A Bayesian optimization tunning integrated multi-stacking classifier framework for the prediction of radiodermatitis from 4D-CT of patients underwent breast cancer radiotherapy.

Purpose: In this study, we aimed to develop a novel Bayesian optimization based multi-stacking deep learning platform for the prediction of radiation-induced dermatitis (grade ≥ two) (RD 2+) before radiotherapy, by using multi-region dose-gradient-related radiomics features extracted from pre-treatment planning four-dimensional computed tomography (4D-CT) images, as well as clinical and dosimetric characteristics of breast cancer patients who underwent radiotherapy. Materials and methods: The study retrospectively included 214 patients with breast cancer who received radiotherapy after breast surgeries. Six regions of interest (ROIs) were delineated based on three PTV dose -gradient-related and three skin dose-gradient-related parameters (i.e., isodose). A total of 4309 radiomics features extracted from these six ROIs, as well as clinical and dosimetric characteristics, were used to train and validate the prediction model using nine mainstream deep machine learning algorithms and three stacking classifiers (i.e., meta-learners). To achieve the best prediction performance, a Bayesian optimization based multi-parameter tuning technology was adopted for the AdaBoost, random forest (RF), decision tree (DT), gradient boosting (GB) and extra tree (XTree) five machine learning models. The five parameter -tuned learners and the other four learners (i.e., logistic regression (LR), K-nearest neighbors (KNN), linear discriminant analysis (LDA), Bagging) whose parameters cannot be tuned, all as the primary week learners, were fed into the subsequent meta-learners for training and learning the final prediction model. Results: The final prediction model included 20 radiomics features and eight clinical and dosimetric characteristics. At the primary learner level, on base of Bayesian parameter tuning optimization, the RF, XGBoost, AdaBoost, GBDT, and LGBM models with the best parameter combinations achieved AUC of 0.82, 0.82, 0.77, 0.80, and 0.80 prediction performance in the verification data set, respectively. In the secondary meta-learner lever, compared with LR and MLP meta-learner, the best predictor of symptomatic RD 2+ for stacked classifiers was the GB meta-learner with an area under the curve (AUC) of 0.97 [95% CI: 0.91-1.0] and an AUC of 0.93 [95% CI: 0.87-0.97] in the training and validation datasets, respectively and the 10 top predictive characteristics were identified. Conclusion: A novel multi-region dose-gradient-based Bayesian optimization tunning integrated multi-stacking classifier framework can achieve a high-accuracy prediction of symptomatic RD 2+ in breast cancer patients than any other single deep machine learning algorithm.