Medical imaging system and method based on deep learning

By using perturbation signal and loss function training in deep learning networks, the problems of noise changes and non-uniform point expansion functions in medical imaging are solved, and high-fidelity images are generated and faster scanning speeds are achieved.

CN118941654APending Publication Date: 2024-11-12GE PRECISION HEALTHCARE LLC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410505882.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2023-05-12
Filing Date
2024-04-25
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Existing deep learning networks have problems with noise changes and non-uniform point expansion functions in medical imaging, resulting in poor image quality and limited acceleration rates.

Method used

By using input image training data, including raw image data and perturbation signals, train deep learning networks, use perturbation signals to improve image reconstruction effects, and adjust network parameters through loss functions to minimize differences between reconstructed images, achieving high-fidelity images generate.

Benefits of technology

Improve image quality in medical imaging, reduce artifacts, improve signal-to-noise ratio and clarity, achieve higher acceleration rates and faster scanning speeds.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941654A_ABST
    Figure CN118941654A_ABST
Patent Text Reader

Abstract

A medical imaging system, the medical imaging system comprising at least one medical imaging device for providing image data of a subject. The processing system is programmed to train a deep learning (DL) network using input image training data. The input image training data includes original image data and at least one perturbation signal. The trained DL network is used to determine reconstructed image data from the image data of the subject and generate a medical image of the subject based on the reconstructed image data.
Need to check novelty before this filing date? Find Prior Art

Description

Background Art

[0001] Embodiments disclosed herein relate to medical devices, and more particularly to deep learning (DL) based medical imaging systems and methods.

[0002] Currently, medical imaging equipment is increasingly used to scan a subject (such as a human body) to obtain medical images of a specific part (such as all or part of each organ in the body, or a specific region of interest) to provide useful information for medical diagnosis. Medical image scanning includes magnetic resonance imaging (MRI), X-ray imaging, computed tomography (CT) imaging, ultrasound imaging, etc.

[0003] As a medical imaging mode, magnetic resonance imaging (MRI) can obtain images of the human body without using X-rays or other ionizing radiation. MRI uses a magnet with a strong magnetic field to generate a static magnetic field B0. When the part of the human body to be imaged is located in the static magnetic field B0, the nuclear spins associated with the hydrogen nuclei in the human tissue are polarized, so that the tissue of the part to be imaged generates a longitudinal magnetization vector at the macroscopic level. After applying a radio frequency field B1 that intersects the direction of the static magnetic field B0, the rotation direction of the protons changes, so that the tissue of the part to be imaged generates a transverse magnetization vector at the macroscopic level. After removing the radio frequency field B1, the transverse magnetization vector decays in a spiral manner until it returns to zero. A free induction decay signal is generated during the decay. The free induction decay signal can be collected as a magnetic resonance signal, and the tissue image of the part to be imaged can be reconstructed based on the collected signal.

[0004] Similar to other technical fields, deep learning (DL) has also made significant progress in the field of MRI. In particular, DL-assisted image reconstruction is becoming state of the art, producing better image quality and / or enabling higher acceleration rates than can be achieved with conventional methods. DL networks are used to mitigate noise amplification while maintaining important signal characteristics. However, conventional loss functions used in DL networks produce object-dependent noise variations and non-uniform point spread functions.

[0005] Therefore, there is a need for an improved method to train DL networks for medical imaging systems and methods. Summary of the invention

[0006] According to an embodiment of the present technology, a medical imaging system is provided. The medical imaging system includes at least one medical imaging device and a processing system, the at least one medical imaging device providing image data of a subject. The processing system is programmed to train a deep learning (DL) network using input image training data, wherein the input image training data includes original image data and at least one perturbation signal. The processing system is further programmed to use the trained DL network to determine reconstructed image data from the image data of the subject and generate a medical image of the subject based on the reconstructed image data.

[0007] According to another embodiment of the present technology, a method for imaging a subject is provided. The method includes training a deep learning (DL) network using input image training data, wherein the input training data includes original image data and at least one perturbation signal. The method also includes acquiring image data of the subject using a medical imaging device and providing the image data of the subject as input to the trained DL network. The method also includes generating a medical image of the subject based on the acquired image data using the DL network.

[0008] According to another embodiment of the present technology, a method for generating an image of an object using a magnetic resonance imaging system is provided. The method includes training a deep learning (DL) network using magnetic resonance (MR) image training data, wherein the MR image training data includes original image data and at least one perturbation signal. The method also includes acquiring MR image data of an object using an MRI system and providing the MR image data of the subject as input to the trained DL network. In addition, the method includes using the DL network to generate an image of the object based on the acquired MR image data. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] These and other features, aspects and advantages of the present invention will be better understood when the following detailed description is read with reference to the accompanying drawings, in which like characters refer to like parts throughout the several views, and in which:

[0010] Figure 1 is a schematic diagram of an exemplary magnetic resonance imaging (MRI) system according to an embodiment of the present technology;

[0011] Figure 2A is an exemplary medical image processing system according to an embodiment of the present technology;

[0012] Figure 2B According to the implementation scheme of the present technology, Figure 2A A flowchart of an exemplary method implemented in a system of;

[0013] Figure 3is a schematic diagram depicting the training of a deep learning (DL) network according to an embodiment of the present technology;

[0014] Figure 4 According to the implementation scheme of the present technology, Figure 2A Schematic diagram of an exemplary DL network used in a medical image processing system;

[0015] Figure 5 is a schematic diagram depicting comparative experimental results of various reconstruction techniques according to an embodiment of the present technology; and

[0016] Figure 6 is a schematic diagram depicting another set of comparative experimental results of various reconstruction techniques according to an embodiment of the present technology. DETAILED DESCRIPTION

[0017] One or more specific embodiments will be described below. In order to provide a concise description of these embodiments, all features of an actual implementation may not be described in the specification. It should be understood that in the development of any such actual implementation, as in any engineering or design project, many implementation-specific decisions must be made to achieve the developer's specific goals, such as complying with system-related and business-related constraints that may vary from implementation to implementation. In addition, it should be understood that such development efforts may be complex and time-consuming, but are still routine tasks for design, fabrication, and manufacturing for ordinary technicians who benefit from this disclosure.

[0018] When introducing the elements of the various embodiments of the present embodiment, the articles "a", "an", "the" and "said" are intended to mean that there are one or more of these elements. The terms "comprise", "comprising" and "having" are intended to be inclusive, and mean that there may be additional elements in addition to the listed elements. In addition, any numerical examples in the following discussion are intended to be non-limiting, and therefore additional numerical values, ranges and percentages are within the scope of the disclosed embodiments. In addition, the terms "circuit" and "circuitry" and "controller" may include a single component or multiple components, which are active and / or passive and are connected or otherwise coupled together to provide the functions.

[0019] Although the systems and methods presented herein are explained with respect to magnetic resonance imaging (MRI), the technology may be equally applicable to other medical imaging modalities, including but not limited to computed tomography (CT) equipment, x-ray imaging equipment, positron emission tomography (PET) equipment, single photon emission computed tomography (SPECT) equipment, ultrasound equipment, or any other suitable medical imaging equipment.

[0020] Embodiments of the present technology will now be described by way of example with reference to the accompanying drawings, in which Figure 1 1 is a schematic diagram of a magnetic resonance imaging (MRI) system 10. The operation of the system 10 can be controlled from an operator console 12, which includes an input device 13, a control panel 14, and a display screen 16. The input device 13 can be a mouse, a joystick, a keyboard, a trackball, a touch-activated screen, a light stick, a voice controller, and / or other input devices. The input device 13 can be used for interactive geometric shape designation. The console 12 communicates with a computer system 20 via a link 18, which enables an operator to control the generation and display of images on the display screen 16. The link 18 can be a wireless or wired connection. The computer system 20 may include modules that communicate with each other via a backplane 20a. The modules of the computer system 20 may include, for example, an image processor module 22, a central processing unit (CPU) module 24, and a memory module 26, which may include, for example, a frame buffer for storing an array of image data. The computer system 20 may be linked to an archival media device, a permanent or backup memory, or a network for storing image data and programs, and communicates with an MRI system control 32 via a high-speed signal link 34. The MRI system control 32 may be separate from or integrated with the computer system 20. The computer system 20 and the MRI system control 32 together form an "MRI controller" 33 or "controller."

[0021] In an exemplary embodiment, the MRI system control 32 includes modules connected by a backplane 32a. These modules include a CPU module 36, a deep learning (DL) network or module 75, and a pulse generator module 38. The CPU module 36 is connected to the operator console 12 via a data link 40. The MRI system control 32 receives commands from the operator via the data link 40 to indicate the scan sequence to be performed. The CPU module 36 operates the system components to implement the desired scan sequence and generates data indicating the timing, intensity and shape of the generated RF pulses and the timing and length of the data acquisition window. The CPU module 36 is connected to the components operated by the MRI controller 32, including the pulse generator module 38 that controls the gradient amplifier 42, the physiological acquisition controller (PAC) 44, and the scan room interface circuit 46.

[0022] In one example, the CPU module 36 receives patient data from a physiological acquisition controller 44, which receives signals from sensors connected to the subject, such as ECG signals received from electrodes attached to the patient. As used herein, a subject is a person (or patient), an animal, or a phantom. The CPU module 36 receives signals from sensors associated with the condition of the patient and the magnet system via a scan room interface circuit 46. The scan room interface circuit 46 also enables the MRI controller 33 to command a patient positioning system 48 to move the patient to a desired position for scanning.

[0023] The whole body RF coil 56 is used to transmit the waveform toward the subject's anatomical structure. The whole body RF coil 56 can be a body coil. The RF coil can also be a local coil, which can be placed closer to the subject's anatomical structure than the body coil. The RF coil 56 can also be a surface coil. An RF coil containing RF receiver channels can be used to receive signals from the subject's anatomical structure. A typical surface coil will have eight receiving channels; however, a different number of channels is possible. It is known to use a combination of both the body coil 56 and the surface coil to provide better image quality.

[0024] The MR signals generated by the excitation of the target are digitized by the transceiver module 58. The digitized signals are then processed by the MR system control 32 by Fourier transform to generate k-space data, which is transmitted to the memory module 66 or other computer-readable medium via the MRI system control 32. "Computer-readable medium" may include, for example, a structure configured so that an electrical, optical, or magnetic state can be fixed in a manner that is perceptible and reproducible by a conventional computer (e.g., text or images printed on paper or displayed on a screen, a compact disk or other optical storage medium, "flash" memory, EEPROM, SDRAM or other electrical storage medium; a floppy disk or other magnetic disk, a tape or other magnetic storage medium).

[0025] The scan is complete when an array of raw k-space data is acquired in the computer readable medium 66. For each image to be reconstructed, the raw k-space data is rearranged into a separate k-space data array, and each of these k-space data arrays is input to the array processor 68, which operates to reconstruct the data into an array of image data using a reconstruction algorithm (such as a Fourier transform or a deep learning algorithm). When complete k-space data is obtained, it represents the entire volume of the subject's body, and the k-space so obtained may be referred to as reference k-space. Similarly, when only partial k-space data is obtained, the image may be referred to as partial k-space. The image data is transmitted to the computer system 20 via the data link 34 and stored in the memory. In response to commands received from the operator console 12, the image data may be archived in a long-term storage device, or may be further processed by the image processor 22 and transmitted to the operator console 12, and presented on the display 16.

[0026] The MR signal is represented by a complex number, where each position at k-space is represented by a complex number, where the I and Q orthogonal MR signals are real and imaginary components. A composite MR image can be reconstructed based on the I orthogonal MR signal and the Q orthogonal MR signal using a process such as Fourier transform of the k-space MR data. In one embodiment, the DL module 75 can be used to reconstruct the MR image. Typically, a composite MR image is an MR image in which each pixel is represented by a complex number, which also has a real component and an imaginary component. The magnitude M of the received MR signal can be determined as the square root of the sum of the squares of the I orthogonal component and the Q orthogonal component of the received MR signal, as shown in Equation 3 below:

[0027]

[0028] And the phase φ of the received MR signal can also be determined as shown in the following equation 2:

[0029]

[0030] The MR data collected from the image processor 22 may include artifacts due to subject motion or due to a pathological condition of the subject (such as a metal implant or a fracture in the human body). In one embodiment of the present technology, the DL network module 75 removes these artifacts without affecting the MR image features. The DL network module 75 can also be used to improve the signal-to-noise ratio (SNR) and clarity in the image data. Typically, the DL module 75 provides a high-fidelity image that maximizes feature retention. In one embodiment, the DL module 75 is trained using perturbed training input data, wherein the output of the DL module is reconstructed data of the perturbed training input data. The details will be explained in subsequent paragraphs.

[0031] Figure 2A is Figure 1 Schematic diagram of an exemplary medical image processing system 100 used in an MRI system 10 of FIG. In an exemplary embodiment, the system 100 includes a processing device 102 configured to provide high-fidelity images, reduce artifacts in medical images, and / or improve SNR and clarity. The processing device 102 includes a DL network model 104. In one embodiment, the processing device 102 may include an image reconstruction module (not shown) that may work in conjunction with the DL network model 104 to reconstruct a medical image with better quality than the original image data of the subject.

[0032] The system 100 also includes a training data module 106 to provide training data according to embodiments of the present technology. The training data from the training data module 106 is used to train the DL network model 104. The processing device 102 can then use the trained DL network model 104 to generate a high-fidelity image. In one embodiment, a high-fidelity image can be defined as an image with an optimized point spread function (PSF), i.e., an image in which blur has been minimized and clarity has been improved compared to the original image.

[0033] The processing device 102 may be included in the workstation 12 of the MRI system 10, or may be included on a separate computing device that communicates with the workstation 12. In addition, in another embodiment, the training data module 106 may be included on a separate computing device and may provide training data before the actual operation of the MRI system 10. The training data may include the medical image itself or a transformed version (such as Fourier (i.e., k-space) or Radon transform data). In addition, the training data may include input image training data. In one embodiment, the input image training data may be original data or k-space data together with a perturbation signal, and the output of the DL network may be a reconstructed medical image of the input image training data or some type of reconstructed medical image data. In another embodiment, the input image training data may be the image itself with some artifacts or noise. In yet another embodiment, the output of the DL network 104 may be only some reconstructed k-space data, which are converted into high-fidelity images by a separate reconstruction module, for example, with the aid of Fourier transform. In one embodiment, the input image training data may include perturbed input image training data, which may be some combination of original data, random noise, and at least one perturbation signal. The details will be explained in the following paragraphs.

[0034] Figure 2B 1 is a flow chart of an exemplary method 150. The method 150 may be implemented on the medical image processing system 100. However, the method may also be used with other medical imaging systems, such as an X-ray imaging system, a computed tomography (CT) imaging system, or an ultrasound imaging system. In an exemplary embodiment, the method includes training the DL network using input image training data at step 152. The input image training data may be undersampled image training data of a plurality of patients stored historically. In one embodiment, the undersampled image training data may be obtained based on the fully sampled image data itself.

[0035] In one embodiment, the training of the DL network is unsupervised training, that is, fully sampled ground truth images are not required to train the DL network. The input image training data includes original image data and at least one perturbation signal. In one embodiment, the input image training data may include original image data with multiple perturbation signals. In one embodiment, the original image data is low-resolution or undersampled image data of multiple patients stored historically. In addition, the original image data may be k-space data or a medical image itself. In one embodiment, the perturbation signal may be synthetically generated. In another embodiment, the perturbation signal is a signal of very small amplitude relative to the original image data, but large enough to make it detectable. For example, the perturbation signal may be a very sparse image of a single point (single pixel). Therefore, the perturbation signal does not substantially change the final image appearance. In one embodiment, the perturbation signal is converted into the same format as the original image data, i.e., the k-space data or the image itself. In addition, it should be noted that the perturbation signal may not be a noise signal like Gaussian noise, because the purpose of using the perturbation signal to train the DL network is to retain image details or maximize feature retention even in the presence of noise in the image. In fact, in an embodiment, noise signals are added to the original image data and the perturbation signal used to train the DL network, respectively, so that the DL network should be able to provide clear image details even in the presence of noise. Figure 3 The details of the training process are described.

[0036] Typically, the purpose of a fully trained DL network is to provide high-fidelity image data (i.e., high definition-to-noise ratio) based on input image data (i.e., input signals to the DL network). In one embodiment, during the training process, the DL network reconstructs the input image data multiple times: typically once with or without some noise added to the original image data, and at least another time with a small perturbation added to the original image data. Because the DL network needs to reconstruct the input training image data multiple times, multiple copies of the same DL network (i.e., multiple DL networks with the same parameters) may be used in a specific implementation. For example, if 3 perturbation signals are used to train the DL network, 4 DL networks (4 copies of the same DL network) may be used in one embodiment. In this embodiment, one DL network may reconstruct the original image data with or without some noise added, and the other three DL networks may reconstruct the original image data with one of the three different perturbation signals used. The parameters of the DL network (e.g., weights and offsets for different layers of the DL network) are then modified based on a loss function. The loss function is used to minimize the difference or error between multiple DL network outputs (i.e., reconstructed images or their k-space data generated by multiple DL networks).

[0037] For MRI systems, the MR image reconstruction problem can be defined as the following equation 3:

[0038] Ax=b (3)

[0039] Where x is the desired MR image, b is the acquired MR k-space data or raw image data, and A is a combination of the MR system coil sensitivities (known in advance), the Fourier operator, and the sampling mask (known in advance). The matrix A is then inverted and multiplied by b to obtain x or the desired image. However, the inversion of the matrix A is ill-conditioned. Therefore, the solution is usually regularized by adding some additional constraints, as shown in Equation 4:

[0040]

[0041] Where ∈ is a data consistency threshold determined based on the noise level of the data. These coupled equations in the above equations can be solved using direct inversion or iterative methods to recover the image or some expected solution x (D operator) that is consistent with the known data but satisfies the desired image. In one embodiment, a deep learning network can be used as a D operator or directly solve equation 4 to generate image data x based on the original image data b.

[0042] The entire reconstruction process including the DL network for solving Equation 4 can be represented as an operator or function R that accepts the raw image data (and possible auxiliary information such as coil sensitivities) and returns the estimated image data (x e ):

[0043] R(b)= x e (5)

[0044] Typically, the estimated image data x is converted into e With image data x g The ground truth version is compared to obtain a scalar loss value (l):

[0045] l=L(x e x g ) (6)

[0046] As will be appreciated by those skilled in the art, the loss function L may be an L2 norm (sum of squared errors) function or a structural similarity (SSIM) function. The weights of the DL network R are adjusted via gradient back propagation to minimize the loss l.

[0047] In the proposed DL network training suggested in step 252, a low amplitude image perturbation signal (p) and random noise (σ) are added to the input signal b (i.e., small perturbations overlaid on the undersampled original image data), resulting in an estimated image (x e ′):

[0048] R(b+Ap+σ)= x e ′ (7)

[0049] This leads to a new loss function:

[0050] l=L(x e ′-x e, p) (8)

[0051] The loss function requires that two reconstructed images (i.e., the first reconstructed image x without perturbation) e and the second reconstructed image x with perturbation e ′) (referred to as at least one restored disturbance signal) should be equal to the originally added disturbance signal p. In one embodiment, the first reconstructed image and the second reconstructed image are part of a plurality of reconstructed images. The loss function may be an L2 norm (error square sum) function or a structural similarity (SSIM) function. It should be noted that the function R in equation (5) and equation (7) is the same. The only difference between equation (5) and equation (7) is the input data on which the function R operates, i.e., b+Ap+σ in equation (7), instead of b in equation (5).

[0052] It should also be noted that in another embodiment, random noise (σ) can be added to the original image data b in Equation 5 or to both Equation 5 and Equation 7, instead of adding random noise (σ) to the perturbation p in Equation 7. In addition, as previously discussed, in one embodiment, multiple DL networks with the same parameters can be used to reconstruct the data. In such an embodiment, multiple perturbation signals can be used while executing the DL network and added to the original image data b. In other words, multiple reconstructed images are generated by executing the DL network multiple times using the original image data, the noise data, the multiple perturbation signals, or various combinations thereof. Therefore, in one embodiment, multiple equations such as Equation (7) can be used, and the loss l will be the difference of the perturbations in the following equations:

[0053] l=L(x e1 ′-x e2 ′,p1-p2) (9)

[0054] where p1 and p2 are two different perturbations, and x e1 ′ and x e2 ′ is the image estimated by two DL networks (with the same parameters) with two perturbations p1 and p2 as input respectively along with the original image.

[0055] Furthermore, it should be noted that the DL networks presented here are nonlinear and object-dependent. In order for these DL networks to work, the original image data b is required. If these DL networks are trained only on the perturbation signal without any original image data (i.e., evaluating R(Ap)=p e And directly minimize p and p e The distance between the two images is 2(m / s), then the DL network will optimally reconstruct the perturbation, but will not generalize well to image reconstruction. Therefore, the DL network should be trained on the original image data with small perturbations rather than on the perturbation signal itself. In addition, the perturbation should be very small relative to the original image data. Typically, the perturbation should be hidden in the image, using it as a carrier signal during the reconstruction process.

[0056] By training the DL network to recognize perturbations in the image data, the trained DL network can recognize image features with high accuracy. Furthermore, in one embodiment, when training the DL network, some noise data can be added to the image data so that the DL network should be able to recognize perturbations even in the presence of noise. Figure 3 The details of the training process are described.

[0057] It should be noted that in the present technology, the input image training data with the perturbation signal and the corresponding output image training data are only used to train the DL network. After training, once the DL network is deployed in the system and used for actual subject data, the perturbation signal is not used. Therefore, after the actual deployment of the DL network, the DL network only operates to reconstruct the actual MR image based on the acquired MR data fed to the DL network. Therefore, multiple DL networks with the same parameters may be needed only for training purposes, but during deployment into a real-time system, only one DL network may be sufficient.

[0058] Return to reference Figure 2B , actual image data of the subject is acquired in step 154. Depending on the medical imaging system, the image data may be MR data, CT data, ultrasound data, or X-ray data. In addition, the image data may be for a specific organ or body part of the subject, such as the brain, knee, shoulder, etc. The acquired image data of the subject is then fed to the DL network trained in step 152. At step 156, the method includes generating reconstructed image data from the acquired image data by executing the trained DL network. Finally, at step 158, a medical image of the subject is generated based on the reconstructed image data. The reconstructed image data may be reconstructed k-space data with optimized feature retention. It should be noted that if the DL network itself is trained to generate a medical image of the subject based on the input data, step 158 is redundant, i.e., step 158 is optional depending on the nature of the DL network.

[0059] Figure 3 is a block diagram 200 illustrating the training of a DL network according to an embodiment of the present technology. Figure 3 5, a sampling mask 203 is applied to the training image data to generate raw image data 204. In the illustrated embodiment, the raw image data 204 is undersampled k-space data of the patient. The raw image data 204 is fed as an input signal to a DL network 208. The output of the DL network 208 then generates a first reconstructed image 210. In one embodiment, the DL network also receives a sensitivity map 212 as an input to generate the first reconstructed image 210. In general, the DL network represents the function R of Equation 5 and converts the raw image k-space data 204 into the first reconstructed image 210.

[0060] also, Figure 3 A disturbance signal 214 is also shown in FIG. As mentioned earlier, when training a DL network, some noise data can be added to the image data so that the DL network should be able to recognize the disturbance or image features even in the presence of noise. Figure 3 A noise signal 202 is shown. The noise signal image 202 shown is a Gaussian noise image, but other noise signals may be used. The perturbation signal image 214 and the noise signal image 202 are combined and converted in the same format as the original image data 204 using the same sampling mask 203. The resulting image data 216 (referred to as perturbation image data) is therefore k-space data of the sampled perturbation signal with Gaussian noise. It should be noted that if the original image data 204 is an image rather than k-space data, the noise signal and the perturbation signal do not need to be converted to k-space data.

[0061] The perturbed image data 216 is then added to the original image data 204 using a summing block 210. The output of the summing block 210 is referred to as the perturbed input training data 220. The perturbed input training data 220 is fed to the DL network 222 as an input signal. It should be noted that both the top and bottom DL networks 208 and 222 are exactly the same and represent the function R of Equation 5 and Equation 7, respectively. Therefore, the same sensitivity map 212 is also provided to the DL network 222. In one embodiment, the same network 208 can be used twice (rather than using two DL networks 208 and 222 separately), for example, the first time without the perturbed input training data (the output of the summing block 206), and then the second time with the perturbed input training data 220.

[0062] The output of the DL network 222 is a second reconstructed image 224. A subtraction block 226 subtracts the second reconstructed image 224 from the first reconstructed image 210. The subtraction results in a restored perturbation signal 228. The DL network 222 is trained to produce a restored perturbation signal 228 that is similar to the perturbation signal 214. Therefore, a subtraction block 230 is used to subtract the restored perturbation signal 228 from the perturbation signal 214 to generate a loss image 232 (also referred to as a loss metric). The loss function then operates on the loss metric 232 to minimize the loss, and the parameters (such as "weights" or "biases") of the DL network 222 are adjusted based on the loss function. As previously described, since the DL networks 208 and 222 are exactly the same, changing the parameters of the DL network 222 will also result in changing the parameters of the DL network 208. It should be noted that if, instead of generating images, the DL networks 208 and 222 generate some reconstructed k-space data of the corresponding original image data 204 and the perturbed input training data 220, then the loss image 232 will also be the reconstructed k-space data of the loss image 232, rather than the image itself.

[0063] Figure 4 314, 316. The exemplary DL network model 300 may be used as a DL network model in any of the previous embodiments. The exemplary DL network model 300 includes layers 320, 340, 360, and 380. Layers 320 and 340 are connected using neural connections 330. Layers 340 and 360 are connected using neural connections 350. Layers 360 and 380 are connected using neural connections 370. Data flows forward from input layer 320 to output layer 380 via inputs 312, 314, 316 and arrives at output 390. Inputs 312, 314, 316 may be original image data 204, perturbation signals 214, sensitivity maps 212, etc.

[0064] Layer 320 is the input layer, which Figure 4 The example includes multiple nodes 322, 324, 326. Layers 340 and 360 are hidden layers and are Figure 4 The example includes nodes 342, 344, 346, 348, 362, 364, 366, 368. The DL network model 300 may include more or fewer hidden layers 340 and 360 than shown. Layer 380 is the output layer and is Figure 4The example of DL network model 300 includes a node 382 with an output 390. Each input 312 to 316 corresponds to a node 322-326 of the input layer 320, and each node 322-326 of the input layer 320 has a connection 330 to each node 342-348 of the hidden layer 340. Each node 342-348 of the hidden layer 340 has a connection 350 to each node 362-368 of the hidden layer 360. Each node 362-368 of the hidden layer 360 has a connection 370 to the output layer 380. The output layer 380 has an output 390 to provide an output from the example DL network model 300.

[0065] In the connections 330, 350 and 370, some example connections 332, 352, 372 may be given increased weights, while other example connections 334, 354, 374 may be given less weights in the DL network model 300. For example, input nodes 322-326 are activated by receiving input data via inputs 312-316. Nodes 342-348 and 362-368 of hidden layers 340 and 360 are activated by data flowing forward through the network model 300 via connections 330 and 350, respectively. After the data processed in the hidden layers 340 and 360 is sent via connection 370, a node 382 of the output layer 380 is activated. When the output node 382 of the output layer 380 is activated, the node 382 outputs an appropriate value based on the processing completed in the hidden layers 340 and 360 of the DL network model 300.

[0066] Typically, a deep learning network is characterized by using one or more artificial neural network architectures to extract or simulate data of interest. A deep learning network can be implemented using one or more processing layers (e.g., layers 320, 340, 360, and 380), wherein the configuration and number of each layer allows the deep learning network to handle complex information extraction and modeling tasks. Specific parameters (or referred to as "weights" or "biases") of the network are usually estimated through a so-called learning process (or training process). The learned or trained parameters usually result in (or output) a network corresponding to layers of different levels, so that the extraction or simulation of different aspects of the initial data or the output of the previous layer can usually represent a hierarchical structure or a cascade of layers. Therefore, processing can be performed layer by layer. That is, "simple" features can be extracted from the input data of an earlier or higher-level layer, and then these simple features are combined into a layer presenting features of higher complexity. In practice, each layer (or more specifically, each "neuron" in each layer) can use one or more linear and / or nonlinear transformations (so-called activation functions) to process the input data into output data for representation. The number of "neurons" may be constant across layers, or may vary from layer to layer.

[0067] As discussed herein, a training data set with known input values ​​(e.g., a known MRI image or known k-space data, a known perturbation, etc.) can be used as part of the initial training of a deep learning process to solve a specific problem (i.e., a high-fidelity image). In this way, a deep learning algorithm can process a known data set or a training data set (in an unsupervised or unguided manner in this embodiment) until a mathematical relationship between the initial data and the expected output is identified (e.g., the R function discussed above with respect to Equations 5 and 7). The created output is then compared with the expected (target) output of the data set, and the generated difference from the expected output (i.e., loss) is then used to iteratively update the network parameters (weights and offsets) using a loss function (e.g., L Equation 8). One such update / learning mechanism uses a stochastic gradient descent method to update the parameters of the DL network. Obviously, those skilled in the art will appreciate that other methods known in the art may also be used.

[0068] Figure 5 A schematic diagram 400 depicting comparative experimental results of various reconstruction techniques according to an embodiment of the present technology is shown. Schematic diagram 400 shows a comparison of reconstruction techniques for a 6x accelerated magnetization prepared rapid gradient echo (MPRAGE) MRI scan. In general, a 6X accelerated MPRAGE scan is faster than a traditional MPRAGE scan and involves acquiring a series of low-resolution images of the brain in parallel (compared to a traditional MPRAGE scan), and then using image processing techniques (such as a deep learning network model) to generate a high-resolution image compared to the acquired low-resolution images.

[0069] Specifically, in the schematic diagram 400, column 402 represents three reconstructed MRI images obtained using three different reconstruction techniques. For example, image 404 corresponds to a reconstructed image acquired using the non-regularized iterative sensitivity encoding (SENSE) technique (which is a parallel imaging technique without DL regularization), while image 406 corresponds to a reconstructed image acquired using the unfolded self-supervised deep learning network technique (which utilizes a feed-forward network trained using back-propagation and a mean absolute error loss function). In addition, image 408 corresponds to a reconstructed image acquired using the deep learning technique presented herein (which utilizes perturbed input training data).

[0070] Furthermore, in diagram 400, column 410 represents relative sharpness maps, column 412 represents relative noise maps, and column 414 represents sharpness-noise maps for images 404, 406, and 408. For example, for image 404, the three corresponding maps are 416, 418, 420. Furthermore, for image 406, the three corresponding maps are 422, 424, and 426, and for image 408, the three corresponding maps are 428, 430, and 432.

[0071] The iterative SENSE technique provides the clearest image as shown in Figure 416, but suffers from extreme noise amplification (Figure 418), resulting in a low clarity-noise map 420. The extreme noise amplification of this technique makes it unclear whether the pixels in the image are real signals or noise. On the other hand, the unfolding self-supervised training technique produces a very low noise image (Figure 424), but has relatively poor clarity (Figure 422), resulting in a very high clarity-noise map 426. In this technique, clarity and noise are related to anatomical structures, making the detectability of low contrast regionally variable. The proposed deep learning technique using perturbed input training data presented in this article balances clarity and noise amplification to maximize feature detectability. Using this technique, clarity (Figure 428) is generally higher than using the unfolding self-supervised training technique, while noise amplification (Figure 430) is not as high as the iterative SENSE technique. The relative clarity-noise (Figure 432) is appropriately high, and interestingly, it is not related to anatomical structures.

[0072] Figure 6 A schematic diagram 500 depicting another set of comparative experimental results of various reconstruction techniques according to an embodiment of the present technology is shown. Specifically, in schematic diagram 500, rows 502, 504, and 506 correspond to three reconstruction techniques: 1) iterative SENSE; 2) unfolding self-supervised deep learning; and 3) the proposed deep learning technique with perturbed input training data. In addition, columns 508, 510, 512, and 514 respectively show the reconstructed images obtained using these techniques with 4 times, 6 times, 8 times, and 10 times the expected acceleration. It can be seen that the iterative SENSE technique (row 502 image) looks clear, but the noise is very large. The self-supervised technique (row 504 image) produces clear images that are not optimized for small feature recognition. The proposed DL technique with perturbed input training data (row 506 image) maintains a natural appearance with a balance of noise and subtle features.

[0073] One of the advantages of this technique is that it provides better image quality, especially for using low-resolution images to optimize small feature retention. Since this technique utilizes low-resolution images, this technique is able to achieve higher acceleration rates, resulting in faster scans for patients. Faster scans also result in fewer motion artifacts in the final image. In addition, this technique can be combined with other loss functions or DL ​​techniques to further improve image quality.

[0074] This written description uses examples to disclose the invention, including the best mode, and also to enable those skilled in the art to practice the invention, including making and using any devices or systems and performing any included methods. The patentable scope of the invention is defined by the claims, and may include other examples that occur to those skilled in the art. Such other examples are intended to fall within the scope of the claims if they have structural elements that do not differ from the literal language of the claims, or if they include equivalent structural elements with insignificant differences from the literal language of the claims.

Claims

1. A method for imaging a subject, the method comprising: training a deep learning network using input image training data, wherein the input image training data comprises original image data and at least one perturbation signal; Using a medical imaging device to acquire image data of the subject; providing the image data of the subject as input to the trained deep learning network; A medical image of the subject is generated based on the acquired image data using the deep learning network.

2. The method of claim 1, wherein the medical imaging device comprises a magnetic resonance imaging system, an X-ray imaging system, a computed tomography imaging system, or an ultrasound imaging system. The method of claim 1 , wherein the original image data is undersampled image data. The method of claim 1 , wherein the at least one disturbance signal is a small amplitude signal relative to the original image data.

5. The method of claim 1 , further comprising updating parameters of the deep learning network based on a loss function, wherein the loss function uses a difference between the at least one perturbation signal and the at least one restored perturbation signal as a loss metric. The method according to claim 5 , wherein the loss function is a sum of squared errors function or a structural similarity function. The method of claim 5 , wherein the at least one recovered disturbance signal is determined based on differences between data of a plurality of reconstructed images.

8. The method of claim 7, wherein a first reconstructed image of the plurality of reconstructed images is determined by executing the deep learning network, wherein an input signal to the deep learning network comprises the original image data.

9. The method of claim 7, wherein a second reconstructed image of the plurality of reconstructed images is determined by executing the deep learning network with a combination of the original image data, noise data, and the at least one perturbation signal as input signals.

10. The method of claim 7, wherein the plurality of reconstructed images are generated by executing the deep learning network multiple times using the original image data, noise data, a plurality of perturbation signals, or a combination thereof. The method of claim 1 , wherein the reconstructed image data comprises a reconstructed image or k-space data thereof.

12. A medical imaging system, comprising: at least one medical imaging device providing image data of a subject; A processing system programmed to perform any of the methods according to claims 1 to 11.

13. A method for generating an image of a subject using a magnetic resonance imaging system, the method comprising: training a deep learning network using magnetic resonance image training data, wherein the magnetic resonance image training data comprises original image data and at least one perturbation signal; using the magnetic resonance imaging system to acquire magnetic resonance image data of the subject; providing the magnetic resonance image data of the subject as input to the trained deep learning network; The image of the subject is generated based on the acquired deep learning image data using the deep learning network.