J. Kim1, M. Kim1, H. Park1, O. Noh1, Y. Seol2, C. Park3, B. O. Choi2, Y. S. Kim2, S. H. Son2, J. Song2, Y. K. Lee2, and Y. N. Kang2; 1Department of Radiation Oncology, Ajou University School of Medicine, Suwon, Korea, Republic of (South), 2Department of Radiation Oncology, Seoul St. Mary’s Hospital, College of Medicine, The Catholic University of Korea, Seoul, Korea, Republic of (South), 3Department of Biomedicine & Health Sciences, College of Medicine, The Catholic University of Korea, Seoul, Korea, Republic of (South)
Purpose/Objective(s):This study aims to identify potential DQA failures in Tomotherapy plans pre-DQA execution and determine contributing factors. By utilizing plan parameters and Dosiomics from dose distributions, a predictive model is being developed to assess DQA success or failure likelihood and identify critical contributing factors, with the goal of streamlining replanning processes and enhancing radiation therapy treatment efficiency and quality. Materials/
Methods: Tomotherapy plans from two hospitals, involving over 600 patients treated with three Tomotherapy versions, were analyzed. Plan parameters and Dosiomics features were extracted from the plan DICOM file, respectively. DQA outcomes were obtained using a patient-specific quality assurance. Data were categorized into three groups: plan parameters, Dosiomics features, and a combined set. The preprocessing involved normalization and handling missing values and outliers. Data were classified as Pass or Fail based on a gamma index criterion (3mm/3% with a 97% threshold), and analyzed using machine learning algorithms, including Logistic Regression, Decision Trees, Random Forest etc. and ensemble methods, with cross-validation to prevent overfitting. SMOTE was used to address data imbalance, and model performance was assessed via Precision, Recall, F1-Score, and PR-AUC (Precision-Recall Area Under the Curve). Feature selection and hyperparameter optimization were conducted to refine models and improve prediction accuracy. Results: This study evaluated the efficacy of Plan Parameters (Group1), Dosiomics (Group2), and their Integration (Group3) in predicting DQA outcomes. The Random Forest algorithm outperformed in Groups 1 and 3, whereas the Extra Trees algorithm was most effective in Group 2. Among ensemble methods, the Stacking Classifier emerged as superior. Performance comparison revealed Group1 outdoing Group2, with Group3 demonstrating the best overall efficacy, evidenced by F1-scores of 0.823 for Group1, 0.807 for Group2, and 0.851 for Group3. The most influential plan parameters were Target Coverage, Modulation Factor, Field Width, Red Laser Position, and Treatment Duration. The Dosiomics features with the greatest significance in the model, indicating substantial predictive value, were image_Size_z, Mask_CenterOfMassIndex_z, glcm_ClusterShade, and firstorder_10Percentile etc. Conclusion: Given the inherent variability in therapy machine conditions and operator practices, achieving perfect accuracy in DQA results using phantoms is unattainable. Consequently, this studys findings are regarded as satisfactorily robust. These insights enable the advance identification of potential DQA failures and the determination of influential factors, streamlining the replanning process. This strategic advantage improves the precision and efficiency of radiation therapy planning, thereby elevating the standard of patient care and treatment success.
