2073 - Orthogonal MRI Motion Prediction to Mitigate Latencies in MRI-Guided Radiotherapy

02:30pm - 04:00pm PT

Hall F

Screen: 15

POSTER

Presenter(s)

John Ginn, PhD - Duke University School of Medicine, Durham, NC

J. Ginn, Z. Wu, C. Wang, and D. Yang; Duke University, Durham, NC

Purpose/Objective(s): MRI-guided radiotherapy image acquisition, reconstruction and gating systems create a latency between the time the target moves and when the radiation is gated. Motion prediction algorithms may help minimize system latency by preemptively gating the radiation delivery before the target moves. We apply an image regression prediction technique to images acquired in orthogonal planes and hypothesize the algorithm will provide a more accurate estimate of motion than assuming no motion occurs between image acquisitions. The primary advantage of this algorithm is that it can readily be integrated with a previously published manifold alignment technique to predict motion in both planes simultaneously.

Materials/Methods: Five healthy volunteers were scanned under an IRB approved protocol using a Siemens Skyra simulator. Images of the liver were acquired for 10 min in alternating sagittal and coronal planes with a 2.6 mm x 2.6 mm in-plane resolution. The prediction method is based on a dimensionality reduction technique known as locally linear embedding and estimates the future motion based on a weighted combination of similar motion states in the training set. The estimates were compared to the ground-truth motion derived in each newly acquired image using deformable registration. All motion was predicted in the same imaging orientation one frame (~0.5 s) in the future. The method was compared to assuming no motion occurred between image acquisitions and a linear extrapolation of motion based on the two most recently acquired images. A simulated target was defined on the dome of the liver and used to evaluate model performance. The Dice similarity coefficient and distance between the predicted and ground-truth contour centroids were evaluated.

Results: The average Dice coefficient and centroid distance between the image regression predictions and ground-truth target contours were 0.93 and 0.96 mm respectively across all volunteer studies. The average results for motion extrapolation were 0.84 and 2.29 mm respectively whereas assuming no motion occurred yielded 0.85 and 2.27 mm respectively. The results for individual volunteers are reported in Table 1. A paired t-test indicated the difference between the image regression prediction and other methods was statistically significant.

Conclusion: The healthy volunteer studies indicate the image regression predictions may mitigate system latencies. On average, the predictions were < 1 mm and yielded a Dice > 0.90 compared to the ground-truth motion.

Abstract 2073 - Table 1: The average Dice coefficient and centroid distance using the image regression, motion extrapolation and no motion assumptions for all volunteer studies

Volunteer		1	2	3	4	5
Dice (AU)	Image Regression	0.94	0.93	0.96	0.93	0.91
	Extrapolation	0.86	0.79	0.89	0.85	0.82
	No Motion	0.89	0.81	0.90	0.86	0.82
Centroid Distance (mm)	Image Regression	0.82	0.97	0.67	1.04	1.33
	Extrapolation	1.83	2.84	1.76	2.36	2.68
	No Motion	1.59	2.70	1.73	2.33	2.98