Model-free Cardiorespiratory Motion Prediction from X-ray Angiography Sequence with LSTM Network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Cardiorespiratory motion significantly impacts X-ray angiography image quality and accuracy.
• Accurate prediction of coronary vessel displacement is crucial for precise 2D X-ray registrations.
• Existing methods may struggle with the complex and irregular nature of cardiorespiratory motion.

Purpose of the Study:

• To develop a novel model-free approach for predicting cardiorespiratory motion from X-ray angiography time series.
• To estimate future displacement of coronary vessels using Long Short-Term Memory Recurrent Neural Networks (LSTM-RNN).
• To enable accurate 2D X-ray registrations by predicting displacement parameters.

Main Methods:

• Utilized Long Short-Term Memory Recurrent Neural Networks (LSTM-RNN), a type of Recurrent Neural Network (RNN) adept at sequential data.
• Represented vessel displacement as a sequence of 2D affine transformation matrices.
• Validated the method using simulated data from the realistic XCAT cardiorespiratory motion simulator.

Main Results:

• The LSTM-RNN model demonstrated rapid convergence.
• The method successfully predicted complex and irregular cardiorespiratory motion in angiography sequences.
• Achieved mean prediction errors of 0.29 mm for combined motion, 0.51 mm for cardiac motion, and 0.44 mm for respiratory motion.

Conclusions:

• The proposed model-free LSTM-RNN approach offers an effective solution for cardiorespiratory motion prediction in X-ray angiography.
• This method holds potential for improving the accuracy and reliability of 2D X-ray registrations in clinical practice.
• The model's ability to handle complex, irregular motions suggests broad applicability in cardiovascular imaging.