Resnet Deep Learning Convolutional Neural Networks for Halcyon Based Portal Dosimetry 2D Dose Plane Gamma Passing Rate Predictions
Abstract
Purpose
Developing and researching methods of implementing deep learning neural networks capable of predicting gamma passing rates from retroactive 2D dose plane images for QA procedure optimization, workflow improvements, and plan delivery.
Methods
A dataset built from 184 DICOM files collected from 87 patients that underwent IMRT and FSRT prostate cancer procedures on a Halcyon Linear Accelerator was created in conjunction with the use of ResNet architecture to act as a feature extractor. Dose normalised data was loaded and preprocessed into an appropriate format to match the expected inputs for the used ResNet model, and an 80/20 training validation split was used. ResNet layers/weights were frozen to prevent overfitting from a small sample size by limiting changeable parameters, while loss and optimisers were chosen to prevent large errors and optimal regression. The training loop then looked at small batches of data and optimized for regression loss per epoch and for evaluation, validation, and backpropagation. Fine-tuning is then done by unfreezing layers and experimenting with parameters and learning rates. The results are then measured in mean absolute error and root mean squared error, and graphed to show predicted gamma passing rates vs the actual gamma passing rates.
Results
Error metrics of around 4 percent mean absolute error and 5 percent root mean squared error are found after training and validation, while avoiding overfitting and keeping complexity. Our coefficient of determination, R-squared, varied around zero with noise.
Conclusion
There is potential in the use of Deep Learning for predicting 2D dose plane GPR, although further testing and improvements are needed. Experimenting with weights, hyperparameters, learning rate, larger datasets, and testing higher-layer models of ResNet, such as ResNet50 or ResNet34 are likely to increase the performance of the model and reach a higher correlation.