Deep learning based scatter correction
An imaging system includes a computed tomography (CT) imaging device ( 10 ) (optionally a spectral CT), an electronic processor ( 16, 50 ), and a non-transitory storage medium ( 18, 52 ) storing a neural network ( 40 ) trained on simulated imaging data ( 74 ) generated by Monte Carlo simulation ( 60 ) including simulation of at least one scattering mechanism ( 66 ) to convert CT imaging data to a scatter estimate in projection space or to convert an uncorrected reconstructed CT image to a scatter estimate in image space. The storage medium further stores instructions readable and executable by the electronic processor to reconstruct CT imaging data ( 12, 14 ) acquired by the CT imaging device to generate a scatter-corrected reconstructed CT image ( 42 ). This includes generating a scatter estimate ( 92, 112, 132, 162, 182 ) by applying the neural network to the acquired CT imaging data or to an uncorrected CT image ( 178 ) reconstructed from the acquired CT imaging data.
1. A non-transitory storage medium configured to store instructions readable and executable by a processor to perform a method comprising:
inputting an uncorrected computed tomography (CT) image into a neural network to output an image domain scatter estimate, wherein the neural network is trained to transform a simulated uncorrected CT image to a simulated scatter estimate for which a value of a difference metric, comparing the simulated scatter estimate to a simulated scatter image component of the simulated uncorrected CT image, is minimized by the training;
reconstructing CT projection data to generate the uncorrected CT image; and
correcting the uncorrected CT image by subtracting the image domain scatter estimate from the uncorrected CT image to generate the scatter-corrected CT image.
2. The non-transitory storage medium of claim 1 , wherein the simulated uncorrected CT image is generated by reconstruction of Monte Carlo simulated CT imaging data including Monte Carlo simulation of at least one scattering mechanism producing simulated scatter projection data that is reconstructed into the simulated scatter image component of the simulated uncorrected CT image.
3. The non-transitory storage medium of claim 1 , wherein the CT projection data is processed by a basis function decomposition of acquired spectral or dual energy CT imaging data.
4. The non-transitory storage medium of claim 3 , wherein the basis function decomposition of the acquired spectral or dual energy CT imaging data produces the CT projection data as water projection data or iodine projection data.
5. The non-transitory storage medium of claim 1 , wherein the scatter-corrected CT image is displayed on a display device.
6. The non-transitory storage medium of claim 1 , wherein the CT projection data is generated by performing a convolution- or deconvolution-based scatter correction on acquired CT projection data received from a CT imaging device.
7. A computer-implemented method, comprising:
inputting an uncorrected computed tomography (CT) image into a neural network to output an image domain scatter estimate, wherein the neural network is trained to transform a simulated uncorrected CT image to a simulated scatter estimate for which a value of a difference metric, comparing the simulated scatter estimate to a simulated scatter image component of the simulated uncorrected CT image, is minimized by the training;
reconstructing CT projection data to generate the uncorrected CT image; and
correcting the uncorrected CT image by subtracting the image domain scatter estimate from the uncorrected CT image to generate the scatter-corrected CT image.
8. The method of claim 7 , further comprising:
acquiring the CT projection data using a CT imaging device.
9. An imaging system, comprising:
a computed tomography (CT) imaging device;
a processor configured to:
input an uncorrected computed tomography (CT) image into a neural network to output an image domain scatter estimate, wherein the neural network is trained to transform a simulated uncorrected CT image to a simulated scatter estimate for which a value of a difference metric, comparing the simulated scatter estimate to a simulated scatter image component of the simulated uncorrected CT image, is minimized by the training;
reconstruct CT projection data to generate the uncorrected CT image; and
correct the uncorrected CT image by subtracting the image domain scatter estimate from the uncorrected CT image to generate the scatter-corrected CT image.
10. The imaging system of claim 9 , wherein:
the neural network includes two or more neural networks each trained on different simulated imaging data generated by Monte Carlo simulation including simulation of imaging of human subjects of different anatomical dimensions or body weights.
11. The imaging system of claim 9 , wherein:
the neural network includes two or more neural networks each trained on different simulated imaging data generated by Monte Carlo simulation including simulation of scattering by different filters installable on the CT imaging device.