Low dose digital tomosynthesis system and method using artificial intelligence
A mobile radiography apparatus is configured to sparsely sample radiographic projection images to generate high resolution tomosynthesis volume images using a digital radiographic detector that is mechanically uncoupled from the x-ray source and an artificial intelligence network. The artificial intelligence network is trained to correct a volume image generated from sparsely sample projection images to generate the high resolution tomosynthesis volume images.
1. A computer implemented radiographic image processing method, the method comprising:
training a convolutional neural network to correct image defects in an image volume reconstructed from sparsely captured tomosynthesis projection images;
sparsely capturing tomosynthesis projection images of a patient anatomy within an angular range between about two-thirds of a full angular range to about four-fifths of the full angular range using at least one x-ray source and a digital radiographic (DR) detector;
reconstructing a sparse volume image of the patient anatomy using the sparsely captured tomosynthesis projection images of the patient anatomy and providing the reconstructed sparse volume image of the patient anatomy to the trained convolutional neural network; and
the trained neural network outputting a corrected reconstructed sparse volume image of the patient anatomy.
2. The method of claim 1 , wherein the step of training comprises:
reconstructing a full tomosynthesis volume image from a full set of acquired tomosynthesis projection images;
reconstructing a sparse tomosynthesis volume image from a preselected subset of the full set of acquired tomosynthesis projection images
determining differences between the full tomosynthesis volume and the sparse tomosynthesis volume image; and
providing the reconstructed sparse tomosynthesis volume image and the determined differences to the convolutional neural network.
3. The method of claim 2 , wherein the step of sparsely capturing comprises at least one of capturing a plurality of tomosynthesis projection images over a preselected limited angular range within the angular range between about two-thirds to about four-fifths of the full angular range and capturing a preselected sparse number of tomographic projection images within the angular range between about two-thirds to about four-fifths of the full angular range.
4. The method of claim 1 , further comprising:
positioning a plurality of radiopaque markers in a path of an x-ray beam used for the step of sparsely capturing; and
calculating a position of the DR detector relative to the at least one x-ray source according to positions of the radiopaque markers in the sparsely captured tomosynthesis projection images.
5. The method of claim 4 , further comprising positioning a plurality of x-ray sources in a linear or curvilinear pattern for the step of sparsely capturing the tomosynthesis projection images of the patient anatomy.
6. A mobile radiography system comprising
a mobile base;
a support arm attached to the mobile base;
an x-ray tube head attached to one end of the support arm, the x-ray tube head comprising one or more x-ray sources;
a portable DR detector mechanically uncoupled from the x-ray tube head, the portable DR detector and the x-ray tube head configured to sparsely capture tomosynthesis projection images of a patient anatomy over a preselected angular range at about every two degrees to about every five degrees over the preselected angular range;
a processing system for reconstructing a sparse volume image of the patient anatomy using the sparsely captured tomosynthesis projection images; and
an artificial intelligence module trained to correct the reconstructed sparse volume image and to output a corrected reconstructed sparse volume image.
7. The system of claim 6 , wherein the tube head comprises a plurality of x-ray sources arranged in a linear or curvilinear pattern.
8. The system of claim 6 , further comprising a plurality of radiopaque markers, the radiopaque markers disposed in one or more radiation paths each extending from one of the x-ray sources to the portable DR detector, and wherein the processing system is programmed to calculate a position of the portable DR detector relative to the one or more x-ray sources according to positions of the radiopaque markers appearing in the sparsely captured tomosynthesis projection images of the patient anatomy.
9. The system of claim 8 , wherein the processing system is further programmed to digitally remove the radiopaque markers appearing in the sparsely captured tomosynthesis projection images.
10. The system of claim 6 , further comprising a mobile cone beam computed tomography system.
11. A computer implemented radiographic image processing method, the method comprising:
training a convolutional neural network to correct image defects in an image volume reconstructed from sparsely captured tomosynthesis projection images;
sparsely capturing tomosynthesis projection images of a patient anatomy at about one-third to about one-fourth of a standard number of tomosynthesis projection images of the patient anatomy using at least one x-ray source and a digital radiographic (DR) detector;
reconstructing a sparse volume image of the patient anatomy using the sparsely captured tomosynthesis projection images of the patient anatomy and providing the reconstructed sparse volume image of the patient anatomy to the trained convolutional neural network; and
the trained neural network outputting a corrected reconstructed sparse volume image of the patient anatomy.
12. The method of claim 11 , wherein the step of training comprises:
reconstructing a full tomosynthesis volume image from a full set of acquired tomosynthesis projection images;
reconstructing a sparse tomosynthesis volume image from a preselected subset of the full set of acquired tomosynthesis projection images
determining differences between the full tomosynthesis volume image and the sparse tomosynthesis volume image; and
providing the reconstructed sparse tomosynthesis volume image and the determined differences to the convolutional neural network.
13. The method of claim 11 , wherein the step of sparsely capturing comprises at least one of capturing a plurality of tomosynthesis projection images over a preselected limited angular range and capturing a preselected sparse number of tomographic projection images over a full angular range.
14. The method of claim 11 , further comprising sparsely, capturing tomosynthesis projection images of the patient anatomy within an angular range between about two-thirds of a full angular range to about four-fifths of the full angular range.
15. The method of claim 11 , further comprising sparsely capturing tomosynthesis projection images of the patient anatomy at about every two degrees to about every five degrees over a preselected angular range.
16. The method of claim 11 , further comprising:
positioning a plurality of radiopaque markers in a path of an x-ray beam used for the step of sparsely capturing; and
calculating a position of the DR detector relative to the at least one x-ray source according to positions of the radiopaque markers in the sparsely captured tomosynthesis projection images.
17. The method of claim 16 , further comprising positioning a plurality of x-ray sources in a linear or curvilinear pattern for the step of sparsely capturing the tomosynthesis projection images of the patient anatomy.