Author: Amanda J. Deisher, Andrew YK Foong, Witold Matysiak, Jing Qian, Xueyan Tang, Erik J. Tryggestad, Mi Zhou 👨🔬
Affiliation: Mayo Clinic 🌍
Purpose: Phase gating is commonly employed to mitigate the impact of tumor motion in radiotherapy. Due to the machine-specific time delay between triggering and radiation delivery, the triggering signal must be sent in advance, requiring accurate prediction of the breathing phase within a specific time window. This task is particularly challenging because patients exhibit varying breathing cycles, amplitudes, and levels of consistency. Existing vendor algorithms have demonstrated low accuracy, often leading to incorrect delivery timing and frequent manual interventions. In this study, we developed and tested several deep learning (DL) algorithms to enhance the accuracy of breathing phase prediction.
Methods: We retrospectively collected breathing traces sampled at 30 Hz from patients treated with proton therapy at our institution. The phases were retrospectively processed to establish ground truth. Breathing amplitudes were normalized in each lookback window of 256 points (8.5 seconds) that was utilized as input to predict the phase 330 ms ahead. Models with varying complexity were trained using mean squared error (MSE) loss and evaluated for accuracy after binning phase values into 10 categories (1–10).
Results: A total of 30,881 seconds and 5,633 seconds of breathing trace data was collected for training and validation, respectively. Another 16,679 seconds of data, from patients not included in the training, were used for testing. The Long Short-Term Memory (LSTM) model achieved an accuracy of 90.5%, significantly outperforming the vendor's algorithm, which achieved 42.6% accuracy. In comparison, simpler linear models achieved 30.6% accuracy, while the more complex transformer model reached 69.6%. All models demonstrated inference times of less than 1 millisecond.
Conclusion: The DL algorithms achieved significantly higher accuracy in phase prediction, enhancing the precision of radiation delivery. Notably, model performance did not consistently correlate with model complexity or size. Currently we are working with engineering team to implement the model clinically.