Cone Beam CT Reconstruction

Imaging dose reduction in cone beam CT (CBCT) is an important topic for image guided radiation therapy. In contrast to conventional approaches that employ novel reconstruction techniques to retrieve a CBCT image from data acquired at a low-exposure setting, this project uses temporal correlations of CBCT images for a patient at different treatment fractions for dose reduction.

We created a progressive dose-control scheme that maximally employs prior information from existing CBCT images. It uses image-processing techniques to maintain the quality of a newly acquired CBCT image at a low dose level. With this approach, it becomes possible to use a gradually reduced exposure along a treatment course while achieving a constant image quality (see figure). This scheme allows significantly reduced overall dose in a radiotherapy course compared to a conventional approach, which uses the same exposure level at each fraction.

4D cone beam CT (4D-CBCT) provides phase-resolved volumetric imaging for lung radiotherapy. Conventional reconstruction approaches from image intensity domain are susceptible to the problem of low projection number due to phase binning. We propose to use a new reconstruction approach via motion-vector optimization to overcome this problem and achieve high image quality. This method uses a prior image acquired from treatment planning stage and optimizes a deformation vector field for each phase, such that the projections of the deformed volumetric image match the measurements. A novel algorithm is developed to solve the optimization problem. High-quality 4D-CBCT imaging based on the clinically standard 1-minute 3D CBCT scanning protocol is feasible via the proposed hybrid reconstruction algorithm.

While 4D-cone beam CT provides phase-resolved volumetric images for lung region, it assumes periodic motion of the anatomy and hence neglect irregular motions. To overcome this problem, we develop novel real-time volumetric imaging approach using machine-learning approach. Due to the limited amount of measurements, it is impossible to reconstruct an instantaneous volumetric image from conventional CT reconstruction perspective. Instead, we employ machine-learning methods, such as sparse learning and manifold learning, to build a model that relates respiration related measurements (x-ray projection, surface imaging, etc) to the volumetric image. Our method is able to automatically identify the most relevant data to predict an instantaneous volumetric image with clinically acceptable accuracy.