[Prediction of respiratory motion based on sequential embedding combined with relational embedding].

Area of Science:

• Medical physics
• Artificial intelligence
• Biomedical engineering

Context:

• Accurate modeling of respiratory motion is crucial for image-guided radiotherapy.
• Existing methods often struggle with the complex, multi-directional nature of respiratory movements.

Purpose:

• To develop and validate a novel deep learning model for predicting respiratory motion in all directions.
• To improve the accuracy and efficiency of radiotherapy by reducing system delays associated with motion prediction.

Summary:

• A deep learning model integrating Long Short-Term Memory (LSTM) and self-attention mechanisms was developed to capture sequential and relational dependencies in respiratory motion.
• The model combines sequential and relational embeddings with an autoregressive module for accurate motion prediction.
• Trained using a pre-training and fine-tuning approach, the model demonstrated superior performance over existing methods in validation experiments.

Impact:

• The proposed model achieved over 70% reduction in absolute deviations across 3D directions for respiratory motion prediction.
• This advancement has the potential to significantly reduce system delay in precise radiotherapy, improving treatment outcomes.