3126 - Delta-Dosiomics-Based Multi-Level Precision Identification of Radiotherapy Treatment Errors In Cervical Cancer

Purpose/Objective(s):

Conventional gamma analysis in dose verification has inherent limitations, as it cannot accurately identify treatment error types or quantify their magnitudes. This study aimed to develop a multi-level error classification model based on delta-dosiomics features to precisely detect and quantify radiotherapy treatment errors, thereby ensuring the safety and precision of cervical cancer radiotherapy.

Materials/Methods:

A total of 52 cervical cancer patients who underwent VMAT were retrospectively included. Based on the original treatment plans and actually delivered clinical dose distributions, eight types of treatment errors with varying magnitudes were deliberately introduced (including setup treatment errors, random and systematic MLC errors, MU errors, gantry position errors, bladder filling variations, patient weight loss, and dose calculation algorithm differences), resulting in 1,984 delta-dosiomics datasets. Delta-dosiomics features were extracted using Python-based radiomics tools from four evaluation volumes: PTV, BODY, and the 10% and 50% prescription isodose volumes. The optimal feature subset was selected through Spearman correlation analysis, random forest-based feature importance assessment and recursive feature elimination (RFE). Four machine learning algorithms—SVM, RF, ANN, and XGBoost —were employed to develop an error classification model. The training set utilized 5-fold cross-validation across three levels: Level 1 (main error type), Level 2 (error direction), and Level 3 (error magnitude). The models were evaluated on the independent and clinical test sets, and their performance was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F1-score.

Results:

Feature selection analysis identified first-order statistical features (e.g., maximum, percentiles) as the most dominant predictors, followed by texture features derived from GLRLM and GLSZM. The XGBoost model, utilizing delta-dosiomics features derived from the 50% prescription isodose volume, demonstrated the most robust and stable performance across all classification levels. On the independent test set, this model achieved accuracies of 94.19% (AUC=0.998) for Level 1, 83.87% (AUC=0.993) for Level 2, and 74.84% (AUC=0.985) for Level 3. For the clinical test set, the model maintained high discriminative power, achieving accuracies of 72.81%, 64.75%, and 51.61% for Levels 1, 2, and 3, respectively, with corresponding AUCs of 0.945, 0.935, and 0.911.

Conclusion:

The dose-difference map in 3D dose verification serves as a unique “fingerprint”;delta-dosiomics characterizes different types of radiotherapy treatment errors. Combining delta-dosiomics with machine learning provides an effective complement to conventional gamma analysis. This approach helps medical physicists rapidly locate and quantify treatment errors, advancing patient safety and the efficacy of radiotherapy.