Developing Dosiomics Models for the Prediction of Postoperative Radiotherapy-Induced Esophagitis in Patients with Non-Small Cell Lung Cancer

Purpose/Objective(s) Symptomatic postoperative radiotherapy-induced esophagitis (RE) (Grade≥2) develops in up to 31% of patients undergoing thoracic radiotherapy, lowering the quality life and sometimes leading treatment interruption, which negatively affects overall survival. We adopted common omics convolutional neural network (CNN) methods extract dosiomics features predict RE with non-small cell lung cancer (NSCLC), compared them dosimetric features-based model. Materials/Methods Patients histologically confirmed pIIIA-N2 NSCLC who underwent resection followed by radiotherapy from 2011 2015 were enrolled. The endpoint was Grade≥2 according Common Terminology Criteria for Adverse Events (CTCAE v4.0). Dosiomics extracted 3D dose distribution within esophagus using methods. screened an entropy-based method further selected minimum redundancy maximum relevance (mRMR) method. Prediction models built machine learning algorithms these as input. area under curve (AUC), accuracy, sensitivity, specificity, F1 score used evaluate performance prediction models. model input also comparison. Results A total 189 eligible. Models trained hyper-tuned on 80% (n = 151) dataset validated remaining 20% 38). In total, 107 4096 derived CNN methods, respectively. Four selected, including glcm_ClusterShade, shape_MinorAxisLength, firstorder_Kurtosis, Shape_Flatness. Three CNN-extracted while not interpretable. (Dmax), equivalent uniform (EUD), percentages volume receiving >20 Gy (V20), >60 (V60). AUC, 0.74, 0.70, 0.62, 0.80 model, 0.79, 0.77, 0.78, 0.68, 0.85 0.64, 0.58, 0.70 Conclusion Both outperformed more predictive but less interpretable than features.