Optimizing the Accuracy and Training Speed of An Artificial Neural Network for Three-Dimensional Tumour Trajectory Prediction in Linac-MR Image-Guided Radiation Therapy
N W Johnson1*, B G Fallone1,2, J Yun1,2, (1) University of Alberta, Edmonton, AB, CA, (2) Cross Cancer Institute, Edmonton, AB, CA
Presentations
PO-GePV-M-194 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: To predict 3D tumor trajectories during respiratory motion using an artificial neural network (ANN) for applications to real-time IGRT using a hybrid linac-MR, and to improve the accuracy and training speed of the ANN through hyperparameter optimization, adaptive learning, data augmentation, ensemble learning and GPU acceleration.
Methods: External surrogate estimates of 3D internal thoracoabdominal tumor trajectories during normal respiration are obtained from Suh et al. (Phys. Med. Biol. 53 3623 (2008)). Long short-term memory recurrent neural networks (LSTM-RNNs) with a range of structures and hyperparameter selections are trained on a 30s subset of data obtained at the start of each fraction, and their accuracy in predicting the tumor position at a forecasting delay of 200ms is evaluated for the following 450s of motion.
Results: Preliminary results indicate that shallow, narrow LSTM-RNNs can perform as well as or better than the larger LSTM-RNNs employed previously in this work, significantly reducing computational expense and training time. Across all fractions, the mean absolute error at a 200ms forecasting time is found to be 0.40 mm, increasing by a factor of 1.5 between 0-30s post-training and 420-450s post-training. Early experiments with ensemble learning indicates that improvements up to 30% are attainable when the output of 5 randomly initialized models are averaged, and since the individual networks are small, large ensembles may be trained simultaneously on a single GPU.
Conclusion: Sub-mm mean absolute error (30s training, 450s testing, 240ms input intervals, 200ms forecasting interval) is achievable with small LSTM-RNNs when applied to a dataset of 3D thoracoabdominal tumor trajectories during normal respiratory motion. This suggests that computationally inexpensive, patient- and fraction-specific ANNs may be practically trained on-demand prior to each fraction of IGRT on a linac-MR.
Funding Support, Disclosures, and Conflict of Interest: Funding: The Cathy and Harold Roozen Endowment Entrance Award (Johnson); CIHR Project Grant (Yun). Disclosures: Stock, MagnetTx Oncology Solutions (Yun); Stock, CEO, Board Chair, MagnetTx Oncology Solutions (Fallone)
ePosters
Keywords
Taxonomy
IM/TH- MRI in Radiation Therapy: MRI/Linear accelerator combined- IGRT and tracking
Contact Email