An image processing method, apparatus, device and storage medium

Abstract

This application provides an image processing method, apparatus, device, and storage medium applicable to various scenarios such as cloud technology, artificial intelligence, smart transportation, and vehicle-mounted systems. The method includes: acquiring corresponding undersampled initial data through multiple receiving coils; then employing cascaded multiple target processing networks to perform information supplementation operations on the acquired initial data to obtain corresponding target restoration images; and determining a target reconstruction image based on the acquired target restoration images. Each target processing network includes an image restoration network, a frequency domain completion network, and a sensitivity estimation network. By employing cascaded multiple target processing networks to perform cross-information supplementation on the acquired initial data in the frequency domain, image domain, and sensitivity information dimensions, more comprehensive image information is obtained. Therefore, when reconstructing the image based on the more comprehensive image information, the quality and efficiency of the target reconstruction image are effectively improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an image processing method, apparatus, device, and storage medium. Background Technology

[0002] With the development of medical technology and the widespread use of medical imaging, medical images are now commonly used to understand a patient's current physical condition. Among these, magnetic resonance imaging (MRI) provides high-quality reference information through high-resolution and high-contrast images. However, compared to other medical imaging techniques, MRI requires a relatively long scan time. This longer scan time can cause patient discomfort and produce motion artifacts, thus affecting the quality of the MRI images.

[0003] In related technologies, during magnetic resonance imaging (MRI), the receiving coil acquires undersampled K-space data based on a sampling mask, and then reconstructs the MRI image based on the undersampled K-space data. However, undersampled K-space data suffers from missing information, resulting in poor quality MRI images reconstructed from it. Summary of the Invention

[0004] This application provides an image processing method, apparatus, device, and storage medium for improving the quality and efficiency of reconstructed MRI images.

[0005] On one hand, embodiments of this application provide an image processing method, the method comprising:

[0006] The corresponding undersampled initial data are acquired through multiple receiving coils respectively;

[0007] Multiple cascaded target processing networks are used to perform information supplementation operations on multiple initial data sets to obtain corresponding target restoration images. Based on the obtained target restoration images, the target reconstruction image is determined. Each target processing network includes an image restoration network, a frequency domain completion network, and a sensitivity estimation network. When obtaining the target restoration image corresponding to each initial data set, the following operations are performed:

[0008] For the first target processing network, the initial data is supplemented with image domain information through the currently cascaded image restoration network, and the obtained restored image of the current cascade is input into the next cascaded frequency domain completion network for frequency domain information supplementation; and the initial data is supplemented with frequency domain information through the currently cascaded frequency domain completion network, and the obtained frequency domain completion data of the current cascade is input into the next cascaded sensitivity estimation network for sensitivity supplementation;

[0009] For each non-first target processing network, perform the following operations: supplement the image domain information of the frequency domain completion data output by the frequency domain completion network of the previous cascade and the coil sensitivity output by the sensitivity estimation network of the previous cascade through the current cascaded image restoration network, and obtain the restored image of the current cascade.

[0010] On one hand, embodiments of this application provide an image processing apparatus, which includes:

[0011] The acquisition module is used to acquire the corresponding undersampled initial data through multiple receiving coils respectively;

[0012] The processing module is used to perform information supplementation operations on multiple initial data using cascaded target processing networks to obtain corresponding target restoration images. Based on the obtained target restoration images, the module determines the target reconstruction image. Each target processing network includes an image restoration network, a frequency domain completion network, and a sensitivity estimation network. When obtaining the target restoration image corresponding to each initial data, the module performs the following operations:

[0013] For the first target processing network, the initial data is supplemented with image domain information through the currently cascaded image restoration network, and the obtained restored image of the current cascade is input into the next cascaded frequency domain completion network for frequency domain information supplementation; and the initial data is supplemented with frequency domain information through the currently cascaded frequency domain completion network, and the obtained frequency domain completion data of the current cascade is input into the next cascaded sensitivity estimation network for sensitivity supplementation;

[0014] For each non-first target processing network, perform the following operations: supplement the image domain information of the frequency domain completion data output by the frequency domain completion network of the previous cascade and the coil sensitivity output by the sensitivity estimation network of the previous cascade through the current cascaded image restoration network, and obtain the restored image of the current cascade.

[0015] Optionally, the processing module is specifically used for:

[0016] Perform an inverse Fourier transform on the initial data to obtain the initial time-domain image;

[0017] The initial time-domain image is supplemented with image domain information by the current cascaded image restoration network, and the obtained current cascaded restored image is input into the next cascaded frequency domain completion network for frequency domain information supplementation.

[0018] Optionally, the processing module is further configured to:

[0019] For the first target processing network, the following operations are also performed:

[0020] Target data within a preset frequency range is selected from the initial data, and the target data is subjected to inverse Fourier transform to obtain the initial coil sensitivity.

[0021] The initial coil sensitivity is supplemented by the current cascaded sensitivity estimation network, and the obtained current cascaded coil sensitivity is input into the next cascaded frequency domain completion network to supplement frequency domain information.

[0022] Optionally, the processing module is specifically used for:

[0023] Perform inverse Fourier transform and shrinkage operations on the frequency domain complete data output by the previous-level frequency domain complete network to obtain the time domain image;

[0024] The image domain information of the time-domain image and the coil sensitivity output by the previous cascaded sensitivity estimation network is supplemented by the current cascaded image restoration network to obtain the current cascaded restored image.

[0025] Optionally, the processing module is further configured to:

[0026] For each non-first target processing network, the following operations are also performed:

[0027] The frequency domain information of the restored image output by the previous cascaded image restoration network is supplemented by the frequency domain completion network of the current cascaded network to obtain the frequency domain completion data of the current cascaded network.

[0028] The sensitivity of the current cascaded coil is obtained by supplementing the frequency domain completion data output by the frequency domain completion network of the previous cascaded coil through the current cascaded sensitivity estimation network.

[0029] Optionally, the processing module is specifically used for:

[0030] Perform Fourier transform and expansion operations on the restored image output by the previous stage image restoration network to obtain the corresponding full-frequency domain data to be supplemented.

[0031] The frequency domain information of the frequency domain data to be completed is supplemented by the frequency domain completion network of the current cascaded network to obtain the frequency domain completion data of the current cascaded network.

[0032] Optionally, the processing module is specifically used for:

[0033] Perform a root sum of squares operation on the obtained multiple target restoration images to obtain the target reconstructed image.

[0034] Optionally, a training module may also be included;

[0035] The training module is specifically used for:

[0036] Based on an undersampled dataset, multiple cascaded processing networks to be trained are jointly and iteratively trained to output the multiple target processing networks. During each iteration of training, the following operations are performed:

[0037] The multiple training networks are used to perform information supplementation operations on multiple sample data selected from the sample dataset to obtain the corresponding predicted restored image and the corresponding predicted frequency domain complete data. The predicted reconstructed image is then determined based on the obtained multiple predicted restored images.

[0038] The target loss function is determined based on the predicted reconstructed image and the obtained multiple predicted frequency domain completion data, and the parameters are adjusted using the target loss function.

[0039] Optionally, the training module is specifically used for:

[0040] Based on the multiple predicted frequency domain completion data and the full sample data corresponding to the multiple sample data, a first loss function is determined;

[0041] Based on the predicted reconstructed image and the corresponding reference reconstructed image, a second loss function is determined, wherein the reference reconstructed image is constructed based on the full sample data;

[0042] The target loss function is determined based on the first loss function and the second loss function.

[0043] On one hand, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described image processing method.

[0044] On one hand, embodiments of this application provide a computer-readable storage medium storing a computer program executable by a computer device, which, when run on the computer device, causes the computer device to perform the steps of the above-described image processing method.

[0045] On one hand, embodiments of this application provide a computer program product, the computer program product including a computer program stored on a computer-readable storage medium, the computer program including program instructions, which, when executed by a computer device, cause the computer device to perform the steps of the above-described image processing method.

[0046] In this embodiment, multiple cascaded target processing networks are used to perform information supplementation operations on multiple undersampled initial data to obtain corresponding target restoration images. Based on the obtained multiple target restoration images, the target reconstruction image is determined. In the process of obtaining the target restoration image, the image restoration network, frequency domain completion network, and sensitivity estimation network in the multiple target processing networks perform cross-information supplementation in the frequency domain, image domain, and sensitive information dimensions to obtain more comprehensive image information. Therefore, when performing image reconstruction based on the more comprehensive image information, the quality of the target reconstruction image is effectively improved, and the efficiency of obtaining the target reconstruction image is also improved. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 A schematic diagram of a system architecture provided in an embodiment of this application;

[0049] Figure 2 A schematic flowchart of an image processing method provided in an embodiment of this application;

[0050] Figure 3 A schematic diagram of a network structure provided in an embodiment of this application;

[0051] Figure 4 A schematic diagram of a network structure provided in an embodiment of this application;

[0052] Figure 5 A schematic diagram of a network structure provided in an embodiment of this application;

[0053] Figure 6 A schematic diagram of a network structure provided in an embodiment of this application;

[0054] Figure 7 A schematic diagram of a network structure provided in an embodiment of this application;

[0055] Figure 8 A schematic diagram of a network structure provided in an embodiment of this application;

[0056] Figure 9 A schematic diagram of a network structure provided in an embodiment of this application;

[0057] Figure 10A A schematic flowchart of an image processing method provided in an embodiment of this application;

[0058] Figure 10B This is a schematic diagram of the structure of an image processing apparatus provided in an embodiment of this application;

[0059] Figure 11 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0060] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0061] For ease of understanding, the terms used in the embodiments of this invention are explained below.

[0062] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0063] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0064] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning. For example, in this embodiment, machine learning techniques are used to reconstruct magnetic resonance imaging (MRI) images.

[0065] Magnetic resonance imaging (MRI) is a medical imaging technique that uses magnetic fields and computer-generated radio waves to create detailed images of organs and tissues inside the body.

[0066] K-space is an extension of the Fourier space concept in MRI imaging. K-space represents the two-dimensional or three-dimensional spatial frequency information of an object, defined by the space covered by phase and frequency encoded data.

[0067] Multiple receiver coils: In multichannel MRI, each receiver coil acquires undersampled k-space data of its corresponding body part under the guidance of a sampling mask. Because the undersampled k-space data is only partially sampled, the scan time is shortened. The MRI image is reconstructed from the undersampled k-space data collected by each receiver coil individually.

[0068] The design concept of the embodiments of this application will be introduced below.

[0069] In magnetic resonance imaging (MRI), to reduce MRI scan time, multiple receiving coils acquire undersampled K-space data in parallel according to a sampling mask, and then reconstruct the MRI image based on the undersampled K-space data. However, undersampled K-space data suffers from missing information, resulting in poor quality MRI images reconstructed from it.

[0070] Analysis revealed that coil sensitivity can serve as supplementary information to undersampled K-space data. Furthermore, deep learning methods have demonstrated superior performance in inverse imaging problems such as denoising, compressed sensing, and super-resolution. Therefore, employing deep learning to perform multi-level information compensation on undersampled K-space data based on coil sensitivity to obtain complete K-space data, and then using this complete K-space data to predict and obtain reconstructed MRI images, will effectively improve the quality of MRI images.

[0071] In view of this, embodiments of this application provide an image processing method, which includes: acquiring corresponding undersampled initial data through multiple receiving coils; then using multiple cascaded target processing networks to perform information supplementation operations on the acquired initial data to obtain corresponding target restoration images; and determining a target reconstructed image based on the acquired target restoration images. Each target processing network includes an image restoration network, a frequency domain completion network, and a sensitivity estimation network. Furthermore, when acquiring the target restoration image corresponding to each initial data, the following operations are performed:

[0072] For the first target processing network, the initial data is supplemented with image domain information through the current cascaded image restoration network, and the obtained restored image of the current cascaded network is input into the next cascaded frequency domain completion network for frequency domain information supplementation; and the initial data is supplemented with frequency domain information through the current cascaded frequency domain completion network, and the obtained frequency domain completion data of the current cascaded network is input into the next cascaded sensitivity estimation network for sensitivity supplementation.

[0073] For each non-first target processing network, perform the following operations: supplement the image domain information of the frequency domain completion data output by the frequency domain completion network of the previous cascade and the coil sensitivity output by the sensitivity estimation network of the previous cascade through the current cascaded image restoration network, and obtain the restored image of the current cascade.

[0074] In this embodiment, multiple cascaded target processing networks are used to perform information supplementation operations on multiple undersampled initial data to obtain corresponding target restoration images. Based on the obtained multiple target restoration images, the target reconstruction image is determined. In the process of obtaining the target restoration image, the image restoration network, frequency domain completion network, and sensitivity estimation network in the multiple target processing networks perform cross-information supplementation in the image domain, frequency domain, and sensitivity information dimensions to obtain more comprehensive image information. Therefore, when performing image reconstruction based on the more comprehensive image information, the quality of the target reconstruction image is effectively improved, and the efficiency of obtaining the target reconstruction image is also improved.

[0075] refer to Figure 1 This is a system architecture diagram of an MRI system applicable to the embodiments of this application. The system architecture includes at least a signal acquisition device 101 and an image reconstruction device 102.

[0076] The signal acquisition device 101 includes multiple radio frequency (RF) coils. These RF coils emit RF signals towards the patient and receive MR signals emitted from the patient. Specifically, to induce atomic nuclei to transition from a low-energy state to a high-energy state, the RF coils generate electromagnetic wave signals and apply these signals to the patient. These electromagnetic wave signals are RF signals corresponding to the type of atomic nucleus. When the electromagnetic wave signals generated by the RF coils are applied to the atomic nuclei, the nuclei can transition from a low-energy state to a high-energy state. Then, when the electromagnetic waves generated by the RF coils disappear, the atomic nuclei to which the electromagnetic wave signals were applied transition from a high-energy state to a low-energy state, thereby emitting electromagnetic waves with a Larmor frequency. The RF coils receive electromagnetic wave signals from the atomic nuclei in the patient's body.

[0077] The RF coil transmits a partially discrete-phase RF signal sequence according to the sampling mask to obtain undersampled K-space data. This undersampled K-space data is then sent to the image reconstruction device 102.

[0078] The image reconstruction device 102 employs multiple cascaded target processing networks to perform information supplementation operations on the acquired multiple undersampled K-space data to obtain corresponding target restoration images. Based on the acquired multiple target restoration images, the target reconstruction image is determined. Each target processing network includes an image restoration network, a frequency domain completion network, and a sensitivity estimation network. When obtaining the target restoration image corresponding to each undersampled K-space data, the following operations are performed:

[0079] For the first target processing network, the undersampled K-space data is supplemented with image domain information through the current cascaded image restoration network, and the obtained restored image of the current cascaded network is input into the next cascaded frequency domain completion network for frequency domain information supplementation; and the undersampled K-space data is supplemented with frequency domain information through the current cascaded frequency domain completion network, and the obtained frequency domain completion data of the current cascaded network is input into the next cascaded sensitivity estimation network for sensitivity supplementation.

[0080] For each non-first target processing network, perform the following operations: supplement the image domain information of the frequency domain completion data output by the frequency domain completion network of the previous cascade and the coil sensitivity output by the sensitivity estimation network of the previous cascade through the current cascaded image restoration network, and obtain the restored image of the current cascade.

[0081] The image reconstruction device 102 described above can be a terminal device or a server. The terminal device can be a smartphone, tablet, laptop, desktop computer, intelligent voice interaction device, intelligent in-vehicle device, etc. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The signal acquisition device 101 and the image reconstruction device 102 can be directly or indirectly connected via wired or wireless communication, which is not limited herein.

[0082] based on Figure 1 The system architecture diagram shown in this application illustrates the flow of an image processing method. Figure 2 As shown, the process of this method is executed by a computer device, which can be... Figure 1 The image reconstruction device 102 shown includes the following steps:

[0083] Step S201: Acquire the corresponding undersampled initial data through multiple receiving coils.

[0084] Specifically, the receiving coil can be an RF coil; for each receiving coil, the receiving coil directly acquires undersampled K-space data as initial data according to the sampling mask; or, the receiving coil acquires fully sampled K-space data, and then performs undersampling processing on the fully sampled K-space data to obtain undersampled K-space data as initial data.

[0085] Step S202: Using multiple cascaded target processing networks, information supplementation operations are performed on the multiple initial data obtained to obtain corresponding target restoration images, and the target reconstruction image is determined based on the multiple target restoration images obtained.

[0086] Specifically, the network structure composed of multiple target processing networks is as follows: Figure 3As shown, each target processing network includes an image restoration network, a frequency domain completion network, and a sensitivity estimation network. Since the undersampled K-space data, the restored image, and the coil sensitivity are all complex numbers, the input layers of the image restoration network, the frequency domain completion network, and the sensitivity estimation network each have two input channels: an imaginary input channel and a real input channel. The image restoration network, the frequency domain completion network, and the sensitivity estimation network can be fully convolutional networks such as U-Net, or other types of networks; this application does not impose specific limitations on these. Furthermore, steps S201 and S202 above are processes for processing the initial data acquired by one set of receiving coils to obtain the target reconstructed image. Repeating steps S201 and S202 to process the initial data acquired by other sets of receiving coils can obtain the target reconstructed images corresponding to those other sets of receiving coils. Then, based on the target reconstructed images corresponding to multiple sets of receiving coils, a three-dimensional MRI image is obtained.

[0087] The following is combined with Figure 3 The network structure shown illustrates the processing of initial data by the first target processing network, such as... Figure 4 As shown:

[0088] For the first target processing network, the initial data is supplemented with image domain information through the currently cascaded image restoration network, and the obtained restored image is input into the next cascaded frequency domain completion network for frequency domain information supplementation. Then, the initial data is supplemented with frequency domain information through the currently cascaded frequency domain completion network, and the obtained frequency domain completion data is input into the next cascaded sensitivity estimation network for sensitivity supplementation.

[0089] Specifically, perform an inverse Fourier transform (F) on the initial data. - The initial time-domain image is obtained. Then, the image restoration network in the first target processing network supplements the initial time-domain image with image domain information to obtain the current cascaded restored image. The current cascaded restored image is then subjected to Fourier transform (F) and input into the next cascaded frequency domain completion network (the frequency domain completion network in the second target processing network) for frequency domain information supplementation.

[0090] Since the initial data is undersampled K-space data, i.e., frequency domain data, it can be directly input into the frequency domain completion network in the first target processing network to supplement frequency domain information, thus obtaining the current cascaded frequency domain completion data. The current cascaded frequency domain completion data is then input into the next cascaded sensitivity estimation network (the sensitivity estimation network in the second target processing network) for sensitivity supplementation, and after performing an inverse Fourier transform on the current cascaded frequency domain completion data, it is input into the next cascaded image restoration network (the sensitivity estimation network in the second target processing network) for image domain information supplementation.

[0091] In some embodiments, such as Figure 5 As shown, for the first target processing network, target data within a preset frequency range is selected from the initial data, and an inverse Fourier transform is performed on the target data to obtain the initial coil sensitivity. The initial coil sensitivity is supplemented by the current cascaded sensitivity estimation network, and the obtained current cascaded coil sensitivity is input into the next cascaded frequency domain completion network to supplement frequency domain information.

[0092] Specifically, target data with frequencies below a preset frequency are selected from the undersampled K-space data. Then, an inverse Fourier transform is performed on the target data to obtain the initial coil sensitivity in image form. The initial coil sensitivity is then supplemented by the sensitivity estimation network in the first target processing network to obtain the current cascaded coil sensitivity in image form. The coil sensitivity is then input into the next cascaded frequency domain completion network (the frequency domain completion network in the second target processing network) to supplement frequency domain information.

[0093] For example, such as Figure 6 As shown, the first target processing network includes an image restoration network. Frequency domain completion network and sensitivity estimation network The second target processing network includes an image restoration network. Frequency domain completion network and sensitivity estimation network

[0094] Perform an inverse Fourier transform (F) on the undersampled K-space data. - This process yields a zero-padding image (the initial temporal image). Then, it is processed through an image restoration network. Image domain information is supplemented into the zero-padding image to obtain the currently cascaded restored image. The currently cascaded restored image is then Fourier transformed and input into the frequency domain completion network in the second target processing network.

[0095] Input the undersampled k-space data into the frequency domain completion network Obtain the current cascaded frequency domain complete data. Input the current cascaded frequency domain complete data into the sensitivity estimation network in the second target processing network. And perform an inverse Fourier transform (F) on the currently cascaded frequency domain complete data. - Then input the image restoration network in the second target processing network.

[0096] Target frequency domain data below a preset frequency is selected from the undersampled K-space data, and then inverse Fourier transform (F) is performed on the target frequency domain data. -The initial coil sensitivity is obtained in image form. Then, it is passed through the sensitivity estimation network in the first target processing network. The initial coil sensitivity is supplemented to obtain the current cascaded coil sensitivity in image form. This current cascaded coil sensitivity is then input into the image restoration network within the second target processing network.

[0097] In this embodiment, a frequency domain completion network is used to supplement the frequency domain information of the undersampled K-space data, an image restoration network is used to supplement the image domain information of the undersampled K-space data, and a sensitivity estimation network is used to supplement the coil sensitivity of the undersampled K-space data. This achieves multi-dimensional information supplementation in the frequency and image domains, resulting in more complete K-space data. Therefore, when reconstructing images based on complete K-space data, the quality of the reconstructed image can be effectively improved.

[0098] The following is combined with Figure 3 The network structure shown illustrates the processing of initial data by non-first target processing networks, such as... Figure 7 As shown, it includes the following steps:

[0099] The current cascaded image restoration network supplements the frequency domain completion data output by the previous cascaded frequency domain completion network and the coil sensitivity output by the previous cascaded sensitivity estimation network with image domain information to obtain the currently cascaded restored image. The current cascaded frequency domain completion network supplements the restored image output by the previous cascaded image restoration network with frequency domain information to obtain the currently cascaded frequency domain completion data. The current cascaded sensitivity estimation network supplements the frequency domain completion data output by the previous cascaded frequency domain completion network with sensitivity to obtain the currently cascaded coil sensitivity.

[0100] Specifically, the frequency-domain complete data output from the previous cascaded frequency-domain completeness network undergoes inverse Fourier transform and contraction to obtain the time-domain image. Then, the current cascaded image restoration network supplements the time-domain image and the coil sensitivity output from the previous cascaded sensitivity estimation network with image-domain information to obtain the current cascaded restored image. If the current cascaded image restoration network is located in the last target processing network, then the current cascaded restored image is used as the target restored image for image reconstruction. If the current cascaded image restoration network is not located in the last target processing network, then the current cascaded restored image undergoes a Fourier transform (F) and is input into the frequency-domain completeness network in the next target processing network for frequency-domain information supplementation.

[0101] The restored image output from the previous-level image restoration network undergoes Fourier transform and expansion operations to obtain the corresponding frequency domain data to be completed. Then, the currently cascaded frequency domain completion network supplements the frequency domain data to be completed, obtaining the currently cascaded frequency domain completed data. If the currently cascaded frequency domain completion network is not the last target processing network, the currently cascaded frequency domain completed data is input into the sensitivity estimation network in the next target processing network for sensitivity supplementation, and the currently cascaded frequency domain completed data undergoes an inverse Fourier transform (F... - Then, it is input into the image restoration network in the next target processing network to supplement the image domain information.

[0102] The sensitivity of the coil in the current cascade is obtained by supplementing the frequency domain completion data output by the frequency domain completion network of the previous cascade through the current sensitivity estimation network. If the current cascade sensitivity estimation network is not the last target processing network, the current cascade coil sensitivity is input to the image restoration network in the next target processing network to supplement the image domain information.

[0103] For example, such as Figure 8 As shown, the first target processing network includes an image restoration network. Frequency domain completion network and sensitivity estimation network The second target processing network includes an image restoration network. Frequency domain completion network and sensitivity estimation network The third target processing network includes an image restoration network. Frequency domain completion network and sensitivity estimation network

[0104] Frequency domain completion network in the first target processing network The output frequency domain padded data is subjected to inverse Fourier transform (F - The image is obtained through a shrinking operation and then input into the image restoration network in the second target processing network. The sensitivity estimation network in the first target processing network Output coil sensitivity input image restoration network Image Restoration Network Based on the input time-domain image and coil sensitivity, the currently cascaded restored image is obtained. After performing Fourier transform (F) and expansion operations on the currently cascaded restored image, it is input into the frequency domain completion network in the third target processing network.

[0105] Image restoration network in the first target processing network The output restored image undergoes Fourier transform (F) and expansion operations to obtain the corresponding full-frequency domain data to be completed. This completed-frequency domain data is then input into the frequency domain completion network within the second target processing network. Frequency domain completion network Frequency domain information is supplemented into the frequency domain data to be completed, resulting in the currently cascaded frequency domain completed data. This completed data is then input into the sensitivity estimation network within the third target processing network. And perform an inverse Fourier transform (F) on the currently cascaded frequency domain complete data. - After the shrinking operation, the image is input into the image restoration network in the third target processing network.

[0106] The sensitivity estimation network in the second objective processing network. Frequency domain completion network in the first target processing network The output frequency domain complete data is used for sensitivity supplementation to obtain the sensitivity of the currently cascaded coil. Then, the sensitivity of the currently cascaded coil is input into the image restoration network in the third target processing network.

[0107] In this embodiment, multiple cascaded target processing networks are used to perform information supplementation operations on multiple undersampled initial data to obtain corresponding target restoration images. Based on the obtained multiple target restoration images, the target reconstruction image is determined. In the process of obtaining the target restoration image, the image restoration network, frequency domain completion network, and sensitivity estimation network in the multiple target processing networks perform cross-information supplementation in the image domain, frequency domain, and sensitivity information dimensions to obtain more comprehensive image information. Therefore, when performing image reconstruction based on the more comprehensive image information, the quality of the target reconstruction image is effectively improved, and the efficiency of obtaining the target reconstruction image is also improved.

[0108] In some embodiments, the restored image output by the image restoration network in the last target processing network is used as the target restored image. Then, the root sum of squares operation is performed on the obtained multiple target restored images to obtain the target reconstructed image, as shown in the following formula (1):

[0109]

[0110] Where N represents the number of receiving coils, x i This represents the target reconstruction image corresponding to the i-th receiving coil.

[0111] For example, such as Figure 9 As shown, the first target processing network includes an image restoration network. Frequency domain completion network and sensitivity estimation network The second target processing network includes an image restoration network. Frequency domain completion network and sensitivity estimation network The third target processing network includes an image restoration network. Frequency domain completion network and sensitivity estimation network The third target processing network is the last target processing network.

[0112] Perform an inverse Fourier transform (F) on the undersampled K-space data. - Then input the image restoration network. Image Restoration Network After image domain information supplementation, the first-concatenated restored image is output. The first-concatenated restored image is then subjected to a Fourier transform (F) and input into the frequency domain completion network.

[0113] Input the undersampled k-space data into the frequency domain completion network Obtain the first-stage frequency domain complete data. Input the first-stage frequency domain complete data into the sensitivity estimation network. And perform an inverse Fourier transform (F) on the frequency domain complete data of the first cascade. - Then input the image restoration network.

[0114] Target frequency domain data below a preset frequency is selected from the undersampled K-space data, and then inverse Fourier transform (F) is performed on the target frequency domain data. - Input sensitivity estimation network Obtain the sensitivity of the first cascaded coil. Input the sensitivity of the first cascaded coil into the image restoration network.

[0115] Image Restoration Network Based on the frequency domain completion data and the sensitivity of the first-cascade coil, the restored image of the second-cascade is obtained. The restored image of the second-cascade is then subjected to a Fourier transform (F) and input into the frequency domain completion network.

[0116] Frequency domain completion network After supplementing the frequency domain information, the second-level frequency domain complete data is output. This second-level frequency domain complete data is then input into the sensitivity estimation network. And perform an inverse Fourier transform (F) on the frequency domain complete data of the second cascade. - Then input the image restoration network.

[0117] Sensitivity Estimation Network After supplementing the coil sensitivity, the output is the sensitivity of the second cascaded coil. This second cascaded coil sensitivity is then input into the image restoration network.

[0118] Image Restoration Network Based on the frequency domain completion data of the second cascade and the coil sensitivity of the second cascade, the restored image of the third cascade is output. Then, the root sum of squares (RSS) operation is performed on multiple restored images of the third cascade to obtain the target reconstructed image.

[0119] In this embodiment of the application, after obtaining the target restoration images corresponding to each of the multiple receiving coils, the root sum of squares operation is performed on the multiple target restoration images to obtain the target reconstruction image, so that the image information in the target reconstruction image is more evenly distributed, thereby improving the quality of the target reconstruction image.

[0120] After introducing how to obtain object reconstruction images through multiple object processing networks, the following describes the joint training process of multiple object processing networks, including the following steps:

[0121] Based on an undersampled dataset, multiple cascaded processing networks are jointly and iteratively trained to output multiple target processing networks. During each iteration, the following operations are performed: Multiple processing networks perform information supplementation operations on multiple samples selected from the dataset to obtain corresponding predicted restored images and corresponding predicted frequency domain complete data. Based on the obtained predicted restored images, a predicted reconstructed image is determined. Then, based on the predicted reconstructed image and the obtained predicted frequency domain complete data, a target loss function is determined, and the parameters are adjusted using the target loss function.

[0122] Specifically, the undersampled data is also undersampled K-space data, with multiple sample data corresponding to different receiving coils. Training is performed end-to-end using the Structural Similarity (SSIM) loss function and the Mean Squared Error (MSE) loss function. During training, the complex data in K-space and the image domain are divided into real and imaginary parts, which are then input into the real and imaginary input channels of the network respectively.

[0123] In some embodiments, a first loss function is determined based on multiple predicted frequency domain completion data and full sample data corresponding to multiple sample data, as shown in the following formula (2):

[0124]

[0125] Where, N k k is the number of samples in the sample dataset. T k represents the predicted frequency domain completion data output by the last cascaded network to be trained.n This represents the fully sampled data (fully sampled K-space data).

[0126] Based on the predicted reconstructed image and the corresponding reference reconstructed image, the second loss function is determined as shown in the following formula (3):

[0127]

[0128] Where, N x x is the number of sample images obtained after performing an inverse Fourier transform on the sample data in the sample dataset. T x represents the predicted reconstructed image. n This represents the reference reconstructed image, which is constructed based on the full sample data.

[0129] Based on the first loss function and the second loss function, the target loss function is determined as shown in the following formula (4):

[0130]

[0131] Where, θ={θ x ,θ k ,θ s}, θ x θ represents the network parameters in the image restoration network. k θ represents the network parameters in the frequency domain completion network. s This represents the network parameters in the sensitivity estimation network. This represents the target loss function.

[0132] For example, we set up 10 cascaded training networks. Each training network includes a U-Net network for image restoration, frequency domain completion, and sensitivity estimation. The U-Net network comprises a compression path and an expansion path. The compression path includes four convolutional blocks, each containing a 2D convolutional layer, an activation layer, a normalization layer, and a max-pooling layer. The activation layer has a negative slope coefficient of 0.2, and the 2D convolutional layer has a 3×3 kernel size. The expansion path also includes four convolutional blocks, each containing a 2D convolutional layer, an activation layer, a normalization layer, and an upsampling layer. The activation layer has a negative slope coefficient of 0.2, and the 2D convolutional layer has a 3×3 kernel size. The number of feature maps starts at 32, 32, and 4, doubling after the max-pooling layer and halving after the upsampling layer.

[0133] Complex sample data in the K-space and image domains are concatenated with their real and imaginary parts, and then input into the image restoration network, frequency domain completion network, and sensitivity estimation network through the real and imaginary input channels, respectively, for computation. The target loss value is then calculated using the formula (4) above, and backpropagated to multiple cascaded training networks. By minimizing the target loss value, the multiple cascaded training networks are updated, making the predicted reconstructed image increasingly similar to the reference reconstructed image, and the data difference between the predicted frequency domain completion data and the fully sampled sample data increasingly smaller. Furthermore, to optimize the network parameter adjustment process, the ADAM algorithm is used to optimize and train the network, with an initial learning rate of 1.0 × 10⁻⁶. -4 And it decreases with Epoch.

[0134] In this embodiment, the frequency domain error is determined based on multiple predicted frequency domain completion data and the full sample data corresponding to multiple sample data, and the time domain error is determined based on the predicted reconstructed image and the corresponding reference reconstructed image. Then, the frequency domain error and the time domain error are combined to obtain the target loss function for adjusting the model parameters. Therefore, during the training process, the processing network gradually takes into account both time domain prediction and frequency domain prediction, thereby improving the performance of the training-obtained processing network.

[0135] In some embodiments, in order to ensure the accuracy of network prediction while taking into account the model's generalization ability to diverse data, the embodiments of this application introduce regularization terms during model training.

[0136] Specifically, the frequency domain completion network is trained using the following formula (5):

[0137]

[0138] Where k represents the full sample data, This represents the predicted frequency domain completion data, where x represents the reference reconstructed image. This indicates a contraction operation. This represents the inverse Fourier transform. This represents the regularization term corresponding to the training frequency domain completion network.

[0139] The image restoration network is trained using the following formula (6):

[0140]

[0141] Where x represents the reference reconstructed image, Let M represent the predicted reconstructed image, M represent the sampling mask, and ε represent the spreading operation. This indicates the predicted frequency domain completion data. Indicates Fourier transform, This represents the regularization term corresponding to the image restoration network.

[0142] The sensitivity estimation network is trained using the following formula (7):

[0143]

[0144] Where s represents the actual coil sensitivity, Indicates the predicted coil sensitivity. This represents the consistency term corresponding to the sensitivity estimation network.

[0145] In this embodiment, a regularization term is introduced during the joint training of the frequency domain completion network, the image restoration network, and the sensitivity estimation network. This improves the model's generalization ability to diverse data while ensuring the accuracy of the network's predictions.

[0146] To better explain the embodiments of this application, the following describes an image processing method provided by the embodiments of this application in conjunction with a specific implementation scenario. The process of this method can be described as follows: Figure 1 The signal acquisition device 101 and image reconstruction device 102 shown are used for execution. The signal acquisition device 101 includes an MRI scanner gantry, an examination table, and multiple RF coils, such as... Figure 10A As shown, it includes the following steps:

[0147] When the patient lies supine on the examination bed, multiple RF coils are attached to the corresponding examination sites. The examination bed is then moved into the receiving aperture of the MRI scanner gantry. Each RF coil then transmits a partially discrete-phase RF signal sequence to the patient according to a sampling mask and receives corresponding undersampled K-space data. These multiple undersampled K-space data are then sent to the image reconstruction device 102. The image reconstruction device 102 includes T+1 target processing networks, where the first target processing network includes an image restoration network. Frequency domain completion network and sensitivity estimation network The second target processing network includes an image restoration network. Frequency domain completion network and sensitivity estimation network …; The (T+1)th target processing network includes an image restoration network. Frequency domain completion network and sensitivity estimation network The Tth target processing network is the last target processing network. The image reconstruction device 102 divides the received multiple undersampled K-space data into multiple slices, and performs the following steps for the undersampled K-space data in each slice:

[0148] Perform an inverse Fourier transform (F) on the undersampled K-space data. - Then input the image restoration network. Image Restoration Network After image domain information supplementation, the first-concatenated restored image is output. The first-concatenated restored image is then subjected to a Fourier transform (F) and input into the frequency domain completion network.

[0149] Input the undersampled k-space data into the frequency domain completion network Obtain the first-stage frequency domain complete data. Input the first-stage frequency domain complete data into the sensitivity estimation network. And perform an inverse Fourier transform (F) on the frequency domain complete data of the first cascade. - Then input the image restoration network.

[0150] Target frequency domain data below a preset frequency is selected from the undersampled K-space data, and then inverse Fourier transform (F) is performed on the target frequency domain data. - Input sensitivity estimation network Obtain the sensitivity of the first cascaded coil. Input the sensitivity of the first cascaded coil into the image restoration network.

[0151] Image Restoration Network Based on the frequency domain completion data and the sensitivity of the first-cascade coil, the restored image of the second-cascade is obtained. The restored image of the second-cascade is then subjected to a Fourier transform (F) and input into the frequency domain completion network.

[0152] Frequency domain completion network After supplementing the frequency domain information, the second-level frequency domain complete data is output. This second-level frequency domain complete data is then input into the sensitivity estimation network. And perform an inverse Fourier transform (F) on the frequency domain complete data of the second cascade. - Then input the image restoration network.

[0153] Sensitivity Estimation Network After supplementing the coil sensitivity, the output is the sensitivity of the second cascaded coil. This second cascaded coil sensitivity is then input into the image restoration network.

[0154] And so on, image restoration networks Based on the frequency domain completion data of the T-th cascade and the coil sensitivity of the T-th cascade, the restored image of the T+1-th cascade is output. Then, the root sum of squares (RSS) operation is performed on multiple restored images of the T+1-th cascade to obtain a two-dimensional anatomical image corresponding to a slice.

[0155] By merging the two-dimensional anatomical images corresponding to multiple slices, a three-dimensional MRI image corresponding to the examination site is obtained. This three-dimensional MRI image can be used for the diagnosis of the examination site in the patient.

[0156] In this embodiment, multiple cascaded target processing networks are used to perform information supplementation operations on multiple undersampled initial data to obtain corresponding target restoration images. Based on the obtained multiple target restoration images, the target reconstruction image is determined. In the process of obtaining the target restoration image, the image restoration network, frequency domain completion network, and sensitivity estimation network in the multiple target processing networks perform cross-information supplementation in the image domain, frequency domain, and sensitivity information dimensions to obtain more comprehensive image information. Therefore, when performing image reconstruction based on the more comprehensive image information, the quality of the target reconstruction image is effectively improved, and the efficiency of obtaining the target reconstruction image is also improved.

[0157] Based on the same technical concept, embodiments of this application provide an image processing apparatus, such as... Figure 10B As shown, the device 1000 includes:

[0158] The acquisition module 1001 is used to acquire the corresponding undersampled initial data through multiple receiving coils respectively;

[0159] Processing module 1002 is used to employ multiple cascaded target processing networks to perform information supplementation operations on multiple initial data to obtain corresponding target restoration images, and to determine the target reconstruction image based on the multiple target restoration images obtained. Each target processing network includes an image restoration network, a frequency domain completion network, and a sensitivity estimation network. When obtaining the target restoration image corresponding to each initial data, the following operations are performed:

[0160] For the first target processing network, the initial data is supplemented with image domain information through the currently cascaded image restoration network, and the obtained restored image of the current cascade is input into the next cascaded frequency domain completion network for frequency domain information supplementation; and the initial data is supplemented with frequency domain information through the currently cascaded frequency domain completion network, and the obtained frequency domain completion data of the current cascade is input into the next cascaded sensitivity estimation network for sensitivity supplementation;

[0161] For each non-first target processing network, perform the following operations: supplement the image domain information of the frequency domain completion data output by the frequency domain completion network of the previous cascade and the coil sensitivity output by the sensitivity estimation network of the previous cascade through the current cascaded image restoration network, and obtain the restored image of the current cascade.

[0162] Optionally, the processing module 1002 is specifically used for:

[0163] Perform an inverse Fourier transform on the initial data to obtain the initial time-domain image;

[0164] The initial time-domain image is supplemented with image domain information by the current cascaded image restoration network, and the obtained current cascaded restored image is input into the next cascaded frequency domain completion network for frequency domain information supplementation.

[0165] Optionally, the processing module 1002 is further configured to:

[0166] For the first target processing network, the following operations are also performed:

[0167] Target data within a preset frequency range is selected from the initial data, and the target data is subjected to inverse Fourier transform to obtain the initial coil sensitivity.

[0168] The initial coil sensitivity is supplemented by the current cascaded sensitivity estimation network, and the obtained current cascaded coil sensitivity is input into the next cascaded frequency domain completion network to supplement frequency domain information.

[0169] Optionally, the processing module 1002 is specifically used for:

[0170] Perform inverse Fourier transform and shrinkage operations on the frequency domain complete data output by the previous-level frequency domain complete network to obtain the time domain image;

[0171] The image domain information of the time-domain image and the coil sensitivity output by the previous cascaded sensitivity estimation network is supplemented by the current cascaded image restoration network to obtain the current cascaded restored image.

[0172] Optionally, the processing module 1002 is further configured to:

[0173] For each non-first target processing network, the following operations are also performed:

[0174] The frequency domain information of the restored image output by the previous cascaded image restoration network is supplemented by the frequency domain completion network of the current cascaded network to obtain the frequency domain completion data of the current cascaded network.

[0175] The sensitivity of the current cascaded coil is obtained by supplementing the frequency domain completion data output by the frequency domain completion network of the previous cascaded coil through the current cascaded sensitivity estimation network.

[0176] Optionally, the processing module 1002 is specifically used for:

[0177] Perform Fourier transform and expansion operations on the restored image output by the previous stage image restoration network to obtain the corresponding full-frequency domain data to be supplemented.

[0178] The frequency domain information of the frequency domain data to be completed is supplemented by the frequency domain completion network of the current cascaded network to obtain the frequency domain completion data of the current cascaded network.

[0179] Optionally, the processing module 1002 is specifically used for:

[0180] Perform a root sum of squares operation on the obtained multiple target restoration images to obtain the target reconstructed image.

[0181] Optionally, it also includes a training module 1003;

[0182] The training module 1003 is specifically used for:

[0183] Based on an undersampled dataset, multiple cascaded processing networks to be trained are jointly and iteratively trained to output the multiple target processing networks. During each iteration of training, the following operations are performed:

[0184] The multiple training networks are used to perform information supplementation operations on multiple sample data selected from the sample dataset to obtain the corresponding predicted restored image and the corresponding predicted frequency domain complete data. The predicted reconstructed image is then determined based on the obtained multiple predicted restored images.

[0185] The target loss function is determined based on the predicted reconstructed image and the obtained multiple predicted frequency domain completion data, and the parameters are adjusted using the target loss function.

[0186] Optionally, the training module 1003 is specifically used for:

[0187] Based on the multiple predicted frequency domain completion data and the full sample data corresponding to the multiple sample data, a first loss function is determined;

[0188] Based on the predicted reconstructed image and the corresponding reference reconstructed image, a second loss function is determined, wherein the reference reconstructed image is constructed based on the full sample data;

[0189] The target loss function is determined based on the first loss function and the second loss function.

[0190] In this embodiment, multiple cascaded target processing networks are used to perform information supplementation operations on multiple undersampled initial data to obtain corresponding target restoration images. Based on the obtained multiple target restoration images, the target reconstruction image is determined. In the process of obtaining the target restoration image, the image restoration network, frequency domain completion network, and sensitivity estimation network in the multiple target processing networks perform cross-information supplementation in the image domain, frequency domain, and sensitivity information dimensions to obtain more comprehensive image information. Therefore, when performing image reconstruction based on the more comprehensive image information, the quality of the target reconstruction image is effectively improved, and the efficiency of obtaining the target reconstruction image is also improved.

[0191] Based on the same technical concept, embodiments of this application provide a computer device, which can be... Figure 1 The image reconstruction device 102 shown is, for example Figure 11 As shown, it includes at least one processor 1101 and a memory 1102 connected to at least one processor. In this embodiment, the specific connection medium between the processor 1101 and the memory 1102 is not limited. Figure 11 Taking the connection between processor 1101 and memory 1102 via a bus as an example, the bus can be divided into address bus, data bus, control bus, etc.

[0192] In this embodiment of the application, the memory 1102 stores instructions that can be executed by at least one processor 1101. By executing the instructions stored in the memory 1102, at least one processor 1101 can perform the steps of the above-described image processing method.

[0193] The processor 1101 is the control center of the computer device, capable of connecting to various parts of the computer device via various interfaces and lines. It performs MRI image reconstruction by running or executing instructions stored in the memory 1102 and accessing data stored in the memory 1102. Optionally, the processor 1101 may include one or more processing units. The processor 1101 may integrate an application processor and a modem processor. The application processor primarily handles the operating system, user interface, and applications, while the modem processor primarily handles wireless communication. It is understood that the modem processor may not be integrated into the processor 1101. In some embodiments, the processor 1101 and the memory 1102 may be implemented on the same chip; in other embodiments, they may be implemented on separate chips.

[0194] Processor 1101 can be a general-purpose processor, such as a central processing unit (CPU), digital signal processor, application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0195] Memory 1102, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. Memory 1102 may include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic memory, magnetic disk, optical disk, etc. Memory 1102 can be any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer device, but is not limited thereto. In the embodiments of this application, memory 1102 can also be a circuit or any other device capable of implementing storage functions for storing program instructions and / or data.

[0196] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing a computer program executable by a computer device, which, when run on the computer device, causes the computer device to perform the steps of the above-described image processing method.

[0197] Based on the same inventive concept, this application provides a computer program product, which includes a computer program stored on a computer-readable storage medium. The computer program includes program instructions, which, when executed by a computer device, cause the computer device to perform the steps of the above-described image processing method.

[0198] Those skilled in the art will understand that embodiments of the present invention can be provided as methods or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0199] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer apparatus or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0200] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer device or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0201] These computer program instructions may also be loaded onto a computer device or other programmable data processing equipment to cause a series of operational steps to be performed on the computer device or other programmable equipment to produce a process implemented by the computer device, thereby providing instructions that execute on the computer device or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0202] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0203] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An image processing method, characterized in that, include: The corresponding undersampled initial data are acquired through multiple receiving coils respectively; Multiple cascaded target processing networks are used to perform information supplementation operations on multiple initial data sets to obtain corresponding target restoration images. Based on the obtained target restoration images, a target reconstruction image is determined. Each target processing network includes an image restoration network, a frequency domain completion network, and a sensitivity estimation network. The input layers of the image restoration network, frequency domain completion network, and sensitivity estimation network are all provided with imaginary part input channels and real part input channels. When obtaining the target restoration image corresponding to each initial data set, the following operations are performed: For the first target processing network, the initial data is supplemented with image domain information through the currently cascaded image restoration network, and the obtained restored image of the current cascade is input into the next cascaded frequency domain completion network for frequency domain information supplementation; the initial data is also supplemented with frequency domain information through the currently cascaded frequency domain completion network, and the obtained frequency domain completion data of the current cascade is input into the next cascaded sensitivity estimation network for sensitivity supplementation; target data with frequencies lower than a preset frequency are selected from the initial data, and the target data is subjected to inverse Fourier transform to obtain the initial coil sensitivity in image form; the initial coil sensitivity is supplemented with sensitivity through the sensitivity estimation network of the first target processing network to obtain the current cascaded coil sensitivity in image form; the current cascaded coil sensitivity is input into the next cascaded frequency domain completion network for frequency domain information supplementation; For each non-first target processing network, the following operations are performed: Inverse Fourier transform and contraction are applied to the frequency domain complete data output by the previous cascaded frequency domain complete network to obtain a time-domain image; image domain information is supplemented to the time-domain image and the coil sensitivity output by the previous cascaded sensitivity estimation network using the currently cascaded image restoration network to obtain the currently cascaded restored image; Fourier transform and expansion are applied to the restored image output by the previous cascaded image restoration network to obtain the corresponding frequency domain data to be completed; frequency domain information is supplemented to the frequency domain data to be completed using the currently cascaded frequency domain complete network to obtain the currently cascaded frequency domain complete data; sensitivity supplementation is applied to the frequency domain complete data output by the previous cascaded frequency domain complete network using the currently cascaded sensitivity estimation network to obtain the currently cascaded coil sensitivity.

2. The method as described in claim 1, characterized in that, The step of supplementing the initial data with image domain information through the currently cascaded image restoration network, and inputting the obtained currently cascaded restored image into the next cascaded frequency domain completion network for frequency domain information supplementation, includes: Perform an inverse Fourier transform on the initial data to obtain the initial time-domain image; The initial time-domain image is supplemented with image domain information by the current cascaded image restoration network, and the obtained current cascaded restored image is input into the next cascaded frequency domain completion network for frequency domain information supplementation.

3. The method as described in claim 1, characterized in that, The process of determining the target reconstructed image based on the obtained multiple target restored images includes: Perform a root sum of squares operation on the obtained multiple target restoration images to obtain the target reconstructed image.

4. The method according to any one of claims 1 to 3, characterized in that, The multiple target processing networks are obtained through joint training in the following manner: Based on an undersampled dataset, multiple cascaded processing networks to be trained are jointly and iteratively trained to output the multiple target processing networks. During each iteration of training, the following operations are performed: The multiple training networks are used to perform information supplementation operations on multiple sample data selected from the sample dataset to obtain the corresponding predicted restored image and the corresponding predicted frequency domain complete data. The predicted reconstructed image is then determined based on the obtained multiple predicted restored images. The target loss function is determined based on the predicted reconstructed image and the obtained multiple predicted frequency domain completion data, and the parameters are adjusted using the target loss function.

5. The method as described in claim 4, characterized in that, The step of determining the target loss function based on the predicted reconstructed image and the obtained multiple predicted frequency domain completion data includes: Based on the multiple predicted frequency domain completion data and the full sample data corresponding to the multiple sample data, a first loss function is determined; Based on the predicted reconstructed image and the corresponding reference reconstructed image, a second loss function is determined, wherein the reference reconstructed image is constructed based on the full sample data; The target loss function is determined based on the first loss function and the second loss function.

6. An image processing apparatus, characterized in that, include: The acquisition module is used to acquire the corresponding undersampled initial data through multiple receiving coils respectively; The processing module is used to employ multiple cascaded target processing networks to perform information supplementation operations on multiple initial data sets to obtain corresponding target restoration images. Based on the obtained target restoration images, a target reconstruction image is determined. Each target processing network includes an image restoration network, a frequency domain completion network, and a sensitivity estimation network. The input layers of the image restoration network, frequency domain completion network, and sensitivity estimation network are all provided with imaginary part input channels and real part input channels. When obtaining the target restoration image corresponding to each initial data set, the following operations are performed: For the first target processing network, the initial data is supplemented with image domain information through the currently cascaded image restoration network, and the obtained restored image of the current cascade is input into the next cascaded frequency domain completion network for frequency domain information supplementation; the initial data is also supplemented with frequency domain information through the currently cascaded frequency domain completion network, and the obtained frequency domain completion data of the current cascade is input into the next cascaded sensitivity estimation network for sensitivity supplementation; target data with frequencies lower than a preset frequency are selected from the initial data, and the target data is subjected to inverse Fourier transform to obtain the initial coil sensitivity in image form; the initial coil sensitivity is supplemented with sensitivity through the sensitivity estimation network of the first target processing network to obtain the current cascaded coil sensitivity in image form; the current cascaded coil sensitivity is input into the next cascaded frequency domain completion network for frequency domain information supplementation; For each non-first target processing network, the following operations are performed: Inverse Fourier transform and contraction are applied to the frequency domain complete data output by the previous cascaded frequency domain complete network to obtain a time-domain image; image domain information is supplemented to the time-domain image and the coil sensitivity output by the previous cascaded sensitivity estimation network using the currently cascaded image restoration network to obtain the currently cascaded restored image; Fourier transform and expansion are applied to the restored image output by the previous cascaded image restoration network to obtain the corresponding frequency domain data to be completed; frequency domain information is supplemented to the frequency domain data to be completed using the currently cascaded frequency domain complete network to obtain the currently cascaded frequency domain complete data; sensitivity supplementation is applied to the frequency domain complete data output by the previous cascaded frequency domain complete network using the currently cascaded sensitivity estimation network to obtain the currently cascaded coil sensitivity.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, It stores a computer program executable by a computer device, which, when run on the computer device, causes the computer device to perform the steps of the method according to any one of claims 1 to 5.

9. A computer program product, characterized in that, The computer program product includes a computer program stored on a computer-readable storage medium, the computer program including program instructions that, when executed by a computer device, cause the computer device to perform the steps of the method according to any one of claims 1-5.

Images