Author: Suman Gautam, Tianjun Ma, William Song 👨🔬
Affiliation: Virginia Commonwealth University 🌍
Purpose: We propose an artificial intelligence (AI)-based method to rapidly predict the patient-specific quality assurance (PSQA) results for magnetic resonance (MR)-guided online adaptive radiation therapy (ART) treatment plans.
Methods: A total of 111 intensity modulated radiation therapy adaptive plans from various anatomical sites were delivered and measured with an ArcCHECK device for PSQA. Gamma analysis was performed using 3%/2mm criteria, a 10% dose threshold, and a 90% passing rate. A fully-connected convolutional neural network was developed to predict the gamma passing rate results from the planned dose. The model was trained on 70% and validated on 20% of the data to solve regression and classification problems. To mitigate the likelihood of model overfitting, a dropout technique with different rates was implemented during the model training. The trained model performance was assessed on 10% of test set.
Results: The model achieved a mean absolute error of 4.62 ± 5.54%, a mean absolute percentage error of 0.05%, a median absolute error of 3.85%, and a maximum error of 16.7% when predicting the gamma passing rate values with 3%/2mm criteria. On the other hand, the model achieved 0.75 accuracy, 0.56 precision, 0.75 recall, and 0.64 f1-score in predicting the plan pass/fail results. The model also reports the probability of the passing/failing decision, offering clinicians greater interpretability and insight.
Conclusion: We developed an AI-based model for rapid prediction of PSQA results, which can speed up the MR-linac online ART workflow. With further performance improvement, the model has the potential for clinical deployment and could greatly reduce the necessity for measurement-based PSQA for every online ART treatment plan.