by modeling image formation as image reconstruction of one or more neural networks

Abstract

Systems and methods for image reconstruction based on modeling image formation as one or more neural networks. According to one aspect, one or more neural networks are configured based on physical properties of image formation (202). The one or more neural networks are optimized using acquired test image data (204). Output images can then be reconstructed by applying current image data as input to the one or more optimized neural networks (208).

Description

Technical Field

[0001] This disclosure relates generally to image processing, and more particularly to image reconstruction based on modeling image formation as one or more neural networks. Background Technology

[0002] Single-photon emission computed tomography (SPECT) is a widely used nuclear medicine computed tomography imaging technique. For SPECT imaging, a gamma-emitter-labeled pharmaceutical agent is first administered to the subject. Then, an external device (gamma camera) is used to detect radioactivity from the body from one or more viewpoints. The planar image obtained at one viewpoint is a projection of the three-dimensional (3D) distribution onto the two-dimensional (2D) detector plane. A 3D image of the distribution of radioactive sources within the subject can be reconstructed using a sequence of planar images acquired within a certain angular range around the subject.

[0003] Several techniques exist for performing SPECT image reconstruction. One technique involves iterative reconstruction, which typically begins with a hypothetical image, calculates a projection from that image, compares the original projection data, and updates the image based on the difference between the calculated and actual projections. In this approach, the system is modeled as a probabilistic linear operator that encompasses all image-forming effects that will be considered in the reconstruction: camera or body rotation, attenuation, and even flood correction. This technique is computationally intensive because it requires recalculating the projection operator for each view at each iteration of the reconstruction, since the probability list would otherwise be too large to store.

[0004] Another common technique is based on machine learning. Machine learning-based SPECT image reconstruction typically models the problem by defining a neural network structure and then training that structure to optimize layer weights to increase reconstruction accuracy. Current deep learning methods for SPECT reconstruction are similar to those used in other scientific fields, where the network design is independent of the physics of the image formation model. Summary of the Invention

[0005] This paper describes a system and method for image reconstruction based on modeling image formation as one or more neural networks. One or more neural networks are configured according to one aspect, based on the physical properties of image formation. The one or more neural networks are optimized using acquired test image data. The output image can then be reconstructed by applying current image data as input to the one or more optimized neural networks. Attached Figure Description

[0006] A more complete understanding of this disclosure and its many accompanying aspects will be readily obtained, as this disclosure and its many accompanying aspects become better understood when considered in conjunction with the accompanying drawings by referring to the following detailed description.

[0007] Figure 1 It is a block diagram illustrating an exemplary system;

[0008] Figure 2 An exemplary image reconstruction method performed by a computer system is shown;

[0009] Figure 3 An exemplary SPECT image formation is illustrated;

[0010] Figure 4a An exemplary neural network is shown;

[0011] Figure 4b The illustration shows an exemplary simple rotation;

[0012] Figure 4c An exemplary convolutional layer for modeling motion correction is shown;

[0013] Figure 4d The illustration shows the modeling of motion correction;

[0014] Figure 4e An exemplary convolutional layer is shown that models head misalignment correction;

[0015] Figure 4f An exemplary modeling of point spread function (PSF) filter correction is illustrated;

[0016] Figure 5 An exemplary neural network architecture for SPECT image reconstruction is shown; and

[0017] Figure 6 An exemplary method for neural network optimization is shown. Detailed Implementation

[0018] In the following description, numerous specific details, such as examples of specific components, devices, methods, etc., are set forth to provide a comprehensive understanding of how this framework is implemented. However, it will be apparent to those skilled in the art that these specific details need not be adopted to practice the implementation of this framework. In other instances, well-known materials or methods have not been described in detail to avoid unnecessarily obscuring the implementation of this framework. While this framework allows for various modifications and alternatives, specific embodiments thereof are shown by way of example in the accompanying drawings and will be described in detail herein. However, it should be understood that the invention is not intended to be limited to the specific forms disclosed, but rather, it is intended to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the invention. Furthermore, for ease of understanding, certain method steps are depicted as separate steps; however, these separately depicted steps should not be construed as necessarily depending on the order in which they are performed.

[0019] As used herein, the term "x-ray image" can refer to a visible x-ray image (e.g., displayed on a video screen) or a digital representation of an x-ray image (e.g., a file corresponding to the pixel output of an x-ray detector). The term "x-ray image during treatment" as used herein can refer to an image captured at any point in time during the treatment delivery phase of an interventional or therapeutic procedure, which may include the time when the radiation source is turned on or off. Sometimes, for ease of description, CT imaging data (e.g., cone-beam CT imaging data) may be used herein as an exemplary imaging modality. However, it will be appreciated that data from any type of imaging modality can also be used in a variety of implementations, including but not limited to x-ray radiography, MRI (magnetic resonance imaging), PET (positron emission tomography), PET-CT (computed tomography), SPECT (single-photon emission computed tomography), SPECT-CT, MR-PET, 3D ultrasound images, etc.

[0020] Unless otherwise stated, as will be apparent from the following discussion, terms such as “segmentation,” “generation,” “registration,” “determination,” “alignment,” “positioning,” “processing,” “computation,” “selection,” “estimation,” “detection,” and “tracking” can refer to actions and processes of a computer system or similar electronic computing device that are represented as data manipulation and transformation of physical (e.g., electronic) quantities within the registers and memories of the computer system into other data similarly represented as physical quantities within the computer system's memory or registers or other such information storage, transmission, or display devices. Embodiments of the methods described herein can be implemented using computer software. If written in a programming language conforming to recognized standards, sequences of instructions designed to implement these methods can be compiled for execution on various hardware platforms and interface with various operating systems. Furthermore, the implementation of this framework is not described with reference to any particular programming language. It will be understood that a variety of programming languages ​​can be used.

[0021] As used herein, the term "image" refers to multidimensional data composed of discrete image elements, such as pixels for a 2D image, voxels for a 3D image, and doxels for a 4D dataset. The image can be, for example, a medical image of a subject collected by a CT (computed tomography), MRI (magnetic resonance imaging), ultrasound, or any other medical imaging system known to those skilled in the art. The image can also be provided from non-medical contexts, such as, for example, remote sensing systems, electron microscopes, etc. Although an image can be considered as originating from R... 3 Functions in R, or to R 3 This method is not limited to this type of image and can be applied to images of any dimension, such as 2D pictures, 3D volumes, or 4D datasets. For 2D or 3D images, the domain of the image is typically a 2D or 3D rectangular array, where each pixel or voxel can be addressed by referencing a set of 2 or 3 mutually orthogonal axes. The terms "digital" and "digitized" as used herein will refer to images in digital or digitized formats obtained via a digital acquisition system or via conversion from analog images, or, where appropriate, volumes.

[0022] The terms "pixel," conventionally used for image elements in 2D imaging and image display, "voxel," commonly used for volumetric image elements in 3D imaging, and "doxel," used for 4D datasets, are used interchangeably. It should be noted that a 3D volumetric image is itself synthesized from images obtained as pixels on a 2D sensor array and displayed as a 2D image from a certain viewpoint. Therefore, 2D image processing and image analysis techniques can be applied to 3D volumetric images. In the following description, techniques described as operating on doxels can alternatively be described as operating on 3D voxel data, which is stored and represented for display as 2D pixel data. Similarly, techniques operating on voxel data can also be described as operating on pixels. In the following description, the variable x is used to indicate the subject image element at a specific spatial location, or alternatively, is considered as the subject pixel. The terms "subject pixel," "subject voxel," and "subject doxel" are used to indicate a specific image element operated on as described herein.

[0023] One aspect of this framework uses the physical properties of image formation to define one or more neural networks. To understand the difference between this framework and existing methods, consider a simple linear equation:

[0024] y = a · x + b (1)

[0025] in y This represents the predicted output value. x Indicates the input value. a It is a coefficient, and b It's the intercept. The goal is usually to produce the most accurate prediction possible. y In traditional machine learning, {( x i , y i )} i=1,…,N To estimate a and b In algebraic reconstruction, obtain And use physical properties to explain a and b One aspect of this framework uses physical properties to model the neural network, while employing deep learning methods to compute the solution.

[0026] In some implementations, this framework models image formation as one or more neural networks to achieve faster gradient descent reconstruction of image data. By modeling image formation as a cascade of operations rather than linear operators as in traditional iterative reconstruction, the overall scale of the computational task is reduced. Time and storage requirements are significantly reduced because the neural network only needs to be computed once and stored in memory (e.g., 500MB-4GB memory space). The neural network formulation can be directly applied to any image reconstruction and can run natively on, for example, a graphics processing unit (GPU) without special effort.

[0027] Additionally, this framework enables the development of research branches of interest, such as iterative intra-attenuation optimization, generative adversarial network (GAN) research, or optimization of image formation estimation. The derivatives for each image formation step can be automatically computed using neural network formulations within deep learning frameworks (such as GANs). These and other exemplary features and advantages will be described in more detail in this paper. It is to be understood that while specific applications for SPECT image reconstruction may be shown in this paper, the technique is not limited to the specific implementations illustrated. This framework can also be applied to images acquired through other modalities.

[0028] Figure 1 This is a block diagram illustrating an exemplary system 100. System 100 includes a computer system 101 for implementing the framework described herein. Computer system 101 may be a desktop personal computer, a portable laptop computer, another portable device, a minicomputer, a mainframe computer, a server, cloud infrastructure, a storage system, a dedicated digital appliance, a communication device, or another device having a storage subsystem configured to store a collection of digital data items. In some implementations, computer system 101 operates as a standalone device. In other implementations, computer system 101 may be connected (e.g., using a network) to other machines, such as imaging device 102 and workstation 103. In a network deployment, computer system 101 may operate as a server or client user machine in a server-client user network environment, or as a peer machine in a peer-to-peer (or distributed) network environment.

[0029] Computer system 101 may include a processor device or central processing unit (CPU) 104, which is coupled via an input-output interface 121 to one or more non-transitory computer-readable media 105 (e.g., computer storage devices or memories), a display device 108 (e.g., a monitor), and various input devices 110 (e.g., a mouse or keyboard). Computer system 101 may further include supporting circuitry such as caches, power supplies, clock circuitry, and communication buses. Various other peripheral devices, such as additional data storage devices and printing devices, may also be connected to computer system 101.

[0030] This technology can be implemented in various forms of hardware, software, firmware, dedicated processors, or combinations thereof, executed via an operating system, or as part of microinstruction code, or as part of an application or software product, or a combination thereof. In one implementation, the technology described herein is implemented as computer-readable program code tangibly embodied in one or more non-transitory computer-readable media 105. In particular, this technology can be implemented by an image reconstruction module 106 and a neural network module 111. The non-transitory computer-readable medium 105 may include random access memory (RAM), read-only memory (ROM), floppy disk, flash memory, and other types of memory, or combinations thereof. The computer-readable program code is executed by a processor device 104 to process images acquired by, for example, an imaging device 102. Thus, computer system 101 is a general-purpose computer system, which becomes a special-purpose computer system when the computer-readable program code is executed. The computer-readable program code is not intended to be limited to any particular programming language or implementation thereof. It will be appreciated that the teachings of the disclosure contained herein can be implemented using a variety of programming languages ​​and their encodings.

[0031] The same or different computer-readable media 105 can be used to store image sets, patient records, knowledge bases, etc. This data can also be stored in external storage devices or other memories. External storage devices can be implemented using a database management system (DBMS) managed by the processor device 104 and residing in memory such as a hard disk, RAM, or removable media. External storage devices can be implemented on one or more additional computer systems. For example, external storage devices may include data warehouse systems, picture archiving and communication systems (PACS), or any other hospital, medical facility, clinic, testing facility, pharmacy, or other medical patient record storage systems residing on separate computer systems.

[0032] Imaging device 102 may be a nuclear imaging system for acquiring images, such as a single-photon emission computed tomography (SPECT) scanner. Workstation 103 may include a computer and suitable peripherals, such as a keyboard and a display device, and may be operated in conjunction with the entire system 100. For example, workstation 103 may communicate with imaging device 102 such that images collected by imaging device 102 can be presented at workstation 103 and viewed on a display device.

[0033] Workstation 103 can communicate directly with computer system 101 to display processed images and / or output image processing results. Workstation 103 may include a graphical user interface to receive user input via input devices (e.g., keyboard, mouse, touchscreen, voice or video recognition interface, etc.) to manipulate the visualization and / or processing of images. For example, a user can view processed images and specify one or more view adjustments or preferences (e.g., zoom, crop, panning, rotation, changing contrast, changing color, changing view angle, changing view depth, changing rendering or reconstruction techniques, etc.), navigate to a specific area of ​​interest by specifying a "goto" location, navigate the temporal volume of an image set (e.g., stop, play, step through, etc.), and so on.

[0034] It is important to further understand that, because some of the system components and method steps depicted in the accompanying figures can be implemented in software, the actual connections between system components (or process steps) can vary depending on how this framework is programmed. Given the teachings provided herein, those skilled in the art will be able to envision these and similar implementations or configurations of this framework.

[0035] Figure 2 An exemplary image reconstruction method 200 performed by a computer system is illustrated. It should be understood that the steps of method 200 may be performed in the order shown or a different order. Additional, different, or fewer steps may also be provided. Furthermore, method 200 may utilize… Figure 1 The system can be implemented through 100 different systems or combinations thereof.

[0036] At 202, one or more neural networks in neural network module 111 are configured based on the physical properties of image formation. The physical properties of image formation depend on the imaging modality used to acquire images of the subject of interest. In some implementations, one or more neural networks are defined based on SPECT image formation. SPECT is a nuclear medicine tomographic imaging technique that represents the distribution of a radioactive tracer administered within an organ. Other types of imaging modalities, such as computed tomography (CT), SPECT / CT, and positron emission tomography (PET), can also be used.

[0037] Figure 3 The illustration depicts an exemplary SPECT image formation. The SPECT system uses one or more gamma cameras, typically equipped with collimators, mounted on a gantry so that the detector can rotate around the patient. A set 302 of two-dimensional (2D) projection data is acquired at many evenly spaced angles around the patient. The projected images are typically displayed at intervals of 360 degrees or 180 degrees. n Degree (for example, n The values ​​are obtained by rotating the projection dataset 302 by a given angle (e.g., 3-6 degrees). The rotated dataset 304 is then corrected for depth-related blur (e.g., Gaussian blur) by performing a point spread function (PSF) filter to generate a filtered dataset 306. A mathematical algorithm can then be used to estimate the 2D projection image 308 based on the filtered dataset 306.

[0038] Figure 4a An exemplary neural network 410 is shown. The neural network 410 can act as a black box, explained by physical properties. Input data. f i It is applied to neural network 410 to generate output data. g i ,in i This represents the row index. Each neural network can be modeled based on different physical effects, including but not limited to rotation correction, depth-dependent PSF filter error correction, scattering correction, attenuation correction, motion correction, and head misalignment correction. Each effect can be modeled by initializing or modifying the weights of the neural network, or by adding one or more layers to the neural network.

[0039] For example, for attenuation correction, input coefficients for attenuation correction can be derived from CT images. However, these attenuation correction coefficients may need to be changed when, for example, the arm moves or the bladder becomes full during the imaging process. The neural network 410 can update its internal weights to values ​​more suitable for the projected image data.

[0040] Figure 4b An exemplary simple rotation 402 is shown. Rotation 402 can be represented using the following function:

[0041] (2)

[0042] in f This indicates the volume data as 404. This indicates that the obtained data is 406. psf Represents the point spread function. R Represents the rotation function. iIndicates row index, k Indicates the rotation index, and j This is a pixel column index. The retrieved data... This represents the depth integral of the image after rotation correction. The initial weights of the neural network can be determined by calculating this rotation function (2).

[0043] Figure 4c An exemplary convolutional layer 420 for modeling motion correction is shown. Layer 420 can be provided as one of the terminal layers of, for example, a neural network 410. Convolutional layer 420 generates motion-corrected projection data K' based on input projection data K.

[0044] Figure 4d The illustration shows the modeling of motion correction. Convolutional kernels 424a-b are applied to the input image 422 to generate output images 426a-b. Placing the "1" in the middle of convolutional kernel 424a produces no change in the output image 426a. However, if the "1" is placed in another position, as shown in convolutional matrix 424b, patient motion correction can be modeled, as shown in output image 426b.

[0045] Figure 4e An exemplary convolutional layer 430 is shown that models head misalignment correction. Layer 430 can be provided as one of the terminal layers of, for example, a neural network 410. Convolutional layer 430 generates alignment-fixed (or corrected) projection data K' for the head based on the input projection data K.

[0046] Figure 4f The illustration shows an exemplary modeling of point spread function (PSF) filter correction. PSF filter correction can be depth-dependent. PSF describes the imaging system's response to a point source. In depth... d At this point, it can be calculated using the following convolution equation. g i Value:

[0047] g i = f depth d * PSF d (3)

[0048] in g i This represents the estimated projected image. f depth d This represents the estimated volume distribution, and PSF d Indicates depth d The point spread function at that location.

[0049] The PSF (Pressure Sensitivity) can be estimated to provide a measure of the amount of blur added to any given object due to optical defects in the imaging system. Filtering using the PSF can be performed to remove blur from an image. However, the PSF estimate may be incorrect for various reasons. For example, the collimator may be defective, or it may actually be farther away than expected. Errors in PSF estimation and filtering can be corrected by updating the weights of at least one convolutional layer of the neural network to optimize the PSF filter output.

[0050] Figure 5 An exemplary neural network architecture 500 for SPECT image reconstruction is illustrated. Architecture 500 includes a rotation correction neural network 506, a PSF filter error correction neural network 510, a cumulative integration neural network 512, and a product and projection neural network 514. The exemplary neural network architecture 500 accounts for attenuation and scattering correction effects by modeling neural networks 506, 512, and 514 based on attenuation and scattering correction effects. Attenuation of gamma rays within the body can lead to a significant underestimation of activity in deep tissues. Optimal correction is obtained using measured attenuation data. Attenuation data 502 can be measured in the form of an X-ray CT image, which is used as an attenuation map of the tissue. Scattering of gamma rays and the random nature of gamma rays can also degrade the quality of SPECT images, resulting in resolution loss. Scattering data 504 can be measured to perform scattering correction and resolution restoration, thereby improving the resolution of the SPECT image.

[0051] More specifically, the acquired SPECT volume image 302 and the measured attenuation data 502 can be used as inputs to a rotation neural network 506 to generate a rotated image volume 304 and rotated attenuation data 508. The rotated image volume is applied to a PSF neural network 510 to generate a filtered image volume 306, while the rotated attenuation data is applied to a cumulative integral neural network 512 to generate processed attenuation data 513. The cumulative integral neural network 512 projects the 3D input data to generate 2D attenuation data 513 for data comparison. The filtered image volume 306, the processed attenuation data 513, and the measured scattering data 504 are then applied to a product and projection neural network 514 to generate an output estimated projected image 308. The neural network 514 may include, for example, modeling motion correction or head misalignment correction (as previously referenced). Figure 4c One or more layers (as discussed in -e).

[0052] Return to Figure 2At position 204, image reconstruction module 106 uses the acquired test image data to optimize one or more neural networks in neural network module 111. Each layer of the neural network has a simple gradient. Whether a layer is used alone or in combination with another layer, the gradient calculation is independent of the layer's weights. Therefore, by defining the layer as type A, its gradient can be implicitly defined. Thus, the gradient is implicitly defined for each operation, and gradient descent can be used to perform reconstruction without any additional steps. The gradient for each operation can be computed to optimize each step of the image forming model, thereby finding a better estimate with the predicted decay, which helps the iterative reconstruction converge.

[0053] Iterative optimization of the projection operator can also be performed. If convergence has not been achieved after several iterations, it may be due to a mismatch between the modeling of the image formation and the real-world setting. In some implementations, the input to the neural network is modified to maximize the agreement between the output and the acquired data. Convergence can be achieved by providing updates to the attenuation map, head alignment, or motion correction.

[0054] Figure 6 An exemplary method for neural network optimization 204 is shown. It should be understood that the steps of method 204 can be performed in the order shown or a different order. Additional, different, or fewer steps may also be provided. Furthermore, method 204 can utilize... Figure 1 The system can be implemented through 100 different systems or combinations thereof.

[0055] At 601, the image reconstruction module 106 receives one or more neural networks from the neural network module 111 to be optimized.

[0056] At position 602, the image reconstruction module 106 propagates the input data via a neural network. f i To generate output data g i ,in i Indicates the number of iterations.

[0057] At position 604, the image reconstruction module 106 calculates the difference ( ) to output data g i With the acquired test data The comparison is performed. This difference represents the error and is used to determine the weights used for backpropagation through the neural network.

[0058] At position 606, the image reconstruction module 106 backpropagates the difference via a neural network. Backpropagation is performed, making...f i+1 = f i +∆ f about( To minimize ( ).

[0059] At point 608, image reconstruction module 106 determines whether the number of iterations is greater than (or equal to) a predetermined number. m1 If the predetermined number of iterations has been reached... m1 If the condition is met, then method 204 proceeds to step 610. If not, steps 602 through 608 are repeated.

[0060] At position 610, the image reconstruction module 106 uses deep learning techniques to retrain the weights within the neural network. f i and Deep learning techniques estimate changes in the weights of the neural network to achieve the improvements in step 604 (e.g., convergence). For example, in Figure 5 In this configuration, the weights of one or more of the following neural networks can be optimized: 506, 510, 512, or 514. Individual neural networks (506, 510, 512, 514) can be selectively optimized while other neural networks are "frozen." If changes to the weights improve performance by a predetermined threshold... g i and If the difference is between them, then method 600 continues to 611.

[0061] Return to Figure 6 At point 611, image reconstruction module 106 determines whether the number of iterations is greater than (or equal to) a predetermined number. m2 If the predetermined number of iterations has been reached... m2 If the condition is met, then method 204 proceeds to step 612. If not, steps 602 through 610 are repeated.

[0062] At position 612, the image reconstruction module 106 outputs an optimized neural network.

[0063] Return to Figure 2 At 206, current image data of the object of interest is acquired using, for example, imaging device 102. The current image data is acquired using the same modality (e.g., SPECT) upon which the physical properties of image formation are based. The object of interest can be any biological object identified for investigation or examination, such as a part of a patient's organ, brain, heart, leg, arm, etc.

[0064] At 208, the image reconstruction module 106 applies the current image data as input to one or more optimized neural networks in the neural network module 111 to reconstruct the output image.

[0065] At position 210, the output image is presented. The output image can be presented and displayed via, for example, a user interface at workstation 103.

[0066] While this framework has been described in detail with reference to exemplary embodiments, those skilled in the art will appreciate that various modifications and substitutions can be made therein without departing from the spirit and scope of the invention as set forth in the appended claims. For example, within the scope of this disclosure and the appended claims, elements and / or features of different exemplary embodiments may be combined with and / or substituted for each other.

Claims

1. A system for image reconstruction, comprising: Memory, which is used to store computer-readable program code; as well as A processor that communicates with the memory, the processor operating together with the computer-readable program code to perform steps, the steps including: Multiple neural networks are configured based on several different physical effects of image formation, wherein image formation is modeled as a cascade of the multiple neural networks. The acquired test image data is used to optimize multiple configured neural networks, and The output image is reconstructed by applying the current image data as input to multiple optimized neural networks.

2. The system of claim 1, wherein the processor operates together with the computer-readable program code to configure the plurality of neural networks based on the physical properties of single-photon emission computed tomography (SPECT) images formed.

3. The system of claim 1, wherein the processor operates together with the computer-readable program code to configure the plurality of neural networks based on rotation correction, point spread function (PSF) filter error correction, scattering correction, attenuation correction, motion correction, head misalignment correction, or a combination thereof.

4. The system of claim 1, wherein the processor operates together with the computer-readable program code to configure the plurality of neural networks by updating the weights of at least one neural network for attenuation correction.

5. The system of claim 1, wherein the processor operates together with the computer-readable program code to configure the plurality of neural networks by initializing the weights of at least one of the plurality of neural networks based on a rotation function.

6. The system of claim 1, wherein the processor operates together with the computer-readable program code to configure the plurality of neural networks by providing at least one terminal layer to at least one of the plurality of neural networks that models motion correction.

7. The system of claim 1, wherein the processor operates together with the computer-readable program code to configure the plurality of neural networks by providing at least one terminal layer to at least one neural network that models head misalignment correction among the plurality of neural networks.

8. The system of claim 1, wherein the processor operates together with the computer-readable program code to configure the plurality of neural networks by updating the weights of at least one neural network for PSF filter error correction in the plurality of neural networks.

9. The system of claim 1, wherein the processor operates together with the computer-readable program code to optimize the plurality of configured neural networks by performing gradient descent optimization.

10. The system of claim 9, wherein the processor operates together with the computer-readable program code to perform gradient descent optimization by: iteratively propagating input data through the plurality of configured neural networks to generate output data, comparing the output data with acquired test image data, and backpropagating the difference between the output data and the acquired test image data through the plurality of configured neural networks.

11. The system of claim 10, wherein the processor operates together with the computer-readable program code to retrain the weights in the plurality of configured neural networks using deep learning techniques.

12. The system of claim 1, wherein the processor operates together with the computer-readable program code to optimize the plurality of configured neural networks by performing intra-iterative optimization.

13. A computer-based method for image reconstruction, comprising: Multiple neural networks are configured based on multiple different physical effects of image formation, wherein the image formation is modeled as a cascade of the multiple neural networks; The acquired test image data is used to optimize multiple configured neural networks; as well as The output image is reconstructed by applying the current image data as input to multiple optimized neural networks.

14. The method of claim 13, further comprising: The multiple neural networks are configured based on rotation correction, point spread function (PSF) filter error correction, scattering correction, attenuation correction, motion correction, head misalignment correction, or a combination thereof.

15. The method of claim 14, wherein configuring the plurality of neural networks comprises: The weights of at least one of the plurality of neural networks are initialized based on a rotation function.

16. The method of claim 13, wherein using the acquired test image data to optimize the plurality of configured neural networks comprises: Perform gradient descent optimization.

17. A computer program product comprising instructions that, when executed by at least one processor of a device, cause the device to perform the method according to any one of claims 13 to 16.

Images