Automated bolus ROI placement using 3D CT surveys

Abstract

A system and associated method for facilitating contrast-based tomography, the system may comprise an input interface for receiving specifications for the imaging operation of a tomography-type imaging device, and a 3D survey image volume acquired by the patient imaging device in a preparatory phase prior to the contrast-assisted imaging phase; a segmenter segment for a surveillance area m-ROI within the 3D survey image volume; the m-ROI is associated with a target anatomical feature, as may be specified in the specifications; and an output interface providing a reference position for the segmented m-ROI within the 3D survey volume to facilitate monitoring of the presence of contrast agent relative to the target anatomical feature.

Description

[Technical Field]

[0001] The present invention relates to a system for facilitating contrast-based tomography, a training system for training based on training data, a machine learning model for use in such a system, a training data generation system, related methods, imaging apparatus, computer program elements, and computer-readable media. [Background technology]

[0002] Two-dimensional (2D) projection survey images acquired prior to diagnostic CT (computed tomography) imaging play a crucial role in assisting clinical technicians ("users"), for example, in proper patient positioning or in achieving optimal image quality.

[0003] In addition to conventional 2D surveys (e.g., frontal and sagittal), it is now possible to acquire 3D (reconstructed) survey images (image volumes) at clinically acceptable doses. These offer several advantages over conventional 2D surveys and open up new possibilities for optimizing CT workflows that were previously impossible.

[0004] In CT imaging, it is common practice to perform a 2D survey scan covering the entire region of interest, including the expected "bolus" tracking location. The bolus is a volume of contrast agent material, such as iodine, that is administered. The contrast agent may be used to boost the image contrast of a weakly radioactive anatomical structure of interest, such as the cardiovascular system in cardiac imaging. The user then manually selects a z-position on such a 2D survey, and a corresponding axial image, also known as a "locator," is acquired and reconstructed for defining ("placement") a region of interest ("ROI") for bolus tracking or monitoring bolus arrival. The bolus travels with the blood flow and, at one point, accumulates in the anatomical structure of interest. Capturing an image with high contrast before the bolus washout is intended to trigger imaging at the appropriate moment.

[0005] Multiple locator scans may be required to locate the appropriate ROI for tracking. The ROI to be captured for bolus tracking is positioned within this locator slice, and the arrival of the contrast agent bolus is monitored through repeated acquisitions of this single axial slice. Subsequent diagnostic CTA (CT angiography) acquisitions are triggered when a contrast-induced density threshold is reached within this ROI.

[0006] The main drawback of conventional bolus tracking using 2D survey images, or bolus tracking in general, is that planning the bolus tracking during the scan procedure is complex and heavily dependent on the skill and experience of the clinical user. This process is prone to errors. For example, there are several manual interventions required by the user, such as manually selecting z-slices and placing ROIs on acquired locator slices. If there are human errors in the localization of anatomical structures and the placement of ROIs within the localizer image, the diagnostic CT scan may be initiated at a suboptimal time. This can subsequently lead to poor image quality and misdiagnosis, potentially resulting in unnecessary re-administration of contrast agents and re-imaging, with potential side effects for the patient and / or medical staff. Furthermore, the manual localization process is subject to intra- and inter-individual variability. Another drawback is the need to acquire specialized locator scans, which require extra time and dosage. [Overview of the project] [Problems that the invention aims to solve]

[0007] More efficient operation of imaging devices may be required. In particular, it may be necessary to ensure that precise timing is established for triggering imaging operations for such imaging devices. [Means for solving the problem]

[0008] The object of the present invention is achieved by the subject matter of the independent claims, and further embodiments are incorporated into the dependent claims. It should be noted that the embodiments described below of the present invention are equally applicable to training systems for training, training data generation systems, related methods, imaging devices, computer program elements, and computer-readable media.

[0009] According to a first aspect of the present invention, A system for facilitating contrast-based tomography, wherein the system, when in use, The specifications for the imaging operation of the tomography type imaging device, and an input interface for receiving a 3D survey image volume obtained by the imaging device of at least a portion of the patient during a preparation phase prior to the contrast-assisted imaging phase. A segmenter configured to segment the 3D survey image volume with respect to a surveillance region ("m-ROI"), wherein the m-ROI is associated with a target anatomical feature identified in the specification, To facilitate monitoring for the presence of contrast agent with respect to the target anatomical features, an output interface is provided that provides a reference position for segmented m-ROIs in the 3D survey volume. A system is provided that has the following features.

[0010] The specification may, for example, be part of an imaging protocol. The specification can identify the target anatomical feature that is the objective for imaging. Based on medical knowledge, the specification defines the types of anatomical landmarks or other auxiliary anatomical features (preferably parts of (blood) vessels) that function as m-ROIs as a function of the intended target anatomical feature. The segmenter translates this medical and anatomical knowledge into specific location definitions for (one or more) m-ROIs in a given 3D survey image / volume, enabling medical users to assist in this task.

[0011] The target anatomical feature may be related to an organ / anatomical structure, a part of an organ / anatomical structure, a group of organs / anatomical structures, a tissue type, etc. The target anatomical feature may be a blood vessel, or a part / part thereof. The m-ROI is generally different from other target anatomical features located upstream of the target anatomical feature with respect to blood flow. The m-ROI may be a specific part of a blood vessel, a section thereof, etc. The blood vessel may supply the target anatomical feature.

[0012] In one embodiment, the system includes a controller configured to instruct an imaging device to acquire a first set of projection data while a contrast agent propagates within the patient, and a reconstructor that reconstructs a cross-sectional tracker image in a plane passing through a reference position with a first image quality based on the projection data.

[0013] In the embodiment, the in-image monitoring unit is configured to perform such monitoring based on tracker images in an image neighborhood within the tracker image, the image neighborhood being based on a reference position. The image neighborhood is preferably generated by a facilitator, or otherwise can be sourced. The size of the image neighborhood may be predefined or adaptable.

[0014] Preferably, the facilitator generates the image neighborhood. For example, the image neighborhood is defined within a synthesized tracker image ("locator") synthesized from the survey image, and such an image neighborhood is then used by an in-image monitoring unit to monitor within the reconstructed tracker image. Thus, in the proposed system, the locator can be synthesized from the survey image. Therefore, there is no need to spend a dose on the locator. This results in a saving in medication for the patient and / or staff. The in-image monitoring unit may be part of the system.

[0015] In an embodiment, based on an in-image monitoring unit in which a controller of a system issues a trigger signal at a second image quality higher than a first image quality based on one or more image values monitored within the vicinity of an image, the imaging device is configured to be commanded to acquire a second set of projection data.

[0016] In an embodiment, the reference position is based on at least one reference point related to the segmentation of the m-ROI.

[0017] In an embodiment, the at least one reference point includes the centroid of the segmentation of the m-ROI.

[0018] In an embodiment, the system includes an adjuster for adjusting any one of the position, orientation, and / or spatial extent in the vicinity of the image. The adjustment can be performed according to a user request, and the user can control the adjustment through a user interface ("UI").

[0019] In an embodiment, the adjuster is operable in response to a command receivable via a user interface configured to enable the user to request adjustment of any one or more of i) the vicinity within the image, ii) the reference position, iii) the reference point, and iv) the segmentation of the m-ROI.

[0020] In an embodiment, the adjustment by the adjuster is based on a waiting time related to the imaging stage. The waiting time may be related to the machine-side waiting time, that is, how quickly the imaging device can start acquisition, the delay in reconstruction from projection data, etc. Also, the latency may be related to the patient-side aspect of how quickly a bolus propagates through the patient's vascular system, such as bolus hemodynamics.

[0021] In an embodiment, the segmented m-ROI has a spatially identical extent to the geometric aspect of the target anatomical feature.

[0022] In this embodiment, the segmenter is based on a trained machine learning model. The facilitator may, if necessary, be used in deployment to testing after training.

[0023] In this embodiment, the model is of the artificial neural network type.

[0024] In this embodiment, the model is configured for multiscale processing of image information. This enables better processing of low-resolution (high-noise) images, such as in the case of low-dose survey images.

[0025] The proposed system uses 3D survey image volumes as input (tomography reconstruction) for processing, in contrast to mere 2D projected survey images. This 3D information is preferably used to enable the system, if not entirely automatically, to select a reference position (e.g., z-position) and define the proximity of m-ROI bolus on axial (or any other view) slices synthesized from the 3D survey image at the corresponding z-position. The neighborhood defines the location, preferably spatial extent, of the explored m-ROI. Therefore, z-slice selection and m-ROI neighborhood placement can be performed in one-to-one steps using 3D bolus m-ROI segmentation. Thus, m-ROI definition and its representation in tracker images are achieved solely through image segmentation. This significantly simplifies the user's workflow, allowing for safer and faster imaging of more patients.

[0026] Preferably, the user still has the option to adjust / modify the automatically suggested reference position (z position), and therefore the slice selection. The user can move or adapt the size of the automatically placed bolus m-ROI neighborhood.

[0027] User-requested adjustments may be embedded in a dynamic feedback mechanism. Therefore, when a user requests adjustment of any of the m-ROI components, such as the z-position, the corresponding axial slice selection is automatically updated, and the bolus ROI neighborhood on the new / updated slice is adapted to the neighborhood positioned on this new slice. The system's ability to position m-ROIs for various locations via 3D bolus segmentation, and the user's given options for adapting the relevant slice selection and neighborhood placement at both stages, enables a highly efficient clinical workflow.

[0028] Furthermore, in the proposed system and method (see below), image registration is not required because, in the embodiment, image segmentation is performed end-to-end by a machine learning model. Thus, the machine learning model maps the input (survey image) to segmentation within the same survey image. A single model can be used to output the z position, which automatically derives the definition of the neighborhood of the bolus m-ROI in the synthesized tracker image for guided image acquisition. Any geometric shape deemed appropriate can be used to define the position / degree of the m-ROI, e.g., a circle, ellipse, square, or any other shape that can be selected for the neighborhood.

[0029] The user can intervene in either of these two steps and adapt the workflow as needed. The proposed system closely follows the CT imaging workflow but is configured to follow it in a more efficient way. Thus, the user is still shown automated results for both z-slice selection on the coronal view and bolus tracking ROIs on axial slices. Both outputs, i.e., the z-position and neighborhood of the m-ROI, are found by the same segmentation of the 3D survey volume, and the segmentation map is post-processed to obtain reference points, on which the selected z-slice forms the survey image.

[0030] This system can be used to generate z-slice selections on the coronal plane of the survey image, while bolus tracking m-ROIs are represented on the axial tracker image (slice), very similar to current CT workflow implementations. However, any combination of such views is possible.

[0031] This model is preferably trained on labeled ground truth data because it facilitates a single-step operation of (i) the z position in the survey image and (ii) the neighborhood arrangement of the bolus m-ROI on the composite locator slice. The neighborhood can be defined in image space, in transformed space such as in the distance transformation of the composite tracker image, or in a follow-up tracker image reconstructed from newly acquired projection images.

[0032] Other advantages and benefits of the proposed system and method (see below) include greater consistency in performance across clinical user cases. For example, the spatial distance from the bolus m-ROI to the target organ in a CT angiography or other imaging session can vary considerably from patient to patient. By using a 3D survey as proposed herein, it is possible to consistently position the m-ROI at a predetermined distance upstream of the target structure to be imaged. By making latency information readily available from the examination or vendor specifications, the scan length and timing can be optimized to be as short as possible by positioning the m-ROI closer to the target anatomical structure and consistently across the patient cohort. Such latency information may include the effect duration regarding how long it takes to start the scanner (to perform the acquisition operation) and how long it takes for the contrast agent to reach the target anatomical structure. The latter can be estimated from medical knowledge or patient characteristics (e.g., height, BMI), mean blood flow data, etc.

[0033] The system and related methods (see below) are primarily intended herein for use in the medical field. However, the principles described herein may also be used in fields other than medical, such as contrast-aided inspection (e.g., dyes) of inaccessible hydraulic or piping systems, or hydraulic inspections to understand groundwater movement. As in medical imaging, the system and methods can be used at a better timing to trigger acquisition at the moment of best image acquisition, thus avoiding repeated imaging and reducing the effects of wear and tear on the imaging device.

[0034] In another embodiment, a training system is provided for training the machine learning model based on training data.

[0035] In another embodiment, a training data generation system is provided that is configured to generate training data based on which training system is used to train a machine learning model.

[0036] In one embodiment, the training data generation system is configured to facilitate the generation of training data by including annotations in the preceding images, where the annotations indicate the location and spatial extent of the sample m-ROI within the preceding image.

[0037] In another embodiment, A computer implementation method for facilitating contrast-enhanced tomography, The specifications for the imaging operation of the tomography type imaging device, and the step of receiving a 3D survey image volume obtained by the imaging device of at least a portion of the patient in a preparation stage prior to the imaging stage assisted by a contrast agent, The 3D survey image volume is segmented with respect to a surveillance region m-ROI, wherein the m-ROI is associated with a target anatomical feature identified in the specification. A computer implementation method having the above is provided. Target anatomical features can be identified, for example, herein.

[0038] This method involves providing a reference position for segmented m-ROIs in the 3D survey volume in order to facilitate monitoring the presence of contrast agent with respect to the target anatomical features. It may also include the following.

[0039] In another embodiment, a computer implementation method is provided for training the machine learning model based on training data. This method may include receiving training data and adapting the model parameters based on the training data in order to obtain a trained machine learning model. The training may be supervised, self-supervised, or unsupervised.

[0040] In another embodiment, a computer implementation method is provided for generating training data on which a training method operates in order to train a machine learning model. This method may include annotating data to obtain the training data. Such annotated data may include existing historical survey images. Alternatively, or in addition, noisy images may be obtained as training (input) data by artificially applying noise to existing volumetric image data, thereby generating a sample of survey volume that can serve as training input. Annotations may be made to the existing volumetric image data, which are generally of high quality / low noise, if any. In some embodiments, the generation of the training data itself may be machine learning-based. For example, a generative model such as a GAN (Generative Adversarial Network) may be used to generate a sample of the training data.

[0041] In one embodiment, a computer implementation method for generating training data may include receiving definitions of seed markers for existing survey images and generating a convex hull containing all such seed markers based on the seed markers, where the voxels enclosed by the convex hull represent annotations for instances of the survey region.

[0042] The seed markers can be geometrically configured as circles, ellipses, or any shape that can be made to conform at least substantially to at least a partial and local contour of the segmentation.

[0043] In another embodiment, an imaging apparatus is provided comprising the system described in any one of the above claims and one or more contrast-enhancing devices and contrast-administering devices for administering a contrast agent.

[0044] Any of the above systems may be provided on one or more appropriately configured computing systems.

[0045] In another embodiment, a computer program element is provided which, when executed by at least one processing unit, is adapted to cause the processing unit to perform any given one of the methods described above.

[0046] In yet another embodiment, at least one computer-readable medium is provided that stores program elements or machine learning models.

[0047] The term "user" refers to a person who operates the imaging device or monitors the imaging procedure, such as a healthcare professional or another person. In other words, the user is generally not the patient.

[0048] "Survey image" as used herein refers to a 3D image volume obtained with a larger field of view (FOV) at a lower dose cost and radiation dose cost than the FOV of a subsequent diagnostic image volume. Diagnostic image volumes are required to accomplish or assist medical tasks such as diagnosis, therapy, and planning. Survey images are not used for such tasks. The FOV of a survey image volume not only includes the anatomical structures of interest that are the very object of the medical task, but also captures other landmark anatomical structures that generally do not serve the medical task, although they generally only have an auxiliary function for imaging. The FOV may be a whole-body scan, but this is not necessary in all cases. Generally, the FOV of a survey image captures larger anatomical sections such as the abdomen and rib cage.

[0049] As used herein, “z position” or similar terms refer to an example of a reference position in the image region assumed herein for the definition of a tracker image. “z position” generally refers to a position on one of three spatial coordinate axes (the “Z-axis”) across the image region. Generally, this Z-axis corresponds to the rotation axis of a rotational tomography system, or a virtual rotation axis in the imaging device of a fifth-generation scanner. However, the Z-axis may be different from the rotation axis for a reformatted survey volume, as assumed herein.

[0050] Generally, the term “machine learning” includes computerized devices (or modules) that implement machine learning ("ML") algorithms. Some such ML algorithms work to tune machine learning models configured to perform tasks ("learn"). Other ML works directly on training data without necessarily using models or the like. This tuning or updating of the training data corpus is called “training.” Generally, the task performance of an ML module can be measurably improved using training experience. Training experience may include exposure of appropriate training data and models to such data. Task performance can be improved so that the data better represents the task to be learned. “Training experience helps improve performance if the training data better represents the distribution of examples on which the final system performance will be measured.” Performance may be objectively test-measured based on the output produced by the module in response to being fed test data. Performance may be defined in relation to a specific error rate that should be achieved on given test data. For example, see TM Mitchell, "Machine Learning", page 2, section 1.1, page 6 section 1.2.1, McGraw-Hill, 1997. Some machine learning approaches are based on various forms of backpropagation algorithms, and some are based on gradient descent.

[0051] "Image processing" refers to a type, method, or process relating to information captured by an image. However, in this regard, any image processing assumed and described herein in any embodiment and configuration or setting includes not only processing in the image domain, such as image values, but also, in particular, processing in the frequency domain, or any other domain to which the image V may first be transformed, and it is observed that at least a portion of the processing takes place there, with, if necessary, an optional re-transformation to the image / class map domain. Transformations assumed herein include any Fourier-based transforms (such as Laplace transforms and discrete cosine transforms), wavelet transforms (such as Haar transforms), Hilbert transforms, as well as distance transforms and the like.

[0052] Herein, exemplary embodiments of the present invention will be described with reference to the following drawings, which are not to a fixed scale unless otherwise specified. [Brief explanation of the drawing]

[0053] [Figure 1] A schematic block diagram of a medical imaging device is shown. [Figure 2] This document describes an imaging protocol based on contrast agents. [Figure 3] This document outlines the steps of a facilitator system for facilitating contrast-assisted tomography imaging. [Figure 4] A more detailed block diagram of the facilitator system is shown. [Figure 5] This shows the images that can be generated by the facilitator system. [Figure 6] This document presents machine learning models that may be used in embodiments of a facilitator system segmenter. [Figure 7] This document describes a training system for training machine learning models. [Figure 8] This shows a flowchart of a computer implementation method to facilitate contrast-assisted tomography. [Figure 9] This shows a flowchart of a computer implementation method for training machine learning models. [Modes for carrying out the invention]

[0054] First, referring to Figure 1, Figure 1 shows a schematic block diagram of a medical imaging apparatus (MAR) assumed herein in an embodiment. The configuration IAR may include a medical imaging apparatus (IA) (abbreviated as "imager"), preferably of a tomography-based type, and therefore the imager may be a computed tomography (CT) scanner, but intervention systems such as C-arm / U-arm types are not excluded herein. Other tomography modalities such as MRI, PET, and others are not further excluded herein in some other embodiments. The medical imaging apparatus is preferably assumed for contrast-based imaging protocols, such as angiography.

[0055] The IAR device further includes a computing system CS that is widely operable to process data, including image data, as provided by the imager. The computing system can further enable control over the operation of the imaging device. The computing system may be located remotely from the imaging device IA, or it may be located close to it, such as integrated into the operator console CS from which a user can operate the imaging device, and in particular control the imaging operation to acquire medical images for diagnosis, treatment, planning (such as radiotherapy), etc.

[0056] More broadly, and as will be described in more detail below, the computing system CS may include a facilitator system FS that can be used to establish the correct timing for triggering the imaging operation in order to ensure that good quality (contrast) images are obtained.

[0057] The computing system CS, and in particular its facilitator system FS, may be implemented as a cloud solution running on one or more servers. The imaging device IA may be installed in a clinical facility such as a hospital. The computing system may be installed or used in a control room adjacent to the imaging room where the imaging device IA is located. In some embodiments, the computing system may be integrated with the imager IA. The imager IA may be communicably coupled to the computing system CS via a wired or wireless (or partially both) communication channel CC. The computing system CS may be configured as a fixed computing system such as a desktop PC, as the server, or as a mobile device such as a laptop, smartphone, or tablet.

[0058] Before describing the operation of the facilitator system FS in more detail, refer to the first mode of the components of the imager IA, which will be described below in relation to the facilitator system FS.

[0059] The imaging device IA is capable of generating, and specifically acquiring, projection data λ, which is transmitted to the computing system CS via a communication channel for reconstruction into tomographic (cross-sectional) images. The computing system CS runs one or more reconstruction units RECON, which implement one or more reconstruction algorithms. Generally, the reconstruction algorithm implements mapping, which maps projection data λ located in the projection region to the image region. The image region is part of 3D space and is located in the inspection region ER of the imaging device, while the projection region is located in 2D and is located in the (X-ray) detector XD of the imaging device IA.

[0060] As described above, the imaging device IA is preferably of the tomographic type and preferably configured for acquiring multi-directional projection images. Therefore, the imaging device IA can acquire projection images λ along different projection directions α with respect to the examination area ER, and thus can acquire the anatomical region of interest ("ROI") of the patient. Acquisition is performed, in embodiments, by a rotating system in which at least the X-ray source XS is positioned within a movable gantry MG.

[0061] The movable gantry (and, in embodiments, the X-ray source XS) is rotatable within the stationary gantry SG around the examination area ER where the patient / ROI is located during imaging. Opposite the X-ray source in the movable gantry is an X-ray detector XD which can rotate with the gantry and X-ray source around the examination area ER to achieve different projection directions α.

[0062] As schematically shown in Figure 1, the patient's longitudinal axis or image axis Z may extend within the examination area ER in the image. The patient PAT can lie on a patient support PS, such as a bed, which is at least partially positioned within the examination area ER during imaging. In some, but not all, embodiments, a helical imaging protocol is assumed herein in which there is relative lateral motion along the longitudinal axis Z between the X-ray source XS and the patient PAT. For example, the patient support PS can be advanced through the examination area ER during multi-directional projection image acquisition, for example, during the rotation of the X-ray source XS around the patient.

[0063] The CT scanner setup shown in Figure 1 is merely one embodiment, and other tomography devices such as C-arm or U-arm scanners, cone-beam CT configurations, and mammography imaging devices are not excluded herein. In some embodiments, C-arm cone-beam imagers are preferred herein. Furthermore, multi-directional acquisition capability may not necessarily be obtained from a rotating system such as the one shown in Figure 1. Non-rotating imaging systems, such as fourth or fifth-generation CT scanners, are also conceivable, where, for example, multiple X-ray sources are arranged around the examination area in a source ring. In addition, or instead, detectors XD may be arranged as a detector ring around the examination area. Thus, in such a system, there is no rotation of the X-ray sources XS or detectors XD, or both.

[0064] An operator console (OC) may be provided for users, such as medical professionals, to control the imaging operation. For example, the user may request the start of image acquisition, or request reconstruction or other operations, or initiate the transmission of data to the computing system (CS), or stop such transmission as needed.

[0065] During imaging, the X-ray beam XB is emitted from the focal point of the X-ray source XS along different projection directions α. The beam XB passes through the examination area with the patient in it. The X-rays interact with the patient's tissues. The X-ray beam XB is modified as a result of this interaction. Generally, such modification of the X-ray beam XB involves attenuation and scattering of the original incident X-ray beam. The modified X-rays are then detected as a spatial distribution of varying intensities in the X-ray sensitive pixels of the detector XD.

[0066] It should be noted that, in this context, it is not essential to acquire a projected image λ over the entire 360-degree angular region around the inspection area ER. Acquisition over a partial angular range, such as 270 degrees, 180 degrees, or less, may suffice. The X-ray detector is preferably configured to acquire a 2D projected image having rows and columns of intensity values ​​registered by the detector pixels. That is, the detector pixels themselves may be arranged in a matrix layout. Such a 2D layout can be used with divergent image geometries such as cones or fan beams. However, one-dimensional detector pixel layouts (such as along a single line) are not excluded herein, and neither is a parallel beam geometry.

[0067] The reconstructor RECON implements one or more reconstruction algorithms for processing projection images. Specifically, the reconstructor RECON can compute a cross-sectional image V of an examination area (in which the patient is located) for diagnostic, therapeutic, or other purposes. The reconstructor RECON may be capable of generating cross-sectional volumetric image data ("image volume") V. However, this does not preclude generating a single image slice within the examination area as needed. Thus, the reconstructed image may be expressed herein as V, which may include the total volume, a partial volume, or a specific portion through which it passes. Volumetric reconstruction can be facilitated by the helical motion and / or 2D layout of the X-ray detector XD.

[0068] The reconstructed volumetric image V may be stored in memory MEM or processed in other ways, as needed. The reconstructed volumetric image V can be visualized by a visualizer VIZ. The visualizer VIZ can generate a graphic representation of the volume or a given slice. The graphic display may be displayed on a display device DD. The visualizer VIZ can map the image volume V or a portion of the image volume V as a whole to a grayscale value or a color palette. The visualizer VIZ can control the video circuitry via a suitable interface to display the graphic display on the display device DD. In addition to, or instead of, such a display, the reconstructed image V may be stored in memory for later review or for other types of processing. Such memory may include an image repository such as a database (e.g., PACS) or other (preferably) non-volatile data storage device.

[0069] The reconstructed volume image V can be manipulated, for example, by reformatting to define a cross-sectional plane different from the cross-section defined by the imaging geometry. Such reformatting may allow medical users to better identify tissue types or anatomical details within a patient's body, depending on their medical purpose at hand, such as for diagnosis or preparation for certain treatments.

[0070] The reconstructed volume V may, in this specification, be referred to as the target image acquired during the target or operational phase of a contrast-assisted imaging procedure. This reconstructed image V is intended as a diagnostic image representing the target anatomical feature TAF, such as a target anatomical structure, organ, part of an organ, or group of organs, or a different tissue type, and is intended to have sufficient contrast to safely inform treatment or diagnostic decisions, or other medical decisions, such as taking the image into consideration, using other imaging modalities, or other tasks (such as tests), or performing further imaging sessions.

[0071] The (target) projection image λ from which the target volume V is reconstructed is acquired with a sufficient dose by appropriately controlling the dose used via the operator console OC. This can be done by controlling the voltage and / or amperage settings of the X-ray source XS tube to ensure a certain diagnostic image quality. As previously stated, the imaging protocols assumed herein are preferably contrast-based to ensure that target anatomical features TAF, which may be inherently radiopaque, can still be imaged with sufficient contrast. Such target projection images acquired with sufficiently high doses may also be referred to herein as diagnostic projection images λ in order to follow established terminology. However, this nomenclature does not preclude the use of this projection data λ and its reconstruction V for non-diagnostic tasks, such as for therapy (in Cath Lab, for example), planning, or any other task.

[0072] The target anatomical feature (TAF) is relevant to the medical purpose of the image. Therefore, if a patient's liver needs to be examined, the object anatomical feature TAF is the liver, and it is the liver that is the object and target of the abdominal scan to be performed.

[0073] Each target anatomical feature (TAF) is generally associated with a set of standards that define it as a function of the imaging protocol, preferably the patient's biometric characteristics (age, weight, sex, height, BMI, medical records, etc.), specific preferred imaging settings, the required image contrast to be achieved, the radiation dose to be used for the target anatomical feature TAF for a given purpose, the voltage / ampere setting of the source X used, collimation, etc. In short, the imaging protocol encapsulates the medical knowledge for any given imaging task, purpose, target anatomical feature TAF, etc.

[0074] Before acquiring such a diagnostic projection image λ, and for such low radiopaque anatomical feature TAFs, a volume of contrast agent CA, such as iodine or other suitable material, is administered to the patient PAT via an administration device ADA (e.g., a pump). This volume of contrast agent CA (sometimes called a "bolus") then propagates through the patient via the bloodstream and accumulates in the target anatomical feature TAF. Ideally, once sufficient contrast agent CA has accumulated in the target anatomical feature TAF, a diagnostic projection image of diagnostic quality should be acquired. Therefore, the timing of acquisition of the diagnostic projection image λ relative to the target image V by the imager IA is a critical consideration, as it must be ensured that the concentration of contrast agent in the target anatomical feature of interest is sufficient. Only then can the reconstructible image region target volume V be expected to have the required IQ (image quality) according to the protocol or otherwise specified image contrast. Otherwise, re-imaging may be required, which should be avoided due to cost, time and effort, increased dose exposure, and machine wear (especially of the anode disk of the tube XS). Therefore, in this specification, it is considered important to obtain the diagnosis at the "correct time".

[0075] The facilitator system FS envisioned herein facilitates the acquisition of such images at the correct timing in contrast-assisted imaging protocols. The facilitator FS ensures that diagnostic projection image acquisition is reproducibly and reliably triggered at the precise moment when sufficient contrast agent has actually accumulated in the target anatomical feature (TAF).

[0076] The facilitator system operates across two phases: the exploration or preparation phase PP and the monitoring phase, both of which precede the target phase in which the target volume V is obtained. Prior to all such stages, there is an initial stage in which an initial survey image V0 is obtained. This initial image V0 is preferably a 3D (image region) survey volume reconstructed from a first set of projection images. However, this first / initial set of projection images λ0 is acquired later in the target phase and is acquired at a lower quality than the projection image λ used for reconstructing the target image V. In particular, acquiring λ0 results in a lower radiation dose compared to the later dose which would be reduced for the diagnostic scan to acquire the diagnostic projection image λ. This is to conserve dose on the patient PAT, and since the 3D survey volume V0 serves a completely different purpose than the target volume V, the purpose of the survey volume / image V0 is essentially one of navigation, and as will be elaborated herein, the survey image V0 is used to find the appropriate position for monitoring bolus arrival and to ensure that imaging of the target anatomical feature TAF is initiated in a timely manner, so that the target volume V has the expected diagnostic grade contrast.

[0077] The contrast-enhanced imaging protocol is schematically shown in Figure 2, which is referenced here before a more detailed explanation of the operation of the facilitator system FS.

[0078] As previously observed, the contrast agent CA enhances image contrast for target structures (TAFs) that have low natural radiopaqueness. In this regard, referring more specifically to Figure 2, this is a schematic diagram of a portion of a blood vessel at an access point (indicated by "X") into which a bolus CA is administered by the contrast agent delivery device ADA or otherwise.

[0079] In cardiac imaging, blood vessels of interest, such as arteries or veins, are soft tissue and therefore have poor radiopaqueness. Consequently, when using non-contrast scans, they are represented by low contrast. The volume of contrast agent CA moves with the blood flow and propagates through the target feature TAF until a concentration of contrast agent accumulates there, allowing the target phase to begin and a higher quality projection image for the target volume V to be obtained.

[0080] Preferably, the upstream of the target anatomical feature TAF (where the blood flow direction is a vector in Figure 2). In the example shown in TIFF2026510075000002.tif23158, the monitoring region m-ROI is defined by an in-image neighborhood U, which will be described in more detail later. This neighborhood U is based on 3D segmentation s, as seen in the initial survey image V0. The neighborhood U and segmentation may relate to anatomical landmarks where contrast agent concentration is monitored. That is, the monitoring region m-ROI may initially be defined in relation to anatomical landmarks according to medical knowledge, but the precise location within a given volume is provided by the facilitator FS.

[0081] Therefore, the concentration of the contrast agent CA is preferably monitored upstream of the actual target anatomical feature TAF, for example, to account for image acquisition latency. As the concentration monitored in the monitoring region m-ROI is measurable and specific by thresholding in the time series (t) of the tracker image r(t) (explored more thoroughly below), once a certain minimum concentration is reached, acquisition of the current diagnostic projection data λ of the target volume V can begin by that time thanks to blood flow, and it can be expected that the concentration at the target anatomical feature TAF has reached an acceptable minimum concentration.

[0082] The spatial distance between the monitored region (m-ROI) and the actual target anatomical feature (TAF) is primarily based on clinical knowledge and / or determined by patient characteristics, blood flow velocity, etc. Such medical contextual information includes the m-ROI-TAF distance, and all landmarks used to monitor the region (m-ROI, etc.) can be encoded in the aforementioned imaging protocol / standard for the current imaging task / purpose.

[0083] Generally, the facilitator system FS is configured to quickly, reliably, and accurately find the correct area of ​​observation (m-ROI) for a given patient and target anatomical feature (TAF). Preferably, the operation of the facilitator system FS is designed to integrate well with existing CT workflows. The facilitator system FS is intended to operate fully automatically and provide the area of ​​observation (m-ROI) without any further user input (other than specifying the imaging protocol or target anatomical feature (TAF)), although in some embodiments such further user input is specifically assumed. In embodiments that invite such user interaction, it is specifically provided that certain components of the area of ​​observation discovery procedure can be specifically adjusted to the clinical user's comfort via an appropriate user interface UI, such as a graphical user interface (GUI). User requests for such adjustments can be passed to the system SYS via a touchscreen, pointer tool (computer mouse, stylus), etc. Such user interactions are preferably envisioned herein as dynamic real-time experiences, and when such a user requests adjustments related to the discovered monitoring area m-ROI, a recalculation of the relevant components is triggered, and their displays may be updated as many times as the user requests such changes or adjustments.

[0084] As will be explained in more detail below, the machine learning model M can be used when implementing the facilitator system FS. The facilitator system FS enables finding the correct moment t=t0 for the target phase to begin and controls the imaging instrument IA to ensure that a higher quality projection image λ can be obtained, which can then reconstruct the target volume V.

[0085] Figure 3 shows some basic operating modes of the proposed facilitator system FS. Specifically, as shown in a), the initial 3D survey volume V0 is segmented to suit one or more anatomical landmarks that function as bolus monitoring regions m-ROIs.

[0086] Based on the segmentation s, a reference position z0 within the volume is identified, which may be, for example, a set position along the rotation / imaging axis Z of the imaging device IA.

[0087] Next, from each set of low-dose projection images acquired at appropriate sampling intervals around the reference position z0, the time series r of the tracker image is obtained. t It will be reconfigured.

[0088] In some or each such tracker image r(t), the monitoring neighborhood U (shown as a small circle) in b) is automatically defined based on the reference position z0. In the user-interactive embodiment described above, an event handler can be used to allow the user to change z0 or U as needed. Thus, it will be understood that the monitoring region m—ROI—may be defined by two elements: its position in 3D, as given by the reference position z0, and a spatial extent for monitoring purposes, as represented in the tracker image rt, whose representation is defined by the neighborhood U.

[0089] In a series of tracker imagers (t), the concentration of contrast agent arrival is monitored after bolus administration. An example is the contrast curve c) for a given tracker image and the vicinity U of the image.U This is shown as (t). In such a curve, the contrast value HU (Haunsfield units) in the vicinity U is recorded over time t. The CA concentration, and therefore the contrast, is expected to increase over time in the ramp-up phase, reach a plateau in the plateau phase, reach saturation in the plateau phase, and then drop off in the drop-up phase as the contrast agent washes out. Only the ramp-up phase is shown in Figure 3.

[0090] Image value-based thresholding policies may be used in surveillance of the neighboring U to trigger an acquisition signal for obtaining a reconstructible diagnostic projection image of the target volume V. This acquisition of λ should be triggered after the contrast agent concentration (or HU value in the contrast curve c) reaches a certain minimum value which may be lower than the expected maximum value in the target anatomical feature TFA, as this is expected to be located slightly downstream of the surveillance region m—ROI / U. Instead of basing the thresholding described above on monitoring absolute HU values, the gradient of HU values ​​may be monitored instead, or in addition to them, as needed.

[0091] However, in some embodiments where latency is not an issue, the monitoring region and the target anatomical feature may coincide, in which case thresholding can be performed so that diagnostic projection image acquisition is triggered at the maximum density value in the plateau phase. The maximum value can be found, for example, by using a gradient-based method.

[0092] The set of projection images acquired at the set sampling rate from which the tracker image is reconstructed is of equal lower quality (lower dose) than the projection image λ for the later diagnostic acquisition phase. For example, the image quality may be the same as that of the projection image from which the survey image V0 is reconstructed.

[0093] The size of the survey (also called a "scout") image volume V0 (field of view ["FOV"]) is preferably selected to include not only the target feature TAF but also at least one of the anatomical landmarks on which bolus monitoring should be performed. Where in doubt, the survey scan may be a whole-body scan, but this is not always necessary, as in some cases a scan across applicable body regions such as an abdominal scan, head scan, thoracic scan, or leg scan may suffice.

[0094] Contrast is usually very poor in such survey images V0 due to low doses; nevertheless, it may be sufficient to broadly locate landmarks for m-ROIs and, optionally, targeted anatomical features (TAFs). The latter may also be manually marked up by the user (see below).

[0095] Once the reference position Z0 is discovered by the system FS based on segmentations, the imaging geometry may need to be adjusted, for example, by moving the patient table relative to the gantry, and as a result, tracker images around the reference position can be obtained by reconstruction from a series of projection images acquired at the aforementioned sampling intervals. However, in a preferred embodiment, such rearrangement of the table or gantry is not necessary, as the distance between the target anatomical feature and the surveillance position is known, and it is preferable that the survey volume V0 is acquired from the outset to cover a sufficiently large volume. For example, the FOV of the survey scout volume V0 may be linked to the autoplan, and therefore, in general, there is no need for a very large FOV survey to collect anatomical context information, for example, for landmark detection. Such an autoplan is a facility that enables the detection of target anatomical features TAF in the survey image V0 or enables the definition of the FOV to be used. The autoplan facility may be implemented by appropriate image processing, for example, by segmentation again if necessary. However, the detection and knowledge of the TAF location may also be provided by any means, whether manually by the user or based on machine learning. Generally, this specification assumes that the location of the target anatomical feature (TAF) is known, and the main focus of this disclosure is the reliable localization of the surveillance region (m-ROI).

[0096] As will be discussed in more detail below, the segmentation s is a 3D segmentation, i.e., defined by a 3D subvolume within the survey image, where s⊂V0. This subvolume, having spatial extensions in all three spatial directions (X, Y, Z), is preferably recognized by the anatomical structure as corresponding in shape, size, and orientation to the spatial structure of the anatomical structure at its location. For example, the segmentation may conform to the spatial structure of an anatomical landmark associated with the target anatomical structure. Therefore, the segmentation can, at least partially, follow the anatomical / tissue boundaries of the landmark. This allows the user to quickly verify at a glance whether the segmentation proposed by the FS system is medically meaningful.

[0097] Generally, relatively small segmented anatomical structures can be used to define a reference position z0, and separately, a neighborhood U in the tracker slice image passing through the reference position z0. As described above, the reference position z0 may be located on the imaging axis Z. However, this is not required herein, as the initial survey volume V0 may also be provided as a reformatting, and the reference position z0 may be a point on any geometric line for the reformatting, and therefore may be a point different from the rotation / imaging axis Z.

[0098] The facilitator system FS is preferably capable of operating primarily over two phases: the preparation phase PP and the target phase TP, which are described in more detail in the block diagram of Figure 4, which is referenced here.

[0099] The initial survey image V0, reconstructed from projection images acquired at a lower quality (lower dosage) λ0 than the dosage used for the diagnostic projection image λ for the later target volume V, is received at input port IN.

[0100] The segmenter component SEG is used to segment the volume V0 of locations, such as anatomical landmarks, to act as a surveillance area m—ROI—for the bolus to reach the target anatomical feature TAF. Preferably, since the segmentation operation is based on machine learning, it includes a machine learning model M pre-trained on training data, as will be described in more detail below. Optionally, the model may be further trained to segment the survey volume V0 about the target anatomical TAF as well, or the facilitator FS may include a different ML model trained to segment about the target anatomical TAF, as described above in relation to automated planning. However, as previously stated, the location of the target anatomical TAF is assumed to be generally known herein and can be specified by either means, manually, or by ML or any other method.

[0101] The segmentation (result), i.e., subvolumes s, are output via the (internal) output port OUT and are then available for subsequent monitoring phases, but may also be used for other purposes. For example, it may be desirable to visualize the segmentation within the survey volume immediately or later for cross-checking purposes. This may cause the visualizer VIZ to be instructed by the system to render a composite visualization of volume V0 (selectable views, axial, sagittal, etc.) with the segmentation appropriately marked, using color or grayscale coding of the volume and its boundaries, as an overlay, or by any other visualization technique. To visualize such segmentation, any suitable view of volume Vo may be rendered sagittally, coronally, or axially, as needed. Alternatively, true 3D rendering may be optionally rotated downwards (e.g., automatically or during user interaction with a pointer tool), allowing the user to quickly visually inspect the segmentation from various directions and verify how it fits into the surrounding anatomical structure. As mentioned earlier, at this point, due to the low dose spent on survey V0, any contrast, if any, is only present. As previously observed, the shape, orientation, etc., of the segmentations are those of the surrounding anatomical structures, such as vascular sections.

[0102] Segmented subvolumes s representing the correct landmarks associated with the m-ROI can be processed in the output interface to generate a reference position Z0. For example, the reference position of the tracker image RT can be defined using the centroid or other geometric reference point P within or on the subvolumes s. For example, the z-coordinate of the geometric reference point can be the reference position z0 of the m-ROI.

[0103] The reference position z0 is optional, but preferably, it may first be used to synthesize a “locator” (or “prototype”) tracker image r0 from the survey volume V0 in a plane passing through the reference position z0. This prototype tracker image r0 may optionally be visualized by a visualizer VIZ and displayed on a display device in any desired view, such as axial, coronal, or sagittal. In this case as well, this provides useful cross-check feedback for the user to check whether the calculated segmentation meets the user’s expectations. The reference point may be visually marked up in the displayed locator tracker imager r0, and / or it is the neighborhood that is marked up in this way.

[0104] It should be noted that since the prototype tracker image r0 can be synthesized purely computationally from the survey volume V0, separate projection data acquisition is not required, thus saving on dosage. Specifically, the neighbor U for use in the “live” tracker image for live monitoring of bolus arrival can already be defined on this proto-tracker image synthesized from the already available survey 3D image V0. No dosage is required to determine the neighbor U that represents the spatial extension of the region of interest m—ROI—in a representable way in the tracker image. Furthermore, both the determination of the reference position and the neighbor U can be performed in a single step based only on the survey image V0 and the thus synthesized locator tracker image r0. This is more convenient and faster for clinical users, and this one-step operation can still be easily integrated into existing CT workflows such as those performed in healthcare facilities worldwide. It will be understood that the use of the locator tracker image r0 is independent of its display, and such display is not actually required in all embodiments herein for the purpose of neighbor U definition as described in the single-step setup.

[0105] The reference position z0 derived from segmentation may be used by the controller CL to instruct the imager to begin acquiring a series of low-dose tracker projection images λ” around the reference position Z0 at an appropriate (e.g., user-adjustable) sampling rate. For some or each set of such tracker projection images λ, the reconstructor RECON may be instructed via the interface CL to reconstruct a series of tracker images r(t) from the series of low-dose tracker projection images, and each tracker image may be sequentially displayed on the display device DD in the video feed as they are calculated. However, such display is optional, and monitoring operations for bolus arrival (described in more detail below) can be performed without display. The subsequent tracker images r(t) thus reconstructed are essential copies of the location-composite locator tracker image r0, but are updated versions of r0 with changes in image values ​​due to the gradual accumulation of contrast agent.

[0106] In this case as well, the image quality of this series of tracker projection images λ” acquired during the monitoring phase is of lower quality (lower dose) than the projection images λ acquired in the later diagnostic phase. For example, the image quality may be similar to one of the projection images λ0 for the initial survey volume V0, or the radiation dose may be adjusted accordingly, triggering acquisition, and still being lower or somewhat higher.

[0107] Each tracker image r = r(t) = r t In this case, the appropriate size image neighborhood U = U around the reference point P = P(s) in segmentation s. t, s = U t, P(s) The (image subset) is automatically defined by the neighbor definer ND to define m-ROU in each or some instances of the tracker image. Thus, in some or preferably each of the tracker images r(t), the neighbor U t It is defined around the same reference point P associated with the same reference position z0. Each tracker image r tIt may be an axial slice. The increased accumulation of the contrast agent in the m-ROI (within the field of view of the imaging device IA) is offset for the low-dose setting when obtaining the tracker projection image r of the serial tracker image r t When obtaining the tracker projection image r of t , it is offset for the low-dose setting.

[0108] The reference point P can be extended to such a neighborhood U of any desired shape and size. For example, the neighborhood Ut can be defined as a circle, ellipse, or other shape arranged around the detected center of gravity P in each tracker image r t In t , the size of the circle can be predefined or automatically set so that it completely contains and preferably matches the anatomical structure referred to by the segmentation, such as the target section of the target blood vessel. The user can adjust any one of the size, position, and direction of the system-proposed neighborhood Ut. Alternatively, for simplicity, the neighborhood around P can be extended to a circle, ellipse, etc. of a predefined default size, such as a diameter of 20 mm. Then, the user can request through the UI to adjust the size interactively, by keying in data via the keyboard, or by selecting from a list of sizes (10 mm, 5 mm, 2 mm, etc.).

[0109] The in-image monitoring unit IMU is configured to perform the above-described bolus arrival based on examining the change in the image values in each neighborhood U t For this purpose, the in-image monitoring unit IMU can examine each voxel value within the range of the neighborhood U, or, for example, can examine only the average of such values. The in-image monitoring unit IMU can use any suitable threshold setting policy to monitor the image values representing the contrast agent concentration via a series of tracker imaging devices within each neighborhood U t The in-image monitoring unit IMU can examine each voxel value within the range of the neighborhood U, or, for example, can examine only the average of such values. The in-image monitoring unit IMU can use any suitable threshold setting policy to monitor the image values representing the contrast agent concentration via a series of tracker imaging devices within each neighborhood U t within a series of tracker imaging devices within each neighborhood U t For monitoring the image values representing the contrast agent concentration via a series of tracker imaging devices within each neighborhood U, any suitable threshold setting policy can be used.

[0110] The policy can include predetermined trigger conditions formulated with respect to the magnitude or gradient of the threshold, or a combination thereof. The neighborhood Ut When a trigger condition is discovered by the In-Image Monitoring Unit (IMU) based on the investigated evolution of values ​​within the image, a control signal is issued to the Controller CL. The Controller CL then acts to trigger the acquisition of a diagnostic projection image at diagnostic quality. The Controller CL can then instruct the Reconstructor (RECON) to reconstruct the target image V from the currently acquired diagnostic projection image λ.

[0111] The acquisition trigger signal issued by the in-image monitoring unit (IMU) may again include a requirement for adaptation of the imaging geometry to ensure that the diagnostic projection image λ to be acquired completely covers the area of ​​the target anatomical feature (TAF). However, as mentioned above, in most preferred cases, such adaptation is not required (preferred herein) because the target anatomical feature is sufficiently close to the monitoring region m-ROI, or the FOV is selected as such from the outset, for example, if the section along the rotation axis Z covered by the scan ("scan box") is sufficiently wide, which may depend, for example, on the width of the detector XD in the Z direction.

[0112] Figure 5 is an explanatory diagram of the data generated by the facilitator system FS. Specifically, pane a) shows a coronal view of the 3D survey volume V0, with the segmentation of the monitoring region highlighted, and the reference point P (e.g., centroid) taken from the subvolume / region marked by the segmentation. Thus, the corresponding sections of the subvolume segmentation s are shown in the coronal view pane a). The segmentation may have a different appearance / geometric structure when rendered in a different view of V0.

[0113] Based on the segmented subvolumes, a reference position l=z0 can be defined. For example, the output interface OUT can be operated (via a localizer component / function) to determine a reference point P as the centroid or other point of the area mapped by the segmentation s, by taking the z-coordinate of the point P. Once the reference position l=z0 is reached, this determines the 3D position of the monitoring region m—ROI—and the tracker image slice r, which is selected to pass through the reference position l=z0 or reference point P. The reference position also determines where the tracker projection data should be acquired. The image geometry (e.g., one or more of the gantry position and bed position) may need to be adapted to ensure that projection data can be acquired near z0. The trigger for tracker projection data acquisition is generally made by the user, in response to the start when the bolus dosing device ADA is activated. To be on the safe side, tracker projection data acquisition can be triggered at the start of bolus dosing, or shortly thereafter, for example, after the entire bolus volume has been administered in this manner.

[0114] Monitoring by the In-Image Monitoring Unit (IMU) is performed on functions of voxels, s, and P within the neighborhood U, as shown in pane b). In fact, the facilitator system FS may be able to invoke the visualizer VIZ to visualize any desired view of the segmented survey image Vo along any other reformatting direction, such as sagittal, coronal, or axial. The reference position P from which the monitored neighborhood U is extended is marked by a cross section ("X") in Figure 5, and such can be actually visualized in embodiments.

[0115] The neighbor U can be defined as an ellipse, a circle, or any other geometrically appropriate structure. The spatial dimension of the neighbor U is generally 2D, as is defined in the tracker image, which is generally a 2D slice image. However, if necessary, the neighbor U may be rendered as a 3D object such as an ellipsoid, a ball, etc., if the tracker image r is 3D, such as a slab rather than a slice, and such a 3D view is desired.

[0116] As mentioned above, the calculation of segmentation and monitored neighborhoods U around a reference position L can be performed automatically by using, for example, a machine learning model M or any other arbitrary method. In embodiments, ML is the segmentation, while the determination of P(s) and neighborhoods U may be performed by geometric analysis methods. If the size and orientation of neighborhoods U should be constrained to anatomical boundaries between tissue types, or vascular boundaries, or any other such anatomically driven, classical analytical segmentation approaches such as region growth methods, model-based segmentation, and others may be used, or in fact, a separate ML segmentation (different from model M that discovers anatomical landmarks) may be used in embodiments.

[0117] However, user interaction in the feedback loop regarding the results is also assumed herein, where the adjuster AJ can operate to adjust any one or more of the following: reference position P, segmentation, or the size or rotation shape of the monitoring region U on which the monitoring region m—ROI—is mapped. For example, a user interface UI, such as a graphical user interface GUI, may be used when exemplary visualizations shown in Figures 5a) and 5b) can be rendered. The visualization may be configured with interactive features, for example, the user may request such adjustments by touchscreen actions or by using a pointer tool (stylus, comp mouse, etc.) to adjust the size of the neighborhood, e.g., the position, shape, or size of the segmentation, or by clicking, pinching, or repositioning the reference position P by any other user interaction, gesture-based, or otherwise.

[0118] If the segmentation or reference position P is adjusted by the user based on the initially proposed segmentation and reference position, it will be understood that the associated neighborhood U will then be recalculated in response to such adjustment. For example, adjustment of the segmentation can trigger a recalculation of the reference position, for example, if the segmentation volume is changed, the centroid changes and triggers a new neighborhood U being calculated around the updated position. Thus, any change in any of the elements z0, U, P, s may trigger a redetermination of the remaining elements of such elements. In addition, or instead, modifications to the tracker image may be made by the user at any time, for example, by adjusting the position reference position, and the neighborhood U will be made to co-adjust. In some embodiments, adjustment of the neighborhood U can equally trigger a recalculation of the segmentation and / or reference position, etc. Thus, user input is embedded and operates dynamically, preferably in real time.

[0119] The example in Figure 5 shows that the organ of interest TAF is the kidney, while the monitoring region of the kidney may be selected as the descending aorta-diaphragm. Therefore, the location of the bolus m-ROI may vary depending on the target anatomical structure TAF, and thus the imaging protocol. For example, for a target anatomical TAF / dilation in the coronary arteries, the m-ROI is located in the ascending aorta near the heart. In cardiac applications, which are primarily (but not exclusively) assumed herein, a suitable m-ROI would include one or more of the following: (a) the aortic arch, (b) the ascending aorta, (c) the descending aorta-mediothoracic region, and (d) the pulmonary trunk. The proposed method may be applied to any (additional or alternative) anatomical landmark tracking location m-ROI as desired or specified in the examination protocol and / or CT workflow.

[0120] Next, an embodiment of the machine learning model M is shown, with reference to Figure 6, which may be used by the segmenter SEG.

[0121] Specifically, Figure 6 shows the components of a convolutional neural network (CNN) type model assumed herein, in an embodiment, which includes a convolutional filter CV to better explain the aforementioned spatial correlations in the pixel data constituting the input V, survey image. This setup will be described primarily in terms of post-training (deployment or testing), but the following also applies when the input is a training survey image x.

[0122] Specifically, Figure 6 shows a convolutional neural network M in a feedforward architecture. Network M comprises multiple computing nodes arranged in layers in a cascaded manner, where the data flow proceeds from left to right, and therefore from layer to layer. Recurrent networks are not excluded herein.

[0123] During deployment or training, input data including survey volume V0 is applied to the input layer IL. Thus, the input data x=V0 is fed into the model M in the input layer IL, then propagates through a series of hidden layers L1 to L3 (only three are shown, but there may be one, two, or more), and then appears in the output layer OL as the training data output M(x), or in deployment as the final segmentation M(V0)=s. The network M can be said to have a deep architecture because it has two or more hidden layers. In a feedforward network, "depth" is the number of hidden layers between the input layer IL and the output layer OL, while in a time network, depth is the number of hidden layers multiplied by the number of paths.

[0124] Network layers, specifically input and output maps, as well as hidden layer inputs and outputs (referred to herein as feature maps), can be represented as matrices ("tensors") of two or more dimensions for computational and memory allocation efficiency.

[0125] The output layer OL determines the model type, and since segmentation can be understood, for example, as a classification at the voxel level, this is preferably one of the classifications. Therefore, the output in the output layer OL is configured to output a binary or continuous segmentation map with voxel-level classification results regarding whether (or to what extent) each voxel or some voxels represents a portion of an acceptable m-ROI of a given task. Thus, in terms of dimensions and size (m × n × o), the classification map / segmentation map has the same dimensions as the input survey volume V0,x. The output layer may be configured based on a soft-max function.

[0126] Preferably, the system includes a sequence of convolutional layers, the hidden layers being represented herein as layers L1 to LN-k, k>1. The number of convolutional layers is at least one, such as 2, 3, 4, or 5, or any other number. The number may be represented by two digits.

[0127] In some embodiments, there may be one or more fully connected layers downstream of the sequence of convolutional layers, but this is not necessarily required in all embodiments, and in fact, in preferred embodiments, fully connected layers are not used in the architecture assumed herein.

[0128] Each hidden layer Lm and input layer IL implements one or more convolution operators CV. Each layer Lm can implement the same number of convolution operators CV, or the number may differ for some or all layers.

[0129] The convolution operator CV performs the convolution operation to be performed on each of its inputs. The convolution operator can be conceptualized as a convolution kernel. It can be implemented as a matrix containing entries that form filter elements called weights θ in this specification. In particular, it is these weights that are adjusted during the learning phase. The first layer IL processes the input data by its one or more convolution operators. The feature map FM is the output of the convolution layer, with one feature map for each convolution operator in the layer. The feature map from the previous layer is then input to the next layer to generate higher-generation feature maps FMi, FMi+1, etc., until the last layer OL combines all the feature maps to produce the output M(x),s. Each feature map entry can be described as a node. The final coupler layer OL may also be implemented as a convolution that provides a corrected front image.

[0130] The convolutional operator CV within a convolutional layer is distinguished from a fully connected layer in that the entries in the output feature map of the convolutional layer are not combinations of all the nodes received as input to that layer. In other words, the convolutional kernel applies only to a subset of the input volume V, or to the feature map received from a previous convolutional layer. The subset differs for each entry in the output feature map. Thus, the operation of the convolutional operator can be conceptualized as a "slide" on the input, similar to the discrete filter kernel in classical convolution operations known from classical signal processing. Hence the naming "convolutional layer". In a fully connected layer, the output node is generally obtained by processing all the nodes of the input layer.

[0131] The stride of the convolution operator can be selected as 1 or greater than 1. The stride defines how a subset is selected. A stride greater than 1 reduces the dimension of the feature map relative to the dimensions of the input in that layer. A stride of 1 is preferred herein. A zero-padding layer P is applied to maintain the dimensioning of the feature map to correspond to the dimensions of the input image. This allows for the convolution of feature map entries located at the edges of the processed feature map.

[0132] In a preferred embodiment (see schematic Figure 6A), a convolutional neural network (CNN) is used as model M configured for multiscale processing. Such models include, for example, the U-unit architecture or its derivatives, as described by O. Ronneberger et al. in "U-Net: Convolutional Networks for Biomedical Image Segmentation," which is available online on the arXiv repository under citation code arXiv:1505.04597(2015).

[0133] In multiscale analogous NN models of this or a similar architecture, the deconvolution operator DV is used. More specifically, such a multiscale-enabled architecture of model M may have a downscale path DPH of layers in series with an upscale path UPH of layers downstream thereof. Convolution operators with varying strides gradually reduce the feature map size / dimension (m×n) by passing through layer Lj in the downscale path, reducing it to the lowest dimensional representation by the feature map, i.e., the latent representation κ. The latent representation feature map κ has dimensions that gradually increase as it passes through layer Lk of the upscale path UPH, returning to the dimensions of the original input data x. In this case, the output is a segmentation s represented as a map having the same size (3D) as the input survey volume VO, x. This dimensional bottleneck structure has been found to result in better learning because it forces the system to distribute its learning down to the simplest possible structure so that it is represented by the latent representation.

[0134] An additional regularization channel may be defined when intermediate outputs (intermediate feature maps) from the scale level in the downscale path are fed to the corresponding layer at the corresponding scale level in the upscale path UPH. Such interscale interconnects ISX have been found to promote learning and performance robustness and efficiency.

[0135] The multiscale processing of Figure 6A or a similar model is particularly useful herein. Specifically, although contrast enhancement is insufficient in the initial 3D survey image, structures of interest such as surveillance ROI m-ROI / landmarks (and optionally target anatomical features) can still be sufficiently segmented for multiscale processing, where even a small contrast gradient may be sufficient for the multiscale model M to find the desired segmentations s of the m-ROI. In addition, diagnostic training images with sufficient contrast can be used to train model M, as will be explored in more detail below in relation to the training data generation system TDGS.

[0136] Naturally, the monitoring locations s from which the initial survey V0 is segmented may differ for different imaging tasks. Therefore, each imaging task or objective defines a target anatomical feature TAF, defining an appropriate (one or more) monitoring region m—ROI—as an anatomical landmark associated with the target feature TAF for that medical imaging task, based on clinical knowledge. Thus, at least one different model Mj can be trained for each given imaging task / protocol / standard. Therefore, each model Mj in the thus-obtained set of models {Mj} is configured to segment for different monitoring regions, such as i) different target anatomical feature TAFs, or ii) different tasks for the same TAF, etc. More specifically, the set of models {Mj} can be stored and managed as a bank or library of different such trained model Mj. A database management system MEM' can provide a platform for automated or user-initiated queries to retrieve, or otherwise access, the correct model Mj' for a given task j'.

[0137] The specifications for the image task j' to be performed may be provided by the user prior to acquisition. The specifications may be an XML file, a message, or any other file. For example, the user may select an event handler (not shown) that uses a selection signal to provide access to the image task j' and a trained model Mj' associated with this task, either selectively or through a menu structure in a GUI or other interface, and the input at input port IN is supplied to the model Mj' thus selected. For example, more specifically, the specifications for task j' (XML file, flags, etc.) are provided alongside the survey image V0, both of which are received at input port IN. The specifications allow for the selection of a corresponding machine learning model Mj, or more generally, an associated segmenter SEGj, which is configured to find appropriate landmarks for the surveillance region m—ROI—and a target anatomical feature TAF and / or task j'.

[0138] Next, refer to Figure 7, which shows the training system TS for training the ML model M, which the segmenter SEG will use, based on the training data.

[0139] The training data preferably includes annotations down to the voxel level to ensure the desired level of accuracy assumed herein. The training data generator system TDGS can supply labeled training instances (x,y), where x represents a historical survey image pulled from an image database and y specifies its ground truth label / annotation. The label may be a voxel map that unifies voxels representing acceptable definitions of m-ROI. The label can be applied to the 3D volume of x, or to one of its views, such as corona, as needed. The training data generator system TDGS may be configured as an interactive user interface system (UI) to support the user in annotation tasks. For example, the training data generator system TDGS may be configured to allow the user to define seed markers SM1 to 2 in the training input survey volume x. Such seed markers SM1 to 2 may be used to define a convex hull CH from which a map is automatically generated. Seed markers, such as pairs (or more) of circles or other geometric figures, are set by experts on the training input survey image x via the graphical user interface GUI of the training data generation system TDGS. This is explained in more detail in Figure 9 below.

[0140] As previously observed, the survey input image x, like any arbitrary survey image, has very poor contrast, and even a human expert cannot always reliably or completely annotate the ground truth m-ROI. In such situations, the training data generation system TDGS may be able to operate to generate training data pairs (x,y) based on noise simulation as follows: A set of existing historical 3D diagnostic images (spectral or non-spectral images) of a common body part of interest on which the model should be trained is localized within a database query in a medical database such as PACS. In such diagnostic images, annotation can be easily achieved by a clinical expert to define the location and extent of the m-ROI. Next, the high IQ (image quality) of the diagnostic image used for annotation simulates noise, which is artificially reduced by adding this noise to the diagnostic image, thereby simulating the low-dose effect in survey imaging and thus obtaining an artificially generated sample that represents a good approximation of the survey volume case. Thus, the noise-damaged 3D image sample can serve as the training input x, while the annotation on the high-quality image serves as the associated ground truth y. Therefore, the training data generation system TDGS may include a noise simulator and a noise adder to generate pairs of training data as needed, following the line described above.

[0141] The training system TS is configured to train an ML model M based on such training data TD = {(x, y)}. The training data is optionally provided by the training data generator system TDGS.

[0142] In a supervised setting, training data can include k pairs of data (xk, yk). For each pair of k, the training data comprises training input data xk and associated target or ground truth yk. Therefore, training data is organized in pairs of k, particularly for the supervised learning plans assumed herein. However, it should be noted that unsupervised learning plans are not excluded herein. x and y are as defined above.

[0143] In the training phase, the initial set of weights is pre-inputted into the architecture of the machine learning model M, such as the CNN network shown in Figure 6 below. The weights θ of the model NN are parameterized M. θ The goal of the training system TS is to optimize and thus adapt the parameter θ based on the training data (xk, yk) pairs. In other words, learning can be mathematically formulated as an optimization scheme that minimizes the cost function F, but a dual formation that maximizes the utility function may be used instead.

[0144] Here, assuming a cost function paradigm F, this measures the aggregated residuals, i.e., the error that occurs between the data estimated by the neural network model NN and some or all of the training data targets k. argmin θ F=Σ k |M θ (x k ),y k (1)

[0145] In equation (1) and hereafter, the function M() represents the result of the model NN applied to the input x. The cost function may be pixel / voxel based. For classification tasks, as assumed herein in some embodiments, the sum of the cost function F in (1) may be based on cross-entropy, weighted cross-entropy, or negative log-likelihood (NLL) divergence, or similar. As a further alternative, the cost function may be based on a Dice coefficient function that measures the difference between sets with respect to the size of their overlap (crossing), in which case the input training output and label segmentation map are compared as sets. The Jaccard index function may also be used.

[0146] However, it may be possible to represent the segmentation task not as a classification but as a regression, in which case the cost function may be based on the squared Euclidean distance between the prediction map M(x) and the label map y. Thus, an L2 loss function may be used, or in some cases even an L1 loss function may be used.

[0147] During training, the training input data xk of a training pair is propagated through an initialized network M. Specifically, the training input xk of the k-th pair is received at input IL, passes through model M, and is output at output OL as output training data Mθ(x). A suitable distance measure |·| is used to measure the difference (also called the residual herein) between the actual training output Mθ(xk) generated by the model NN and the desired target yk, such as p-norm, squared difference, or other difference. However, the distance measure |·| is not necessarily based on the underlying norm definition. The notation "||" used herein is a general notation referring to any suitable distance measure of any kind.

[0148] The output training data M(xk) is an estimate of the target yk associated with the applied input training image data xk. Generally, there is an error between this output M(xk) and the associated target yk of the k-th pair currently being considered. Then, the parameters θ of the model NN can be adapted using an optimization scheme such as backpropagation or other gradient-based methods to reduce the residuals of the pair (xk, yk) being considered or a subset of the training pair from the full training dataset.

[0149] After one or more iterations in the first inner loop, in which the model parameter θ is updated by the updater UP for the current batch of pairs {(xk, yk)}, the training system TS enters the second outer loop, where the next batch of training data pairs {xk+1, yk+1} is processed accordingly. While it is possible to loop through the outer loop over individual pairs until the training dataset is exhausted, it is preferred herein to loop over a set of pairs (i.e., over the aforementioned batch of training pairs) at once. The iterations in the inner loop are over the parameters of all pairs that make up the batch. Such batch-style leaning has been shown to be more efficient than proceeding pair by pair. Hereafter, the subscript "k" indicating individual instances of training data pairs will be reduced to a largely meaningless notation.

[0150] The structure of the updater UP depends on the optimization scheme used. For example, an internal loop managed by the updater UP may be implemented by one or more forward and backward passes in the backpropagation algorithm. To improve the objective function while adapting the parameters, the aggregated, e.g., summed residuals of all training pairs in a given batch are considered up to the current pair. The aggregated residuals can be formed by constructing the objective function F as the sum of squared residuals, such as equation (1), of some or all of the residuals considered for each pair. Other algebraic combinations are also possible instead of the sum of squares.

[0151] GPUs can be used to implement the training system TS to improve efficiency.

[0152] While voxel-wise training is preferred herein, this is not necessarily the case in all embodiments, and training with coarser granularity on patches or tiles (subsets of voxels) is not excluded herein for the purpose of saving CPU time and memory, or for faster training.

[0153] Next, we refer to the flowchart in Figure 8, which shows the steps of a computer implementation method for carrying out the facilitator system FS described above. Figure 9 is a flowchart relating to a model training mode, and therefore to the training system TS. However, the following flowchart may be understood as teaching in itself, not necessarily in all embodiments related to the above.

[0154] However, in some embodiments, within the context of the embodiments described above, it should be understood that the method is performed by computer and, in particular, facilitates contrast-assisted imaging protocols such as angiography for tomographic imaging sessions such as angiography-CT or C-arm. That is, X-ray based modalities such as radiography, CT, C-arm, mammography, and tomosynthesis are all assumed in different embodiments herein. Attenuation imaging is primarily assumed, i.e., other imaging techniques such as phase contrast and dark-field imaging are excluded herein. They are not. Also, methods and systems assumed and proposed in the specification for non-X-ray based imaging modalities such as MRI and nuclear imaging, which can use contrast agents in some form, may also be used with their benefits therein.

[0155] In step S800, a 3D survey image V0 with lower image quality than the subsequent diagnostic volume V is acquired. This can be done by acquiring an initial projection image λ' with a lower dose than the dose of the diagnostic projection image λ that will later be used for the diagnostic volume V. The low-dose projection image λ' thus acquired in the projection region is then reconstructed in the image region by a tomography reconstruction algorithm such as FBP, iterative, algebraic, or any other.

[0156] In step 805, the survey image V0 is received by the computing system on which the method is performed.

[0157] In step S810, the low-dose (low-IQ) 3D survey volume V0 is segmented with respect to one (or more) surveillance areas m-ROIs, such as anatomical landmarks associated with the target anatomical feature TAF. The target anatomical feature TAF is the medical purpose of the image, or defines it. The purpose or target anatomical feature TAF is, in embodiments, alongside the 3D survey volume V0, either received in step S805 or specified in the volume specification. However, the specification may be pre-configured or otherwise predetermined by context, and therefore, receiving this specification in step S805 does not need to be an explicit action, but may be implicitly provided in the computing environment in which the method is performed, or by any other method or scheme.

[0158] The specification may determine a particular type of segmenter or segmentation algorithm to be used for step S810, in any form provided or received, either explicitly or implicitly. The segmenter or segmentation algorithm may be specifically tailored to find specific relevant landmarks as monitoring regions in low-dose survey images, taking into account the purpose or nature of the target anatomical feature TAF. The segmentation in step S810 may be ML-based, such as being based on a trained NN model, such as a CNN, preferably configured for multiscale processing. Any other ML models capable of such segmentation are assumed herein, and are therefore more classically configured segmentation techniques, such as region growth, or preferably (shape) model-based segmentation ("MBS"), which are not typically considered ML. However, ML-based methods have been found to cope well with the generally low IQ in survey images.

[0159] The field of view (FOV) for the survey image VO in S800 may be a priori or based on the margin of view (ML), and is selected so that the intended m-ROI and target anatomical feature (TAFS) are roughly known to be covered by imaging based on general medical knowledge, but the exact location of the m-ROI is unknown before step S810.

[0160] In step S820, segmented landmarks for the surveillance region m-ROI are made available. The segmentation s of the surveillance region m-ROI is a sub-volume (subset) of the survey volume V0. Therefore, the segmentation is also a 3D character. Preferably, the segmentation of the surveillance region m-ROI matches in shape and to the extent of the surrounding anatomical tissue boundary with which it borders.

[0161] Generally, the bolus monitoring region m-ROI is based on segmentation s. The monitoring region m-ROI has a location (reference location z0) and a spatial extension to the tracker image. The tracker image is acquired to monitor bolus arrival. The location and extent of each m-ROI are determined as described in the following steps.

[0162] The initial tuning of the m-ROI and the reference position z0 of the tracker image rt is determined in step S830 based on the segmented region s. For example, the reference point P of the segmentation s, e.g., the centroid, can be calculated for the 3D subvolume as defined in the segmentation s. The z-coordinate of P in the image region can be used as the z-position of the tracker image r(t) to be acquired, and thus defines the reference position z0.

[0163] Generally, the reference position is located along the rotation axis Z of the imaging device, which is the position where the tracker projection image should be acquired, and this is the position in the image region where the tracker image is positioned.

[0164] As described above, the reference position z0 of a segmentation defines its location (as opposed to its spatial extent) within the image region of interest m-ROI. The use of the centroid P of the segmentation as described above is merely one example. Other methods for defining the location of m-ROI, such as drawing the envelope of the segmentation structure s, may also be considered herein. Alternatively, P may be defined as the center point of the largest circle / ellipse that fits within the region marked by the segmentation. However, pinning the m-ROI to a single coordinate system (such as using the centroid P) may be advantageous in terms of computational responsiveness and memory resources. Any other method for calculating the reference position z0 based on the segmentations s may be used instead, which may include two or more such reference points P.

[0165] In step S840, a tracker image rt is acquired. This may involve acquiring a series of time sets of tracker projection images λ'' at a given sampling rate around the reference position, and at a dose lower than that required later for the diagnostic image. The tracker image rt is then reconstructed from the set as each axial slice t in a plane passing through the reference position z0, as calculated in step S830. Alternatively, a non-axial format may be used, defined along any axial direction as needed and obtainable by a reformatting algorithm.

[0166] Preferably, the administration of the contrast agent can determine when the acquisition of tracker projection images should begin in step S840. This can be started before or immediately after CA administration. Care should be taken not to "miss" the point at which CA begins to accumulate in the m-ROI.

[0167] In step S850, the range of the monitoring area m-ROI for a part or each part of the tracker image rt is defined.

[0168] Preferably, this is done in step S850 by defining each neighborhood U around a reference point P (one or more) in some or each instance of the tracker image, where P ∈ Ut ⊂ rt.

[0169] Preferably, instead of defining neighbor U in the tracker image rt, it can already be defined in the locator image which can be synthesized from the survey volume V0 before the “live” tracker image is acquired. That is, no new acquisition is required to define neighbor U. The neighbor thus defined on the synthesized locator tracker image r0 can then be used repeatedly in the live tracker image instance reconstructed from the newly acquired tracker projection data in step S840. Therefore, step S850 can be performed before step S840, in which the live tracker image rt is acquired for monitoring.

[0170] The neighborhood U may be, for example, an ellipse, a square, or any other geometric shape or form. Thus, each instance of the tracker image has a neighborhood Ut defined within it around the same reference position P.

[0171] Preferably, the neighborhood U is defined in this way with respect to the anatomical tissue boundary. For example, U may be defined so that it does not extend beyond the image representation of the vessel wall to which the bolus should move, and is captured in each instance of the tracker image rt. The neighborhood may be extended as far as possible, as long as it remains within the image representation of the vessel wall. The neighborhood U needs to be determined only once, for example, on the locator tracker image or the first line tracker image, and then it can be reused for all follow-up tracker images. The same applies to the reference position z0, which needs to be determined once based on the segmentation s. If the view changes (e.g., from axial to sagittal or other directions), any of the above components s, U, P, etc. may need to be transformed accordingly simultaneously.

[0172] Preferably, the neighbor U definition step S850 based on the reference position definition z0 in step S830 and the locator image r0 is performed in a single procedural step, and steps S830 and S850 are merged because both depend only on the survey image V0 information, which includes segmentation.

[0173] In step S860, a series of tracker images are monitored for the arrival of the contrast agent. This can be done by thresholding or any other policy that analyzes image values ​​in the neighborhood U around the image position z,P in the tracker image.

[0174] If the contrast agent concentration is deemed sufficient in step S860 based on the monitored image values ​​within the vicinity U, a control signal may be issued in step S870 to acquire the target volume TAF at the full dose, and it is assumed that its location is known and included in the current FOV used for the tracker image rt, as described above. Thus, since the FOV has been appropriately and broadly selected for acquiring the tracker projection data, the same FOV may be used to acquire the diagnostic image V. Therefore, step S870, upon receiving the trigger signal according to step S860, may include acquiring a diagnostic projection image λ at a higher diagnostic dose at the location of the anatomical feature TAF of interest, and then the target volume can be reconstructed from the projection image λ.

[0175] In any, but preferred, step S835, a user interaction request may be received requesting modification of any one or more of the segmentations s, ii) a reference position z0 derived therefrom, iii) a reference point P, iv) and / or the monitored neighborhood U in the tracker image as a function of the segmentations and / or position z0.

[0176] A change to one or more components i) through iv) may trigger the corresponding adaptations to one, two or more, or all of the remaining components i) through iv).

[0177] Furthermore, the above method can optionally include various display operations.

[0178] For example, optionally, before step S840, a locator tracker image synthesized from the survey volume may be displayed to the user as a quality check, based on a newly acquired tracker projection image, as a locator tracker image synthesized from the survey volume before it is actually acquired.

[0179] The adjacent U of m-FROI in each instance of the tracker image can be graphically outlined, for example, by color coding or by line style modulation of the contour curve (dashed lines, bold lines, etc.).

[0180] The survey image V0 in any desired view may be displayed simultaneously with the video stream of tracker image instances, or simultaneously with each instance, such as a still image. In the survey image, segmentation may be outlined in color or visualized by graphically modulated contour curves (nords, dashed lines, etc.). In embodiments, a graphics display such as those shown in Figures 5a), b) is envisioned to display in any arrangement of panes a), b) deemed suitable. It is envisioned to display on a single display device DD or across multiple display devices. Each of the display options described may be implemented on its own, or in combination with any one or more of the other options, as required by the user.

[0181] Next, refer to Figure 9, which shows a flowchart illustrating how to train a machine learning model based on training data.

[0182] This method may include an initial step S905 for generating such training data.

[0183] Generating machine learning training data may include voxel-level based annotations of existing survey images, such as those obtained from a medical database of previous examinations. Medical users / experts may use appropriate pointer tools to perform annotations, and graphical user interface-based tools are preferred, for example, as they allow for more efficient annotation operations.

[0184] Specifically, the ground truth data y or labels of individual bolus tracking m-ROIs can be annotated by forming a binary voxel mask within an existing historical 3D survey image x. The mask indicates the tolerance along the vessel where the bolus tracking m-ROI can be placed. The use of a continuous probability map, whose values ​​vary at unit intervals, instead of a binary map, is also contemplated herein.

[0185] One efficient way to annotate such a range is to place spheres at the start and end points of such an acceptable range and adjust the radius of the spheres to match the size of the blood vessel, as shown in the survey training image x. As mentioned above, the diagnostic IQ image can be used for annotation to define y, while its artificially noisy version is used as the associated x. The range can then be calculated as the convex hull CH (see Figure 7) surrounding the connected spheres SM1 to SM3. Again, as mentioned above in Figure 7, an interactive GUI setup that provides graphical drawing input may be preferred to carry out this step in relation to the training data generator system TDGS. Thus, the acceptable region thus defined automatically indicates that all voxels covered by the blood vessel, unlike voxels outside the blood vessel, represent m-ROI voxels. The spheres SM1 to SM3 mentioned above are thus placed from seed markers for the convex hull CH. Depending on the image structure of interest, other seed markers of other shapes can be used.

[0186] In step S910, the training data is received in the form of a set (batch) of pairs (xk, yk), either to be generated in step S905 or to be retrieved from another location. Each pair contains a training input xk and an associated target yk, where xk is a 3D image survey volume and yk is an annotated segmentation map.

[0187] In step S920, the training input xk is applied to the initialized machine learning model M to generate the training output.

[0188] The deviation or residual of the training output M(xk) from the associated target yk is quantified by a cost function F (e.g., equation (1) above). One or more parameters of the model are adapted in one or more iterations in the inner loop in step S930 to improve the cost function. For example, model parameters are adapted to reduce the residuals measured by the cost function. The parameters include, in particular, the weights W of the convolution operator CV and an arbitrary bias term. Initially, the model may be used with pre-input parameters such as random parameters, or a pre-trained model may be used.

[0189] Next, the training method returns to step S910 in the outer loop, where the next batch of training data pairs is supplied.

[0190] In step S930, the model parameters are fitted so that the aggregated residuals of all pairs in a given batch under consideration are reduced, in particular, minimized. The cost function quantifies the aggregated residuals. Propagation, especially backpropagation, or similar gradient-based techniques may be used in the inner loop. Backpropagation allows for the consideration of the prediction errors used to adapt the parameters. Thus, for each batch of pairs under consideration, more information and better generalization ability can be extracted thanks to backpropagation.

[0191] More generally, the parameters of the model M are adjusted to improve the cost function F. In embodiments, the cost function is configured to measure aggregated residuals, such as an integer term of (weighted) cross-entropy, a Dice function, or another cost function configured for the purpose of classification / segmentation tasks. In embodiments, residual aggregation is performed by summing over all or some residuals for all pairs considered. The method may be implemented on one or more general-purpose processing units TS having parallel processing processors to accelerate training.

[0192] The components of the system SYS may run as one or more software modules on one or more general-purpose processing unit PUs, such as workstations associated with imager IAs, or on server computers associated with groups of imagers.

[0193] Alternatively, some or all components of the system SYS may be configured as hardware such as a properly programmed microcontroller or microprocessor, for example as an FPGA (Field-Programmable Gate Array), or as a hardwired IC chip, application-specific integrated circuit (ASIC), integrated into the imaging system IA. In further embodiments, the system SYS may be implemented partially in software and partially in hardware.

[0194] Different components of the system SYS can be implemented on a single data processing unit (PU). Alternatively, several or more components can be implemented on different processing units (PUs), potentially located remotely within a distributed architecture and connected within an appropriate communication network, such as a cloud setup or client-server setup.

[0195] One or more features described herein may be configured or implemented as a circuit encoded in a computer-readable medium, or using a circuit and / or a combination thereof. A circuit may include discrete and / or integrated circuits, a system-on-a-chip (SOC), and combinations thereof, a machine, a computer system, a processor and memory, and a computer program.

[0196] In another exemplary embodiment of the present invention, a computer program or computer program element is provided, characterized in that it is adapted to perform a method step of a method according to one of the embodiments described above on a suitable system.

[0197] Accordingly, the computer program elements may be stored in a computer unit, which may be part of an embodiment of the present invention. This computing unit may be adapted to perform or trigger the execution of the steps of the method described above. Furthermore, it may be adapted to operate the components of the apparatus described above. The computing unit may be adapted to operate automatically and / or to execute user sequences. The computer program may be loaded into the working memory of a data processor. Accordingly, the data processor may be equipped to perform the method of the present invention.

[0198] This exemplary embodiment of the present invention encompasses both computer programs that use the present invention from the outset and computer programs that, through updates, transform existing programs into programs that use the present invention.

[0199] Furthermore, computer program elements can provide all the steps necessary to satisfy the procedure of the exemplary embodiment of the procedure described above.

[0200] According to a further exemplary embodiment of the present invention, a computer-readable medium such as a CD-ROM is presented, having computer program elements stored thereon, which are described in the previous section.

[0201] Computer programs may be stored and / or distributed on suitable media (in particular, but not necessarily, non-temporary media) such as optical storage media or solid-state media supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems.

[0202] However, computer programs may also be presented over a network such as the World Wide Web and downloaded from such a network into the working memory of a data processor. According to a further exemplary embodiment of the present invention, a medium is provided for making a computer program element available for download, and this computer program element is configured to perform a method according to one of the aforementioned embodiments of the present invention.

[0203] It should be noted that embodiments of the present invention are described with reference to different subject matter. In particular, some embodiments are described with reference to method-type claims, and other embodiments are described with reference to apparatus-type claims. However, unless otherwise notified, those skilled in the art will find that any combination of features belonging to one type of subject matter, as well as any combination of features relating to different subject matter, are gathered from the above and below descriptions and are deemed to be disclosed in this application. However, all features can be combined to provide a synergistic effect greater than the simple sum of the features.

[0204] Although the present invention is illustrated and described in detail in the drawings and the foregoing description, such illustrations and descriptions should be considered illustrative or exemplary and not limiting. The present invention is not limited to the disclosed embodiments. Other variations of the disclosed embodiments can be understood and practiced by those skilled in the art in carrying out the claimed invention from a study of the drawings, disclosure and dependent claims.

[0205] In the claims, the word “comprising” does not exclude other elements or steps, and the indefinite article “a” or “an” does not exclude plurals. A single processor or other unit may fulfill the functions of several items enumerated in the claims. The mere fact that certain means are referenced in different dependent claims does not imply that combinations of these means cannot be used advantageously. No reference numeral in the claims should be construed as limiting in scope. Such reference numerals may consist of numbers, letters, or any combination of alphanumeric characters.

Claims

1. A system for facilitating contrast-based tomography, wherein the system, when in use, The specifications for the imaging operation of the tomography type imaging device, and an input interface for receiving a 3D survey image volume obtained by the imaging device of at least a portion of the patient during a preparation phase prior to the contrast-assisted imaging phase. A segmenter configured to segment the 3D survey image volume with respect to a surveillance region m-ROI, wherein the m-ROI is associated with a target anatomical feature identified in the specification, To facilitate monitoring for the presence of contrast agent with respect to the target anatomical features, an output interface is provided that provides a reference position for segmented m-ROIs in the 3D survey volume. A system that has

2. The system according to claim 1, comprising a controller configured to instruct an imaging device to acquire a first set of projection data while a contrast agent is propagating within a patient, and a reconstructor that reconstructs a cross-sectional tracker image in a plane passing through the reference position with first image quality based on the projection data.

3. The system according to claim 2, comprising an in-image monitoring unit configured to perform monitoring based on the tracker image in the vicinity of the image in the tracker image, wherein the vicinity of the image is based on the reference position.

4. The system according to claim 3, wherein the image neighborhood is defined in a composite tracker image synthesized from the survey image volume, and the image neighborhood is used by the in-image monitoring unit to monitor in the reconstructed tracker image.

5. The system according to any one of claims 2 to 4, wherein the controller is configured to instruct the imaging device to acquire a second set of projection data on an in-image monitoring unit that issues a trigger signal based on one or more monitored image values ​​in the vicinity of the image, at a second image quality higher than the first image quality.

6. The system according to any one of claims 1 to 5, wherein the reference position is based on at least one reference point with respect to the segmentation of the m-ROI.

7. The system according to any one of claims 1 to 6, comprising a adjuster for adjusting the position, orientation, and / or spatial extent of a vicinity within the image.

8. The system according to claim 7, wherein the adjuster is operable in response to commands that can be received via a user interface configured to allow a user to request adjustment of one or more of the following: i) the vicinity in the image, ii) the reference position, iii) the reference point, and iv) the segmentation of the m-ROI.

9. The segmenter is based on a trained machine learning model, according to any one of claims 1 to 8.

10. An imaging device comprising the system according to any one of claims 1 to 9, a contrast agent administration device for administering a contrast agent, and one or more of the imaging devices.

11. A method for facilitating contrast-based tomography, The specifications for the imaging operation of the tomography type imaging device, and the step of receiving a 3D survey image volume obtained by the imaging device of at least a portion of the patient in a preparation stage prior to the imaging stage assisted by a contrast agent, The 3D survey image volume is segmented with respect to a surveillance region m-ROI, wherein the m-ROI is associated with a target anatomical feature identified in the specification. To facilitate monitoring the presence of contrast agent with respect to the target anatomical features, the steps include providing a reference position for segmented m-ROIs in the 3D survey volume. A method having.

12. A computer implementation method comprising the step of training a machine learning model according to claim 9 based on training data.

13. A computer implementation method comprising the step of generating training data, wherein the method according to claim 12 is a method for training the machine learning model based on the step.

14. A computer program element adapted to cause at least one processing unit to perform the method according to any one of claims 11 to 13, when executed by said processing unit.

15. At least one computer-readable medium storing the program element described in claim 14, or the machine learning model described in claim 9.

Images