Patient anatomy and task-specific automatic exposure control in computed tomography

By training a deep learning model using machine learning techniques, the automatic exposure control of CT scans was optimized, solving the problems of uneven radiation dose and image quality degradation in conventional CT imaging. This enabled personalized adjustment of CT scan parameters, improving image quality and reducing radiation dose.

CN115797252BActive Publication Date: 2026-11-17GE PRECISION HEALTHCARE LLC +1
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202211064986.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2021-09-10
Filing Date
2022-09-01
Publication Date
2026-11-17
Estimated Expiration
2042-09-01

AI Technical Summary

Technical Problem

In current CT imaging, conventional automatic exposure control algorithms fail to effectively consider patient-specific anatomical structures and clinical task requirements, resulting in uneven radiation dose and decreased image quality.

Method used

Machine learning techniques are used to train a deep learning model, which estimates organ radiation dose and image quality based on localization images, optimizes the automatic exposure control mode of CT scans, and predicts radiation dose and image quality under different AEC modes through the deep learning model. CT scan parameters are then dynamically adjusted to optimize image quality and reduce radiation dose.

Benefits of technology

It enables personalized CT scans based on patient-specific anatomy and clinical tasks, improving image quality and reducing radiation dose, and providing consistent high-quality image output.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115797252B_ABST
    Figure CN115797252B_ABST
Patent Text Reader

Abstract

Techniques for customizing automated exposure control (AEC) settings for specific patient anatomy and clinical tasks are described. According to an embodiment, the computer-implemented method includes receiving one or more localized images of a patient's anatomical region captured in association with the execution of a computed tomography (CT) scan. The method also includes employing a first machine learning model to estimate, based on the one or more localized images, an expected organ dose representing the expected radiation dose to organs exposed in the anatomical region under different AEC modes used for the CT scan. The method may further include employing a second machine learning model to estimate, based on the one or more localized images, an expected metric of image quality in the target and background regions of the scan images captured under the different AEC modes, and determining an optimal AEC mode based on the expected organ dose and the expected image quality metric.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the optimization of medical image computed tomography (CT) scans, and more specifically to techniques for customizing automatic exposure control settings for specific patient anatomy and clinical tasks. Background Technology

[0002] Computed tomography (CT) has been one of the most successful imaging modalities and has facilitated countless image-based medical procedures since its invention decades ago. CT results in significant exposure to ionizing radiation, especially with the rapid increase in CT examinations. Therefore, reducing CT radiation dose is desirable. However, dose reduction also introduces additional noise, and diagnostic performance may be compromised due to decreased image quality.

[0003] In this regard, radiation dose and image quality have traditionally been competing goals in CT imaging. To balance these two, modern CT systems use automatic exposure control (AEC) based on localization images, particularly tube current modulation (TCM). However, conventional AEC algorithms aim to provide a uniform noise level across the entire imaging volume. While this is generally successful in avoiding excessive radiation dose and maintaining image quality, it is primarily based on the patient's approximate size and body shape. Furthermore, conventional AEC algorithms use dose metrics that reflect tube output rather than the patient's actual dose. Conventional AEC techniques still lack awareness of patient-specific anatomy affecting organ dose, do not consider task-based image quality, and are prone to problems such as patient centering. Summary of the Invention

[0004] The following summary is presented to provide a basic understanding of one or more embodiments of the invention. This summary is not intended to identify key or essential elements, nor is it intended to depict any scope of the different embodiments or any scope of the claims. Its sole purpose is to present concepts in a simplified form as a prelude to the more detailed description that follows. In one or more embodiments described herein, systems, computer-implemented methods, apparatuses, and / or computer program products are provided to facilitate the customization of AEC settings for specific patient anatomy and clinical tasks using machine learning techniques.

[0005] According to one embodiment, a system is provided including a memory and a processor, the memory storing computer-executable components, the processor executing the computer-executable components stored in the memory. The computer-executable components include a receiving component that receives one or more localization images of the anatomical region of a patient, captured in association with the execution of a CT scan of that anatomical region, wherein the anatomical region includes one or more organs. The computer-executable components also include a dose estimation component that employs one or more first machine learning models to estimate, based on the one or more localization images, a predicted organ dose representing the expected radiation dose to the one or more organs under different AEC modes used for the CT scan.

[0006] The one or more first machine learning models include deep learning models trained using a supervised machine learning process and benchmark ground truth data, the benchmark ground truth data including multiple CT volumes generated from CT scans of the anatomical region under the different AEC modes and organ segmentation dose maps generated from the CT volumes. In some embodiments, the computer-executable component further includes a training component that trains the deep learning model using a supervised machine learning process.

[0007] In some embodiments, the computer-executable component further includes a quality estimation component that employs one or more second machine learning models to estimate expected metrics of image quality in target and background regions of scan images captured in the different AEC modes based on the one or more localization images. The quality estimation component determines the target and background regions based on a defined clinical task associated with the execution of the CT scan, wherein the target and background regions vary for different clinical tasks. The one or more second machine learning models include a second learning model trained using a supervised machine learning process and benchmark ground truth image quality data generated using multiple CT volumes generated from CT scans of the anatomical region in the different AEC modes, wherein the benchmark ground truth image quality data provides metrics of image quality in the target and background regions as represented in the CT volumes in the different AEC modes. In some embodiments, the computer-executable component further includes a training component that trains a second deep learning model using a supervised machine learning process. The computer-executable component may also include a simulation component that generates simulated noise projections using the CT volumes and generates the benchmark ground truth image quality data based on the simulated noise projections.

[0008] The computer-executable component may further include an optimization component that determines an optimal AEC mode among the different AEC modes performed for the CT scan based on the expected organ dose and an expected metric of image quality in the target region, wherein the optimal AEC mode maximizes image quality in the target region and minimizes radiation dose to the organ. In this regard, different acquisition parameter values ​​selected from the group consisting of: tube current modulation (TCM), tube voltage, collimation, filtering, arcuate junction, pitch, and initiation angle.

[0009] According to one embodiment, another system is provided, the system including a memory and a processor, the memory storing computer-executable components, and the processor executing the computer-executable components stored in the memory. The computer-executable components include a training component that trains a first deep learning network to determine a predicted organ dose, representing the expected radiation dose to one or more organs exposed to the anatomical region in different AEC modes of a CT scan, based on one or more localization images of a different patient's anatomical region captured. The training component further trains a second deep learning network to determine a predicted metric of image quality in a target region and a background region of the scan images captured in the different AEC modes, based on the localization images. The computer-executable components also include an optimization component that uses the first and second deep learning networks to determine an optimal AEC mode among the different AEC modes for the CT scan, the optimal AEC mode maximizing image quality in the target region and minimizing the radiation dose to the one or more organs.

[0010] The target region and the background region vary for different clinical tasks. In some embodiments, the training component may further train a third deep learning network to determine the optimal AEC pattern for each localization image combined with the clinical task. The computer-executable component may also include an inference component that employs the third deep learning network to determine the optimal automatic exposure pattern for a new CT scan of the patient's anatomical region based on one or more newly captured localization images of the patient's anatomical region and the selected task of the different clinical tasks.

[0011] According to another embodiment, a different system is provided, comprising a memory and a processor, the memory storing computer-executable components, the processor executing the computer-executable components stored in the memory. The computer-executable components include a receiving component that receives task information identifying a clinical task associated with the execution of a CT scan of a patient's anatomical region, and one or more captured localization images of the anatomical region. The computer-executable components further include an optimization component that determines an optimal AEC mode for the execution of the CT scan based on the task information, the one or more localization images, and using one or more machine learning techniques, wherein the optimal automatic control mode maximizes image quality in a target region of the scan images generated from the CT scan and minimizes radiation dose to one or more organs included in the anatomical region.

[0012] The one or more machine learning techniques include employing a first deep learning network and a second deep learning network to determine the optimal AEC pattern from different candidate AEC patterns for the CT scan, the optimal AEC pattern maximizing the image quality in the target region and minimizing the radiation dose to the one or more organs. The first deep learning network includes one or more first machine learning models that estimate a expected organ dose (AID) representing the expected radiation dose to the one or more organs under different AEC patterns based on the one or more localization images. The second deep learning network includes one or more second machine learning models that estimate an expected metric of image quality in the target region and the background region of the scan images captured under the different AEC patterns based on the one or more localization images.

[0013] In various embodiments, in addition to determining the optimal AEC mode, the optimization component further determines the expected image quality in the target region under the optimal AEC mode, and the expected radiation dose to the one or more organs under the optimal AEC mode. In some embodiments, the computer-executable component further includes an interface component that facilitates receiving user input adjusting one or more parameters of the optimal exposure control mode to generate a modified AEC mode, and wherein, based on the received user input, the optimization component determines an updated expected image quality in the target region and an updated expected radiation dose to the one or more organs under the modified AEC mode. The interface component may also facilitate receiving user input identifying at least one of the following: the desired image quality of the target region and the desired radiation dose to the one or more organs, and wherein, based on the received user input, the optimization component determines a modified AEC mode that achieves the desired image quality and the desired radiation dose. The computer-executable component also includes a control component operatively coupled to an imaging device that performs the CT scan and controls the execution of the CT scan performed by the imaging device based on the optimal automatic exposure mode.

[0014] In some implementations, the elements described in the disclosed systems and methods may be embodied in different forms, such as computer-implemented methods, computer program products, or another form. Attached Figure Description

[0015] Figure 1 A diagram illustrates an exemplary end-to-end process for customizing ACE settings for CT scans based on patient and clinical tasks, according to one or more embodiments of the disclosed subject matter.

[0016] Figure 2 A high-level flowchart of an exemplary machine learning process for generating patient- and task-customized AEC settings, based on one or more embodiments of the disclosed subject, is presented.

[0017] Figure 3 An exemplary computational system is presented, based on one or more embodiments of the disclosed subject matter, which facilitates the customization of AEC settings for specific patient anatomy and clinical tasks.

[0018] Figure 4 Exemplary dose estimation models based on one or more embodiments of the disclosed subject matter are presented.

[0019] Figure 5 Exemplary dose estimation model architectures based on one or more embodiments of the disclosed subject matter are presented.

[0020] Figure 6 An exemplary pipeline for organ dose estimation based on CT data is shown, according to one or more embodiments of the disclosed subject matter.

[0021] Figure 7 Organ-specific dose reporting information generated from CT scans is shown, based on one or more embodiments of the disclosed subject matter.

[0022] Figure 8 Exemplary dose estimation components based on one or more embodiments of the disclosed subject matter are presented.

[0023] Figure 9 The training of a dose estimation model based on one or more embodiments of the disclosed subject is shown.

[0024] Figure 10 A high-level flowchart of an exemplary process for estimating organ dose based on one or more localization images in a selected AEC mode, according to one or more embodiments of the disclosed subject matter, is presented.

[0025] Figure 11 Exemplary quality estimation models based on one or more embodiments of the disclosed subject are presented.

[0026] Figure 12 The training of a quality estimation model based on one or more implementations of the disclosed topic is shown.

[0027] Figure 13 An exemplary quality estimation component is shown according to one or more embodiments of the disclosed subject matter.

[0028] Figure 14 An exemplary pipeline for organ noise estimation based on CT data is shown, according to one or more embodiments of the disclosed subject matter.

[0029] Figure 15 Organ-specific noise report information generated from CT scans is shown, based on one or more embodiments of the disclosed subject matter.

[0030] Figure 16 Exemplary CT scan images generated from CT volume data at different radiation dose percentages according to one or more embodiments of the disclosed subject matter are shown.

[0031] Figure 17 The generation of a differential image from an analog noisy image is illustrated according to one or more embodiments of the disclosed subject matter.

[0032] Figure 18 A graph is provided illustrating noise in a uniform region of a simulated dose measurement according to one or more embodiments of the disclosed subject matter.

[0033] Figure 19 A high-level flowchart of an exemplary process for estimating a measure of image quality in a target region and a background region based on one or more localization images in a selected AEC mode, according to one or more embodiments of the disclosed subject matter, is presented.

[0034] Figure 20 Exemplary optimized components based on one or more embodiments of the disclosed subject matter are shown.

[0035] Figure 21 The optimal AEC processing is shown, which determines the optimal AEC based on one or more localization images and a specific task, according to one or more embodiments of the disclosed subject.

[0036] Figure 22 An exemplary AEC optimization model based on one or more implementations of the disclosed subject is shown.

[0037] Figure 23 A high-level flowchart is presented of an exemplary computer implementation of a process for determining the optimal AEC, based on one or more embodiments of the disclosed subject matter.

[0038] Figure 24 A high-level flowchart is presented of an exemplary computer implementation of a process for determining the optimal AEC, based on one or more embodiments of the disclosed subject matter.

[0039] Figure 25 Another exemplary computational system is presented, based on one or more embodiments of the disclosed subject matter, which facilitates the customization of AEC settings for specific patient anatomy and clinical tasks.

[0040] Figure 26 A high-level flowchart of an exemplary computer implementation of a process for performing a CT scan using optimal AEC, based on one or more embodiments of the disclosed subject matter, is presented.

[0041] Figure 27 A block diagram is shown illustrating an exemplary, non-limiting operating environment in which one or more embodiments described herein may be facilitated. Detailed Implementation

[0042] The following specific embodiments are merely exemplary and are not intended to limit the implementation and / or application or use of the embodiments. Furthermore, they are not intended to be construed as being bound by any express or implied information set forth in the foregoing "Background Art" section, "Summary of the Invention" section, or "Detailed Description" section.

[0043] The disclosed subject matter relates to systems, computer-implemented methods, devices, and / or computer program products that facilitate the customization of AEC settings for specific patient anatomy and clinical tasks. Conventional AEC techniques still lack awareness of patient-specific anatomy affecting organ dosing, do not consider task-based image quality, and are prone to problems such as patient centering. The proposed method brings an understanding of how different AEC settings affect organ dosing and organ noise to optimize tube current distribution and other AEC parameters. This is achieved using deep learning techniques. Access to organ-level indicators makes this method more task- and patient-specific, and further provides additional knobs for tube current optimization, such as dosing and advanced image quality features.

[0044] The disclosed AEC optimization techniques aim to maximize relevant image quality for a fixed patient dose, or conversely, minimize the dose while achieving the desired level of image quality. In imaging protocols ranging from general to highly customized, the optimal AEC mode for a specific CT scan focuses image quality (and radiation) where it is most needed, while avoiding radiation-sensitive organs. Customization will be fully automated to seamlessly integrate with the workflow. This type of personalized imaging takes into account the variability of both patient and technology, which in turn reduces the variability of image quality to produce consistently high-quality images.

[0045] Even with complete volumetric information about the patient, this is not a simple task, especially if it is fast and automated. Current methods achieve patient-specific organ dose calculation by computationally generating dense dose maps (e.g., Monte Carlo simulations) and with the aid of manual organ segmentation. They further require manual input to locate areas where image quality should be prioritized and areas where image quality is less important. However, clinical workflows require determining the AEC (Area Center of Occurrence) based on localization images and in a fast and automated manner. Using novel AI methods and deep learning techniques, the disclosed AEC optimization strategy leverages information from localization images to rapidly and accurately predict patient-specific organ dose and image quality, thereby selecting the optimal AEC for clinical tasks.

[0046] To facilitate this purpose, the disclosed technique trains a first deep learning model to directly infer organ dose based on localization images and a prescribed AEC pattern. The training data uses matched CT volumes to calculate Monte Carlo dose maps from scans and segments organs (e.g., using one or more organ segmentation models) to provide organ dose and total effective dose. The disclosed technique further trains a second deep learning model to infer image quality based on localization images and the AEC pattern. Image quality reflects the expected image quality in different anatomical regions, allowing for the customization of an optimal AEC pattern to optimize radiation in the target region used for scanning and the resulting image quality, while minimizing radiation to background regions. The training data uses the same matched CT volumes to calculate noise in the projection and the resulting reconstructed image. In various embodiments, organ segmentation can also be used to map image quality metrics to different organs, where the target region and background region are based on different organs.

[0047] The disclosed technology further employs a first deep learning model and a second deep learning model to determine the optimal AEC pattern for a specific patient and clinical task, said optimal AEC pattern delivering the requested image quality with minimal dose. The disclosed technology further constructs an end-to-end optimization model that maps a given localization image and task to the optimal AEC in real time. This optimization model can be integrated into an imaging workflow to automatically determine the optimal AEC for CT scans in real time based on acquired localization images.

[0048] One or more embodiments will now be described with reference to the accompanying drawings, wherein the same reference numerals are always used to denote the same elements. In the following description, numerous specific details are set forth for purposes of explanation in order to provide a more thorough understanding of one or more embodiments. However, it will be apparent, in various cases, that one or more embodiments may be practiced without these specific details.

[0049] Figure 1 A diagram illustrates an exemplary end-to-end process 100 for customizing CT scans based on patient and clinical tasks, according to one or more embodiments of the disclosed subject matter. Process 100 is illustrated by way of the execution of a CT scan of a patient 110 via an imaging system including a CT scanner 108. The type of CT scanner can vary. For example, CT scanner 108 may include a single-energy scanner, a spectral CT scanner, and other types of CT scanners available in the industry.

[0050] According to a conventional CT imaging procedure, the patient is placed on the scanner table, and the anatomical structures of interest are positioned around the center of the CT gantry. One or more positioning images are then acquired (e.g., positioning image 102), and the operator marks the scan area to be acquired (e.g., start and end positions). The operator then selects the appropriate scanning protocol (e.g., contrast versus non-contrast) based on the clinical indications of the examination, and then selects specific scanning parameters to produce scan images of the required quality for the diagnostic task. Finally, the scan images are captured and reconstructed using a set of parameters that determine the characteristics of the images, such as image sharpness and field of view (FOV).

[0051] Given this context, process 100 begins by acquiring positioning images 102 of the patient 110 in association with the positioning of the patient 110 relative to the gantry of CT scanner 108. These positioning images 102 may include one or more low-resolution images of the ROI to be scanned, captured prior to capturing subsequent high-resolution (e.g., isotropic) three-dimensional (3D) image data of the patient's region of interest (ROI) and generating corresponding high-resolution reconstructed CT scan images. These positioning images 102 are typically used to position / align the scanner with the desired scan prescription plane relative to the capture of subsequent 3D image data. Positioning images 102 are also used to define the 3D imaging coverage volume or scan extent for capturing subsequent high-resolution scans. In this respect, positioning images 102 may have a lower resolution / image quality relative to the subsequently acquired high-resolution scan images, but depict the same or a wider region of interest. The terms “positioning image,” “positioner image,” “calibration image,” “trial image,” “reference image,” etc., are used interchangeably herein unless the context guarantees a specific distinction between the terms. Each of these terms refers to a captured medical image of the patient's anatomical region at a lower resolution relative to a captured CT scan image of the patient's anatomical region. In the illustrated embodiment, the positioning image 102 comprises two distinct images captured from different anatomical orientations (e.g., lateral and frontal) relative to the patient and the 3D imaging space. However, the number and orientation of the positioning images 102 can vary. For example, in some embodiments, a single positioning image may be used, while in other embodiments, two or more positioning images may be used.

[0052] At 104, one or more AI models are used to process the localization images to determine the optimal AEC mode 106 for patient 110 and the specific clinical task being performed on the CT scan. The optimal AEC 106 defines the parameter settings for the CT scan, which control the radiation dose delivered to the patient and the resulting image quality of the scanned image 114. In CT scans, there is always a trade-off between dose and image quality because images are acquired using two competing objectives: maximizing image quality and minimizing radiation dose. The AI ​​model processing at 104 customizes these AEC parameter settings for patient 110 to take into account the patient's specific anatomy, which can vary based on patient body size, stature, and the ROI being captured. The AI ​​model processing at 104 also customizes these AEC settings to consider one or more specific regions of the resulting high-resolution scan image that are more important for clinical assessment (e.g., detecting liver lesions on a non-contrast scan), which can vary depending on the patient and the ROI being scanned. Specifically, the optimal AEC mode is tailored to the patient's anatomy to account for the estimated radiation dose to organs within the ROI exposed to the patient 110 under different AEC settings, and how these different AEC settings affect image quality in one or more specific areas of clinical importance for assessment in the scanned images. In this regard, the AI ​​model processing at 104 involves determining the optimal AEC setting that focuses radiation exposure on areas most needed for image quality and clinical tasks, while avoiding or minimizing the amount of radiation delivered to radiation-sensitive organs.

[0053] In the illustrated implementation, the AI ​​model processing at 104 involves organ segmentation 1041, dose estimation 1042, and image quality estimation 1043. At a high level, organ segmentation 1041 involves using machine learning to automatically identify and segment organs included in the scanned ROI based on the localization image 102. Dose estimation 1042 involves using machine learning to predict how different AEC settings will affect the radiation dose to the organs. Image quality estimation 1043 involves using machine learning to predict how these different AEC settings and radiation doses will affect the image quality in different regions of the scanned image. The optimal AEC setting for the patient is then determined based on the organ dose and image quality estimates, which minimizes the radiation dose to the organs while also providing the desired level of image quality to specific target areas important for the clinical task. Optimal AEC setting evaluation and analysis may involve using one or more optimization functions as well as machine learning techniques. These machine learning and AI model processing techniques typically involve training one or more deep learning models to perform segmentation 1041, dose estimation 1042, and image quality estimation 1043 prior to their use in the clinical workflow depicted in process 100. See below for reference. Figures 2 to 26Further details are provided regarding these machine learning and AI model processing techniques.

[0054] The specific parameters used in processing the optimal AEC mode determined at 104 locations using an AI model can vary. AEC has many aspects, including the selection of tube voltage or tube current modulation (TCM), collimation, filtering, bowing, pitch, and starting angle. TCM refers to the modulation mode of the tube current during a CT scan. Conventional AEC algorithms use TCM to increase the tube current (i.e., the number of X-ray photons) for thicker body regions and decrease the tube current for thinner body regions. Because the patient is not a uniform cylinder, the end result is typically that the tube current oscillates up and down within a single rotation of the gantry, increasing on average through thicker body regions (e.g., shoulders and hips) and decreasing on average through thinner body regions (e.g., chest). In this respect, TCM can be adjusted to dynamically control the number of photons passing through each projection point of the patient.

[0055] TCM significantly affects radiation dose and corresponding image quality, and can change rapidly during multiple CT acquisition rotations and within a single gantry rotation. Therefore, in various implementations, the optimal AEC mode 106 defines the optimal TCM setting for CT scanning that minimizes organ dose and maximizes image quality in the target area. However, the disclosed techniques are not limited to optimizing only the TCM. In this respect, the optimal AEC 106 can reflect different acquisition parameter values ​​for various additional CT acquisition parameters, including but not limited to: tube voltage (or tube potential), collimation, filtering, arcuate junction, pitch (or pitch factor), Z-axis coverage, source-to-detector distance, source-to-rotation axis distance, starting angle, angle between projections, vertical translation between projections, vertical translation between projections, polar aperture and azimuth aperture, and voxel pitch (isotropic). The starting angle is particularly important for helical scans. For example, because helical scans have a highly non-uniform dose distribution, the starting angle of a helical CT scan is chosen so that the beam avoids direct exposure to radiation-sensitive organs. The optimal AEC 106 can also consider extending beyond the capabilities of current hardware with different flux field modulation strategies, such as laterally shifted bow junctions, or even dynamic bow junctions. In this regard, the term "AEC mode" is used herein to refer to a specific value (e.g., TCM) of one or more CT scanner acquisition parameters that can be adjusted in each scan and / or during a CT scan. The terms "AEC mode," "AEC setting," "AEC parameter," etc., are used interchangeably herein unless the context ensures a specific distinction between the terms.

[0056] Once the optimal AEC is determined based on the positioning image 102, the system (or operating technician) can configure the CT scanner to perform a CT scan using the parameter values / settings defined by the optimal AEC. This configuration is performed while the patient is positioned on the scanning stage, prior to the actual CT scan (e.g., the actual acquisition of 3D imaging CT data). In this respect, according to process 100, the optimal AEC can be determined and applied in real time based on the positioning image 102 at the start of the imaging session. At 112, the imaging device (e.g., CT scanner 108) can perform a CT scan with high-resolution image capture using the optimal AEC mode. The resulting scan image 114 will include the optimal CT scan image for the patient and clinical task. For example, the image quality (e.g., noise, resolution, etc.) of the scan image will be higher in the target area compared to the background area, where the target area represents a clinically relevant area for the CT scan purpose (e.g., an anatomical area as a clinical assessment target).

[0057] Figure 2 A high-level flowchart of an exemplary machine learning process 200 for generating patient- and task-customized AEC settings is presented according to one or more embodiments of the disclosed subject matter. In various embodiments, one or more aspects of process 200 may be incorporated into AI model processing 104 of process 100 to determine optimal AEC 106.

[0058] According to procedure 200, at point 202, a first deep learning network is trained to determine the expected organ dose based on localization images under different AEC patterns, where the localization images represent different patients, and where the expected organ dose reflects the expected radiation dose to organs included in the anatomical ROI represented in the localization images. This first deep learning model is referred to herein as the dose estimation model. See below for reference. Figures 4 to 11In more detail, this training process uses matched CT volumes and their localization images captured for each patient to calculate Monte Carlo dose maps from the scan and segment organs (e.g., using one or more organ segmentation models) to provide organ dose and total effective dose. Once trained, the input to the dose estimation model can include one or more localization images, the scan range (or ROI) of the CT scan, and the selected AEC mode. The output of the dose estimation model includes the estimated radiation dose to the organs exposed in the ROI under the selected AEC mode. In some implementations, the dose estimation model can be customized for specific anatomical ROIs (e.g., body parts). Using these implementations, multiple different dose estimation models can be generated for different anatomical ROIs and / or CT scan types (e.g., abdomen, body, abdomen / pelvis, pelvis, CAP (chest, abdomen, and pelvis), chest / abdomen, flow, head, head / neck, cervical spine, chest, extremities, thoracolumbar spine, lumbar spine, thoracic spine, and facial bones). At runtime, the system can select and apply an appropriate dose estimation model based on the ROI being scanned.

[0059] At position 204, a second deep learning model is trained to determine, based on the localization images, expected metrics of image quality in target and background regions of high-resolution scan images captured in different AEC modes for different tasks, where the target and background regions vary for different tasks. This second deep learning model is referred to herein as the quality estimation model. In one or more embodiments, the training data for the quality estimation model uses the same matched CT volume to compute projections and noise in the resulting reconstructed images across different AEC modes. The reconstructed images are further segmented to identify target and background regions, and the corresponding noise levels are mapped onto them. In various embodiments, target and background regions may be based on organ segmentation. Through these embodiments, organ segmentation can also be used to map image quality metrics (e.g., noise levels) to different organs. Once trained, the input to the quality estimation model may include one or more localization images, the scan extent (or ROI) of the CT scan, the selected AEC mode, and the selected task (e.g., indicating the desired target region). The output of the quality estimation model includes estimated metrics of image quality in the target and background regions under the selected AEC mode (e.g., measured as a function of noise levels and / or other quality assessment metrics). In implementations where the target region corresponds to one or more organs, the image quality metric can reflect the image quality of one or more organs in the scanned image. In some implementations, the quality estimation model can also be customized for a specific anatomical ROI (e.g., body part). Using these implementations, multiple quality estimation models can be generated for different anatomical ROIs and / or CT scan types (e.g., abdomen, body, abdomen / pelvis, pelvis, CAP (chest, abdomen, and pelvis), chest / abdomen, flow, head, head / neck, cervical spine, chest, limbs, thoracic and lumbar spine, lumbar spine, thoracic spine, and facial bones). At runtime, the system can select and apply an appropriate quality estimation model based on the ROI being scanned.

[0060] Once the first (e.g., dose estimation) and second (e.g., quality estimation) deep learning networks have been trained to the desired performance level, at point 206, both networks can be used to determine their optimal AEC pattern based on localization images for different patients and tasks. In this regard, based on the localization images, the optimal AEC pattern (e.g., the optimal ACE for each localization image / task combination) is determined for each clinical task, where the optimal ACE pattern maximizes image quality in the target region and minimizes radiation dose to the organ. In various embodiments, the optimization analysis performed at point 206 models the AEC using one or more optimization functions that use defined optimization criteria of dose and image quality metrics to balance competing objectives of dose and image quality. For example, in some embodiments, the optimization function may be configured to find the optimal ACE that maximizes the relevant image quality for a fixed patient dose. In other embodiments, the optimization function may be configured to find the optimal ACE that minimizes the dose while achieving the desired image quality level.

[0061] In some implementations, the optimization analysis described above can be performed at runtime (e.g., within clinical workflow process 100) to determine the optimal AEC for a new patient based on localization images. Using these implementations, the system iteratively applies dose estimation and quality estimation models to localization images under different AEC modes to determine how different AEC mode options affect the expected organ dose and image quality in the target region. The system can further use optimization functions and predefined or user-selected optimization criteria for dose and image quality metrics to automatically select the optimal ACE based on the model output.

[0062] Alternatively or concurrently, at 208, the system may train a third deep learning network to determine the optimal AEC pattern for each training data localization image / task combination evaluated at 206. In this regard, the optimal AEC determined at 206 for each localization image / task combination for the training data can be used as a ground truth for the training data to train the third deep learning network. Once trained, the input to the third deep learning network will include one or more localization images (e.g., with scan extents and / or ROIs indicated relative to the localization images) and the task. The third deep learning network will further output the optimal AEC. In some embodiments, the third deep learning network may also be configured to generate expected organ doses and expected image quality metrics for the target and background regions under the optimal AEC. At 210, this third deep learning network can be employed in a clinical workflow to determine the optimal AEC pattern 210 for a new CT scan based on one or more new localization images and a selected task from different clinical tasks. Utilizing these embodiments, the AI ​​model processing according to 104 of process 100 may involve applying only the third deep learning network.

[0063] Figure 3 An exemplary computing system 300, facilitating the customization of AEC settings for specific patient anatomy and clinical tasks, is presented according to one or more embodiments of the disclosed subject matter. Embodiments of the system described herein may include one or more machine-executable components embodied within one or more machines (e.g., embodied in one or more computer-readable storage media associated with one or more machines). When executed by one or more machines (e.g., processors, computers, computing devices, virtual machines, etc.), such components can enable the one or more machines to perform the operations described herein.

[0064] In this regard, computing system 300 provides an exemplary computing system including machine-executable components configured to perform one or more operations described in processes 100, 200, and additional processes described herein. The computer-executable components include a dose estimation component 302, one or more dose estimation models 304, a mass estimation component 306, one or more mass estimation models 308, an optimization component 310, one or more AEC optimization models 312, a receiving component 314, a training component 316, and an inference component 318. These computer / machine-executable components (and other components described herein) may be stored in memory associated with one or more machines. The memory may also be operatively coupled to at least one processor such that the components can be executed by the at least one processor to perform the operations described herein. For example, in some embodiments, these computer / machine-executable components may be stored in memory 324 of computing system 300, which may be coupled to processing unit 322 for its execution. The computing system 300 also includes a system bus 320, which communicatively and operatively couples a dose estimation unit 302, one or more dose estimation models 304, a quality estimation unit 306, one or more quality estimation models 308, an optimization unit 310, one or more AEC optimization models 312, a receiving unit 314, a training unit 316, an inference unit 318, a processing unit 316, and a memory. Examples of the memory, processor, and other suitable computer or computing-based elements may be found in [reference]. Figure 27 Find, and can be combined with implementation Figure 3 It may be used in accordance with one or more of the systems or components shown and described in the other accompanying drawings disclosed herein.

[0065] The deployment architecture of computing system 300 can vary. In some embodiments, computing system 300 may correspond to a single computing device (e.g., real or virtual). In other embodiments, computing system 300 may correspond to two or more separate, communication-coupled computing devices operating in a distributed computing environment. Using these embodiments, one or more of the following can be deployed at a separate computing device: dose estimation component 302, one or more dose estimation models 304, quality estimation component 306, one or more quality estimation models 308, optimization component 310, one or more AEC optimization models 312, receiving component 314, training component 316, and inference component 318. The separate computing devices may be communicatively coupled via one or more wired or wireless communication networks. In some embodiments, computing system 300 may include or be operatively coupled to a medical image scanning device (e.g., a CT machine, such as CT scanner 108, etc.), which performs CT scanning procedures in association with the capture of localization images and high-resolution CT scan images. Alternatively, the computing system 300 may include separate devices / machines (e.g., physical or virtual / cloud-based) communicatively coupled to the CT scanning equipment and / or an external image data storage system that provides training data. Various alternative deployment architecture variations may also be used.

[0066] refer to Figures 1 to 3 The dose estimation component 302 can perform reference Figure 1 and Figure 2 The described organ dose estimation operation. The mass estimation component 306 can perform the reference... Figure 1 and Figure 2 The described target region and background region scan image quality estimation operation, and the optimization component 310 can perform the reference... Figure 1 and Figure 2 The AEC optimization operation described.

[0067] For reference Figure 2As described, organ dose estimation, image quality estimation, and (optionally) AEC optimization can involve training separate machine learning models to perform the corresponding tasks. In the illustrated embodiment, these machine learning models correspond to one or more dose estimation models 304, one or more quality estimation models 308, and one or more AEC optimization models 312, respectively. To facilitate this purpose, the computing system 300 may include a training unit 310 that performs (or facilitates) the training and development of machine learning models using one or more supervised or semi-supervised machine learning techniques. The computing system 300 may also include an inference unit 312 that executes the trained models to generate their corresponding inference outputs. The computing system 300 also includes a receiving unit 308 that receives training data (and / or information for generating the training data) for training the respective models and input data for performing model execution after training. Reference Figures 4 to 10 Describe the features and functionality of dose estimation component 302 and one or more dose estimation models 304. (Reference) Figures 11 to 19 Describe the features and functionality of the quality estimation component 306 and one or more quality estimation models 308, and refer to Figures 20 to 26 Describe the characteristics and functionality of optimization component 310 and one or more AEC optimization models. For the sake of brevity, repeated descriptions of similar components used in the corresponding implementations are omitted.

[0068] Organ dose estimation

[0069] Figure 4 Exemplary dose estimation models 304 based on one or more embodiments of the disclosed subject matter are presented. Figure 4 As shown, the input to dose estimation model 304 may include one or more localization images 402 and an AEC pattern. The one or more localization images 402 may include localization images of the anatomical region captured in association with the execution of a CT scan of the patient's anatomical region. In some embodiments, the localization images 402 may provide different perspectives of the anatomical region (e.g., axial, sagittal, coronal, etc.). In other embodiments, a single localization image may be used. The AEC pattern 404 may define selected parameter values ​​for one or more AEC parameters in the CT scan. In various embodiments, the AEC pattern 404 may define a specific TCM scheme for the CT scan. Additionally or alternatively, the AEC pattern 404 may define specific parameter values ​​for one or more other AEC parameters, including but not limited to: tube voltage (or tube potential), collimation, filtering, arcuate junction, pitch (or pitch factor), Z-axis coverage, source-to-detector distance, source-to-rotation axis distance, starting angle, angle between projections, vertical translation between projections, projection, polar aperture and azimuth aperture, and vertical translation between voxel pitches (isotropic).

[0070] The output of dose estimation model 304 includes organ dose 406, which represents the estimated radiation dose absorbed by the patient's organs in connection with the execution of a CT scan using AEC mode 404. Dose estimation model 304 can also provide an estimate of the total effective dose in AEC mode. The whole-body effective dose is calculated by adding equivalent doses to all organs, each adjusted to account for the organ's sensitivity to radiation. The effective dose is expressed in millisieverts (mSv). For example, dose estimation model 304 can generate the effective dose (mSv) by combining organ doses through a weighted sum of organ doses using a known International Commission on Radiological Protection (ICRP) tissue weighting factor.

[0071] The specific organ evaluated by the dose estimation model is based on the specific organ included in the ROI of the scanned patient. In some embodiments, the ROI may take into account the entire anatomical region represented in the positioning image 402. In other embodiments, the ROI may include a subregion of the anatomical region depicted in the positioning image. Using these embodiments, information defining the subregion or scan range may also be provided as input to the positioning image 402. In this regard, the positioning image is typically used to position the patient relative to the CT gantry and define the scan range or body part to be captured in the CT scan. The positioning is almost always longer than the actual scan. Since our goal is to predict the dose in the actual scan, a scan range signal may be included in the input data to inform the dose estimation model 304 about the focal region in the input positioning. This additional information, taken into account the input positioning image, informs the model about where to view the dose estimation. As described in more detail below, when estimating organ dose based on the positioning image 402, the dose estimation model 304 will also implicitly learn organ segmentation to identify relevant organs.

[0072] In this example, the organs being evaluated include the lungs, liver, spleen, pancreas, and kidneys. These organs are typically scanned in conjunction with the execution of a CT scan of the abdomen. Therefore, in this example, the ROI includes the patient's abdomen. However, the dose estimation model 304 can be trained to evaluate any anatomical ROI and its associated organs. The associated organs for each anatomical ROI can be predefined. As described above, in some embodiments, different dose estimation models 304 can be customized for different anatomical ROIs. Using these embodiments, during inference mode, the inference component 318 can select and apply an appropriate dose estimation model 304 for a specific ROI represented in the localization image. Alternatively or additionally, the dose estimation model 304 can include a general model configured to evaluate any anatomical ROI. Using these embodiments, the dose estimation model 304 can be trained to identify associated organs in any ROI and customize organ dose estimates for those associated organs.

[0073] The type of machine learning model used for dose estimation model 304 can vary. In some embodiments, dose estimation model 304 may include one or more deep learning models, such as convolutional neural networks (CNNs), RESNET-type models, and other neural network models with different types of pooling techniques. For example, in one embodiment, dose estimation model 304 may employ a CNN architecture followed by fully connected layers and regression layers that output organ-specific dose and total effective dose. Other suitable types of machine learning models may include, but are not limited to, generative adversarial neural network (GAN) models, long short-term memory (LSTM) models, attention-based models, transformers, decision tree-based models, Bayesian network models, regression models, etc. These models can be trained using supervised, unsupervised, and / or semi-supervised machine learning techniques.

[0074] Figure 5 An exemplary dose estimation model architecture that can be used for dose estimation model 304 according to one or more embodiments of the disclosed subject matter is presented. According to this example, dose estimation model 304 includes a CNN-based model. In this example, dose estimation model 304 is used to predict the CT dose distribution at the organ of interest based on input frontal and lateral localization images of the patient's body within an indicated scan range under different AEC settings. However, it should be understood that the anatomical ROI represented in the input localization images can vary, and the number and type (e.g., orientation) of the input localization images can vary.

[0075] In the illustrated implementation, the dose estimation model 304 includes two modules: a feature learning module (FLM) 502 and multiple dose modules (DM) 5041-504. L+1 Where L corresponds to the number of different organs being assessed. FLM 502 is used to extract shared features, which are then used via individual organ-specific DMs. In addition to organ dose prediction, the model also estimates the patient's overall body dose via individual DMs. Therefore, the dose outputs for L different organs and the patient (body) are transmitted via L+1 DMs (e.g., DM1, DM2…DM). L+1 The predicted outputs are generated using the average dose.

[0076] In one exemplary implementation, the FLM module may include eight 3×3 convolutions, with feature maps of 16, 16, 32, 32, 64, 128, 256, and 512, respectively, and a stride of 2 in the second and fourth convolutions. Each convolution is followed by instance normalization and a leaky ReLU activation with a negative slope of 0.2. After every two convolutions, the feature maps may be halved via a 2×2 maxpooling operation. The CNN model may also include global average pooling layers and may draw features across channels. The extracted features are then shared on the DM. Each DM module may consist of two fully connected (FC) layers (512 and 1.0 neurons, respectively), a leaky ReLU activation with a slope of 0.2, and a final sigmoid to output the average dose prediction in a normalized scale.

[0077] According to this exemplary implementation, the dose is predicted based on a reference average dose d1 and a model at an organ labeled l (l∈L). Therefore, the loss of dose estimation model 304 is calculated according to Equation 1 below, where M represents the mini-batch size and the patient body dose is represented as (L+1).

[0078]

[0079] The reference organ dose values ​​correspond to the baseline truth information used to train the dose estimation model 304. In one or more embodiments, the disclosed techniques generate these reference values ​​from actual CT scan data.

[0080] Specifically, training data used to train and generate (e.g., test and validate) the dose estimation model 304 may include multiple CT scans for different patients and corresponding localization images of said patients captured prior to the CT scans. In some embodiments, the localization images used for the CT scans may include synthetically generated localization images. The CT scans (and their paired localization images) may depict the same anatomical ROI in embodiments that train the model for a specific anatomical ROI (e.g., abdomen). Alternatively or additionally, the CT scans (and their paired localization images) may depict a variety of different anatomical ROIs. The CT scan data provides a 3D volumetric representation of the anatomical ROI.

[0081] In one or more embodiments, GT (i.e., reference) organ dose values ​​can be generated by processing the CT scan volume to calculate dose maps of the CT scan volume under different AEC modes. In various embodiments, these dose maps can be generated using Monte Carlo (MC) simulations. MC simulations can generate and track particles at the voxel level and calculate the deposited energy for patient-specific estimation of the absorbed dose. Accelerated MC simulations have been introduced through computation using a graphics processing unit (GPU) known as MC-GPU, which can accurately simulate the physics of X-ray photon transport in voxelized geometry. The disclosed technique configures this MC-GPU simulation method to account for different AEC modes of CT scanners (e.g., existing and future CT scanners) and employs the method to calculate dose maps of the CT volume. In this regard, for a given protocol, the trajectory can be discretized into multiple source locations for each rotation. The source output is then modeled for each location, taking into account AEC parameter values, including tube voltage, collimation, filtering, and bow junction. The dose map for each source location is further normalized for each source photon and then scaled according to the tube current and rotation time (mAs) represented by said source location. This allows for flexible adjustment of the tube current modulation without having to rerun the MC simulation.

[0082] Organs are further segmented from these dose maps using one or more segmentation models to provide organ-specific GT dose values ​​and total effective GT dose. In this regard, organ dose model 304 will implicitly learn organ segmentation when estimating organ dose based on localization images. However, for training and testing purposes, training component 316 will require baseline ground truth organ segmentation for each CT volume, such as lung, heart, esophagus, stomach, intestine, liver, pancreas, spleen, kidney, bladder, gonads, muscle, skin, and bone. In some embodiments, the organ segmentation model may include a separate machine learning organ segmentation model (e.g., CNN) specific to a particular organ. Using these embodiments, the applied organ-specific segmentation model may be based on relevant organs included in the ROI. In other embodiments, the organ segmentation model may include an ROI-specific organ segmentation model suitable for segmenting all relevant organs included in a particular ROI. Still in other embodiments, the organ segmentation model may include a general organ segmentation model suitable for segmenting relevant organs from CT data from any ROI. Organ segmentation and dose maps are combined to estimate organ dose (mGy), and the organ doses are combined into an effective dose (mSv) by weighted summation of organ doses using an ICRP tissue weighting factor. The dose estimation model 304 is then trained using corresponding localization images of CT volumes to infer these organ doses and effective doses for different AEC patterns.

[0083] Figure 6An exemplary pipeline 600 for organ dose estimation based on CT data, according to one or more embodiments of the disclosed subject matter, is illustrated. Pipeline 600 begins with CT scan data 602, which corresponds to a CT volume comprising multiple scan slices. Using an MC-GPU simulation method, the CT scan data 602 is used to calculate a voxel density map 604, which is then used to calculate a body dose map 606. The body dose map 606 is then segmented using an organ segmentation model to generate a corresponding organ dose map 608. The organ dose map 608 is further used to generate an organ dose report 610, which provides individual organ doses and effective doses for a specific AEC pattern.

[0084] according to Figure 6 In the illustrated implementation, for each scan slice, the input CT scan data 602 is interpolated and resampled to 128 × 128 inches. A voxel density map 604 is created as a voxelized patient phantom by mapping CT Henle units (HU) to water equivalent density, where isotropic voxels are 4 × 4 × 4 mm according to Equation 2. 3 .

[0085]

[0086] The MC-GPU process is further modified to include bow-junction filtering and the anode-root effect. The anode-root effect causes photon intensity variations at azimuth angles that can be modeled as probability functions. The bow-junction model is based on fundamental material decomposition, allowing the MC-GPU to take material attenuation coefficients and material thicknesses at different sector angles as input. Using these implementations, once the source spectrum is input, the filtered spectra and relative photon intensity distributions for all sector angles can be calculated. Therefore, photon orientation and energy can be sampled accordingly.

[0087] The modified MC-GPU enables the analysis and computation of the baseline ground truth of detected images given an arcuate knot model, a root effect model, and input spectra. For individualized patient organ doses, a phantom is generated from patient CT voxel data. In this regard, the input phantom data for MC simulations can include spatial maps of both material type and mass density. Density mapping is performed following a piecewise linear curve that defines the density of the mixture of water and bone.

[0088] exist Figure 6 In the illustrated embodiment, the slices from top to bottom simulate a helical scan with a pitch factor of 0.99, an 80 mm Z-axis coverage, 24 views per rotation, a 120 kVp spectrum, and a constant tube current. However, it should be understood that these AEC parameter values ​​are merely exemplary, and the disclosed techniques can be applied to generate organ dose maps for different TCM protocols and other AEC parameters. Figure 6 In the implementation shown, the MC-GPU dose calculation is repeated, where N = 4 starting angles (θ(i)) are evenly spaced and averaged to obtain the dose map according to Equation 3.

[0089]

[0090] In Equation 3, ψ represents the set of all parameters used to configure the MC-GPU simulation. The dose is reported as eV / g / photon and can be scaled to mGy (1 eV / g / photon = 3.79 mGy / 100 mAs) using scanner-specific calibration (e.g., CTDI measurement). A threshold t is used with respect to the voxel geometry V. 空气 A patient body mask with a diameter of 0.1 is used to shield against air to obtain a body dose map 606. Therefore, the final dose map D can be represented according to Equation 4.

[0091] D = D 平均 ·(V>t 空气 )

[0092] Equation 4.

[0093] The CT scan 602 was then segmented using a segmentation model employing a 3D context encoder U-Net network. The context encoder utilizes dilated convolutions at varying rates within the encoder network, enabling the capture of information over longer distances compared to the standard U-Net. The decoder of the segmentation model employs residual multi-kernel pooling, which performs max pooling at multiple fields of view (FOV). The segmentation model was trained separately for L different organs. The encoder and decoder networks utilized 5 convolutional blocks, followed by downsampling and upsampling, respectively. Softmax activation followed the final convolution. The segmentation model was trained using focus classification cross-entropy loss according to Equation 5, where i iterates over the number of channels and t iterates over the number of voxels in the image patch.

[0094]

[0095] Calculate the weight graph w according to Equation 6. t To emphasize voxels near the reference boundary, where d is the Euclidean distance to the nearest label, and a and b are weighting factors. Weighting factors a = 2.0 and b = 0.5 were chosen for the experiment. The weighted graph was further added to the loss function equation 5.

[0096] w t = a·exp(-b·d t )+1

[0097] Equation 6.

[0098] To calculate organ-specific doses, the segmentation mask predicted by the model is... Interpolation and resampling are performed to align the voxel coordinates with the CT voxel geometry prepared for the MC-GPU. Organ-specific dose map O l , l∈L (e.g., organ dose map 608) is derived from body dose map D (e.g., body dose map 606) according to Equation 4 and segmentation mask according to Equation 7. generate.

[0099]

[0100] Then, organ-specific reference dose distribution was determined, and the mean dose (d) was used to determine the organ-specific reference dose distribution. l =mean(O l )) and average body dose (d L+1 =mean(O L+1 )) to represent.

[0101] Figure 7 The diagram illustrates organ-specific dose reporting information generated from CT data using the techniques described above. In various embodiments, the organ dose report 610 may include… Figure 7 The diagram shows organ-specific dose information for each segmented organ within the corresponding segmented organ. In this example, the organ dose report data shows six different organs associated with the body and abdominal scan, including the lungs, kidneys, liver, bladder, spleen, and pancreas. The organ dose report data plots the absorbed dose in mGy at different voxel locations and provides the average dose at all voxel locations.

[0102] Figure 8 An exemplary dose estimation component 302 according to one or more embodiments of the disclosed subject matter is presented. For the sake of brevity, repeated descriptions of similar elements used in the respective embodiments are omitted.

[0103] In some embodiments, the dose estimation component 302 may include a training data generation component 802 to generate baseline true organ and effective body dose values ​​based on the MC simulation and organ segmentation methods described above. In this regard, the training data generation component 802 may receive paired training localization images and CT volumes 801 from different patients. In embodiments where the dose estimation model 304 is trained for a specific anatomical ROI, the paired training localization images and CT volumes 801 may depict the same anatomical ROI. Using the CT volumes, the training data generation component 802 may generate baseline true organ dose data 803 using the MC simulation and organ segmentation methods described above. For each training localization image in the training localization images (or where the input localization images comprise two or more, e.g., a group of two or more embodiments of lateral and frontal views), this GT organ dose data 803 may include organ dose and effective dose for one or more AEC patterns. To facilitate this purpose, the training data generation component 802 may include a CT dose map generation component 804, an organ segmentation component 806, and one or more organ segmentation models 808.

[0104] In this regard, using the aforementioned MC-GPU simulation method and CT volume 801, the GT dose map generation component 804 can calculate the dose map for each CT scan volume within the CT scan volume. The calculated dose map for each CT volume can provide an estimated 3D body dose map (e.g., body dose map 606) for the anatomical ROI under different AEC modes. For example, in one or more embodiments, for a given protocol, the CT dose map generation component 804 can discretize the photon trajectory into multiple source locations for each rotation. The CT dose map generation component 804 can further model the source output at each location, taking into account tube voltage, collimation, filtering, and bow junction. The CT dose map generation component 804 can further normalize the body dose map for each source location for each source photon. The CT dose map generation component 804 can further scale the body dose map based on the tube current and rotation time (mAs) represented by the source location. The CT dose map generation component 804 can further adjust the TCM for different AEC modes to generate corresponding body dose map values ​​without having to rerun the Monte Carlo simulation.

[0105] Then, the organ segmentation component 806 can apply one or more organ segmentation models 808 to the body dose map to segment relevant organs included in the ROI (e.g., which may be predefined) and generate organ-specific doses for different AEC modalities and the total effective body dose per CT scan volume. As described above, in some embodiments, the organ segmentation model 808 may include a multi-organ segmentation model adapted to segment all relevant organs of a specific anatomical ROI or any anatomical ROI. In other embodiments, the organ segmentation model 808 may include a separate model for each organ of interest.

[0106] One or more dose estimation models 304 can then be trained using the GT organ dose data 803 and paired localization images (e.g., by training component 316) to infer the doses to these organs and the total effective dose, such as Figure 9 As shown.

[0107] in this regard, Figure 9 The training of a dose estimation model based on one or more embodiments of the disclosed subject is illustrated. For example... Figure 9 As shown, dose estimation model training involves training a dose estimation model to predict organ doses 904 (which may include the total effective dose) for a selected AEC pattern based on one or more localization images using organ doses generated from the aforementioned baseline ground truth CT. Reference Figure 3 and Figure 9 In various implementations, the training component 316 may be performed using one or more supervised or semi-supervised machine learning training, testing, and validation processes. Figure 9 The dose estimation model shown is trained.

[0108] like Figure 9 As shown, training data 900 can include multiple different training datasets, each labeled as 902. 1-N The number N of training datasets can vary. Each training dataset includes three components indicated by the letters A (component 1), B (component 2), and C (component 3), respectively. Specifically, the first component A includes one or more localization images. In some implementations, the localization images may also include or otherwise be associated with information defining the scan range. The second component B includes known / defined AEC patterns, and the third component C includes GT organ doses (and effective doses) determined for said AEC patterns. As discussed above, these GT organ doses are generated using CT volume data paired with localization images A and MC simulations and organ dose maps generated for the AEC patterns. In this regard, with respect to an exemplary training dataset 9021, the input to dose estimation model 304 includes AEC pattern 901. 1-B and one or more positioning images 901 1-AThen, the dose estimation model 304 was trained to use the AEC model 901. 1-C The estimated organ dose 904 is generated from the GT organ dose.

[0109] It should be understood that the collective training dataset is 902. 1-N This represents one or more localization images and paired CT volumes for different patients and different AEC patterns. Additionally, there is a collective training dataset of 902. 1-N This can include groups of training datasets that consist of the same localization images (e.g., for the same patient) but paired with different AEC patterns generated from the same paired CT volume data for GT organ dose (and effective dose). When training dose estimation model 304 for a specific anatomical ROI (e.g., abdomen), the collective training dataset 902... 1-N The same anatomical ROI will also be depicted.

[0110] In one or more implementations, for organs directly exposed to the beam, the target accuracy of the dose estimation model can be set within 5%, as this is a typical error even for MC dose calculations. Organs not directly exposed but receiving dose from scattering may have a higher percentage of error, although their absolute dose will be very low. The total effective dose estimate can be set to be accurate within 1%, as it averages the error across all organs. To improve network performance and robustness, the training data can also be augmented by randomly changing the 3D location of the CT volume (and generating corresponding localization images) and by using different AEC modes.

[0111] Figure 10 A high-level flowchart of an exemplary process 1000 for estimating organ dose based on one or more localization images in a selected AEC mode, according to one or more embodiments of the disclosed subject matter, is presented. Process 1000 provides an exemplary procedure that can be performed by a computing system 300 using a receiving unit 314, a training unit 316, a dose estimation unit 302, and an inference unit 318. For brevity, repeated descriptions of similar elements employed in corresponding embodiments are omitted.

[0112] According to process 1000, at 1002, a system operatively coupled to the processor (e.g., system 300, etc.) can be trained (e.g., by dose estimation component 302 using...) Figure 9The training component 316) of the deep learning network (e.g., dose estimation model 304) determines the expected organ dose, representing the expected radiation dose to one or more organs (or those absorbed by them) in the anatomical region under different AEC modes of a CT scan, based on one or more localization images of the anatomical region captured from different patients. As described above, the training component 316 can be trained using GT organ doses determined for different patient and AEC mode combinations. In some embodiments, prior to training at 1002, process 1000 may also include generating GT organ doses from matching CT volumes used for localization images by simulating 3D dose maps and subsequent organ segmentation of 3D dose maps using MC simulations under different AEC modes.

[0113] At 1004, once the deep learning network (i.e., dose estimation model 304) has been trained, tested, and validated, the system can (e.g., by using inference component 318 by dose estimation component 302) apply the deep learning network to one or more newly localized images of the captured anatomical regions of the patient to determine the estimated organ dose to one or more organs in the selected AEC mode.

[0114] Image quality estimation

[0115] The organ dose estimation techniques described above provide tools for rapidly estimating organ doses for a given AEC pattern directly from localization images. In this section, we evaluate the impact of different AEC patterns and / or their resulting organ doses on image quality. For a given task, such as liver lesion detection on contrast-enhanced scans, the disclosed techniques seek to maximize task-relevant image quality, such as contrast-to-noise ratio (CNR), while maintaining adequate image quality elsewhere. To facilitate this, the disclosed techniques train a second deep learning network (e.g., one or more quality estimation models 308) to predict image quality in target and background regions based on localization images of the patient. These techniques utilize the same training dataset of paired localization images and CT volumes used above for organ dose estimation, using CT volumes to establish ground truth benchmarks for image quality (IQ) in both the target (IQ-target) and background (IQ-background) regions. Quantifying the image quality in these regions separately allows CT scanners to be selectively configured with AEC patterns that provide high-quality images in the target region while providing baseline image quality in the background region.

[0116] Figure 11 Exemplary quality estimation models 308 based on one or more embodiments of the disclosed subject matter are presented. For the sake of brevity, repeated descriptions of similar elements used in the respective embodiments are omitted.

[0117] like Figure 11As shown, the input to the image quality estimation model 308 may include one or more localization images 402, a selected AEC mode 404, and a selected task 1102. Task 1102 controls or otherwise defines the target region and background region of the CT scan image. In various embodiments, task 1102 may include a clinical task for performing the CT scan. In this regard, the specific organs, anatomical regions, etc., that are clinically important for assessment captured in the scan image may vary depending on the clinical indications of the CT example. Alternatively or additionally, task 1102 may include a specific post-processing task to be used to process the CT scan image, wherein the post-processing task is facilitated by a higher level of image quality in the target region relative to the background region. For example, in some embodiments, the post-processing task involves processing the scan image using one or more automated medical image inference models suitable for performing automated image processing tasks such as, but not limited to, disease / symptom classification, disease region segmentation, organ segmentation, disease quantification, disease / symptom staging, risk prediction, temporal analysis, anomaly detection, anatomical feature characterization, medical image reconstruction, medical image enhancement, etc.

[0118] In one or more embodiments, task 1102 may be selected from a defined set of different tasks. In many exemplary embodiments, the target region and background region defined for different tasks are based on organs included in the anatomical ROI. For example, for a particular task, the target region may include one or more specific organs (e.g., the liver in relation to the detection of liver lesions), while other organs and anatomical features included in the ROI may be defined as background regions. However, the disclosed techniques are not limited to defining target and background regions in scanned images solely based on organs. In this regard, target and background regions may be based on other defined anatomical landmarks, specific scan lines, specific prescription planes, specific points / locations in the scanned image, etc. The term target region is used herein to refer to the entirety of the scanned image portion considered as the target region. In this regard, "target region" may include two or more separate regions of the same CT scan image. Similarly, "background region" may include two or more separate regions of the same CT scan image.

[0119] The output of the quality estimation model 308 includes one or more image quality metrics or indices that represent a measure of the expected image quality in the target region (IQ-target) and background region (IQ-background) of a CT scan image to be acquired in the selected AEC mode 404. In some embodiments, the image quality metric may be based on the image noise level determined by segmenting the target and background regions and distinguishing morphological variations and pixel differences due to photon noise. For example, in some exemplary embodiments, such as... Figure 11As shown, expected image quality can be provided based on organ noise levels (measured in HU) and overall noise levels (e.g., patient). Alternatively, image quality metrics can be based on one or more other factors, including but not limited to: average noise, maximum noise, CNR, signal-to-noise ratio (SNR), noise power spectrum (NPS), detection quantum efficiency (DQE), and detectability index. Figure 11 In the example shown, the target region and background region are not explicitly identified, but are understood to be based on a specific organ. For example, depending on task 1102, the target region may include all organs or a subset of organs (one or more organs). Similarly, the background region may include a subset of organs (e.g., one or more). Still in another example, the patient column may represent the background region and correspond to the average noise level of the excluding organ region of the scanned image.

[0120] Similar to organ dose estimation, in addition to task 1102, specific target and background regions can vary based on specific organs included in the scanned patient's ROI. In some embodiments, the ROI may take into account the entire anatomical region represented in the localization image 402. In other embodiments, the ROI may include a subregion of the anatomical region depicted in the localization image. Using these embodiments, information defining subregions or scan extents can also be provided as input to the localization image 402. As described in more detail below, when estimating the image quality of target and background regions based on the localization image 402, the quality estimation model 308 implicitly learns the relevant target and background regions for each task and ROI, and how different AEC modes affect image quality in the CT volume. Similar to dose estimation model 304, in some embodiments, the quality estimation model 308 can be customized for a specific anatomical ROI. Using these embodiments, different quality estimation models 308 can be generated for different ROIs and selectively applied based on the scanned ROI.

[0121] The type or more of the machine learning models used for one or more quality estimation models 308 can vary. In some embodiments, one or more quality estimation models 308 may include one or more deep learning models, such as CNN, RESNET-type models, and other neural network models with different types of pooling techniques. In some embodiments, the quality estimation model 308 may employ a pooling technique corresponding to... Figure 5The CNN architecture is shown. In another example, one or more dose estimation models can employ a CNN architecture, which includes fully connected layers and regression layers that then output estimated quality metrics for the target and background regions. Other suitable types of machine learning models may include, but are not limited to, GANs, LSTMs, attention-based models, transformers, decision tree-based models, Bayesian network models, regression models, etc. These models can be trained using supervised, unsupervised, and / or semi-supervised machine learning techniques.

[0122] Figure 12 The training of a quality estimation model based on one or more implementations of the disclosed topic is illustrated. Figure 12 As shown, training the quality estimation model involves training a dose estimation model to predict one or more metrics of image quality in the target and background regions 1204 (e.g., target and background image quality metrics generated using benchmark ground truth CT based on IQ-target and IQ-background of one or more localized images for a selected AEC mode and task). The techniques used to generate these benchmark ground truth image quality metrics are discussed in more detail below. References Figure 3 and Figure 12 In various implementations, the training component 316 may be performed using one or more supervised or semi-supervised machine learning training, testing, and validation processes. Figure 12 The quality estimation model shown is trained.

[0123] like Figure 12 As shown, training data 1200 can include multiple different training datasets, each labeled as 1202. 1-NThe number N of training datasets can vary. Each training dataset includes four components indicated by the letters A (component 1), B (component 2), C (component 3), and D (component 4), respectively. Specifically, the first component A includes one or more training localization images. These training localization images may include the same localization images used to train the dose estimation model 304 and / or additional or alternative localization images. In some embodiments, the localization images may also include or otherwise be associated with information defining the scan range. The second component B includes known / defined AEC patterns. The known / defined AEC patterns may also include different AEC patterns used to train the dose estimation model 304. The third component C includes GT target and background quality indices determined for the AEC patterns. As discussed below, these GT quality indices may also be generated using CT volume data paired with training localization images, which may include the same localization image-CT volume data pairs used to generate the GT training data for the dose estimation model 304 (e.g., paired training localization images and CT volumes 801). The fourth component D includes a specific task that controls / defines the relevant target and background regions.

[0124] In this regard, regarding an exemplary training dataset 12021, the input to the quality estimation model 308 includes task 1202. 1-D AEC mode 1202 1-B and one or more training localization images 1201 1-A Then, a quality estimation model 1204 was trained to use the ground truth objective and background quality metric 1202. 1-C Based on training localization image 1202 1-A Generate / Estimate AEC Model 1202 1-B And Task 1202 1-D The target and background image quality metrics 1204. The quality estimation model 308 will implicitly learn the relevant task and background regions of the task, as well as how AEC affects image quality in the CT volume. In various implementations, the training component 316 can train the quality estimation model 308 to predict noise levels in these regions to within 1.0 HU, which will provide the level of accuracy required to optimize AEC.

[0125] It should be understood that the collective training dataset is 1202. 1-N This represents one or more localization images and paired CT volumes for different patients and different combinations of AEC modalities and tasks. In this regard, the collective training dataset is 1202. 1-NThis can include groups of training datasets that consist of identical localization images of the same patient and task, but paired with ground truth (GT) targets and background quality indices for different AEC patterns generated from the same paired CT volume data. Additionally, there is a collective training dataset of 1202. 1-N This can include groups of training datasets that consist of identical localization images of the same patient and AEC pattern, but paired with ground truth (GT) targets and background quality metrics for different tasks generated from the same paired CT volume data. When training the quality estimation model 1204 for a specific anatomical ROI (e.g., abdomen), the collective training dataset 1202... 1-N The same anatomical ROI will also be depicted.

[0126] Figure 13 An exemplary mass dose estimation component 306 according to one or more embodiments of the disclosed subject matter is presented. For the sake of brevity, repeated descriptions of similar elements used in the respective embodiments are omitted.

[0127] In some implementations, the quality estimation component 306 may include a training data generation component 1302 to generate ground truth (GT) target and background quality data 1303 for training one or more quality estimation models 308. In this regard, the GT target and background quality data 1303 may correspond to GT target and background quality training data included in training data 1200 for different AEC and task combinations. Figure 8 As shown, in some embodiments, the training data generation component 1302 may employ the same paired training localization images and CT volumes 801 used to generate GT organ dose data 803 to generate GT target and background quality data 1303. In embodiments where the quality estimation model 308 is trained for a specific anatomical ROI, the paired training localization images and CT volumes 801 may depict the same anatomical ROI.

[0128] In one or more embodiments, the training data generation component 1302 can generate GT target and background quality data 1303 using noise maps generated from CT volume data corresponding to a body dose map (e.g., body dose map 606). These noise maps can reflect different noise distributions in the expected CT scan images generated under different AEC modes. These noise maps can be further segmented into target and background regions, and corresponding noise levels can be mapped to the corresponding target and background regions, such as... Figure 14 As shown. To facilitate this purpose, the training data generation component 1302 may include a simulation component 1302, a segmentation component 1306, and one or more anatomical segmentation models 1308.

[0129] in this regard, Figure 14An exemplary pipeline 1400 for organ noise estimation based on CT data, according to one or more embodiments of the disclosed subject matter, is shown. Pipeline 1400 begins with CT scan data 1402 corresponding to one or more CT scan slices constituting a CT scan volume. Reference Figure 13 and Figure 14 In one or more embodiments, simulation component 1302 may use CT scan data 1402 to calculate a body noise map 1404, which represents the noise distribution in the CT scan data 1402 under a specific AEC mode. Segmentation component 1306 may further segment the CT scan data 1402 using one or more anatomical segmentation models 1308 to facilitate the identification of target and background regions in the CT scan data 1402, as shown in the anatomical segmentation map 1406. For example, in various embodiments, one or more anatomical segmentation models 1308 may include or correspond to one or more organ segmentation models (e.g., organ segmentation model 808) for performing organ dose mapping on CT scan data. Using these embodiments, segmenting the anatomical segmentation map can segment organ anatomical features in the CT scan data 1402. Additionally or alternatively, one or more anatomical segmentation models 1406 may include segmentation models adapted to segment various different anatomical landmarks or features (e.g., other than organs) in the CT scan data 1402. For example, other anatomical landmarks or features may include bone, soft tissue, lesions, artifacts, scanning planes, specific locations or points in CT scan data 1402, etc.

[0130] The training data generation component 1302 can further combine the anatomical segmentation map 1406 with the body noise map 1404 to obtain an anatomical-specific noise map 1408. The anatomical-specific noise map 1408 provides a mapping of body noise levels to segmented anatomical regions identified in the anatomical segmentation map. In this respect, using the body noise map and the anatomical segmentation map 1406, the training data generation component 1302 can determine the average noise level associated with each segmented anatomical region within the segmented anatomical region, including specific markers and background regions utilizing those markers. Based on this mapping, the training data generation component 1302 can generate target and background noise report data 1410 that identifies the average noise levels associated with the target and background regions, where the average noise levels reflect a specific AEC pattern. Because the target and background regions can vary based on the task, the target and background noise report data can also vary depending on the task. For example, in some embodiments where the segmented anatomical region includes organs in the ROI, the target and background noise report data 1410 can identify the average noise level of each segmented organ within the segmented organ and the average noise level of regions in the CT scan data that do not include organs. In these implementations, the target region may include all organs or a subset of organs, and the background region may include regions excluding all organs or a subset of organs, depending on the task. In other implementations, the target / background region noise reporting data 1410 may include a specific noise level for each anatomical segmentation marker / feature represented in the anatomical segmentation map 1406 and an average noise level for the remaining region (e.g., excluding anatomical segmentation regions). Based on this data, the target and background regions can be selectively defined for a task under the same AEC mode. It should be understood that the same anatomical segmentation map 1406 can be mapped to different body noise maps generated from the same CT scan data 1402 used for different AEC modes. In this respect, as the body noise map changes according to the modeled AEC mode, the noise values ​​of the selected target and background regions, as defined based on the anatomical segmentation map 1406, will also change.

[0131] Figure 15 An exemplary organ-specific noise report information generated from CT data using the above-described technique is shown. Figure 15 The CT organ noise reporting data shown provides exemplary noise reporting data that can be included in the target and background region noise reporting data 1410. According to this example, organ anatomy segmentation is used to generate... Figure 7The same organ segmentation model 808 is used to segment relevant organs within the ROI for the organ dose reporting data shown. In this regard, the average noise level for each organ within an individual organ is calculated under a specific AEC mode. In this example, the CT organ noise dose reporting data shows the body (e.g., areas that may reflect overall image quality and / or exclude organs) and six different organs associated with an abdominal scan, including the lungs, kidneys, liver, bladder, spleen, and pancreas. The organ dose reporting data plots the noise levels (e.g., in HU) at different voxel locations and provides the average noise level at all voxel locations for each part (e.g., organs and body).

[0132] Refer again Figure 13 and Figure 14 The techniques used by the simulation component 1302 to generate body noise maps 1404 for different AEC modes from CT volume data can vary. In some embodiments, the simulation component 1302 can use simulation techniques to generate body noise maps 1404 from CT volume data in different AEC modes representing different TCM and / or other different AEC parameter values. In this regard, associated with performing CT scans using TCM, the image quality in the projection changes with the modulation of the tube current, which affects the image quality in the reconstructed scan images. In some embodiments, to establish GT target and background quality data 1303, the simulation component 1302 can determine image quality metrics (e.g., noise and / or other image quality metrics) for different AEC modes by simulating projections at different tube current levels (and therefore different dose levels) using CT volume and generating corresponding reconstructed scan images. For each projection, the simulation component 1302 can insert noise corresponding to spatial variations in tube current and beam (e.g., due to bowing), thereby taking into account quantum noise and electronic noise, such as... Figure 16 As shown. This requires an accurate CT scanner model, including its source spectrum and detector response.

[0133] in this regard, Figure 16Exemplary CT scan images generated from CT volumetric data at different radiation dose percentages are shown. Upper scan image set 1601 corresponds to a scan image captured at 100% dose, while lower scan image set 1602 shows a scan image generated at 10% dose. In various embodiments, the simulation component 1302 can generate different sets of reconstructed scan images from the original CT volumetric data at different dose levels ranging from 100% to 10% by inserting noise into the raw CT data. In this example, upper scan image set 1601 corresponds to a reconstructed scan image without noise, and lower scan image set 1602 corresponds to a reconstructed scan image simulated without 90% noise. As can be seen by comparing the corresponding scan image sets, the high-dose image 1601 at 100% dose and without added noise has better tissue visualization due to its lower noise compared to the low-dose image 1602 with added noise. The scan images in both sets are marked with boxes to indicate exemplary regions of interest.

[0134] According to these embodiments, the simulation component 1302 can generate reconstructed scan images from simulation projections with and without added noise, such as Figure 16 As shown, the simulation component 1302 can generate multiple noise implementations at different dose levels (e.g., 100% dose, 75% dose, 50% dose, 25% dose, 5% dose, etc.) to better estimate the image noise level. For filtered backprojection reconstruction, the noise should be highly realistic due to its quasi-linear nature. The simulation component 1302 can also employ iterative reconstruction (e.g., ASIR) and deep learning image reconstruction (DLIR) techniques to facilitate the generation of simulated dose levels.

[0135] Simulation component 1302 can further determine, based on the difference between noise projections with and without added noise, the expected image noise distribution value of scan images generated under a specific AEC mode and / or at a specific dose level associated with the AEC mode, such as... Figure 17 As shown.

[0136] in this regard, Figure 17 The diagram illustrates the generation of differential images from a high-dose image at 100% dose and a simulated low-dose image. A first differential image 1701 depicts a differential noise image generated based on the difference between the high-dose image at 100% dose and another scan image simulated at 90% dose. A second differential image 1702 depicts another differential noise image generated based on the difference between the high-dose image at 100% dose and another scan image simulated at 10% dose. The differential images show the added noise, but the original noise at full dose is not shown.

[0137] In one or more embodiments, simulation component 1302 may use empirical curve fitting to estimate the noise level from the difference images (e.g., first difference image 1701, second difference image 1702, etc.) to estimate the standard deviation at any dose. To facilitate this purpose, simulation component 1302 may first calculate a standard deviation (stdev) plot of the difference images at two simulated doses according to Equation 8 (this helps to capture behavior within the dose range).

[0138] σ Diff1 = stdev image Im of the difference image D1 -Im 1.0 ;as well as

[0139] σ Diff2 = stdev image Im of the difference image D2 -Im 1.0

[0140] Equation 8.

[0141] Then, the simulation component 1302 can estimate the standard deviation of any dose D≤1 according to Equation 9.

[0142]

[0143] Figure 18 A graph is provided showing the measured noise across a uniform region according to the technique described above for all simulated doses. Graph 1801 plots the noise levels of multiple different image regions as a function of dose (e.g., measured as a function of the standard deviation of the difference image), where each line corresponds to a different region. Graph 1802 plots normalized or arbitrary noise levels normalized to 100% dose.

[0144] Figure 19 A high-level flowchart of an exemplary process 1900 for estimating a measure of image quality in a target region and a background region based on one or more localization images in a selected AEC mode, according to one or more embodiments of the disclosed subject matter, is presented. Process 1900 provides an exemplary process that can be performed by a computing system 300 using a receiving unit 314, a training unit 316, a quality estimation unit 306, and an inference unit 318. For brevity, repeated descriptions of similar elements employed in corresponding embodiments are omitted.

[0145] According to process 1900, at 1902, the system operatively coupled to the processor (e.g., system 300, etc.) can be trained (e.g., by the quality estimation unit 306 using...) Figure 12The training component 316) of the deep learning network (e.g., quality estimation model 306) determines expected metrics (e.g., noise level and / or one or more other image quality metrics) of image quality in target and background regions of scan images captured in different AEC modes, based on localization images captured for different patients, where the target and background regions are used for different clinical tasks. In some embodiments, prior to training at 1902, process 1900 may also include generating GT training and background image quality metrics from the matched CT volume used for localization images by segmenting 3D dose-map anatomy into target and background regions using simulated image quality maps from different AEC modes and subsequently for different clinical tasks.

[0146] At 1904, once the deep learning network (i.e., the quality estimation model 308) has been trained, tested, and validated, the system can (e.g., by the quality estimation component 306 using the inference component 318) apply the deep learning network to one or more newly localized images of the captured anatomical regions of the patient to determine the estimated measure of image quality in the target and background regions of the CT scan images based on the selected AEC pattern and clinical task.

[0147] AEC optimization

[0148] The previous section provided deep learning techniques that enable the individual estimation of patient-specific organ doses under different AEC settings and how different organ doses and AEC settings affect the expected image quality in scan images for a specific task. Using two separate deep learning networks (i.e., dose estimation model 304 and quality estimation model 308), the following describes the use of these networks to facilitate the determination of optimal AEC settings for a specific patient and task, while balancing the competing goal of minimizing radiation dose exposure to the patient and achieving the desired image quality level in the target area required for clinical assessment. To facilitate this purpose, the following AEC optimization technique employs dose estimation model 304 and quality estimation model 308 to determine the optimal AEC settings for different patients and tasks based on training localization images that deliver the requested image quality with the minimum dose, thereby formulating the optimal AEC as an optimization problem. These optimal AEC settings are then used as ground truth examples to build an end-to-end model that maps localization and task to optimal AEC in real time.

[0149] refer to Figure 20 An exemplary optimized component 310 according to one or more embodiments of the disclosed subject matter is shown. For the sake of brevity, repeated descriptions of similar elements used in the respective embodiments are omitted.

[0150] The optimization component 310 may include an optimal AEC evaluation component 2002 and a deep learning optimization component 2008. The optimal AEC evaluation component 2002 may employ a trained dose estimation model 304 and a trained quality estimation model 308 to determine and evaluate how different AEC settings directly affect organ dose and image quality in the target and background regions from one or more localization images. The optimal AEC evaluation component 2002 may further employ one or more combined optimization techniques to iteratively evaluate how different AEC patterns affect organ dose and image quality in the target and background regions, focusing on an optimal AEC for a specific patient and task (where the task defines / controls the target and background regions) that best achieves and / or balances the defined AEC optimization criteria 2004, according to a defined weighting scheme.

[0151] In various implementations, the AEC optimization criterion 2004 can model AEC as an optimization problem balancing competing objectives of dose minimization and quality maximization. Specifically, the AEC optimization criterion 2004 can provide an optimization function that defines the optimal AEC as a function of optimization criteria for organ dose and image quality in the target region, where the optimization criteria are based on minimizing organ dose and maximizing image quality. For example, in some implementations, the optimization function can be modeled as follows: given location and task, maximize image quality in the target region for a fixed dose constrained by system constraints (i.e., the feasible space of AEC). In another implementation, the optimization function can be modeled as follows: given location and task, minimize organ dose for a fixed image quality in the target region constrained by system constraints (i.e., the feasible space of AEC). Still in other implementations, the optimization criteria of the optimization function can be based on defined thresholds and / or ranges for maximum and minimum organ dose and image quality in the target / background region.

[0152] In some implementations, the optimal AEC evaluation component 2002 can employ a derivative-free optimization function, such as a covariance matrix adaptive evolutionary strategy (CMA-ES), to determine the optimal AEC for a specific localization image / task combination. Such optimization methods employ repeated evaluations of objective functions (dose and image quality), such as... Figure 21 As shown.

[0153] in this regard, Figure 21An optimal AEC process 2106 is illustrated, which can be performed by the optimal AEC evaluation component 2002 to determine the optimal AEC 2110 based on one or more localization images 2102 and a specific task 2104. The localization images 2102 may include one or more localization images of the patient's anatomical ROI (e.g., a single fluoroscopic localization image, lateral and frontal localization images, etc.). For example, localization images 2102 may correspond to localization image 102, localization image 402, and training localization image 902. 1-A Training localization image 1202 1-A Task 2104 may include tasks selected from the defined set of tasks for which the quality estimation model 308 is trained.

[0154] According to this implementation, optimal AEC processing 2106 involves iteratively applying dose estimation model 304 and quality estimation model 308 to the localization image under different candidate AEC modes 2108, and determining the expected organ dose and target and background image quality metrics under each candidate AEC mode in the candidate AEC modes 2108. While each evaluation of the expected deep learning model is in the sub-second range, the overall optimization processing time can range from seconds to minutes, depending on the processing speed of the processing hardware used for the many evaluations of the objective function. The optimal AEC evaluation component 2002 then selects the optimal AEC 2110 from the candidate AEC modes that best achieve the optimization criteria of organ dose and target / background region image quality. In this regard, assuming the optimization criteria are based on a maximum fixed dose, the optimal AEC evaluation component 2002 can find the optimal AEC mode that achieves the highest image quality in the target region under the maximum fixed dose. In other implementations, the optimization criteria can be tailored to balance dose and target region image quality based on defined weights for organ dose and target region image quality. Using these implementations, the optimal AEC 2110 may include an AEC that provides a high level of image quality (e.g., but not the highest) in the target area while also providing a low organ dose (e.g., but not the lowest).

[0155] In some implementation schemes, Figure 21 The optimal AEC processing 2106 shown can be used in conjunction with the execution of an actual CT scan before capturing high-resolution scan images (e.g., while the patient is on the scanning table). For example, refer again... Figure 1 The optimal AEC processing 2106 can correspond to the AI ​​model processing at position 104. Using these implementations, the optimal AEC processing 2106 can be used to determine the optimal AEC for the patient and task in the actual imaging procedure within a clinical workflow.

[0156] In other embodiments, the localization image 2102 may correspond to a training localization image included in a training localization image set. For example, the training localization image set may include the same training localization images used to train the dose estimation model 304 and / or the mass estimation model (e.g., training localization images included in training data 900 and / or training data 1200). In other embodiments, the training localization images used as input to the optimal AEC processing 2106 may include new localization images. Regardless of the approach, using these embodiments, the optimal AEC processing 2106 can be applied to each training localization image in the training localization images to determine the optimal AEC pattern for each training localization image and task combination. These optimal AEC patterns can then be used as GT examples to train and develop end-to-end models that take localization and task as input and directly output the optimal AEC.

[0157] Refer again Figure 20 In the illustrated implementation, this end-to-end model may include one or more AEC optimization models 312. Using these implementations, the optimal AEC data 2006 may correspond to the ground truth optimal AEC pattern for different combinations of localization images / tasks. In this regard, the deep learning optimization component 2008 may employ a training component 316 to train one or more AEC optimization models 312 to take one or more localization images and the selected task as input and directly output the optimal AEC. The training data will be generated by the optimal AEC evaluation component 2002 and a reference... Figure 21 The described technique comprises a predetermined optimal AEC composition for different localization image / task combinations. In this regard, while the optimal AEC evaluation component 2002 explicitly estimates the dose and image quality for each AEC candidate, the AEC optimization model 312 can be trained to implicitly model these doses and image qualities in its prediction of the optimal AEC. This direct prediction should run in real-time in inference mode.

[0158] The type of machine learning model used for one or more AEC optimization models 312 can vary. In some embodiments, the variations of one or more AEC optimization models 312 may include one or more deep learning models, such as CNNs, RESNET-type models, and other neural network models with different types of pooling techniques. Other suitable types of machine learning models may include, but are not limited to, GANs, LSTMs, attention-based models, transformers, decision tree-based models, Bayesian network models, regression models, etc. These models can be trained using supervised, unsupervised, and / or semi-supervised machine learning techniques. The number of AEC optimization models 312 can also vary. In this regard, in some embodiments, multiple different AEC optimization models 312 can be trained and customized for different anatomical ROIs that are synonymous with the corresponding dose estimation model 304 and quality estimation model 308. In other embodiments, a general AEC optimization model 312 that can be applied to any anatomical ROI can be developed.

[0159] Figure 22 An exemplary AEC optimization model 312, once trained, is illustrated according to one or more embodiments. In various embodiments, the learning optimization component 2008 (and / or the inference component 318) can be applied to the AEC optimization model 312 in a clinical workflow to provide real-time AEC optimization for CT imaging scans. For example, refer again... Figure 1 The AI ​​model processing at position 104 can correspond to the application of AEC optimization model 312, such as... Figure 22 As shown. In this respect, once trained, the input to the AEC optimization model 312 can include one or more localization images 2202 of the captured anatomical regions of the patient and the selected task 2204. The output of the AEC optimization model 312 includes the optimal AEC 2206.

[0160] In some implementations, the AEC optimization model 312 can also be configured to generate the expected organ dose 2208 and the expected target and background image quality 2210 for the optimal AEC mode 2206. Alternatively, once the optimal AEC 2206 has been generated by the AEC optimization model 312, the dose estimation model 304 and / or quality estimation model 308 can be applied individually to the localization image 2202 using the optimal AEC 2206 as the selected AEC mode to determine the organ dose 2208 and the expected target and background image quality 2210 for the optimal AEC mode 2206. Therefore, as the user selects a task or modifies other acquisition parameters, the resulting effects on dose and IQ are dynamically generated, and the optimized AEC is prepared for scanning the patient. The delivered AEC, along with the predicted dose and IQ, can be included in the scan report.

[0161] Figure 23 A high-level flowchart of an exemplary computer implementation of process 2300 for determining optimal AEC according to one or more embodiments of the disclosed subject matter is presented. Process 2300 provides exemplary procedures for execution by optimal AEC evaluation component 2002 according to various embodiments described herein. For brevity, repeated descriptions of similar elements employed in the respective embodiments are omitted.

[0162] According to process 2300, at 2302, a system operatively coupled to a processor (e.g., system 300, etc.) (e.g., via receiving component 314) receives task information identifying a task associated with the execution of a CT scan of a patient's anatomical region, and one or more localization images of the captured anatomical region. For example, in some embodiments, the localization images and task information may be received in association with the execution of an actual CT scan prior to the capture of high-resolution scan images (e.g., while the patient is on the scanning table). Using these embodiments, process 2300 can be used to determine the optimal AEC for the patient and task in an actual imaging procedure within a clinical workflow. In other embodiments, the localization images may correspond to training localization images included in a set of training localization images. Using these embodiments, process 2300 can be applied to each training localization image in the training localization images to determine the optimal AEC pattern for each training and task combination. These optimal AEC patterns can be used by deep learning optimization component 2008 (and training component 316) to train and develop one or more AEC optimization models 312.

[0163] At 2304, the system employs a first deep learning network (e.g., dose estimation model 304) to determine, based on one or more localization images (e.g., using the optimal AEC evaluation component 2002), the expected organ dose, representing the expected radiation dose exposed to (or absorbed by) one or more organs in different AEC modes used for CT scanning. At 2306, the system employs a second deep learning network (e.g., quality estimation model 308) to determine, based on one or more localization images (e.g., using the optimal AEC evaluation component 2002), an expected metric of image quality in target and background regions of scanned images captured in different AEC modes, where the target and background regions are task-based. At 2308, the system determines, based on the expected organ dose and the expected metric of image quality in different AEC modes (e.g., using the optimal AEC evaluation component 2002), the optimal AEC mode using an optimization function (e.g., provided by AEC optimization criterion 2004) to maximize image quality in the target region and minimize radiation dose to one or more organs.

[0164] Figure 24A high-level flowchart of another exemplary computer implementation of a process 2400 for determining the optimal AEC according to one or more embodiments of the disclosed subject matter is presented. Process 2400 provides exemplary procedures performed by the optimal AEC evaluation component 2002 and the deep learning optimization component 2008 according to various embodiments described herein. For brevity, repeated descriptions of similar elements employed in the respective embodiments are omitted.

[0165] According to process 2400, at 2402, a system operatively coupled to the processor (e.g., system 300, etc.) determines an optimal AEC mode for different combinations of localization images and tasks based on the expected organ dose under different AEC modes and expected metrics of image quality in the target and background regions (e.g., using optimal AEC evaluation component 2002), wherein the determination includes using an objective function that maximizes image quality in the target region and minimizes organ dose. At 2404, the system trains a deep learning network (e.g., AEC optimization model 312) to determine the optimal AEC mode for different combinations of localization images and tasks (e.g., via deep learning optimization component 2008 and training component 316). At 2406, the system employs a deep learning network to determine an optimal AEC mode for a new CT scan based on a new localization image (or image) and a selected task of different tasks (e.g., via deep learning optimization component 2008 and inference component 318).

[0166] Figure 25 Another exemplary computing system 2500, facilitating the customization of AEC settings for specific patient anatomy and clinical tasks according to one or more embodiments of the disclosed subject matter, is presented. The computing system 2500 includes a computing device 2502 and a CT scanner 108. The computing device 2502 may be operatively and / or communicatively coupled to the CT scanner 108 and provides one or more operations for controlling the CT scanner 108. The computing device 2502 may include components identical or similar to those of the computing system 300, with the addition of an interface component 2504, a parameter adjustment component 2506, an acquisition control component 2508, and a display 2506. For brevity, repeated descriptions of similar elements employed in corresponding embodiments are omitted.

[0167] According to system 2500, training component 316 has been excluded from computing device 2502 to indicate that the training and development of one or more dose estimation models 304, one or more quality estimation models 208, and one or more AEC optimization models 312 have been completed. In this regard, system 2500 provides an exemplary operating environment in which one or more of these models can be executed in real time by inference component 318 in association with the actual CT scan of patient 110 performed according to process 100. Using these embodiments, computing device 2502 can correspond to a computing device used by a CT scanner operator (or another suitable entity).

[0168] For example, before performing a CT scan and acquiring high-resolution CT scan images, receiving unit 314 may receive one or more low-resolution background images of the anatomical ROI of the patient to be scanned. Receiving unit 314 may also receive information selecting / defining relevant tasks for performing the CT scan, which controls / defining relevant target and background regions for the high-resolution scan images. Inference unit 318 may further apply one or more AEC optimization models 312 to the localization images under the selected task to generate an optimal AEC pattern (e.g., optimal AEC settings) for the patient and task. In some embodiments, inference unit 318 may also apply one or more dose estimation models 304 to the localization images under the optimal AEC pattern to generate an estimated organ dose that will be absorbed under the optimal AEC pattern. Inference unit 318 may also apply one or more quality estimation models 308 to the localization images under the optimal AEC pattern and task to generate an estimated metric of the image quality of the target and background regions. Information identifying the optimal AEC pattern, the estimated organ dose, and / or the metrics of the image quality of the target and background regions may be presented to the operator via a display. The acquisition control unit 2508 can further automatically configure the CT scanner to perform CT scans using the optimal AEC mode (e.g., high-resolution CT scan procedures).

[0169] To facilitate this purpose, interface component 2504 may provide an interactive graphical user interface (GUI) that may be presented to an operator via display 2506 in association with performing AEC and configuring and controlling CT acquisitions performed by the CT scanner. The interactive GUI may facilitate receiving user input to select / define tasks (and / or adjust AEC optimization criteria), performing one or more models, presenting model outputs, and controlling one or more operations of the CT scanner 108. For example, the GUI may provide controls for receiving user input to identify one or more parameters of the CT acquisition, receiving localization images, and defining / selecting tasks for the CT examination. In some embodiments, the GUI may include optimal AEC controls that automatically generate an optimal AEC for the patient and task based on localization images using the techniques described herein. Acquisition control component 2508 may further configure the CT scanner to automatically and / or, in response to a request from an operator, perform a CT examination using an optimal AEC mode provided by the CT scanner 108. For example, the acquisition control unit can be operatively coupled to the imaging device (e.g., CT scanner 108) that performs the CT scan, and control the execution of the CT scan by the imaging device based on the optimal AEC mode.

[0170] The parameter adjustment unit 2506 may also provide additional tools (e.g., accessible via a GUI) that allow the operator to provide inputs for adjusting one or more parameters for optimal AEC and / or for providing optimization criteria for optimal AEC (e.g., desired image quality level and / or desired organ dose distribution). For example, the parameter adjustment unit 2506 may facilitate receiving user input for adjusting one or more parameters for optimal AEC, resulting in a modified AEC pattern. Based on the received user input, the inference unit 318 may reapply one or more dose estimation models 304 and one or more quality estimation models 308 to the localization image to determine, in real time, an updated expected radiation dose to one or more organs, and an updated metric of expected image quality in the target / background region under the modified AEC pattern. The modified AEC can then be configured for the CT scanner via the acquisition control unit 2508. In another example, the parameter adjustment unit 2506 may facilitate receiving user input for defining optimization criteria for the optimal AEC pattern. Specifically, parameter adjustment component 2506 can provide adjustment controls that allow an operator to provide input of at least one of the desired image quality for identifying the target and / or background area and the desired radiation dose for one or more organs. Based on the received user input, optimization component 310 can determine a modified AEC pattern that achieves the desired image quality and desired radiation dose. For example, in some embodiments, optimization component 310 can employ optimal AEC evaluation component 2002 to re-evaluate optimal ACE using modified ACE optimization criteria in association with the reapplication of dose estimation model 304 and quality estimation model 308 (e.g., by inference component 318) to the localization image. Alternatively or additionally, one or more AEC optimization models 312 may include models configured to generate an optimal AEC pattern under user-defined AEC optimization criteria and a selected task based on the localization image. Thus, as the user selects a task or modifies other acquisition parameters, the resulting effects on dose and IQ, and the optimized AEC, are prepared for scanning the patient. In addition to configuring and controlling the CT scanner to perform CT acquisitions using the optimal AEC, information identifying the AEC used, predicted dose, and predicted target / background IQ can be included in the scan report.

[0171] Figure 26 A high-level flowchart of another exemplary computer implementation of a process for performing a CT examination using optimal AEC, according to one or more embodiments of the disclosed subject matter, is presented. For brevity, repeated descriptions of similar elements employed in the respective embodiments are omitted.

[0172] According to process 2600, at 2602, a system operatively coupled to a processor (e.g., system 2500, etc.) (e.g., via receiving unit 314) receives task information identifying a task associated with the execution of a CT scan of a patient's anatomical region, as well as one or more localization images of the captured anatomical region. For example, in some embodiments, the localization images and task information may be received in association with the execution of an actual CT scan prior to the capture of high-resolution scan images (e.g., while the patient is on the scanning table). Using these embodiments, process 2600 can be used to determine the patient's optimal AEC and the task of the actual imaging procedure in a clinical workflow, and reconnaissance images can be received directly from CT scanner 108 (e.g., via acquisition control unit 2508).

[0173] At 2604, the system uses one or more deep learning networks (e.g., one or more optimization models 312, one or more dose estimation models 304, and / or one or more quality estimation models 308) to determine (e.g., using optimization component 310) an optimal AEC mode for the CT scan, which maximizes image quality in a target region of the scanned images to be captured during the scan and minimizes radiation dose to organs included in the anatomical region, where the target region is task-based. At 2606, the system configures the CT scanner (e.g., CT scanner 108) to perform the CT scan using the optimal AEC mode.

[0174] Exemplary operating environment

[0175] One or more embodiments may be systems, methods, and / or computer program products at any possible level of technical detail in integration. A computer program product may include a computer-readable storage medium (or media) having computer-readable program instructions thereon for causing a processor to execute aspects of the invention.

[0176] Computer-readable storage media can be tangible devices that hold and store instructions for use by instruction execution devices. Computer-readable storage media can be, for example, but not limited to, electronic storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. A non-exhaustive list of more specific examples of computer-readable storage media includes: portable computer floppy disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable optical disc read-only memory (CD-ROM), digital versatile optical disc (DVD), memory sticks, floppy disks, mechanical encoding devices (such as punch cards or raised structures in grooves on which instructions are recorded), and any suitable combination of the foregoing. As used herein, computer-readable storage media should not be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses passing through fiber optic cables), or electrical signals transmitted through wires.

[0177] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to a suitable computing / processing device, or downloaded via a network (e.g., the Internet, a local area network, a wide area network, and / or a wireless network) to an external computer or external storage device. The network may include copper transmission cables, optical fiber transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them for use in a computer-readable storage medium stored within the suitable computing / processing device.

[0178] Computer-readable program instructions used to perform the operations of this invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, configuration data of an integrated circuit system, or source code or object code written in any combination of one or more programming languages ​​(including object-oriented programming languages ​​such as Smalltalk, C++, etc.), procedural programming languages ​​(such as the "C" programming language or similar programming languages), and machine learning programming languages ​​(such as CUDA, Python, Tensorflow, PyTorch, etc.). The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or a server using suitable processing hardware. In the latter scenario, the remote computer may be connected to the user's computer via any type of network (including a local area network (LAN) or a wide area network (WAN)) or may be connected to an external computer (e.g., via the Internet through an Internet service provider). In various embodiments involving machine learning programming instructions, the processing hardware may include one or more graphics processing units (GPUs), central processing units (CPUs), etc. For example, one or more machine learning models among the disclosed machine learning models (e.g., one or more dose estimation models 304, one or more quality estimation models 308, one or more optimization models 312, one or more organ segmentation models 808, and one or more anatomical structure segmentation models 1308) can be written in a suitable machine learning programming language and executed via one or more GPUs, CPUs, or combinations thereof. In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs) can execute computer-readable program instructions to personalize the electronic circuitry in order to perform aspects of the invention by utilizing state information of computer-readable program instructions.

[0179] This document describes aspects of the invention with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0180] These computer-readable program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions / actions specified in one or more blocks of a flowchart and / or block diagram. These computer-readable program instructions may also be stored in a computer-readable storage medium that can direct a computer, programmable data processing apparatus, and / or other devices to function in a particular manner, such that the computer-readable storage medium having the instructions stored therein includes an article of writing comprising instructions that implement aspects of the functions / actions specified in one or more blocks of a flowchart and / or block diagram.

[0181] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer-implemented process, such that the instructions executed on the computer, other programmable apparatus or other device implement the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0182] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible specific implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of instructions comprising one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions indicated in the blocks may not occur in the order shown in the figures. For example, depending on the functions involved, two blocks shown consecutively may be executed substantially simultaneously, or sometimes the blocks may be executed in reverse order. It will also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented by a dedicated hardware-based system that performs the specified function or action or implements a combination of dedicated hardware and computer instructions.

[0183] Combination Figure 27 The systems and programs described below may be embodied in hardware, such as a single integrated circuit (IC) chip, multiple ICs, application-specific integrated circuits (ASICs), etc. Furthermore, the order in which some or all program blocks appear in each program should not be considered restrictive. Rather, it should be understood that some program blocks can be executed in various orders, and not all of the orders described can be explicitly shown herein.

[0184] refer to Figure 27An example environment 2700 for implementing various aspects of the claimed subject matter includes a computer 2702. Computer 2702 includes a processing unit 2704, system memory 2706, a codec 2735, and a system bus 2708. System bus 2708 couples system components, including but not limited to system memory 2706, to processing unit 2704. Processing unit 2704 can be any of a variety of available processors. Dual microprocessors, one or more GPUs, and other multiprocessor architectures may also be used as processing unit 2704.

[0185] The system bus 2708 can be any of a variety of bus architectures, including memory bus or memory controller, peripheral bus or external bus, or local bus using any of the available bus architectures, including but not limited to Industry Standard Architecture (ISA), Micro Channel Architecture (MSA), Extended ISA (EISA), Intelligent Drive Electronic Devices (IDE), VESA Local Bus (VLB), Peripheral Component Interconnect (PCI), Card Bus, Universal Serial Bus (USB), Advanced Graphics Port (AGP), PCMCIA Bus, FireWire (IEEE 1394), and Small Computer System Interface (SCSI).

[0186] In various embodiments, system memory 2706 includes volatile memory 2710 and non-volatile memory 2712, which may employ one or more memory architectures among the disclosed memory architectures. The Basic Input / Output System (BIOS) (containing basic routines such as transferring information between elements within computer 2702 during startup) is stored in non-volatile memory 2712. Additionally, according to the invention, codec 2735 may include at least one of an encoder or a decoder, wherein at least one of the encoder or decoder may be composed of hardware, software, or a combination of hardware and software. Although codec 2735 is depicted as a separate component, codec 2735 may be contained within non-volatile memory 2712. By way of illustration and not limitation, non-volatile memory 2712 may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, 3D flash memory, or resistive memory, such as resistive random access memory (RRAM). In at least some embodiments, the non-volatile memory 2712 may employ one or more of the disclosed memory devices. Furthermore, the non-volatile memory 2712 may be computer memory (e.g., physically integrated with computer 2702 or its motherboard) or removable memory. Examples of suitable removable memory that can be used to implement the disclosed embodiments may include a Secure Digital (SD) card, a Compact Flash (CF) card, a Universal Serial Bus (USB) Memory Stick, etc. The volatile memory 2710 includes random access memory (RAM) that acts as external cache memory, and may also employ one or more of the disclosed memory devices in various embodiments. By way of illustration and not limitation, RAM can be provided in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), and enhanced SDRAM (ESDRAM), etc.

[0187] Computer 2702 may also include removable / non-removable, volatile / non-volatile computer storage media. Figure 27Disk storage device 2714 is illustrated, for example. Disk storage device 2714 includes, but is not limited to, devices such as disk drives, solid-state drives (SSDs), flash memory cards, or memory sticks. Additionally, disk storage device 2714 may include storage media, alone or in combination with other storage media, including but not limited to optical disc drives, such as optical disc ROM devices (CD-ROM), CD recordable drives (CD-R drives), CD rewritable drives (CD-RW drives), or digital versatile optical disc ROM drives (DVD-ROM). To facilitate connection of disk storage device 2714 to system bus 2708, removable or non-removable interfaces, such as interface 2716, are typically used. It should be understood that disk storage device 2714 may store user-related information. Such information may be stored on a server or provided to applications running on a server or user device. In one embodiment, the type of information stored on disk storage device 2714 or transmitted to a server or application may be notified to the user (e.g., via output device 2736). Users may be given the opportunity to opt in or out of the collection or sharing of such information via a server or application (e.g., via input from input device 2728).

[0188] It should be understood that Figure 27 Software is described that acts as an intermediary between the user and the basic computer resources described in the suitable operating environment 2700. This software includes an operating system 2718. The operating system 2718, which may be stored on a disk storage device 2714, is used to control and allocate the resources of the computer 2702. Application programs 2720 utilize the operating system 2718 to manage resources through program modules 2724, and program data 2726, such as power-on / power-off transaction tables, stored in system memory 2706 or on disk storage device 2714. It should be understood that the claimed subject matter can be implemented using various operating systems or combinations of operating systems.

[0189] Users input commands or information into computer 2702 via input device 2728. Input device 2728 includes, but is not limited to, pointing devices such as mouse, trackball, stylus, touchpad, keyboard, microphone, joystick, gamepad, satellite dish, scanner, TV tuner card, digital camera, digital camcorder, webcam, etc. These and other input devices are connected to processing unit 2704 via interface port 2730 through system bus 2708. Interface port 2730 includes, for example, serial port, parallel port, game port, and Universal Serial Bus (USB). Output device 2736 uses some of the same type of ports as input device 2728. Thus, for example, a USB port can be used to provide input to computer 2702 and output information from computer 2702 to output device 2736. Output adapter 2734 is provided to illustrate some output devices 2736 such as monitors, speakers, and printers, as well as other output devices 2736 that require special adapters. By way of illustration and not limitation, output adapter 2734 includes a video card and a sound card that provide a means of connection between output device 2736 and system bus 2708. It should be noted that other devices or systems of devices provide both input and output capabilities, such as remote computer 2738.

[0190] Computer 2702 can operate in a networked environment using a logical connection to one or more remote computers, such as remote computer 2738. Remote computer 2738 can be a personal computer, server, router, network PC, workstation, microprocessor-based device, peer-to-peer device, smartphone, tablet, or other network node, and typically includes many of the elements described with respect to computer 2702. For simplicity, only memory storage device 2740 is shown for remote computer 2738. Remote computer 2738 is logically connected to computer 2702 via network interface 2742 and subsequently connected via communication connection 2744. Network interface 2742 covers wired or wireless communication networks, such as local area networks (LANs) and wide area networks (WANs), as well as cellular networks. LAN technologies include Fiber Distributed Data Interface (FDDI), Copper Distributed Data Interface (CDDI), Ethernet, Token Ring, etc. WAN technologies include, but are not limited to, point-to-point links, such as Integrated Services Digital Network (ISDN) and its variants, circuit-switched networks, packet-switched networks, and Digital Subscriber Line (DSL).

[0191] Communication connection 2744 refers to the hardware / software used to connect network interface 2742 to bus 2708. Although communication connection 2744 is shown inside computer 2702 for clarity, the communication connection may also be external to computer 2702. For illustrative purposes only, the hardware / software required to connect to network interface 2742 includes internal and external technologies such as modems, including conventional telephone-grade modems, cable modems and DSL modems, ISDN adapters, and wired and wireless Ethernet cards, hubs and routers.

[0192] Although the subject matter has been described above in the general context of computer executable instructions for a computer program product running on one or more computers, those skilled in the art will recognize that the present disclosure may also be implemented in combination with other program modules. Generally, program modules include routines, programs, components, data structures, etc., that perform specific tasks and / or implement specific abstract data types. Furthermore, those skilled in the art will understand that the computer implementation of the methods of the present invention can be practiced using other computer system configurations, including single-processor or multi-processor computer systems, small computing devices, mainframe computers, and computers, handheld computing devices (e.g., PDAs, telephones), microprocessor-based or programmable consumer or industrial electronic devices, etc. The illustrated aspects can also be practiced in a distributed computing environment where tasks are performed by remote processing devices linked via a communication network. However, some (if not all) aspects of the present disclosure can be practiced on a standalone computer. In a distributed computing environment, program modules can reside in both local and remote memory storage devices.

[0193] As used herein, the terms “component,” “system,” “platform,” “interface,” etc., may refer to and / or include computer-related entities or entities associated with an operator having one or more specific functions. Entities disclosed herein may be hardware, a combination of hardware and software, software, or software in execution. For example, a component may be, but is not limited to, a program running on a processor, a processor, an object, an executable file, an execution thread, a program, and / or a computer. By way of illustration, an application running on a server and a server may both be components. One or more components may reside within a program and / or an execution thread, and components may be located on a single computer and / or distributed across two or more computers. In another example, a corresponding component may execute on various computer-readable media on which various data structures are stored. Components may communicate via local and / or remote programs, such as based on signals having one or more data packets (e.g., data from one component that interacts with another component in a local system, a distributed system, and / or a network (such as the Internet with other systems). As another example, a component can be a device having specific functions provided by mechanical parts operated by an electrical or electronic circuitry system, which is operated by a software or firmware application executed by a processor. In this case, the processor can be internal or external to the device and can execute at least a portion of the software or firmware application. As yet another example, a component can be a device that provides specific functions through electronic components rather than mechanical parts, wherein the electronic components can include a processor or other means for executing software or firmware that at least partially gives the electronic components functionality. In one aspect, a component can be emulated, for example, via a virtual machine within a cloud computing system.

[0194] Furthermore, the term "or" is intended to mean an inclusive "or" rather than an exclusive "or." That is, unless otherwise specified or clear from the context, "X adopts A or B" is intended to mean any natural inclusive substitution. That is, if X adopts A; X adopts B; or X adopts both A and B, then "X adopts A or B" is satisfied in any of the foregoing cases. Additionally, unless otherwise specified or clear from the context to be directed to the singular form, the articles "a" and "an" used in this specification and accompanying drawings should generally be construed as meaning "one or more." As used herein, the terms "example" and / or "exemplary" are used to mean serving as an example, instance, or illustration, and are intended to be non-limiting. For the avoidance of doubt, the subject matter disclosed herein is not limited to such examples. Furthermore, any aspect or design described herein as "example" and / or "exemplary" is not necessarily to be construed as preferred or advantageous over other aspects or designs, nor does it exclude equivalent exemplary structures and techniques known to those skilled in the art.

[0195] As used herein, the term "processor" can refer substantially to any computing processing unit or device, including but not limited to a single-core processor; a single processor with software multithreading capabilities; a multi-core processor; a multi-core processor with software multithreading capabilities; a multi-core processor with hardware multithreading technology; a parallel platform; and a parallel platform with distributed shared memory. Additionally, a processor can refer to an integrated circuit, an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a field-programmable gate array (FPGA), a programmable logic controller (PLC), a complex programmable logic device (CPLD), discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. Furthermore, a processor can utilize nanoscale architectures (such as, but not limited to, molecular and quantum dot-based transistors, switches, and gates) to optimize space usage or enhance the performance of user devices. A processor can also be implemented as a combination of computing processing units. In this disclosure, terms such as "storage," "storage device," "data storage," "data storage apparatus," "database," and substantially any other information storage component associated with the operation and function of a component are used to refer to a "memory component," an entity embodied in "memory," or a component that includes memory. It should be understood that the memory and / or memory components described herein may be volatile or non-volatile memory, or may include both volatile and non-volatile memory. By way of illustration and not limitation, non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable ROM (EEPROM), flash memory, or non-volatile random access memory (RAM) (e.g., ferroelectric RAM (FeRAM)). For example, volatile memory may include RAM that can act as external cache memory. By way of illustration and not limitation, RAM can be provided in a variety of forms, such as synchronous RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), direct Rambus RAM (DRRAM), direct Rambus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM). Furthermore, the memory components disclosed in the systems or computer-implemented methods described herein are intended to include, but are not limited to, these and any other suitable types of memory.

[0196] The foregoing description includes only examples of systems and computer-implemented methods. Of course, it is impossible to describe every conceivable combination of components or computer-implemented methods for the purposes of describing this disclosure, but those skilled in the art will recognize that many other combinations and substitutions are possible. Furthermore, regarding the extent to which the terms “comprising,” “having,” “possessing,” etc., are used in the detailed description, claims, appendices, and drawings, such terms are intended to be inclusive in a manner similar to the term “comprising,” as interpreted when “comprising” is used as a transitional word in the claims. Various embodiments have been described for illustrative purposes, but these descriptions are not intended to be exhaustive or limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein has been chosen to best explain the principles of the embodiments, the practical application or improvement of technology superior to that found in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A system comprising: Memory, which stores computer-executable components; and A processor that executes computer-executable components stored in the memory, wherein the computer-executable components include: A receiving component receives one or more localized images of the anatomical region of the patient, captured in association with the execution of a computed tomographic scan of the anatomical region of the patient, wherein the anatomical region includes one or more organs; A dose estimation component employs one or more first machine learning models to estimate, based on the one or more positioning images, a expected organ dose representing the expected radiation dose to the one or more organs under different automatic exposure control modes for the computed tomography scan; A quality estimation component, which employs one or more second machine learning models to estimate, based on the one or more localization images, a predicted measure of image quality in the target and background regions of computed tomography scans captured under the different automatic exposure control modes; and An optimization component determines, based on the expected organ dose and an expected metric of image quality in the target region, an optimal automatic exposure control mode among the different automatic exposure control modes performed for the computed tomography scan, wherein the optimal automatic exposure control mode maximizes image quality in the target region and minimizes radiation dose to the organ. The quality estimation component determines the target region and the background region based on a defined clinical task associated with the execution of the computed tomography scan, and the target region and the background region vary for different clinical tasks.

2. The system of claim 1, wherein the one or more first machine learning models comprise deep learning models trained using a supervised machine learning process and benchmark ground truth data, the benchmark ground truth data comprising multiple computed tomographic volumes generated from computed tomographic scans of the anatomical region under the different automatic exposure control modes and organ segmentation dose maps generated from the computed tomographic volumes.

3. The system of claim 2, wherein the computer-executable component further comprises: A training component that uses the supervised machine learning process to train the deep learning model.

4. The system of claim 1, wherein the one or more second machine learning models comprise a deep learning model trained using a supervised machine learning process and benchmark ground truth image quality data, the benchmark ground truth image quality data being generated using multiple computed tomographic volumes generated from computed tomographic scans of the anatomical region under the different automatic exposure control modes, wherein the benchmark ground truth image quality data provides a measure of image quality in the target region and the background region as represented in the computed tomographic volumes under the different automatic exposure control modes.

5. The system of claim 4, wherein the computer-executable component further comprises: A training component that uses the supervised machine learning process to train the deep learning model.

6. The system of claim 4, wherein the computer-executable component further comprises: The simulation component generates a simulated noise projection using the computed tomography scan and generates the reference ground truth data image quality data based on the simulated noise projection.

7. The system according to claim 1, wherein the different automatic exposure control modes reflect different tube current modulations.

8. The system of claim 1, wherein the different automatic exposure control modes reflect different acquisition parameter values ​​of one or more acquisition parameters selected from the group consisting of: tube voltage, collimation, filtering, bow junction, pitch, and starting angle.

9. A method, the method comprising: A system operatively coupled to a processor receives one or more localized images of the anatomical region of the patient, captured in association with the execution of a computed tomographic scan of the anatomical region of the patient, wherein the anatomical region includes one or more organs; The system employs one or more first machine learning models to estimate, based on the one or more localization images, the expected organ dose, representing the expected radiation dose to the one or more organs under different automatic exposure control modes used for the computed tomography scan; The system employs one or more second machine learning models to estimate, based on the one or more localized images, a desired measure of image quality in the target and background regions of computed tomography scans captured under the different automatic exposure control modes; as well as The system determines, based on the expected organ dose and an expected metric of image quality in the target region, the optimal automatic exposure control mode among the different automatic exposure control modes performed for the computed tomography scan, wherein the optimal automatic exposure control mode maximizes image quality in the target region and minimizes radiation dose to the organ. The target region and the background region are determined by the system based on a defined clinical task associated with the execution of the computed tomography scan, and the target region and the background region vary for different clinical tasks.

10. The method according to claim 9, further comprising: The system trains one or more first machine learning models using a supervised machine learning process and benchmark ground truth data, the benchmark ground truth data including multiple computed tomographic volumes generated from computed tomographic scans of the anatomical region under the different automatic exposure control modes and organ segmentation dose maps generated from the computed tomographic volumes.

11. The method according to claim 9, further comprising: The system trains one or more second machine learning models using a supervised machine learning process and benchmark ground truth image quality data, which is generated using multiple computed tomographic volumes generated from computed tomographic scans of the anatomical region under different automatic exposure control modes. The benchmark ground truth image quality data provides a measure of image quality in the target region and the background region as represented in the computed tomographic volumes under the different automatic exposure control modes.

12. The method according to claim 11, further comprising: The system generates a simulated noise projection using the computed tomography scan; as well as The system generates the reference true data image quality data based on the simulated noise projection.

Citation Information

Patent Citations

  • Medical image quality evaluation method, device, equipment and storage medium

    CN110428415A

  • Systems and methods for identifying medical image acquisition parameters

    US20150199478A1

  • System and Method for Reconstructing Image Volumes from Sparse Two-Dimensional Projection Data

    US20200094074A1