Patent Application
20250054206
This disclosure relates to medical image reconstruction, such as reconstruction in magnetic resonance (MR) imaging.
Medical imaging, such as magnetic resonance (MR), computed tomography (CT), positron emission tomography (PET), or single photon emission computed tomography (SPECT), use reconstruction to estimate an image or real-space object from measurements. These scans may be time consuming. For example, MR imaging (MRI) is intrinsically slow, and numerous methods have been proposed to accelerate the MRI scan. One acceleration method is the under-sampling reconstruction technique (i.e., MR compressed sensing (CS)), where fewer samples are acquired in the MRI data space (k-space), and prior knowledge is used to restore the images in reconstruction. Parallel Imaging (PI) combined with CS techniques provides faster MR Imaging (MRI) scan times.
Current deep learning MRI reconstruction is formulated as a trainable unrolled optimization framework with several cascades of regularization networks and varying data consistency layers. However, MR images reconstructed from sub-sampled Cartesian data with integrated reference lines using deep learning techniques produce images with some unnatural structures referred to as banding artifacts. Banding is characterized by a streaking pattern precisely aligned with the phase-encoding direction. This banding is anisotropic and non-homogenous across the image and most visible in high-noise or low-contrast areas. Banding results from the signal subsampling process used during Cartesian accelerated MRI, whereby subsampling occurs in one spatial direction only with a fully sampled k-space center.
In one approach to deal with banding, an adversarial loss penalizes banding structures in training the regularization networks. However, this approach provides limited reduction of the appearance of banding without complete removal. Moreover, when combined with additional adversarial losses that target improving the image sharpness, the overall training becomes challenging with increased potential of introducing hallucinations.
By way of introduction, the preferred embodiments described below include methods, systems, instructions, and computer readable media for reconstruction in medical imaging, such as reconstruction in MR imaging. In reconstruction, the measurements from the scan are used in each iteration. By masking parts of the sub-sampled measurements (i.e., sub-sampling the acquired sub-sampled data) used in one or more iterations of reconstruction, banding is reduced or eliminated.
In a first aspect, a method is provided for reconstruction of a medical image in a medical imaging system. The medical imaging system scans a patient, resulting in measurements. An image processor reconstructs the medical image from the measurements. The reconstruction includes iterations. A mask of the measurements in applied in at least one of the iterations. The medical image is displayed.
In a second aspect, a method is provided for reconstruction in magnetic resonance imaging. A magnetic resonance imaging system scans a patient using parallel imaging with compressed sensing, resulting in k-space measurements having fully sampled center lines and sub-sampling of other lines. An image processor reconstructs a magnetic resonance image from the measurements. The reconstruction uses an unrolled iterative sequence of machine-learned networks. Each of the iterations of the sequence having an input for the k-space measurement. Different masks sub-sampling the fully sampled center lines of the k-space measurements are applied to the inputs for the iterations. The magnetic resonance image is displayed.
In a third aspect, a system is provided for reconstruction in medical imaging. A magnetic resonance scanner is configured to scan a region of a patient, the scan providing sparsely sampled k-space data. An image processor is configured to reconstruct a representation of the region from the sparsely sampled k-space data. The image processor is configured to reconstruct by a sequence of cascades in an unrolled iterative reconstruction. The image processor is configured to sub-sample the sparsely sampled k-space data from the scan differently for different cascades. A display is configured to display a magnetic resonance image of the region from the reconstructed representation.
Various other embodiments, enhancements, improvements, combinations, aspects, advantages, or approaches are summarized at the end of the detailed description in the illustrative embodiments.
The present invention is defined by the following claims, and nothing in this section should be taken as a limitation on those claims. Further aspects and advantages of the invention are discussed below in conjunction with the preferred embodiments and may be later claimed independently or in combination.
Banding artifacts may be related to the point spread function of the uniform sampling pattern with fully sampled center lines.
One naive solution would be not to acquire or drop the center lines. However, a fully-sampled k-space center has several advantages, including a better signal-to-noise ratio (SNR) and fewer aliasing artifacts at higher accelerations. Instead, the center fully-sampled k-space is uniformly sub-sampled. Different uniform sub-sampling of the center lines may be used, such as a set of masks with all possible offsets depending on the acceleration factor for the scanning. A sub-mask is selected for data consistency layers of different cascades or iterations in reconstruction to ensure a cover of the acquired k-space center while not adding additional computational overhead. With unrolled iterative reconstruction frameworks that end with less computationally expensive cascades (e.g., without a regularization network), all possible sub-masks can be utilized with results averaged for an SNR boost with slight computational overhead.
The system may use a machine-learned model in reconstruction. The machine-learned model is formed from one or more networks and/or other machine-learned architecture. For example, the machine-learned model is a deep learned neural network used for multiple iterations of the reconstruction. The machine-learned model is used in any aspect of reconstruction. In one embodiment, the machine-learned model is formed as a convolutional neural network for use as a regularizer or denoiser in the reconstruction. Image or object domain data is input, and image or object domain data with less artifact is output. The machine-learned model assists in compressed, parallel sensing, and/or other MR imaging for more rapid scanning of the patient with less artifacts. The reconstruction may also includes extrapolation. The remaining portions or stages of the reconstruction (e.g., Fourier transform and gradients in iterative optimization) are performed using reconstruction algorithms and/or other machine-learned networks included in the machine-learned model. In other embodiments, the machine-learned model replaces, at least in part, the Fourier transform so that k-space measurements are input, and image or object domain data is output. One or more parts of the reconstruction (e.g., gradient update/data consistency) uses k-space measurements. The k-space measurements are masked to reduce or eliminate banding artifacts.
The system is implemented by an MR scanner or system, a computer based on data obtained by MR scanning, a server, or another processor. MR scanning system 100 is only exemplary, and a variety of MR scanning systems can be used to collect the MR data. In the embodiment of
RF (radio frequency) module 20 provides RF pulse signals to RF coil(s) 18, which in response produces magnetic field pulses that rotate the spins of the protons in the imaged body of the patient 11 by ninety degrees, by one hundred and eighty degrees for so-called “spin echo” imaging, or by angles less than or equal to 90 degrees for so-called “gradient echo” imaging. Gradient and shim coil control module 16 in conjunction with RF module 20, as directed by central control unit 26, control slice-selection, phase-encoding, readout gradient magnetic fields, radio frequency transmission, and magnetic resonance signal detection, to acquire magnetic resonance signals representing planar slices of patient 11.
In response to applied RF pulse signals, the RF coil(s) 18 receives MR signals, i.e., signals from the excited protons within the body as they return to an equilibrium position established by the static and gradient magnetic fields. The MR signals are detected and processed by a detector within RF module 20 and k-space component processor unit 34 to provide an MR dataset to an image processor for processing into an image (i.e., for reconstruction in the object domain from the k-space data in the scan domain). In some embodiments, the image processor is located in or is the central control unit 26. In other embodiments, such as the one depicted in
A magnetic field generator (comprising coils 12, 14 and 18) generates a magnetic field for use in acquiring multiple individual frequency components corresponding to individual data elements in the storage array. The individual frequency components are successively acquired using a Cartesian acquisition strategy as the multiple individual frequency components are sequentially acquired during acquisition of an MR dataset representing an MR image. A storage processor in the k-space component processor unit 34 stores individual frequency components acquired using the magnetic field in corresponding individual data elements in the array. The row and/or column of corresponding individual data elements alternately increases and decreases as multiple sequential individual frequency components are acquired. The magnetic field acquires individual frequency components in an order corresponding to a sequence of substantially adjacent individual data elements in the array, and magnetic field gradient change between successively acquired frequency components is substantially minimized. The central control processor 26 is programmed to sample the MR signals according to a predetermined sampling pattern. Any MR scan sequence may be used, such as for T1, T2, or other MR parameter.
In one embodiment, a compressive sensing scan sequence is used, such as a parallel imaging combined with compressive sensing. The scan provides sparsely sampled k-space data. Any under sampling or sub-sampling pattern may be used. In one embodiment, the center lines (e.g., autocalibration region) of the Cartesian format scan are fully sampled. Any number of center lines may be fully sampled, depending on the acceleration factor of the scan. The other lines are sub-sampled or under sampled, such as sampling every 4th line.
The central control unit 26 also uses information stored in an internal database to process the detected MR signals in a coordinated manner to generate high quality images of a selected slice(s) of the body (e.g., using the image data processor) and adjusts other parameters of system 100. The stored information comprises predetermined pulse sequence and magnetic field gradient and strength data as well as data indicating timing, orientation, and spatial volume of gradient magnetic fields to be applied in imaging.
The central control unit 26 (i.e., controller) and/or processor 27 are an image processor that reconstructs a representation of the patient from the k-space data. The image processor is a general processor, digital signal processor, three-dimensional data processor, graphics processing unit, application specific integrated circuit, field programmable gate array, artificial intelligence processor, digital circuit, analog circuit, combinations thereof, and/or another now known or later developed device for reconstruction. The image processor is a single device, a plurality of devices, or a network. For more than one device, parallel or sequential division of processing may be used. Different devices making up the image processor may perform different functions, such as reconstructing by one device and volume rendering by another device. In one embodiment, the image processor is a control processor or other processor of the MR scanner 100. Other image processors of the MR scanner 100 or external to the MR scanner 100 may be used.
The image processor is configured by software, firmware, and/or hardware to reconstruct. The image processor operates pursuant to stored instructions to perform various acts described herein.
The image processor is configured to reconstruct a representation in an object domain. The object domain is an image space and corresponds to the spatial distribution of the patient. A planar area or volume representation is reconstructed as an image representing the patient. For example, pixels values representing tissue in an area or voxel values representing tissue distributed in a volume are generated.
The representation in the object domain is reconstructed from the scan data in the scan domain. The scan data is a set or frame of k-space data from a scan of the patient. The k-space measurements resulting from the scan sequence are transformed from the frequency domain to the spatial domain in reconstruction. In general, reconstruction is an iterative process, such as a minimization problem. This minimization can be expressed as:
where x is the target image to be reconstructed, and y is the raw k-space data. A is the MRI model to connect the image to MRI-space (k-space), which can involve a combination of an under-sampling matrix U, a Fourier transform F, and sensitivity maps S. T represents a sparsifying (shrinkage) transform. A is a regularization parameter. The first term of the right side of equation 1 represents the image (2D or 3D spatial distribution or representation) fit to the acquired data, and the second term of the right side is a term added for denoising by reduction of artifacts (e.g., aliasing) due to under sampling. The l1 norm is used to enforce sparsity in the transform domain. ∥Ax−y∥22 is the l2 norm of the variation of the under-sampled k-space data. Generally, the lp norm is
In some embodiments, the operator T is a wavelet transform. In other embodiments, the operator T is a finite difference operator in the case of Total Variation regularization.
The image processor may be configured to optimize as part of reconstruction, such as performing a sequence of iterations. A same machine-learned network may be used in each iteration. Alternatively, a machine-learned network is not used. In another approach, the image processor is configured to reconstruction by a sequence of cascades or iterations in an unrolled iterative reconstruction. Each iteration uses a separate model, such as separate machine-learned networks. In any of these approaches, each iteration uses the k-space measurements, such as in a data-consistency operation.
The image processor is configured to sub-sample the sparsely sampled k-space data from the scan differently for different cascades or iterations. A mask is applied to the k-space measurements, such as masking out one or more center lines. In one approach, the image processor is configured to generate different sub-sample patterns (masks) for the sub-sampling. The sub-sample patterns are based on an acceleration factor of the scan, such as generating four masks M1-4 where the acceleration factor (e.g., PAT4) provides for uniformly sub-sampling every fourth line. Each mask is formed by using a different offset 0-3 corresponding to the acceleration factor. Each offset shifts the sub-sampling by a line. For reconstruction, the image processor is configured to select (e.g., randomly) one of the different sub-sample patterns for each of the cascades or iterations. A preset pattern may instead be used matching different masks to different cascades or iterations. Each of the cascades includes a machine-learned network or function (operation) that has an input for the sub-sampled sparsely sampled k-space data.
In one approach, the sequence of cascades of the unrolled iterative reconstruction model includes first cascades with data consistency and regularization functions and includes at least a second cascade with the data consistency function and no regularization function. Different masks are selected for different ones of the first cascades. The image processor is configured to perform the second cascade multiple times with different ones of the sub-sampled (masked) sparsely sampled k-space data and average results from the performance of the second cascade.
The image processor may be configured to generate an MR image from the reconstructed representation. Where the representation is of an area, the values of the representation may be mapped to display values (e.g., scalar values to display color values) and/or formatted (e.g., interpolated to a display pixel grid). Alternatively, the output representation is of display values in the display format. Where the representation is of a volume, the image processor performs volume or surface rendering to render a two-dimensional image from the voxels of the volume. This two-dimensional image may be mapped and/or formatted for display as an MR image. Any MR image generation may be used so that the image represents the measured MR response from the patient. The image represents a region of the patient.
Generated images of the reconstructed representation for a given patient are presented on a display 40 of the operator interface. Computer 28 of the operator interface includes a graphical user interface (GUI) enabling user interaction with central control unit 26 and enables user modification of magnetic resonance imaging signals in substantially real time. Display processor 37 processes the magnetic resonance signals to provide image representative data for display on display 40, for example.
The display 40 is a CRT, LCD, plasma, projector, printer, or other display device. The display 40 is configured by loading an image to a display plane or buffer. The display 40 is configured to display the reconstructed MR image.
The method is performed by the system of
The method is performed in the order shown (top to bottom or numerical) or other orders. For example, the masks may be generated prior to the scan of the patient. Additional, different, or fewer acts may be provided. For example, a preset, default, or user input settings are used to configure the scanning prior to act 200. As another example, the image is stored in a memory (e.g., computerized patient medical record) or transmitted over a computer network instead of or in addition to the display of act 240.
In act 200, the medical system scans a patient. For example, an MR scanner or another MR system scans the patient with an MR compressed (e.g., under sampling) or another MR sequence. For example, a compressed sensing in parallel imaging scan is performed. The amount of under sampling is based on the settings, such as the acceleration (acceleration factor). Based on the configuration of the MR scanner, a pulse sequence is created. The pulse sequence is transmitted from coils into the patient. The resulting responses are measured by receiving radio frequency signals at the same or different coils. The scanning results in k-space measurements as the scan data. In another example, a computed tomography scanner scans a patient by transmitting x-rays from different angles through the patient. The scanning results in detected projections for a given patent as the scan data.
In act 210 of
The set includes any number of masks, such as one or more. In one approach, the number of the different masks is based on the acceleration factor. For example, the acceleration factor is 4, so 4 masks are generated. Other ratios, such as 2 masks for an acceleration factor of 4, may be used. Fractional acceleration factors may be rounded to the nearest even integer or handled in a different way.
In one embodiment, the acceleration factor provides the number of skipped lines in under sampling. For example, an acceleration factor of 4 provides for every fourth line being sampled (skip 3 lines). This under sampling may be for some of the lines and not others, such as not the center lines. The result is 3 center lines fully sampled. The masks select different ones or combinations of the center lines to use. For example, an offset is provided. The offset is a beginning line to pass with the next X (e.g., 3) lines to not select (i.e., mask). The offset is 0-3 so that each of the four masks sub-samples to provide a different or no center line. Different offsets are used where the offset is based on the acceleration factor. The number of offsets is provided by the acceleration factor for the scanning.
In one embodiment, the center fully-sampled k-space is uniformly sub-sampled with all possible offsets (depending on the acceleration PAT factor). This creates corresponding sub-masks (
The masks, as generated, are fed to the network as an input. The masks and/or measurements as masked are used in reconstruction.
In act 220 of
The image processor reconstructs a representation of the patient from the scan data. For MR reconstruction, the k-space data is Fourier transformed into scalar values representing different spatial locations, such as spatial locations representing a plane through or volume of a region in the patient. Scalar pixel or voxel values are reconstructed as the MR image. The spatial distribution of measurements in object or image space is formed. This spatial distribution represents the patient. The reconstruction forms an object or patient representation in two- or three-dimensional space from the measurements in a scan space (e.g., k-space). For example, a three-dimensional distribution of voxels representing a volume of the patient is reconstructed from the measurements.
In one approach, the reconstruction uses optimization based on functions. In another approach, one or more functions in the reconstruction are replaced by or performed by a machine-learned model, such as a neural network. In yet another approach, an unrolled iterative reconstruction is used. Each iteration is formed separately, such as having one or more parameters set differently. Where machine-learning is used, each iteration may use a different machine-learned network. In any of these approaches, the measurements used in one or more iterations are masked.
In one embodiment, the machine-learned network of the machine-learned model implements a regularizer. The reconstruction is performed iteratively with gradients (gradient update or data consistency check), a Fourier transform, and the regularizer. The regularizer receives image space information from the Fourier transform or after the gradient operation and outputs denoised image space information. The machine-learned network may be an image-to-image network with DenseNet blocks or have another architecture, such as a CNN.
In another embodiment, the machine-learned model includes an extrapolation function or algorithm with a machine-learned parameter. For example, the value for the weight applied in a Nesterov or Polyak heavy ball extrapolation was machine learned. The extrapolation operates on an input image data and outputs to the gradient update. The learned value of the parameter is used in the extrapolation. For example, the machine-learned parameter is a weight for a difference between current and previous image values. The extrapolation provides momentum to the gradient descent.
The reconstruction is iterative. Each iteration or cascade determines an updated image object from an input image object, with the gradient operation comparing fit with the measurements. For example, an unrolled iterative reconstruction is performed. Different machine-learned networks and/or extrapolation weights are used for the different iterations. Some iterations may not include regularization. For example, an initial sequence of iterations includes extrapolation with learned weights and does not include regularization, and a subsequent sequence of iterations includes regularization with or without extrapolation with learned weights. After the last iteration, the output representation by the regularizer or gradient update is provided for imaging or the medical record of the patient.
The network 800 also includes bias field correction (BCF). The input bias field correction map, B, is applied to correct for biases in the image object output after the iterations. Other filtering and/or operations for reconstruction and/or post-reconstruction may be provided.
The iterations 822 with regularization includes the regularizer operating on the output of the gradient update. The regularizer is a machine-learned network. In one embodiment, deep learning is used to train a convolutional neural network as the regularizer. Machine learning is an offline training phase where the goal is to identify an optimal set of values of learnable parameters of the model that can be applied to many different inputs (i.e., image domain data after gradient calculation in the optimization or minimization of the reconstruction). These machine-learned parameters can subsequently be used during clinical operation to rapidly reconstruct images. Once learned, the machine-learned model is used in an online processing phase in which MR scan data y (e.g., k-space measurements) for patients is input and the reconstructed representations for the patients are output based on the model values learned during the training phase. Other functions (e.g., extrapolation and/or gradient update) may use machine-learned models or networks. Coil sensitivity maps C for the one or more coils and/or other information representing the model of the scanner may be input for use in each, one, or more iterations.
During application to one or more different patients and corresponding different measurements, the same learned weights or values for machine-learned network for each iteration are used. The model and values for the learnable parameters are not changed from one patient to the next, at least over a given time (e.g., weeks, months, or years) or given number of uses (e.g., tens or hundreds). These fixed values and corresponding fixed model are applied sequentially and/or by different processors to scan data for different patients. The model may be updated, such as retrained, or replaced but does not learn new values as part of application for a given patient.
The model has an architecture. This structure defines the learnable variables and the relationships between the variables. In one embodiment for the regularization, a neural network is used, but other networks may be used. For example, a convolutional neural network (CNN) is used. Any number of layers and nodes within layers may be used. A DenseNet, U-Net, encoder-decoder, Deep Iterative Down-Up CNN, and/or another network may be used. In one embodiment, an image-to-image neural network (spatial distribution input and spatial distribution output) is used. The image-to-image neural network may include convolution layers or be a CNN. Some of the network may include dense blocks (i.e., multiple layers in sequence outputting to the next layer as well as the final layer in the dense block). Any know known or later developed neural network may be used.
Deep learning is used to train the model for each iteration where machine learning is used. The training learns both the features of the input data and the conversion of those features to the desired output (i.e., denoised or regularized image domain data). Backpropagation, RMSprop, ADAM, or another optimization is used in learning the values of the learnable parameters. Where the training is supervised, the differences (e.g., L1, L2, or mean square error) between the estimated output and the ground truth output are minimized. Where a discriminator is used in training, the ground truth is not needed. Instead, the discriminator determines whether the output is real or estimated as an objective function for feedback in the optimization. The characteristic is one that likely distinguishes between good and bad output by examining the output rather than by comparison to a known output for that sample. Joint training (e.g., semi-supervised) may be used.
The training uses multiple samples of input sets, such as object domain data representing patients after Fourier transform and/or gradient calculation. The measurements for these samples are generated by scanning a patient and/or phantom with different settings or sequences, scanning different patients and/or phantoms with the same or different settings or sequences, and/or simulating MR scanning with an MR scanner model. By using many samples, the model is trained given a range of possible inputs. The samples are used in deep learning to determine the values of the learnable variables (e.g., values for convolution kernels) that produce outputs with minimized cost function and/or maximized likelihood of being a good representation (i.e., discriminator cannot tell the difference) across the variance of the different samples. Masking of the measurements may or may not be used for training.
In one embodiment, the image processor is configured to reconstruct with the CNN as trained being used as a regularizer in the reconstruction. The iterative reconstruction may be unrolled where a given number of iterations is used. The same CNN is used for each iteration. Alternatively, a different CNN is provided for each iteration, whether a different architecture or same architecture but with different values for one or more of the learnable parameters of the CNN. Different CNNs are trained for different iterations in the reconstruction. Each CNN may have the same architecture, but each is separately learned so that different values of the learnable parameters may be provided for different iterations of the reconstruction.
Once trained, the machine-learned model is used for reconstruction of a spatial representation from input k-space measurements for a patient. Some of or all the iterations use k-space measurements as an input. The k-space measurements may be input to a machine-learned model or to a function/operation of the reconstruction (e.g., gradient update). The measurements, y, used or input are masked to reduce banding artifacts. For example, rather than using all the center lines acquired in the compressed sensing scan, only center lines passing through or filtered by a mask are used.
In application of the already trained network, the reconstruction process is followed. The machine-learned model is used in the reconstruction. For example, extrapolation is performed in every or some iterations using weights learned for those iterations, and regularization is performed in every or only some iterations using the deep learned network (e.g., CNN). In response to the input for a given patient, a patient specific image is reconstructed. The machine-learned model outputs the image as pixels, voxels, and/or a display formatted image in response to the input. The learned values and network architecture, with any algorithms (e.g., extrapolation and gradient update) determine the output from the input.
Different or the same masks may be used for different iterations. The different masks may be applied to the k-space measurements of the center lines for input to respective different iterations. The masked measurements may be input to a machine-learned network, such as a machine-learned network to implement the gradient update. Alternatively, the masked measurements are input to the gradient update (data consistency) function or operation without being input to a machine-learned network. Each or at least some of the iterations in the unrolled iterative sequence use measurements as masked.
In the example of
In act 230, the image processor selects masks to be used for one or more iterations (cascades). The mask to use for any iteration may be set. A pattern may be applied to mask selection for each sequence 810, 820, 830. In one approach, the masks applied for the iterations of one or more sequences (e.g., 810 and 820) is random. One of the different masks is randomly selected for each iteration in the unrolled iterative sequence. For example, the sub-mask M3 is randomly selected for data consistency layers of different cascades (iterations) to ensure a cover of the acquired k-space center while not adding additional computational overhead. A sub-mask will be randomly selected from such fixed length set, allowing for a unified framework handling different acceleration factors. In the example of
For iterations without regularization, another approach may be used. For a given iteration, the operation may be performed multiple times with different masks. In the example of
In the example of
The image processor reconstructs a magnetic resonance image from the measurements. The reconstruction uses an unrolled iterative sequence of cascades, which may include machine-learned networks. Each or some of the cascades have an input for the k-space measurements. Different masks sub-sampling the fully sampled center lines of the k-space measurements are applied to the inputs for the cascades.
The output of the machine-learned network is a two-dimensional distribution of pixels representing an area of the patient and/or a three-dimensional distribution of voxels representing a volume of the patient. The output from the last iteration may be used as the output representation of the patient.
Other processing may be performed on the output representation or reconstruction, such as spatial filtering, color mapping, and/or display formatting. In one embodiment, the reconstruction outputs voxels or scalar values for a volume spatial distribution as the medical image. Volume rendering is performed to generate a display image as a further display image. In alternative embodiments, the machine-learned network outputs the display image directly in response to the input
In act 240 of
The displayed image may represent a planar region or area in the patient. Alternatively, or additionally, the displayed image is a volume or surface rendering from voxels (three-dimensional distribution) to the two-dimensional display.
Banding artifacts are reduced or entirely removed (not present) in the medical image. The banding artifacts are not present by directly adapting the input measurements to different network layers without requiring customized adversarial loss functions. The proposed generic solution can be utilized in various k-space-to-image networks, including 2D, 2D+t, 3D, 3D, and 3D+t applications with Cartesian sampling.
Below are illustrative embodiments:
Illustrative embodiment 1. A method for reconstruction of a medical image in an medical imaging system, the method comprising: scanning, by the medical imaging system, a patient, the scanning resulting in measurements; reconstructing, by an image processor, the medical image from the measurements, the reconstructing including iterations, wherein a mask of the measurements in applied in at least one of the iterations; and displaying the medical image.
Illustrative embodiment 2. The method of illustrative embodiment 1 wherein scanning comprises scanning with the medical imaging system being a magnetic resonance (MR) scanner using compressed sensing and the measurements being k-space measurements.
Illustrative embodiment 3. The method of illustrative embodiment 2 wherein the measurements include a first center line sampling, and wherein the mask sub-samples the first center line sampling, resulting in a second center line sampling, the measurements from the second center line sampling used in the one of the iterations.
Illustrative embodiment 4. The method of illustrative embodiment 3 wherein the first center line sampling comprises a full sampling of center lines.
Illustrative embodiment 5. The method of any of illustrative embodiments 1-4 wherein reconstructing comprises reconstructing a three-dimensional distribution of voxels representing a volume of the patient, and wherein displaying comprises volume or surface rendering from the voxels to a two-dimensional display.
Illustrative embodiment 6. The method of any of illustrative embodiments 1-5 wherein reconstructing comprises reconstructing as an unrolled iterative reconstruction, each iteration using a different machine-learned network.
Illustrative embodiment 7. The method of illustrative embodiment 6 wherein reconstructing comprises reconstructing with the different machine-learned networks, the mask being for one of the iterations and corresponding inputs to that one iteration, other masks being used for other of the iterations.
Illustrative embodiment 8. The method of illustrative embodiments 7 further comprising generating the masks based on different offsets, the different offsets based on an acceleration factor for the scanning, the mask for the one iteration selected randomly from the masks.
Illustrative embodiment 9. The method of any of illustrative embodiments 7-8 wherein the other masks are used for other of the iterations.
Illustrative embodiment 10. The method of any of illustrative embodiments 7-9 wherein all of the masks are used for each iteration where the corresponding iteration is free of regularization.
Illustrative embodiment 11. A method for reconstruction in magnetic resonance imaging, the method comprising: scanning, by a magnetic resonance imaging system using parallel imaging with compressed sensing, a patient, the scanning resulting in k-space measurements having fully sampled center lines and sub-sampling of other lines; reconstructing, by an image processor, a magnetic resonance image from the measurements, the reconstructing using an unrolled iterative sequence of machine-learned networks, each of the iterations of the sequence having an input for the k-space measurement, wherein different masks sub-sampling the fully sampled center lines of the k-space measurements are applied to the inputs for the iterations; and displaying the magnetic resonance image.
Illustrative embodiment 12. The method of illustrative embodiment 11 wherein scanning comprises scanning with an acceleration factor; and further comprising: generating a set of the different masks with different offsets, a number of the different masks based on the acceleration factor; and randomly selecting one of the different masks for each of at least some of the iterations of the unrolled iterative sequence.
Illustrative embodiment 13. The method of any of illustrative embodiments 11-12 wherein a first part the unrolled iterative sequence includes gradient update where each of the gradient updates of the first part have inputs of the k-space measurements masked by ones of the different masks and a second part of the unrolled iterative sequence does not include regularization where multiple of the masks are applied for one of the gradient updates of the second part.
Illustrative embodiment 14. The method of illustrative embodiment 13 wherein results derived from the one gradient update in response to inputs of the k-space measurements from each of the multiple masks are averaged.
Illustrative embodiment 15. The method of any of illustrative embodiments 11-14 wherein reconstructing comprises applying the different masks to the k-space measurements of the center lines for input to the respective iteration.
Illustrative embodiment 16. A system for reconstruction in medical imaging, the system comprising: a magnetic resonance scanner configured to scan a region of a patient, the scan providing sparsely sampled k-space data; an image processor configured to reconstruct a representation of the region from the sparsely sampled k-space data, the image processor configured to reconstruct by a sequence of cascades in an unrolled iterative reconstruction, wherein the image processor is configured to sub-sample the sparsely sampled k-space data from the scan differently for different cascades; and a display configured to display a magnetic resonance image of the region from the reconstructed representation.
Illustrative embodiment 17. The system of illustrative embodiment 16 wherein the magnetic resonance scanner is configured to scan with parallel imaging combined with compressed sensing along a Cartesian format where the provided sparsely sampled k-space data includes fully sampled center lines, and wherein the sub-sampling is of the fully sampled center lines.
Illustrative embodiment 18. The system of any of illustrative embodiments 16-17 wherein the image processor is configured to generate different sub-sample patterns for the sub-sampling based on an acceleration factor of the scan and configured to randomly select one of the different sub-sample patterns for each of the cascades.
Illustrative embodiment 19. The system of any of illustrative embodiments 16-18 wherein the sequence of cascades includes first cascades with data consistency and regularization functions and includes at least a second cascade with the data consistency function and no regularization function, wherein the image processor is configured to perform the second cascade multiple times with different ones of the sub-sampled sparsely sampled k-space data and average results from the performance of the second cascade.
Illustrative embodiment 20. The system of any of illustrative embodiments 16-19 wherein each of the cascades includes an input for the sub-sampled sparsely sampled k-space data.
Although the subject matter has been described in terms of exemplary embodiments, it is not limited thereto. Rather, the appended claims should be construed broadly, to include other variants and embodiments, which can be made by those skilled in the art.
