Systematic method for a deep learning‐based prediction model for gamma evaluation in patient‐specific quality assurance of volumetric modulated arc therapy

Abstract

Purpose This study aimed to develop and evaluate a novel strategy for establishing a deep learning‐based gamma passing rate (GPR) prediction model for volumetric modulated arc therapy (VMAT) using dummy target plan data, one measurement process, and a multicriteria prediction method. Methods A total of 147 VMAT plans were used for the training set (two sets of 48 dummy target plans) and test set (51 clinical target plans). The dummy plans were measured using a diode array detector. We developed an original convolutional neural network that accepts coronal and sagittal dose distributions to predict the GPRs of 36 pairs of gamma criteria from 0.5%/0.5 mm to 3%/3 mm. Sixfold cross‐validation and model averaging were performed, and the mean training result and mean test result were derived from six trained models that were produced during cross‐validation. Results Strong or moderate correlations were observed between the measured and predicted GPRs in all criteria. The mean absolute errors and root mean squared errors of the test set (clinical target plan) were 0.63 and 1.11 in 3%/3 mm, 1.16 and 1.73 in 3%/2 mm, 1.96 and 2.66 in 2%/2 mm, 5.00 and 6.35 in 1%/1 mm, and 5.42 and 6.78 in 0.5%/1 mm, respectively. The Pearson correlation coefficients were 0.80 in the training set and 0.68 in the test set at the 0.5%/1 mm criterion. Conclusion Our results suggest that the training of the deep learning‐based quality assurance model can be performed using a dummy target plan.

Applications of machine and deep learning to patient‐specific IMRT/VMAT quality assurance

“…The models obtained a MAE of 4.2% in predicting the gamma passing rate at 2%/2 mm criteria. Tomori et al 50 examined a CNN model trained on features derived from dose distribution images to predict the gamma passing rates for patient‐specific QA measurements using a 2D array detector. The model achieved a prediction performance of MAE of 0.63%.…”

“…For instance, the front‐end layers in the CNN encode low‐level features in the image common to most computer vision applications, whereas the subsequent layers learn high‐level features which are more application‐specific. Few studies 36 , 39 , 50 , 52 implemented feature‐less approach in training ML/DL models for patient‐specific IMRT/VMAT QA outcome predictions. CNN features extracted from dose distribution, 50 dose difference, 52 fluence map, 36 and gamma map images 39 , 52 were applied to train various ML/DL models.…”

“…Few studies 36 , 39 , 50 , 52 implemented feature‐less approach in training ML/DL models for patient‐specific IMRT/VMAT QA outcome predictions. CNN features extracted from dose distribution, 50 dose difference, 52 fluence map, 36 and gamma map images 39 , 52 were applied to train various ML/DL models. Results have shown that using the feature‐less approach for training the models provide comparable or higher prediction performance compared to using radiomics features derived by the hand‐crafting approach 39 or complexity metrics.…”

“…Eleven studies have implemented ML/DL models for IMRT/VMAT QA gamma passing rate predictions (Tables 1 and 2 ) including Poisson regression, 33 , 34 , 45 XGBoost, 38 , 49 AdaBoost, 38 , 48 random forest, 38 , 48 regression tree analysis, 46 multiple regression analysis, 48 and ANN, 46 , 48 , SVMs, 48 decision tree, 48 gradient boosting, 48 CNNs, 35 , 36 , 50 and DNNs. 51 All the models were trained on datasets consisting of multiple treatment sites, except the study by Tomori et al 35 where the dataset was site‐specific (prostate).…”

“…In this review, model validation was performed using the k ‐fold cross‐validation 33 , 35 , 36 , 38 , 40 , 42 , 45 , 47 , 48 , 49 , 50 , 51 , 52 and the hold‐out validation 37 , 39 , 41 , 43 , 44 , 46 techniques. In k ‐fold cross‐validation, the dataset is equally partitioned into k subparts or folds.…”