Supervised artifact reduction in x-ray images

Abstract

The presently disclosed subject matter addresses the intrinsic limitation that artifact- free ground-truth X-ray images cannot be physically acquired, restricting the use of direct supervised learning for artifact correction. To overcome this constraint, the disclosure provides multiple mechanisms for developing machine-learning models dedicated to attenuating overlapping-layer artifacts in radiographs.

Description

[0001] SUPERVISED ARTIFACT REDUCTION IN X-RAY IMAGES

[0002] FIELD OF THE PRESENTLY DISCLOSED SUBJECT MATTER

[0003] The presently disclosed subject matter relates to computer implemented methods and systems for X-ray image processing and particularity to the reduction of artifacts in such images.

[0004] BACKGROUND

[0005] Two-dimensional (2D) X-ray imaging of the body results from projecting a three- dimensional (3D) anatomical region-of-interest (ROI) onto a 2D detector or, in the case of panoramic radiography, a moving one-dimensional detector array. Two commonly used types of 2D X-ray images (also referred to as ’radiographs') are chest and dental panoramic X-rays. A chest X-ray is often among the first images acquired at an emergency facility as it may be useful in patients' triage. Panoramic radiography is a frequently prescribed dental imaging technique. The ability to provide a "full-mouth image" at relatively low cost and with a low radiation dose makes it a valuable tool for dentists. A modern panoramic radiography machine comprises an X-ray source and detector that rotate simultaneously around a patient's head, generating a full panoramic view of the patient's maxillofacial region. Modern panoramic radiography machines may dynamically move the center of rotation for the source-receiver line to optimize image quality and sharpness of the dental structures.

[0006] OVERVIEW

[0007] Projecting the 3D maxillofacial region onto a 2D X-ray image may lead to artifacts caused by the superposition of partially radio-opaque 3D structures projected onto the same 2D image area. The resulting image may be difficult to interpret and may lead to incorrect diagnoses. A similar problem may be encountered in the case of chest X-rays where the rib cage overlaps with the lungs and mediastinum, potentially limiting the diagnostic value of the image in the corresponding areas. Three common structures which are artifact sources in panoramic X-rays include the contralateral mandible (often appearing as a "ghost jaw"), the cervical spine, and pharyngeal air spaces (also referred to as "air gaps"). Fig. 1 shows a panoramic radiograph displaying the full maxillofacial region at the top, with three corresponding images below that highlight different artifact types, each caused by a different artifact-causing structures, through colored overlays, from left to right: the ghost jaw, the spine, and the air gap. The first artifact, the "ghost jaw," is the ghost image of the side of the jaw closer to the X-ray source. It may appear as a radiopaque, brighter region on either side of the image, often with a bright line that cuts horizontally through the center of the jaw that is meant to be imaged. The ghost jaw artifact is highlighted by the green overlays on the bottom left of Fig. 1.

[0008] The second artifact, the spine, appears as a bright, radiopaque region in the center of most panoramic radiographs. In some cases, the brightness of the spine can be overwhelming, hiding other layers or details in the same region, near the chin. The spine is highlighted by the red overlay that cuts vertically through the bottom center image in Fig. 1.

[0009] The third artifact, the "air gap," appears as a dark, radiolucent region that arcs from either bottom corner of the image, diagonally up through the jaw, and finally converges in the vicinity of the nose. This artifact is caused by the air in the patient's mouth and pharynx, which provides a region where X-rays can travel unimpeded. The air gap is highlighted by the blue overlay in the bottom right image in Fig. 1.

[0010] In order to train a supervised model for the correction of the artifacts, a dataset of image pairs is required, where the artifacts appear only in one image of the pair, which is the input to the model, while the artifact does not appear in the second image of the pair, which is the ground-truth. This requirement is not possible to fulfill, since the artifacts are created by body tissue or areas that cannot merely disappear during image acquisition, for example, the rib cage cannot disappear during a chest radiography acquisition, and neither can the pharyngeal air do so during a panoramic image acquisition.

[0011] The presently disclosed subject matter addresses the fundamental limitation that artifact-free ground-truth X-ray images cannot be physically acquired, preventing direct supervised training of machine learning models for artifact correction. To overcome this constraint, the disclosure provides multiple approaches that enable the development and application of such models. In a first approach, the need for explicit ground-truth data is avoided by repurposing a model, pre-trained for another image-processing task to perform artifact removal in X-ray images. In a second approach, synthetic ground-truth data are generated from three-dimensional imaging sources, such as computed tomography (CT) scans, thereby enabling supervised training or fine-tuning of artifact-correction models. In a third approach, outputs generated from various data sources, such as de-shadowed, de-hazed, sequentially enhanced, and CT-derived synthetic radiographs, are combined into a diverse training dataset. This dataset forms pairs of X-ray images— one with artifacts and the other artifact-attenuated— used to train a model dedicated to artifact removal. By incorporating these diverse sources, the model is equipped to effectively learn how to reduce or eliminate overlapping layer artifacts from X-ray images

[0012] As used herein, the term "attenuated" is used in a general sense to refer to the reduction or removal of contributions of a given structure or degradation in a radiograph. Accordingly, the term "artifact-attenuated," as used herein, refers to the attenuation of overlapping-layer artifacts, including both partial reduction and complete suppression of their effect in the resulting image.

[0013] SUMMARY

[0014] According to a first aspect of the presently disclosed subject matter there is provided a computer-implemented method of attenuating overlapping-layer artifacts from a radiograph, the method comprising: receiving the radiograph comprising overlapping-layer artifacts resulting from superposition of anatomical (e.g., spine) and / or non-anatomical structures (e.g., an air gap); segmenting, using a first machine-learning model, the radiograph to generate an artifactsegmentation mask identifying respective artifact types; applying the artifact-segmentation mask together with the radiograph to a pre-trained machine-learning model for optical deshadowing to attenuate shadow-like regions within a mask, the model thereby generating corrected pixel data for the artifact-segmented regions; and outputting an artifact-attenuated radiograph generated by the model using the corrected pixel data.

[0015] In addition to the above features, the method according to this aspect of the presently disclosed subject matter can optionally comprise one or more of features (i) to (xii) below, in any desired and technically possible combination or permutation: i. Wherein overlapping-layer artifacts include radiopaque artifacts characterized by pixel values higher than surrounding tissue; wherein the method further comprises generating a negative version of the radiograph prior to applying the de-shadowing model to convert the radiopaque regions into shadow-like regions compatible with the de-shadowing process. ii. Wherein the method further comprises: applying a second machine-learning model trained to segment an anatomical structure within the radiograph to generate an anatomical-segmentation mask; comparing the artifact-segmentation mask with the anatomical-segmentation mask to identify regions in which artifact regions spatially overlap the anatomical structure; masking only the overlapping regions identified by the comparison to generate an overlap mask; and providing the overlap mask together with the radiograph to the de-shadowing machine-learning model to enable attenuation of the artifact solely within the regions that overlap the anatomical structure, thereby improving visibility of the anatomical structure. iii. Wherein the method further comprises after applying the de-shadowing deeplearning model, applying a de-hazing deep-learning model to the output of the de-shadowing model to attenuate diffuse, haze-like degradations arising from scattered radiation or overlapping soft-tissue regions, thereby further improving contrast uniformity and visibility of anatomical structures. iv. Wherein the radiograph comprises a plurality of overlapping-layer artifact types, and segmenting the radiograph comprises applying, for each artifact type, a respective machine-learning model trained to segment that artifact type, thereby producing a corresponding set of artifact-specific segmentation masks. v. Wherein the radiograph is a panoramic radiograph, and the plurality of overlapping-layer artifact types comprise at least two of: air-gap artifacts, ghost-jaw artifacts, and spine artifacts, each segmented using its respective machine-learning model. vi. Wherein applying the pre-trained de-shadowing machine-learning model comprises performing a plurality of sequential passes, each pass receiving a different artifactspecific segmentation mask and attenuating the corresponding artifact type, such that the radiograph is progressively corrected across successive passes. vii. Wherein the second machine-learning model is an autoencoder-based machinelearning model (e.g., U-Net) or GAN. viii. Wherein the pre-trained de-shadowing machine-learning model is any one of: a generative adversarial network (GAN); a convolutional neural network (CNN); a diffusion-based neural network, or a transformer- based network. ix. Wherein the radiograph is any one of a panoramic radiograph (e.g., dental X- ray) and standard projection radiographs (e.g., chest X-ray). x. Wherein the artifacts include one or more artifact types including: air gaps, superimposed skeletal structures, and other semi-transparent anatomical layers observed in radiographic imaging. xi. Wherein the radiograph is a dental panoramic radiograph, and the overlappinglayer artifact is caused by the mandible. xii. wherein the artifacts are overlapping layer artifacts that result from the superposition of anatomical (e.g., spine) and / or non-anatomical structures (e.g., an air gap), commonly exhibited as semi-transparent regions in an X-ray image

[0016] According to a second aspect of the presently disclosed subject matter there is provided a computer system comprising at least one processing circuitry configured to execute a method according to the first aspect.

[0017] In some examples the computer system is integrated with or is otherwise operatively connected to an X-ray device to enable the X-ray device to provide artifact-attenuated images.

[0018] According to a third aspect of the presently disclosed subject matter, there is provided a non-transitory program storage device readable by a computer, tangibly embodying a program of instructions executable by the computer to perform a method according to the first aspect disclosed above.

[0019] The system and the non-transitory program storage devices, disclosed with reference to the second and third aspects, can optionally comprise one or more of features (i) to (xii) listed above, mutatis mutandis, in any technically possible combination or permutation.

[0020] The presently disclosed subject matter further contemplates a method, system and non-transitory program storage device as in the first, second and third aspects above, where a de-hazing machine learning model is used instead of the de-shadowing machine learning model.

[0021] According to a fourth aspect of the presently disclosed subject matter there is provided a computer-implemented method of attenuating overlapping-layer artifacts in an input radiograph, the method comprising: receiving the input radiograph; providing the input radiograph to a machine-learning model trained to generate artifact-attenuated radiographs; and obtaining, from the machinelearning model, an artifact-attenuated version of the input radiograph; wherein the machine-learning model has been trained using pairs of digitally reconstructed radiographs (DRRs) generated from three-dimensional computerized- tomography (CT) volumes, each pair comprising an artifact-containing DRR and a corresponding artifact-attenuated DRR obtained by performing radiographic projection of the CT volume under a first projection configuration in which contributions of one or more artifact-causing structures in the DRR are preserved and under a second projection configuration in which contributions of the artifact-causing structures in the DRR are reduced or removed.

[0022] In addition to the above features, the method according to this aspect of the presently disclosed subject matter can optionally comprise one or more of features (i) to (vi) below, in any desired and technically possible combination or permutation: i. Wherein generating the artifact-attenuated DRR comprises segmenting one or more artifact-causing structures within the CT volume to obtain segmented regions and modifying the CT volume within the segmented regions so as to reduce or remove contributions of the artifact-causing structures in the radiographic projection. ii. Wherein modifying the CT volume in the segmented regions comprises inpainting the segmented regions using voxel values representative of adjacent tissue. iii. Wherein generating the artifact-attenuated DRR comprises performing radiographic projection of the CT volume using alternative virtual source locations, including: a first source location outside the CT volume such that corresponding sampling rays intersect one or more artifact-causing structures and produce an artifact-containing DRR; and a second source location inside the CT volume such that corresponding sampling rays bypass said artifact-causing structures and produce an artifact-attenuated DRR. iv. Wherein the machine-learning model trained using the DRR pairs comprises a neural-network architecture selected from an autoencoder, a convolutional neural network, or a transformer-based neural network. v. Wherein the machine-learning model trained using the paired DRRs comprises a de-shadowing model pre-trained to remove shadows in optical images and further trained on the paired DRRs to learn attenuation of artifact-causing structures in radiographic images. vi. Wherein the artifact-causing structures segmented in the CT volume comprise one or more of: bones, air cavities, or soft-tissue regions that produce overlapping-layer artifacts in radiographic projection. vii. Wherein the method further comprises training the machine-learning model using the training dataset generated from the paired digitally reconstructed radiographs. viii. Wherein the method further comprises generating the training dataset by forming, for each three-dimensional computerized-tomography (CT) volume, a pair of digitally reconstructed radiographs comprising an artifact-containing DRR and a corresponding artifact- attenuated DRR.

[0023] According to a fifth aspect of the presently disclosed subject matter there is provided a computer system comprising at least one processing circuitry configured to execute a method according to the fifth aspect.

[0024] In some examples the computer system is integrated with or is otherwise operatively connected to an X-ray device to enable the X-ray device to provide artifact-attenuated images.

[0025] According to a sixth aspect of the presently disclosed subject matter, there is provided a non-transitory program storage device readable by a computer, tangibly embodying a program of instructions executable by the computer to perform a method according to the fifth aspect disclosed above.

[0026] The system and the non-transitory program storage devices, disclosed with reference to the fifth and sixth aspects, can optionally comprise one or more of features (i) to (viii) listed above, mutatis mutandis, in any technically possible combination or permutation.

[0027] The presently disclosed subject matter also contemplates any one of a computer- implemented method, a computer system and a non-transitory program storage device each dedicated to generating a training dataset comprising pairs of digitally reconstructed radiographs (DRRs) generated from three-dimensional computerized-tomography (CT) volumes, each pair comprising an artifact-containing DRR and a corresponding artifact- attenuated DRR obtained by performing radiographic projection of the CT volume under a first projection configuration in which contributions of one or more artifact-causing structures in the DRR are preserved and under a second projection configuration in which contributions of the artifact-causing structures in the DRR are reduced or removed.

[0028] The presently disclosed subject matter also contemplates any one of a computer- implemented method, a computer system and a non-transitory program storage device each dedicated to training a machine learning model dedicated to attenuating artifacts in an input radiograph using the training dataset of pairs of digitally reconstructed radiographs (DRRs) generated from three-dimensional computerized-tomography (CT) volumes.

[0029] According to a seventh aspect of the presently disclosed subject matter there is provided a computer-implemented method of attenuating overlapping-layer artifacts in a radiographic input image, the method comprising, by processing circuitry: receiving the radiographic input image; applying a machine-learning model to the radiographic input image to generate an artifact-attenuated radiographic output; wherein the machine-learning model is a model trained using a training dataset comprising pairs of radiographs produced by a heterogeneous collection of two or more processing pipelines, each pair including a first radiograph containing overlapping-layer artifacts and a second radiograph in which the artifacts are attenuated by the respective processing pipeline, the processing pipelines differing in the processes used to generate the second radiograph of each pair.

[0030] In addition to the above features, the method according to this aspect of the presently disclosed subject matter can optionally comprise one or more of features (i) to (v) below, in any desired and technically possible combination or permutation: i. Wherein the heterogeneous collection of processing pipelines comprises at least two of: de-shadowing pipeline, a de-hazing pipeline, a sequential de-shadowing and de-hazing pipeline, and a CT-based digitally reconstructed radiograph (DRR) pipeline. ii. Wherein the heterogeneous collection of processing pipelines comprises three or four of: de-shadowing pipeline, a de-hazing pipeline, a sequential de-shadowing and de- hazing pipeline, and a CT-based digitally reconstructed radiograph (DRR) pipeline. iii. wherein generating the pairs of radiographs for the CT-based digitally reconstructed radiograph (DRR) processing pipeline comprises: a. generating, from a three-dimensional CT volume, an artifact-containing DRR by performing radiographic projection under a first projection configuration in which contributions of one or more artifact-causing structures are preserved; b. generating, from the same CT volume, a corresponding artifact-attenuated DRR by performing radiographic projection under a second projection configuration in which contributions of the artifact-causing structures are reduced or removed; and c. forming a paired training example by associating the artifact-containing DRR with the corresponding artifact-attenuated DRR, the artifact-containing DRR and the artifact- attenuated DRR forming the first and second radiographs of the pair, respectively. iv. Wherein the machine-learning model comprises a diffusion model configured to learn a mapping the first radiograph of each pair to the second radiograph of each pair provided by the processing pipelines. v. Wherein:

[0031] (a) the de-shadowing processing pipeline comprises applying a de-shadowing machinelearning model to a radiograph of a region of interest to obtain a corresponding radiograph in which overlapping-layer artifacts are attenuated;

[0032] (b) the de-hazing processing pipeline comprises applying a de-hazing machine-learning model to a radiograph of a region of interest to obtain a corresponding radiograph in which diffuse, haze-like artifacts are attenuated; and

[0033] (c) the sequential de-shadowing and de-hazing processing pipeline comprises first applying a de-shadowing machine-learning model to a radiograph of a region of interest and subsequently applying a de-hazing machine-learning model to obtain a corresponding radiograph in which both overlapping-layer artifacts and diffuse, haze-like artifacts are attenuated.

[0034] According to an eighth aspect of the presently disclosed subject matter there is provided a computer system comprising at least one processing circuitry configured to execute a method according to the seventh aspect. In some examples the computer system is integrated with or is otherwise operatively connected to an X-ray device to enable the X-ray device to provide artifact-attenuated images.

[0035] According to a ninth aspect of the presently disclosed subject matter, there is provided a non-transitory program storage device readable by a computer, tangibly embodying a program of instructions executable by the computer to perform a method according to the seventh aspect disclosed above.

[0036] The presently disclosed subject matter further contemplates any one of a computer- implemented method, a computer system, and a non-transitory program storage device each dedicated to training a machine-learning model configured to attenuate overlapping-layer artifacts in radiographic input images, the method comprising, by processing circuitry: generating a training dataset comprising pairs of radiographs produced by a heterogeneous collection of two or more processing pipelines, each pair including a first radiograph containing overlapping-layer artifacts and a second radiograph in which the artifacts are attenuated by the respective processing pipeline, the heterogeneous collection comprising two or more of: (i) a de-shadowing processing pipeline in which a de-shadowing machine-learning model is applied to the first radiograph to generate the second radiograph of each pair; (ii) a de-hazing processing pipeline in which a de-hazing machine-learning model is applied to the first radiograph to generate the second radiograph of each pair; (iii) a sequential de-shadowing and de-hazing processing pipeline in which the de-shadowing model and the de-hazing model are applied in sequence to the first radiograph to generate the second radiograph of each pair; and (iv) a CT-based synthetic radiograph processing pipeline in which a pair of digitally reconstructed radiographs (DRRs) is generated from a three-dimensional CT volume, including an artifact-containing DRR and a corresponding artifact-attenuated DRR obtained under different radiographic projection configurations; and training the machine-learning model using the first radiographs of the pairs as inputs and the corresponding second radiographs as target outputs.

[0037] The presently disclosed subject matter further contemplates any one of a computer- implemented method, a computer system, and a non-transitory program storage device, each dedicated to attenuating artifacts in an input radiograph compromising: receiving the radiograph; applying one or more machine-learning models selected from a collection of machine-learning models, each machine-learning model in the collection being trained using a training dataset comprising radiograph pairs produced by one or more processing pipelines, each pair including a first radiograph containing artifacts and a second radiograph in which the artifacts are attenuated by the respective processing pipeline, the processing pipelines differing in the processes used to generate the second radiograph of each pair; and generating one or more artifact-attenuated radiographic outputs each corresponding to a respective machine-learning models applied on the radiograph.

[0038] According to some examples, the collection of machine-learning models includes multiple models respectively trained on radiograph pairs derived from different combinations of the processing pipelines, and the system is configured to apply two or more of these models to the radiograph to generate multiple artifact-attenuated output images, each output corresponding to a different machine-learning model and reflecting its respective attenuation characteristics.

[0039] According to some examples, the system is further configured to receive user instructions identifying a subset of the machine-learning models from the collection to be applied to the radiograph, and the processing circuitry applies is configured to be responsive to the instructions and apply the machine-learning models indicated by the user instructions to generate the corresponding artifact-attenuated output images.

[0040] According to some examples, at least one of the machine-learning models in the collection is trained using a training dataset comprising radiograph pairs generated from two or more of the processing pipelines, thereby enabling the corresponding artifact-attenuated output to reflect attenuation behavior learned jointly from multiple pipeline domains.

[0041] The computer system can be integrated with or operatively connected to an X-ray device to enable the X-ray device to provide artifact-attenuated images.

[0042] BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to understand the presently disclosed subject matter and to see how it may be carried out in practice, the subject matter will now be described, by way of non-limiting examples only, with reference to the accompanying drawings, in which:

[0044] FIG. 1 is an image showing a panoramic radiograph displaying the full maxillofacial region at the top, with three corresponding images below that highlight different artifact types through colored overlays, from left to right: the ghost jaw, the spine, and the air gap;

[0045] FIG. 2 is a block diagram illustration of a computer system configured with machinelearning enhancement of X-ray images, according to some examples of the presently disclosed subject matter;

[0046] FIG. 3 shows, in a side-by-side view, a generalized flowchart of operations and a corresponding visual demonstration of an artifact-attenuation process for panoramic radiographs, according to some examples of the presently disclosed subject matter;

[0047] FIG. 4 shows examples of original panoramic radiographs and their corresponding deshadowed radiographs, according to some examples of the presently disclosed subject matter;

[0048] FIG. 5 is a block diagram illustration of a computer system configured with selective machine-learning enhancement of X-ray images, according to some examples of the presently disclosed subject matter;

[0049] FIG. 6 illustrates a flowchart of operations for performing selective artifact attenuation in radiographs, in accordance with certain examples of the presently disclosed subject matter;

[0050] FIG. 7 is a block diagram illustration of a computer system configured with sequential machine-learning enhancement of X-ray images, according to some examples of the presently disclosed subject matter;

[0051] FIG. 8 is a block diagram illustration of a computer system configured with CT-based synthetic-radiograph- enhancement of X-ray images, according to some examples of the presently disclosed subject matter;

[0052] FIG. 9 illustrates a flowchart of operations for generating synthetic training data using digitally reconstructed panoramic radiograph (DRR) from 3D CT data and training an artifactcorrection model, according to some examples of the presently disclosed subject matter;

[0053] FIG. 10 illustrates a process flow of the generation of a digitally reconstructed panoramic radiograph (DRR) from 3D CT data, according to some examples of the presently disclosed subject matter;

[0054] FIG. 11 illustrates a process flow of the generation of a digitally reconstructed panoramic radiograph (DRR) after artifact-mask segmentation and inpainting of 3D CT data, according to some examples of the presently disclosed subject matter;

[0055] FIG. 12 illustrates two sampling rays used for digitally reconstructed radiograph (DRR) reconstruction, according to some examples of the presently disclosed subject matter;

[0056] FIG. 13 is a block diagram illustration of a computer system configured with a multisource training architecture of a machine-learning for enhancement of X-ray images, according to some examples of the presently disclosed subject matter.

[0057] DETAILED DESCRIPTION

[0058] In the context of dental radiographs, the three artifact types mentioned above can be conceptually regarded as blurry and partially transparent intensity layers cast onto the targeted dental anatomy image. The inventors have found that although their intensity results from X- ray transmission through different materials (e.g. an air gap or bone) rather than light-source obstruction (e.g. the sun behind a tree), the resulting overlap is visually similar to a sunlight shadow onto a natural scene. Building on this observation, the presently disclosed subject matter introduces the inventive realization that a deep neural network pre-trained for optical shadow removal can be repurposed to correct X-ray artifacts. By adapting the input radiographic data to align with the model's expected domain— such as masking and by intensity inversion for radiopaque artifacts —the optical shadow-removal framework (de-shadowing) originating from other computer domains, can be directly applied to radiographic imagery without retraining, enabling artifact correction.

[0059] According to one example, a two-step process is proposed for removing specific artifacts and structures from panoramic radiographs. Step one involves segmenting each of the artifacts. In step two, the segmentation masks generated in step one and the original panoramic radiograph are forwarded to a de-shadowing deep learning model, where the "shadow" region is defined by the segmentation masks. The final image is produced from passing the radiograph through the de-shadowing model multiple times, each time removing a specific artifact from the image. Publicly available datasets, such as the Image Shadow Triplets Dataset (ISTD), Adjusted ISTD (ISTD+), and the Shadow Removal Dataset (SRD), can be used for training the model. These datasets consist of high-quality image pairs, including shadowed images, their corresponding shadow-free versions, and annotated shadow masks, which serve as ground truth for the model's learning process. Notably, the following description is presented with various details that are intended solely as examples and are not meant to be limiting.

[0060] Bearing the above in mind, attention is drawn to Fig. 2, which is block diagram illustration of a computer system 100 configured with machine learning enhancement of X-ray images, according to some examples of the presently disclosed subject matter.

[0061] System 100 can comprise or be otherwise operatively connected to an X-ray device 120, configured to capture X-ray images that may contain artifacts, such as overlapping layers resulting from the superposition of anatomical and non-anatomical regions. In some examples, system 100 is integrated with the X-ray device (e.g., implemented as dedicated processing circuitry within the device's computer system) to provide the X-ray device with artifact correction capabilities (e.g., in real-time). In other examples, system 100 is not directly connected to the X-ray device and may receive the X-ray images from another source, such as from computer storage.

[0062] System 100 can comprise one or more processing circuitries configured to execute various processing operations on X-ray images for attenuating (e.g., removing) artifacts and enhancing image quality. A processing circuitry can include one or more processors and one or more memories (not shown). In some examples, the processing circuitry is configured to execute several functional modules according to computer-readable instructions implemented on a non-transitory computer-readable memory included in the processing circuitry. These functional modules are hereinafter referred to as being comprised within the processing circuitry.

[0063] According to one non-limiting example, the processing circuitry includes the following functional modules:

[0064] • Input interface (102): The input interface is configured to receive the X-ray image from the X-ray device or from some other external source. This component serves as the entry point for the captured image data into the system.

[0065] • Segmentation training module (104): The segmentation training module is configured to train the segmentation machine learning (ML) model (e.g., and autoencoder such as a U-net network or a generative adversarial network (GAN)) using labeled X-ray image data. This module identifies distinct regions within the X-ray image that correspond to artifact areas. Training can be accomplished, for example, by creating one or more training datasets in which X-ray images are labeled (e.g., manually) to identify artifact regions. In some examples where multiple ML models are used, each model may be trained to segment specific types of artifacts, such as air gaps or overlapping anatomical structures, to improve segmentation accuracy. Notably, in some examples, part or all of the training is not performed by system 100; instead, system 100 receives the trained model after training has been performed elsewhere.

[0066] • Segmentation ML model (106): The segmentation ML model is configured to receive the X-ray image and segment it into regions, identifying specific areas affected by artifacts. The output of the segmentation ML model can include for example a mask that highlights artifact regions, which can then guide the artifact correction process.

[0067] • Artifacts Correction ML model (108): The artifact correction ML model is configured to receive as input the original X-ray images with artifacts and the output of the segmentation ML model, which indicates the areas affected by artifacts (artifact regions). It processes these inputs to reconstruct artifact-attenuated (e.g., artifact-free) data. The model is trained and configured to apply corrections specifically to artifact regions, producing a modified X-ray image with reduced or eliminated artifacts. In some examples, the artifact correction ML model is a pre-trained de-shadowing deep learning model.

[0068] System 100 can further include computer data storage (110), configured to store any data necessary for operating system 100, including, for example, computer software loaded during the execution of any one of the modules described above, intermediate processing results generated by system 100, processed X-ray images with removed or attenuated artifacts (e.g., artifact-free X-ray images), and more.

[0069] In some examples, system 100 can optionally comprise a User Interface configured to enable user interaction with system 100. The user interface can include a display device, user interaction devices (e.g., a computer mouse and keyboard), and a graphical user interface (GUI) configured to enable, inter alia, user-specified inputs related to system 100. For instance, the user may view processing results or intermediate processing results on the display, such as X- ray images with and without artifacts, following artifact extraction.

[0070] It is noted that any reference made herein to a machine-learning model system component, such as models 106, 108, 116 (mentioned below), contemplates the execution environment, including the computational resources, processing hardware, and software framework required for the model's execution and inference.

[0071] FIG. 3 shows, in a side-by-side view, a generalized flowchart of operations and a corresponding visual demonstration of the artifact-attenuation process for panoramic radiographs, in accordance with certain examples of the presently disclosed subject matter. The operations shown in FIG. 3 are presented by way of example only and for ease of understanding, with reference to components of computer system 100 described with reference to FIG. 2.

[0072] At the beginning of the process, one or more initial radiographs are obtained (block 301; e.g., by the input interface 102). For example, a panoramic X-ray imaging system (120) may be operated on a patient to generate one or more panoramic images of the maxillofacial region. The resulting radiographs, which typically include imaging artifacts, are received from the X-ray machine and provided to the processing system for artifact correction. Subsequent processing operations are then performed on the received images, to attenuate one or more artifacts. Image (A) in FIG. 3 illustrates an example of such an initial panoramic radiograph of a maxillofacial region comprising multiple imaging artifacts.

[0073] The one or more radiographs are applied to a neural network configured for artifact segmentation (block 303; e.g., by the segmentation ML model 106). In one example, the neural network comprises a U-Net model, which is a convolutional neural network architecture commonly used for semantic image segmentation tasks due to its reliability and ease of implementation. The U-Net model receives a panoramic radiograph as input and generates one or more semantic segmentation masks corresponding to distinct artifact types, such as the air gap, the spine, and the ghost jaw. In some examples, a separate U-Net may be trained for each of these segmentation tasks to improve precision and task-specific performance.

[0074] The U-Net includes a contracting (encoding) path and an expanding (decoding) path. The contracting path performs, for example, a sequence of convolution and downsampling operations (e.g., repeated 3 x 3 convolutions followed by ReLU activation and 2 2 max-pooling with stride 2) to extract spatial features from the input image at multiple resolutions. The expanding path mirrors the contracting path but performs up-sampling instead of downsampling to reconstruct the semantic segmentation mask to the original image dimensions while preserving spatial context from earlier layers through skip connections. Accordingly, the U-Net performs pixel-level classification of the radiograph, assigning each pixel to one of multiple artifact categories or to background. The output of the U-Net is an artifactspecific semantic segmentation mask reflecting the dimensions of the input image. It is noted that the specific parameters and layer configurations described herein are provided by way of non-limiting example only. Training of each U-Net model may employ, for example, a Dice Loss function to optimize overlap between predicted and reference masks. The central mask images shown to the left of block 203 in FIG. 3 exemplify these outputs— each mask highlights the spatial extent of one artifact class.

[0075] The original radiograph and the generated semantic segmentation masks are then supplied sequentially to a de-shadowing deep-learning model (block 305; e.g., by the artifact correction ML Model 108). For example, the original radiograph and the air-gap semantic segmentation mask are first provided as inputs to the de-shadowing network, which produces a new radiograph with the air-gap "shadow" attenuated (e.g., removed). Similarly, this output image, and the ghost-jaw semantic segmentation mask are re-supplied to the de-shadowing network to produce a new image with the ghost-jaw shadow removed. Lastly, the resulting intermediate image and the spine semantic segmentation mask are given to the de-shadowing network to create the final enhanced (artifact-corrected) radiograph. Images (B) and (C) in FIG. 3 illustrate intermediate examples of these de-shadowing operations, showing progressive attenuation of respective artifact types.

[0076] According to one non-limiting implementation example, the de-shadowing network (corresponding to model 108 in FIG. 2) comprises a state-of-the-art deep model such as ShadowFormer, which was originally developed for optical-domain shadow removal and trained on large datasets of natural images (e.g., ISTD+). ShadowFormer operates in accordance with a Retinex-based formulation [1] and leverages self-attention between similar image patches located inside and outside of a pre-segmented shadow mask to correct for shadow- induced changes in pixel values while maintaining texture continuity across mask borders. In some examples, the network can be applied directly using the pre-trained weights obtained from its optical-domain training.

[0077] Notably, for radiopaque artifacts such as the ghost jaw and spine, the radiograph may be inverted (i.e., converted to a negative representation) prior to processing so that the bright artifact regions become hypointense "shadows" consistent with those encountered during ShadowFormer training. Thus, in some examples, the process may further include an inversion procedure performed in association with block 305— i.e., before or during application of the de-shadowing model to the radiopaque artifact regions identified by segmentation in block 303. This inversion step adapts the radiographic input to the intensity characteristics expected by the optical-domain of the de-shadower and represents an additional distinctive feature of the presently disclosed approach. According to some examples, system 100 may comprise a processing module (122) configured to invert the radiograph immediately before it is provided to the de-shadowing model. Inversion can be performed either before or after segmentation, depending on the processing requirements, with some methods benefiting from inversion prior to segmentation and others applying it afterward.

[0078] In some examples, the de-shadowing network generates corrected pixel values only for the regions indicated by the corresponding semantic segmentation mask, while the unmasked portions of the radiograph remain unmodified. Accordingly, after each de-shadowing pass, the system combines the corrected pixel values for the masked artifact regions with the unmodified pixel values of the intermediate radiograph to form an updated radiograph for the subsequent de-shadowing stage. This integration ensures that only the artifact-affected regions are replaced by the reconstructed values generated by the de-shadowing model, whereas all nonartifact regions retain their original appearance. The sequential merging of corrected regions across the air-gap, ghost-jaw, and spine passes yields the fully enhanced radiograph output depicted in FIG. 3.

[0079] The final output of enhanced radiograph obtained from the de-shadowing model corresponds to block 307, representing the completion of the artifact-removal sequence. In some examples, the output image may be inverted back to its original intensity distribution (e.g., by processing module 122) to restore the radiograph to its natural appearance. The artifact-corrected panoramic radiograph can be stored, for example, in a computer data storage (e.g., 110), as shown in FIG. 2. The resulting artifact-corrected images may be further used for visualization, clinical evaluation, or additional analysis modules. Image (D) in FIG. 3 illustrates an example of the final artifact-attenuated panoramic radiograph, showing substantially improved contrast uniformity and visibility of anatomical structures following sequential removal of the respective artifact types. An example of the output of the artifact correction approach, showing the mandibular canal both before and after de-shadowing in sided by side view, can be seen in Fig. 4. Experimental results supporting the effectiveness of the artifact attenuation technique described above are provided below.

[0080] Turning to FIG. 5, it shows a block diagram illustrating a computer system configured with selective machine-learning enhancement of X-ray images, according to some examples of the presently disclosed subject matter. The architecture shown in Fig. 5 differs from the system described above with reference to Fig. 2 by specifically targeting artifact regions that overlap with anatomical structures, ensuring that corrections are applied only to overlapping artifact regions while preserving artifact regions that do not overlap with anatomy. This approach selectively removes artifacts from the anatomical structure (e.g., the mandible), reducing obstructions within the structure itself while preserving the artifacts in surrounding areas and thus maintaining contextual detail. By targeting only the overlapping regions, the method enhances contrast and visibility of anatomical structures while maintaining adjacent image context, thereby assisting in accurate interpretation and diagnosis.

[0081] According to this example, system 100 further includes, in addition to the components specified above with respect to Fig. 2:

[0082] • Anatomy Segmentation Training Module (112): The anatomy segmentation training module is configured to train an anatomical segmentation ML Model (116) using labeled anatomical data. This training enables the model to recognize anatomical structures within the X-ray image, ensuring they are distinguished from artifact regions. As with the artifact-segmentation model, training may occur externally, with system 100 receiving the pretrained anatomical segmentation model. This model may be implemented, for example, as an autoencoder architecture such as a U-Net, or as a GAN-based model.

[0083] • Anatomical Segmentation ML Model (116): The anatomical segmentation ML model is configured to identify and segment anatomical structures within the X-ray image. By distinguishing anatomical regions from artifacts, this model helps ensure that only artifact regions overlapping with anatomical parts are targeted during the artifact correction process.

[0084] • Segmentation Integration Module (118): The segmentation integration module is configured to combine the outputs from the artifact segmentation ML model (106) and the anatomical segmentation ML model (116). This integration creates a comprehensive segmentation map that precisely identifies artifact regions overlapping with anatomical structures, guiding the artifact correction model to selectively remove these regions while preserving non-overlapping artifacts.

[0085] • Image Processing Module (122): The image processing module is configured to generate a negative version of the original X-ray image. As mentioned above, in example, where artifacts that are brighter than the surrounding tissues this module can be used. In such cases, since the de-shadowing model is designed to remove shadow artifacts, which are typically characterized as darker regions compared to their surroundings, a brightness inversion is applied to maintain this assumption for radiopaque artifacts. Examples of such artifacts include the ghost jaw or spine in a panoramic X-ray, or the ribs in a chest X-ray. The brightness inversion ensures that the brighter artifact regions are treated as shadows by the de-shadowing model. Once the de-shadowing process is complete, the resulting image may undergo a second brightness inversion to restore its original appearance, ensuring the corrected image is anatomically accurate and visually consistent. This can be accomplished, for example, by applying the output of the artifact correction ML model to the image processing module 122 to obtain the reverse effect on the images.

[0086] FIG. 6 illustrates a flowchart of operations for performing selective artifact attenuation in radiographs, in accordance with certain examples of the presently disclosed subject matter. The operations shown in FIG. 6 are presented by way of example and for ease of understanding, with reference to components of computer system 100 described with reference to FIG. 5.

[0087] One or more panoramic radiographs are first obtained (block 601; e.g., by the Input Interface 102). As discussed above with reference to FIG. 2, the radiographs may be generated by an X-ray imaging device (120) and may include anatomical structures partially obscured by artifacts such as the air gap, spine, or ghost jaw. The obtained radiographs are provided to system 100 for selective artifact correction.

[0088] The radiographs are applied to an artifact-segmentation model (block 603a; e.g., the Segmentation ML Model 106) configured to detect and delineate artifact regions. The model generates one or more semantic segmentation masks corresponding to respective artifact types. In some examples, the artifact-segmentation model may comprise a deep neural network architecture, such as a U-Net, trained to perform artifact segmentation. Other convolutional or transformer-based segmentation networks may likewise be used. In some implementations, separate models may be trained for each artifact category (e.g., air gap, spine, ghost jaw) to improve segmentation precision. Each model outputs a mask highlighting the spatial extent of the corresponding artifact within each panoramic image.

[0089] In a separate or subsequent operation, the radiographs may be applied to an anatomical-segmentation model (block 603b; e.g., the Anatomical Segmentation ML Model 116) configured to identify anatomical structures of diagnostic interest, for example the mandible. In some examples, the anatomical-segmentation model may comprise a neural network architecture (e.g., a U-Net or transformer-based segmentation network) trained to identify anatomical structures in panoramic radiographs. The anatomical-segmentation model produces a semantic segmentation mask delineating the anatomical regions within the same radiographs. These masks enable the system to differentiate anatomical areas from background regions and non-anatomical artifacts.

[0090] The artifact and anatomical segmentation masks are then integrated (block 605; e.g., by the Segmentation Integration Module 118) to generate an overlap map that identifies regions where the segmented artifacts spatially intersect with the segmented anatomical structures. The overlap map defines the target regions for correction, ensuring that only artifacts overlapping with anatomy are processed while non-overlapping artifact regions remain unmodified.

[0091] As previously explained with reference to FIG. 5, in some examples, the system may perform an optional intensity-inversion step (block 607; e.g., by the Image Processing Module 122) to prepare the radiographs for de-shadowing. For radiopaque artifacts (e.g., ghost jaw or spine) appearing as bright regions, the image is inverted so that these regions become dark "shadows," consistent with the data distribution used to train the pre-trained de-shadowing model.

[0092] The overlap maps and corresponding radiographs (inverted when applicable) are then applied to a de-shadowing deep-learning model (block 609; e.g., the Artifact-Correction ML Model 108) configured to remove artifacts selectively within the overlapping regions. In one example, the de-shadowing model may comprise a ShadowFormer network leveraging selfattention between image patches located inside and outside of the masked regions to correct local intensity distortions while maintaining texture continuity. The de-shadowing model operates selectively over the overlapping regions defined by the overlap map generated in block 607, leaving non-overlapping areas unmodified.

[0093] Following selective artifact correction, the radiographs may undergo a reverse inversion (block 611; e.g., again via the Image Processing Module 122) to restore the original brightness distribution for radiopaque artifacts that were previously inverted. This ensures that the final corrected images maintain anatomically accurate intensity relationships.

[0094] The selectively corrected panoramic radiographs can be provided e.g., displayed and / or stored (block 613; e.g., in Data Storage 110). The resulting images exhibit improved visibility and contrast of anatomical regions such as the mandible, with contextual surroundings preserved.

[0095] In addition to the de-shadowing repursuing in the artifact-removal techniques described above, further radiographic enhancement can be achieved by adapting principles originally developed for optical de-hazing in natural-light imaging. Unlike shadow removal, which compensates for directional illumination occlusion, de-hazing addresses the scattering of light by a semi-transparent medium such as fog or smoke. In panoramic radiography, X-ray scatter and soft-tissue superimposition create an analogous visual degradation— a diffuse, low-frequency veil that reduces image contrast and obscures fine anatomical structures.

[0096] This similarity between optical haze and radiographic scatter has been recognized in medical imaging research, where de-hazing formulations originally developed for natural images— such as the Dark Channel Prior (DCP) — have been adapted to suppress haze-like scattering effects in X-ray projections. Building on this prior work, the inventors conceived that modern deep-learning-based de-hazing architecture, despite being originally trained on natural imagery, could be repurposed for radiographic applications and may generalize to panoramic radiographs to improve visibility in regions affected by scattered radiation.

[0097] The inventors have identified a physical and perceptual similarity between the degradations caused by optical haze and the low-frequency scattering that occurs in panoramic X-ray imaging. In radiographic acquisitions, scattered photons and overlapping soft-tissue layers act as a semi-transparent medium that attenuates the primary X-ray beam, producing a diffuse intensity veil over the anatomical structures. This phenomenon results in a perceptual effect analogous to optical haze, characterized by reduced local contrast, diminished dynamic range, and blurring of fine anatomical detail.

[0098] Accordingly, in some examples, the system architecture and operational flow described with reference to FIGS. 2 and 3 may be modified to utilize a de-hazing model instead of a de-shadowing model. In such examples, the de-hazing model is integrated into system 100 and executed by the Artifact-Correction ML Model (108). In this configuration, the input radiographs are applied directly to the de-hazing model— without prior artifact-segmentation step— because the de-hazing approach does not require segmentation masks or the imageprocessing operations used to generate or apply them. The de-hazing model estimates a spatially varying transmission map corresponding to the contributions of scattered radiation and overlapping anatomical structures. The restored output represents a "de-scattered" radiograph in which both radiolucent and radiopaque veiling effects are attenuated while preserving anatomical fidelity.

[0099] The presently disclosed subject matter further contemplates an approach that combines a de-shadowing model with a de-hazing model in a sequential artifact attenuation pipeline. According to this approach, de-shadowed images— previously processed to suppress structured artifacts such as ghost jaws or spinal overlays— are supplied as inputs to the de- hazing model. Supplying de-shadowed images in this manner enhances the effectiveness of the de-hazing process, as the model receives an input in which structured artifacts have already been attenuated. This sequential configuration reduces both structured and diffuse artifacts, yielding clearer visualization of diagnostically critical regions such as the mandibular canal, condylar margins, and adjacent osseous structures.

[0100] FIG. 7 shows a modified system architecture of that shown in FIG. 2, which has been adapted to support the sequential processing approach described above. In this configuration, system 100 comprises two Artifact-Correction ML Models, a de-shadowing model (108A) and a de-hazing model (108B), enabling the radiograph to undergo two successive image enhancement stages. In the first stage, the segmentation-based pipeline of FIG. 2 is used to generate artifact masks via the Segmentation ML Model (106), after which the Image processing module (122) prepares these masks for application by the de-shadowing model (108A). The de-shadowing model attenuates structured overlap artifacts such as the ghost jaw, spine, or air gap to produce a de-shadowed radiograph.

[0101] This de-shadowed radiograph is then forwarded to the de-hazing model (108B). The dehazing model estimates and removes diffuse low-frequency degradations arising from scattered radiation and overlapping soft-tissue regions. This sequential process can also be realized in the processes demonstrated in FIGS. 3 and 6, where the outputs generated at blocks 305 and 609, respectively, may be subsequently provided to the de-hazing model (108B) to perform a second enhancement stage. By operating on an already de-shadowed input, model 108B can more effectively enhance contrast uniformity and improve visualization of fine anatomical details, including the mandibular canal and condylar margins. Notably, in alternative examples, the order is reversed and de-shadowing is applied after de-hazing.

[0102] GROUND TRUTH CREATION FOR SUPERVISED ARTIFACTS REMOVAL LEARNING

[0103] As noted earlier, a key challenge in developing supervised models for removing overlapping artifacts is the absence of suitable training data in which these artifacts are not present. Supervised learning requires ground-truth images that are free of the artifacts being corrected yet acquired 2D radiographs inherently contain such artifacts due to human anatomy, for example, the air gap, ghost jaw, and spine in panoramic X-rays, or the rib cage in chest X-rays.

[0104] For this purpose, it is proposed herein to use 3D computed tomography (CT) scans of the corresponding anatomical region (e.g., head or chest) and generate a pair of digitally reconstructed radiographs (DRRs): one that includes the artifacts, and another in which the artifacts are removed from the 3D CT volume prior to projection. Unlike X-ray images, which inherently contain artifacts due to overlapping anatomical and non-anatomical regions, CT images are captured in a way that minimizes these artifacts. This makes them suitable for generating artifact-attenuated (e.g., artifact-free) reference data. By modifying CT images to simulate artifact-attenuatedand artifact-containing conditions, the system can create the paired training dataset needed to train the model effectively. By using CT images, the system overcomes the limitations posed by the inherent presence of artifacts in X-ray images, providing a more robust and effective training dataset for the correction model.

[0105] Once generated these images serve as supervised training data for any suitable machine-learning algorithm that, given an input X-ray image affected by the considered artifact(s), generates an output X-ray image in which said artifact(s) are attenuated. This includes a dedicated machine-learning model (e.g., a deep-learning network) trained specifically for this purpose, as well as the re-training of an existing machine-learning model, such as the de-shadowing or de-hazing model mentioned above, to refine its performance in addressing the problem of removing overlap artifacts from X-ray images.

[0106] Turning to Fig. 8 is a block diagram illustration of a computer system 100A configured to provide X-ray artifact correction capabilities according to some examples of the presently disclosed subject matter.

[0107] In some examples, system 100A can be configured to operate in conjunction with an X- ray Imaging Device (120), which captures X-ray images that may contain artifacts arising from overlapping anatomical and non-anatomical regions. System 100A can be directly connected to the X-ray imaging device or can receive X-ray images from an external source. In some configurations, system 100A may be integrated within the X-ray device itself to enable on- device artifact correction.

[0108] System 100A is shown to include, by way of example, components configured to perform operations necessary for CT-based digitally reconstructed radiographs (DRRs) generation, model training and X-ray image attenuation. These components include:

[0109] • Input Interface (102): The input interface is configured to receive X-ray images from the X-ray imaging device. This interface serves as the initial entry point for captured image data into system 100A. • DRRs Synthesizer (130): The DRRs synthesizer is configured to generate digitally reconstructed radiographs (DRRs) that emulate X-ray images. These DRRs can be created to either include or exclude artifacts (artifact-containing and artifact-attenuatedDRRs), providing a comprehensive dataset for training machine learning models. In some examples, 3D CT images of the area of interest are provided to the synthesizer which is configured to generate paired DRRs by projecting an unmodified CT volume to obtain an artifact-containing DRR and by projecting a modified CT volume— produced by suppressing contributions from selected anatomical structures using suitable processing techniques —to obtain an artifact-attenuatedor artifact-attenuated DRR suitable for supervised training.

[0110] • Artifacts Correction Model Training Module (132): This module is configured to train the Artifacts Correction ML Model (108) using the paired artifact-attenuatedand artifactcontaining DRRs created by the DRRs synthesizer. By training on this dataset, the model learns to identify and correct artifact regions in X-ray images, using the artifact-attenuatedDRRs as ground truth references. As mentioned above, in some examples part or all of the training is not performed by system 100A; instead, system 100A simply receives the trained model after training has been performed elsewhere.

[0111] • Artifacts Correction ML Model (108): Once trained, this machine learning model is configured to receive X-ray images containing artifacts, and apply corrections to the identified artifact regions, reconstructing an artifact-attenuatedversion of the image. This results in enhanced X-ray images with reduced or eliminated artifacts, improving image clarity for diagnostic purposes.

[0112] • Computer Data Storage (110): This storage component is configured to store essential data for system 100, including 3D CT images, the training dataset, artifact- attenuatedX-ray images, and any software necessary for the operation of the modules. It ensures that data required for both the training process and the real-time artifact correction is readily available.

[0113] FIG. 9 illustrates a flowchart of operations of a process of generating CT-based DRRs, training, and executing an artifact-correction model using the DRRs, in accordance with certain examples of the presently disclosed subject matter. The operations shown in FIG. 9 are presented by way of example and for ease of understanding, with reference to components of computer system 100 described with reference to FIG. 8.

[0114] A 3D CT volume of the target anatomical region is first obtained (block 901; e.g., through the Input Interface 102). The CT volume may correspond, for example, to the head or chest region, depending on the intended application. The volumetric CT data provides a high- resolution representation of anatomical structures without the projection-domain overlap artifacts characteristic of two-dimensional radiographs.

[0115] The 3D CT volume is then processed to generate a DRR (block 903; e.g., using the DRR Synthesizer 130). To generate artifact-containing and artifact-attenuated DRRs, a first radiographic projection configuration is applied to the unmodified CT volume, and a second radiographic projection configuration is applied to the modified CT volume, wherein the first and second projection configurations share the same projection geometry so as to produce pixel-aligned DRRs. The unmodified CT volume is projected to emulate the formation of artifact-containing radiographs. Because the projection accumulates attenuation values along these ray paths, anatomical structures located outside the focal region naturally produce the same overlap artifacts observed in real radiographs, such as air-gap, ghost-jaw, spine, or ribcage artifacts. FIG. 10 illustrates an example of this projection step, showing a CT slice and its corresponding DRR generated using panoramic projection geometry.

[0116] Next, an artifact-segmentation network is applied (block 905; e.g., by a Segmentation Module within the DRR Synthesizer 130) to identify anatomical regions within the CT volume whose projection through the selected X-ray geometry gives rise to artifacts in the DRR. The segmentation network may comprise a deep neural architecture, such as a U-Net, V-Net, or transformer-based model (e.g., SWIN-UNETR), trained to detect volumetric regions corresponding to these artifact-causing anatomical structures (e.g., air cavities, contralateral mandible, cervical spine, rib cage). Examples of the segmentation process are illustrated in FIG. 11, which depicts the generation of a 3D artifact mask highlighting regions such as the air gap.

[0117] Following segmentation, an inpainting operation is performed (block 907; e.g., by an Inpainting Module within the DRR Synthesizer 130) on the segmented anatomical structures to remove or replace them with values consistent with surrounding tissue, thereby producing a modified CT volume in which the anatomical structures responsible for artifact formation during projection have been removed or replaced. In the example of FIG. 11, this step is shown as "artifact mask inpainting," which removes the air-gap region and restores anatomical continuity in the CT data. In some examples, the inpainting replaces the Hounsfield values (H V) of the segmented region (e.g., air HV close to -1000 in CT images) with values similar to those of the adjacent tissues, so that the removed anatomical structure does not contribute to the DRR generated from the modified CT volume. The inpainting may be performed using any suitable prior-art method.

[0118] The modified CT volume is then projected again using a second radiographic projection configuration (block 909) to generate an artifact-attenuated DRR. The artifact-containing DRR (from block 903) and the artifact-attenuatedDRR (from block 909) form a matched pair suitable for supervised learning.

[0119] The paired DRRs are organized into a training dataset (block 911; e.g., by the Data Storage 110), which may include multiple anatomical examples and artifact types. This dataset is then used to train the artifact-correction machine-learning model (block 913; e.g., the Artifact-Correction Model 108) within the Training Module 132. The model learns to transform artifact-affected X-ray images into corrected outputs by mapping the relationships between the artifact-containing and artifact-attenuated DRR pairs. The model may be implemented using any suitable architecture, such as a convolutional neural network (e.g., U-Net, V-Net) or a transformer-based network (e.g., SWIN-UNETR).

[0120] Once trained, the artifact-correction model is applied to real X-ray radiographs for artifact removal (block 915). This inference stage may occur in real time— when the trained model is integrated into an X-ray imaging device (e.g., 120)— or asynchronously when processing previously acquired radiographs. The resulting corrected images exhibit attenuated overlap artifacts and enhanced visibility of diagnostically relevant anatomical features.

[0121] As an alternative to the inpainting and modified-volume projection sequence described above, artifact-attenuated DRRs may also be generated by modifying the virtual sampling-ray geometry during DRR synthesis (e.g., by DRR synthesizer 130). In this approach, the artifactcontaining DRR generated at block 903 of FIG. 9 is obtained by projecting the unmodified CT volume using a first radiographic projection configuration in which virtual rays originate from a source location outside the CT volume and intersect one or more artifact-causing structure. Instead of performing the inpainting operation of block 907 and projecting a modified CT volume as in block 909, the system applies a second radiographic projection configuration in which virtual rays originate from a source location inside the CT volume and bypass the artifactcausing structures, thereby generating an artifact-attenuated DRR without altering the CT volume.

[0122] FIG. 12 illustrates this concept. A sampling ray originating from a square marker located outside the head intersects air voxels, resulting in an air-gap artifact in the artifact-containing DRR. Repositioning the ray's origin to a star marker located inside the head allows the ray to avoid the air cavity entirely, producing an air-gap-free DRR that serves as the artifact- attenuated DRR in the training pair. Likewise, a ray originating from an external square marker on the opposite side intersects contralateral mandible voxels, generating a ghost -jaw artifact, whereas relocating the ray origin to the corresponding internal star marker prevents this intersection and eliminates the artifact.

[0123] Although such ray-source repositioning cannot be realized in real panoramic imaging systems, it is fully implementable in synthetic DRR generation. This geometric flexibility allows the system to generate artifact-attenuated DRRs without altering CT data. The artifact- attenuated DRRs produced via ray-path modification can be paired with the artifact-containing DRRs generated from the unmodified CT volume, thereby forming matched training pairs suitable for supervised training of the artifact-correction model.

[0124] Regardless of whether the inpainting-based method or the ray-path modification method is used, it is noted that while the system and workflow are illustrated in FIG. 8 and FIG. 9 as performing synthetic image generation, model training, and image correction within a single integrated framework, this is by way of illustrative example only. In some implementations, one or more of these stages— synthetic image synthesis, model training, and image correction— may be performed asynchronously and by different computational systems. For example, certain stages may be co-located (e.g., image synthesis and training performed on the same platform), while another stage (e.g., image correction) is executed on a separate system, such as an X-ray imaging device or remote processing unit. In other implementations, the trained artifact-correction model may be deployed independently of the training environment, for instance integrated into an X-ray imaging device for real-time correction of radiographs as they are acquired or used offline to process previously obtained X-ray images. Accordingly, the stages of data generation, model training, and model application, as shown in FIGS. 8 and 9, may be distributed or temporally decoupled, enabling flexible and asynchronous operation across centralized or decentralized infrastructures.

[0125] The presently disclosed subject matter further contemplates an additional approach for removing artifacts from radiographic images, in which a machine-learning model is trained using a supervised dataset composed of radiograph pairs generated by multiple processing pipelines. Their outputs serve as the target images in the supervised dataset, while the trained model used at inference is a separate artifact-correction model that learns from these target images collectively.

[0126] In some examples, the supervised dataset includes radiograph pairs originating from four distinct processing pipelines mentioned above. A first category comprises radiographs generated by the de-shadowing processing pipeline, in which structured overlapping-layer artifacts such as ghost-jaw, spinal, or air-gap regions are attenuated, as described with reference to FIGS. 2, 3, 6, and 7. A second category comprises radiographs generated by the de-hazing processing pipeline, in which diffuse, low-frequency degradations associated with scattered radiation and soft-tissue superimposition are attenuated. A third category comprises radiographs generated by the sequential de-shadowing and de-hazing processing pipeline, as described with reference to FIG. 7, in which both structured and diffuse artifacts are attenuated. A fourth category comprises radiographs generated by the CT-based synthetic radiographs processing pipeline, consisting of paired digitally reconstructed radiographs (DRRs) including one radiograph containing overlapping-layer artifacts and a corresponding radiograph in which the artifacts are attenuated through CT-domain segmentation, inpainting, or ray-path modification techniques, as described with reference to FIGS. 8-12.

[0127] Together, these four categories form a heterogeneous collection of radiograph pairs used for training a unified artifact-correction model. Two or more, and in some examples three or all four, of these categories are combined to form a mixed-domain supervised dataset, exposing the model to a broad spectrum of artifact characteristics, attenuation behaviors, and intensity distributions.

[0128] As illustrated in FIG. 13, the processed outputs from two or more of the de-shadowing, de-hazing, sequential de-shadowing and de-hazing, and CT-synthetic pipelines are collected into a unified data repository (31) together with their corresponding input radiographs. This mixed-domain dataset is provided to a training module (33), implemented on processing circuitry configured to train a machine-learning model— such as a convolutional neural network, a transformer-based network, a diffusion-based image-enhancement model, or another suitable architecture— and to learn a mapping from artifact-affected radiographs to their corresponding enhanced outputs.

[0129] In FIG. 13, three functional groupings are illustrated— a Data Collection Layer, a Training Layer, and an Inference Layer. These groupings are provided solely for explanatory convenience. Each "layer" represents a respective computing device or computing environment comprising one or more processing circuitries configured to execute relevant operations. The illustrated division into layers is by way of example only; other divisions, reorganizations, or consolidations are likewise contemplated.

[0130] One example is a diffusion model, which is well suited for this multi-source training configuration because it explicitly models the forward noising and reverse denoising processes, allowing it to learn a conditional distribution over heterogeneous target domains while maintaining high-fidelity anatomical detail. By leveraging this generative structure, a diffusion-based artifact-correction model can unify disparate correction styles and generate anatomically consistent, artifact-attenuated radiographs without the over smoothing, regression-induced blurring, or mode collapse that may occur with conventional deterministic regression models.

[0131] During training, input images from the different pipelines are progressively diffused and then reconstructed under a learned denoising trajectory, conditioned on the artifact- affected radiographs. This process enables training of the model (34) to handle multiple artifact styles, including structured overlap artifacts and diffuse haze-like degradations, within a single unified framework.

[0132] During inference, the trained model (34) receives an input panoramic radiograph and generates an output radiograph in which artifacts are attenuated, and anatomical structures exhibit improved clarity and contrast uniformity.

[0133] Accordingly, combining radiograph pairs from multiple processing pipelines provides a combined training benefit: the de-shadowing, de-hazing, and sequential processing pipelines contribute different attenuation behaviors, and CT-derived paired DRRs supply physicsconsistent supervision. The integration of these heterogeneous radiograph pairs within a unified diffusion-based training framework enables the trained model to generalize effectively across diverse imaging conditions and to produce radiographs with improved anatomical detail, reduced artifact burden, and enhanced diagnostic quality.

[0134] In some examples, the system is further configured to generate and present to a client device— such as a clinician workstation — multiple artifact-attenuated versions of the same input radiograph. Instead of training a single artifact-correction model using a specific combination of training data generated by the processing pipelines, the system may train multiple machine-learning models, with each model being trained on radiograph pairs originating from one or more of the processing pipelines described above. For instance, one model may be trained exclusively on radiograph pairs produced by the de-shadowing pipeline, another on de-hazing pipeline pairs, another on sequential de-shadowing and de-hazing pipeline pairs, and another on CT-derived DRR pairs. Additional models may be trained using mixed-domain datasets assembled from two, three, or all four of the pipelines, thereby capturing different attenuation behaviors, artifact characteristics, and intensity distributions. Each trained model therefore provides a distinct attenuation profile when applied to the same input radiograph.

[0135] Accordingly, for a given radiograph, the system can apply this collection of machinelearning models to generate multiple attenuated outputs, each reflecting the attenuation characteristics learned from its a respective combination of pipeline-derived training data. These outputs may emphasize different anatomical structures or reduce different categories of artifacts depending on the training composition.

[0136] The system is configured to present these multiple attenuation variants through a client-facing interface operatively connected to the processing circuitry. The interface can be configured to receive and display all or part of the attenuated images associated with the radiograph, with each image corresponding to a respective machine-learning model trained on radiograph pairs assembled from one or more of the processing pipelines. In some examples, the interface enables the clinician to perform operations such as browsing through the available attenuation variants, comparing two or more variants side-by-side, adjusting visualization parameters (for instance, zoom, contrast, or brightness), annotating or marking regions of interest, tagging or bookmarking preferred variants, and selecting a particular attenuation output for clinical review, export, or inclusion in a patient record. In some examples, the interface may further indicate, for each displayed attenuation result, the specific combination of processing pipelines used to generate the training dataset for the corresponding machine-learning model.

[0137] In some examples, the system can additionally be configured to receive instructions indicating which of the machine-learning models should be applied to a given radiograph. Such instructions may be provided through the client-facing interface or via an external connected system. For instance, a clinician may specify that only models trained on certain pipeline combinations are to be executed or may request attenuation outputs from a selected subset of the available models. Upon receiving these instructions, the system can apply the indicated models to the input radiograph and generate the requested attenuation results accordingly.

[0138] By generating multiple artifact-correction models based on different combinations of pipeline-derived training data, and by providing an interface through which the clinician can review, select, and instruct the system regarding the attenuation variants to be produced, the presently disclosed subject matter allows the clinician to identify and obtain the attenuation result that reveals the most diagnostically useful information for the imaging task at hand.

[0139] EXPERIMENTAL IMPLEMENTATION RESULTS

[0140] The presently disclosed subject matter was implemented and evaluated using empirical datasets and quantitative image-quality metrics, as described below. The following examples are provided for illustration purposes and to demonstrate one possible implementation of the disclosed methods and systems.

[0141] Datasets

[0142] Artifact Segmentation Datasets: In order to create a supervised learning algorithm to perform semantic segmentation of the interfering artifacts, panoramic radiograph datasets labeled with masks of the ghost jaw, air gap, and spine wre required. As there are no public datasets with labels for these artifacts, panoramic radiographs were labelled (in the current case by amnually labelin effort). Separate datasets were created for the ghost jaw, spine, and air gap, each consisting of 100 panoramic radiographs with corresponding segmentation masks. Images were annotated using ITK-SNAP [2] by a graduate student in Electrical Engineering under the supervision validation of two senior maxillofacial surgeons, each with extensive expertise in panoramic radiography.

[0143] Post-Enhancement Comparison Datasets: To validate de-shadowing enhancements, new semantic segmentation networks were trained to segment the mandible and the mandibular canal after image enhancement. The mandible analysis was conducted using a publicly available dataset of 116 panoramic radiographs taken at Noor Medical Imaging Center in Iran

[0010] , For the mandibular canal analysis, 100 panoramic radiographs from Sheba Medical Center were manually labelled using the same process used for the artifact detection datasets.

[0144] Implementation Details

[0145] All experiments were conducted with an NVIDIA GeForce RTX 2070 GPU. PyTorch was used for the implementation of all models and code infrastructure; additionally, MONAI was used [3] for dataset management and pre-processing steps. Each of the aforementioned datasets was split following an 80 / 10 / 10 train / validation / test split strategy. During training and evaluation, the radiographs were converted from DICOM files to 960x512 pixel PNG files. All of the semantic segmentation networks were trained for 100 epochs using an Adam optimizer [4] with betas - 0.9, 0.99. During training, random additions of Gaussian noise, coarse pixel dropout, contrast adjustments, rotations, and horizontal image flipping to diversify the model's input data were used. For de-shadowing implementation, the basic ShadowFormer code and infrastructure created by Guo [3] was recycled, with minor modifications to accommodate panoramic radiograph data, as well as to allow for removal of both "dark" (radiolucent) and "bright" (radiopaque) shadows. Weights were taken from the original ShadowFormer model trained on the ISTD+ dataset with no extra training or modifications applied.

[0146] Results

[0147] A subsample of the results is displayed in FIG. 4.

[0148] Artifact Segmentation Results

[0149] The Dice Scores of the U-Net segmentation of the artifacts were about 0.9 or higher: 0.8985 for the ghost jaw, 0.9153 for the air gap, and 0.9680 for the spine. Given the small training datasets, 80 images each, these scores are remarkably high.

[0150] Weber Contrast of Mandibular Canal

[0151] The Weber contrast [6] of the mandibular canal was calculated, both before and after de-shadowing enhancement. The Weber contrast, Cw, is defined by Equation 1, where Lsis the luminance of the symbol being assessed for, and Lb is the luminance of the background.

[0152] The mandibular canal is most easily observed by looking for its edges, which are relatively radiopaque compared to surrounding structures in the jaw and to the canal itself. To assess the change in Weber contrast across the dataset after the de-shadowing enhancements, the focus was on the contrast that defines the upper edge of the mandibular canal.

[0153] This analysis was performed twice: first on the full length of the mandibular canal, then focusing specifically on the region where the ghost jaw overlaps with the mandibular canal. The Weber contrast results, calculated for all 100 images that had mandibular canal segmentations, are shown in Table 1.

[0154] Table 1. The Weber contrast along the upper edge of the mandibular canal before and after artifact removal.

[0155] Mandible Segmentation Results

[0156] The quantitative performance of the enhancements was assessed by comparing U-Net training and performance for mandible segmentation on original and enhanced panoramic radiographs. Forthis task, the U-Net was trained to segment the mandible, which encompasses the full breadth of the panoramic radiograph and intersects with all of the artifacts of concern, making it a good proxy source of quantitative performance. In one case, the U-Net was trained on the original radiographs; in the second, it was trained on the sameradio graphs, but with the artifacts de-shadowed. The Dice Score results of the highest performing model, compared with the existing state of the art by Machado [7], are shown in Table 2. The U-Net, with the deshadowing enhancements, achieved the highest performance on the publicly available Abdi dataset. Note, however, that Machado's models were trained and tested on a dataset consisting of a combination of the Abdi dataset and of privately annotated data. Therefore, the results in this table cannot be seen as an exact comparison. Table 2. A comparison of Dice Scores for panoramic radiograph mandible segmentation.

[0157] Mandibular Canal Segmentation Results

[0158] Table 3. A comparison of Dice Scores for panoramic radiograph mandibular canal segmentation

[0159] Conclusion

[0160] The experimental results demonstrated the effectiveness of the disclosed two-step artifact-removal framework, which combines semantic segmentation of artifact regions with a transformer-based de-shadowing process. The approach provides measurable improvements in image quality and segmentation performance across multiple types of panoramic- radiography artifacts, including, for example, the air gap, the spine, and the ghost jaw. REFRENCES

[0161] [1] E. Land, "The retinex theory of color vision," Scientific American, vol. 237, no. 6, pp. 108-128, 12 1977.

[0162] [2] P. Yushkevich, J. Piven, C. Hazlett, et al., "User-guided 3D active contour segmentation of anatomical struc- tures: Significantly improved efficiency and reliability,"

[0163] Neuroimage, vol. 31, no. 3, pp. 1116-1128, 2006.

[0164] [3] M. Cardoso, W. Li, R. Brown, et al., "MONAI: An open-source framework for deep learning in healthcare," Nov. 2022.

[0165] [4] D. Kingma and J. Ba, "Adam: A method for stochas- tic optimization," International Conference on Learning Representations, 12 2014.

[0166] [5] L. Guo, S. Huang, D. Liu, et al., "Shadowformer: Global context helps shadow removal," Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, pp. 710-718,06 2023.

[0167] [6] D. Pel li and P. Bex, "Measuring contrast sensitivity," Vision research, vol. 90, pp. 10-14, September 2013.

[0168] [7] L. Machado, P. Watanabe, G. Rodrigues, et al., "Deep learning for automatic mandible segmentation on dental panoramic x-ray images," Biomedical Physics Engl- neering Express, vol. 9, no. 3, pp. 035015, 03 2023.

[0169] It is noted that while the examples provided in this description primarily pertain to dental and chest radiographs, these are intended solely as non-limiting examples. The principles disclosed herein can likewise be applied to images of other anatomical areas or in other applications where X-rays are used to capture non-human objects. For example, it can be applied in applications including other medical imaging tasks, orthopedic imaging, veterinary imaging, industry, and security, etc.

[0170] Those versed in the art will readily appreciate that the teachings of the presently disclosed subject matter are not bound by the system illustrated in FIGs. 2, 5, 7, 8 and 13. Different system components and modules in FIGs. 2, 5, 7, 8 and 13 can be made up of any combination of software, hardware, and / or firmware, as relevant, executed on a suitable device or devices, which perform the functions as defined and explained herein. Equivalent and / or modified functionality, as described with respect to each system component and module, can be consolidated or divided in another manner. Thus, in some embodiments of the presently disclosed subject matter, the system may include fewer, more, modified, and / or different components, modules, and functions than those shown in FIGs. 2, 5, 7, 8 and 13.

[0171] Each component in FIGs. 2, 5, 7, 8 and 13 may represent a plurality of the particular components, which are adapted to independently and / or cooperatively operate to process various data and electrical inputs. In some cases, multiple instances of a component may be utilized for reasons of performance, redundancy, and / or availability. Similarly, in some cases, multiple instances of a component may be utilized for reasons of functionality or application. For example, different portions of the particular functionality may be placed in different instances of the component.

[0172] While certain examples of the present disclosure refer to processing circuitry configured to perform the operations described above, these operations may be carried out in various ways by one or more processors included within the processing circuitry. For example, the operations associated with different functional modules may be executed by separate processors, or by a shared processor operating in a time-multiplexed manner. The present disclosure should not be construed as requiring a single processor to perform all operations. Furthermore, any reference to a single processing circuitry should be interpreted to optionally include multiple processing circuitries. The systems illustrated in FIGs. 2, 5, 7, 8 and 13 can be implemented in a distributed computing environment, in which one or more of the aforementioned components and functional modules shown in FIGs. 2, 5, 7, 8 and 13 can be distributed over several local and / or remote devices. In some examples, certain components of the system are implemented in a cloud-based architecture, which may include private or public cloud infrastructure. Communication between system components— when distributed across multiple physical locations or devices— may be facilitated using any suitable signaling mechanisms or communication technologies. These may include various combinations of hardware interfaces, software protocols, communication modules, programming languages, and transmission signals, and may operate over wired and / or wireless links, as appropriate.

[0173] Unless specifically stated otherwise, as apparent from the above discussions, it is appreciated that, throughout the specification, discussions utilizing terms such as "using", "receiving", "applying", "generating", "segmenting", "comparing", "making" or the like, include an action and / or processes of a computer that manipulate and / or transform data into other data, said data represented as physical quantities, e.g. such as electronic quantities, and / or said data representing the physical objects.

[0174] The terms "system", "computer", "computer system", "computer device", "computerized device" or the like that may be used herein, should be expansively construed to include any kind of hardware-based electronic device with one or more data processing circuitries. Each processing circuitry can comprise, for example, one or more processors operatively connected to computer memory, loaded with executable instructions for executing operations, as described above.

[0175] The one or more processors referred to herein can represent, for example, one or more general-purpose processing devices, such as a microprocessor, a central processing unit, or the like. More particularly, a given processor may be one of a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, a processor implementing other instruction sets, or a processor implementing a combination of instruction sets. The one or more processors may also be one or more special-purpose processing devices such as an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital signal processor (DSP), a graphics processing unit (GPU), a network processor, or the like. The memories referred to herein can comprise, for example, one or more of the following: internal memory, such as, e.g., processor registers and cache, etc., main memory such as, e.g., read-only memory (ROM), flash memory, dynamic random-access memory (DRAM) such as synchronous DRAM (SDRAM) or Rambus DRAM (RDRAM), etc.

[0176] As used herein, the phrase "for example," "such as", "for instance" and variants thereof, describe non-limiting embodiments of the presently disclosed subject matter. Reference in the specification to "one case", "some cases", "other cases", or variants thereof, means that a particular feature, structure, or characteristic described in connection with the embodiment(s), is included in at least one embodiment of the presently disclosed subject matter. Thus, the appearance of the phrase "one case", "some cases", "other cases", or variants thereof, does not necessarily refer to the same embodiment(s).

[0177] It is appreciated that certain features of the presently disclosed subject matter, which are, for clarity, described in the context of separate embodiments, may also be provided in combination in a single embodiment. Conversely, various features of the presently disclosed subject matter, which are, for brevity, described in the context of a single embodiment, may also be provided separately, or in any suitable sub-combination.

[0178] In embodiments of the presently disclosed subject matter, fewer, more and / or different stages than those shown in FIGs. 3, 6 and 9 may be executed. In some examples of the presently disclosed subject matter, one or more stages illustrated in the figures may be executed in a different order, and / or one or more groups of stages may be executed simultaneously. For example, although FIG. 9 illustrates operations for generating artifact-containing DRRs followed by operations for generating artifact-attenuated DRRs, the order of these operations may be reversed, or the two types of DRRs can be created concurrently.

[0179] It will also be understood that the system according to the presently disclosed subject matter may be a suitably programmed computer. Likewise, the presently disclosed subject matter contemplates a computer program being readable by a computer for executing the method of the presently disclosed subject matter. The presently disclosed subject matter further contemplates a machine-readable non-transitory memory tangibly embodying a program of instructions executable by the machine for executing the method of the presently disclosed subject matter. It is to be understood that the presently disclosed subject matter is not limited in its application to the details set forth in the description contained herein or illustrated in the drawings. The presently disclosed subject matter is capable of other embodiments and of being practiced and carried out in various ways. Hence, it is to be understood that the phraseology and terminology employed herein are for the purpose of description and should not be regarded as limiting. As such, those skilled in the art will appreciate that the conception upon which this disclosure is based may readily be utilized as a basis for designing other structures, methods, and systems for carrying out the several purposes of the presently disclosed subject matter.

Claims

CLAIMS:

1. A computer-implemented method of attenuating overlapping-layer artifacts from a radiograph, the method comprising, using at least one processing circuitry for: receiving the radiograph comprising overlapping-layer artifacts resulting from superposition of anatomical and / or non-anatomical structures; segmenting, using a first machine-learning model, the radiograph to generate an artifact-segmentation mask identifying respective artifact types; applying the artifact-segmentation mask together with the radiograph to a pre-trained de-shadowing machine learning model to attenuate shadow-like regions within a mask, the model thereby generating corrected pixel data for the artifact-segmented regions; and outputting an artifact-attenuated radiograph generated by the model using the corrected pixel data.

2. The method of claim 1, wherein the overlapping-layer artifacts include radiopaque artifacts characterized by pixel values higher than surrounding tissue, and the method further comprises generating a negative version of the radiograph prior to applying the pre-trained de-shadowing machine learning model, thereby converting the radiopaque artifact regions into shadow-like regions compatible with the de-shadowing model.

3. The method of any one of claims 1 and 2, comprising: applying a second machine-learning model trained to segment an anatomical structure within the radiograph to generate an anatomical-segmentation mask; comparing the artifact-segmentation mask with the anatomical-segmentation mask to identify regions in which artifact regions spatially overlap the anatomical structure; masking only the overlapping regions identified by the comparison to generate an overlap mask; and providing the overlap mask together with the radiograph to the pre-trained de- shadowing machine learning model to enable attenuation of the artifact solely within the regions that overlap the anatomical structure, thereby improving visibility of the anatomical structure.

4. The method of any one of the preceding claims, further comprising: afterapplying the pre-trained de-shadowing machine learning model, applying a de-hazing deeplearning model to the output of the de-shadowing model to attenuate diffuse, haze-like degradations arising from scattered radiation or overlapping soft-tissue regions, thereby further improving contrast uniformity and visibility of anatomical structures.

5. The method of any one of claims 1 to 4, wherein the radiograph comprises a plurality of overlapping-layer artifact types, and segmenting the radiograph comprises applying, for each artifact type, a respective machine-learning model trained to segment that artifact type, thereby producing a corresponding set of artifact-specific segmentation masks.

6. The method of claim 5, wherein the radiograph is a panoramic radiograph, and the plurality of overlapping-layer artifact types comprise air-gap artifacts, ghost-jaw artifacts, and spine artifacts, each segmented using its respective machine-learning model.

7. The method of any one of claims 5 and 6, wherein applying the pre-trained de- shadowing machine learning model comprises performing a plurality of sequential passes, each pass receiving a different artifact-specific segmentation mask and attenuating the corresponding artifact type, such that the radiograph is progressively corrected across successive passes.

8. The method of any one of claims 3 to 7, wherein the second machine-learning model is an autoencoder-based machine-learning model.

9. The method of any one of the preceding claims wherein the radiograph is any one of a panoramic radiograph (e.g., dental X-ray) and standard projection radiographs (e.g., chest X-ray).

10. A computer-implemented method of attenuating overlapping-layer artifacts in an input radiograph, the method comprising, by processing circuitry: receiving the input radiograph; providing the input radiograph to a machine-learning model trained to generate artifact-attenuated radiographs; and obtaining, from the machine-learning model, an artifact-attenuated version of the input radiograph; wherein the machine-learning model has been trained using pairs of digitally reconstructed radiographs (DRRs) generated from three-dimensional computerized-tomography (CT) volumes, each pair comprising an artifact-containing DRR and a corresponding artifact-attenuated DRR obtained by performing radiographic projection of the CT volume under a first projection configuration in which contributions of one or more artifact-causing structures in the DRR are preserved and under a second projection configuration in which contributions of the artifact-causing structures in the DRR are reduced or removed.

11. The method of claim 10, wherein generating the artifact-attenuated DRR comprises segmenting one or more artifact-causing structures within the CT volume to obtain segmented regions, and modifyingthe CT volume within the segmented regions so as to reduce or remove contributions of the artifact-causing structures in the radiographic projection.

12. The method of claim 11, wherein modifying the CT volume in the segmented regions comprises inpainting the segmented regions using voxel values representative of adjacent tissue.

13. The method of claim 10, wherein generating the artifact-attenuated DRR comprises performing radiographic projection of the CT volume using alternative virtual source locations, including: a first source location outside the CT volume such that corresponding sampling rays intersect one or more artifact-causing structures and produce an artifact-containing DRR; and a second source location inside the CT volume such that corresponding sampling rays bypass said artifact-causing structures and produce an artifact-attenuated DRR.

14. The method of any one of claims 10 to 13, wherein the machine-learning model trained using the DRR pairs comprises a neural-network architecture selected from an autoencoder, a convolutional neural network, a diffusion-based neural network, or a transformer-based neural network.

15. The method of any one of claims 10 to 14, wherein the machine-learning model trained using the paired DRRs comprises a de-shadowing model originally trained to remove shadows in optical images and further trained on the paired DRRs to learn attenuation of artifact-causing structures in radiographic images.

16. The method of any one of claim 10 to 15, wherein the artifact-causing structures segmented in the CT volume comprise one or more of: bones, air cavities, or soft -tissue regionsthat produce overlapping-layer artifacts in radiographic projection.

17. The method of any one of claims 10 to 16, wherein the method further comprises training the machine-learning model using the training dataset generated from the paired digitally reconstructed radiographs.

18. The method of any one of claims 10 to 17, wherein the method further comprises generating the training dataset by forming, for each three-dimensional computerized-tomography (CT) volume, a pair of digitally reconstructed radiographs comprising an artifact-containing DRR and a corresponding artifact-attenuated DRR.

19. A computer-implemented method of training a machine-learning model to attenuate overlapping-layer artifacts in radiographs, the method comprising, by processing circuitry: obtaining a training dataset comprising pairs of digitally reconstructed radiographs (DRRs), each pair including an artifact-containing DRR and a corresponding artifact-attenuated DRR; and training the machine-learning model using the artifact-containing DRRs as inputs and the artifact-attenuated DRRs as target outputs; wherein the training dataset is generated by performing radiographic projection of a three-dimensional computerized-tomography (CT) volume under a first projection configuration in which contributions of one or more artifact-causing structures in the DRR are preserved, and under a second projection configuration in which contributions of the artifact - causing structures in the DRR are reduced or removed.

20. A computer-implemented method of generating a training dataset for training a machine-learning model dedicated to attenuating overlapping-layer artifacts in radiographs, the method comprising, by processing circuitry: obtaining a three-dimensional computerized-tomography (CT) volume of an anatomical region; generating, from the CT volume, an artifact-containing digitally reconstructed radiograph (DRR) by performing radiographic projection under a first projection configuration in which contributions of one or more artifact-causing structures are preserved; generating, from the same CT volume, a corresponding artifact-attenuated DRR byperforming radiographic projection under a second projection configuration in which contributions of the artifact-causing structures are reduced or removed; forming a paired training example by associating the artifact-containing DRR with the corresponding artifact-attenuated DRR; and storing the paired training example in the training dataset.

21. A computer-implemented method of attenuating overlapping-layer artifacts in a radiograph, the method comprising using at least one processing circuitry for: receiving the radiograph; applying a machine-learning model to the radiographic input image to generate an artifact-attenuated radiographic output; wherein the machine-learning model is a model trained using a training dataset comprising pairs of radiographs produced by a heterogeneous collection of two or more processing pipelines, each pair including a first radiograph containing overlapping-layer artifacts and a second radiograph in which the artifacts are attenuated by the respective processing pipeline, the processing pipelines differing in the processes used to generate the second radiograph of each pair.

22. The method of claim 21, wherein the heterogeneous collection of processing pipelines comprises at least two of: de-shadowing pipeline, a de-hazing pipeline, a sequential de-shadowing and de-hazing pipeline, and a CT-based digitally reconstructed radiograph (DRR) pipeline.

23. The method of claim 22, wherein the heterogeneous collection of processing pipelines comprises three or four of: de-shadowing pipeline, a de-hazing pipeline, a sequential de-shadowing and de-hazing pipeline, and a CT-based digitally reconstructed radiograph (DRR) pipeline.

24. The method of any one of claims 21 to 23, wherein generating the pairs of radiographs for the CT-based digitally reconstructed radiograph (DRR) processing pipeline comprises: generating, from a three-dimensional CT volume, an artifact-containing DRR by performing radiographic projection under a first projection configuration in whichcontributions of one or more artifact-causing structures are preserved; generating, from the same CT volume, a corresponding artifact-attenuated DRR by performing radiographic projection under a second projection configuration in which contributions of the artifact-causing structures are reduced or removed; and forming a paired training example by associating the artifact-containing DRR with the corresponding artifact-attenuated DRR, the artifact-containing DRR and the artifact-attenuated DRR forming the first and second radiographs of the pair, respectively.

25. The method of any one of claims 21 to 24, wherein the machine-learning model comprises a diffusion model configured to learn a mapping of the first radiograph of each pair to the second radiograph of each pair provided by the processing pipelines.

26. The method of any one of claims 21 to 25, wherein:(a) the de-shadowing processing pipeline comprises applying a de-shadowing machinelearning model to a radiograph of a region of interest to obtain a corresponding radiograph in which overlapping-layer artifacts are attenuated;(b) the de-hazing processing pipeline comprises applying a de-hazing machine-learning model to a radiograph of a region of interest to obtain a corresponding radiograph in which diffuse, haze-like artifacts are attenuated; and(c) the sequential de-shadowing and de-hazing processing pipeline comprises first applying a de-shadowing machine-learning model to a radiograph of a region of interest and subsequently applying a de-hazing machine-learning model to obtain a corresponding radiograph in which both overlapping-layer artifacts and diffuse, haze-like artifacts are attenuated.

27. A computer system comprising at least one processing circuitry configured to execute a method according to any one of claims 1 to 9.

28. The computer system of claim 27, wherein the computer system is integrated with or operatively connected to an X-ray device to enable the X-ray device to provide artifact- attenuated images.

29. A non-transitory computer-readable storage device tangibly embodying instructions executable by a computer to perform a method according to any one of claims 1 to 9.

30. A computer system comprising at least one processing circuitry configured to execute a method according to any one of claims 10 to 18.

31. The computer system of claim 30, wherein the computer system is integrated with or operatively connected to an X-ray device to enable the X-ray device to provide artifact- attenuated images.

32. A non-transitory computer-readable storage device tangibly embodying instructions executable by a computer to perform a method according to any one of claims 10 to 18.

33. A computer system comprising at least one processing circuitry configured to execute the method according to claim 19.

34. The computer system of claim 33, wherein the computer system is integrated with or operatively connected to an X-ray device to enable the X-ray device to provide artifact- attenuated images.

35. A non-transitory computer-readable storage device tangibly embodying instructions executable by a computer to perform the method according to claim 19.

36. A computer system comprising at least one processing circuitry configured to execute the method according to claim 20.

37. The computer system of claim 35, wherein the computer system is integrated with or operatively connected to an X-ray device to enable the X-ray device to provide artifact- attenuated images.

38. A non-transitory computer-readable storage device tangibly embodying instructions executable by a computer to perform the method according to claim 20.

39. A computer system comprising at least one processing circuitry configured to execute a method according to any one of claims 21 to 26.

40. The computer system of claim 39, wherein the computer system is integrated with or operatively connected to an X-ray device to enable the X-ray device to provide artifact- attenuated images.

41. A non-transitory computer-readable storage device tangibly embodying instructions executable by a computer to perform a method according to any one of claims 21to 26.

42. A computer-implemented method for attenuating artifacts in an input radiograph, the method comprising, using at least one processing circuitry for: receiving, by one or more processors, an input radiograph; applying, one or more machine-learning models selected from a collection of machinelearning models, each machine-learning model in the collection being trained using a training dataset comprising radiograph pairs produced by one or more processing pipelines, each radiograph pair including a first radiograph containing artifacts and a second radiograph in which the artifacts are attenuated by a respective processing pipeline, the processing pipelines differing from one another in the processes used to generate the second radiograph of each pair; and generating one or more artifact-attenuated radiographic outputs, each artifact- attenuated radiographic output corresponding to a respective machine-learning model applied to the input radiograph.

43. A computer system comprising at least one processing circuitry configured to execute a method according to claim 42.

44. The computer system of claim 43, wherein the computer system is integrated with or operatively connected to an X-ray device to enable the X-ray device to provide artifact- attenuated images.

45. A non-transitory computer-readable storage device tangibly embodying instructions executable by a computer to perform a method according to claim 42.

Images