The Quality Evaluation Of Radiotherapy Treatment Plans For Dual-Layer MLC Linac And Quality Assurance Of Treatment Plans Based On Machine Learning

VMAT treatment plans comparison between dual-layer MLC linac Halcyon2.0 and single-layer MLC linac Truebeam for different treatment sites:dosimetric quality and aperture complexityPURPOSE To study the differences in dosimetric quality and aperture complexity of VMAT plans based on two different linacs,Halcyon2.0 and Truebeam,for 4 different treatment sites:head&neck,chest,abdomen,and pelvis.METHODS 49 cases of Halcyon2.0 VMAT plans from four treatment sites were randomly selected,and the VMAT plans were re-designed based on Truebeam linac with the same optimization parameters.The differences in dosimetric metrics of the PTV and OARs of the two types of plans were compared and analyzed,as well as the differences in aperture complexity metrics,with P<0.05 means statistically significant.RESULTS In terms of dosimetry metrics,compared with Truebeam VMAT plans,Halcyon2.0 VMAT showed better homogeneity index(HI)and conformity index(CI)in the head&neck plans and chest plans.But Truebeam VMAT plans showed better gradient index(GI)in the abdominal plans and pelvic plans(P<0.05).For chest plans,Halcyon2.0 VMAT plans showed lower D20%,Dmean of the double lung,and Dmean of the heart.In terms of aperture complexity metrics,the median small aperture scores(SAS)of the Halcyon2.0 VMAT plans in all treatment sites were smaller than those of the Truebeam VMAT plans,and the median plan average beam area(PA)of the Halcyon2.0 VMAT plans were larger than those of the Truebeam VMAT plans(P<0.05).CONCLUSION Compared with VMAT plans based on Halcyon2.0 and Truebeam linac,Halcyon2.0 VMAT plans have similar or even better dosimetric quality.However,the Halcyon2.0 VMAT plans showed lower aperture complexity,which made it an advantage in clinical application.Patient-specific quality assurance prediction models based on machine learning for linac equipped with novel dual-layer MLCPURPOSE To evaluate the performance of machine learning models in predicting gamma passing rates(GPRs)of fixed field intensity-modulated radiation therapy(FFIMRT)plans for linac equipped with dual-layer MLC.METHODS AND MATERIAL A total of 213 FF-IMRT treatment plans were selected,and 33 complexity metrics were extracted and calculated for each beam.Gamma analysis was performed by portal dosimetry(PD)verification using 1%/1mm,2%/2mm,and 3%/2mm criteria with a 10%threshold.Machine learning models’ training set(TS)consisted of 1106 beams,and an independent evaluation set(ES)consisted of 277 beams.Three machine learning algorithms(Gradient Boosting Decision Tree/GBDT,Random Forest/RF,and Poisson Lasso/PL)were used to build models and predict GPRs.To improve prediction accuracy in the rare samples with low measured GPRs,the study performed a weighted method on the training set.Using the permutation importance method to analyze the importance of plan complexity metrics.RESULTS The GBDT model had the best performance in this study.In ES,the minimal mean absolute errors for prediction results were 1.93%,1.16%,and 0.78%under 1%/1mm,2%/2mm,and 3%/2mm criteria,respectively,from the GBDT model.After using the weighted method,the prediction accuracy of the machine learning models for the rare samples with low measured GPRs was improved.In contrast,overall prediction errors slightly worsened depending on the model type.For feature importance,two treebased models(GBDT and RF)had in common the top 10 most important metrics and the same metric(p-C/A)with the largest impact.CONCLUSION The machine learning models developed on the GBDT algorithm show a certain degree of accuracy for GPRs prediction for linacs equipped with novel duallayer MLC.Using the weighted method can improve the prediction accuracy of machine learning models for samples with low measured GPRs to a certain extent.The specialized machine learning model for dual-layer MLC configuration could be useful for physicists identifying PSQA measurement failure.