Real-time prediction of tumor motion using a dynamic neural network

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Computational Biology

Background:

• Respiratory motion significantly impacts radiotherapy accuracy by displacing tumors in the thoracic and abdomen cavities.
• System latency in radiotherapy equipment further challenges precise radiation dose delivery.
• Accurate tumor motion prediction is essential to compensate for latency and ensure targeted treatment.

Purpose of the Study:

• To investigate the efficacy of spatio-temporal and dynamic neural networks for predicting respiratory-induced tumor displacement.
• To evaluate different neural network designs for real-time tumor motion prediction.
• To assess the prediction accuracy and computational efficiency of the developed models.

Main Methods:

• Examined nine distinct neural network designs, including spatio-temporal and dynamic architectures.
• Focused on a prediction horizon of 665 milliseconds to capture rapid tumor movements.
• Evaluated models based on mean absolute error (MAE) and root mean square error (RMSE).

Main Results:

• A dynamic 35-to-3 neural network achieved the highest prediction accuracy.
• The best model yielded a mean absolute error of 0.54 ± 0.13 and a root mean square error of 0.57 ± 0.20.
• The proposed predictor model demonstrated independence from time-consuming processes like real-time retraining.

Conclusions:

• Dynamic neural networks show significant promise for real-time tumor motion prediction in radiotherapy.
• The developed model offers comparable or superior accuracy to existing methods.
• This approach can enhance the precision of radiation dose delivery by compensating for respiratory motion and system latency.