MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Machine Learning Based Respiratory Motion Prediction for MR-IGRT

A Curcuru1 , Gach1*, (1) Washington University in Saint Louis, Saint Louis, MO

Presentations

(Sunday, 7/14/2019) 4:00 PM - 4:30 PM

Room: Exhibit Hall | Forum 8

Purpose: To compare the performance of three algorithms for predicting respiratory motion for MRI-guided radiation therapy (MR-IGRT): 1) an artificial neural network (ANN); 2) a nonlinear autoregressive neural network (NAR); and 3) a long short term memory neural network (LSTM).

Methods: Motion traces were extracted from: 1) pencil-beam navigator echoes acquired in five volunteers undergoing arterial spin labeled MRI; and 2) respiratory bellows data acquired from 12 patients undergoing CyberKnife treatments. Training data was taken from three minutes of the CyberKnife dataset and two minutes of the navigator echo dataset. Both were resampled to 8 Hz corresponding to the anticipated MR-IGRT frame rate. Training data was normalized such that the data mean was zero and the standard deviation was unity. Testing data was normalized using the mean and standard deviation of the training data to aid in learning. Fifteen percent of the training data was used as validation data to check the model for overfitting. Each network was fed the 20 previous position time samples that were used to predict the next eight time points. Root mean square error (RMSE) was calculated for prediction horizons varying ranging from 0.1250 s to 1 s in intervals of 0.125 s.

Results: For the Cyberknife data, the RMSE was highest for the ANN and lowest for the LSTM over the range of prediction horizons. For the navigator echo data, the RMSE was highest for the NAR and lowest for the LSTM over the range of prediction horizons. However, input/output cross-correlation showed that the NAR frequently returned the input value resulting in a lagged copy of the original trace.

Conclusion: The LSTM consistently showed the best results of the three algorithms without the lag issues that the NAR demonstrated. Compounding errors resulted in poor predictions for the longer prediction horizons.

Funding Support, Disclosures, and Conflict of Interest: This research was conducted with the support of National Institutes of Health National Cancer Institute grant R01 CA159471.

Keywords

MRI, Patient Movement, Gating

Taxonomy

IM- MRI : Development (New technology and techniques)

Contact Email