177 CT RADIOMICS BASED ON MACHINE LEARNING PREDICTING PATHOLOGIC COMPLETE RESPONSE AFTER NEOADJUVANT CHEMORADIOTHERAPY FOR ESOPHAGEAL CANCER. (17th September 2021)

• Record Type: Journal Article
• Authors: Leng, Xue-Feng; Wang, Qi-Feng; Lang, Jin-Yi; Li, Tao; Han, Yong-Tao; Peng, Lin; Yue, H; Cheng, J; Liu, J
• Is Part Of: Diseases of the esophagus. Volume 34(2021)Supplement 1
• Publication Date: 2021-09-17
• Subjects: Esophagus -- Diseases -- Periodicals; 616.32
• DOI: 10.1093/dote/doab052.177
• Languages: English
• ISSNs: 1120-8694

Abstract

Abstract: : Neoadjuvant chemoradiotherapy (nCRT) has become the standard treatment for patients with locally advanced squamous cell esophageal cancer (LA-ESCC). Patients with the complete pathologic response (pCR) have significantly improved long-term survival. All efforts should be to improve the accuracy of predicting pCR. In this study, we investigate the use of radiomics based on machine learning to identify the pathologic complete response of patients with esophageal squamous cell carcinoma (ESCC) based on Computed Tomography (CT). Methods: The study included 155 patients with pathologically confirmed LA-ESCC. All 155 patients underwent simulation CT before nCRT, Quantitative radiomics features were extracted from CT images of each patient. To explore the relationship between radiomics features and the pCR, we used five-fold cross validation to classify the training and the testing cohorts. The Least Absolute Shrinkage and Selectionator operator (Lasso) were used to select features useful for the grading of pCR in the training cohort. Different models were measured in the training cohort using accuracy, sensitivity, specificity, and area under the curve of the receiver operating characteristic curve (AUC). Results: There were 155 patients. The pretreatment clinical stage was II in 16 patients (10.3%), III in 132 (85.2%), and IV in 7 (4.5%). Pathologic response was complete in 69 patients (44.5%), near-partial complete in 86 (55.5%). A total of 2193 radiomics features were extracted in the training set. After the use of statistical dimensionality reduction, five radiomics features were selected by Lasso to build radiomics signature. Prediction models for pCR were developed, and the model was able to predict pCR well in the training set(AUC = 0.902). In the testing cohorts, the model had a good performance in predicting pCR (AUC = 0.78). Conclusion: This study showed that CT-based radiomics features could be used as biomarkers to predict the complete pathological response of esophageal cancer underwent Neoadjuvant chemoradiotherapy.