Room: Exhibit Hall | Forum 6
Purpose: Treatment errors in radiation oncology occur at a rate of approximately 2% per patient and a safeguard is urgently needed. The goal is to develop and validate a treatment plan classifier that can automatically identify the treatment site based on plan characteristics using support vector machine (SVM). This can be used as an additional safety layer for safely delivery radiation treatment.
Methods: A multi-class Support Vector Machine (SVM) was developed to classify the treatment plans. Data of 14 head-neck, breast, and prostate cancers treatment plans were collected in this study. Though the data structure is very complex containing hundreds of possible variables, 2 features from treatment plans were extracted that summarize the effects of all variables including machine gantry angle and leaf variations during the treatment. These two features were taken as inputs to the SVM model. Eleven plans were used to train the SVM model. One hundred combinations (5x5x4) of mixed treatment plans were used to test the model. A test is called successful if the model can individually and correctly identify three testing sets. A test is failed if any one of three testing data is misidentified.
Results: The multi-class SVM successfully identified all three testing sets with a 100% success rate with a 0% failure rate. For the clinical application, if an incorrect treatment plan were offered, SVM would identify and warn that the pending treatment plan is inconsistent with the intended use, and flag the treatment plan.
Conclusion: The preliminary results are very encouraging; they demonstrated that machine learning algorithms can serve as appropriate tools in treatment plan classification and provide a safeguard for radiation therapy treatment.