352 - Virtual Monoenergetic CBCT Volumes Using a Clinical Dual-Layer Imager

Purpose/Objective(s): The on-board cone beam CT (CBCT) imagers of external beam treatment systems commonly provide single-energy imaging for therapy guidance. A dual-layer detector, however, could collect two spectral channels of x-ray measurements, bringing dual-energy CBCT capability to the treatment room. A benefit to this is the ability to generate virtual mono-energetic image (VMI) volumes over a range of energies. By examining how tissue attenuation changes with energy, better contrast between anatomical regions of interest might be obtained, which could aid on-treatment recontouring, image registration, and other adaptive radiotherapy (ART) tasks. A difficulty, though, is that energy separation between the spectral channels is narrower with dual-layer scans than with other acquisition techniques. This can amplify VMI noise at very high/low virtual energy selections.

Materials/Methods: A prototype dual-layer kV imager (DLI) has been installed on a clinical medical linear accelerator. To obtain low-noise VMIs, we have designed an algorithm that uses deep learning (U-Net) models in conjunction with physics-based post-processing. The U-Nets transform the acquired DLI projections directly to mono-energetic x-ray measurements at 55 and 100 keV, from which VMIs are reconstructed. A fast, iterative physics-based algorithm is then applied, which derives material basis volumes from the two VMIs. With the basis volumes, VMIs at arbitrary energies can then be generated. The algorithm is regularized with a novel, edge-preserving noise penalty to help maintain image contrast-to-noise (CNR) ratios at extreme energies. The technique was tested in a head & neck patient scan and compared to a baseline material decomposition method based on table-lookup. Performance was quantified by generating VMIs over a range of virtual energies and examining CNR between regions of interest in adipose, muscle, and spinal trabecular bone.

Results: Between muscle and adipose, the proposed method gave at least twice the CNR of the baseline method over the energy range from 30 to 150 keV. The accompanying table shows data up to 100 keV. Between muscle and trabecular bone, our method achieved at least the same CNR as the baseline, and significantly higher CNR at low virtual energies where the CNR of standard CBCT (~10) was also well-exceeded. Conversely, the CNR of the baseline method deteriorated at lower energies.

Conclusion: We have demonstrated an effective technique for increasing CNR in VMIs generated from an actual, clinical dual-layer kV-imager. This stands to benefit ART tasks such as on-treatment recontouring, as well as registration with pre-treatment imaging.