Prediction and classification of gamma passing rate for patient-specific quality assurance using machine learning models based on radiomics features and beam parameters

The aim of this work is to predict and classify the gamma passing rate (GPR) values for intensity-modulated radiation therapy plans at the pelvis site utilizing radiomics features and beam features combined with machine learning. Dosimetric verification of 486 fields was performed using the portal dosimetry system. Three types of models were constructed using support vector machines: radiomics models based on radiomics features derived from fluence images, beam models based on beam parameters related to dose delivery accuracy, and hybrid models that integrated both feature sets. For the radiomics, beam, and hybrid models, the mean absolute errors in the test set were 1.62 %, 1.61 %, and 1.45 % at the 2 %/2 mm criterion, and 1.09 %, 1.18 %, and 1.02 % at the 3 %/2 mm criterion, respectively. Similarly, for classification models, the area under the curve values were 0.80, 0.76, and 0.83 for 2 %/2 mm, and 0.79, 0.74, and 0.82 for 3 %/2 mm, respectively. Moreover, radiomics features, particularly the first-order statistics, contributed more significantly than beam features in hybrid models. In conclusion, both radiomics and beam features showed promising value in predicting and classifying the GPR, while the hybrid models achieved the best performance, potentially improving plan quality and reducing quality assurance workload.