Apparatus and method that uses deep learning to correct computed tomography (CT) with sinogram completion of projection data
A deep learning (DL) network corrects/performs sinogram completion in computed tomography (CT) images based on complementary high- and low-kV projection data generated from a sparse (or fast) kilo-voltage (kV)-switching CT scan. The DL network is trained using inputs and targets, which respectively generated with and without kV switching. Another DL network can be trained to correct sinogram-completion errors in the projection data after a basis/material decomposition. A third DL network can be trained to correct sinogram-completion errors in reconstructed images based on the kV-switching projection data. Performance of the DL network can be improved by dividing a 3D convolutional neural network (CNN) into two steps performed by respective 2D CNNs. Further, the projection data and DLL can be divided into high- and low-frequency components to improve performance.
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 high-kV projection data interleaved as a function of projection view with low-kV projection data in which projection views of the high-kV projection data are disjoint from projection views of the low-kV projection data, the low-kV projection data being acquired using a first voltage applied to an X-ray source, and the high-kV projection data being acquired using a second voltage applied to the X-ray source, the second voltage being greater than the first voltage,
acquire a first neural network that has two input channels, one input channel for high-kV data and another input channel for low-kV data, and the first neural network having been trained using a training dataset that includes input data and target data, the input data being based on kV-switching sinograms in which projection views of kV-switching low-kV sinograms are disjoint from projection views of kV-switching high-kV sinograms, and the target data being based on complete sinograms that include complete low-kV sinograms and complete high-kV sinograms in which each of the complete sinograms includes projection data for all projection views,
perform sinogram completion to the high-kV projection data and the low-kV projection data, and
apply, to the first neural network, a low-kV input based on the low-kV projection data and a high-kV input based on the high-kV projection data, the first neural network outputting a result in which errors from the sinogram completion are reduced.
2. The apparatus according to claim 1 , wherein the processing circuitry is further configured to perform the sinogram completion integrally with applying the low-kV and high-kV inputs to the first neural network, wherein
the high-kV input to the first neural network is the high-kV projection data prior to the sinogram completion,
the low-kV input to the first neural network is the low-kV projection data prior to the sinogram completion, and
the first neural network is trained to use the complementary information of the high-kV projection data and the low-kV projection data to perform sinogram completion, generating complete high-kV projection data that includes projection data for all projection views and complete low-kV projection data that includes projection data for all projection views.
3. The apparatus according to claim 1 , wherein the processing circuitry is further configured to perform the sinogram completion prior to applying, to the first neural network, the low-kV and high-kV inputs, wherein
performing the sinogram completion generates complete high-kV projection data that includes projection data for all projection views and complete low-kV projection data that includes projection data for all projection views, and
applying, to the first neural network, the complete high-kV projection data as the high-kV input and the complete low-kV projection data as the low-kV input to output the result in which sinogram-completion errors are reduced in the complete high-kV projection data and the complete low-kV projection data that are output from the first neural network.
4. The apparatus according to claim 1 , wherein the processing circuitry is further configured to
decompose the high-kV projection data and the low-kV projection data into first-basis projection data and second-basis projection data,
reconstruct a first-basis image from the first-basis projection data, and
reconstruct a second-basis image from the second-basis projection data.
5. The apparatus according to claim 1 , wherein the processing circuitry is further configured to decompose the high-kV projection data and the low-kV projection data into first material-component projection data and second material-component projection data using a material decomposition.
6. The apparatus according to claim 4 , wherein the processing circuitry is further configured to
acquire a second neural network that has two channels, one input channel for the first-basis projection data and another input channel for the second-basis projection data, wherein
the second neural network has been trained using a training dataset that includes input data and target data, the input data being based on a basis decomposition of the kV-switching sinograms, and the target data being based on a basis decomposition of the complete sinograms.
7. The apparatus according to claim 4 , wherein the processing circuitry is further configured to
acquire a second neural network that has two channels, one input channel for the first-basis image and another input channel for second-basis image wherein
the second neural network has been trained using a training dataset that includes input data and target data,
the input data being based on images that have been reconstructed from a basis decomposition of the kV-switching sinograms, and
the target data being based on images that have been reconstructed from a basis decomposition of the complete sinograms.
8. The apparatus according to claim 1 , wherein the processing circuitry is further configured to
acquire a second neural network that has two input channels, one input channel for high-kV data and another input channel for low-kV data,
partition the low-kV projection data and the high-kV projection data into respective first slices along a first plane,
apply the respective first slices to the first neural network to generate partially-corrected low-kV projection data and partially-corrected high-kV projection data,
partition the partially-corrected low-kV projection data and the partially-corrected high-kV projection data into respective second slices along a second plane, which is orthogonal to the first plane, and
apply the respective second slices to the second neural network to generate corrected low-kV projection data and corrected high-kV projection data.
9. The apparatus according to claim 1 , wherein the processing circuitry is further configured to
acquire a second neural network that has two input channels, one input channel for high-kV data and another input channel for low-kV data,
transform the low-kV projection data into low-kV frequency components, including high-spatial-frequency low-kV projection data and low-spatial-frequency low-kV projection data,
transform the high-kV projection data into high-kV frequency components, including high-spatial-frequency high-kV projection data and low-spatial-frequency high-kV projection data,
apply, to the first neural network, high frequency components of the high-kV frequency components and of the low-kV frequency components,
apply, to the second neural network, low frequency components of the high-kV frequency components and of the low-kV frequency components, and
combine the respective high frequency components with the corresponding low frequency components to generate corrected low-kV projection data and corrected high-kV projection data.
10. The apparatus according to claim 9 , wherein the processing circuitry is further configured to transform projection data into frequency components by one of
applying a low-pass filter to the projection data to generate low-spatial-frequency projection data, and subtracting the low-spatial-frequency projection data from the projection data to generate high-spatial-frequency projection data,
applying a wavelet transformation to projection data to separate the projection data into low-spatial-frequency projection data and high-spatial-frequency projection data, and
transforming projection data into a frequency domain and separating the projection data in the frequency domain into low-spatial-frequency projection data and high-spatial-frequency projection data.
11. The apparatus according to claim 9 , wherein the processing circuitry is further configured to apply the low frequency components to the second neural network, by
downsampling the low-spatial-frequency low-kV projection data and the low-spatial-frequency high-kV projection data to generate downsampled projection data,
apply the downsampled projection data to the second neural network to generate a downsampled result, and
upsample the downsampled result to generate corrected low-spatial-frequency low-kV projection data and corrected low-spatial-frequency high-kV projection data.
12. The apparatus according to claim 1 , wherein the processing circuitry is further configured to
obtain a kV-switching mask that represents at which projection views high-kV projection data was acquired versus at which projection views low-kV projection data was acquired, wherein
the first neural network has a third input channel for receiving the kV-switching mask.
13. 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 high-kV projection data and output the low-kV 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 the first voltage and the second voltage, a voltage applied to the X-ray source, wherein
the control circuitry configured to perform sparse-view kV-switching by changing the voltage applied to the X-ray source only after the voltage applied to the X-ray source has not changed during two or more most recent projection views of the series of projection views.
14. 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 high-kV projection data interleaved as a function of projection view with low-kV projection data in which projection views of the high-kV projection data are disjoint from projection views of the low-kV projection data, the low-kV projection data being acquired using a first voltage applied to an X-ray source, and the high-kV projection data being acquired using a second voltage applied to the X-ray source, the second voltage being greater than the first voltage,
acquire a first neural network that has a first channel for high-kV data, and the first neural network having been trained using a first training dataset that includes input data and target data, the input data being based on kV-switching sinograms in which projection views of kV-switching low-kV sinograms are disjoint from projection views of kV-switching high-kV sinograms, and the target data being based on complete sinograms that include complete high-kV sinograms in which each of the complete sinograms including projection data for all projection views,
acquire a second neural network that has a first channel for low-kV data, and the second neural network having been trained using a second training dataset that includes input data and target data, the input data being based on kV-switching sinograms in which the projection views of the kV-switching low-kV sinograms are disjoint from the projection views of the kV-switching high-kV sinograms, and the target data being based on complete sinograms that include complete low-kV sinograms including projection data for all projection views,
perform sinogram completion to the high-kV projection data, and sinogram completion to the low-kV projection data,
apply, to the first neural network, a high-kV input based on the high-kV projection data, the first neural network outputting a result in which errors from the sinogram completion of the high-kV projection data are reduced, and
apply, to the second neural network, a low-kV input based on the low-kV projection data, the second neural network outputting a result in which errors from the sinogram completion of the low-kV projection data are reduced.
15. The apparatus according to claim 14 , wherein the processing circuitry is further configured to
acquire the first neural network, wherein the first neural network has a second channel for a kV-switching mask that represents at which projection views the kV-switching low-kV sinograms were acquired versus at which projection views the kV-switching high-kV sinograms were acquired,
acquire the second neural network, wherein the second neural network has a second channel for the kV-switching mask,
apply, to the first neural network, the high-kV input and a kV-switching mask of the obtained projection data, and
apply, to the second neural network, the low-kV input and the kV-switching mask of the obtained projection data, wherein
the first neural network has been trained using the first training dataset that includes the input data that includes respective kV-switching masks corresponding to the kV-switching sinograms, and
the second neural network has been trained using the second training dataset that includes the input data that includes respective kV-switching masks corresponding to the kV-switching sinograms.
16. The apparatus according to claim 1 , wherein the processing circuitry is further configured to
obtain the projection data, wherein the projection data represents middle-kV projection data interleaved, as the function of the projection view, with the high-kV projection data and the low-kV projection data, and
acquire the first neural network that has three input channels, the one input channel for the high-kV data, the another input channel for the low-kV data, and a third input channel for middle-kV data.
17. An apparatus, comprising:
processing circuitry configured to train a sinogram-completion neural network having two channels by
initializing network coefficients of the sinogram-completion neural network,
obtaining input data including pairs of high-kV and low-kV inputs based on kV-switching sinograms in which projection views of kV-switching low-kV sinograms are disjoint from projection views of kV-switching high-kV sinograms and the kV-switching low-kV sinograms are interleaved as a function of projection view with the kV-switching high-kV sinograms,
obtaining target data including pairs of high-kV and low-kV targets based on complete sinograms that include complete low-kV sinograms and complete high-kV sinograms in which each of the complete sinograms includes projection data for all projection views, and
iteratively updating the network coefficients to optimize a loss function representing agreement between the target data and outputs from applying the input data to the sinogram-completion neural network.
18. The apparatus according to claim 17 , wherein the processing circuitry is further configured to train the sinogram-completion neural network by
obtaining a kV-switching mask that represents at which projection views the kV-switching low-kV sinograms were acquired versus at which projection views the kV-switching high-kV sinograms were acquired, and
iteratively updating the network coefficients to optimize the loss function, wherein the loss function includes a weighting map based on the kV-switching mask.
19. The apparatus according to claim 18 , wherein the processing circuitry is further configured to train the sinogram-completion neural network, wherein the kV-switching mask further represents, for a missing value in a kV-switching sinogram, a proximity to a nearest acquired value in the kV-switching sinogram.
20. A method of sinogram completion of computed tomography (CT) projection data acquired via kV-switching, the method comprising:
obtaining projection data representing an intensity of X-ray radiation at a plurality of detector elements, the projection data representing high-kV projection data interleaved as a function of projection view with low-kV projection data in which projection views of the high-kV projection data are disjoint from projection views of the low-kV projection data, the low-kV projection data being acquired using a first voltage applied to an X-ray source, and the high-kV projection data being acquired using a second voltage applied to the X-ray source, the second voltage being greater than the first voltage;
acquiring a first neural network that has two input channels, one input channel for high-kV data and another input channel for low-kV data, and the first neural network having been trained using a training dataset that includes input data and target data, the input data being based on kV-switching sinograms in which projection views of kV-switching low-kV sinograms are disjoint from projection views of kV-switching high-kV sinograms, and the target data being based on complete sinograms that include complete low-kV sinograms and complete high-kV sinograms in which each of the complete sinograms includes projection data for all projection views;
performing sinogram completion to the high-kV projection data and the low-kV projection data; and
applying, to the first neural network, a low-kV input based on the low-kV projection data and a high-kV input based on the high-kV projection data, the first neural network outputting a result in which errors from the sinogram completion are reduced.