, overall survival (OS) and disease free survival (DFS). C-index was calculated to evaluation model performance.
Results: The cluster results showed patients can be split into two groups by Non-negative matrix factorization(NMF) based cluster. The chi-square test results for this cluster showed that there were no clinical features correlated to radiomics feature cluster results. These two groups of patient have significant difference in OS (p = 0.032, logrank test). For other outcome, including local control, distant control and DFS, there were no statistical difference between these two groups. In supervised modeling, the OS was improved by radiomcis features from 0.672 [0.617 0.728], which was clinical features only, to 0.730 [0.658 0.801]. This indicated radiomics features were independent factors compared to clinical features for OS prediction.
Conclusion: We developed a radiomics prediction model which can predict OS for locally advanced rectal cancer patients received neoadjuvant chemoradiation.
Contact Email: