Author: Fangfen Dong, Jiaming Li, Xiaobo Li, Weipei Wang, Zhixin Wang, Bing Wu, Benhua Xu, Yong Yang, Yifa Zhao 👨🔬
Affiliation: Department of Radiation Oncology, Fujian Medical University Union Hospital/Fujian Key Laboratory of Intelligent Imaging and Precision Radiotherapy for Tumors/Clinical Research Center for Radiology and Radiotherapy of Fujian Province (Digestive, Hematologi, Zhangpu County Hospital, School of Medical Imaging, Fujian Medical University 🌍
Purpose: To explore the construction and clinical application value of a deep learning-based positioning error prediction model, providing a reference for optimizing iSCOUT system-guided precision radiotherapy for breast cancer.Methods: This study analyzed 1,200 scans from the iSCOUT system involving 80 breast cancer patients. Initially, factors influencing patient positioning errors were identified through literature review and clinical experience, leading to the collection of 13 feature values as input for the deep learning model. The output focused on classifying the maximum positioning error in three translational directions, categorized into two classes based on a 3mm threshold. Feature selection utilized XGBoost to assess feature importance, ranking the most significant features for further analysis. The dataset was divided into training and validation sets in an 80:20 ratio. A deep neural network was trained on the training set, with preliminary performance evaluation conducted using the validation set. Lastly, the predictive outcomes of the deep neural network were compared with those from reference models, including SVM-SVC and DecisionTreeClassifier, to evaluate performance differences. This comprehensive approach aimed to enhance understanding and accuracy in predicting positioning errors during treatment.Results: Among the various training models, XGBoost showed higher accuracy. The deep neural network constructed in this study achieved the highest prediction accuracy for patient positioning errors on the 4th to 7th days of treatment, with a maximum accuracy of 71.87%.Conclusion: The deep learning-based method can predict instances where the patient's positioning error exceeds 3mm, although accuracy needs improvement. To optimize iSCOUT image-guided frequency, enhance radiotherapy efficacy, and reduce radiotherapy side effects, more feature parameters need to be identified for model training.