Comparison of initial learning algorithms for long short-term memory method on real-time respiratory signal...

Area of Science:

• Medical Physics
• Machine Learning
• Signal Processing

Background:

• Accurate respiratory signal prediction is crucial for image-guided radiation therapy.
• Long short-term memory (LSTM) models show promise for real-time respiratory motion prediction.
• The choice of weight initializers can impact LSTM model performance.

Purpose of the Study:

• To evaluate the impact of different weight initializers on LSTM-based respiratory signal prediction.
• To compare the predictive performance of Glorot, He, Orthogonal, and Narrow-normal initializers.

Main Methods:

• Respiratory signals were collected using the CyberKnife Synchrony device across 304 breathing motion traces.
• LSTM models were trained and evaluated using four distinct weight initializers.
• Prediction performance was quantified using normalized root mean square error (NRMSE).

Main Results:

• The He initializer demonstrated superior performance in respiratory signal prediction.
• He initializer achieved a mean NRMSE 7.5% to 11.3% better than other tested initializers at a 385-ms prediction horizon.
• Confidence intervals for NRMSE varied across initializers, with He showing a narrower range.

Conclusions:

• The He initializer is a valuable choice for enhancing LSTM model performance in respiratory signal prediction.
• This finding supports the use of the He initializer for improving real-time motion tracking in radiotherapy.