47 APPLICATION OF MACHINE LEARNING TO PREDICT NEOADJUVANT CHEMORADIOTHERAPY TREATMENT RESPONSE IN ESOPHAGEAL SQUAMOUS CELL CARCINOMA USING COMPUTED TOMOGRAPHY. (14th September 2020)

• Record Type: Journal Article
• Authors: Hu, Y; Xie, C; Vardhanabhuti, V; Fu, J
• Is Part Of: Diseases of the esophagus. Volume 33(2020)Supplement 1
• Publication Date: 2020-09-14
• Subjects: Esophagus -- Diseases -- Periodicals; 616.32
• DOI: 10.1093/dote/doaa087.05
• Languages: English
• ISSNs: 1120-8694

Abstract

Abstract: : Predicting the neoadjuvant chemoradiotherapy (NCRT) treatment response in esophageal squamous cell carcinoma (ESCC) patients remains challenging. This study aims to evaluate the value of CT imaging-based machine learning models for predicting pathologic complete response (pCR) in ESCC patients receiving NCRT and to establish correlations with their underlying biology via radiogenomics analysis. Methods: We identified 231 eligible patients from two centers. Handcrafted radiomics features (analyzed by PyRadiomics) and deep learning features (analyzed by transfer learning using the convolutional neural network, Xception) were extracted from pretreatment CT images. A handcrafted radiomics model and deep learning model were built with support vector machine. The models were trained in the training cohort (n = 161) and validated in an external testing cohort (n = 70). The radiological model with better performance was incorporated into a nomogram. Pathway enrichment analysis in a subset of the cohort (n = 28) with gene expression profiles revealed the potential biological processes correlated with the radiological prediction. Results: We constructed a seven-feature handcrafted radiomics model and an eight-feature deep learning model to predict NCRT response. The deep learning model outperformed its counterpart (AUC: 0.76 vs. 0.73; accuracy: 71.4% vs. 65.7%) and was used to establish a nomogram model incorporating cN staging, which showed good discrimination in the testing cohort (C-index = 0.76, accuracy = 72.9%). Radiogenomics analysis illustrated a potential association between the radiological prediction and underlying molecular processes involving multiple factors, including WNT and TGF- signaling pathways, the microenvironment, radiation response and mitotic nuclear division. Conclusion: We developed a CT imaging-based machine learning model that showed satisfactory performance in NCRT response prediction for ESCC patients. Further radiogenomics analysis provided useful insights into the biological mechanisms of therapy resistance in ESCC. These findings could enhance the applications of radiological characteristics in precision oncology and clinical practice.