Radiomics and Dosiomics for Predicting Complete Response to Definitive Chemoradiotherapy Patients with Esophageal Squamous Cell Cancer with Hybrid Institution Model

Abstract

<h3>Purpose/Objective(s)</h3> The stability of treatment outcome prediction can be increased by using multi-institution data. In the current study, we proposed the prediction model of local pathological complete response for the definitive chemoradiotherapy patients with radiomics features from computed tomography (CT) image, 18-fluorodeoxyglucose positron emission tomography (FDG PET) images and dosiomics features with multi-institution dataset. Moreover, we proposed the hybrid model to improve the prediction performance with the multi-institution data. <h3>Materials/Methods</h3> Local response was categorized into two groups (noncomplete response and complete response). Two models, external validation model and hybrid model, were proposed. For the external validation model, 112 patients from one institution were divided into training/validation and testing datasets. The model performance was evaluated with external validation dataset with 28 patients at second institution. For the hybrid model, patients across two institutions were randomly mixed into training/validation, testing, and final validation at a fixed ratio. A total of radiomics and dosiomics features were extracted using 5 segmentations on the CT, PET images, and dose distribution. Machine-learning based prediction models were constructed by using a decision tree (DT), sup-port vector machine (SVM), k-nearest neighbor algorithm (kNN), and neural network classifier (NN) classifiers. <h3>Results</h3> A total of 11063 features extracted from the radiomics analysis were reduced to 8630 features for the external validation model and 7629 features for the hybrid model by the VIF. A total of 7 CT-based radiomic features, 17 PET-based radiomics features, and 13 dosiomics features were selected in the external validation model. The accuracy was the highest 65.4% with the NN model from the CT-based radiomics, 77.9% with the NN classifier from the PET-based radiomics, and 72.1% with the NN classifier from the dosiomics. For the hybrid model, 18 CT-based radiomic features, 19 PET-based radiomics features, and 18 dosiomics features were selected in the hybrid model. The accuracy was the highest 84.4% with the kNN classifier from the CT-based radiomics, 86.0% with the NN classifier from the PET-based radiomics, and 79.0% with the NN classifier from the dosiomics. <h3>Conclusion</h3> The proposed prediction model showed promising prediction performance of the local response of definitive chemoradiotherapy for esophageal cancer. The hybrid model using NN classifier and PET-based radiomics features showed a possibility to improve the prediction performance.