Room: ePoster Forums
Purpose: To extract quantitative image features from radiomics to analyze gamma distributions generated after patient specific quality assurance procedures and apply them to a neural network to predict the gamma analysis-based plans approval.
Methods: Twenty-six IMRT plans from head and neck and prostate patients and their QA dose distributions measured using MatrixX ionization chamber array, were analyzed using OmniPro Iâ€™mRT software. For each patient, we constructed gamma analysis maps considering the 2%/2mm criteria and plans with more than 95% of points approved in the gamma analyses were considered approved in the QA. Those gamma maps were exported as images to Slicer-3D software and then the 107 Radiomics features were extracted from the images using the Radiomics plugin. Those characteristics were putted in a neural network of ensemble bagged tree type available in Matlab software. This type of neural network combines different models outcomes to reach a final global characteristic. In this case, we combined the Radiomics features extracted from the gamma images trying to analyze if this neural network was able to predict the plan approval or not.
Results: For the 2%/2mm gamma analysis 11 plans were not approved on the QA and the other 15 plans were approved. The radiomics features data were applied to the model, in the confusion matrix we obtained an 88.5% accuracy and the under the curve area calculation (AUC) resulted in 0,895.
Conclusion: After the data analyses using that neural network, we could conclude that the model is suitable for our purpose. To complement this research and improve the neural network results, we intend to increase the number of patients and vary the treatment sites.
Not Applicable / None Entered.