Benchmarking machine learning-based real-time respiratory signal predictors in 4D SBRT.

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Respiratory motion significantly impacts thoracic and abdominal tumor radiotherapy.
• External breathing signals are used as surrogates for tumor motion, but system latencies cause temporal lags.
• Respiratory signal prediction models are crucial for improving imaging and dose delivery by compensating for time delays.

Purpose of the Study:

• To compare six state-of-the-art machine and deep learning models for respiratory signal prediction.
• To evaluate real-time and real-world applicability of these models.
• To ensure reproducibility by providing open-source models and data.

Main Methods:

• Utilized 2502 clinical breathing signals for training, validation, and testing.
• Compared Linear, Dlinear, Xgboost, Lstm, Trans-Enc, and Trans-TSF models.
• Assessed prediction performance across 480, 680, and 920 ms horizons and analyzed robustness on out-of-distribution (OOD) signals.

Main Results:

• Lstm model demonstrated the lowest prediction errors across all horizons.
• Prediction accuracy decreased for out-of-distribution (OOD) signals, with Lstm and Trans-Enc showing less performance loss.
• Most models, except Trans-Enc, exhibited inference times suitable for real-time application.

Conclusions:

• The Lstm model offers the most accurate respiratory signal prediction.
• Simpler prediction models struggle with OOD signals due to limited signal history access.