MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Design of a Deep Learning Algorithm to Track Intrafractional Lung Tumor Motion Via the On-Board KV Imaging System

C Wang*, M Hunt , L Zhang , A Rimner , E Yorke , D Lovelock , X Li , T Li , G Mageras , P Zhang , Mem Sloan-Kettering Cancer Ctr, New York, NY

Presentations

(Tuesday, 7/16/2019) 1:45 PM - 3:45 PM

Room: Stars at Night Ballroom 2-3

Purpose: To design a deep learning algorithm (T-net) that tracks 3D positions of lung tumor in realtime using a continuous feed of kV images acquired via the on-board imager during CBCT acquisition and VMAT delivery.

Methods: On an IRB-approved lung free-breathing protocol, kV projections of the setup CBCT and fluoroscopic images acquired during VMAT were collected in realtime. Concurrently, Calypso system recorded motion traces of three electromagnetic transponders implanted in or near the tumor. T-net was designed to utilize a convolutional neural network for extracting relevant features of the kV images around the tumor, followed by a recurrence neural network for analyzing the temporal patterns of the moving features. T-net was trained on the simultaneously collected kV images and Calypso traces, subsequently utilized to calculate realtime motion traces solely based on the continuous feed of kV images. To enhance performance, T-net was also facilitated by frequent calibrations (every 10° gantry rotation) derived from cross-correlation based registrations between kV images and planning 4DCT. T-net was validated on a leave-one-out strategy using data from seven lung patients, including 4000 kV images. The root mean square error between the T-net and Calypso traces were calculated to evaluate the localization accuracy.

Results: 3D displacements relative to setup shown in the Calypso traces were 3.4±1.7mm. Compared to Calypso traces, the 3D tracking accuracy of T-net with calibration was 1.0±1.2mm, vs 4.3±3.7mm without calibration. T-net had an accuracy of 86±8% in determining whether the motion was within a 3D displacement window of 2mm. The latency was 20ms when T-net ran on a high-performance computer cluster.

Conclusion: T-net is able to provide realtime tracking of lung tumors without frequent interruptions caused by poor image contrast in a conventional cross-correlation based registration approach, and has the potential to remove the reliance on the implanted fiducials.

Funding Support, Disclosures, and Conflict of Interest: Research is partially supported by Varian Medical System

Keywords

Not Applicable / None Entered.

Taxonomy

TH- External beam- photons: Motion management (intrafraction)

Contact Email