Room: AAPM ePoster Library
Purpose: Twenty to thirty percent of esophageal squamous cell carcinoma (ESCC) can achieve a pathological complete response (pCR) after neoadjuvant chemoradiotherapy (nCRT) followed by surgery. However, one-third of these pCR patients may still experience recurrence. The aim of this study is to predict the postoperative recurrence in those pCR patients by incorporating multi-section CT radiomics with machine learning.
Methods: 206 ESCC patients were enrolled in this study and were divided into a training cohort (n=146) and a validation cohort (n=60). Scanned transversal sectional CT images were included and then reconstructed into sagittal section images, like three-view drawing in engineering field. The binary contour masks were reconstructed with corresponding CT images. Then, 114 radiomic features were extracted from transversal section, reconstructed sagittal section and coronal section CT images, including Gray-Level Co-Occurrence Matrix (n=22), Gray-Level Run Length Matrix (n=11) and Neighborhood Gray Tone Difference Matrix (N=5). For robustness intends, reproducibility of radiomics features were evaluated to against the uncertainty caused by segmentation variability and volume-dependent risk factor. Feature selection was performed by RELIEFF algorithm and top three ranking features included in next predictive model building. Support vector machine (SVM) with 10-fold cross-validation was employed to build a prediction model and performance was assessed by area under the curve (AUC), sensitivity, and specificity.
Results: The AUC, sensitivity, and specificity for predicting recurrence and recurrence-free were 0.862 (95% confidence interval, 0.809-0.915), 0.784, 0.883 in training cohort, respectively, while they were 0.837 (95% confidence interval, 0.801-0.873), 0.841, 0.857 in validation cohort, respectively.
Conclusion: The proposed novel predictive model incorporate multi-section CT images with machine learning can accurately predict postoperative recurrence of pCR patients. It indicated that the prediction model incorporates multi-sections provide holistic information of intratumor heterogeneity may improve precision therapy delivery.