Machine learning based detachment correction for front layer

Abstract

A system (SYS) and related methods for acquisition lag correction in X-ray imaging. The system comprises an input interface (IN) for receiving an input comprising an input image acquired by a front layer detector (FLD) in a multi-energy X-ray imaging device (IA). A corrector component (CC) processes the input image into a corrected version representing an estimate of image information at a lower acquisition lag than the input image experiences an acquisition lag. Specifically, hysteresis effects and associated artifacts are removed, and the corrected version is an estimate of the input image in acquisition settings without such acquisition hysteresis.

Description

Technical Field

[0001] The present invention relates to a system for acquisition lag correction in X-ray based imaging, an imaging arrangement comprising such a system, a training system for training a machine learning model used in such a system, a training data generator system for generating training data used in such a training system, related methods, a computer program unit and a computer-readable medium. Background Art

[0002] For 2D and 3D imaging, lag artifacts are an important problem in X-ray based imaging. They originate from the delayed signal accumulation in detector pixels, especially in the detector areas that are in the shadow of an attenuating object and are then irradiated by the direct X-ray beam. Another frequent source of lag artifacts is a sequence where some images are acquired at a high X-ray dose, followed by low-dose images. In this case, there is a delayed signal spillover from the high-dose images interfering with the low-dose images. Both lag effects result in artifacts in 2D imaging and tomographic 3D imaging, such as in cone-beam CT (CBCT) volumes or other volumes.

[0003] Some such artifacts can be mitigated by applying hardware solutions, such as the dorsal reset optical system described, for example, in "Latest advancements in state-of-the-art aSi-based X-ray Flat Panel Detectors" by T. Ducourant et al. (published in Proc. SPIE 10573, Medical Imaging (2018)).

[0004] However, it has been found that artifacts still remain in spectral imaging. Spectral imaging is multi-energy imaging where multiple sets of projection data are acquired at different radiation energy ranges. These sets can be computationally combined to produce 2D or 3D images with material-specific contrast, which is particularly useful in therapy and diagnosis. Summary of the Invention

[0005] There may be a need for improved X-ray imaging, particularly improved spectral imaging.

[0006] The object of the present invention is achieved by the subject matter of the independent claims, where further embodiments are incorporated in the dependent claims. It should be noted that the aspects described below of the present invention apply equally to an imaging arrangement comprising a system, a training system for training a machine learning model used in the system, a training data generator system for generating training data used in the training system, related methods, a computer program unit and a computer-readable medium.

[0007] According to a first aspect of the present invention, there is provided a system for acquisition lag correction in X-ray imaging, comprising:

[0008] An input interface for receiving inputs, the inputs including an input image acquired by a front-layer detector in a multi-energy X-ray imaging device and an additional image acquired by a back-layer detector of the multi-energy X-ray imaging device. The back layer preferably has circuitry (such as a back-side reset light or other hardware solution) for reducing the time lag effect, and it is preferably activated when the additional input image is acquired by the back-layer detector, the back-layer detector including circuitry for reducing the time lag effect, and the additional image having been acquired by the back-layer detector when the circuitry was activated.

[0009] Furthermore, the system includes a corrector component configured to process the input image and the additional input image for generating a corrected input image, the corrected input image representing an estimate of the image information of the front-layer detector at an acquisition lag lower than the acquisition lag experienced by the input image.

[0010] Thus, 2-channel input data is provided to the correction component, allowing for better performance and in particular higher accuracy, as will be explained below.

[0011] In an embodiment, the processing performed by the corrector component includes scaling the input image using a scaling factor obtained based on a series of previous input images and additional previous input images preferably acquired respectively by the front-layer detector and the back-layer detector of the multi-energy X-ray imaging device.

[0012] In an embodiment, the corrector component is implemented based on a trained machine learning model.

[0013] In an embodiment, the machine learning ML model was previously trained based on a supervised or unsupervised learning setting.

[0014] In an embodiment, the trained ML model includes an artificial neural network.

[0015] In an embodiment, the artificial neural network is configured for multi-scale processing. For example, an artificial neural network (NN) of the U-net type architecture can be used. Multi-scale processing using filtering or other techniques (not necessarily of the NN type) is also envisioned herein. In an embodiment, the ML model is trained in a supervised setting based on training data including training input data and associated ground truth data, wherein the training input data includes previous images acquired by the front-layer detector or the front-layer detector, and wherein the ground truth data includes previous images acquired by the back-layer detector or the back-layer detector including circuitry for reducing the time lag effect in cases where such circuitry was activated.

[0016] Such a circuit for reducing the time lag effect may include, for example, the aforementioned backside reset optical facility coupled to the far side of the back layer detector. As further elaborated herein, in a multi-layer detector, the front layer detector typically cannot provide such a lag correction circuit. The training input data may include previous images acquired by a front layer detector (or a single-layer detector) in which there is no lag correction circuit (or, in the case of a single-layer detector, present but not activated) and / or previous images acquired by a single-layer detector in which there is and is activated a lag correction circuit. Thus, in effect, the training input data may include front layer previous image data having a time lag effect.

[0017] Thus, in an embodiment, 2-channel training data input, a front layer image, and a back layer image are used. This allows for better consideration of artifacts: typically, the observed artifacts are the result of at least two effects and their interaction: i) lag caused by the way detector signals accumulate due to signal spillover from one detection event to another, and ii) the spectral distribution of the detection signals caused by the various materials of the imaged object. The proposed two-channel ML method represents a regularization scheme that promotes the model to learn to distinguish between the two effects i), ii): i) is eliminated or reduced, while ii) is better retained in a purer form. This helps for better spectral processing through spectral imaging algorithms as envisioned downstream. Such spectral imaging algorithms include one or more of the following: material decomposition, decomposition into Compton and photoelectric images, calculation of virtual mono-energetic images, virtual contrast-only images, virtual non-contrast images, and any one of other images that require multi-energy information. The two-channel method can also be used during deployment, post-training.

[0018] In an embodiment, the front layer detector is one of the single-layer detector X-ray imaging devices, or where the front layer detector is the multi-energy X-ray imaging device or one of the multi-energy X-ray imaging devices with the back layer detector removed.

[0019] Specifically, the lag effect and associated artifacts are removed. Thus, the corrected version is an estimate of the input image in an acquisition setting without (or only negligible) such acquisition lag.

[0020] The proposed system solves the problem of acquisition lag-based artifacts in projection images and their reconstruction in the image domain. Different from the back layer detector, in a multi-layer spectral detector system (such as a double-layer detector system), its front layer detector module cannot be equipped with the aforementioned backside reset light or a similar hardware-based lag artifact solution because the X-ray shadow of its components (such as power lines, etc.) will be fed into the associated back layer X-ray image to cause its own artifacts there. The proposed system computationally solves this problem.

[0021] The system can be used for all imaging protocols and is expected to be most relevant in angiographic DSA protocols and CBCT scans with asymmetric or non-isocentric objects, where the hysteresis artifacts are most severe.

[0022] The system is not limited to use in dual-energy imaging, but is also capable of being used in spectral imaging setups with three or more layers, where only the last layer is equipped with a dorsal reset light or a similar hardware-based hysteresis error correction function.

[0023] Due to its hardware-based hysteresis error correction function, the dorsal layer of the dual-layer detector system does not record hysteresis artifacts or records significantly fewer hysteresis artifacts compared to the front layer. This gives a strong indication of whether the front layer data has suffered from delayed signal accumulation or attenuation during temporal (video) 2D or 3D (e.g., CBCT) imaging.

[0024] On the other hand, there is provided an imaging arrangement including the system and a multi-energy X-ray imaging device, the multi-energy X-ray imaging device including a front layer detector and a dorsal layer detector, preferably as part of a dual-layer or multi-layer detector system.

[0025] In yet another aspect, there is provided a training system configured to train an ML model based on training data.

[0026] In yet another aspect, there is provided a training data generator system configured to generate training data on which the training system trains a model.

[0027] In yet another aspect, there is provided a computer-implemented method for acquisition hysteresis correction in X-ray imaging, including:

[0028] Receiving an input including an input image acquired by a front layer detector in a multi-energy X-ray imaging device and an additional input image acquired by the dorsal layer detector of the multi-energy X-ray imaging device, the dorsal layer detector including circuitry for reducing the time lag effect, the additional image having been acquired by the dorsal layer detector when the circuitry was activated; and

[0029] Processing the input image and the additional input image to generate a corrected input image, the corrected input image representing an estimate of the image information of the front layer detector at an acquisition lag lower than the acquisition lag experienced by the input image.

[0030] In another aspect, there is provided the use of the corrected version in computer-implemented reconstruction operations and / or computer-implemented spectral image processing algorithms. Preferably, the associated dorsal layer image (preferably acquired with the delay reduction circuitry activated) is used in this way together with the corrected front layer image.

[0031] In yet another aspect, a method for training an ML model based on training data is provided.

[0032] In yet another aspect, a method for generating training data for a training method is provided.

[0033] The method for generating training data may include i) imaging a phantom from multiple relative directions with and / or without activating such a lag reduction circuit (LRC) of the detector for a single-layer detector in the same imaging geometry, and / or ii) imaging the phantom using a multi-layer detector with the phantom repositioned.

[0034] The method can be used for supervised or unsupervised learning.

[0035] The method may further include imaging a physical phantom with the lag reduction circuit of the detector backplane activated or deactivated.

[0036] To implement the above scaling embodiments, the "lag-free" backplane signal is compared with the frontplane signal to estimate the lag component when imaging the phantom. The frontplane lag artifact can be estimated by recording a 2D image sequence with and without dorsal lag correction for the backplane. By comparing the deviation between the frontplane image and the backplane image with and without lag correction of the 2D image sequence, the correct scaling correction can be estimated to account for the lag artifact of the frontplane detector.

[0037] In yet another aspect, a computer program unit is provided that, when executed by at least one processing unit, is adapted to cause the processing unit to perform any of the above methods.

[0038] In yet another aspect, at least one computer-readable medium is provided, having a program unit stored thereon, or having a trained machine learning model stored thereon.

[0039] "User" refers to a person who operates an imaging device or supervises an imaging process, such as a medical staff member or other person. In other words, the user is generally not a patient.

[0040] In general, the term "machine learning" encompasses computerized arrangements (or modules) that implement machine learning ("ML") algorithms. Some such ML algorithms operate to adjust machine learning models configured to perform ("learn") tasks. Other ML operates directly on training data and does not necessarily use models such as a training data corpus. This adjustment or update of the training data is referred to as "training". Generally, the task performance of an ML module can be measurably improved through training experience. Training experience can include suitable training data and the exposure of the model to such data. Task performance can be improved if the data better represents the task to be learned. "If the training data well represents the distribution of examples on which the final system performance is measured, the training experience helps to improve performance". Performance can be measured by an objective test based on the output produced by the module in response to feeding it test data. Performance can be defined according to a specific error rate to be achieved for given test data. See, for example, "Machine Learning" by T.M. Mitchell (page 2, section 1.1, page 61.2.1, McGraw-Hill, 1997).

[0041] "Lag" involves a delay in the accumulation of signals in an X-ray detector when exposed to X-ray radiation of a given intensity. Thus, the values recorded at the pixels are typically incorrect and do not represent the true signal intensity. Consequently, lag results in temporal artifacts in the images recorded by such detectors. The cause of lag may be found in defects in some components of such detectors, such as charge trapping in semiconductor crystals. The artifacts in the projection images caused by such lag are transferred to the images derived from such projection images, such as the images reconstructed by a tomographic reconstruction algorithm from such projection images.

[0042] In the context of machine learning embodiments (particularly with respect to the training of machine learning models), a "front layer detector" can refer not only to a multi-detector system but also to a more conventional detector system (single layer detector / system), where there is only a single such detector layer. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Exemplary embodiments of the present invention will now be described with reference to the following drawings, in which:

[0044] Figure 1 A schematic block diagram of a medical imaging arrangement is shown;

[0045] Figure 2 A multi-layer X-ray detection system is shown;

[0046] Figure 3 A block diagram of a system for correcting projection data for lag artifacts is shown;

[0047] Figure 4illustrates a training system for training a machine learning model to be used in the system of Figure 3 ;

[0048] Figure 5 illustrates various training configurations that can be used as training data and combined with the model trained for the system above during inference; Figure 4 during inference;

[0049] Figure 6 shows a block diagram of a machine learning model that can be used in the embodiments contemplated herein; and

[0050] FIG. 7 shows a flow chart of a method for reducing imaging artifacts, a method for training a machine learning model, and a method for generating training data. DETAILED DESCRIPTION

[0051] Referring now initially to the block diagram of Figure 1 which shows a medical imaging arrangement IAR as contemplated herein in an embodiment.

[0052] The imaging arrangement IAR may include an X-ray based imaging device IA which is operable for example for therapeutic or diagnostic purposes to produce projection data Λ in an acquisition operation that can be processed by a computing system SYS.

[0053] The computing system SYS is operable to process the projection data into a tomographic image which can for example be stored, displayed or otherwise processed to assist in treatment and / or diagnosis, for statistical or educational purposes, or other applications such as planning, control tasks etc.

[0054] The imaging device IA (also referred to herein as the "imager" for brevity) is X-ray based and is configured herein for multi-energy X-ray imaging, also known as spectral imaging. Although pure projection based imaging such as radiography is not excluded herein, tomographic imaging is mainly contemplated herein. The imagers contemplated herein that are configured for tomographic imaging include a CT scanner as shown in Figure 1 or an interventional imager of the C-arm or U-arm type as sometimes used for example in a CathLab or others. Tomosynthesis based imaging is also contemplated herein such as in mammography or dentistry etc.

[0055] In short, and as will be expanded more fully, the system SYS as contemplated herein is configured to reduce (if not eliminate) certain image artifacts that may corrupt the acquired projection data in a certain type of spectral imaging setup.

[0056] Spectral imaging, which will be mainly mentioned hereinafter, allows material-specific imaging. That is to say, the obtainable images have a contrast that varies mainly or uniquely with the concentration of the specific material of interest. Specifically, different from typical projection-based imaging of the traditional energy integration type, spectral imaging avoids the unwanted contrast contributions of intermediate structures or other materials.

[0057] Such materials of interest whose specific distribution exists in the field of view of the imager can be sought, including types of contrast agent materials (iodine, barium or their compounds), such as those that can be used in contrast agent-assisted protocols to image structures that are otherwise nearly transparent to X-radiation. The contrast agent accumulates particularly in the regions or organs of interest, such as in specific vascular segments in the cardiovascular system or elsewhere for cardiac imaging, or in segments of the intestine in abdominal imaging, etc. This allows for more accurate images, which can better assist in treatment, diagnosis, planning, control of medical robots, etc.

[0058] Spectral imaging requires an imaging setup in which the imager IA is configured for multi-energy projection data acquisition. That is to say, for each pixel in the projection data, there are at least two intensities recorded at different energy levels. Thus, in practice, (at least) two sets of projection data are acquired, at least one set for each energy range. This is achieved in an embodiment by exposing the region of interest (such as a specific body or organ part) to X-radiation with different energy spectra. This can in turn be achieved by a detector-side solution or a source-side solution, etc. The detector-side solution is of primary interest herein and thus includes a multi-layer detector module. This allows the region of interest to be exposed to radiation of different spectra, thus enabling multi-energy spectral imaging.

[0059] The acquired multi-energy projection data can be processed by a spectral processor component SP to obtain a spectral image with the desired material-specific contrast of the target material of interest.

[0060] It has been observed that in such multi-energy projection data acquisition required in spectral imaging, inaccurate detection of intensity signals may occur, which in turn leads to image artifacts in the projection images, or due to the fact that these situations are usually amplified, leads to image artifacts in the 3D reconstruction of such material-specific spectral images. Other types of spectral images that are also affected by artifacts include virtual mono-energy images, virtual non-contrast images, etc.

[0061] One such cause of this type of artifact has been found to lie in the manner of image acquisition, i.e., the manner of recording multiple energy intensities by certain multi-detector systems XD, which is an example of the detector-side spectral imaging solution mentioned. This paper presents a computer-implemented correction system CS that corrects such artifacts that cause errors in the acquired multi-energy projection data. The thus-corrected projection data can then be passed through the output port OUT for further processing. Such further processing may include spectral processing of the corrected input image as needed and / or reconstructing the corrected input image into a tomographic image in the image domain.

[0062] The corrector system CS can be implemented by machine learning, as will be described in more detail below. The corrector system CS can be implemented on a single computing platform or across multiple computing platforms or systems, such as on multiple computing devices, servers, etc., preferably at least partially interconnected in a data communication network in a distributed manner (cloud setting). The artifact corrector system CS can be integrated into the imaging device IA, such as in the operation console OC, or in the workstation of the imager IA, or can otherwise be integrated into the entire computing system SYS for processing projection data mentioned above. Alternatively, the corrector system CS is integrated into the detector system DX of the imager.

[0063] Before explaining in more detail the artifact corrector system CS for multi-energy projection data, first refer to the components of the imaging device IA to better provide a functional context for the operation of the corrector system CS.

[0064] Preferably, the X-ray based imager IA allows for obtaining cross-sectional images of the ROI, preferably 3D images. To this end, the imager IA is configured for multi-directional (schematically indicated as "α" in the figure) projection data acquisition around the ROI. The acquisition (path) or "scan" may not necessarily define a full 360° angular range, but an angular range of 180° or even less may be sufficient in this paper. The acquisition path may not necessarily define a circular arc, although this may be the case in most instances. Helical paths are also envisioned in this paper. Thus, the present disclosure herein is not limited to axial paths, but such paths as in a "step-and-shoot" setting are not excluded herein.

[0065] In some embodiments, the imaging device IA can be of a rotational type to enable multi-directional projection data acquisition operations. Such a rotational imaging device allows for the collection of projection images λ along different projection directions α during a scan around an area of interest (such as around a stenosis in the aforementioned coronary vessel, etc.). The imaging device IA includes an X-ray source XS and an X-ray sensitive detector system XD. In some embodiments, the X-ray source XS and the X-ray sensitive detector XD opposite thereto are preferably arranged on a rotatable gantry MG mounted in an optional stationary gantry SG. The examination area can be defined by the gantries SG, MG and is defined between the source XS and the detector system DX, within which the patient to be imaged or at least the area of interest is located. Some imagers do not have a stationary gantry SG, and their rotatable gantry G (appropriately journaled) can have a recess within which the examination area ER is so defined, as Figure 1 shown in the CT scanner shown. Thus, the rotatable gantry MG can have a "C", "U" or similar shape, such as in the C-arm imaging device contemplated herein. A biplane imaging setup with two detectors having intersecting imaging axes (e.g., at right angles) is not excluded herein. Instead of using a C-arm / U-arm system, in some embodiments herein, the CT scanners shown in the figures or similar are contemplated, and they can be of any generation, but preferably at least the 3rd generation, wherein the X-ray source XS and the detector D are mounted in the rotatable gantry MG such that they rotate together around the patient. However, for multi-directional projection data acquisition, higher generation CT setups are also contemplated, wherein there is not necessarily a mechanical rotation of the source and the detector. Instead, a fixed ring of detectors around the ROI and / or a fixed plurality of X-sources arranged in a ring around the ROI are used. Cone-beam imaging or other divergent beam geometries are also contemplated herein, and thus imaging using a helical scan path caused by a circular path of the X-ray source is also contemplated, e.g., using a translational movement along the patient longitudinal axis Z relative to the imaging device IA and the ROI. In most cases, such translational movement is caused by the patient bed PS on which the patient PAT translates during projection data acquisition and passes through the examination area ER. However, as previously mentioned, scan paths having geometries other than circular / helical are not excluded herein.

[0066] Generally, imaging / acquisition includes exciting the X-ray source XS such that an X-ray beam XB emanates from the focal point of the source XS, passes through the examination area (within which is the area of interest) and interacts with the patient tissue substance. The interaction of the X-ray beam XB with the tissue substance causes the beam to be modified. The modified beam can then be detected at the detector system XD.

[0067] The images obtainable by the imager IA can be stored in the memory DB, for example, in an image database, can be displayed on the display device DD by the visualizer VIZ, or can be otherwise processed as required according to the medical task at hand.

[0068] The spectral processing component SP (referred to simply as "spectral processor") implements spectral (image) processing algorithms in suitable software or hardware (or partly in both). The spectral processor component SP can be part of the computing system SYS, the console OC, or the workstation. The spectral processor SP processes multi-energy projection data to compute spectral images, i.e., material-specific tomographic or projection images of interest. One such spectral image algorithm is the 1976 method of R. Alvarez and A. Macovski (e.g., published as "Energy-selective reconstructions in X-ray computerized tomography", Phys Med Biol, Vol. 21(5), pp. 733-44) and similar methods, where two (or more) sets of projection images for different energy ranges per pixel are fed into a system of simultaneous linear equations that allows the solution of the unknown material-specific attenuation coefficients. Then, N (≥2) energy levels can help solve for N different materials. Other types of material decomposition techniques, related or unrelated to the Alvarez method, have been devised, and each method is envisioned in embodiments of the spectral processor SP herein.

[0069] If spectral tomographic images are sought, the system SYS can also include a reconstructor RECON configured to implement tomographic reconstruction algorithms (such as FBP, iterative, algebraic, etc.) to transform the projection images from the projection domain to the image domain (the part of the 3D space where the patient PAT or object to be imaged during imaging is located). The spectral processing performed by the spectral processor SP can occur in pre-reconstruction in the projection domain or post-reconstruction in the image domain. Alternatively, spectral processing and reconstruction are combined.

[0070] The multi-layer X-ray detector system XD that can be used for dual-energy imaging in this article is shown in more detail in the cross-sectional view provided by the Figure 2 now being referred to.

[0071] The multi-layer X-ray detector system XD can include a housing H that houses a plurality of detector modules, each detector module being capable of detecting X-radiation. The plurality of detector modules are such as Figure 1The arrangement shown is in layers and includes a front layer detector module FLD (also simply referred to herein as the "front layer") and a back layer detector module BLD (also referred to herein as the "back layer" for brevity). This is the basic setup for dual-energy imaging and will be mainly referred to herein, but those skilled in the art will understand that the principles described herein equally apply to multi-energy detector systems having more than two such layers formed by multiple such modules (such as three, four or more) in a multi-energy imaging setup for more than two (N) energy ranges (one above the other). Such layers include a front layer and a back layer and one or more intermediate layers therebetween. However, for the current purpose, all layers other than the front layer may be referred to herein as the "one / the" back layer module because when viewed from the X-ray source XS, they are all arranged behind the front layer. Thus, the front layer detector FLD is spatially closer to the X-ray source XS and is located proximal to the X-ray source XS compared to any one of the other one or more back layers (such as a single back layer BLD) distal to the X-ray source. That is, the qualifiers "front", "back" and "intermediate", etc. will be understood herein as references to the relative proximity of the respective layers to the X-ray source XS.

[0072] Each of the layers FLD and BLD includes an array of X-radiation sensitive pixels, preferably in a 2D layout of rows and columns. Each layer acts as a transducer to convert the incident X-ray intensity into a corresponding varying charge. As the incident X-radiation that exits the patient tissue PAT after interaction, it is detected at the front layer first FLD to generate such a charge. The charge is then converted into a numerical format, such as a matrix of numbers (intensity values), by a suitable A / D circuit, with each number representing the corresponding X-ray intensity that generated the specific charge and varying therewith. After interacting with the first layer, the X-radiation passes through the first layer and interacts with each subsequent layer (such as with the back layer BLD) in the same manner to similarly generate their own set of charges, which are also converted by a suitable A / D circuit into a second set of projection data / projection image, and so on for any other layers (if any).

[0073] Since the X-radiation passes through the first layer FLD, the spectrum of the radiation leaving the front layer has been changed, and it is this radiation with the changed spectrum that then interacts with the back layer BLD, thereby allowing spectral imaging because the two sets of projection data obtained are derived from X-radiation of different spectra. Then, the two intensity values of each pixel included in at least two sets of projection images obtained by the multi-energy detection system XD can be processed by the spectral processor SP using the spectral imaging algorithm as described above. The artifacts mentioned in the context of a hierarchical multi-energy X-ray detection system XD as described are due to a certain acquisition lag caused by the way the BLD and FLD layers record the intensity of the incident radiation and convert the incident radiation intensity into charge. Indirect or direct conversion type modules are envisaged for the front FLD and the back layer BLD. In each such layer of any type of module, a semiconductor crystal layer is used. The interacting radiation is converted into a cloud of electron-hole pairs in such a crystal, and the signal chain in each detector module uses these pairs in generating the associated measurement signal (i.e., the charge mentioned). In a detector module of the direct conversion type, such pairs diffuse across the crystal to generate a corresponding charge that can be sensed. In a detector module of the indirect conversion type, such pairs are used to generate visible light, which is then sensed by a photoelectric sensor. However, in each case of direct or indirect conversion, the operation depends on the generation and / or movement (diffusion) of such electron pairs in the corresponding crystal layer. The generation or movement of such electron-hole pairs and their interaction within the detector module setup can be hindered due to defects in the crystal. Such defects lead to the formation of charge traps that unfavorably interfere with such electron-hole pairs (EHP). In particular, such EHP or parts thereof (electrons or holes) may be trapped in such traps and cannot respond to the incident X-radiation as expected to generate the associated charge or photons that should represent the corresponding local radiation intensity. That is, the expected charge is not generated at all, or if it is generated, it is provided with a delay. This effect is called acquisition lag, a time effect that causes artifacts. Thus, each pixel exposed to different radiation intensities may respond in a delayed manner, resulting in incorrect recording of intensity values for some pixels in each layer. This effect is significant in cases where a large dynamic range is required and as described above in Figure 1This is more common in medical images that record image signals at high speeds in similar rotational tomography systems as described. High image gradients can contribute to such lag artifacts. For example, during the rotation of the detector, at a certain moment, a pixel may be behind a highly attenuating object such as bone, but at the next moment, once the pixel moves out of the shadow of the highly attenuating object, the pixel may be exposed to much higher radiation within a fraction of a millisecond. For the described type of detection module that suffers from charge trapping, such a high-intensity gradient may be too difficult to handle, which in turn leads to delayed signal accumulation. As a result, the recorded projection images may suffer from artifacts, that is, incorrectly recorded and resulting intensity values, which are even fed into the reconstructed images.

[0074] The proposed corrector system CS is configured herein to eliminate or at least reduce, in short, correct such lag artifacts. This described effect and temporal artifacts of such lag due to delayed signal capture can be mitigated to some extent using hardware solutions. One such solution can include a lag reduction circuit LRC attached to the back layer detector module BLD. Such a circuit (e.g., a backside reset optical circuit) can be implemented as a matrix of visible light emitting diodes placed at the distal side of the back layer detector module BLD. When the lag reduction circuit LRC is activated, the semiconductor (crystal) layer of the back layer BLD is exposed to visible light before some or each X-ray exposure. It has been found that this exposure of the crystal to visible light reduces the obstructive effect of such charge traps by "filling" the potential wells at a large depth in the semiconductor lattice energetically.

[0075] Due to its setup as explained above, the light reduction circuit LRC cannot be arranged at any front layer detector FLD because the materials of the components of such a circuit LRC may in turn cause artifacts and a reduction in efficiency, especially due to the shadow that such a circuit will project onto the underlying back layer detector BLD. Unwantedly, the back layer detector module BLD will then record such a shadow, thus completely compromising the projection data.

[0076] Therefore, what is proposed herein is to configure the corrector system CS to computationally remove such lag-related artifacts from any front layer projection images detected by any front layer detector FLD. The newly proposed corrector system CS does not require intervention in signal interference hardware. The lag reduction circuit LRC is generally effective enough to reduce or completely avoid the described lag-related artifacts in the projection images recorded by the back layer BLD. For the current purpose, the projection images detected by the back layer can be considered artifact-free. Therefore, the projection images detected by the back layer do not have to be separately processed by the corrector system CS, and only the front layer images detected by one or more front layer detector modules FLD need to be so processed.

[0077] For simplicity, the projection image detected by the front layer detector module FLD can be simply referred to as "front image / image" in this article, rather than the projection image recorded by the back layer detector module BLD, which will sometimes be correspondingly referred to as "back image / image" therein.

[0078] Now refer to Figure 3 the flowchart in

[0079] which gives more details of the hysteresis-induced artifact reducer or corrector system CS. In particular, a front image X recorded by one of the front layer detector modules FLD is received at the input port IN of the corrector system CS. The corrector component CC processes the thus received front projection image and produces its corrected version X', which is output at the output port OUT. The thus corrected front image X' can then be stored, sometimes visualized by a visualizer VIZ on a display device DD, or can be passed together with the back image to a spectral processor SP for spectral processing and / or a reconstructor RECON to transform into the image domain to obtain a spectral cross-section / tomographic image of the 3D distribution of a specific material of interest as described. If there are more than one front layer, the system SYS can jointly or separately process more than one front layer as needed. If it is necessary to better process the specific characteristics of each such front layer, there can be a separate constructor system CS for each front layer.

[0080] Thus, a conventional energy-integrating single-layer detector system uses a similar detector module, even having the same construction and manufacturing as the detector module used in the dual-energy setup described in Figure 2 Therefore, the front layer detector FLD of the spectral detector XD can be well used in a conventional single detector layer image, and the single detector layer image can be well used as the FLD module described for dual / spectral imaging. This construction fact can be used in this article to easily obtain suitable training data to train the machine learning model M, as will be described in more detail below.

[0081] However, this machine learning model implementation may not necessarily be required in this document, as the applicant has found that there is sometimes a surprisingly simple relationship between the back layer image that has been artifact-corrected and the corresponding front layer image corrupted by hysteresis artifacts. Experiments have found that these two are related by scaling with a constant scaling factor for the entire image, or by a scaling map with scaling entries for each individual pixel location. The (one or more) scaling factors can be determined by a linear fit based on recording a series of such front and back image pairs. Then, this linear modeling setup can reveal a good approximation of the scaling factor, which can be used by the corrector component CC in a simpler embodiment. Specifically, this scaling method is based on the scaling of the "hysteresis-free" back image and compares it with the associated front image to estimate the hysteresis component. Alternatively, it can be the back image that is scaled. More specifically, the front layer hysteresis artifacts can be estimated by recording a 2D image sequence with and without the hysteresis reducer LRC for the back layer BLD. By comparing the deviation between the front layer image and the back layer image with and without hysteresis correction for such a 2D projection image sequence, the scaling factor for correction can be estimated by a linear fit in order to account for the hysteresis artifacts that damage the front layer detector module FLD. Then the (one or more) scaling factors can be applied to the corrupted front image to reveal the corrected version.

[0082] However, depending on the properties of the semiconductor lattice, the distribution of charge traps causing hysteresis, or any other factors, this relationship may be more complex. Thus, there may be a highly non-linear, more complex relationship between the back image and the corrupted front image. It has been found that while the analysis setup described by a linear fit can be one-way, a more general modeling that does not rely on specific assumptions can produce better results. And this has been discovered by using a neural network type model, which does not require a specific modeling setup but is generally capable of better learning this relationship between the back image and the front image in terms of hysteresis behavior.

[0083] Due to the spatially correlated nature of the images, it has been found that convolutional type neural networks ("CNNs") work particularly well. As will be explained in more detail below, some of these neural network models allow for multi-scale processing, which has been found to produce better results and is specifically contemplated in some embodiments in this document. Again, the details will be described below.

[0084] Some general aspects of machine learning embodiments are now explained in more detail, which generally require two phases: a prior learning phase / training phase and a subsequent inference / deployment phase. The training of the ML model M is a mathematical process that can be done as a one-time operation or can be done in multiple phases when new training data appears. The training phase is based on training data, which can be artificially generated by a training data generator system TDGS, as actually envisioned in the embodiments, or can be obtained otherwise from an inventory of existing images, such as those that can be found in a medical database of a hospital, medical facility, etc.

[0085] Broadly, the training of a machine learning model requires adapting the parameters of model M according to a set of appropriately varying training data (x, y). In this context, x represents the training data input, which is processed by model M to produce an output M(x), and this output M(x) is then compared with the ground truth or label y (the prior associated with input x). Thus, x can be a historical or synthetic pre-image, and y is an estimate of what the artifact-free version of x should be. This corresponds to the supervised learning setting envisioned in some embodiments of this document; however, unsupervised learning settings based on clustering or other methods are not excluded in this document. Another characteristic of ML learning (not common in analytical modeling methods) is the backpropagation feature, where the mismatch between M(x) and y will be "backpropagated", or more generally, used to adapt the current model parameters, preferably over an iterative cycle. Due to the "more flexible" modeling, this allows for efficient learning and extraction of a good approximation of the sought-after relationships without the tendency to fix unknown relationships to a specific form of explicit analytical modeling that may not always be appropriate.

[0086] Lowercase letter symbols such as "x", "y" will be used in this document for training data, while data in post-deployment training will be referred to by uppercase letters such as "X", "Y".

[0087] Reference Figure 4 , shows a training system TS for training an ML model M based on training data TD = {(x, y)}. The training data is optionally provided by a training data generator system TDGS.

[0088] In a supervised setting, the training data can include pairs k of data ((x k , y k ). For each pair k (the index k is independent of the index used above to specify the generation of the feature map), the training data includes training input data x k and the associated target or ground truth y k . Thus, the training data is organized into pairs k, especially for the supervised learning scheme mainly envisioned in this document. However, it should be noted that unsupervised learning schemes are not excluded in this document.

[0089] During the training phase, the architecture of the machine learning model M, such as an artificial neural network (which will be discussed more fully below), is pre-populated with an initial set of weights or other parameters. The weights θ of the model M represent the parameterization of M Figure 6 , and the purpose of the training system TS is to optimize and thus adapt the parameters θ based on the training data (x θ , y k , y k ). In other words, learning can be mathematically formulated as an optimization problem, where a cost function F is minimized, although an alternative dual formulation maximizing a utility function can be used instead.

[0090] Now assume an example of the cost function F, which measures the aggregate residuals (s), i.e., the error generated between the data estimated by the neural network model M and the target for k based on some or all of the training data:

[0091] argmin θ F = ∑ k || M θ (x k ) - y k ||² (1)

[0092] In equation (1) and below, the function M() refers to the result of applying the model M to the input x. The cost function can be pixel / voxel-based, such as the L1 or L2 cost function. During training, the training input data x of the training pairs k propagates through the initialized network M. Specifically, for the training input x of the k-th pair k is received at the input IL, passed through the model and then output at the output OL as the output training data M θ (x). A suitable similarity measure ||·||, such as the derived p-norm, squared difference or others, is used to measure the difference between the actual training output M θ (x k ) produced by the model M and the desired target y k , which is also called the residual in this article.

[0093] The output training data M(x k ) is an estimate of the target y k associated with the applied input training image data x k . Typically, there is an error between this output M(x k ) and the target y k associated with the k-th pair currently under consideration. Then an optimization scheme such as backpropagation or other gradient-based methods can be used to adapt the parameters θ of the model M in order to reduce the error for the considered pair (x k, y k ) or the residuals of a subset of the training pairs from the complete training data set.

[0094] In which, after one or more iterations in the first inner loop where the updater UP updates the model parameters θ for the current batch of pairs {(x k , y k )}, the training system TS enters the second outer loop, where the next batch of training data pairs {x k+1 , y k+1} is processed accordingly. Although it is possible to loop over the outer loop on individual pairs until the training data set is exhausted, it is preferred herein to loop over a set of pairs at a time (i.e., on the said batch of training pairs). The iterations in the inner loop are on the parameters for all pairs that make up the batch. This batch learning has been shown to be more efficient than learning by pairs. Hereinafter, the subscript notation "k" for indicating individual instances of training data pairs will be largely discarded to eliminate notation.

[0095] The structure of the updater UP depends on the optimization scheme used. For example, the inner loop managed by the updater UP can be implemented by one or more forward and backward passes in the backpropagation algorithm. When adapting the parameters, the aggregated (e.g., summed) residuals of all training pairs in a given batch are taken into account for the current pair to improve the objective function. The aggregated residuals can be formed by configuring the objective function F as the sum of squared residuals, for example, in Equation (1) for some or all of the considered residuals for each pair. Other algebraic combinations rather than the sum of squares are also contemplated.

[0096] A GPU can be used to implement the training system TS to improve efficiency.

[0097] Now refer to Figure 5 , which shows the learning phase LP and the subsequent inference / deployment phase IP in a schematic block diagram. The deployment phase IP can involve the phase where the trained model is used by the corrector component CC in clinical practice, where the pattern is to process new inputs X that have not been seen before, i.e., input data X that has not been extracted from the training data set TD (e.g., a new artifact-corrupted pre-image). There can also be a test phase, where some of the training data is set aside for later use in the test phase to test the performance of the system. Generally, the test phase and the deployment / inference are very similar.

[0098] As mentioned, the training data (x, y) is either obtained from a stockpile of raw images or artificially generated, for example, by imaging a phantom PH. The labels can be generated by an experienced human expert using graphics touch software, but this can be cumbersome and expensive. Other ways of specifically providing the associated label y will be described in more detail below. The training data (x, y) obtained in any way is fed into an initial model M0 = M with a default set of parameters, such as a convolutional neural network. For reasons that will soon become apparent, given the nature of the special type of operation (convolution) used by such a model, these CNN parameters are commonly referred to as filter parameters. Thus, the initial model M0 is initialized with a default or even arbitrary set of initial parameters, which is true not only for CNN / NN models but also for other model types. The output of the initial model M0(x) is compared with the label y, and preferably the parameters are iteratively adapted in one or more epochs during the learning phase LP. When adapting the parameters for a given batch of training data being processed, preferably error backpropagation is used in one form or another. The two curved arrows schematically indicate the most common iterative training process, in which the model parameters are adapted according to the training data instances (x, y) in the iterative epochs. Once fully trained (which can be established based on various criteria, such as convergence or a preset number of iterative epochs, etc.), the now trained model M is then deployed in the inference phase or testing, as described. In the deployment, given a new pre-image X with hysteresis damage, the desired hysteresis / artifact-corrected version X’ of the pre-image X is computed by the model M, where the artifacts otherwise caused by the described hysteresis mechanism are corrected (at least reduced or completely eliminated). Thus, the version M(X) = X’ is an approximation of the pattern of intensity values that would be obtained without the damage caused by hysteresis.

[0099] In some embodiments, generative models such as the GAN model M can also be beneficially used herein, where the model M is trained to “redraw” X as X’ in a hysteresis-free parlance learned from a stockpile of training data (X, Y). Such generative models are sometimes used to artificially generate motivational images for artists that were not initially drawn as such. But in their “style”, it has been found that such GAN setups can be beneficially used herein to learn to draw pre-images in a hysteresis-free manner. In such a GAN-type setup, the cost function F driving the training, eq (1), represents the adversarial interaction between the generator network and the discriminator network that together make up the model M in the GAN. See “Generative Adversarial Networks” by I Goodfellow et al., available on arXiv with reference code arXiv:1406.2661 (2014).

[0100] InFigure 5 In the figures and subsequent illustrations, the subscript "f" indicates "front image", and the subscript "b" represents "back image". In this text, an asterisk symbol "*" is used in the superscript to indicate whether the back image was acquired with the hardware lag reducer circuit activated: for example, whether the back reset optical circuit was activated prior to the acquisition of a given back image frame, or whether any other such type of lag reduction circuit LRC was activated. Thus, the presence of this asterisk symbol "*" in this text will indicate that the back image was indeed recorded through the prior activation of the circuit LRC, while the absence of the asterisk indicates that no such LRC was activated, and thus no hardware-based temporal artifact correction was used. Thus, the asterisk relates to the use of hardware-induced correction associated with back image acquisition. Preferably, regardless of the type of lag reduction circuit LRC, it can be deactivated in this text to allow the acquisition of uncorrected back images, which can be used to facilitate the acquisition of training data TD, as will be explained in more detail below.

[0101] It has been found that in some embodiments, the above-described generally linear relationship between the scaling of the front image damaged by lag and the LRC-corrected back image can be used in this text to implement data-based regularization in a training algorithm through the regulator channel RL. This data-based regularization allows for better and more robust learning and better deployment or test performance. In this regularization-based embodiment, not only is the front image X to be corrected f fed into the training model M for processing, but also the backlight-corrected image X b * is fed into the model M through a second regularization channel RL. Thus, in the proposed training embodiment, when the model M processes the front image X f as input, the model M uses a dual-channel setup, and the hardware-corrected back image X b *(due to the LRC circuit) is jointly processed by the model to derive a corrected version X' of the front image X f . This dual-channel processing allows providing context for the model M to better learn the relationship between the artifacts that damage the front image and the corrected back image, in order to better learn the way lag damage affects the image.

[0102] Continuing to refer Figure 5 , and as shown, multiple training configurations based on different combinations of training data X are envisioned in this text for training the system TS (for more details, see Figure 6 below) to train the model M.

[0103] For example, in one embodiment, the pre-image xf of a single-layer detector without lag correction is used, optionally together with the post-image xb* of a dual-layer detector system XD, as the training input. The optional use of the post-image xb* implements the regularization channel RL described above. Thus, the training input data in this embodiment of the training data configuration is x = xf, or optionally with regularization, x = (xf, xb*).

[0104] Without abusing language logic, for the sake of unified terminology, the projection image recordable by a conventional energy-integrating imager with a single-layer detector system (not shown) may also be equally referred to herein as the "pre-image".

[0105] The single-layer pre-image xf* with LRC circuit lag correction can be used as the ground truth y in a supervised training setting, y = xf*.

[0106] Such a single pre-image xf* can be obtained by physically removing the front layer FLD in the dual imager IA and using only the remaining back layer BLD with the lag corrector circuit LRC, effectively transforming the dual-layer imager IA into a single-layer detector system temporarily. Thus, the previous back layer module BLD is now the front layer module. Alternatively, the ground truth pre-image xf* with lag correction can be acquired in a different natural single-layer imager with the same type and manufacture of detector modules as those used in the dual imager IA. For example, this is usually the case for a given manufacturer, as previously observed.

[0107] The above training data set (x = x f , y = x f *) or (x = (x f , x b *), y = x f *) can be obtained in the training data generator setting TDGS as follows: The input training image x can be obtained by the normal operation of the dual imager with its multi-layer detector XD, for example, by imaging the changes of a properly configured phantom PH in a sufficient number of instances. The ground truth part y = x f * of the training data TD can be obtained by using the described dual-energy imager IA, where the front layer FLD is removed by acquiring the same number of images of the phantom PH, preferably in the same imaging geometry as that for the training input x and the ground truth y.

[0108] Alternatively, the single-layer data without lag correction x f can be from a natural single-layer detector imager with the lag correction LRC deactivated, while the x f * image is acquired with the lag correction LRC activated before exposure, as described above.

[0109] Preferably, training images are acquired in sequence, for example imaging for training the input x is followed by imaging for the associated ground truth y or vice versa, to maintain the correct association of pairs.

[0110] Turning now more specifically to training data generation, using the protocol and training data configuration (x,y) described above, generally, the single-layer pre-images and double-layer data (pre-image and post-image) for training should preferably be acquired in the same imaging geometry and preferably using the same acquisition-related imaging settings (rotation speed, etc.).

[0111] Accordingly, the training data generator system TDGS includes operating a multi-energy imager IA or a single-layer imager in the protocol described above. The training data generator system TDGS also includes acquiring a set of different training images by changing the configuration (such as the spatial configuration) of the phantom PH imaged in the protocol over time or by changing the imaging context of the static phantom (such as the IA settings of the imager). In particular, various time-related instances of phantom images are contemplated herein to be controlled and managed by the training system TDGS. For example, the phantom PH can be arranged as a rotating disk, where the normal vector of the disk is parallel to the detector normal (2D imaging). Or, the phantom PH can be arranged on a rotating table for a rotational imaging geometry, such as for cone beam CT (CBCT) or any other tomographic imaging geometry, divergent or parallel.

[0112] Alternatively, it is possible to acquire an image sequence with strongly varying doses but using a static phantom PH. The double-layer data should be acquired in the same positioning as the single-layer detector data.

[0113] In any of the above cases, the phantom PH can include illustrations of different materials (such as water, fat, calcium, etc.) commonly found in the human body.

[0114] In the pre-image x f and the post-image x b * are passed together as inputs into the model for data-based regularization in the above-described dual-channel learning to obtain a corrected version x', which can also be done during deployment. Alternatively, the dual-channel method is only used for learning and not during deployment.

[0115] The above learning settings mainly focus on supervised learning. However, unsupervised settings for training ML models are also contemplated herein.

[0116] For example, an unsupervised framework for network learning strategies can be implemented by different localizations of the phantom PH, resulting in different realizations of the hysteresis artifacts that are only from the bilayer data (xf, xb*). Based on multiple independent such artifact realizations of the same underlying scanned object PH, the unsupervised learning scheme is able to recover the underlying ground truth image y = xf* and reduce / remove the hysteresis artifacts of the front layer detector FLD. Thus, in this way, several images xf with hysteresis artifacts from the phantom acquired by the front layer detector FLD and the corresponding images xb* of the back layer detector BLD with the hysteresis corrector circuit LRC for applying hysteresis correction are recorded at different positions. Then, the training system TS can continue to distinguish the artifacts from the expected image content, which obviates the need for prior knowledge of xf*, and thus an unsupervised learning scheme is implemented herein. For example, the training system TS can compare the images via registration to align the images, so as to distinguish the artifacts from the actual image content (represented by the phantom here), because several versions / realizations of the hysteresis artifacts have been recorded. If there are no other means to generate available ground truth data, then the substantially artifact-free images thus arrived at can then be used for training. Therefore, the proposed phantom imaging will provide a self-help way to generate ground truth data, which can then be used in any training method to adapt the parameters of the model M.

[0117] Now refer to Figure 6 , which shows the components of a convolutional neural network CNN type model as envisioned herein in an embodiment, including convolutional filters CV to better account for the mentioned spatial correlations in the pixel data (x, y).

[0118] Specifically, Figure 6 shows a convolutional neural network M in a feed-forward architecture. The network M includes a plurality of computational nodes arranged in layers in a cascaded manner, where the data stream proceeds from left to right and thus layer by layer. Recurrent networks are not excluded herein.

[0119] In deployment or training, input data including the pre-image x f .X f (and optionally together with the post-image xb*) is applied to the input layer IL. The input data is fed into the model M at the input layer IL and then propagates through a series of hidden layers L1 - L3 (only three are shown, but there can be one or two, or more than three), and then appears at the output layer OL as the training data input M(x) or the corrected pre-image X'. It can be said that the network M has a deep architecture because it has more than one hidden layer. In a feed-forward network, "depth" is the number of hidden layers between the input layer IL and the output layer OL, while in a recurrent network, depth is the number of hidden layers multiplied by the number of passes.

[0120] For computational and memory allocation efficiency, the layers of the network, and indeed the input and output images, as well as the inputs and outputs between the hidden layers (referred to herein as feature maps) can be represented as two-dimensional or higher-dimensional matrices (“tensors”).

[0121] Preferably, the hidden layer includes a sequence of convolutional layers, represented herein as layers L1-L N-k,k>1 . The number of convolutional layers is at least one, such as 2, 3, 4, or 5, or any other number. This number can go into double digits figure.

[0122] In embodiments, one or more fully connected layers may exist downstream of the sequence of convolutional layers, but this is not necessarily the case in all embodiments, and in fact preferably in some embodiments no fully connected layers are used in the architectures contemplated herein. Each hidden L m layer and the input layer IL implement one or more convolutional operators CV. Each layer L m may implement the same number of convolutional operators CV, or the number may be different for some or all layers.

[0123] The convolutional operator CV implements the convolution operation to be performed on its corresponding input. The convolutional operator can be conceptualized as a convolution kernel. It can be implemented as a matrix that includes entries forming filter units referred to herein as weights θ. In particular, these weights are adjusted during the learning phase. The first layer IL processes the input data through one or more of its convolutional operators. The feature map FM is the output of the convolutional layer, one feature map for each convolutional operator in the layer. The feature maps of the earlier layer are then input into the next layer to produce higher-generation feature maps FM i 、FM i+1 , and so on, until the last layer OL combines all the feature maps into the output M(x), X’. The entries of each feature map can be described as nodes. The final combiner layer OL can also be implemented as a convolution that provides the corrected pre-image.

[0124] The convolutional operator CV in the convolutional layer differs from the fully connected layer in that the entries in the output feature map of the convolutional layer are not a combination of all the nodes received as inputs to that layer. In other words, the convolution kernel is applied only to a subset of the input volume V, or to the feature maps received from earlier convolutional layers. The subset is different for each entry in the output feature map. Thus, the operation of the convolutional operator can be conceptualized as a “sliding” over the input, similar to the discrete filter kernel in the classical convolution operation known from classical signal processing. Hence, the name “convolutional layer”. In a fully connected layer, the output nodes are typically obtained by processing all the nodes of the input layer.

[0125] The stride of the convolutional operator can be selected as 1 or greater than 1. The stride defines how the subsets are selected. A stride greater than 1 reduces the size of the feature map relative to the size of the input in that layer. In this article, a stride of 1 is preferred. To maintain the size of the feature map corresponding to the size of the input image, a zero-padding layer P is applied. This even allows convolution of the feature map entries located at the edges of the processed feature map.

[0126] The number of pixels of the input data processed in this way for each output node in a given layer describes the size of the receptive field of the neural network of the given cell at that layer and is determined by the kernel size, stride, and padding of all convolutional operators CV before that layer where the node is located. More specifically, the receptive field of the entire convolutional neural network M is the size of the action region of the input data that is "seen", i.e., processed to determine a given node of the output data. The action region is the number of nodes across the layers of the network M that contribute to a given node in the output data X', M(x).

[0127] In a preferred embodiment (see the schematic illustration Figure 6A ), a convolutional neural network CNN is used for a model M configured for multi-scale processing. Such a model includes, for example, a U-cell architecture or its ilk, as described by O. Ronneberger et al. in "U-Net: Convolutional Networks for Biomedical Image Segmentation" (available online at the arXiv repository with citation code arXiv:1505.04597 (2015)).

[0128] In an NN model supporting multi-scale of this or a similar architecture, a deconvolutional operator DV is used. More specifically, such a multi-scale enabling architecture of the model M can serially include a downsampling path DPH of layers and an upsampling path UPH of layers downstream thereof. Convolutional operators with varying stride lengths gradually reduce the feature map size / dimension (m×n) as passing through layers L j in the downsampling path, down to the lowest size representation of the feature map, the latent representation κ. The dimension of the latent representation feature map κ gradually increases as passing through layers L k in the upsampling path UPH, returning to the size of the original input data X. This size bottleneck structure has been found to lead to better learning because the system is forced to extract its learning down to the simple structures represented by the low-size latent representation.

[0129] Additional regularization channels can be defined, where the intermediate outputs (intermediate feature maps) from the scale levels in the downsampling path are fed into the corresponding layers at the corresponding scale levels in the upsampling path UPH. It has been found that such inter-scale cross-connections ISX promote the robustness and efficiency of learning and performance.

[0130] Preferably, in any NN-type model M, the receptive field of the network is chosen to cover the feature artifact size.

[0131] Other regression-type models can be used instead of or in combination with the NN-type model, and thus the present disclosure is not limited to NN models.

[0132] Now refer to the flowchart in FIG. 7, which relates to the above-described processes of deployment, ML training, and (if needed) training data TD generation.

[0133] First turn to Figure 7A the flowchart in, which represents the steps of a hysteresis correction method for processing projection images acquired by the spectral imager IA. The data includes two sets of projection data, referred to as the pre-image and the post-image / image, as described above. The input pre-image X is the result of projection data acquisition using a multi-energy detector system, such as can be used in an X-ray-based imager IA of the radiography or multi-directional tomography type. Thus, the multi-energy projection images (pre-image and post-image) record the same FOV with the same imaging geometry but with different energy ranges or spectra.

[0134] In step S710, such a potentially hysteresis-corrupted pre-image X acquired by the front layer FLD of the multi-energy detection system DX in the spectral imager IA is received.

[0135] The pre-image X thus received is processed into a corrected image X' in step S720.

[0136] Then, in step S730, the corrected pre-image X' is made available for storage, display (if needed), or for further processing, for educational purposes, for statistical analysis, etc.

[0137] For example, as an optional step S740, and as mainly contemplated herein, such a corrected pre-image X' is then passed together with the naturally LRC-circuit-corrected, co-acquired post-image (acquired by the back layer BLD) for processing by a spectral image processing algorithm to obtain spectral-processed projection data.

[0138] As described above in connection with Figure 1 In such spectral projection images (such as "virtual-only contrast" images), the contrast is then only material-specific and corresponds to, for example, a pre-selected material type chosen by the user. At this stage, the spectral projection images are no longer multi-energy, but now are a single set, having a single value for each pixel position for any given projection direction (regardless of the effects from a cone-beam geometry or other divergent imaging geometries).

[0139] In step S750, the spectral projection image obtained in step S740 can then be reconstructed into a tomographic image in the image domain, such as reconstructed into a 3D image volume or into one or more of its parts (slices).

[0140] Machine learning can be used to implement the correction operation S720. In such a method, a trained ML model M previously trained on training data can be used.

[0141] If machine learning is used, the steps of the Figure 7A method can be used for deployment (such as in clinical daily practice) or testing prior to deployment.

[0142] Once the model is sufficiently trained (as judged by certain criteria as mentioned above), the model can be used in the correction step S720.

[0143] The correction operation in step S720 is actually preferably based on the trained machine learning model as already described.

[0144] In a simpler non - necessarily ML method or a simpler ML method with more explicit modeling, a scaling factor can be determined by fitting a linear model to a series of acquired pre - images and their co - acquired post - image counterparts. A single such scaling factor can be used, or a matrix (map) of such scaling factors, where the scaling factors are pixel - specific. The scaling factor can then be stored in the memory MEM’. Then step S720 can be implemented by retrieving the (one or more) scaling factors from the memory MEM’ and applying it / them to the anterior image X by pixel - by - pixel multiplication or division (or by other algebraic combinations) to obtain its lag - corrected version X’.

[0145] If a machine learning model M is used in step S720, such a model can be obtained by a training method as shown in the Figure 7B flowchart now referred to.

[0146] Broadly, this method proceeds to step S810, where training data TD is received. The training data is retrieved from an existing inventory of medical images such as can be found in a medical database, preferably including human - generated annotations, or generated synthetically as will be described in more detail below.

[0147] In step S820, the parameters of a predefined architecture of a machine learning model M (such as a convolutional neural network) are adapted based on the training data.

[0148] Training can be carried out in one or more iterations to produce a fully - trained model, which is then made available for deployment or testing in step S830, where it can be as Figure 7Afor use as shown. Training the S820 can be done once, or repeated whenever new training data becomes available.

[0149] The training can be supervised or unsupervised, such as by clustering a training data set, or by processing with an autoencoder network (AE) (including variational types of AE (VAE)). As previously mentioned, generative models such as I Goodfellow's GAN method or any of its ilk can be used.

[0150] Rather than (only) relying on existing historical images to train the data, training data can be synthetically generated in a training data generation setting, as Figure 7C shown in the flowchart of Figure 7C .

[0151] In such a method, at step S910, images in various combinations of pre-images and post-images are generated by imaging a phantom PH, for example, as described above in connection with Figure 5 in the case of activating and deactivating the hysteresis reducer circuit.

[0152] A dual imager and / or a single-layer imager can be used. The pre-image xf / xf* can be acquired by using a single-layer imager instead of the dual imager IA, and then imaging the phantom PH with the hysteresis reducer circuit LRC activated and deactivated. The imager settings (such as dose) can be changed for a static phantom, or the same settings can be used, but the phantom is moved during imaging. The single-layer detector imager preferably uses the same type and / or manufactured detector module as that used in the multi-energy detector system XD ( Figure 2 ). If this is not possible, care should be taken that at least the detector module specifications are the same or at least comparable. Preferably, at least, the same type (such as doping method, etc.) of semiconductor crystal should be used in both imagers.

[0153] As an alternative to using such a single-layer imager as an aid, the dual imager IA itself can be operated to acquire x b *, and x f .x f* can be obtained by removing the front layer FLD of the dual imager and using only the back layer BLD with the LRC circuit. In this regard, a dual imager IA equipped with a removable front layer detector module FLD may be beneficial. For example, the front layer FLD can be arranged to be slid in / slid out of the unit, etc., which can be slid out and returned to the housing H as needed to make the training data generation trouble-free. Alternatively, a laboratory setup can be used, which is configured to simulate the properties of a real double-layer detector. Similarly, during such imaging, the phantom PH is moved, or the settings of the imager are changed to generate different realizations of the training data, especially training data pairs. Other training data generation options are not excluded herein.

[0154] In step S920, once a sufficient supply of appropriately varied training data images of the front and back images has been obtained, they can then be used as training data for a machine learning model, such as as Figure 7B shown.

[0155] It has been observed that if, at the input during inference / testing 7A and / or training 7B, not only a corrupted front image is provided for the machine learning model to process, but also in combination with an associated naturally corrected back image, then Figure 7A and 7B the methods described above in

[0156] work particularly well. Then both can be jointly processed during inference or training to provide better context to the method, thereby improving inference and / or training. Using such a dual-channel method during training rather than during inference may be sufficient. However, preferably, it is used for both training and inference.

[0157] Alternatively, some or all components of the corrector system CS can be arranged in hardware integrated into the imaging system IAR, such as a suitably programmed microcontroller or microprocessor, such as an FPGA (Field Programmable Gate Array), or a hardwired IC chip, an Application Specific Integrated Circuit (ASIC). In yet another embodiment, the system SYS can be implemented partly in software and partly in hardware both.

[0158] Different components of the corrector system CS for the entire computation SYS can be implemented on a single data processing unit PU. Alternatively, some or more components are implemented on different processing units PU, which may be remotely arranged in a distributed architecture and connectable in a suitable communication network, such as in a cloud setup or a client-server setup, etc.

[0159] One or more features described herein can be configured as or implemented as circuitry and / or combinations thereof encoded within a computer-readable medium. The circuitry can include discrete and / or integrated circuits, system-on-a-chip (SOC), and combinations thereof, machines, computer systems, processors and memories, computer programs.

[0160] In another exemplary embodiment of the present invention, there is provided a computer program or computer program unit, characterized in that it is adapted to execute the method steps of a method according to one of the foregoing embodiments on a suitable system.

[0161] Thus, the computer program unit can be stored on a computer unit, which can also be part of an embodiment of the present invention. The computing unit can be adapted to execute the steps of the above method or cause the execution of the steps of the above method. In addition, it can be adapted to operate the components of the above device. The computing unit can be adapted to automatically operate and / or execute the commands of the user. The computer program can be loaded into the working memory of the data processor. Thus, the data processor can be equipped to execute the method of the present invention.

[0162] This exemplary embodiment of the present invention covers computer programs that use the present invention from the beginning and computer programs that turn existing programs into programs using the present invention with the help of the latest programs.

[0163] In addition, the computer program unit may be able to provide all the necessary steps to implement the process of the exemplary embodiment of the method as described above.

[0164] According to another exemplary embodiment of the present invention, there is provided a computer-readable medium such as a CD-ROM, wherein the computer-readable medium has stored thereon a computer program unit as described in the previous section.

[0165] The computer program can be stored and / or distributed on a suitable medium (in particular but not necessarily a non-transitory medium), such as an optical storage medium or a solid-state medium provided together with or as part of other hardware, but can also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems.

[0166] However, the computer program can also be presented via a network such as the World Wide Web and can be downloaded from such a network into the working memory of the data processor. According to another exemplary embodiment of the present invention, there is provided a medium for making a computer program unit available for download, the computer program unit being arranged to execute a method according to one of the foregoing embodiments of the present invention.

[0167] It should be noted that the embodiments of the present invention are described with reference to different subjects. In particular, some embodiments are described with reference to method-type claims, while other embodiments are described with reference to device-type claims. However, those skilled in the art will learn from the above and the following descriptions that, unless otherwise stated, any combination between features related to different subjects is also considered to be disclosed in the present application, in addition to any combination of features belonging to the subject matter of one type. However, all features can be combined to provide a synergistic effect that is more than the simple sum of the features.

[0168] Although the present invention has been illustrated and described in detail in the accompanying drawings and the foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive. The present invention is not limited to the disclosed embodiments. By studying the drawings, the disclosure, and the dependent claims, those skilled in the art can understand and realize other variations of the disclosed embodiments when practicing the claimed invention.

[0169] In the claims, the word "comprising" does not exclude other elements or steps, and the words "a" or "an" do not exclude a plurality. A single processor or other unit can implement the functions of several items recited in the claims. The fact that certain measures are recited only in mutually different dependent claims does not indicate that a combination of these measures cannot be used advantageously. Any reference signs in the claims should not be construed as limiting the scope. Such reference signs may include numbers, letters, or any alphanumeric combination.

Claims

1. A system (SYS) for acquisition lag correction in X-ray imaging, comprising: An input interface (IN) for receiving an input, the input including an input image acquired by a front-layer detector (FLD) in a multi-energy X-ray imaging device (IA) and an additional image acquired by a back-layer detector (BLD) of the multi-energy X-ray imaging device (IA), the back-layer detector including a circuit (LRC) for reducing a time lag effect, the additional image having been acquired by the back-layer detector when the circuit was activated; And A corrector component (CC) configured to process the input image and the additional input image to generate a corrected input image, the corrected input image representing an estimate of the image information of the front-layer detector at an acquisition lag lower than the acquisition lag experienced by the input image.

2. The system according to claim 1, wherein The processing by the corrector component (CC) includes scaling the input image using a scaling factor obtained based on a series of previous input images and additional input images previously acquired by the multi-energy X-ray imaging device (IA) or a multi-energy X-ray imaging device.

3. The system according to claim 1 or 2, wherein The corrector component (CC) is implemented based on a trained machine learning model (M).

4. The system according to claim 3, wherein, The ML model is trained in a supervised setting based on training data including training input data and associated ground truth data, wherein the training input data includes previous images with a time lag effect acquired by the front-layer detector or a front-layer detector, and wherein the ground truth data includes previous images acquired by the front-layer detector or a front-layer detector including a circuit (LRC) for reducing a time lag effect when such a circuit was activated.

5. The system according to claim 4, wherein, The training input data further includes additional previous images acquired by the back-layer detector or a back-layer detector, the additional previous images having been acquired by the back-layer detector when a circuit for reducing a time lag effect was activated.

6. The system according to claim 3, wherein, The ML model is trained in an unsupervised setting based on training data including previous images with a time lag effect acquired by the front-layer detector or a front-layer detector, and previous images acquired by the back-layer detector or a back-layer detector including a circuit (LRC) for reducing a time lag effect when such a circuit was activated.

7. The system according to any of the preceding claims, further comprising an output port (OUT) for transmitting the corrected input image for further processing by a spectral processing algorithm and / or a tomographic image reconstruction algorithm.

8. An imaging arrangement (IAR) comprising the system according to any one of the preceding claims, and the multi-energy X-ray imaging device including the front-layer detector and the back-layer detector.

9. The imaging arrangement according to claim 8, wherein, The front-layer detector and the back-layer detector are provided as modules of a multi-layer detector system.

10. A computer-implemented method for acquisition lag correction in X-ray imaging, comprising: Receive (S710) an input that includes an input image acquired by a front layer detector (FLD) in a multi-energy X-ray imaging apparatus (IA) and an additional image acquired by a back layer detector (BLD) of the multi-energy X-ray imaging apparatus (IA), the back layer detector including a circuit (LRC) for reducing a time lag effect, the additional image having been acquired by the back layer detector while the circuit was activated; and Process (S720) the input image and the additional input image for generating a corrected input image that represents an estimate of the image information of the front layer detector at an acquisition lag lower than the acquisition lag experienced by the input image.

11. A method of training the machine learning model in the system according to claim 4 or 5, comprising generating training data by imaging a phantom from a plurality of relative directions in the same imaging geometry with and without activating a lag reduction circuit (LRC) for the front layer detector.

12. A method of training the machine learning model in the system according to claim 6, comprising generating training data by imaging the phantom with different localizations of the phantom using the front layer detector and the back layer detector.

13. A computer program unit which, when executed by at least one processing unit, is adapted to cause the processing unit to perform the method according to any one of claims 10 - 12.

14. At least one computer-readable medium having stored thereon the program unit according to claim 13, or having stored thereon the trained machine learning model according to claims 3 - 6.

Images