Tomographic image processing using artificial intelligence (AI) engine

Abstract

Example methods and systems for tomographic image reconstruction are provided. One example method can include obtaining two-dimensional (2D) projection data (310) and processing the 2D projection data using an AI engine (301) that includes a plurality of first processing layers (311), an interposed back-projection module (312), and a plurality of second processing layers (313). Example processing using the AI engine can include generating 2D feature data (320) by processing the 2D projection data using the plurality of first processing layers, reconstructing first three-dimensional (3D) feature volume data (330) from the 2D feature data using the back-projection module, and generating second 3D feature volume data (340) by processing the first 3D feature volume data using the plurality of second processing layers. Methods and systems for tomographic data analysis are also provided.

Description

Technical Field

[0001] This invention relates to tomographic image processing using an AI engine. Tomographic image processing may include tomographic image reconstruction using an AI engine. Tomographic image processing may also include tomographic image analysis using an AI engine. Background Technology

[0002] Computed tomography (CT) involves imaging the internal structures of a target object (e.g., a patient) by collecting projection data in a single scan operation (“scan”). CT is widely used in the medical field to observe the internal structures of selected parts of the human body. In an ideal imaging system, radiation rays travel along a corresponding straight path from the radiation source, through the target object, and then to the corresponding pixel detectors of the imaging system to produce volumetric data (e.g., volumetric images) without artifacts. In addition to artifact reduction, radiotherapy planning (e.g., segmentation) can be performed based on the obtained volumetric data. However, in practice, the reconstructed volumetric data may contain artifacts, which in turn lead to image degradation and affect subsequent diagnostic and radiotherapy planning. Summary of the Invention

[0003] This invention provides a method for enabling a computer system to perform tomographic image reconstruction using a first AI engine. This invention also provides a method for enabling a computer system to perform tomographic data analysis using a second AI engine.

[0004] The present invention provides a non-transitory computer-readable storage medium including an instruction set that, in response to execution by a processor of a computer system, causes the processor to perform a tomographic image reconstruction method using an AI engine as defined in claim .

[0005] The present invention provides a non-transitory computer-readable storage medium including an instruction set that is responsive to execution by a processor of a computer system, causing the processor to perform a method for tomographic data analysis using an AI engine as defined in the claim.

[0006] The present invention provides a computer system configured to perform tomographic image reconstruction using an AI engine, wherein the computer system comprises: a processor as described in claim 1 and a non-transitory computer-readable medium having instructions stored thereon.

[0007] The present invention provides a computer system configured to perform computed tomography data analysis using an AI engine, wherein the computer system comprises: a processor as described in claim 1 and a non-transitory computer-readable medium thereon storing instructions thereon.

[0008] References to elements such as “second,” “third,” or “fourth” are provided as convenient labels to distinguish one element from another. The use of “second,” “third,” or “fourth” should not be construed as implying a claim for the lower-numbered element as a feature of the claim.

[0009] According to one aspect of this disclosure, an example method and system for tomographic image reconstruction are provided. An example method may include: obtaining two-dimensional (2D) projection data and processing the 2D projection data using an AI engine, the AI ​​engine including a plurality of first processing layers, an interposed post-projection module, and a plurality of second processing layers. Example processing using the AI ​​engine may include: generating 2D feature data by processing the 2D projection data using the plurality of first processing layers; reconstructing first three-dimensional (3D) feature volume data from the 2D feature data using the post-projection module; and generating second 3D feature volume data by processing the first 3D feature volume data using the plurality of second processing layers. During a training phase, the plurality of first processing layers and the plurality of second processing layers may be trained together to learn corresponding first and second weight data, with the post-projection module interposed therebetween.

[0010] According to another aspect of this disclosure, example methods and systems for tomographic image analysis are provided. An example method may include: obtaining first three-dimensional (3D) feature volume data and processing the first 3D feature volume data using an AI engine, the AI ​​engine including a plurality of first processing layers, an interposed forward projection module, and a plurality of second processing layers. Example processing using the AI ​​engine may include: generating second 3D feature volume data by processing the first 3D feature volume data using the plurality of first processing layers, transforming the second 3D volume data into 2D feature data using the forward projection module, and generating analytical output data by processing the 2D feature data using the plurality of second processing layers. During a training phase, the plurality of first processing layers and the plurality of second processing layers may be trained together to learn corresponding first and second weight data, wherein the forward projection module is interposed therebetween. Attached Figure Description

[0011] Figure 1 This is a schematic diagram illustrating an example process flow for radiotherapy treatment;

[0012] Figure 2 This is a schematic diagram illustrating an example imaging system;

[0013] Figure 3 This is a schematic diagram of an example artificial intelligence (AI) engine for tomographic image reconstruction and tomographic image analysis;

[0014] Figure 4This is a flowchart illustrating an exemplary process by which a computer system uses a first AI engine to perform tomographic image reconstruction;

[0015] Figure 5 This is a schematic diagram illustrating an example training and inference phase of a first AI engine used for tomographic image reconstruction;

[0016] Figure 6 This is a flowchart illustrating an exemplary process by which a computer system uses a second AI engine to perform tomographic image analysis;

[0017] Figure 7 This is a schematic diagram illustrating an example training and inference phase of a second AI engine used for tomographic image analysis;

[0018] Figure 8 This is a schematic diagram illustrating an example training phase of an AI engine used for integrated tomographic image reconstruction and analysis;

[0019] Figure 9 This is a schematic diagram of an exemplary treatment plan for radiotherapy delivery; and

[0020] Figure 10 This is a schematic diagram of an example computer system used to perform tomographic image reconstruction and / or tomographic image analysis. Detailed Implementation

[0021] In the following detailed description, reference is made to the accompanying drawings, which form part of the specification. In the drawings, similar symbols generally denote similar parts unless the context otherwise requires. The illustrative embodiments described in the detailed description, drawings, and claims are not intended to be limiting. Other embodiments may be utilized, and other changes may be made, without departing from the spirit or scope of the subject matter presented herein. It will be readily understood that aspects of this disclosure, as generally described herein and illustrated in the drawings, can be arranged, substituted, combined, and designed in a wide variety of different configurations, all of which are explicitly covered herein.

[0022] Figure 1 This is a schematic diagram illustrating an example processing flow 110 for radiotherapy. The exemplary process 110 may include one or more operations, functions, or actions illustrated by one or more boxes. Various boxes may be combined into fewer boxes, divided into additional boxes, and / or eliminated based on the desired implementation. Figure 1 In the examples, radiotherapy typically includes various stages, such as: an imaging system that performs image data acquisition for the patient (see 110); a radiotherapy planning system that generates an appropriate treatment plan for the patient (see 156) (see 130); and a treatment delivery system that delivers treatment according to the treatment plan (see 160).

[0023] More specifically, in Figure 1 At point 110, an imaging system can be used to perform image data acquisition to capture image data 120 associated with the patient (specifically, the patient's anatomical structures). Any suitable medical imaging modality can be used, such as computed tomography (CT), cone-beam computed tomography (CBCT), positron emission tomography (PET), magnetic resonance imaging (MRI), magnetic resonance imaging (MRT), single-photon emission computed tomography (SPECT), any combination thereof, etc. For example, when using CT or MRI, image data 120 may include a series of two-dimensional (2D) images or slices, each representing a cross-sectional view of the patient's anatomy, or may include volumetric or three-dimensional (3D) images of the patient, or may include time-series of 2D or 3D images of the patient (e.g., four-dimensional (4D) CT or 4D CBCT).

[0024] exist Figure 1 At point 130 in the image data 120, radiotherapy planning can be performed during the planning phase to generate a treatment plan 156 based on the image data 120. Any suitable number of treatment planning tasks or steps can be performed, such as segmentation, dose prediction, projection data prediction, treatment plan generation, etc. For example, segmentation can be performed to generate structural data 140, which can identify various segments or structures from the image data 120. In practice, a three-dimensional (3D) volume of the patient's anatomy can be reconstructed from the image data 120. The irradiated 3D volume is referred to as the treatment or radiation volume, and it can be divided into multiple smaller volume pixels (voxels) 142. Each voxel 142 represents a 3D element associated with a location (i, j, k) within the treatment volume. The structural data 140 can include any suitable data relating to the contour, shape, size, and location of the patient's anatomy 144, target 146, organ of risk (OAR) 148, or any other structure of interest (e.g., tissue, bone). For example, using image segmentation, a line can be drawn around a portion of the image and labeled as target 146 (e.g., labeled with the tag = "prostate"). Everything inside the line will be considered target 146, while everything outside the line will not be considered target 146.

[0025] In another example, dose prediction can be performed to generate dose data 150, which specifies the radiation dose to be delivered to target 146 (denoted as "D" at 152). TAR The radiation doses of OAR 148 (denoted as "D" at 154) and OAR 148 are also mentioned. OARIn practice, target 146 can represent a malignant tumor requiring radiation therapy (e.g., prostate tumor, etc.), while OAR 148 represents a nearby healthy structure or non-target structure that may be adversely affected by treatment (e.g., rectum, bladder, etc.). Target 146 is also referred to as the planned target volume (PTV). Although in Figure 1 An example is shown where the treatment volume may include multiple targets 146 and OAR 148 with complex shapes and sizes. Furthermore, although shown as having regular shapes (e.g., cubes), voxels 142 can have any suitable shape (e.g., irregular). Depending on the desired implementation, radiotherapy planning at box 130 can be performed based on any additional and / or alternative data, such as prescriptions, disease staging, biological or radiological data, genetic data, laboratory data, biopsy data, past treatment or medical history, any combination thereof, etc.

[0026] Based on structural data 140 and dose data 150, a treatment plan 156, including 2D fluence map data, can be generated for a set of beam orientations or angles. Each fluence map specifies the intensity and shape of the radiation beam emitted from the radiation source at a specific beam orientation and at a specific time (e.g., determined by a multi-leaf collimator (MLC)). For example, in practice, intensity-modulated radiotherapy (IMRT) or any other treatment technique may include varying the shape and intensity of the beam while keeping the gantry and treatment bed angles constant. Alternatively or additionally, the treatment plan 156 may include machine control point data (e.g., jaw and leaf positions), volume-modulated arc therapy (VMAT) trajectory data for controlling the treatment delivery system, etc. In practice, box 130 can be executed based on a target dose prescribed by a clinician (e.g., oncologist, dosimeter, planner, etc.), such as based on the clinician's experience, tumor type and extent, patient geometry and condition, etc.

[0027] exist Figure 1 At location 160, during the treatment phase, treatment delivery is performed to deliver radiation to the patient according to treatment plan 156. For example, the radiotherapy treatment delivery system 160 may include a rotatable gantry 164 to which a radiation source 166 is attached. During treatment delivery, the gantry 164 rotates about a patient 170 supported on a structure 172 (e.g., a table) to emit radiation beams 168 in various beam orientations according to treatment plan 156. A controller 162 can be used to retrieve treatment plan 156 and control the gantry 164, radiation source 166, and radiation beams 168 to deliver radiotherapy treatment according to treatment plan 156.

[0028] It should be understood that any suitable radiotherapy delivery system can be used, such as robotic arm-based systems, tomography systems, brachytherapy, Sirex balls, any combination thereof, etc. Furthermore, the examples of this disclosure can be applied to particle delivery systems (e.g., protons, carbon ions, etc.). Such systems can employ a scattered particle beam subsequently shaped by a device similar to an MLC, or a scanning beam with adjustable energy, spot size, and residence time. Moreover, OAR segmentation can be performed, and automated segmentation of the applicator can be expected.

[0029] Figure 2 This is a schematic diagram illustrating an example imaging system 200. Although an example is shown, the imaging system 200 may have replaceable or additional components, depending on the desired implementation in practice. (See the exemplary appendix.) Figure 2 In the imaging system 200, there are radiation sources 210; detectors 220 having pixel detectors disposed opposite to radiation sources 210 along projection lines (defined below; see 285); a first set of fan blades 230 disposed between radiation sources 210 and detectors 220; and a first fan blade driver 235 for holding the fan blades 230 and setting their positions.

[0030] The imaging system 200 may also include a second set of fan blades 240 disposed between the radiation source 210 and the detector 220, and a second fan blade driver 245 for holding and positioning the fan blades 240. The edges of the fan blades 230-240 may be oriented substantially perpendicular to the scan axis 280 and substantially parallel to the trans-axial dimension of the detector 220. The fan blades 230-240 are typically positioned closer to the radiation source 210 than the detector 220. They may remain wide open to expose the entire area of ​​the detector 220 to radiation, but may be partially closed in certain situations.

[0031] The imaging system 200 may further include: a frame 250 that holds at least the radiation source 210, the detector 220, and fan blade actuators 235 and 245 in a fixed or known spatial relationship with each other; and a mechanical actuator 255 that rotates the frame 250 about a target object 205 disposed between the radiation source 210 and the detector 220, wherein the target object 205 is disposed between the fan blades 230 and 240 on one side and the detector 220 on the other side. The term "frame" may cover all configurations of one or more structural components capable of holding the aforementioned components in a fixed or known (but possibly movable) spatial relationship. For visual simplicity in the figures, the frame housing, frame supports, and fan blade supports are not shown.

[0032] Additionally, the imaging system 200 may include a controller 260, a user interface 265, and a computer system 270. The controller 260 may be electrically coupled to the radiation source 210, mechanical actuator 255, fan blade actuators 235 and 245, detector 220, and user interface 265. The user interface 265 may be configured to enable a user to at least initiate a scan of the target object 205 and collect measured projection data from the detector 220. The user interface 265 may be configured to present a graphical representation of the measured projection data. The computer system 270 may be configured to perform any suitable operations, such as tomographic image reconstruction and analysis according to examples of this disclosure.

[0033] The rack 250 can be configured to rotate around the target object 205 during scanning, such that the radiation source 210, fan blades 230 and 240, fan blade drivers 235 and 245, and detector 220 rotate around the target object 205. More specifically, the rack 250 can rotate these components around the scan axis 280. Figure 2 As shown, the scanning axis 280 intersects the projection line 285 and is generally perpendicular to the projection line 285. The target object 205 is generally aligned with the scanning axis 280 in a substantially fixed relationship. This configuration provides a relative rotation between the projection line 285 on one side and the scanning axis 280 on the other side and the target object 205 aligned therewith, which is measured by an angular displacement value θ.

[0034] Mechanical actuator 255 can be coupled to rack 250 to provide rotation according to commands from controller 260. The pixel detector array on detector 220 can be periodically read to obtain data for radiographic projection (hereinafter also referred to as "measurement projection data"). Detector 220 has an X-axis 290 and a Y-axis 295 perpendicular to each other. X-axis 290 is perpendicular to the plane defined by scan axis 280 and projection line 285, while Y-axis 295 is parallel to the same plane. Each pixel on detector 220 is assigned discrete (x, y) coordinates along X-axis 290 and Y-axis 295. For visual clarity, a smaller number of pixels are shown in the figure. Detector 220 can be centered on projection line 285 to achieve full fan-shaped imaging of target object 205, can be offset from projection line 285 to achieve half-fan-shaped imaging of target object 205, or can be moved relative to projection line 285 to allow both full fan-shaped and half-fan-shaped imaging of target object 205.

[0035] Traditionally, the task of reconstructing 3D volumetric data (e.g., representing target object 205) from 2D projection data is often daunting. As used herein, the term "2D projection data" (which may be used interchangeably with "2D projection image") generally refers to data representing the characteristics of the irradiated radiation rays transmitted through target object 205 using any suitable imaging system 200. In practice, 2D projection data can be considered as a set of line integrals from the output of imaging system 200. 2D projection data can contain imaging artifacts and originate from different 3D configurations due to motion, etc. Any artifacts in the 2D projection data can affect the quality of subsequent diagnostic and radiotherapy treatment planning.

[0036] Artificial Intelligence (AI) Engine

[0037] According to examples in this disclosure, AI engines can be used to improve tomographic image reconstruction and analysis. As used herein, the term "AI engine" can refer to any suitable hardware and / or software component of a computer system capable of executing algorithms based on any suitable AI model. Depending on the desired implementation, an "AI engine" can be a machine learning engine based on a machine learning model, a deep learning engine based on a deep learning model, etc. Typically, deep learning is a subset of machine learning, where multi-layered neural networks are used for feature extraction as well as pattern analysis and / or classification. A deep learning engine can include a hierarchical structure of "processing layers" for non-linear data processing, which includes an input layer, an output layer, and multiple (i.e., two or more) "hidden" layers between the input and output layers. The processing layers can be trained end-to-end (e.g., from the input layer to the output layer) to extract features from the input and classify the features to produce an output (e.g., a classification label or category).

[0038] Depending on the desired implementation, any suitable AI model can be used, such as convolutional neural networks, recurrent neural networks, deep belief networks, generative adversarial networks (GANs), autoencoders, variational autoencoders, long short-term memory architectures for tracking purposes, or any combination thereof. In practice, neural networks are typically formed using networks of processing elements (called "neurons," "nodes," etc.) interconnected via connections (called "synapses," "weight data," etc.). For example, convolutional neural networks can be implemented using any suitable architecture, such as UNet, LeNet, AlexNet, ResNet, VNet, DenseNet, OctNet, etc. The "processing layers" of a convolutional neural network can be convolutional layers, pooling layers, depooling layers, rectified linear unit (ReLU) layers, fully connected layers, loss layers, activation layers, dropout layers, transposed convolutional layers, cascaded layers, or any combination thereof. Due to the large amount of data associated with tomographic image data, non-uniform sampling of 3D volumetric data can be achieved, such as using OctNet, patch / block processing, etc.

[0039] In more detail, Figure 3 This is a schematic diagram illustrating an example system 300 for reconstruction and analysis of tomographic images using corresponding AI engines 301-302. As used herein, the term "tomographic image" can generally refer to any suitable data generated using a computed tomography procedure such as CT, CBCT, PET, MRT, SPECT, etc. In practice, tomographic images can be 2D (e.g., slice images depicting cross-sections of an object); 3D (e.g., volume data representing an object); or 4D (e.g., 3D volume data varying over time).

[0040] exist Figure 3 At position 301 (left path), a first AI engine can be trained to perform tomographic image reconstruction. The first AI engine 301 may include a first processing layer (see 311) forming a first neural network labeled "A", a post-insertion projection module (see 312), and a second processing layer (see 313) forming a second neural network labeled "B". Network "A" 311 includes layers denoted as A1, A2, ... A N1 Multiple (N1>1) first processing layers, and network “B”313 includes layers denoted as B1, B2, ... B N2 Multiple (N2>1) second processing layers.

[0041] If will use Figure 4 and Figure 5Further described, the first AI engine 301 can be trained to perform 2D-to-3D transformations by transforming input = 2D projected image data (see 310) into output = 3D feature volume data (see 340). During the training phase, the first processing layer 311 and the second processing layer 313 can be linked by the post-projection module 312 and trained together. Thus, during the subsequent inference phase, the first AI engine 301 can utilize data in both the 2D projection space and the 3D volumetric space during tomographic image reconstruction.

[0042] exist Figure 3 At point 302 (right path), a second AI engine can be trained to perform tomographic image analysis. The second AI engine 302 may include a first processing layer (see 314) forming a first neural network labeled "C", an inserted forward projection module (see 315), and a second processing layer (see 316) forming a second neural network labeled "D". Network "C" 314 includes denoted as C1, C2, ... C M1 Multiple (M1>1) first processing layers. The network “D”316 includes layers denoted as D1, D2, ... D M2 Multiple (M2>1) second processing layers.

[0043] If will use Figure 6 and Figure 7 Further described, a second AI engine 302 can be trained to transform the input—3D feature volume data (see 340)—into 2D feature data (see 360) for analysis. During the training phase, the first processing layer 314 (C1, C2, ... C... M1 ), and the second processing layer 316 (D1, D2, ... D M2 The second AI engine 302 can be linked and trained together with the forward projection module 315. Thus, during the subsequent inference phase, the second AI engine 302 can utilize data in both the 2D projection space and the 3D volumetric space during tomographic image analysis. In fact, network “D” 316 can be trained based on 2D feature data 360 and the original 2D projection data 310 (see...). Figure 3 (The dashed line in the middle) performs the analysis.

[0044] According to an example of the invention, AI engines 301 / 302 can learn from data in both 2D projection space and 3D volumetric space. Thus, the transformation between 2D projection space and 3D volumetric space can be performed in a substantially lossless manner compared to conventional reconstruction methods, reducing the likelihood of losing necessary features. Using examples of this disclosure, different building blocks for tomographic image reconstruction can be combined with neural networks (i.e., the example "AI engine"). Potential application areas can include automatic segmentation of 3D volumetric data or 2D projection data, object / feature detection, classification, data augmentation (e.g., completion, artifact reduction), any combination thereof, etc.

[0045] Unlike traditional methods, the examples disclosed herein utilize the 3D space of volume data and the 2D space of projection data. Since the 2D projection data and 3D volume data are two representations of the same target object, it can be assumed that analysis or processing may be beneficial in one or the other. In practice, the output 3D volume data 340 / 350 can be a 3D / 4D volume with CT(HU) values, dose data, segmentation / structural data, deformation vectors, 4D time-resolved volume data, any combination thereof, etc. The output 2D feature data 360 (projection) can be X-ray intensity data, attenuation data (both potentially energy-resolved), their modified forms (removed objects), segments, any combination thereof, etc.

[0046] Based on the examples in this disclosure, the first assumption is that the raw data of the tomographic image contains more information than the resulting 3D volumetric data. In practice, image reconstruction can be slightly adjusted for different tasks, such as noise suppression, spatial resolution, edge enhancement, Henle unit (HU) accuracy, and any combination thereof. These slight adjustments typically involve trade-offs, meaning that potentially useful information for any subsequent image analysis (e.g., segmentation) is lost. Other information (e.g., motion) can be reconstructed to suppress it. In fact, once image reconstruction is performed, the 2D projection data is only reviewed in more detail after problems are seen or understood regarding features (e.g., metallic or artifact-like features) in the 3D volumetric image data.

[0047] Based on examples from this disclosure, the second assumption is that the analysis of 2D projection data benefits from knowledge about the 3D image domain. Typical examples may include prior reconstruction with volumetric manipulation, followed by forward projection for purposes such as background subtraction and tumor detection. Unlike traditional methods, the 2D-3D relationship can be an inherent part or component of a machine learning engine. The processing layer can learn any suitable information from both the 2D projection data and the 3D volume data to accomplish the task.

[0048] Depending on the desired implementation, the first AI engine 301 and the second AI engine 302 can be trained and deployed independently (see [link to implementation details]). Figure 4-7). Alternatively, the first AI engine 301 and the second AI engine 302 can be trained and deployed in an integrated manner (see Figure 8 The first AI engine 301 and the second AI engine 302 can be implemented using computer system 270 or a separate computer system. The computer system can be connected to the controller 260 of the imaging system 200 via a local area network (LAN) or wide area network (WAN) (e.g., the Internet). (The last sentence appears to be incomplete and possibly refers to a different implementation.) Figure 10 As described, computer system 270 can provide planning services (PaaS) for users (e.g., clinicians) to perform tomographic image reconstruction and / or analysis.

[0049] Tomography image reconstruction

[0050] According to the first aspect of this disclosure, training is possible. Figure 3 The first AI engine 301 in the system is used to perform tomographic image reconstruction. It will use... Figure 4 Explain some examples, Figure 4 This is a flowchart of an example process 400 in which a computer system performs tomographic image reconstruction using a first AI engine 301. The exemplary process 400 may include one or more operations, functions, or actions illustrated by one or more boxes such as 410 to 440. Various boxes may be combined into fewer boxes, divided into additional boxes, and / or eliminated based on desired implementation methods. The example process 400 can be implemented using any suitable computer system; examples of it will use… Figure 10 Let's discuss this.

[0051] (a) Inferring phase

[0052] exist Figure 4 At point 410, 2D projection data 310 associated with a target object (e.g., the patient's anatomy) can be obtained. Here, the term "obtain" can generally refer to receiving or retrieving 2D projection data 310 from a source (e.g., controller 260, storage device, another computer system, etc.). For example, using... Figure 2 As explained, imaging system 200 can be used to acquire 2D projection data 310 by rotating radiation source 210 and detector 220 around target object 205.

[0053] In practice, the 2D projection data 310 can be raw data from the controller 260 or preprocessed raw data. Example preprocessing algorithms may include defect pixel correction, dark field correction, conversion from transmission integral to attenuation integral (e.g., logarithmic normalization using the air norm), scattering correction, beam hardening correction, decimation, etc. The 2D projection data 310 can be multichannel projection data including various preprocessed instances from the acquisition sequence and additional projections. It should be understood that any suitable tomographic imaging modality can be used to capture the 2D projection data 310, such as X-ray tomography (e.g., CT and CBCT), PET, SPECT, MRT, etc. Although digital tomography synthesis (DTS) imaging is not a direct tomography method, the same principles apply. This is because DTS also uses the relative geometry between projections to calculate a relative 3D reconstruction with finite (scan arc angle dependent) resolution in the imaging direction.

[0054] exist Figure 4 In the 420, the first processing layer (A1, A2, ..., A311) of the preprocessing network "A"311 can be used. N1 The network "A" 311 is used to process the 2D projection data 310 to generate 2D feature data 320. In practice, the network "A" 311 may include a convolutional neural network with convolutional layers, pooling layers, etc. All projections in the 2D projection data 310 can be processed using the network "A" 311 or several instances of different subsets used for projection.

[0055] exist Figure 4 In the diagram 430, the first 3D feature volume data 330 can be reconstructed from the 2D feature data 320 using the post-projection module 312. As used herein, "post-projection" generally refers to a transformation from 2D projection space to 3D volume space. Any suitable reconstruction algorithm can be implemented by the post-projection module 312, such as non-iterative reconstruction (e.g., filtered post-projection), iterative reconstruction (e.g., algebraically and statistically based reconstruction), etc. In practice, the 2D feature data 320 can represent the multi-channel output of network "A" 311. In this case, the post-projection module 312 can perform multiple post-projection operations on the corresponding channels to form the corresponding 3D feature volume data 330 with a multi-channel 3D representation.

[0056] exist Figure 4 In the 440, the second processing layer (B1, B2, ..., B) of network "B"313 can be used. N2 The network "B" 313 is used to process the first 3D feature volume data 330 to generate the second 3D feature volume data 340. Depending on the desired implementation, the network "B" 313 can be implemented based on the UNet architecture. Having a general "U" shape, the left path of the UNet is called the "encoding path" or "shrinking path," in which high-order features are extracted at several downsampling resolutions.

[0057] In one example, network "B" 313 can be configured to implement the encoding path of UNet, in which case the second processing layer (B1, B2, ..., B...) N2 This may include convolutional and pooling layers that form a volumetric processing chain. Network "B" 313 can be viewed as another type of encoder that finds a representation of the 2D projection data 310. (As will be used...) Figure 6 The right path of UNet discussed is referred to as the "decoding path" or "expansion path" and can be implemented by the network "C" 314 of the second AI engine 302.

[0058] (b) Training Phase

[0059] Figure 5 This is a schematic diagram of example 500 illustrating the training and inference phases of a first AI engine 301 for tomographic image reconstruction. During the training phase 501, the first processing layer (A1, A2, ..., A...) of network "A" 311... N1 ) and the second processing layer (B1, B2, ..., B) of network "B" 313 N2 The first AI engine 301 can be linked together and trained together via the post-projection module 312 to learn associated weight data. Based on the training data 510-520, the first AI engine 301 can learn the weights associated with the first processing layer (A1, A2, ..., A...). N1 The first weighted data associated with ) A1 ,w A2 ,...,w AN ), and the second processing layer (B1, B2, ..., B N2 The associated second weight data (w) B1 ,w B2 ,...,w BN ).

[0060] Depending on the desired implementation, networks "A" 311 and "B" 313 can be trained using supervised learning methods. The purpose of training phase 501 is to train the AI ​​engine 301 to map the input training data = 2D projection data 510 to the output training data = 3D feature volume data 520, which represents the desired result or the volume to which it belongs. In practice, the 3D feature volume data 520 represents the labels used for supervised learning, and annotations such as contours can be used as labels. For each iteration, a subset or all of the 2D projection data 510 can be processed using network "A" 311 to generate 2D feature data 530, a subset or all of the 2D projection data 510 can be processed using the post-projection module 312 to generate 3D feature volume data 540, and a subset or all of the 2D projection data 510 can be processed using network "B" 313 to generate a prediction result (see 550).

[0061] Figure 5 The training phase 501 can be guided by estimating and minimizing the loss between the predicted result 550 and the expected result specified by the output training data 520. See also Figure 5 The comparison operation at position 560. Thus, the first weighted data (w) A1 ,w A2 ,...,w AN ) and second weighted data (w B1 ,w B2 ,...,w BN The loss function can be improved during the training phase 501, for example, through backpropagation of the loss. A simple example of a loss function would be the mean squared error between the true and predicted results, but loss functions can have more complex formulas (e.g., dice loss, Jaccard loss, focus loss, etc.). The loss can be estimated from the model's output or from any discrete point within the model.

[0062] Depending on the desired implementation, network "A" 311 can be trained to perform preprocessing on 2D projected data 310, such as by applying convolutional filters to the 2D projected data 310. Typically, network "A" 311 can learn any suitable feature transformation necessary for network "B" 313 to generate its output (i.e., the second 2D feature volume data 350). For example, using the Feldkamp-Davis-Kress (FDK) reconstruction algorithm, network "A" 311 can be trained to learn the convolutional filter portion of the FDK algorithm. In this case, network "B" 313 can be trained to generate the second 2D feature volume data 350 representing the 3D FDK reconstruction output. During training phase 501, network "A" 311 can learn any suitable task that can be optimally performed on line integrals based on the 2D projected data 310 in the 2D projected space.

[0063] Once trained and validated, the first AI engine 301 can be deployed to perform tomographic image reconstruction on the current patient during the inference phase 502. (As used...) Figure 3 and Figure 4 As described, the first AI engine 301 can operate in both 2D projection space and 3D volumetric space to transform input = 2D projection data 310 into output = feature volume data 340 using network "A" 311, post-projection module 312, and network "B" 313. [Already used] Figure 3 and Figure 4 Various examples related to 2D to 3D conversion using the first AI engine 301 were discussed, and will not be repeated here for the sake of brevity.

[0064] It depends on how the expectation is achieved. Figure 3 and Figure 5 Network "B" 313 in the diagram can represent a combination of encoding network "B" 313 and decoding network "C" 314. These networks 313-314 together form an autoencoder network to find a 3D representation of 2D projection data 310 in the form of output = 3D feature volume data 340 / 350. Once trained and validated, tomographic image reconstruction can be performed using network "A" 311, post-projection module 212, and combined networks 313-314. Depending on the desired implementation, the final output of combined networks 313-314 (i.e., feature volume data 340 / 350) can be used as input to an AI engine or algorithm for 3D analysis. The output of combined networks 313-314 can also include 4D temporally resolved volume data or other suitable representations.

[0065] Tomography image analysis

[0066] According to the second aspect of this disclosure, training is possible. Figure 3 The second AI engine 302 in the system is used to perform tomographic image analysis. It will use... Figure 6 Explain some examples, Figure 6 This is a flowchart of an example process 600 in which a computer system uses a second AI engine 302 to perform tomographic image analysis. The exemplary process 600 may include one or more operations, functions, or actions illustrated by one or more boxes such as 610 to 640. Various boxes may be combined into fewer boxes, divided into additional boxes, and / or eliminated based on desired implementation methods. The example process 600 can be implemented using any suitable computer system and will use… Figure 10 Let's discuss an example. In the following text, from the perspective of network "C" 314, the input feature volume data 340 and the output feature volume data 350 will be used as example "first" and "second" 3D feature volume data.

[0067] (a) Inference stage

[0068] exist Figure 6 In step 610, input = 3D feature volume data 340 can be obtained. In one example, the input 3D feature volume data 340 can be the output of the first AI engine 301 or its algorithmic equivalent. In the latter case, any suitable algorithm can be used, such as 3D or 4D reconstruction algorithms. The term "obtain" can generally refer to receiving or retrieving the 3D feature volume data 340 from a source (e.g., the first AI engine 301, a storage device, another computer system, etc.). The input 3D feature volume data 340 can be generated based on 2D projection data 210, which is acquired using the imaging system 200 by rotating the radiation source 210 and the detector 220 around the target object 205.

[0069] exist Figure 6 At position 620, the first processing layer (C1, C2, ..., C) of network "C" 314 can be used. M1 The input 3D feature volume data 340 is processed to generate output 3D feature volume data 350. Typically, network "C" 314 is trained to prepare features that can be forward-projected by the forward projection module 315 and processed by network "D" 316, for example, to reproduce the input projection or segmentation. Depending on the desired implementation, network "C" 314 can be implemented based on the UNet architecture. The right path of UNet is referred to as the "decoding path" or "expansion path," where lower-resolution features are upsampled to a higher resolution.

[0070] In one example, network "C" 314 can be configured to implement the decoding path, in which case the first processing layer (C1, C2, ..., C...) M1 It can include one or more convolutional layers and one or more depooling layers. When combined with Figure 3 When network "B" 313 is connected, encoding is done using network "B" 313 and decoding is done using network "C" 314. Both networks 313 and 314 can be viewed as a type of encoder-decoder. In this case, compared to the input of network "C" 314 = 3D feature volume data 340 (i.e., the output of network "B" 313), the output = 3D feature volume data 350 can have potentially higher (upsampled) resolution or denoised, higher-dimensional features.

[0071] exist Figure 6 At position 630, the forward projection module 315 can be used to forward project or transform the 3D feature volume data 350 (i.e., the output of network "C" 314) into 2D feature data 360. As used herein, "forward projection" can generally refer to a transformation from 3D volume space to 2D projection space. Forward projection (also known as synthetic projection data) can include data such as attenuated path integrals (original signal), Rayleigh scattering, and Compton scattering. The forward projection module 315 can implement any suitable algorithm, such as monochrome or multicolor; source driver or destination driver; voxel-based or blob-based; employing ray tracing, Monte Carlo methods, finite element methods, etc.

[0072] exist Figure 6 At position 640, the second processing layer (D1, D2, ..., D) of network "D"316 can be used. M2Network "D" processes 2D feature data 360° to produce analytical output data. Depending on the analysis performed by network "D", any suitable architecture can be used, such as UNet, LeNet, AlexNet, ResNet, V-net, DenseNet, etc. Example analyses performed by network "D" 316 will be discussed below.

[0073] (b) Training Phase

[0074] Figure 7 This is a schematic diagram of example 700 illustrating the training and inference phases of a second AI engine 302 for tomographic image analysis. During the training phase 701, the first processing layer (C1, C2, ..., C...) of network "C" 314... M1 ) and the second processing layer (D1, D2, ..., D) of network “D”316 M2 The first AI engine 301 can be linked and trained together by the forward projection module 315 to learn associated weight data. Based on the training data 710-720, the first AI engine 301 can learn the weights associated with the first processing layer (C1, C2, ..., C...). M1 The first weighted data associated with ) C1 w C2 , ..., w CM ), and with the second processing layer (D1, D2, ..., D M2 The associated second weight data (w) D1 w D2 , ..., w DM ).

[0075] Depending on the desired implementation, networks "C" 314 and "D" 316 can be trained using supervised learning methods. The purpose of training phase 701 is to train AI engine 302 to map input training data = 3D feature volume data 710 to output training data = analysis output data 720, which represents the desired result. For each iteration, (a) network "C" 314 can be used to process a subset of the 3D feature volume data 710 to generate decoded 3D feature volume data 730, (b) a forward projection module 315 can be used to generate 2D feature data 740, and (c) network "D" 316 can be used to generate prediction results (see 750).

[0076] Similar to Figure 5 Examples in, Figure 7 The training phase 701 can be guided by estimating and minimizing the loss between the predicted result 750 and the expected result specified by the output training data 720. See also Figure 7 The comparison operation is performed at position 760. Thus, the first weighted data (w) C1 w C2 , ..., wCM ) and second weighted data (w D1 w D2 , ..., w DM The loss function can be improved during the training phase, for example, through backpropagation of the loss. Similarly, a simple loss function (e.g., mean squared error) or a more complex function can be used.

[0077] Depending on the desired implementation, training data 710–720 can be used to train network “D”316 to generate analytical output data associated with one or more of the following: automatic segmentation, object detection (e.g., organs or bones), feature detection (e.g., edges / contours of organs, small-scale 3D structures within bones such as the skull, etc.), image artifact suppression, image enhancement (e.g., resolution enhancement using super-resolution), detrusion by learning volumetric image content (voxels), prediction of moving 2D segments, object or tissue removal (e.g., bones, patient tables or fixation devices, etc.), and any combination thereof. These examples will be discussed further below.

[0078] Once trained and validated, the second AI engine 302 can be deployed to perform tomographic image analysis on the current patient during the inference phase 702. (As used...) Figure 3 and Figure 6 As described, the second AI engine 302 can operate in both 2D projection space and 3D volumetric space to transform the input = 3D feature volume data 340 into the output = analysis output data 370 using network "C" 314, forward projection module 315, and network "D" 316. Already used Figure 3 and Figure 6 Example details related to tomographic image analysis using the second AI engine 302 are discussed, and will not be repeated here for the sake of brevity.

[0079] Integrated Tomography Image Reconstruction and Analysis

[0080] According to the third aspect of this disclosure, they can be trained together. Figure 3 The first and second AI engines 301-302 in the system are used to perform integrated tomographic image reconstruction and analysis. This will utilize... Figure 8 Let's discuss some examples. Figure 8 This is a schematic diagram illustrating example training phase 800 of the AI ​​engines 301-302 used for integrated tomographic image reconstruction and analysis.

[0081] exist Figure 8In the example, AI engines 301-302 can be connected to form an integrated AI engine, which includes networks "A" 311 and "B" 313 with a back projection module 312 inserted, followed by networks "C" 314 and "D" 316 with a forward projection module 315 inserted. The purpose of the training phase 801 is to train the integrated AI engines 301-302 to map the input training data = 2D projected data 810 to the output training data = analysis output data 820, which represents the desired result. For each iteration, a subset of the 2D projected data 810 can be processed using networks "A" 311, back projection module 312, network "B" 313, network "C" 314, forward projection module 315, and network "D" 316 to generate a prediction result (see 830).

[0082] Figure 8 The training phase 801 can be guided by estimating and minimizing the loss between the predicted result 830 and the expected result specified by the output training data 820. Using an end-to-end training method, the weight data associated with the corresponding networks 311, 313-314, and 316 can be improved, for example, through backpropagation of the loss. By embedding a post-projection module 312 and a forward projection module 315, the training phase 901 can be guided by an end-to-end loss function in the 2D projection space and / or 3D volumetric space. See also Figure 8 Comparison between the output training data 820 and the prediction results 830 in section 860.

[0083] exist Figure 8 In the example, an optional copy of the data from the first AI engine 301 can be sent to the second AI engine 302 to “skip” the processing layers in between. This provides a shortcut for the data flow, such as allowing high-frequency features to skip or bypass lower layers of the neural network. In one example (see 840), an optional copy of the data from a processing layer (Ai) in network “A” 311 can be provided to another processing layer (Dj) in network “D” 316. In practice, the scattering data removed by network “A” 311 will skip networks “B” 313 and “C” 314 and be added again to reproduce the input projection or patient motion removed by network “B” 313. In this way, static image data can be generated, and network “C” 314 can reproduce the input. In another example (see 850), an optional copy of the data from a processing layer (Bi) in network “B” 313 can be provided to another processing layer (Cj) in network “C” 314. This skipping method is a possibility provided by convolutional neural networks, which offer the possibility of skipping layers.

[0084] It depends on how the expectation is achieved. Figure 3 -8 The first AI engine 301 and / or the second AI engine 302 in the form of AI engine 301 can be implemented to facilitate at least one of the following:

[0085] (a) Identifying imaging artifacts associated with 2D projection data 310 and / or 3D feature volume data 340 / 350, such as when the exact source of the artifact is unknown. In one approach, an autoencoding method can be implemented using AI engines 301-302. The loss function used during training phase 701 can be used to ensure that the 2D projection data 310 and the analysis output data 370 are substantially identical, and that the volume data 330-350 therebetween have the desired quality (e.g., reduced noise or motion artifacts). Another approach is to provide an ideal reconstruction as a label and train the model to predict substantially artifact-free volume data based on reduced or degraded (e.g., simulated noise or scattering) projection data.

[0086] (b) Identify the motion regions associated with 2D projection data 310 and / or 3D feature volume data 340 / 350. In this case, training data 710-720 may include 2D / 3D information of the motion occurring to train network “D” 316 to identify the motion regions.

[0087] (c) Identify regions with artifacts associated with 2D projection data 310 and / or 3D feature volume data 340 / 350. In this case, training data 710-720 may include segmented artifacts to train network “D” 316 to identify regions with artifacts.

[0088] (d) Identify anatomical and / or non-anatomical structures from 2D projection data 310 and / or 3D feature volume data 340 / 350. Anatomical structures such as tumors and organs can be identified through automatic segmentation. Non-anatomical structures may include implants, fixation devices, and other materials in 2D / 3D image regions. In this case, training data 710-720 may include data for identifying such anatomical and / or non-anatomical structures.

[0089] (e) Reduce noise associated with 2D projection data 310 and / or 3D feature data 340 / 350. In practice, this may include identifying projection sequences for further processing (e.g., marker tracking, soft tissue tracking).

[0090] (f) Tracking the movement of patient structures identifiable from 2D projection data 310 or 3D feature volume data 340 / 350. A feasible output of the first AI engine 301 and the second AI engine 302 could be a set of projections, where each pixel indicates the probability of identifying a reference (or any other structure) center point or segment. This would provide the location of the structure for each protrusion. An advantage of this method is that occurrence probabilities can be combined in 3D volume space to make a 2D prediction dependent on each projection. Any suitable tracking method can be used, such as 3D volume data in the form of Long Short-Term Memory (LSTM).

[0091] (g) The 2D slices associated with the 2D projection data 310 are divided into different movement modules (or stages). By identifying the modules, the network “D” 316 can be trained to use data belonging to a particular module.

[0092] (h) Generate 4D image data with motion associated with 2D projection data 310 or 3D feature volume data 340 / 350. In this case, network “D” 316 can be trained to compute a volume with several channels (4D) for different chambers in (g) to resolve motion. Other possibilities include using variational autoencoders in 3D volume space (e.g., networks “B” 313 and “C” 314) to learn deformable models.

[0093] Automated segmentation using AI engines 301-302 should contrast with traditional manual methods. For example, teams of highly skilled and trained oncologists and dosimetry specialists typically manually delineate structures of interest by drawing contours or segments on image data 120. These structures are manually reviewed by physicians and may require adjustments or redrawing. In many cases, segmentation of critical organs can be the most time-consuming part of radiotherapy planning. After these structures are agreed upon, additional labor-intensive steps exist to process them to generate a clinically optimal treatment plan that specifies treatment delivery data, such as beam orientation and trajectory, and corresponding 2D fluence maps. These steps are often complicated by a lack of consensus among different physicians and / or clinical areas regarding what constitutes a “good” contour or segmentation. In fact, the way different clinical experts draw structures or segments can vary considerably. This variation can lead to uncertainty regarding the precise proximity, size, and shape of the target area and the OAR (Organizational Area) to which the minimum radiation dose should be received. Even for a specific expert, the way segments are drawn may vary from day to day.

[0094] Example treatment plan

[0095] Figure 9 Based on Figure 3 A schematic diagram of an exemplary treatment plan 156 / 900 generated or improved from the output data of the AI ​​engine 301 / 302. The treatment plan 156 can be delivered using any suitable treatment delivery system, which includes a radiation source 910 to project a radiation beam 920 at various beam angles 930 onto a treatment volume 960 representing the patient's anatomy.

[0096] Although for the sake of simplicity Figure 9Not shown, but radiation source 910 may include a linear accelerator for accelerating radiation beam 920 and a collimator (e.g., MLC) for modifying or modulating radiation beam 920. In another example, radiation beam 920 can be modulated by scanning radiation beam 920 on a target patient in a specific mode having various energies and residence times (e.g., as in proton therapy). A controller (e.g., a computer system) may be used to control the operation of radiation source 920 according to treatment plan 156.

[0097] During treatment delivery, the radiation source 910 can be rotatable, either by using a gantry surrounding the patient or by rotating the patient (as in some proton radiotherapy solutions), to emit a radiation beam 920 relative to the patient at various beam orientations or angles. For example, a deep learning engine configured to perform treatment delivery data estimation can be used to select five equally spaced beam angles 930A–E (also labeled “A”, “B”, “C”, “D”, and “E”). In practice, any suitable number of beam and / or table or chair angles 930 can be selected (e.g., five, seven, etc.). At each beam angle, the radiation beam 920 is associated with a fluence plane 940 (also called a cross plane) located outside the patient envelope along a beam axis extending from the radiation source 910 to the treatment volume 960. Figure 9 As shown, the injection plane 940 is typically at a known distance from the isocenter.

[0098] In addition to beam angles 930A-E, fluence parameters of the radiation beam 920 are also required for treatment delivery. The term "fluence parameter" can generally refer to the characteristics of the radiation beam 920, such as its intensity distribution represented using fluence maps (e.g., 950A-E corresponding to beam angles 930A-E). Each fluence map (e.g., 950A) represents the intensity of the radiation beam 920 at each point on the fluence plane 940 at a specific beam angle (e.g., 930A). Treatment delivery can then be performed according to the fluence maps 950A-E, for example, using IMRT. The radiation dose deposited according to the fluence maps 950A-E should correspond as closely as possible to the treatment plan generated according to the examples of this disclosure.

[0099] Computer System

[0100] The examples disclosed herein can be deployed in any suitable manner, such as standalone systems, web-based PaaS (PaaS) systems, etc. In the following text, [the examples will be used]. Figure 10 To describe the example computer system (also known as the "planning system"), Figure 10 This is a schematic diagram illustrating an example network environment 1000 in which tomographic image reconstruction and / or tomographic image analysis can be performed. Depending on the desired implementation, the network environment 1000 may include, in addition to... Figure 10Additional and / or replaceable components not shown. Examples of this disclosure may be implemented by hardware, software, or firmware, or a combination thereof.

[0101] Processor 1020 executes here (refer to the following) Figures 1 to 9 The described process. The computer-readable storage medium 1030 may store computer-readable instructions 1032, which, in response to execution by the processor 1020, cause the processor 1020 to perform the various processes described herein. The computer-readable storage medium 1030 may also store any suitable data 1034, such as data related to the AI ​​engine, training data, weight data, 2D projection data, 3D volume data, analysis output data, etc. Figure 10 In this example, computer system 1010 can be accessed by multiple user devices 1041-1043 via any suitable physical network (e.g., local area network, wide area network, etc.). In practice, user devices 1041-1043 can be operated by various users located at any suitable clinical site.

[0102] Computer system 1010 can be implemented using a multi-tiered architecture, comprising a web-based user interface (UI) layer 1021, an application layer 1022, and a data layer 1023. The UI layer 1021 can be configured to provide any suitable interface for interaction with user devices 1041-1043, such as a graphical user interface (GUI), command-line interface (CLI), application programming interface (API) calls, or any combination thereof. The application layer 1022 can be configured to implement the examples of this disclosure. The data layer 1023 can be configured to facilitate access to and from data originating from storage medium 1030. Through interaction with the UI layer 1021, user devices 1041-1043 can generate and send appropriate service requests 1051-1053 for processing by computer system 1010. In response, computer system 1010 can execute the examples of this disclosure, generating service responses 1061-1063 and sending them to the corresponding user devices 1041-1043.

[0103] Depending on the desired implementation, computer system 1010 can be deployed in a cloud computing environment, in which case multiple virtualized computing instances (e.g., virtual machines, containers) can be configured to implement various functions of layers 1021-1023. The cloud computing environment can be supported by on-premise cloud infrastructure, public cloud infrastructure, or a combination of both. Computer system 1010 can be deployed in any suitable manner, including service-oriented deployments within on-premise cloud infrastructure, public cloud infrastructure, or combinations thereof. Computer system 1010 can represent a computing cluster comprising multiple computer systems, with various functions distributed across these multiple computer systems. Computer system 1010 may include... Figure 10Any replaceable and / or additional components not shown, such as graphics processing unit (GPU), message queue for communication, blob storage or database, load balancer, dedicated circuitry, etc.

[0104] The foregoing detailed description has illustrated various embodiments of the apparatus and / or process using block diagrams, flowcharts, and / or examples. Where such block diagrams, flowcharts, and / or examples contain one or more functions and / or operations, those skilled in the art will understand that each function and / or operation in such block diagrams, flowcharts, or examples can be implemented individually and / or collectively with a wide range of hardware, software, firmware, or virtually any combination thereof. Throughout this disclosure, the terms “first,” “second,” “third,” etc., do not indicate any order of importance but are used to distinguish one element from another.

[0105] Those skilled in the art will recognize that certain aspects of the embodiments disclosed herein may be implemented, in whole or in part, equivalently in an integrated circuit as one or more computer programs running on one or more computers (e.g., as one or more programs running on one or more computer systems), as one or more programs running on one or more processors (e.g., as one or more programs running on one or more microprocessors), as firmware, or as any combination thereof, and that, in view of this disclosure, designing the circuit and / or writing the software and / or firmware code will be entirely within the skill of those skilled in the art.

[0106] While this disclosure has been described with reference to specific exemplary embodiments, it will be appreciated that this disclosure is not limited to the described embodiments, but can be practiced with modifications and variations within the spirit and scope of the appended claims. Therefore, the specification and drawings are to be considered illustrative rather than restrictive.

Claims

1. A method for enabling a computer system to perform tomographic image reconstruction using a first AI engine (301), wherein the method comprises: Obtain 2D projection data (310) associated with the target object (146) and acquired using the imaging system. The 2D projection data (310) is processed using the first AI engine (301), which includes multiple first processing layers (311), an inserted post-projection module (312), and multiple second processing layers (313). The processing is achieved by performing the following steps: First 2D feature data (320) is generated by processing the 2D projection data (310) using the plurality of first processing layers (311) that are associated with the first weight data and operate in the 2D projection space. The first 3D feature volume data (330) is reconstructed from the first 2D feature data (320) using the post-projection module (312); and Second 3D feature volume data (340) is generated by processing the first 3D feature volume data (330) using the plurality of second processing layers (313) associated with the second weight data and operating in 3D volume space; and The second AI engine (302) is used to process the second 3D feature volume data (340) generated by the first AI engine (301). The second AI engine (302) is trained to perform tomographic image analysis to transform the second 3D feature volume data (340) into analysis output data (370). The processing of the second 3D feature volume data (340) using the second AI engine (302) includes: The third 3D feature volume data (350) is generated by processing the second 3D feature volume data (340) using multiple third processing layers (314) of the second AI engine (302) which is associated with the third weight data and operates in the 3D volume space. The third 3D feature data (350) is transformed into second 2D feature data (360) using the forward projection module (315); and The analysis output data (370) is generated by processing the second 2D feature data (360) using multiple fourth processing layers (316) associated with the fourth weight data and operating in the 2D projection space.

2. The method according to claim 1, wherein generating the second 3D feature volume data (340) comprises: The second 3D feature data (340) is generated by processing the first 3D feature data (330) using the plurality of second processing layers (313), which are trained to encode the first 3D feature data (330).

3. The method according to claim 1 or 2, wherein generating the first 2D feature data (320) comprises: The first 2D feature data (320) is generated by processing the 2D projection data (310) using the plurality of first processing layers (311), which are trained to perform preprocessing on the 2D projection data (310) in the 2D projection space.

4. The method according to claim 3, wherein generating the first 2D feature data (320) comprises: The first 2D feature data (320) is generated by processing the 2D projection data (310) using the plurality of first processing layers (311), which are trained to apply one or more convolutional filters to the 2D projection data (310).

5. The method according to claim 1, 2 or 4, wherein the method further comprises: Obtain training data including training 2D projection data (310) and training 3D feature volume data (330); as well as The plurality of first processing layers (311) and the plurality of second processing layers (313) are trained together using the interposed post-projection module (312) to transform the training data of the training 2D projection data into the training data of the training 3D feature volume data, and to learn the corresponding first weight data and second weight data.

6. The method of claim 1, wherein the method further comprises: Obtain training data including training 2D projection data and training analysis output data; as well as The first AI engine (301) and the second AI engine (302) are trained together to perform integrated tomographic image reconstruction and analysis to transform the training data of the training 2D projection data into the training data of the training analysis output data.

7. A non-transitory computer-readable storage medium comprising an instruction set that, in response to execution by a processor of a computer system, causes the processor to perform the method according to any one of claims 1 to 6.

8. A computer system configured to perform tomographic image reconstruction using an AI engine, wherein the computer system comprises: A processor and a non-transitory computer-readable medium thereon storing instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1 to 6.

Images