Adaptive image filtering for volume reconstruction using partial image data
A method of generating an image synthesis process is disclosed, where the image synthesis process improves image quality of degraded volumetric images. In the method, a machine learning process is trained in a supervised learning framework as the image synthesis process. In the supervised learning process, a lower-quality partial-data reconstruction of a target volume is employed as an input object in the supervised learning process and a higher-quality full data reconstruction of the target volume is employed as an expected output. The full data reconstruction is generated based on a first set of projection images of the three-dimensional volume and the partial-data reconstruction is generated based on a second set of projection images of the three-dimensional volume, where the second set of projection images includes projection images that have less image information and/or are of a lower image quality than the first set of projection images.
1. A computer-implemented method of training an image synthesis process, the method comprising:
receiving a first set of projection images of a target volume;
generating a second set of modified projection images of the target volume, wherein the second set of modified projection images of the target volume includes less image information associated with the target volume than the first set of projection images;
generating a full data reconstruction of the target volume based on the first set of projection images;
generating a partial data reconstruction of the target volume based on the second set of modified projection images;
modifying the image synthesis process based on the full data reconstruction and the partial data reconstruction;
generating a synthesized reconstruction of the target volume via the modified image synthesis process; and
modifying the modified image synthesis process by reducing a loss function that quantifies a summed difference between the full data reconstruction and the synthesized reconstruction.
2. The computer-implemented method of claim 1 , wherein modifying the image synthesis process comprises adjusting at least one parameter included in a plurality of parameters of the image synthesis process.
3. The computer-implemented method of claim 1 , wherein modifying the image synthesis process comprises performing a supervised learning process on the image synthesis process.
4. The computer-implemented method of claim 3 , wherein the partial data reconstruction comprises an input object in the supervised learning process and the full data reconstruction comprises an expected output in the supervised learning process.
5. The computer-implemented method of claim 3 , wherein an additional input object of the supervised learning process includes imaging information that is associated with the target volume and is acquired separately from the first set of projection images.
6. The computer-implemented method of claim 1 , wherein reducing the loss function comprises applying a gradient descent method to modify at least one parameter included in a plurality of parameters of the image synthesis process.
7. The computer-implemented method of claim 6 , wherein modifying the at least one parameter included in the plurality of parameters of the modified image synthesis process comprises generating a second image synthesis process.
8. The computer-implemented method of claim 7 , wherein reducing the loss function further comprises generating a second synthesized reconstruction of the target volume via the second image synthesis process.
9. The computer-implemented method of claim 1 , further comprising, prior to modifying the image synthesis process, measuring the loss function by performing a voxel-to-voxel comparison of the synthesized reconstruction with the full data reconstruction.
10. The computer-implemented method of claim 1 , wherein at least one modified projection image in the second set is degraded in image quality compared to a corresponding projection image in the first set.
11. The computer-implemented method of claim 10 , wherein generating the second set of modified projection images of the target volume comprises generating the at least one modified projection image via a projection image simulation process.
12. The computer-implemented method of claim 1 , wherein the second set of modified projection images includes fewer total projection images of the target volume than the first set of modified projection images.
13. The computer-implemented method of claim 12 , wherein the second set of modified projection images includes a subset of projection images from the first set of projection images.
14. The computer-implemented method of claim 13 , wherein the first set of projection images have been acquired over a first acquisition arc and the subset of projection images consists of images that have been acquired over a second acquisition arc that is smaller than the first acquisition arc.
15. The computer-implemented method of claim 13 , wherein the subset of projection images consists of each Nth projection image of the first set of projection images, where N is an integer greater than 1.
16. The computer-implemented method of claim 1 , wherein the second set of modified projection images of the target volume is based on the first set of projection images of the target volume.
17. A computer-implemented method of training an image synthesis process, the method comprising:
receiving a first set of projection images of a target volume;
generating a second set of modified projection images of the target volume, wherein the second set of modified projection images of the target volume includes less image information associated with the target volume than the first set of projection images;
generating a full data reconstruction of the target volume based on the first set of projection images;
generating a partial data reconstruction of the target volume based on the second set of modified projection images;
generating a synthesized reconstruction of the target volume via a generator function of a generative adversarial network;
determining an assessed quality value of the synthesized reconstruction of the target volume via a discriminator function of the generative adversarial network; and
modifying the generator function based on the assessed quality value of the synthesized reconstruction of the target volume.
18. The computer-implemented method of claim 17 , further comprising:
determining an actual quality value of the synthesized reconstruction; and
modifying the discriminator function based on the actual quality value and the assessed quality value.
19. The computer-implemented method of claim 17 , further comprising, generating a subsequent synthesized reconstruction of the target volume via the modified generator function.
20. The computer-implemented method of claim 17 , wherein determining the actual quality value of the synthesized reconstruction comprises using the discriminator function to compare the synthesized reconstruction to the full data reconstruction.