Room: AAPM ePoster Library
Purpose: investigate the potential of using AI-based dual energy (DE) imaging to improve markerless lung tumor tracking with an on-board kV imaging system.
Methods: convolutional neural network (CNN) was trained to produce low energy from high energy projection images, allowing AI-DE images to be created using log subtraction. The CNN model was trained on simulated forward projections (60 and 120 kVp) using 1000 CTs. A different set of 84 4D-CTs with known GTV contours were used to simulate projections in the same way. A sequence of projections was created from simulated projections of a 4D-CT at a given gantry angle for three imaging modes single energy (SE), DE, and AI-DE, respectively. A machine-learning-based algorithm was applied to track the GTV motion in the sequence. Tracking accuracy was evaluated against ground truth motion (4D-CT) using precision curves representing the percentage of correctly tracked GTV center-of-mass (COM) for a range of distances and overlap (Dice coefficient) thresholds.
Results: area-under-the-curve (AUC) scores of COM precision curves normalized to a distance threshold of 12mm (30 pixels) were 0.72, 0.75, 0.74 for SE, DE, and AI-DE, respectively. The scores of Dice precision were 0.90, 0.91, 0.91 for SE, DE, and AI-DE, respectively.
Conclusion: the performance of GTV tracking in SE was promising, both DE and AI-DE further improves the tracking accuracy with comparable performance. The use of AI-DE imaging may provide a cost-effective method for providing DE capacity on any linear accelerator with an on-board kV imaging system.
Funding Support, Disclosures, and Conflict of Interest: Research reported in this publication was partially supported by the National Cancer Institute of the National Institutes of Health under Award Number R01CA207483. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.