Real-time respiratory motion prediction using photonic reservoir computing.

Area of Science:

• Medical Physics
• Computational Neuroscience
• Biomedical Engineering

Background:

• Respiration-induced motion poses significant challenges in clinical applications like upper body imaging, lung tumor tracking, and radiation therapy.
• Accurate prediction of respiratory motion is crucial for improving treatment efficacy and patient safety.

Purpose of the Study:

• To develop and evaluate a real-time algorithm for predicting respiratory motion using a photonic delay-line reservoir computer (RC).
• To assess the effectiveness of the RC in handling non-linear distortions inherent in respiratory signals.
• To enable personalized, real-time motion prediction for individual patients.

Main Methods:

• Implementation of a recurrent neural network algorithm within a photonic delay-line reservoir computer (RC).
• Utilization of a double-sliding window technology for real-time patient-specific model training and data processing.
• Investigation of prediction accuracy across various look-ahead times (66.6, 166.6, 333 ms) using a dataset from 76 patients.

Main Results:

• The RC model demonstrated effective short-to-medium range respiratory motion prediction within practical timescales.
• At a 333 ms look-ahead time, the model achieved an average Normalized Mean Square Error (NMSE) of 0.025 and Mean Absolute Error (MAE) of 0.34 mm.
• High therapeutic beam efficiency (TBE) was observed, with 94.14% accuracy for absolute error < 1 mm and 99.89% for absolute error < 3 mm.

Conclusions:

• Real-time reservoir computing (RC) provides an efficient and high-precision framework for predicting respiratory motion.
• The developed algorithm shows significant potential for enhancing accuracy in image-guided radiation therapy and other clinical practices affected by breathing motion.
• This technology facilitates personalized, real-time adaptation to patient-specific breathing patterns.