A prediction model for pathological findings after neoadjuvant chemoradiotherapy for resectable locally advanced esophageal cancer based on PET images using radiomics and machine-learning.

Abstract

456 Background: The pathologic complete response (PCR) rate by neoadjuvant chemoradiotherapy (NCRT) for resectable locally advanced esophageal squamous cell carcinoma (ESCC) is about 40%. If we could predict a PCR from pre-treatment image data, it might be possible to select patients who can be cured by organ-preserving CRT. The purpose of this study is to construct a predictive model for PCR by NCRT in patients with locally advanced ESCC using radiomics and machine-learning. Methods: We used data of 98 ESCC patients who underwent NCRT and surgery from 2003 to 2016. Firstly, we fused the radiotherapy treatment planning CT images and PET images scanned before treatment. Then using target delineations on planning CT images, we created eight kinds of target regions on PET images. Secondly, we generated a total of 6968 features per patient using the PET image data within these target regions that were preprocessed by radiomics technique. Among them, we extracted the optimal features for machine-learning using the least absolute shrinkage and selection operator (LASSO) logistic regression. Thirdly, artificial neural networks were used as a machine-learning method to create a predictive model. The extracted radiomics features were used as input values, and the information of ‘PCR’ or ‘not PCR’ was used as output values. We used data of randomly selected 58 patients for training and constructed a predictive model. Then we used data of 15 patients to validate the models and created the optimal model. Finally, we evaluated the predictive model using the test data of 25 patients. Results: By the LASSO analysis, 32 radiomics features were extracted for machine-learning classification. This predictive model predicted pathological findings after NCRT in 24 of 25 test data. The accuracy, specificity and sensitivity in the prediction of PCR after NCRT by this predictive model were 96.0%, 93.8%, and 100%, respectively. Conclusions: A prediction model based on PET images using radiomics and machine-learning could predict pathological findings after NCRT for resectable locally advanced ESCC.