Apparatus and method for dual-energy computed tomography (CT) image reconstruction using sparse kVp-switching and deep learning
A deep learning (DL) network reduces artifacts in computed tomography (CT) images based on complementary sparse-view projection data generated from a sparse kilo-voltage peak (kVp)-switching CT scan. The DL network is trained using input images exhibiting artifacts and target images exhibiting little to no artifacts. Another DL network can be trained to perform image-domain material decomposition of the artifact-mitigated images by being trained using target images in which beam hardening is corrected and spatial variations in the X-ray beam are accounted for. Further, material decomposition and artifact mitigation can be integrated in a single DL network that is trained using as inputs reconstructed images having artifacts and as targets material images without artifacts with beam-hardening corrections, etc. Further, the target material images can be transformed using a whitening transform to decorrelate noise.
1. An apparatus, comprising:
processing circuitry configured to
obtain projection data representing an intensity of X-ray radiation at a plurality of detector elements, the projection data representing first sparse-view data acquired using a first voltage applied to an X-ray source and second sparse-view data acquired using a second voltage applied to the X-ray source, the second voltage being greater than the first voltage and the first sparse-view data being acquired at views that are different from views at which the second sparse-view data is acquired,
reconstruct, from the first sparse-view data, a first low-energy image,
reconstruct, from the second sparse-view data, a first high-energy image,
acquire an artifact-mitigating neural network that has two channels, one input channel for a low-energy image and another input channel for a high-energy image, the artifact-mitigating neural network having been trained using a training dataset that includes input images that are respective pairs of sparse-view reconstructed images that have complementary streak artifacts and corresponding target images in which the streak artifacts are mitigated, and
apply the first low-energy image and the first high-energy image to the artifact-mitigating neural network to generate artifact-mitigated images.
2. The apparatus according to claim 1 , wherein
the artifact-mitigated images include a second low-energy image and a second high-energy image, the second low-energy image representing the first low-energy image after mitigation of the streak artifacts, and the second high-energy image representing the first high-energy image after mitigation of the streak artifacts.
3. The apparatus according to claim 2 , wherein the processing circuitry is further configured to
acquire a material-decomposition neural network that has two channels, one input channel for the low-energy image and another input channel for the high-energy image, the material-decomposition neural network having been trained to generate a first material image and a second material image when a pair of a low-energy image and a high-energy image, which have streak artifacts mitigated, are applied to the material decomposition neural network, and
apply the second low-energy image and the second high-energy image to the material-decomposition neural network to generate material-decomposed images.
4. The apparatus according to claim 3 , wherein the processing circuitry is further configured to acquire the material-decomposition neural network, wherein the material-decomposition neural network has been trained using a training dataset in which material-decomposed target images include corrections for beam hardening and account for spatial variations in an X-ray beam used to generate the target images.
5. The apparatus according to claim 3 , wherein the processing circuitry is further configured to acquire the material-decomposition neural network, wherein
the material-decomposition neural network has been trained using the training dataset in which the target images are training material-decomposed images that have been transformed using a whitening transform, and
the generated material-decomposed images are generated by applying an inverse whitening transform to resultant images output from the material-decomposition neural network.
6. The apparatus according to claim 1 , wherein the processing circuitry is further configured to acquire the artifact-mitigating neural network, wherein the artifact-mitigating neural network has been trained using a training dataset in which target images are generated based on first full-view projection data acquired using a low-energy X-ray beam and second full-view projection data acquired using a high-energy X-ray beam.
7. The apparatus according to claim 1 , wherein the processing circuitry is further configured to acquire the artifact-mitigating neural network, wherein the artifact-mitigating neural network is also a material-decomposition neural network that has been trained to both mitigate streak artifacts and decompose spectral images into material components as an integrated process when sparse-view reconstructed images are applied to the artifact-mitigating neural network.
8. The apparatus according to claim 7 , wherein the processing circuitry is further configured to acquire the artifact-mitigating neural network, wherein
the artifact-mitigating neural network has been trained using a training dataset in which target images are material images decomposed from spectral images reconstructed from first full-view projection data acquired using a low-energy X-ray beam and from second full-view projection data acquired using a high-energy X-ray beam, and
the input images are spectral images reconstructed from first sparse-view projection data acquired using the low-energy X-ray beam and from second sparse-view projection data acquired using the high-energy X-ray beam, the first sparse-view projection data being at projection views that are different angles than projection views of the second sparse-view projection data.
9. The apparatus according to claim 7 , wherein the processing circuitry is further configured to acquire the artifact-mitigating neural network, wherein
the artifact-mitigating neural network has been trained using the training dataset in which the target images are training material-decomposed images that have been transformed using a whitening transform, and
the generated material-decomposed images are generated by applying an inverse whitening transform to resultant images output from the material-decomposition neural network.
10. The apparatus according to claim 1 , wherein the processing circuitry is further configured to acquire the artifact-mitigating neural network, wherein the artifact-mitigating neural network has been trained using a training dataset in which the input images are reconstructed using a same reconstruction method as used to reconstruct the first low-energy image and the first high-energy image.
11. The apparatus according to claim 1 , further comprising
an X-ray source configured to radiate X-rays having an energy spectrum that depends on a voltage applied to the X-ray source,
an energy-integrating X-ray detector having the plurality of detector elements configured to output the projection data, and
control circuitry configured to
rotate the X-ray source and the energy-integrating X-ray detector through a series of projection views, and
change, back-and-forth between a high kilo-voltage and a low kilo-voltage, the voltage applied to the X-ray source at increments of two or more projection views of the series of projection views.
12. The apparatus according to claim 11 , wherein the control circuitry is configured to change the voltage applied to the X-ray source such that a number of projection views at the high kilo-voltage is different from a number of projection views at the low kilo-voltage.
13. An apparatus, comprising:
processing circuitry configured to train an artifact-mitigating neural network having two channels by
initializing the artifact-mitigating neural network,
obtaining input images including pairs of sparse-view reconstructed images that have complementary streak artifacts, the sparse-view reconstructed images representing reconstructed images from projection data of a sparse-view kilo-voltage peak (kVp)-switching computed tomography (CT) scan,
obtaining target images including pairs of images corresponding to the sparse-view reconstructed images except with the streak artifacts having been mitigated, and
iteratively updating network coefficients to optimize a loss function representing agreement between the target images and output images resulting from the input images applied to the artifact-mitigating neural network.
14. The apparatus according to claim 13 , wherein the processing circuitry is further configured to train the artifact-mitigating neural network, wherein
the obtained target images include pairs of full-view reconstructed images that represent reconstructed images from projection data of a full-view kVp-switching CT scan.
15. The apparatus according to claim 13 , wherein the processing circuitry is further configured to train the artifact-mitigating neural network, wherein
the obtained target images include pairs of material component images generated by performing material decomposition on in the sinogram domain using projection data of a full-view kVp-switching CT scan to generate material-component projection data, and reconstructing material images from the material-component projection data.
16. The apparatus according to claim 15 , wherein the processing circuitry is further configured to train the artifact-mitigating neural network, wherein
the obtained target images further include that the pairs of material component images are generated by performing a whitening transform on the reconstructed material images.
17. A method of mitigating artifacts in computed tomography (CT) images from sparse-view kilo-voltage peak (kVp)-switching CT scans, comprising:
obtaining projection data representing an intensity of X-ray radiation at a plurality of detector elements, the projection data representing first sparse-view data acquired using a first voltage applied to an X-ray source and second sparse-view data acquired using a second voltage applied to the X-ray source, the second voltage being greater than the first voltage and the first sparse-view data being acquired at views that are different from views at which the second sparse-view data is acquired,
reconstructing, from the first sparse-view data, a first low-energy image,
reconstructing, from the second sparse-view data, a first high-energy image,
acquiring an artifact-mitigating neural network that has two channels, one input channel for a low-energy image and another input channel for a high-energy image, and the artifact-mitigating neural network having been trained using a training dataset that includes input images that are respective pairs of sparse-view reconstructed images that have complementary streak artifacts and corresponding target images in which the streak artifacts are mitigated, and
applying the first low-energy image and the first high-energy image to the artifact-mitigating neural network to generate artifact-mitigated images.
18. The method according to claim 17 , wherein the artifact-mitigated images include a second low-energy image and a second high-energy image, the second low-energy image representing the first low-energy image after mitigation of the streak artifacts, and the second high-energy image representing the first high-energy image after mitigation of the streak artifacts, the method further comprising
acquiring a material-decomposition neural network that has two channels, one input channel for a low-energy image and another input channel for a high-energy image, and the material-decomposition neural network having been trained to generate a first material image and a second material image when a pair of a low-energy image and a high-energy image, which have streak artifacts mitigated, are applied to the material decomposition neural network, and
applying the second low-energy image and the second high-energy image to the material-decomposition neural network to generate material-decomposed images.
19. The method according to claim 18 , wherein the acquiring of the artifact-mitigating neural network further includes that the artifact-mitigating neural network is also a material-decomposition neural network that has been configured to both mitigate streak artifacts and decompose spectral images into material components as an integrated process when sparse-view reconstructed images are applied to the artifact-mitigating neural network.
20. A non-transitory computer readable storage medium including executable instructions, wherein the instructions, when executed by circuitry, cause the circuitry to perform the method according to claim 17 .