Deep learning CT-based radiomics for prediction of treatment response to neoadjuvant chemoradiation in oesophageal squamous cell carcinoma

Abstract

Purpose: Neoadjuvant chemoradiotherapy (NCRT) plus surgery improves long-term survival of patients with locally advanced esophageal squamous cell carcinoma (ESCC). Treatment response prediction remains a great challenge. We aimed to evaluate the value of deep learning radiomics models based on computed tomography (CT) for predicting pathologic complete response (pCR) in ESCC patients receiving NCRT. Methods and materials: We identified 161 patients with ESCC (mean age: 58. mate: 83.5%. pCR: 46.0%). A total of 2048 deep learning radiomics features were analysed by the convolutional neural network (Xception) from CT images. After feature selection, a radiomics signature was butt with an extreme gradient boosting (XGBoost) algonthm. Two models were built. Model A. for post-NCRT evaluation, incorporates both pre-NCRT and post-NCRT CT images into the analysts, while Model B. for pretreatment assessment, was built based on pre- NCRT CT images only. Results: Model A comprised 9 selected features and showed good discrimination performance in test set for treatment response to NCRT. with an accuracy of 0.78. area under the receiver operating characteristic curve (AUC) of 0.89. sensitivity of 0.70, and specificity of 0.96. Calibration curves demonstrated good agreement between the prediction probability and the observed pCR (Hosmer-Lemeshow test. P-value = 0.66). Decision curve analysts confirmed the clinical benefits. Model B was also found to have predictrve potential, with an accuracy of 0.67. AUC of 0.73. sensitrvity of 0.87. and specificity of 0.62. Conclusion: Deep learning radiomics analysis based on CT demonstrated promising predictive value for NCRT treatment response in locally advanced ESCC. Both pretreatment and post-NCRT models could be potentially used for treatment strategy decision-making. Limitations: Future studies in larger prospective cohorts are needed to further confinn dmical practicability.