Spatial variation artifact removal method for computed tomography

Abstract

A spatially varying artifact removal method for 3D computed tomography involves thresholding a current reconstruction to create a thresholding reconstruction and then creating a simulated reconstruction from the thresholding reconstruction. These simulated reconstructions are subtracted from the current reconstruction to create a current reconstruction for the next iteration. A final reconstruction is then created by adding the thresholding reconstructions. The method can gradually remove artifacts. In addition, the method can be used for generating high-quality training data so as to further improve speed and robustness. These methods are generally applicable to other non-Orlov complete computed tomography imaging, such as high cone angles, missing views.

Description

[0001] Related applications

[0002] This application claims the benefit under 35 U.S.C. 119(e) of U.S. Provisional Application No. 63 / 550,077, filed February 6, 2024, the entire contents of which are incorporated herein by reference. Background Art

[0003] X-ray microtomography systems provide high-resolution, non-destructive imaging of a sample's internal structure. These systems are used in a variety of industrial and research applications, such as semiconductor device manufacturing, mining, manufacturing, materials science, clinical research, and failure analysis. These systems provide the ability to visualize features within a sample without cutting or slicing the sample.

[0004] The X-ray microtomography system includes an X-ray projection system that generates sample projection data and a computer system that reconstructs a volume of sample tomography based on the projection data. In operation, the X-ray projection system scans the sample at different angles to generate projection data. During the scanning process, X-rays are directed to the sample and are absorbed or scattered by the sample as the X-rays pass through the sample. The X-rays that are not absorbed or scattered are transmitted through the sample and modulated by the sample. The detector system receives the transmitted X-rays and creates a pixelated image with digitized pixel values of the received X-rays. A series of X-ray projections at different angles generated by the scan form projection data of the sample. The computer system then accesses the projection data and applies a reconstruction algorithm to the projection data to reconstruct a volume data set of the sample. These volume data sets are three-dimensional (3D) representations of the sample, and slices are two-dimensional (2D) cross-sectional images of the sample based on the volume data set.

[0005] The most commonly used reconstruction algorithms belong to a class of reconstruction techniques known as analytical reconstruction. Their goal is to find a closed-form solution to the problem of reconstructing the internal structure of an object from its projections. The most common analytical method is filtered back projection (FBP). The projections are first processed in the frequency or spatial domain using a high-pass filter (usually a ramp filter). Each filtered projection is then "smeared" back to the imaging plane, as if each data point were uniformly projected back in the shape of the original beam. These back projections at all angles are summed to obtain a reconstructed volume that approximates the internal structure of the object. The filtering and back projection operations together help to obtain a more accurate and unblurred reconstruction of the original object. One type of filtered back projection is the FDK reconstruction algorithm. See Feldkamp, Lee A., Lloyd C. Davis, and James W. Kress, "Practical cone-beam algorithm," J. Am. Soc. Am. 1.6 (1984): 612-619. It is often used with X-ray microtomography systems because it reduces artifacts associated with typical cone beams. However, such algorithms generally assume that the imaging geometry satisfies the Tuoy condition, which states that every plane intersecting the object (within the 3D region of interest to be reconstructed) must contain at least one vertex. See Feldkamp, LA, Davis, LC, and Kress, JW (1984) Practical Cone-Beam Algorithm in Optical Society of America A, 1, 612-619.

[0006] Compared to traditional computed tomography, tomographic reconstruction in a tomographic setting presents some unique challenges. Tomographic imaging involves capturing X-ray images within a limited angular range around a specific plane or axis within the object. Unlike conventional CT, tomographic imaging typically has a limited angular range. Tomographic tomography often rotates the sample 360 degrees. However, the beamline is no longer perpendicular to the axis of rotation (RA). This limited angular data leads to incomplete information, making it more challenging to achieve high-quality 3D reconstruction. Specifically, the limited angular data can lead to artifacts (such as streaks or blurring) in the reconstructed image. These artifacts can obscure details and affect the accuracy of the reconstructed volume. Specialized reconstruction algorithms are often required to mitigate these artifacts.

[0007] More specifically, for imaging large planar objects, it is often beneficial to use a complete computed tomography geometry that does not conform to Tuy's or Orlov's conditions (in the parallel beam case). See Tuy, Heang K., "An inversion formula for cone-beam reconstruction." SIAM Journal on Applied Mathematics 43.3 (1983): 546-552; Orlov, SS, "Theory of three-dimensional image reconstruction: I Conditions for a complete set of projections." Sov. Phys. Crystallogr. 20 (1976): 312-314. Tuy's sufficient condition for CT (parallel beam and cone beam) reconstruction is that every plane that intersects the object (within the 3D region of interest to be reconstructed) must contain at least one vertex. It is widely used to determine whether the image geometry can theoretically provide a good reconstruction for each voxel. The work is demonstrated for cone beams, but is certainly applicable to parallel beams as well (at low cone angles, cone beams become parallel beams). Some complete examples of non-Tuey conditions are (1) tomography (2) normal CT scans, but with high cone beam angles (e.g., moving both the source and detector very close to the sample, with a large field of view, so the angles of the rays joining the source and detector edges / corners will be very steep). Tomographic geometries improve the signal-to-noise ratio, thus providing better throughput. These geometries make the inverse reconstruction problem more ill-posed. Consequently, noticeable artifacts can appear in the reconstructed tomograms.

[0008] Different approaches have been proposed to remove artifacts caused by tomographic geometry and other similar uncertainties. One approach uses machine learning to remove noise from the reconstruction. Summary of the Invention

[0009] However, many artifact removal methods are unable to remove structured or correlated artifacts (e.g., tomographic artifacts and more generally, high-angle cone-beam artifacts). Indeed, tomographic artifacts are notoriously difficult to remove due to spatially varying structure. While spatially varying artifact removal methods exist (both data-driven and constrained optimization-based), these methods either rely heavily on careful supervision, i.e., 1) only work for similar objects with predictable shapes, or 2) require prior knowledge of the spatial distribution of the objects. Indeed, to date, there appear to be no unbiased artifact removal methods for spatially varying structure in 3D computed tomography.

[0010] The present invention relates to a method for removing or mitigating spatially varying artifacts, such as tomographic artifacts and / or high-angle cone-beam artifacts for 3D computed tomography (CT). In normal cone-beam tomography, the top and bottom of the scan are affected exactly the same as in tomographic imaging and exhibit similar artifacts.

[0011] It can progressively remove artifacts using an unbiased approach. Furthermore, the method can be used to generate high-quality training data to further improve the speed and robustness of artifact removal or mitigation. These methods will be further applicable to other non-Orlov full-length computed tomography imaging settings (e.g., high cone angles, missing views).

[0012] For example, in rotational tomography for scanning wafers or chips, the lack of accurate side views makes it mathematically infeasible to fully reconstruct the tomographic volume of an arbitrary object from in-plane projections, violating the Orlov and Toury rules. However, isolated objects can still be recovered by performing an FDK reconstruction, which shows the approximate location of the object. The brightest parts of the reconstruction are found. These parts indicate the presence of a real (but distorted) object at that location. A threshold is then set just below this brightness, and the thresholded tip of the object is taken. This small tip is added to a new overall tomographic image to show the location of this brightest part. This bright object is forward-projected to reveal views of the object. These views are back-projected using FDK, depicting an accurate image of the bright object at that location, along with all its artifacts. This reconstruction is then subtracted from the main first reconstruction, removing the small bright object and its artifacts. After these steps, the overall first reconstruction is now reduced by some amount (e.g., 10%). The top 10% of the bright spots are now in a different overall buffer, and 10% of the artifacts and streaks are removed from the entire volume. Now, a reduced tomogram exists, and it has the next set of bright spots to threshold, but these have fewer artifacts than the larger bright spots before. Decomposing the image repeatedly in this way creates a fairly self-correcting process.

[0013] In general, according to one aspect, the present invention features a method for reconstructing a tomographic volume of a sample from projection data. The method includes thresholding a current reconstruction to create a thresholded reconstruction, and then creating simulated reconstructions based on the thresholded reconstruction. These simulated reconstructions are subtracted from the current reconstruction to create the current reconstruction for a next iteration. A final reconstruction is then created by summing the thresholded reconstructions.

[0014] The method may further include training a neural network based on the reconstruction.

[0015] Preferably, a threshold is set based on the extent of the initial reconstruction, and the current reconstruction can then be thresholded by setting values below the threshold to zero. This implements the assumption that the highest value is the signal; that is, the brightest pixels are most likely to be real objects, and removes artifacts caused by the largest "signal".

[0016] A simulated reconstruction can be created by forward and back projection from the thresholded reconstruction.

[0017] Thus, spatially varying artifacts such as tomographic artifacts and / or high-angle cone-beam artifacts may be removed or at least mitigated.

[0018] In general, according to another aspect, the present invention features a computer-implemented method for improving the reconstruction of a tomographic volume from projection data. The method includes thresholding a current reconstruction to create a thresholded reconstruction, creating a simulated reconstruction based on the thresholded reconstruction, subtracting the simulated reconstruction from the current reconstruction to create an updated current reconstruction for a subsequent iteration, and summing the thresholded reconstructions to form a final reconstruction. By iteratively refining the reconstructed data in this manner, spatially varying artifacts can be reduced or eliminated, resulting in a clearer, more accurate volumetric image.

[0019] In some embodiments, thresholding of the current reconstruction involves selecting multiple thresholds based on a range of absorption values determined from an initial reconstruction of the projection data. This approach recognizes that the most important signals are often located in the upper range of absorption values, so applying hierarchical or multiple thresholds can more effectively isolate these signals. Iteratively applying these thresholds and removing associated artifacts helps ensure that important features are preserved while progressively reducing lower levels of noise or artifact signals with each subsequent iteration.

[0020] In a further example, simulated reconstructions can be created by forward-projecting each thresholded reconstruction to generate simulated projection data, which are then back-projected to form simulated reconstructions. By simulating the underlying imaging geometry during the forward and back-projection steps, these simulated reconstructions accurately capture artifacts introduced by the sample geometry and scanning trajectory. Subtracting these simulated artifacts from the current reconstruction reduces the artifact content in the final result without assuming prior knowledge of the sample geometry or composition.

[0021] In certain preferred embodiments, negative density values may appear in the updated current reconstruction after the subtraction step, which can produce a non-physical representation in the volumetric image. To address this issue, the method optionally includes correcting these negative density values to ensure a physically meaningful density distribution, for example by resetting them to zero or applying an adaptive correction based on expected material properties. This correction step helps maintain the integrity of the resulting volumetric data and avoids misleading interpretations of structural features.

[0022] Furthermore, the projection data used in this method can be generated in tomographic or high cone-beam imaging geometries, which are often affected by missing views or incomplete coverage as described by non-Orlov complete criteria. Therefore, the thresholding and iterative subtraction techniques described compensate for these geometric limitations, significantly reducing the spatially varying artifacts characteristic of such acquisition settings. By iteratively removing artifacts introduced by limited angles or nonstandard geometries, this method provides more reliable reconstructions for subsequent visualization or quantitative analysis.

[0023] The above and other features of the present invention, including various novel construction details and combinations of parts, as well as other advantages, will now be described in more detail with reference to the accompanying drawings, and pointed out in the claims. It should be understood that the specific methods and apparatus embodying the present invention are shown by way of illustration and not as limitations of the present invention. The principles and features of the present invention may be employed in various and numerous embodiments without departing from the scope of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In the accompanying drawings, reference characters refer to the same components throughout the different views. The drawings are not necessarily drawn to scale; emphasis is instead placed upon illustrating the principles of the invention. In the drawings:

[0025] Figure 1 is a schematic diagram of an X-ray microtomography system to which the method of the present invention is applicable;

[0026] Figure 2A is a schematic perspective view of a tomography setup shown according to one embodiment;

[0027] Figure 2B A reconstruction of a sample, specifically a double data rate synchronous dynamic random access memory (DDR4) chip, scanned using the full Orlov geometry as a reference, is shown, showing the infamous butterfly artifacts that plague rotational tomography. These artifacts cannot be removed by simple deconvolution because they are different for every part of the sample, as shown at the various angles labeled in the figure. Furthermore, these artifacts can extend throughout the entire volume and are unresponsive to typical deconvolution methods.

[0028] Figure 2Dand 2E The sample reconstruction using the principles of the present invention is shown to reduce Figure 2C The spatial variation artifacts shown in ;

[0029] Figure 3A is a flow chart of removing artifacts using a progressive method according to the present invention;

[0030] Figure 3B is a flow chart illustrating the training and use of a convolutional neural network (CNN) according to the present invention; and

[0031] Figures 4A-4D Different reconstructions related to the method are shown. DETAILED DESCRIPTION

[0032] The present invention will now be described more fully hereinafter with reference to the accompanying drawings, in which exemplary embodiments of the invention are shown. However, the invention may be embodied in many different forms and should not be construed as limited to the embodiments described herein; rather, these embodiments are provided so that this disclosure will be thorough and complete and will fully convey the scope of the invention to those skilled in the art.

[0033] As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items. Furthermore, unless expressly stated otherwise, the singular and the articles "a," "an," and "the" also include the plural. It should be further understood that the terms: includes, comprises, including, and / or comprising, when used in this specification, specify the presence of the features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or combinations thereof. Furthermore, it should be understood that when an element comprising a component or subsystem is referred to as and / or shown as being connected or coupled to another element, it may be directly connected or coupled to the other element, or there may be intervening elements.

[0034] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. It should also be understood that terms (such as those defined in commonly used dictionaries) should be interpreted as having a meaning consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0035] Figure 1 FIG. 1 is a schematic diagram of an X-ray microtomography system 100 to which the method and workflow of the present invention are applicable.

[0036] Generally, the X-ray microtomography system 100 combines an X-ray microscope system 101 and a computer system 200 to receive projections and calculate a volumetric dataset from these projections.

[0037] The X-ray microscope system 100 includes an X-ray source system 102 that produces a typically polychromatic X-ray beam 104 and a rotating stage 110 having a sample holder 112 for holding and rotating a sample 114 in the X-ray beam from the X-ray source system 102. Images, or X-ray projections, are captured by a detector system 118. The X-ray source system 102, the rotating stage 110, and the detector system 118 are mounted to a base 108 of the X-ray CT system 100. A computer system 200 typically receives and processes these projections and provides overall control of the system 100. The computer system 200 or another computer typically performs tomographic reconstruction using the X-ray projections to create a volumetric dataset.

[0038] In one example, the X-ray source 102 is a cone-beam polychromatic X-ray source. The polychromatic X-ray source is preferably a laboratory X-ray source because of its ubiquity and relatively low cost. However, synchrotron sources or accelerator-based sources are other alternatives.

[0039] Common laboratory X-ray sources include X-ray tubes, in which electrons are accelerated in a vacuum by an electric field and fired into a target piece of metal, emitting X-rays as the electrons decelerate in the metal.

[0040] In one example, X-ray source 102 is a rotating anode reflection target type or microfocus source with a tungsten target. Target materials including molybdenum, gold, platinum, silver, or copper may also be used. Preferably, in some examples, a transmission target configuration of X-ray source 102 is used, in which the electron beam strikes thin target 103 from the back side of the target. X-rays emitted from the other side of target 103 are then used as beam 104.

[0041] When sample 114 is exposed to X-ray beam 104, X-ray photons transmitted through the sample form an attenuated X-ray beam 106 and are received by detector system 118. In some other examples, an objective lens, such as a zone plate lens, is used to form an image on detector system 118 of X-ray imaging system 100.

[0042] In the most common configuration of the detector system 118, a magnified projected image of the sample 114 is formed on the detector system 118 at a geometric magnification that is equal to the inverse ratio of the source-to-sample distance and the source-to-detector distance. Typically, the geometric magnification provided by the X-ray stage is between 2x and 100x, or higher. In this case, the resolution of the X-ray image is limited by the focal spot size, or virtual size, of the X-ray source system 102.

[0043] To achieve high resolution, one embodiment of the X-ray microtomography system 100 also utilizes an extremely high resolution detector 124-1 of the detector system 118 and positions the sample 114 close to the X-ray source system 102. In one embodiment of the high resolution detector 124-1, a scintillator is used in conjunction with an optical microscope objective to provide additional magnification in the range of 2x to 100x or greater.

[0044] Other possible detectors may be included in the illustrated X-ray CT system 100 as part of the detector system 118. For example, the detector system 118 may include a lower resolution detector 124-2. In an example, this may be a flat panel detector or a detector with a lower magnification microscope objective. Configurations of the detector system 118 with one, two, or even more detectors 124 are possible.

[0045] Preferably, two or more detectors 124 - 1 , 124 - 2 are mounted on a rotating turret 122 of the detector system 118 so that they can be alternately rotated into the path of the attenuated light beam 106 from the sample 114 .

[0046] Typically, based on operator-defined parameters, the controller 210 of the computer system 200 instructs the rotary stage 110 via the control interface 130 to position the sample 114 and the detector 124. Upon completion, the controller 210 rotates the sample 114 relative to the beam 104 to perform a CT scan of the sample 114 and saves the projection data 262 to the data store 260.

[0047] In one example, the computer system 200 or other computer includes a graphics or other accelerated processor 220 that analyzes the X-ray projections and may perform calculations required for tomographic reconstruction or for creating a volumetric data set 264 from the X-ray projections 262. A display device 240 connected to the computer system 200 displays information from the X-ray CT system 100. An input device 250, such as a touch screen, keyboard, and / or computer mouse, enables interaction between an operator, the computer system 200, and the display device 240.

[0048] In one example, an operator defines / selects CT scan or calibration parameters using a user interface application executed on computer system 200, whose interface is displayed on display device 240. These include settings for the X-ray acceleration voltage, and settings for the X-ray energy spectrum that define the scan and exposure times on X-ray source system 102. The operator also typically selects other settings, such as the number of X-ray projection images to be created for sample 114 and the rotation angle of rotating stage 110.

[0049] Computer system 200 receives image or projection information associated with each rotation angle of sample 114 from detector system 118, possibly with the assistance of its image processor or other co-processor 220. Typically, image processor 220 combines the projection images using a reconstruction algorithm to create a 3D tomographic reconstructed volume of information of the sample.

[0050] The computer system 200 will typically execute a reconstruction application 254 to reconstruct a volumetric dataset 264 of the sample from its projection data 262 .

[0051] The reconstruction application 254 runs on top of an operating system 252. The operating system 252 is in turn executed by the computer's central processing unit (CPU) 250. In accordance with the present invention, an artifact removal application 256 is also provided.

[0052] Figures 2A-2E Problems encountered when using a tomographic X-ray microtomography device are described.

[0053] Figure 2A The basic tomography geometry is shown. The X-ray source 102 is shown as a point source and is located below or to the side of the sample 114. A planar 2D detector 124 (a direct flat panel X-ray detector, or any indirect X-ray detection module based on photon energy conversion) is located above the sample.

[0054] In the figure, several rays are shown passing through the sample, connecting the source to different pixels of the detector. These rays have different, non-zero exit angles. In practical systems, these rays typically have exit angles between 10 and 30 degrees. Note that angles between -10 and -30 degrees are also possible, as the source can be positioned above and the detector below, resulting in negative exit values.

[0055] Figure 2B A reconstruction of a sample from a scan of the full geometry of Orlov is shown, specifically a DDR4 chip, as a reference.

[0056] Figure 2C A classic FDK reconstruction of a sample is shown. This reconstruction shows the infamous butterfly artifacts 310 that plague rotational tomography. These spatially varying artifacts cannot be eliminated by simple deconvolution because they are different for each part of the sample, as shown by the various angles marked in the figure. Furthermore, these artifacts can extend throughout the entire volume and are unresponsive to typical deconvolution methods.

[0057] The main ray (the ray connecting the source and the detector center) in the tomographic geometry has a 20-degree departure angle. Spatially varying oblique elongation artifacts are primarily caused by the different ray angles distributed across the sample. Such artifacts are generally well-known in tomographic reconstruction. They arise from the fact that non-parallel beams undersample the sample.

[0058] Furthermore, other non-Orlov full geometry computed tomography will also produce similar artifacts. For example, in normal cone-beam tomography, the top and bottom of the scan are affected in exactly the same way as in tomography and show similar artifacts.

[0059] Figure 2D and 2E Shows a sample reconstruction using the principles of the present invention to reduce Figure 2C Spatially varying artifacts shown in .

[0060] Will Figure 2B and 2C A key observation when comparing these two methods is that artifacts in the reconstruction tend to be lower in value than the signal and are proportional to the signal strength. This approach typically relies on assuming that the highest values are the signal. These signals are then used to simulate reconstruction artifacts and subtracted from the original reconstruction.

[0061] Figure 3A is a flowchart of an artifact removal method using a progressive approach according to the present invention, the method being performed by the artifact removal application 256 executing on the computer system 200 .

[0062] In step 410, the computer system 200 or another computer system executing the reconstruction application 254 reconstructs the initial volume image x0 from the projection data 262 using a classical inverse solver. In the illustrated example, the inverse solver is the FDK analytical reconstruction algorithm. Other examples of suitable inverse solvers include the Grangeat algorithm (Grangeat, Pierre. "Mathematical framework of cone beam 3D reconstruction via the first derivative of the Radon transform." Mathematical Methods in Tomography: Proceedings of a Conference held in Oberwolfach, Germany, June 5-11, 1990. Springer Berlin Heidelberg, 1991) and the Katsevich algorithm (Katsevich, Alexander. "Theoretically exact filtered backprojection-type inversion algorithm for spiral CT." SIAM Journal on Applied Mathematics 62.6 (2002): 2012-2026).

[0063] Threshold list T 0,1..M It is also defined based on the initial volume image x0. In one example, the threshold T 0,1..M is chosen to be uniformly distributed over the entire range of CT reconstruction values of the initial reconstruction to represent the linear absorption of the sample. This division occurs in k segments and stops at the Mth threshold (M <K)。

[0064] Then, in the first iteration (k=0), the artifact removal application 256 applies a first or lowest threshold value T0 to the initial volume image x0 in step 412. One possible approach is to consistently select T 0，1..m , in this case, if the lowest value (grayscale) of the 3D volume is "a" and the highest value is "b", then T0 will be (ab) / M. On the other hand, T 0，1..m A non-uniform distribution can also be selected.

[0065] In the first pass, the first thresholded volume center image is shown in inset 450. Maintaining the threshold T k Above the value, y k =max(0,x k -T k ).

[0066] Thresholded volume image y k It is then forward projected, and the forward projected image is reconstructed into a new volume by the reconstruction application 254 in step 414. k When a signal of the current threshold level is contained, a forward projection is performed and a new volume 264 is reconstructed from its projection 262 .

[0067] Illustration 452 shows a simulated reconstruction based on the top value. It contains reconstruction artifacts similar to the tail w k .

[0068] Next, the volume is updated by subtracting the artifact from the volume image: k+1 =x k -w k Inset 454 shows the volume after the first iteration.

[0069] The process moves to the next iteration, k=k+1, and uses the new value of k and the updated value x k＝k+1 Execute steps 412, 424, and 416.

[0070] Typically, the iteration is repeated until a predetermined threshold T is reached. M After stop

[0071] According to one embodiment, additional steps are added to handle certain types of features that cannot be reconstructed correctly. Planar objects in the reconstruction plane (such as a smooth wide copper plate) will not be represented in the reconstruction. Small holes in this smooth layer may appear in the reconstruction as having negative density values (this is a non-physical result that can be corrected as described herein). This violates the laws of physics (negative density materials do not exist). Step 418 addresses this issue by adjusting the reconstruction to add a planar material layer to offset the negative values to zero. The thin sheet with the hole can then be visualized correctly.

[0072] Finally, as shown in step 420, all signals y are accumulated. k To obtain the final artifact-free reconstruction 264, i.e. In other words, for each step k, y k are stored in memory for summation. By creating an auxiliary variable z k , in each iteration, z k Use z k +y k Update and finally return z k The result is shown in Figure 480.

[0073] Demonstrated using experimental results obtained from the same DDR4 chip Figure 3AThe pipeline shown in . Illustrations 450, 452, 454, 456, 458, and 460 are the z M and x k Representative results.

[0074] An artifact-free result using this method is shown in illustration 480 .

[0075] Furthermore, in order to speed up this artifact removal process, by applying the orthographic projection, the M Re-simulate the reconstruction where artifacts are present and reconstruct the volume based on these projections It can be used to train convolutional neural networks (CNNs).

[0076] Figure 3B is a flow chart illustrating the training and use of a CNN implemented by the artifact removal application 256 executing on the computer system 200 in accordance with the present invention.

[0077] In step 510, FDK reconstruction is performed on one or more samples to be used for training in step 510. Inset 511 shows the reconstruction.

[0078] Then in step 512, execute Figure 3A The progressive artifact removal process described above produces the artifact removal shown in inset 513.

[0079] Optionally, any negative density values resulting from thresholding can be corrected (eg, reset to zero or replaced with a physically meaningful baseline).

[0080] Then in step 514, forward projection and filtered back projection are performed.

[0081] This information is then used to train a CNN in step 516 to predict the artifacts to be removed.

[0082] Apply the trained neural network to the initial reconstruction x0 and get the inferred result CNN θ (x0). The trained network will generalize to samples with similar structures, which can greatly reduce the time of artifact removal.

[0083] In more detail, in step 518, FDK reconstruction is performed on the sample of interest. This produces a reconstruction, albeit with artifacts as shown in inset 519. Inference is then performed using a CNN trained on similar samples in step 520. This results in artifact removal in step 522. Inset 523 shows the result.

[0084] It should be noted that this speedup of deep learning is not essential, as it also makes the pipeline parameter-free. Note that a stopping criterion T needs to be pre-selected. m Our learning-based approach allows the network to be pre-trained with a carefully chosen stopping criterion to provide the best reconstructions and includes other image enhancement methods we developed, which can be easily combined by training an end-to-end neural network.

[0085] Figures 4A-4D Different reconstructions related to the method are shown.

[0086] Figure 4A An FDK reconstruction of the tomographic device raw projections is shown.

[0087] Figure 4B An iterative artifact-removed reconstruction is shown.

[0088] Figure 4C The FDK reconstruction result x0 of the first iteration is shown.

[0089] Figure 4D The reconstruction results based on high-level learning are shown.

[0090] Furthermore, to reiterate, although a tomographic slice is provided as an example, the method is generally applicable to other non-Orlov complete computed tomography slices (eg, high cone angle, missing views).

[0091] While the invention has been particularly shown and described with reference to preferred embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made without departing from the scope of the invention as encompassed by the appended claims.

Claims

1. A method for reconstructing a tomographic volume from projection data, the method comprising: thresholding the current reconstruction to create a thresholded reconstruction; creating simulated reconstructions from the thresholded reconstructions; subtracting the simulated reconstruction from the current reconstruction to create the current reconstruction for the next iteration; as well as The final reconstruction is created by adding the thresholded reconstructions.

2. The method according to claim 1, wherein The step of creating the final reconstruction removes spatially varying artifacts such as tomographic artifacts and / or high-angle cone-beam artifacts.

3. The method according to any one of claims 1 or 2, wherein The threshold is set based on the extent of the initial reconstruction.

4. The method according to any one of claims 1 to 3, wherein Threshold the current reconstruction by zeroing out values below the threshold.

5. The method according to any one of claims 1 to 4, wherein Creating the simulated reconstruction includes forward and back projection from the thresholded reconstruction. The method according to claim 1 , further comprising assuming that the highest value is a signal.

7. The method according to any one of claims 1 to 6, wherein Prior knowledge of non-Orlov missing objects is added back into the tomography to make the reconstruction more complete.

8. The method of any one of claims 1-7, further comprising training a neural network based on the reconstruction; and further comprising adjusting the reconstruction for negative density values.

9. An X-ray microtomography system for performing an artifact removal application implementing the method of any one of claims 1 to 8.

10. A computer-implemented method for reconstructing a tomographic volume from projection data, the method comprising: thresholding the current reconstruction to create a thresholded reconstruction; creating a simulated reconstruction based on the thresholded reconstruction; subtracting the simulated reconstruction from the current reconstruction to create an updated current reconstruction for use in subsequent iterations; and The thresholded reconstructions are added to form the final reconstruction.

11. The method according to claim 10, wherein: Thresholding the current reconstruction includes selecting a plurality of thresholds based on a range of absorption values in an initial reconstruction of the projection data.

12. The method according to claim 10, wherein: Creating the simulated reconstruction includes forward projecting the thresholded reconstruction to generate simulated projection data, and back projecting the simulated projection data to form the simulated reconstruction.

13. The method of claim 10, further comprising correcting negative density values in the updated current reconstruction after the subtracting step to ensure a physically meaningful density distribution.

14. The method according to claim 10, wherein: The projection data are generated in tomographic or high cone-beam imaging geometries, and spatially varying artifacts arise from missing views or non-Orlov complete data acquisition.

Images