Predicting Gamma Passing Rates for Portal Dosimetry-Based IMRT QA Using Deep Learning

• Published in: International journal of radiation oncology, biology, physics Vol. 111; no. 3; pp. e110 - e111
• Main Authors: Fondevila, D.M., Rios, P.J., Peñalva, D.M. Dodat, Arbiser, S.
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.11.2021
• ISSN: 0360-3016; 1879-355X
• DOI: 10.1016/j.ijrobp.2021.07.515

To assess if portal dosimetry-based IMRT QA gamma passing rates can be predicted with a Convolutional Neural Network (CNN) Model trained directly on the fluence maps without any domain expert knowledge. A total of 4268 beams were collected from various treatment site plans and analyzed using gamma criteria of 3%/3 mm local normalization with a 10% threshold. A Transfer Learning Scheme was employed with different Imagenet pre-trained CNN architectures (VGG16 and ResNet18) using five-fold cross-validation to tune and evaluate the models. An ensemble of the best performing CNN models was also created. A linear relationship was found between the measured and predicted values in all of the models. The Pearson Correlation Coefficient and the Mean Absolute Error (MAE) with their respective 95% Confidence Intervals can be seen for the validation set in Table 1 below. These results demonstrate a strong or moderate correlation between the predicted and measured values in all of the models. Even though the global metrics are quite acceptable, it can be seen that the models perform worse in the range of measured gammas below 95%. It was hypothesized that this was due to the problem of an imbalanced dataset (measured gammas < 95% summed up only 26% of the total dataset). State-of-the-art techniques for counteracting this problem, such as, synthetic oversampling of the measured gamma < 95% maps through rotational and translational transformations, and stratified and unstratified under sampling of the measured gammas > 95% were carried out with unfavorable results. Further investigation of this issue is warranted. Nonetheless it must be noticed that carefully characterized models even with a medium performance can be of great value for either safety enhancement as a redundant second check to measurements or workload reduction scheme for triaging highly confident passing maps. For example, the subclass of all maps with a prediction value above 98% with our ensemble model can be shown with a high level of confidence of having the measured gamma values passing a 95%-mark threshold. In our case this equates to over 40% of all maps. It is not uncommon in proper working facilities to have highly skewed distribution of results with the overwhelming majority of points well within the acceptance criteria. We have shown that CNN Models can predict portal dosimetry-based IMRT QA gamma passing rates training directly on the fluence maps without any domain expert knowledge. These predictive modeling methodologies can be highly beneficial if thoughtfully adopted following institutional criteria.