Magnetic Resonance Imaging (MRI) Examination Acceleration Based on Deep Learning

Abstract

The present invention provides a system and method for accelerating magnetic resonance imaging (MRI) examinations based on deep learning. The method for accelerating magnetic resonance imaging (MRI) examinations based on deep learning (DL) includes acquiring at least one fully sampled reference k-space data of a subject and acquiring multiple partial k-spaces of the subject. The method also includes grafting the multiple partial k-spaces with the at least one fully sampled reference k-space data to generate a grafted k-space for accelerated examination. The method also includes training a deep learning (DL) module using the fully sampled reference k-space data and the grafted k-space to remove grafting artifacts.

Description

Technical Field

[0001] Embodiments of the present disclosure relate generally to improved imaging systems, and more particularly to deep learning-based acceleration of magnetic resonance imaging (MRI) exams. Background Art

[0002] Magnetic resonance imaging (MRI) systems are used in the field of medical imaging to acquire image data of anatomical portions of a human subject (hereinafter referred to as the subject). Several image processing techniques are commercially available to generate images of improved quality from the image data acquired by MRI systems. While images produced by these image processing techniques have good quality, many images are adversely affected by the operating conditions of the MRI system and subject motion.

[0003] Imaging artifacts, or MRI artifacts, are imaging defects that occur during a scan and adversely affect image quality. These artifacts are often caused by the movement of the subject or its organs, such as the heart, lungs, or blood vessels, during imaging. These artifacts are known as motion artifacts. Furthermore, due to the movement of the subject's organs, subsequent images obtained during a long MRI scan can differ from previous images of the same subject. These and other artifacts create challenges during image analysis by radiologists.

[0004] Different image acquisition and processing techniques are known in the art to minimize artifacts. Accelerated MRI examinations are a technique that has shown substantial improvements in image artifacts. Scan time reduction in MRI examinations is primarily focused on accelerating each comparison from different images acquired during the scan. In current methods, MRI acceleration is achieved by undersampling the original data space and sharing information acquired from another protocol (or comparison) in the same location. MRI images reconstructed by techniques such as zero padding and compressed sensing have defects such as structural artifacts, missing structures, and blurring.

[0005] Among the different approaches to improving image quality, deep learning (DL) algorithms are trained algorithms used in the field of medical imaging for computer-aided detection and diagnosis of medical conditions in subjects. A DL module comprising a DL algorithm is typically trained with several MRI images, and the DL module can be used to analyze subsequent MRI images of a subject to identify image patterns. Artifacts such as missing structures and blurring may not be overcome by currently available deep learning (DL) techniques. In the case where deep learning (DL) techniques overcome these artifacts, this is due to data synthesis performed by the DL module rather than artifact removal.

[0006] Existing techniques require a long time to perform multiple contrast MR exams and do not exploit data redundancy through efficient processing.What is needed is an MRI exam acceleration technique that will quickly provide high-resolution images for multiple contrasts in a single exam. Summary of the Invention

[0007] This summary introduces concepts that are described in greater detail in the detailed description. It is not intended to identify essential features of the claimed subject matter, nor is it intended to limit the scope of the claimed subject matter. Its sole purpose is to present the concepts in a simplified form as a prelude to the more detailed description that is presented later.

[0008] According to one aspect of the present disclosure, a method for accelerating a magnetic resonance imaging (MRI) examination is provided. The method includes acquiring at least one fully sampled reference k-space data of a subject and acquiring multiple partial k-spaces of the subject. The method also includes grafting the multiple partial k-spaces with the at least one fully sampled reference k-space data to generate a grafted k-space for accelerating the examination.

[0009] According to another aspect of the present disclosure, a method for accelerating magnetic resonance imaging (MRI) examinations based on deep learning (DL) is provided. The method includes acquiring at least one fully sampled reference k-space data of a subject and acquiring multiple partial k-spaces of the subject. The method also includes grafting the partial k-space of the subject with the fully sampled reference k-space data to generate a grafted k-space for accelerating the examination. The method also includes training a deep learning (DL) module using the fully sampled reference k-space data and the grafted k-space to predict and remove grafting artifacts.

[0010] According to another aspect of the present disclosure, a magnetic resonance imaging (MRI) system is provided. The magnetic resonance imaging (MRI) system includes at least one radio frequency (RF) body coil, which is suitable for transmitting radio frequency (RF) signals to a subject and receiving radio frequency (RF) signals from the subject. The MRI system also includes a transceiver module, which is configured to digitize the signals received by the radio frequency (RF) body coil. The MRI system also includes a control system, which is configured to process the digitized signals and generate k-space data corresponding to an imaging volume of the subject, wherein the MRI system is configured to acquire at least one fully sampled reference k-space data of the subject and multiple partial k-spaces of the subject. The MRI system also includes a computer processor, which is configured to graft the partial k-space with the fully sampled reference k-space data to generate a grafted k-space for accelerated examination. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] These and other features of the embodiments of the present disclosure will be better understood when the following non-limiting embodiments in the detailed description are read with reference to the accompanying drawings, wherein:

[0012] Figure 1An exemplary embodiment of a magnetic resonance imaging (MRI) system according to one aspect of the present disclosure is shown.

[0013] Figure 2 A method for accelerating k-space grafting of MRI image examination according to one aspect of the present disclosure is shown.

[0014] Figure 3 A method for artifact correction of an MRI image according to one aspect of the present disclosure is shown.

[0015] Figure 4 A deep learning module trained using two input channels for predicting artifacts according to one aspect of the present disclosure is shown.

[0016] Figure 5 An exemplary k-space grafted MRI image is shown according to one aspect of the present disclosure.

[0017] Figure 6 A fully sampled reference image, a grafted image, and an artifact-corrected image generated by a DL module according to one aspect of the present disclosure are shown.

[0018] Figure 7 An exemplary accelerated examination performed by a magnetic resonance imaging (MRI) system according to one aspect of the present disclosure is shown.

[0019] Figure 8 An exemplary grafted image without artifacts generated by a DL module according to one aspect of the present disclosure is shown.

[0020] Figure 9 A method for accelerating magnetic resonance imaging (MRI) examinations based on deep learning according to one aspect of the present disclosure is shown. DETAILED DESCRIPTION

[0021] The following detailed description of exemplary embodiments refers to the accompanying drawings. The same reference numerals in different figures identify the same or similar elements. In addition, the drawings are not necessarily drawn to scale. In addition, the following detailed description does not limit the present invention. Instead, the scope of the present invention is defined by the appended claims.

[0022] Reference throughout this specification to "one embodiment," "another embodiment," or "some embodiments" means that a feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the disclosed subject matter. Thus, the appearances of the phrases "in one embodiment," "in an embodiment," or "in some embodiments" in various places throughout this specification are not necessarily all referring to the same embodiment. Furthermore, the features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0023] The present disclosure provides a method for accelerating magnetic resonance imaging (MRI) examinations based on deep learning (DL). The method includes acquiring at least one fully sampled reference k-space data of a subject and acquiring multiple partial k-spaces of the subject. The method grafts the partial k-spaces with the fully sampled reference k-space data to generate grafted k-spaces for accelerating the examination. The method also includes training a deep learning (DL) module using the fully sampled reference k-space data and the grafted k-spaces to predict and remove grafting artifacts.

[0024] Medical imaging devices, such as magnetic resonance imaging (MRI) systems, generate images representing body parts (e.g., organs, tissues) to diagnose and treat diseases. The transmission, acquisition, processing, analysis, and storage of medical image data play an important role in the diagnosis and treatment of patients in healthcare settings.

[0025] The methods, apparatus, and articles described herein can be applied to a variety of healthcare and non-healthcare systems. In one example, the methods, apparatus, and articles described herein can be applied to the components, configuration, and operation of a magnetic resonance imaging (MRI) system. Figure 1 An exemplary embodiment of a magnetic resonance imaging (MRI) system is shown in which the methods, apparatus, and articles of manufacture disclosed herein may be used.

[0026] Embodiments of the present disclosure will now be described by way of example with reference to the accompanying drawings, in which Figure 11 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 wand, a voice controller, and / or other input devices. The input device (13) can be used for interactive geometric shape specification. The console (12) communicates with a computer system (20) via a link (18), which enables the 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) can include modules that communicate with each other via a backplane (20a). The modules of the computer system (20) can include, for example, an image processor module (22), a central processing unit (CPU) module (24), and a memory module (26), which can include a frame buffer for storing image data arrays. The computer system (20) can be connected to an archival media device, permanent or backup storage, 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) can be separate from the computer system (20) or integrated therewith. The computer system (20) and the MRI system control (32) together form an "MRI controller" (33) or "controller."

[0027] In an exemplary embodiment, the MRI system control (32) includes modules connected by a baseboard (32a). These modules include a CPU module (36) 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 a scan sequence to be performed. The CPU module (36) operates the system components to perform 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).

[0028] In one example, the CPU module (36) receives patient data from a physiological acquisition controller (44) that receives signals from sensors connected to the subject, such as ECG signals received from electrodes attached to the patient. The CPU module (36) receives signals from the 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.

[0029] The whole body RF coil (56) is used to transmit the waveform toward the subject's anatomy. The whole body RF coil (56) can be a body coil (such as Figure 1 ). The RF coil may also be a local coil, which may be placed closer to the subject's anatomy than a body coil. The RF coil (56) may be a surface coil. A surface coil containing receive channels may be used to receive signals from the subject's anatomy. A typical surface coil will have eight receive channels; however, a different number of channels is possible. It is known to use a combination of both a body coil (56) and a surface coil to provide better image quality.

[0030] The pulse generator module (38) can operate the gradient amplifier (42) to achieve the desired timing and shape of the gradient pulses generated during the scan. The gradient waveforms generated by the pulse generator module (38) can be applied to the gradient amplifier system (42) having Gx, Gy and Gz amplifiers. Each gradient amplifier excites a corresponding physical gradient coil in the gradient coil assembly (50) to generate a magnetic field gradient for spatially encoding the acquired signals. The gradient coil assembly (50) can form part of the magnet assembly (52), which also includes a polarizing magnet (54) (in operation, the polarizing magnet provides a longitudinal magnetic field B0 throughout the target space (55) surrounded by the magnet assembly (52)) and a whole-body RF coil (56) (in operation, the coil provides a transverse magnetic field B1 that is approximately perpendicular to B0 throughout the target space (55)). The transceiver module (58) in the MRI system control (32) generates pulses, which are amplified by the RF amplifier (60) and coupled to the RF coil (56) through the transmit / receive switch (62). The resulting signals emitted by the excited nuclei in the subject's anatomy can be sensed by a receive coil (not shown) and provided to a preamplifier (64) via a transmit / receive switch (62). The amplified MR signals are demodulated, filtered, and digitized in the receiver portion of the transceiver (58). The transmit / receive switch (62) is controlled by a signal from the pulse generator module (38) to electrically connect the RF amplifier (60) to the coil (56) during transmit mode and to connect the preamplifier (64) to the receive coil during receive mode.

[0031] 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 perceptible and reproducible manner by a conventional computer (e.g., text or images printed on paper or displayed on a screen, a compact disc or other optical storage medium, "flash" memory, EEPROM, SDRAM, or other electrical storage medium; a floppy disk or other magnetic disk, a magnetic tape, or other magnetic storage medium).

[0032] The scan is complete when an array of raw k-space data is acquired in a 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 an array processor (68) which operates to reconstruct the data into an array of image data using a reconstruction algorithm such as a Fourier transform. 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 central k-space data is obtained, the image may be referred to as central k-space. The image data is transmitted to the computer system (20) via a data link (34) and stored in memory. In response to commands received from the operator console (12), the image data may be archived in long-term storage or may be further processed by the image processor (22) and transmitted to the operator console (12) and presented on a display (16).

[0033] According to one aspect of the present disclosure, Figure 2 A method (200) is shown for grafting a partial k-space (220) with fully sampled reference k-space data (210) to obtain an accelerated image (230). The duration of an MRI scan can be extended from a few minutes to several minutes. Because an MRI scan involves acquiring the entire k-space (210) during each scan, acquiring images in the shortest possible time is critical for accelerated examinations. According to one aspect of the present disclosure, k-space (220) can be divided into a central portion (221) of k-space (220) and a peripheral portion (222) of k-space (220). Among image parameters, the central portion (221) of k-space determines image contrast and image brightness; and the peripheral portion (222) of k-space determines the quality of the edges of the image (220). Traditionally, an MRI examination consists of acquiring multiple contrasts or images (220) of a subject, and accelerated MRI examinations focus on accelerating each contrast.

[0034] According to one aspect of the present disclosure, fully sampled reference k-space data (210) may be obtained at the beginning of a scan. The acquisition of the reference data may include the acquisition of the entire k-space (210). The fully sampled reference k-space data (210) may be stored for future reference and used throughout the scan. This is also referred to as a fully sampled image (210). It is desirable to reduce the acquisition time of the fully sampled reference k-space data (210) for faster imaging. For accelerated examinations, only a partial k-space (221) representing only a portion of the entire k-space (220) or a central portion of the k-space may be acquired instead of acquiring the entire k-space (220). Acquiring only the partial k-space (221) instead of the entire k-space (220) during a subsequent scan speeds up the image acquisition process. The partial k-space (221) acquired during the subsequent scan may be grafted with the outer k-space of the fully sampled reference data (210). In this way, structural information from the reference k-space data (210) may be shared during another phase of the MRI examination without compromising contrast. By grafting partial k-space (221) with external k-space from reference data (210), complete k-space data (230) is obtained - grafting partial k-space (221) with reference external k-space data (210) provides accelerated inspection, which increases scan throughput and reduces artifacts during longer scans.

[0035] According to Figure 3 One aspect of the present disclosure is disclosed in a method (300) for artifact correction of MRI images. A grafting operation can be performed before k-space data is processed by a deep learning (DL) module. The grafting process can generate artifacts, such as sharp high-frequency artifacts, which can be removed using a deep learning network according to one aspect of the present disclosure.

[0036] Deep learning is a type of machine learning technique that uses representation learning methods. This allows a machine to be given raw data and determine the representation needed to classify it. Deep learning uses the backpropagation algorithm to detect structure within a dataset. Deep learning machines can utilize a variety of multi-layer architectures and algorithms. For example, while machine learning involves identifying features to train a network, deep learning processes raw data to identify interesting features without external identification.

[0037] Deep learning in a neural network environment involves numerous interconnected nodes called neurons. Input neurons, activated by external sources, activate other neurons based on connections with those neurons, which are controlled by the machine's operating conditions. Neural networks function in a specific way based on their own sequences. Learning improves the machine's output and the connections between neurons in the network, causing the neural network to function in the desired manner.

[0038] Deep learning with convolutional neural networks uses convolutional filters to segment data to locate and identify learned observable features in the data. Each filter, or layer, of the CNN architecture transforms the input data to increase its selectivity and invariance. This abstraction of the data allows the machine to focus on the features in the data it is trying to classify and ignore irrelevant background information.

[0039] Deep learning operates on the understanding that many datasets consist of both high-level and low-level features. For example, at a high level, when examining an image, rather than looking for objects, it's more effective to look for edges, which form motifs, which form parts, and which form the objects we're looking for. This hierarchy of features can be seen in many different forms of data, such as speech and text.

[0040] The learned observable features include the objects and quantifiable regularities learned by the machine during supervised learning. Compared to traditional algorithms that have not been continuously trained to classify data, machines that are provided with a larger set of well-classified data are better equipped to distinguish and extract features in successfully classifying new data.

[0041] A deep learning machine using transfer learning can correctly link data features to certain classifications confirmed by human experts. Conversely, the same machine can update the system used for classification when a human expert reports a classification error. For example, settings and / or other configuration information can be guided by the use of learned settings and / or other configuration information, and the number of changes and / or other possibilities for settings and / or other configuration information for a given situation can be reduced as the system is used more times (e.g., repeatedly and / or by multiple users).

[0042] For example, an exemplary deep learning neural network can be trained using an expert classification dataset. This dataset builds the neural network and is a supervised learning phase. During the supervised learning phase, the neural network can be tested to see if it has achieved the desired behavior.

[0043] Once the desired neural network behavior has been achieved (e.g., the machine has been trained to operate according to specified thresholds, etc.), the machine can be deployed for use (e.g., using "real" data to test the machine, etc.). During operation, the neural network classification can be confirmed or rejected (e.g., by an expert user, an expert system, a reference database, etc.) to continue to improve the neural network behavior. The exemplary neural network is then in a state of transfer learning because the classification conditions that determine the neural network behavior are updated based on ongoing interactions. In some examples, the neural network can provide direct feedback to other patient examination processes in a healthcare facility that can be connected to the neural network. In some examples, data output by the neural network is buffered (e.g., via the cloud, etc.) and validated before being provided to another process.

[0044] Deep learning machines using convolutional neural networks (CNNs) can be used for image analysis. CNN analysis stages can be used for facial recognition in natural images, computer-aided diagnosis (CAD), and more.

[0045] Medical image data can be acquired using imaging modalities such as magnetic resonance imaging (MRI). Medical image quality is often affected not by the machine producing the image, but by the patient. For example, patient movement during an MRI can create blurred or distorted images, which can hinder accurate diagnosis.

[0046] Interpreting medical images without regard for quality is a recent development. Medical images are largely interpreted by physicians, but these interpretations can be subjective, influenced by physician experience and / or fatigue in the field. Image analysis via machine learning can help improve the efficiency of healthcare practitioners' workflows.

[0047] For example, deep learning machines can provide computer-aided detection support to improve image analysis in terms of image quality and classification. However, when applied to the medical field, deep learning machines often face challenges, resulting in numerous misclassifications. For example, deep learning machines must overcome small training datasets and repeated adjustments.

[0048] For example, deep learning machines can be used to determine the quality of medical images with minimal training. Semi-supervised and unsupervised deep learning machines can be used to quantitatively measure aspects of image quality. For example, deep learning machines can be used after an image has been acquired to determine whether the image quality is sufficient for diagnosis. Supervised deep learning machines can also be used for computer-assisted diagnosis. For example, supervised learning can help reduce susceptibility to misclassification. These deep learning machines can improve computer-assisted diagnosis over time through training and transfer learning.

[0049] According to Figure 3In one aspect of the present disclosure shown, fully sampled reference k-space data (310) containing the entire k-space data may be acquired, and an accelerated image containing only a portion of the k-space may be acquired. In addition, the accelerated partial k-space may be grafted with the fully sampled reference k-space data (310) to generate accelerated image data (320). The accelerated image data (320) and the fully sampled reference k-space data (310) may be input to an artifact prediction network (330). The artifact prediction network (330) may be a deep learning (DL) network. According to one aspect of the present disclosure, the artifact prediction network (330) may employ dual regularization and an adaptive smart loss function (331). The artifact prediction network (330) may use a smart loss function (331) to predict grafted artifacts. Predicting the grafted artifact includes comparing the grafted data (320) with the fully sampled reference k-space data (310) using the smart loss function (331). Providing fully sampled reference k-space data (310) as an additional channel along with the grafted data (320) and processing these images through a dual-regularized adaptive loss function can more accurately predict grafting artifacts. Sharp high-frequency artifacts can be removed by a DL module using the fully sampled reference k-space data (310) and the grafted data (320). These artifacts cannot be removed by a DL module trained only using the grafted data (320).

[0050] At the beginning of the calculation, the loss function used in the conventional DL module assigns a fixed weight to the regularization term. However, the smart loss function according to one aspect of the present disclosure is a dynamic smart loss function that can be constructed to change the weight assigned to the regularization term based on the relevance of the regularization term as training proceeds. In one example, the regularization weight of the loss term is modulated based on the training loss after each epoch. A transfer function can be adopted by the artifact prediction network to determine the regularization weights of the loss function components. The loss function components may include structural similarity index measure (SSIM) loss, perceptual (e.g., based on latent features) loss, mean absolute error (MAE) loss, which can be dynamically adjusted as training proceeds to obtain better image quality.

[0051] In one example, a higher use of SSIM at the beginning can result in fine banding artifacts in MR images. However, using a dynamic regularization method according to one aspect of the present disclosure, the role of SSIM in the loss function can be controlled through training. As the prediction gets closer to the ground truth, more weight can be placed on SSIM. In addition, the weight of the regularization term after each epoch can be updated by a transfer function. In an exemplary embodiment of the present disclosure, the intelligent loss function calculation can be based on the following formula:

[0052]

[0053] The MAE can be calculated based on the predicted residuals, and the SSIM can be calculated between the ground truth and the artifact-corrected image. The weight of the regularization term α can be updated after each epoch based on the MAE and SSIM values. As the prediction gets closer to the ground truth, the weight of the SSIM loss can be increased. In addition, the user can decide how soon the weight of the SSIM starts to increase. According to the dynamic regularization method of the present disclosure, the weight of the regularization term (α) can be dynamically adjusted using the following formula:

[0054]

[0055] in is the ratio offset term, which determines the nature of the curve to obtain the corresponding r m The higher the R value, the greater the effect of SSIM.

[0056] The method of the present disclosure accelerates MRI examinations, so that the time saved by using the method can be used to obtain higher resolution images for multiple comparisons. The method can more efficiently collect high-resolution data for multiple comparisons. According to one aspect of the present disclosure, Figure 4 A deep learning (DL) module trained using dual input channels for predicting artifacts is shown. The DL module can be trained using only the grafted data (401) and using both the fully sampled reference k-space data (402) and the grafted data (dual input) (401) to analyze and predict (403) the artifact removal efficiency of the DL module. It can be seen that artifact removal by the DL module is improved when dual input is provided. It can be seen that the visual information fidelity (VIF) and structural similarity index metric (SSIM) of the output are improved in the case of the dual input model.

[0057] Model VIF average VIF standard deviation Average pSNR pSNR standard deviation SSIM average SSIM standard deviation Dual input 0.672 0.032 36.325 2.099 0.962 0.042 Single input 0.659 0.045 35.554 2.157 0.957 0.051

[0058] According to one aspect of the present disclosure, Figure 5 The available fully sampled reference k-space data (510) and partial k-space (520) are shown. The partial k-space (520) can be grafted with the fully sampled reference k-space data (510) to obtain a grafted k-space (530). According to one aspect of the present disclosure, if the partial k-space (520) has artifacts (521), a deep learning module trained using the partial k-space (520) may not be able to identify these artifacts (521). In one example, if there is a Figure 5If artifacts such as blur (521) in the portion of k-space (520) shown are used to train the DL module, the DL module may not recognize these image artifacts (521) for all images that may be presented to the DL module for comparison. Since only the portion of k-space (520) with artifacts (521) is used to train the DL module, the structural integrity of the image may be lost.

[0059] According to one aspect of the present disclosure, when partial k-space (520) is grafted with fully sampled reference k-space data (510), the grafted data (530) has improved structural integrity and is free of undersampling artifacts (521). In the presence of artifacts (521) in the partial k-space (520), the DL module may not be able to restore full structural integrity. However, grafting enables DL correction even for higher levels of acceleration. Thus, the grafting process can remove artifacts (521) to provide data (530) with improved image quality.

[0060] According to one aspect of the present disclosure, Figure 6 An example of image data obtained using an MRI system is shown. The data may include fully sampled reference k-space data (610), grafted data (620), and artifact-corrected data (630) generated by a DL module trained using the fully sampled reference k-space data (610) and the grafted data (620). As can be seen, the grafted data (620) may have grafting artifacts, and if the DL module is trained using only the grafted data (620), the DL module may not be able to detect these artifacts during further analysis of subsequent images presented to the DL module. According to one aspect of the present disclosure, when the DL module is trained using the fully sampled reference k-space data (610) and the grafted data (620), grafting artifacts such as sharp high-frequency artifacts can be removed to generate artifact-corrected data (630).

[0061] According to one aspect of the present disclosure, Figure 7An example of an accelerated examination performed by a magnetic resonance imaging (MRI) system is shown. Acquiring fully sampled reference k-space data (710) using an exemplary MRI system may require approximately up to six minutes and thirty-five seconds. This data (710) may be used as reference k-space data to represent ground truth. A partial k-space (720) represents only the central k-space acquired according to one aspect of the present disclosure. In one example, the partial k-space (720) may be acquired in approximately three minutes and eighteen seconds. The partial k-space (720) and subsequently acquired k-space may be grafted with the complete k-space data stored on a computer memory to obtain grafted k-space (720). Similarly, accelerated k-space (730) may be obtained using partially acquired k-space data and grafting the partial k-space data with reference k-space data to generate accelerated k-space data (730). The approximate time required to obtain the grafted image may be up to two minutes and thirteen seconds. Improvements in the time required to acquire MR images have a significant impact on image quality. The shorter duration of image acquisition minimizes motion artifacts caused by movement of patient organs such as the lungs, heart, and blood vessels. The acceleration of MR imaging according to one aspect of the present disclosure reduces data redundancy in k-space data processing. Multiple contrasted high-resolution isotropic images can be obtained in a single examination with reduced net imaging time.

[0062] According to one aspect of the present disclosure, Figure 8 An exemplary grafted image (820) is shown without artifacts generated by a DL module trained using both fully sampled reference k-space data and grafted data. Figure 8 As shown, the grafted data (820) shows the removal of grafting artifacts and the reduced image acquisition time. The method according to the present disclosure not only accelerates MR imaging, but also allows the generation of high-quality images by acquiring multiple contrasted high-resolution data. Grafting allows sufficient structure to be retained for the DL module to correct the accelerated data and provides easy scalability throughout the segmentation and application process.

[0063] According to one aspect of the present disclosure, Figure 9A method (900) for accelerating a magnetic resonance imaging (MRI) examination based on deep learning (DL) is shown. The method (900) includes acquiring (910) at least one fully sampled reference k-space data of a subject using a magnetic resonance imaging (MRI) system. The method (900) also includes acquiring (920) a plurality of accelerated partial k-spaces of the subject using the MRI system. The method (900) also includes grafting (930) the partial k-spaces with the fully sampled reference k-space data to generate grafted k-spaces for accelerated examination of the subject. The grafting (930) process may be employed before the partial k-spaces are processed by the DL module. According to one aspect of the present disclosure, the grafting (930) process may generate artifacts such as sharp high-frequency artifacts that can be removed using a deep learning network. The method (900) also includes training (940) the deep learning (DL) module using the fully sampled reference k-space data and the grafted data to remove the grafted artifacts.

[0064] According to one aspect of the present disclosure, training (940) a deep learning (DL) module may include inputting the grafted k-space and the reference k-space into an artifact prediction network. The artifact prediction network may be a deep learning (DL) network. According to one aspect of the present disclosure, the artifact prediction network may employ dual regularization and an adaptive smart loss function. The artifact prediction network may use the smart loss function to predict grafting artifacts. Predicting grafting artifacts includes comparing the grafted image to the fully sampled image using the smart loss function. Providing fully sampled reference k-space data as an additional channel along with the grafted data and processing these images through a dual regularization adaptive smart loss function may accurately predict grafting artifacts. Sharp high frequency artifacts may be removed by the DL module using the fully sampled k-space data and the grafted data. These artifacts may not be removed by a DL module trained using only the grafted data.

[0065] The loss function used in conventional DL modules assigns fixed weights to the regularization terms at the beginning of the calculation. However, the intelligent loss function according to one aspect of the present disclosure is dynamic and is constructed to change the weight assigned to the regularization term based on the relevance of the regularization term as training progresses. In one example, the regularization weight of the loss term can be modulated based on the training loss after each epoch. A transfer function can be adopted by the artifact prediction network to determine the regularization weights of the loss function components. The loss function components may include structural similarity index measure (SSIM) loss, perceptual (e.g., latent feature-based) loss, mean absolute error (MAE) loss, which can be dynamically adjusted as training progresses to obtain better image quality.

[0066] In one example, a higher use of SSIM at the beginning can result in fine banding artifacts in MR images. However, using a dynamic regularization method according to one aspect of the present disclosure, the role of SSIM in the loss function can be controlled through training. As the prediction gets closer to the ground truth, more weight can be placed on SSIM. In addition, the weight of the regularization term after each epoch can be updated by a transfer function. The intelligent loss function calculation can be based on the following formula:

[0067]

[0068] The MAE can be calculated based on the predicted residuals, and the SSIM can be calculated between the ground truth and the artifact-corrected image. The weight of the regularization term α can be updated after each epoch based on the MAE and SSIM values. As the prediction gets closer to the ground truth, the weight of the SSIM loss can be increased. In addition, the user can decide how soon the weight of the SSIM starts to increase. According to the dynamic regularization method of the present disclosure, the weight of the regularization term (α) can be dynamically adjusted using the following formula:

[0069]

[0070] in is the ratio offset term, which determines the nature of the curve to obtain the corresponding r m The higher the R value, the greater the effect of SSIM.

[0071] The method (900) of the present disclosure not only accelerates MRI examinations, but also provides improved resolution for images. The method (900) can more efficiently collect high-resolution data for multiple comparisons. According to one aspect of the present disclosure, a deep learning (DL) module trained using dual input channels can have better artifact prediction capabilities than a DL module trained using single-channel images. The DL module can be trained using only grafted data (single input) and both reference data and grafted data (dual input) to analyze and predict the artifact removal efficiency of the DL module. It can be seen that when dual input is provided, the removal of artifacts by the DL module is improved. It can be seen that in the case of the dual input model, the visual information fidelity (VIF) and structural similarity index metric (SSIM) of the output are improved.

[0072] This written description uses examples to disclose the invention, including the best mode, and to enable any person skilled in the art to practice the invention, including making and using any computing system 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 be 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 accelerating a magnetic resonance imaging (MRI) examination, the method comprising: acquiring at least one fully sampled reference k-space data of the subject; acquiring a plurality of partial k-spaces of the subject; as well as grafting the plurality of partial k-spaces with the at least one fully sampled reference k-space data to generate grafted k-space data for accelerating the examination; The grafting of the partial k-space with the fully sampled k-space data is performed before reconstructing the grafted data based on a deep learning DL module.

2. The method of claim 1 , wherein the method comprises a subsequent scan comprising acquiring only the portion of k-space of the subject and grafting the portion of k-space with the fully sampled k-space data to generate the grafted k-space data for accelerated examination. 3 . The method of claim 1 , wherein grafting the partial k-space with the fully sampled reference k-space data comprises grafting missing structural information in the partial k-space from the at least one fully sampled reference k-space data.

4. The method of claim 1 , wherein before performing the deep learning (DL) module-based reconstruction on the grafted data, grafting the partial k-space with the fully sampled k-space data provides structural information to the DL module to correct grafting artifacts.

5. A method for accelerating magnetic resonance imaging (MRI) examinations based on deep learning (DL), the method comprising: acquiring at least one fully sampled reference k-space data of the subject; acquiring a plurality of partial k-spaces of the subject; grafting the partial k-space with the at least one fully sampled reference k-space data to generate grafted data for accelerating the examination; and A deep learning (DL) module is trained using the grafted data and the fully sampled reference k-space data to remove grafting artifacts. The method of claim 5 , wherein the DL module comprises an intelligent loss function.

7. The method of claim 6, wherein the smart loss function includes a plurality of regularization terms, and the smart loss function is configured to change weights assigned to the plurality of regularization terms based on correlations of the regularization terms during training of the DL module.

8. The method of claim 7, wherein the weight assigned to each of the regularization terms is dynamically modulated based on the training loss after each imaging epoch. 9 . The method of claim 8 , wherein dynamically modulating the regularization term comprises modulating a structural similarity index metric (SSIM) loss, a perceptual loss, and a mean absolute error (MAE) loss.

10. The method according to claim 6, wherein the intelligent loss function is defined as: in Indicates loss of intelligence; MAE is the mean absolute error; SSIM is the structural similarity index measure; as well as α is the weight of the regularization term; where α is updated after each epoch according to the MAE and SSIM values.

11. The method of claim 5, further comprising acquiring multiple MRI images including simultaneous multi-slice imaging readouts for accelerated examination.

12. A magnetic resonance imaging (MRI) system, comprising: at least one radio frequency (RF) body coil adapted to transmit radio frequency (RF) signals to and receive radio frequency (RF) signals from a subject; a transceiver module configured to digitize the RF signal received by the RF body coil; a control system configured to process the digitized signals and generate k-space data corresponding to an imaging volume of the subject, wherein the MRI system is configured to acquire at least one fully sampled reference k-space data of the subject and a plurality of partial k-spaces of the subject; as well as a computer processor configured to graft the portion of k-space of the subject with the fully sampled reference k-space data of the subject to generate grafted k-space data for accelerating an examination, The MRI system further comprises a deep learning DL module employed on a computer memory, wherein the deep learning DL module is trained using the at least one fully sampled reference k-space data and the grafted k-space data.

13. The MRI system of claim 12, wherein the DL module is an artifact prediction network including an intelligent loss function.

14. The MRI system of claim 13, wherein the smart loss function comprises a plurality of regularization terms, and the smart loss function is configured to change a weight assigned to each of the plurality of regularization terms based on a correlation of the regularization terms during training of the DL module.

15. The MRI system of claim 14, wherein the weight assigned to each of the regularization terms is dynamically modulated based on the training loss after each imaging epoch. 16 . The MRI system of claim 15 , wherein dynamically modulating the regularization term comprises modulating a structural similarity index metric (SSIM) loss, a perceptual loss, and a mean absolute error (MAE) loss.

17. The MRI system of claim 12, wherein the computer processor is configured to graft the partial k-space of the subject with the fully sampled reference k-space data of the subject before performing the DL module-based reconstruction on the grafted data.

18. The magnetic resonance imaging (MRI) system of claim 12, wherein the MRI system is configured to acquire a plurality of MRI images including simultaneous multi-slice imaging readouts for accelerated examination.

Images