Method and apparatus for reconstruction of 3D image volumes from projection images
The invention provides improvements in reconstructive imaging of the type in which a volume is reconstructed from a series of measured projection images (or other two-dimensional representations) by utilizing the capabilities of graphics processing units (GPUs). In one aspect, the invention configures a GPU to reconstruct a volume by initializing an estimated density distribution of that volume to arbitrary values in a three-dimensional voxel-based matrix and, then, determining the actual density distribution iteratively by, for each of the measured projections, (a) forward-projecting the estimated volume computationally and comparing the forward-projection with the measured projection, (b) generating a correction term for each pixel in the forward-projection based on that comparison, and (c) back-projecting the correction term for each pixel in the forward-projection onto all voxels of the volume that were mapped into that pixel in the forward-projection.
1. An imaging apparatus, comprising
an image acquisition device for generating a plurality of measured projection images of a volume,
a digital data processor coupled to the acquisition device for receiving the images, the digital data processor comprising a central processing unit (CPU) in communication with a graphics processing unit (GPU) including a plurality of programmable vertex shaders that are coupled to a plurality of programmable pixel shaders,
the CPU transferring at least some of the measured projection images to the GPU,
wherein the GPU reconstructs at least a portion of the volume based on comparison of a plurality of forward-projection images corresponding to successive iterations of an estimated density distribution assigned to that volume portion with one or more of the measured projection images without transfer of intermediate texture data to the CPU.
2. In the apparatus of claim 1 , the further improvement wherein the image acquisition device comprises a radiation source for illuminating the volume and at least a detector movable about the volume for detecting at least a portion of the illuminating radiation passing through the volume.
3. In the apparatus of claim 2 , the further improvement wherein one or more pixel shaders generate each forward-projection image by employing a projection matrix representing an asymmetric view frustum.
4. In the apparatus of claim 3 , the further improvement wherein the view frustum is defined by the origin of the radiation source and four corners of a plane of the detector.
5. In the apparatus of claim 1 , the further improvement wherein the volume portion comprises an intersection of cone beams corresponding to the measured projection images.
6. In the apparatus of claim 1 , the further improvement wherein the GPU classifies the measured projection images into two groups characterized by substantially orthogonal projection directions.
7. In the apparatus of claim 6 , the further improvement wherein the GPU processes the projection images in one group prior to processing the projection images in the other group.
8. In the apparatus of claim 1 , the further improvement wherein the CPU represents the volume as a voxel-based three-dimensional matrix.
9. In the apparatus of claim 6 , the further improvement wherein the CPU partitions the volume into a plurality of slices.
10. In the apparatus of claim 9 , the further improvement wherein one or more of the vertex shaders map, for each forward-projection image, a plurality of voxels that lie substantially at vertices of one or more selected polygonal portions of one of the slices to a respective pixel in that forward-projection image.
11. In an apparatus comprising a main processing section and a graphics coprocessing section including a plurality of programmable vertex shaders that are coupled to a plurality of programmable pixel shaders, a method for reconstructing a volume from a plurality of measured projection images, the improvement comprising:
transferring the projection images from the main processing section to the graphics coprocessing section, and
iteratively reconstructing the volume in the graphics coprocesssing section based on comparison of a plurality of forward-projection images corresponding to successive iterations of an estimated density distribution assigned to that volume with the measured projection images without transfer of intermediate texture data between the graphics coprocessing section and the main processing section.
12. In the method of claim 11 , the further improvement comprising partitioning the volume into a plurality of slices in the main processing section.
13. In the method of claim 12 , the further improvement comprising:
for generating each forward-projection image, performing the steps of:
with one or more of the vertex shaders, mapping a plurality of voxels that substantially lie at vertices of selected polygonal sections of one or more of the slices to a respective pixel of that forward-projection image, and
with one or more of the pixel shaders, generating an intensity for each pixel in the forward-projection image by forward projecting the voxels mapped to that pixel.