Investigating the Robustness of Long Short-Term Memory Deep Neural Networks for Tumor Motion Tracking at External Surrogates Radiotherapy

Introduction: At radiotherapy, tumor motion is clinically tracked in real-time using external surrogates. To achieve this. A reliable correlation model is used to predict tumor coordinates based on the motion of external markers. In this work, a deep neural networks model is introduced for tumor motion tracking. Material and Methods: A motion database of 20 patients treated with the CyberKnife Synchrony System has been used to train and evaluate the model. The proposed model is based on Long-Short Term Memory neural network developed in a Python software package. The network consists of two layers, each with 40 neurons, and a fully connected layer with a linear activation function. Results: In this study, three-dimensional RMSE which is a common approach for calculating the model error is utilized. The obtained 3D RMSE of the proposed model is compared with the performance accuracy of the CyberKnife modeler. The results show a significant 15.3% reduction in three-dimensional error, indicating that our developed model has a lower error compared to the CyberKnife modeler. Conclusion: In this study, a model based on deep Long-Short Term Memory neural network is used for tumor tracking using a motion database of real patients. The reason for using this model is its robustness to remember information for a long period and its high predictive ability, which makes it promising for future clinical implementation. Unlike previously used models, this model can retain useful information from past time series and use it for training, allowing the model to outperform other models.