System and method for background noise reduction in magnetic resonance images

Abstract

The present invention is entitled "System and method for background noise reduction in magnetic resonance images". The present invention provides methods and systems for reducing background noise in magnetic resonance (MR) images and parametric maps visualization using segmentation and intensity thresholding. Example methods include: segmenting an MR image or a parametric map into a foreground and a background, the foreground including a region of an anatomical structure of interest, the background being outside the region of the anatomical structure of interest; applying an intensity threshold to the background of the MR image or the parametric map, and not applying an intensity threshold to the foreground of the MR image or the parametric map to produce a noise-reduced MR image or a noise-reduced parametric map; and displaying the noise-reduced MR image or the noise-reduced parametric map via a display device.

Description

Technical Field

[0001] Embodiments of the subject matter disclosed herein relate to magnetic resonance imaging (MRI), and more particularly, to background noise reduction in visualized MR images and parametric maps using segmentation and thresholding. Background Art

[0002] Magnetic resonance imaging (MRI) is a medical imaging modality that can create images of the inside of the human body without the use of X-rays or other ionizing radiation. An MRI system includes a superconducting magnet to produce a strong and uniform static magnetic field, B0. When an imaging subject is placed in the magnetic field, B0, the nuclear spins associated with the hydrogen nuclei in the imaging subject become polarized so that the magnetic moments associated with these spins are preferentially aligned in the direction of the magnetic field, B0, resulting in a small net magnetization along that axis. The hydrogen nuclei are excited by a radio frequency signal at or near the resonant frequency of the hydrogen nuclei, which adds energy to the nuclear spin system. When the nuclear spin relaxes back to its resting energy state, it releases the absorbed energy in the form of an RF signal. The RF signal (or MR signal) is detected by one or more RF coil assemblies and converted into an image using a reconstruction algorithm.

[0003] In addition to the contrast achieved by conventional anatomical MRI (e.g., T1 and T2 weighted contrast), unique contrast mechanisms (such as diffusion-weighted imaging (DWI), perfusion-weighted imaging (PWI), and functional MRI (fMRI)) have revolutionized the detection of pathological conditions, providing deep insights into physiological mechanisms and pathological conditions. Low signal-to-noise ratio (SNR) has been a major limitation of these contrast mechanisms. For example, background noise (signal from regions outside the anatomical structure of interest, typically including air) deteriorates the quality of magnetic resonance (MR) images and corresponding parametric maps (such as apparent diffusion coefficient (ADC) maps). Conventional visualization and post-processing methods rely on setting a fixed intensity threshold for suppressing background noise. Pixels or voxels in MR images or parametric maps with intensities below the fixed intensity threshold are masked (their intensities are set to zero), while pixels / voxels with intensities greater than the fixed intensity threshold remain unmasked. However, this may result in loss of information of the imaged anatomical region, since the MR image of the anatomical structure may include pixels with intensity below the threshold, and thus masking these pixels may create holes in the imaged anatomical structure, thereby reducing the MR image quality.

[0004] Improved background suppression techniques for visualizing MR images / maps are generally desired. Summary of the invention

[0005] The present disclosure at least partially solves the above-mentioned problems. In one embodiment, a method for visualizing a magnetic resonance (MR) image or an MR parametric map includes: segmenting the MR image or the parametric map into a foreground and a background, the foreground including a region of an anatomical structure of interest, and the background outside the region of the anatomical structure of interest; applying an intensity threshold to the background of the MR image or the parametric map, and not applying the intensity threshold to the foreground of the MR image or the parametric map to produce a noise-reduced MR image or a noise-reduced parametric map; and displaying the noise-reduced MR image or the noise-reduced parametric map via a display device.

[0006] In this way, background noise can be selectively reduced by first segmenting the MR image / map into a foreground (the region of the anatomical structure of interest in the MR image / map) and a background (outside the region of the anatomical structure of interest). In some embodiments, segmentation is performed by a neural network that has been trained to classify each pixel / voxel of the MR image / map as belonging to the background or the foreground. An intensity threshold is then selectively applied to the background rather than the foreground. This can achieve greater background noise reduction while reducing or eliminating holes in the anatomical region of the MR image, because even if the region of the foreground includes pixels / voxels with intensities below the intensity threshold, the foreground pixels / voxels can remain unmasked because the intensity threshold is only applied to the background pixels / voxels. Therefore, a higher intensity threshold than would otherwise be feasible can be applied to the background, thereby achieving greater background noise reduction without losing information from the anatomical region of interest in the MR image / map.

[0007] The above advantages and other advantages and features of the present specification will be apparent from the following detailed description when taken alone or in conjunction with the accompanying drawings. It should be understood that the above overview is provided to introduce some concepts that are further described in the detailed description in a simplified form. This is not meant to identify key or essential features of the claimed subject matter, the scope of which is uniquely defined by the claims that follow the detailed description. Furthermore, the claimed subject matter is not limited to embodiments that address any disadvantages noted above or in any part of this disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The present disclosure will be better understood by reading the following description of non-limiting embodiments with reference to the accompanying drawings, in which:

[0009] Figure 1 A block diagram of an exemplary embodiment of an MRI system is shown.

[0010] Figure 2 Exemplary embodiments of image processing systems for processing MR images and / or parametric maps are shown.

[0011] Figure 3 Exemplary embodiments of convolutional neural network architectures for segmenting MR images and / or parametric maps are shown.

[0012] Figure 4 A high-level flow chart of a method for reducing background noise in an MR image and / or parametric map according to an exemplary embodiment is shown.

[0013] Figure 5 A flow chart of an example method for determining an intensity threshold is shown.

[0014] Figure 6 A flow chart of an example method for reducing background noise in an MR image and / or a parametric map is shown according to another example embodiment.

[0015] Figure 7 The background and foreground of an exemplary MR image of the abdomen are shown.

[0016] Figure 8 The background and foreground of an exemplary MR image of the brain are shown.

[0017] Fig. 9 A comparison between MR images of the abdomen with background suppression according to various methods is shown.

[0018] Fig.10 Comparisons between MR images of the brain with background suppression according to various methods are shown.

[0019] Fig.11 Shown is a side-by-side comparison between MR images of the abdomen with and without background suppression in a visualization application.

[0020] Fig.12 Shown is a side-by-side comparison between MR images of the brain with and without background suppression in a visualization application.

[0021] Fig.13 A graphical user interface including an intensity threshold control interface is shown according to an exemplary embodiment. DETAILED DESCRIPTION

[0022] Low signal-to-noise ratio (SNR) has become a major limitation of some MR contrast mechanisms (such as fMRI, PWI, DWI) and corresponding parametric maps (such as ADC maps). For example, background noise (signals from regions outside the anatomical structure of interest, usually including air) can deteriorate the quality of magnetic resonance (MR) images and parametric maps. Conventional visualization / post-processing methods rely on setting a fixed intensity threshold for suppressing background noise, where pixels / voxels in the MR image / map with intensities below the fixed intensity threshold are masked (their intensities are set to zero), while pixels / voxels with intensities greater than the fixed intensity threshold remain unmasked.

[0023] There are several problems with this method. In one example, the fixed threshold may be suboptimal for many images, and the user may need to manually adjust the intensity threshold to obtain the desired visualization of the MR image, thereby inconveniencing the user. For example, the area of ​​the anatomical structure that the user wishes to visualize may include anatomical regions whose intensity is below the fixed intensity threshold, and therefore these areas may be masked by applying the fixed intensity threshold to the MR image, thereby forming "holes" in the anatomical structure. The user can manually adjust the intensity threshold to eliminate these "holes" and visualize all areas of the anatomical structure. However, this can significantly increase the visualization of background noise and weaken the contrast in the MR image.

[0024] Visualization applications use a "windowing" approach to visualize medical images captured with higher precision or a higher dynamic range of intensity values ​​than standard images. The approach maps the intensity intervals of interest to the dynamic range of the display. These applications set the range of the intervals used for the mapping and the center of the interval. In some cases, the interval range and interval center are not optimal due to background noise.

[0025] The present disclosure at least partially solves the above-mentioned problems. Figure 2 The MR image processing system 200 shown receives Figure 1 The MR image processing system 200 may be configured to receive an MR image and / or a parameter map generated by the magnetic resonance imaging (MRI) system 10 and to perform one or more of the methods disclosed herein for reducing background noise in the received MR image / map. The MR image processing system 200 may be incorporated into the MRI system 10 or communicatively coupled to the MRI system 10, or may be configured to receive the MR image / map from a storage device storing the MR image / map. The MR image processing system 200 may perform Figure 4 The method 400 for reducing background noise is shown. Figure 5 The method 500 for determining a threshold strength and / or Figure 6 A method 600 for suppressing background noise in MR images is shown.

[0026] In some embodiments, the MR image processing system 200 can be configured to use a trained neural network (such as Figure 3 A convolutional neural network (CNN) as shown in the embodiment of the present invention segments an MR image / map into foreground and background. As used herein, "foreground" refers to a region of an anatomical structure of interest in an MR image or parametric map, while "background" refers to regions outside of the anatomical structure of interest. The neural network has been trained to classify each of a plurality of pixels / voxels of an MR image / map as belonging to the foreground or background. An intensity threshold is then selectively applied to the background of the MR image / map rather than the foreground, thereby mitigating information loss from anatomical features in the MR image / map while enhancing background noise reduction. In some embodiments, the MR image / map is segmented into foreground and background according to the foreground. Figure 5 The illustrated method 500 determines an intensity threshold for background pixels / voxels to be selectively applied to an MR image / map.

[0027] Examples of foreground and background in MR images are shown in Figure 7 (Abdominal MR images) and Figure 8 (Brain MR image) is shown. Fig. 9 and Fig.10 Comparisons between segmentation maps generated by hand (ie, ground truth masks) and segmentation maps generated by a trained neural network according to embodiments of the present disclosure are shown for abdominal and brain MR images, respectively. Fig. 9 and Fig.10 Also shown are MR images (abdomen or brain) processed with a fixed intensity threshold compared to MR images processed according to the noise reduction methods disclosed herein. Fig.11 and Fig.12 It is shown how the contrast and brightness of an MR image are enhanced in an MR image having background noise reduced according to an embodiment of the present disclosure, compared to an MR image having background noise reduced according to a conventional fixed threshold method. Fig.13 An example graphical user interface 1300 is shown in which an intensity threshold control interface is shown, enabling a user to select a background intensity threshold.

[0028] In this way, the background noise of the MR image / map can be reduced to a greater extent than can be achieved using a fixed intensity threshold without losing anatomical detail in the foreground of the MR image. In addition, background detail can be restored by adjusting the background intensity threshold via the intensity threshold control interface.

[0029] Figure 1An MRI system 10 is shown, which includes a static magnetic field magnet unit 12, a gradient coil unit 13, an RF coil unit 14, an RF body or volume coil unit 15, a transmit / receive (T / R) switch 20, an RF driver unit 22, a gradient coil driver unit 23, a data acquisition unit 24, a controller unit 25, a patient couch or bed 26, an image processing system 31, a user input device 32, and a display device 33. In some embodiments, the RF coil unit 14 is a surface coil, which is a local coil that is generally placed near an anatomical structure of interest of an object 16. Here, the RF body coil unit 15 is a transmission coil that transmits RF signals, and the local surface RF coil unit 14 receives MR signals. Therefore, the transmission body coil (e.g., the RF body coil unit 15) and the surface receiving coil (e.g., the RF coil unit 14) are independent but electromagnetically coupled components. The MRI system 10 transmits an electromagnetic pulse signal to an object 16 placed in an imaging space 18, where a static magnetic field is formed to perform scanning to obtain magnetic resonance signals from the object 16. One or more MR images of the subject 16 may be reconstructed based on the magnetic resonance signals thus obtained by scanning. However, the scanning space typically includes non-anatomical features / structures, such as air, that may generate background signals that may obscure anatomical features / structures of interest.

[0030] The static magnetic field magnet unit 12 includes, for example, an annular superconducting magnet installed in an annular vacuum container. The magnet defines a cylindrical space around the object 16 and generates a constant main static magnetic field B0.

[0031] The MRI system 10 further includes a gradient coil unit 13, which forms a gradient magnetic field in an imaging space 18 so as to provide three-dimensional position information for a magnetic resonance signal received by an RF coil array. The gradient coil unit 13 includes three gradient coil systems, each of which generates a gradient magnetic field along one of three spatial axes perpendicular to each other, and generates a gradient field in each of a frequency encoding direction, a phase encoding direction, and a slice selection direction according to an imaging condition. More specifically, the gradient coil unit 13 applies a gradient field in a slice selection direction (or a scanning direction) of the object 16 to select a slice; and the RF body coil unit 15 or the local RF coil array can transmit an RF pulse to the selected slice of the object 16. The gradient coil unit 13 also applies a gradient field in a phase encoding direction of the object 16 to phase encode a magnetic resonance signal from a slice excited by an RF pulse. Then the gradient coil unit 13 applies a gradient field in a frequency encoding direction of the object 16 to frequency encode a magnetic resonance signal from a slice excited by an RF pulse.

[0032] The RF coil unit 14 is provided, for example, to surround a region to be imaged of the object 16. In some examples, the RF coil unit 14 may be referred to as a surface coil or a receiving coil. In a static magnetic field space or an imaging space 18 in which a static magnetic field B0 is formed by the static magnetic field magnet unit 12, the RF coil unit 15 transmits RF pulses as electromagnetic waves to the object 16 based on a control signal from the controller unit 25, and thereby generates a high-frequency magnetic field B1. This excites the proton spins in the slice of the object 16 to be imaged. The RF coil unit 14 receives the electromagnetic waves generated when the proton spins thus excited in the slice of the object 16 to be imaged return to be aligned with the initial magnetization vector as magnetic resonance signals. In some embodiments, the RF coil unit 14 may transmit RF pulses and receive MR signals. In other embodiments, the RF coil unit 14 may be used only to receive MR signals, and not to transmit RF pulses.

[0033] The RF body coil unit 15 is arranged, for example, to surround the imaging space 18, and generates RF magnetic field pulses orthogonal to the main magnetic field B0 generated by the static magnetic field magnet unit 12 in the imaging space 18 to excite the nucleus. The RF coil unit 14 can be disconnected from the MRI system 10 and replaced with another RF coil unit, while the RF body coil unit 15 is fixedly attached and connected to the MRI system 10. In addition, although local coils such as the RF coil unit 14 can transmit or receive signals only from a local area of ​​the object 16, the RF body coil unit 15 generally has a larger coverage area. For example, the RF body coil unit 15 can be used to transmit or receive signals to the whole body of the object 16. The use of a receive-only local coil and a transmission body coil provides uniform RF excitation and good image uniformity at the expense of high RF power deposited in the object. For a transmit-receive local coil, the local coil provides RF excitation to the region of interest and receives MR signals, thereby reducing the RF power deposited in the object. It should be understood that the specific use of the RF coil unit 14 and / or the RF body coil unit 15 depends on the imaging application.

[0034] When operating in the receive mode, the T / R switch 20 may selectively electrically connect the RF body coil unit 15 to the data acquisition unit 24, and when operating in the transmit mode, the T / R switch 20 may selectively electrically connect the RF body coil unit 15 to the RF driver unit 22. Similarly, when the RF coil unit 14 operates in the receive mode, the T / R switch 20 may selectively electrically connect the RF coil unit 14 to the data acquisition unit 24, and when operating in the transmit mode, the T / R switch 20 may selectively electrically connect the RF coil unit 14 to the RF driver unit 22. When both the RF coil unit 14 and the RF body coil unit 15 are used for a single scan, for example, if the RF coil unit 14 is configured to receive MR signals and the RF body coil unit 15 is configured to transmit RF signals, the T / R switch 20 may direct the control signal from the RF driver unit 22 to the RF body coil unit 15 while directing the received MR signal from the RF coil unit 14 to the data acquisition unit 24. The coils of the RF body coil unit 15 may be configured to operate in a transmit-only mode or a transmit-receive mode. The coils of the local RF coil unit 14 may be configured to operate in a transmit-receive mode or a receive-only mode.

[0035] The RF driver unit 22 includes a gate modulator (not shown), an RF power amplifier (not shown), and an RF oscillator (not shown), which are used to drive an RF coil (e.g., the RF coil unit 15) and form a high-frequency magnetic field in the imaging space 18. The RF driver unit 22 modulates the RF signal received from the RF oscillator into a signal of a predetermined timing with a predetermined envelope based on a control signal from the controller unit 25 and using the gate modulator. The RF signal modulated by the gate modulator is amplified by the RF power amplifier and then output to the RF coil unit 15.

[0036] The gradient coil driver unit 23 drives the gradient coil unit 13 based on a control signal from the controller unit 25 and thereby generates a gradient magnetic field in the imaging space 18. The gradient coil driver unit 23 includes three driver circuit systems (not shown) corresponding to the three gradient coil systems included in the gradient coil unit 13.

[0037] The data acquisition unit 24 includes a preamplifier (not shown), a phase detector (not shown), and an analog / digital converter (not shown) for acquiring a magnetic resonance signal received by the RF coil unit 14. In the data acquisition unit 24, the phase detector phase uses the output of the RF oscillator from the RF driver unit 22 as a reference signal to detect the magnetic resonance signal received from the RF coil unit 14 and amplified by the preamplifier, and outputs the phase-detected analog magnetic resonance signal to the analog / digital converter to be converted into a digital signal. The digital signal thus obtained is output to the image processing system 31.

[0038] The MRI system 10 includes a couch 26 for placing the subject 16 thereon. By moving the couch 26 based on a control signal from the controller unit 25, the subject 16 can be moved inside and outside the imaging space 18.

[0039] The controller unit 25 includes a computer and a recording medium on which a program to be executed by the computer is recorded. The program, when executed by the computer, causes the various parts of the system to perform operations corresponding to a predetermined scan. The recording medium may include, for example, a ROM, a floppy disk, a hard disk, an optical disk, a magneto-optical disk, a CD-ROM, or a non-transitory memory card. The controller unit 25 is connected to the user input device 32 and processes an operation signal input to the user input device 32, and also controls the examination bed 26, the RF driver unit 22, the gradient coil driver unit 23, and the data acquisition unit 24 by outputting control signals to them. The controller unit 25 also controls the image processing system 31 and the display device 33 based on the operation signal received from the user input device 32 to obtain a desired image.

[0040] The user input device 32 includes a user input device such as a touch screen, a keyboard, and a mouse. The operator uses the user input device 32 to, for example, input such data as an imaging protocol and set an area to be executed for an imaging sequence. The data on the imaging protocol and the imaging sequence execution area are output to the controller unit 25.

[0041] The image processing system 31 includes a processor and a non-transitory memory on which machine executable instructions may be stored, wherein the machine executable instructions may enable the processor to perform one or more of the steps of one or more methods of the methods disclosed herein. The image processing system 31 may be connected to the controller unit 25 and may perform data processing based on control signals received from the controller unit 25 or the user input device 32. The image processing system 31 is also connected to the data acquisition unit 24 and generates spectral data by applying various image processing operations to the magnetic resonance signals output from the data acquisition unit 24.

[0042] The image processing system 31 can perform background noise reduction in the MR images generated by the MRI system 10. In one example, the image processing unit 31 can perform a method, such as method 400 or 600 (discussed in detail below), to segment the MR image into a foreground and a background using, for example, a neural network, and apply an intensity threshold to the background or the foreground. Thus, the image processing system 31 can generate a noise-reduced MR image, and the noise-reduced MR image can be displayed via the display device 33.

[0043] The display device 33 displays an image on a display screen of the display device based on a control signal received from the controller unit 25. The display device 33 displays, for example, a noise-reduced MR image generated by the image processing system 31. The display device 33 may include a graphical user interface, wherein a user may interact with / input / change one or more data fields via the user input device 32. The display device 33 may display a two-dimensional (2D) slice image or a three-dimensional (3D) image of the object 16 generated by the image processing system 31, wherein the background noise of the 2D or 3D image may be reduced according to one or more methods performed by the image processing system 31.

[0044] During scanning, the RF coil array crossover cable ( Figure 1 The RF coils (not shown) may be used to transmit signals between the RF coils (e.g., the RF coil unit 14 and the RF body coil unit 15) and other aspects of the processing system (e.g., the data acquisition unit 24, the controller unit 25, etc.), for example to control the RF coils and / or receive information from the RF coils. As previously described, the RF body coil unit 15 is a transmission coil that transmits RF signals, and the local surface RF coil unit 14 receives MR signals. More generally, the RF coils are used to transmit RF excitation signals ("transmission coils") and receive MR signals emitted by the imaging object ("reception coils"). In some embodiments, the transmission and reception coils are a single mechanical and electrical structure or array of structures, and the transmission / reception mode may be switched by auxiliary circuits. In other examples, the transmission body coil (e.g., the RF body coil unit 15) and the surface reception coil (e.g., the RF coil unit 14) may include separate components. However, in order to improve image quality, it may be desirable to provide a reception coil that is mechanically and electrically isolated from the transmission coil. In this case, it is desirable that the reception coil is electromagnetically coupled to and resonates with the RF "echo" excited by the transmission coil in its reception mode.

[0045] refer to Figure 2 , which shows an MR image processing system 200 according to an exemplary embodiment. In some embodiments, the MR image processing system 200 is incorporated into an MRI system. In some embodiments, the MR image processing 200 is set at a device (e.g., an edge device, a server, etc.) that is communicatively coupled to the MRI system via a wired and / or wireless connection. In some embodiments, the MR image processing system 200 is set at a separate device (e.g., a workstation), which can receive images / maps from the MRI system or from a storage device storing images / maps generated by the MRI system. The MR image processing system 200 may include an image processing system 31, a user input device 32, and a display device 33.

[0046] The image processing system 31 includes a processor 204 configured to execute machine-readable instructions stored in a non-transitory memory 206. The processor 204 can be single-core or multi-core, and the program executed thereon can be configured to perform parallel or distributed processing. In some embodiments, the processor 204 can optionally include separate components distributed in two or more devices, which can be remotely located and / or configured for coordinated processing. In some embodiments, one or more aspects of the processor 204 can be virtualized and executed by a remotely accessible networked computing device configured with a cloud computing configuration.

[0047] The non-transitory memory 206 may include a segmentation module 208 and a threshold module 210. The segmentation module 208 includes instructions for segmenting the MR image and / or the parameter image into a foreground and a background. In some embodiments, the segmentation instructions implement an analysis method, such as clustering, region growing, edge detection, etc. In some embodiments, the segmentation module 208 implements a trained neural network for segmenting the MR image / map. For example, the segmentation module 208 may store instructions for implementing a neural network (such as a CNN of the CNN architecture 300), and / or routines / methods for initializing, training, and adjusting parameters of the neural network. The segmentation module 208 may include trained and / or untrained neural networks, and may also include various data associated with one or more neural network models stored therein, such as training data, training routines, or parameters (e.g., weights and biases).

[0048] The non-transitory memory 206 also includes a threshold module 210. The threshold module 210 may include instructions for determining and applying an intensity threshold. In some embodiments, the threshold module 210 includes instructions for determining an Otsu threshold for the received MR image and / or parameter map based on the Otsu threshold formula. In some embodiments, the threshold module 210 may use the segmentation map determined by the segmentation module 208 to calculate the highest intensity between background pixels / voxels.

[0049] In some embodiments, non-transitory memory 206 may include components disposed on two or more devices that may be remotely located and / or configured for coordinated processing. In some embodiments, one or more aspects of non-transitory memory 206 may include a remotely accessible networked storage device configured in a cloud computing configuration.

[0050] The image processing system 200 may further include a user input device 32. The user input device 32 may include one or more of a touch screen, a keyboard, a mouse, a trackpad, a motion sensing camera, or other devices configured to enable a user to interact with and manipulate data within the image processing system 31. In one example, the user input device 32 may enable a user to select an intensity threshold or intensity range, such as by interacting with an intensity threshold control interface that may be displayed via the display device 33.

[0051] Display device 33 may include one or more display devices utilizing nearly any type of technology. In some embodiments, display device 33 may include a computer monitor and may display unprocessed and processed MR images and / or parametric maps, as well as an intensity threshold control interface (such as a Fig.13 The display device 33 may be incorporated into the same package as the processor 204, the non-transitory memory 206, and / or the user input device 32, or may be a peripheral display device and may include a monitor, a touch screen, a projector, or other display device known in the art that enables a user to view the MR images generated by the MRI system and / or interact with various data stored in the non-transitory memory 206.

[0052] It should be understood that Figure 2 The image processing system 200 shown in FIG. 2 is for illustration and not limitation. Another suitable image processing system may include more, fewer, or different components.

[0053] Go to Figure 3 , which shows a CNN architecture 300 for segmenting an MR image / map according to an exemplary embodiment. The CNN architecture 300 represents a U-net architecture configured to receive an input MR image / map including a plurality of pixels / voxels, and segment the image / map into foreground and background by classifying each pixel / voxel of the plurality of pixels / voxels of the input MR image / map as foreground or background. The CNN architecture 300 includes a series of mappings from an input image tile 302 that may be received by an input layer through a plurality of feature maps and finally to an output segmentation map 356 that may be produced by an output layer.

[0054] Various elements comprising the CNN architecture 300 are labeled in the legend 358. As indicated in the legend 358, the CNN architecture 300 includes multiple feature maps (and / or replicated feature maps), each of which can receive input from an external file or a previous feature map, and can convert / map the received input into an output to produce the next feature map. Each feature map can include multiple neurons, wherein in some embodiments, each neuron can receive input from a limited number of neurons in the previous layer / feature map, and can calculate a single output based on the received input. Feature maps can be described using spatial dimensions such as length, width, and depth, where the dimensions refer to the number of neurons that comprise the feature map (e.g., how many neurons are long, how many neurons are wide, and how many neurons are deep for a specified feature map).

[0055] In some embodiments, neurons of a feature map may compute outputs by performing dot products of received inputs using a set of learned weights (each set of learned weights may be referred to herein as a filter), wherein each received input has a unique corresponding learned weight, wherein the learned weights are learned during training of the CNN.

[0056] The transformations / mappings performed by each feature map are indicated by arrows, where each type of arrow corresponds to a unique transformation, as indicated in legend 358. A solid black arrow pointing to the right indicates a 3×3 convolution, where the output of a 3×3 grid of feature channels from the immediately previous feature map is mapped to a single feature channel of the current feature map. A downward pointing arrow indicates a 2×2 max pooling, where the maximum value from the 2×2 grid of feature channels from the immediately previous feature map is propagated to a single feature channel of the current feature map, resulting in the spatial resolution of the immediately previous feature map being reduced by half. An upward pointing arrow indicates a 2×2 up-convolution, which includes mapping the output of a single feature channel from the immediately previous feature map to a 2×2 grid of feature channels in the current feature map, thereby increasing the spatial resolution of the immediately previous feature map by a factor of 2. A dashed arrow pointing to the right indicates copying and cropping of the immediately previous feature map for use in the current feature map. Cropping enables the dimensions of the copied feature map to match the dimensions of the feature channels to be concatenated with the copied feature map. A right-pointing arrow with a hollow head indicates a 1×1 convolution, where each feature channel in the immediately previous feature map is mapped to a single feature channel of the current feature map, or in other words, where a 1-to-1 mapping of feature channels occurs between the immediately previous feature map and the current feature map. The processing at each feature map may include the convolution described above as well as activation, where the activation map is a nonlinear function that constrains the output values ​​of the processing to be within bounds.

[0057] In addition to the operations indicated by the arrows in legend 358, CNN architecture 300 also includes a solid filled rectangle corresponding to a feature map, where the feature map includes a height (e.g., Figure 3 The length from top to bottom shown corresponds to the y spatial dimension in the xy plane), the width ( Figure 3 , assuming that the magnitude is equal to the height, corresponding to the x spatial dimension in the xy plane) and the depth (as Figure 3 , corresponding to the number of features in each feature channel). Similarly, CNN architecture 300 includes a hollow (unfilled) rectangle corresponding to the copied and cropped feature map, where the copied feature map includes a height (e.g. Figure 3 The length from top to bottom shown corresponds to the y spatial dimension in the xy plane), the width ( Figure 3 , assuming that the magnitude is equal to the height, corresponding to the x spatial dimension in the xy plane) and the depth (as Figure 3 The length from left to right shown corresponds to the number of features in each feature channel).

[0058] Starting at the input image block 302 (also referred to herein as the input layer), data corresponding to an MR image or a parameter map can be input and mapped to a first feature set. In some embodiments, the input data corresponds to an MR image with some type of contrast, such as a functional MRI (fMRI), a perfusion weighted image (PWI), a diffusion weighted image (DWI), etc. In some embodiments, the input data corresponds to an MR parameter map, such as a map of an apparent diffusion coefficient map (ADC) that reflects diffusion magnitude. The input data can correspond to a two-dimensional (2D) or three-dimensional (3D) MR image / map. In some embodiments, the input data is preprocessed (e.g., normalized) before being processed by the neural network.

[0059] For purposes of illustration, a specific configuration is taken as an example. The input image tile 302 includes a feature map including 572×572 channels, which corresponds to an input image including 572×572 pixels, where there is a 1-to-1 correspondence between feature channels and pixels. Each of the 572×572 channels may include multiple features corresponding to the 572×572 pixels of the input image. In some embodiments, each of the 572×572 channels may include three features, where each feature may correspond to a color intensity of a color, such as in an RGB pixel color model. In some embodiments, each of the 572×572 channels includes a single feature for each feature channel, where the single feature corresponds to a monochrome intensity for each pixel, such as according to a grayscale / black and white color system.

[0060] As indicated by the solid black right-pointing arrow immediately to the right of the input image tile 302, a 3×3 convolution of the feature channels of the input image tile 302 is performed to produce a feature map 304. As described above, the 3×3 convolution includes mapping the input from the 3×3 grid of feature channels to a single feature channel of the current feature map using learned weights. In the CNN architecture 300, a single feature of each channel of the input image tile 302 is used to produce 64 features for each channel in the feature map 304, where 64 different sets of weights (also referred to herein as filters) are applied to each 3×3 grid of feature channels in the input image tile 302 to produce 64 features for each feature channel in the feature map 304. The increase in features for each feature channel is indicated by the increase in the left and right widths of the feature map 304 compared to the input image tile 302. The 3×3 convolution of the CNN architecture 300 includes a stride size of 1, and therefore results in a loss of 1 pixel border in the input image for each application of the 3×3 convolution. Thus, feature map 304 includes 570×570 feature channels (ie, two feature channels are missing in the x-dimension and two feature channels are missing in the y-dimension).

[0061] Feature map 304 includes 570×570 channels, each with 64 features. As indicated by the solid black rightward-pointing arrow immediately to the right of feature map 304 , a 3×3 convolution is performed on feature map 304 to produce feature map 306 .

[0062] The feature map 306 includes 568×568 channels, each with 64 features. As indicated by the downward pointing arrow below the feature map 306, a 2×2 maximum pooling operation is performed on the feature map 306 to produce the feature map 308. In short, the 2×2 maximum pooling operation includes determining the maximum feature value from the 2×2 grid of feature channels of the immediately preceding feature map, and setting a single feature in a single feature channel of the current feature map to the maximum value so determined. The 2×2 maximum pooling employed herein includes a step size of 2. The 2×2 maximum pooling thus combines the outputs from 4 feature channels (2 in the x dimension and 2 in the y dimension) to produce a feature map with reduced resolution (the output feature map will include half the number of feature channels in the x direction and half the number of feature channels in the y direction, or in other words, the output feature map will include one quarter of the number of feature channels of the input feature map). 2×2 max pooling does not change the number of features per channel because corresponding features between feature channels are combined, but features within a single feature channel are not combined. In addition, feature map 306 is cropped, copied, and concatenated with the output from feature map 348 to produce feature map 350, as indicated by the dashed right-pointing arrow immediately to the right of feature map 306.

[0063] Feature map 308 includes 284×284 channels (half the resolution of feature map 306 due to 2×2 max pooling), each with 64 features. As indicated by the solid black right-pointing arrow immediately to the right of feature map 308, a 3×3 convolution is performed on feature map 308 to produce feature map 310.

[0064] Feature map 310 includes 282×282 channels, each with 128 features. As indicated by the solid black rightward-pointing arrow immediately to the right of feature map 310 , a 3×3 convolution is performed on feature map 310 to produce feature map 312 .

[0065] Feature map 312 includes 280×280 channels, each with 128 features. As indicated by the downward pointing arrow below feature map 312, a 2×2 max pooling operation is performed on feature map 312 to produce feature map 314, where feature map 314 is half the spatial resolution of feature map 312. Additionally, feature map 312 is cropped, copied, and concatenated with the output from feature map 342 to produce feature map 344, as indicated by the dashed rightward pointing arrow immediately to the right of feature map 312.

[0066] Feature map 314 includes 140×40 channels, each with 128 features. As indicated by the solid black rightward-pointing arrow immediately to the right of feature map 314 , a 3×3 convolution is performed on feature map 314 to produce feature map 316 .

[0067] Feature map 316 includes 138×138 channels, each with 256 features. As indicated by the solid black rightward-pointing arrow immediately to the right of feature map 316 , a 3×3 convolution is performed on feature map 316 to produce feature map 318 .

[0068] Feature map 318 includes 136×136 channels, each channel having 256 features. As indicated by the downward pointing arrow below feature map 318, a 2×2 max pooling operation is performed on feature map 318 to produce feature map 320, where feature map 320 is half the spatial resolution of feature map 318. In addition, feature map 318 is cropped, copied, and concatenated with the output from feature map 336 to produce feature map 338, as indicated by the dashed rightward pointing arrow immediately to the right of feature map 318.

[0069] Feature map 320 includes 68×68 channels, each with 256 features. As indicated by the solid black rightward-pointing arrow immediately to the right of feature map 320 , a 3×3 convolution is performed on feature map 320 to produce feature map 322 .

[0070] Feature map 322 includes 66×66 channels, each channel having 512 features. As indicated by the solid black rightward-pointing arrow immediately to the right of feature map 322 , a 3×3 convolution is performed on feature map 322 to produce feature map 324 .

[0071] Feature map 324 includes 64×64 channels, each channel having 512 features. As indicated by the downward pointing arrow below feature map 324, a 2×2 max pooling operation is performed on feature map 324 to produce feature map 326, where feature map 326 is half the spatial resolution of feature map 324. In addition, feature map 324 is cropped, copied, and concatenated with the output from feature map 330 to produce feature map 332, as indicated by the dashed rightward pointing arrow immediately to the right of feature map 324.

[0072] Feature map 326 includes 32×32 channels, each channel having 512 features. As indicated by the solid black rightward-pointing arrow immediately to the right of feature map 326 , a 3×3 convolution is performed on feature map 326 to produce feature map 328 .

[0073] Feature map 328 includes 30×30 channels, each channel having 1024 features. As indicated by the solid black rightward-pointing arrow immediately to the right of feature map 328 , a 3×3 convolution is performed on feature map 328 to produce feature map 330 .

[0074] Feature map 330 includes 28×28 channels, each with 1024 features. As indicated by the upward pointing arrow directly above feature map 330, a 2×2 up-convolution is performed on feature map 330 to produce the first half of feature map 332, while features copied and cropped from feature map 324 are used to produce the second half of feature map 332. In short, a 2×2 up-convolution (herein with a stride size of 2) includes convolving a feature map in a single feature channel of the immediately preceding feature map to four features distributed among four feature channels in the current feature map (i.e., the output from a single feature channel is treated as the input of four feature channels).

[0075] Feature map 332 includes 56×56 channels, each with 1024 features. As indicated by the solid black rightward-pointing arrow immediately to the right of feature map 332 , a 3×3 convolution is performed on feature map 332 to produce feature map 334 .

[0076] Feature map 334 includes 54×54 channels, each channel having 512 features. As indicated by the solid black rightward-pointing arrow immediately to the right of feature map 334 , a 3×3 convolution is performed on feature map 334 to produce feature map 336 .

[0077] Feature map 336 includes 52×52 channels, each with 512 features. As indicated by the upward pointing arrow just above feature map 336, a 2×2 upward convolution is performed on feature map 336 to produce one half of feature map 338, while the copied and cropped features from feature map 318 produce the second half of feature map 338.

[0078] Feature map 338 includes 104×104 channels, each with 512 features. As indicated by the solid black rightward-pointing arrow immediately to the right of feature map 338 , a 3×3 convolution is performed on feature map 338 to produce feature map 340 .

[0079] Feature map 340 includes 102×102 channels, each channel having 256 features. As indicated by the solid black rightward-pointing arrow immediately to the right of feature map 340 , a 3×3 convolution is performed on feature map 340 to produce feature map 342 .

[0080] Feature map 342 includes 100×100 channels, each with 256 features. As indicated by the upward pointing arrow just above feature map 342, a 2×2 upward convolution is performed on feature map 342 to produce the first half of feature map 344, while features copied and cropped from feature map 312 are used to produce the second half of feature map 344.

[0081] Feature map 344 includes 200×200 channels, each with 256 features. As indicated by the solid black rightward-pointing arrow immediately to the right of feature map 344 , a 3×3 convolution is performed on feature map 344 to produce feature map 346 .

[0082] Feature map 346 includes 198×198 channels, each with 128 features. As indicated by the solid black rightward-pointing arrow immediately to the right of feature map 346 , a 3×3 convolution is performed on feature map 346 to produce feature map 348 .

[0083] Feature map 348 includes 196×196 channels, each with 128 features. As indicated by the upward pointing arrow just above feature map 348, a 2×2 upward convolution is performed on feature map 348 to produce the first half of feature map 350, while features copied and cropped from feature map 306 are used to produce the second half of feature map 350.

[0084] Feature map 350 includes 392×392 channels, each with 128 features. As indicated by the solid black rightward-pointing arrow immediately to the right of feature map 350 , a 3×3 convolution is performed on feature map 350 to produce feature map 352 .

[0085] Feature map 352 includes 390×390 channels, each channel having 64 features. As indicated by the solid black rightward-pointing arrow immediately to the right of feature map 352 , a 3×3 convolution is performed on feature map 352 to produce feature map 354 .

[0086] Feature map 354 includes 388×388 channels, each with 64 features. As indicated by the rightward-pointing arrow with a hollow head immediately to the right of feature map 354, a 1×1 convolution is performed on feature map 354 to produce feature map 356. In short, the 1×1 convolution includes a 1-to-1 mapping of feature channels in a first feature space to feature channels in a second feature space, wherein no reduction in spatial resolution occurs.

[0087] The feature map 356 includes 388×388 channels, each channel having 2 features. The 2 features of each channel in the feature map 356 correspond to the categories of background and foreground, and the 388×388 channels correspond to a 388×388 pixel segmentation map. Thus, a series of 3×3 convolutions results in a reduction in the resolution of the input MR image from 572×572 pixels to an output segmentation map of 388×388 pixels. The feature map 356 may include an output layer of neurons, wherein each neuron may correspond to a pixel of the segmentation map, and wherein the output of each neuron may correspond to a binary classification scheme, and may therefore indicate whether a given pixel corresponds to the background or foreground of the input MR image or the parametric map.

[0088] In this way, the CNN architecture 300 can enable segmentation / classification of multiple pixels / voxels of an MR image or a parametric map. In one example, the CNN architecture 300 can enable one or more features of a pixel / voxel of an MR image / map to be mapped to a binary space, wherein the binary space includes a first classification and a second classification. In one example, the first classification may correspond to a foreground, and the second classification may correspond to a background. The CNN architecture 300 illustrates the transformation of a feature map that occurs when an input image patch is propagated through a neuron layer of a convolutional neural network to produce a segmentation map of the input image as an output. In one example, the architecture 300 can enable mapping of multiple pixel / voxel intensity values ​​to a binary segmentation map, wherein each pixel intensity value can be uniquely mapped (classified) to a foreground or background. As another example, a segmentation map can include a binary mask comprising a matrix of values, wherein each of the values ​​of the matrix can be one of one or zero, and wherein each pixel / voxel classified as background corresponds to a matrix value of zero, and each pixel / voxel classified as foreground can have a corresponding matrix value of one.

[0089] The weights (and biases) of the convolutional layers in the neural network 300 are learned during training. More specifically, a loss function is defined that reflects the difference between the segmentation / classification mask output by the neural network 300 and the corresponding ground truth mask. The ground truth mask can be generated manually, for example, through a series of thresholding and other image processing operations. In some embodiments, the loss function is the DICE coefficient:

[0090]

[0091] Where S is the ground truth space and T is the neural network output space. The loss can be back-projected to the neural network to update the weights (and biases) of the convolutional layers. Multiple pairs of input MR images / maps and corresponding ground truth masks can be used to train the neural network 300.

[0092] It should be understood that Figure 3The architecture and configuration of the CNN 300 shown are for illustration and not limitation. Any suitable neural network may be used herein for segmenting MR images / maps, such as ResNet, autoencoders, recurrent neural networks, generalized regression neural networks (GRNN), etc.

[0093] Go to Figure 4 , which illustrates a method 400 for reducing background noise in an MR image or parametric map according to an exemplary embodiment. The method 400 may be performed by an image processing system such as the image processing system 200.

[0094] Method 400 begins at 402, where an image processing system receives an MR image or a parametric map. In some embodiments, the image processing system may receive the MR image / map from an MRI system or from a storage device storing MR images / maps generated by the MRI system.

[0095] At 404, the MR image / map is segmented into a foreground and a background, the foreground being the region of the anatomical structure of interest and the background being outside the region of the anatomical structure of interest. In some embodiments, segmentation is achieved by analytical methods such as clustering, region growing, edge detection, etc. In some embodiments, the MR image / map is segmented using a trained neural network. For example, based on the output from the convolutional neural network, each pixel / voxel of the MR image / map received at 402 can be classified as belonging to the foreground or the background. In some embodiments, the neural network generates a segmentation / classification map for the MR image / map, wherein the map is a matrix with values ​​of one (1) or zero (0), and wherein each pixel / voxel classified as background corresponds to a matrix value of zero, and each pixel / voxel classified as foreground corresponds to a matrix value of one.

[0096] At 406, a first intensity threshold is determined. Method 500 discusses the process of determining the first intensity threshold in more detail. In some embodiments, the intensity threshold is calculated based on the intensity distribution of the received MR image / map according to the method of Otsu. In some embodiments, the intensity threshold can be selected based on the intensity of the pixels / voxels classified as background at 404 of method 400. As a more specific example, step 406 can include determining a maximum pixel / voxel intensity of the background, and setting the intensity threshold to the determined maximum background intensity.

[0097] At 408, a first mask is generated based on the first intensity threshold. In some embodiments, generating the first mask may include comparing the pixel / voxel intensity of each background pixel / voxel to the first intensity threshold and, in response to the intensity being below the first intensity threshold, setting the mask value corresponding to the pixel / voxel to zero, or otherwise setting the mask value corresponding to the pixel to one. Setting the mask value corresponding to the foreground pixel / voxel to one.

[0098] At 410, a first de-noised MR image / map is generated by applying the first mask to the MR image / map. In some embodiments, applying the first mask to the MR image / map may include multiplying the intensity of each pixel / voxel of the MR image by a corresponding mask value in the first mask. The mask value includes 1 or 0, wherein background pixels / voxels having intensities less than a first intensity threshold have corresponding mask values ​​of zero (and thus the intensity of such pixels in the de-noised MR image is zero), and all other pixels / voxels of the MR image / map have corresponding mask values ​​of 1 (and thus their pixel intensities are maintained at the same level in the de-noised MR image as in the MR image).

[0099] At 412, the first noise-reduced MR image is displayed via a display device, such as display device 33. Step 412 may also include displaying an intensity threshold control interface along with the noise-reduced MR image, such as Fig.13 Intensity threshold control interface 1300 is shown, which may enable a user to manually select an intensity threshold to apply to the background.

[0100] At 414, it is determined whether the user has manually selected an intensity threshold via the intensity threshold control interface. If it is determined at 414 that the user has not selected an intensity threshold via the intensity threshold control interface, then method 400 continues to apply the first intensity threshold, as indicated at step 416, and then method 400 may end. However, if it is determined at 414 that the user has selected an intensity threshold via the intensity threshold control interface, then method 400 proceeds to 418.

[0101] At 418, a second mask is generated based on the user-selected intensity threshold. The second mask can suppress the output of pixels / voxels (e.g., in the background) having intensities below the user-selected intensity threshold. In some embodiments, the mask can include a matrix of values ​​corresponding to the matrix of pixels / voxels comprising the MR image / map, wherein each pixel / voxel to be exceeded corresponds to a mask value of one in the matrix of values, and wherein each pixel / voxel to be suppressed corresponds to a mask value of zero in the matrix of values.

[0102] At 420, a second denoised MR image is generated by applying the second mask to the MR image. In some embodiments, generating the second denoised MR image by applying the second mask to the MR image includes multiplying each pixel / voxel intensity in the MR image / map by the corresponding mask value, and setting the pixel / voxel intensity in the denoised MR image / map (corresponding to the pixel / voxel in the MR image / map) to the product of the initial pixel intensity and the mask value. In this way, the intensity of each pixel / voxel to be suppressed in the denoised MR image can be multiplied by zero, thereby generating a pixel intensity of zero in the denoised MR image.

[0103] At step 422, the second noise reduced MR image is displayed via a display device. The method 400 may then end.

[0104] In this manner, method 400 can achieve selective noise reduction in those areas indicated as corresponding to the background rather than the foreground, where the foreground may correspond to the anatomical region of interest. Because the MR image is first segmented into foreground and background, and the intensity threshold is applied to the background rather than the foreground, method 400 can reduce the formation of holes in areas of the foreground, which may occur when portions of the foreground have an intensity below the intensity threshold.

[0105] It should be understood that Figure 4 The method 400 shown in is for illustration and not limitation. The method for suppressing background according to the present disclosure may include more, fewer and / or different operations. For example, a method may not include operations 416 to 422, that is, no user interaction is required.

[0106] In method 400, background suppression in an MR parametric map is accomplished by segmenting the parametric map directly into foreground and background and applying an intensity threshold to the background of the map instead of the foreground. In some alternative embodiments, an MR image (e.g., DWI) is segmented into foreground and background, and a first (or second) mask is generated based on the segmentation of the MR image and a first (or second) intensity threshold. The first (or second) mask is then applied to the corresponding parametric map (e.g., ADC map) to suppress the background in the parametric map. In other alternative embodiments, an MR image (e.g., DWI) is segmented into foreground and background. A first (or second) mask is generated based on the segmentation of the MR image and a first (or second) intensity threshold, and the first (or second) mask is applied to the MR image to produce a de-noised MR image. Then, the de-noised MR image is converted into a de-noised parametric map (e.g., ADC map) by methods known in the art.

[0107] Go to Figure 5, which illustrates an example method 500 for determining an intensity threshold based on a received MR image / map. The method 500 may be performed by an image processing system such as the image processing system 200.

[0108] Method 500 begins at 502 where it is determined whether the Otsu thresholding method has been selected. In one example, a user may select a threshold determination method via a user input device, thereby enabling automatic determination of intensity thresholds based on user preferences. For example, the image processing system may store an Otsu method flag in a threshold module in response to a user selection via a user input device, and based on detecting the Otsu method flag, the image processing system may determine that the Otsu thresholding method has been selected.

[0109] If at 502, it is determined that the Otsu threshold method has not been selected, then method 500 can proceed to 504. At 504, the intensity of the background pixels / voxels of the MR image / map is evaluated, that is, only the intensity of those pixels that are classified as belonging to the background is evaluated. In some embodiments, a segmentation map (such as the segmentation map generated by the CNN architecture 300) can be used to determine which pixels / voxels of the MR image / map correspond to the background, and based on this determination, the intensity of each background pixel / voxel can be evaluated.

[0110] At 506, a first intensity threshold is set to the maximum intensity of the background pixels / voxels. That is, in some embodiments, the intensity of the single brightest pixel classified as belonging to the background is set to the first threshold intensity, effectively suppressing the output of each background pixel.

[0111] If it is determined at 502 that the Otsu threshold method has been selected, method 500 proceeds to 508. In some embodiments, determining that the Otsu threshold method has been selected may be based on preset user preferences stored in a non-transitory memory of the image processing system.

[0112] At 508, the intensity distribution of the MR image / map is determined (ie, the intensity distribution of both foreground and background is determined). In one embodiment, a histogram of pixel intensity may be generated where the x-axis includes pixel / voxel intensity and the y-axis includes the number of pixels / voxels at each intensity level.

[0113] At 510, the first intensity threshold is set to an intensity that maximizes the between-class variance of the intensity distribution determined at step 508 (or equivalently, the first intensity threshold is set to an intensity that minimizes the within-class variance). The between-class variances of the first and second classes of the previously determined intensity distribution are calculated by dividing the intensity distribution into a first class including pixels with intensities less than the selected intensity threshold and a second class including pixels with intensities greater than the selected intensity threshold. Intensity thresholds determined in this manner may also be referred to herein as Otsu thresholds, and in some embodiments, the first intensity threshold may be determined at step 510 based on the Otsu formula. Method 500 may then end.

[0114] In this way, method 500 can achieve efficient and user-friendly determination of a first intensity threshold to selectively apply to the background of an MR image / map.By enabling a user to select an intensity determination method preference, a more efficient image processing workflow and less user frustration can be achieved.

[0115] Go to Figure 6 , which illustrates an example method 600 for reducing background noise in an MR image / map according to another example embodiment. The method 600 may be performed by an image processing system such as the image processing system 200.

[0116] Method 600 begins at 602, which includes receiving an MR image or a parametric map. In some embodiments, the image processing system may receive the MR image / map from an MRI system or from a storage device storing MR images / maps generated by the MRI system.

[0117] At 604, the foreground of the MR image / map is determined. In some embodiments, the foreground is determined by analytical methods such as clustering, region growing, edge detection, etc. In some embodiments, a neural network is used to determine the foreground. In some embodiments, the neural network generates a segmentation / classification map based on the MR image / map, wherein the map is a matrix of values ​​of one (1) or zero (0), and wherein each pixel / voxel classified as background corresponds to a matrix value of zero, and each pixel / voxel classified as foreground corresponds to a matrix value of one.

[0118] At 606 , the foreground of the MR image with the background masked / discarded is displayed via a display device.

[0119] At 608, it is determined whether an intensity range including an upper intensity threshold and a lower intensity threshold has been manually selected. In some embodiments, the user may select the intensity range manually through a graphical user interface such as Fig.13The graphical user interface 1300 is shown for selecting an intensity range via a user input device. If it is determined at 608 that an intensity range has not been manually selected, the method 600 may continue to display a foreground of the MR image instead of a background, as at 610 .

[0120] However, if at 608, it is determined that an intensity range has been manually selected (e.g., via user input), method 600 proceeds to 612. At 612, a mask for the foreground is generated based on the selected intensity range and foreground of the MR image / map. In some embodiments, generating a mask for the foreground includes comparing the intensity of each pixel / voxel of the foreground of the MR image / image with a lower intensity threshold and an upper intensity threshold of the selected intensity range, and in response to the intensity of the pixel / voxel being within the selected intensity range (i.e., in response to the intensity of the pixel / voxel being both below the upper intensity threshold and above the lower intensity threshold), setting the mask value corresponding to the pixel / voxel to one, or if the intensity of the pixel / voxel is outside the selected intensity range, setting the mask value corresponding to the pixel / voxel to zero in other ways. In some embodiments, the mask value corresponding to the background pixel / voxel is zero.

[0121] At 614, the mask is applied to the foreground of the MR image / map. In some embodiments, applying the mask to the foreground of the MR image / map includes multiplying the intensity of each foreground pixel / voxel by a corresponding mask value, wherein each pixel corresponds to a unique mask value, determined as indicated at step 612, wherein the product of the pixel / voxel intensity and each pixel / voxel mask value is set to the intensity of the new intensity adjusted foreground image.

[0122] At 616, the method 600 includes displaying the intensity adjusted MR / map image via a display device. The method 600 may then end.

[0123] In this way, method 600 can achieve complete suppression of background noise while also enabling the user to select an intensity range to selectively apply to the foreground. This can also enable the user to more clearly visualize one or more sub-regions within the foreground (corresponding to the anatomical structure of interest) of the MR image / map by selecting an intensity range without considering the background.

[0124] Go to Figure 7, which shows an exemplary MR image 700 of the abdomen. Image 700 is an initial DWI image without background suppression, which includes a background 702 (a relatively dark peripheral area of ​​image 700) separated from a foreground 704 (a relatively bright area bounded by boundary 706, which corresponds to the imaged anatomical region) by a boundary 706 (a relatively bright boundary). Background 702 includes a plurality of pixels, wherein the output / intensity of each of the background pixels has not been suppressed. Foreground 704 includes a second plurality of pixels, wherein the output / intensity of each of the foreground pixels has also not been suppressed. Boundary 706 is merely illustrative in purpose and is used to indicate where background 702 and foreground 704 meet.

[0125] Similarly, Figure 8 An exemplary MR image 800 of a brain is shown, which is a DWI image without background suppression. Image 800 includes a background 802 (a relatively dark peripheral area of ​​the input image 800) separated from a foreground 804 (a relatively bright area bounded by boundary 806, which corresponds to the anatomical region being imaged) by a boundary 806 (a relatively bright boundary). Background 802 includes a plurality of pixels, wherein the output / intensity of each of the background pixels has not been suppressed. Foreground 804 includes a second plurality of pixels, wherein the output / intensity of each of the foreground pixels has also not been suppressed. Boundary 806 is merely illustrative in purpose and is used to indicate where background 802 and foreground 804 meet.

[0126] Fig. 9 An initial MR image 902 of the abdomen, a ground truth mask 904 manually determined based on the initial MR image 902, and a predicted mask 906 (i.e., a segmentation map) determined via a trained neural network are shown. The ground truth mask 904 includes a human-curated segmentation map in which lighter pixels correspond to foreground and darker pixels correspond to background. The ground truth mask 904 represents a gold standard segmentation map produced by a human expert. The predicted mask 906 includes a segmentation map produced by a trained neural network (such as the CNN discussed with reference to the CNN architecture 300), in which lighter (lighter / brighter) pixels correspond to pixels classified as foreground, and in which darker pixels correspond to pixels classified as background. Fig. 9 As shown, there is good agreement between the automatically generated prediction mask 906 and the gold standard ground truth mask 904, demonstrating the ability of the trained neural network to produce segmentation maps that approximate human best efforts.

[0127] Fig. 9Also shown is a default threshold 908, which shows the initial MR image 902 with a predetermined intensity threshold (ie, the default threshold) applied, wherein pixels within the initial MR image 902 below the intensity threshold are suppressed / masked. Fig. 9 908. It can be seen that multiple holes appear within the anatomical region at the default threshold 908. Predicted threshold 910 depicts the initial MR image 902 with the predicted threshold applied (determined as the maximum intensity value of all background pixels). OTSU threshold 912 depicts the initial MR image 902 with the Otsu threshold applied (determined based on the Otsu method using the full intensity distribution of the initial MR image 902). Although both predicted threshold 910 and Otsu threshold 912 reduce background noise to a greater extent, in the illustrated example, the degree of information loss from the anatomical region of the initial MR image 902 is further exacerbated compared to the default threshold 908. However, it should be understood that in other examples, the image to which the Otsu threshold is applied may have less loss of anatomical detail than the associated image to which the default threshold is applied.

[0128] Fig. 9 A comparison between selective background suppression methods is also included. The applied ground truth mask 914 depicts the initial MR image 902 with the ground truth mask 904 applied, and therefore represents a human best effort to selectively suppress background noise within the initial MR image. The applied prediction mask 916 depicts the initial image with the prediction mask 906 applied to selectively suppress background noise. Although some pixels on the boundaries of anatomical features are segmented out, the applied prediction mask 916 seen is considered to be closely consistent with the human best effort shown by the applied ground truth mask 914. Finally, the applied prediction mask + Otsu threshold 918 shows the initial MR image 902, where the Otsu threshold is selectively applied to background pixels determined by the prediction mask 906. As shown Fig. 9 As shown, the applied prediction mask + Otsu threshold 918 is closely consistent with the applied ground truth mask 914.

[0129] Fig.10An initial MR image 1002 of the brain, a ground truth mask 1004 manually determined based on the initial MR image 1002, and a predicted mask 1006 (segmentation map) determined via a trained neural network are shown. The ground truth mask 1004 includes a human-curated segmentation map in which lighter pixels correspond to foreground and darker pixels correspond to background. The ground truth mask 1004 represents a gold standard segmentation map produced by a human expert. The predicted mask 1006 includes a segmentation map produced by a trained neural network (such as the CNN discussed with reference to the CNN architecture 300), in which lighter (lighter / brighter) pixels correspond to pixels classified as foreground, and in which darker pixels correspond to pixels classified as background. Fig.10 As shown, there is good agreement between the automatically generated prediction mask 1006 and the gold standard ground truth mask 1004, demonstrating the ability of the CNN to produce segmentation maps that approximate human best efforts.

[0130] Fig.10 Also shown is a default threshold 1008, which shows the initial MR image 1002 with a predetermined intensity threshold applied, wherein pixels within the initial MR image 1002 below the intensity threshold are suppressed / masked. Fig.10 1008, multiple holes appear within the anatomical region at the default threshold 1008. Prediction threshold 1010 and Otsu threshold 1012 similarly depict the initial MR image 1002 with the prediction threshold applied (determined as the maximum value of all background pixels) or the Otsu threshold applied (determined based on the Otsu method using the full intensity distribution of the initial MR image 1002), respectively. Although both the prediction threshold 1010 and the Otsu threshold 1012 reduce background noise to a greater extent, in the illustrated example, the degree of information loss from the anatomical region of the initial MR image 1002 is further exacerbated compared to the default threshold 1008. However, it should be understood that in other examples, the image to which the Otsu threshold is applied may have less anatomical detail loss than the associated image to which the default threshold is applied.

[0131] Fig.10Comparisons between selective background suppression methods are also included. The applied ground truth mask 1014 depicts the initial MR image 1002 with the ground truth mask 1004 applied, and therefore represents a human best effort to selectively suppress background noise within the initial MR image. The applied prediction mask 1016 depicts the initial image with the prediction mask 1006 applied to selectively suppress background noise. Although some pixels on the boundaries of anatomical features are segmented out, the applied prediction mask 1016 seen is considered to be closely consistent with the human best effort shown by the applied ground truth mask 1014. Finally, the applied prediction mask + Otsu threshold 1018 shows the initial MR image 1002, where the Otsu threshold is selectively applied to background pixels determined by the prediction mask 1006. As shown Fig.10 As shown, the applied prediction mask + Otsu threshold 1018 is closely consistent with the applied ground truth mask 1014.

[0132] Fig.11 A side-by-side comparison of MR images of the abdomen with and without background suppression in a visualization application is shown. The visualization application uses a "windowing" method to visualize MR images captured with higher precision or a higher dynamic intensity value range than standard images, which maps the intensity interval of interest to the dynamic range of the display. These applications set the range of the interval used for the mapping and the center of the interval. In the initial MR image 1102, the interval range and the interval center are not optimal due to background noise, where a large amount of noise from background pixels can be observed, and some boundaries of the anatomical features of the image are lost. An example reference for the visualization of anatomical features of the initial MR image 1102 is shown in LAVA FLEX 1106. LAVA FLEX images can be used to verify correct anatomical details and anatomical feature boundaries. It should be understood that any suitable reference for verifying correct anatomical details and anatomical feature boundaries can be similarly used for the LAVA FLEX images described herein. The denoised image 1104 representing an MR image with background noise selectively reduced according to the method disclosed herein shows improved image quality compared to the initial MR image 1102.

[0133] Similar to Fig.11 , Fig.12 An initial MR image 1202 is shown in which a large amount of noise from background pixels can be observed and some boundaries of anatomical features of the image are lost. An example reference for visualization of anatomical features of the initial MR image 1202 is shown in LAVA FLEX 1206. A denoised image 1204 representing the initial MR image 1202 with background noise selectively reduced according to the methods disclosed herein shows improved image quality compared to the initial MR image 1202.

[0134] Fig.13 An example graphical user interface (GUI) 1300 including an intensity threshold control interface 1302 is shown according to an exemplary embodiment. The intensity threshold control interface 130 includes a bar with a lower intensity threshold 1306 and an upper intensity threshold 1304, wherein a user can adjust the lower intensity threshold 1306 and / or the upper intensity threshold 1304 via a user input device. For example, a user can select an intensity range used in the method 600 by adjusting the lower intensity threshold 1306 and the upper intensity threshold 1304.

[0135] As described above, background noise can be selectively reduced by first segmenting an MR image or parametric map into a foreground (region of the anatomical structure of interest in the MR image) and a background (outside the region of the anatomical structure of interest), and then selectively applying an intensity threshold to the background rather than the foreground. This can achieve greater background noise reduction while reducing or eliminating holes in the anatomical region of the MR image / map because the foreground pixels can remain unmasked even if the region of the foreground includes pixels with intensities below the intensity threshold because the intensity threshold is selectively applied to the background pixels. Thus, a higher intensity threshold than would otherwise be feasible can be applied to the background, thereby achieving greater background noise reduction without losing information from the anatomical region of the MR image.

[0136] The above-described system and method for reducing background noise in MR images can use a trained neural network to segment an MR image / map into a foreground and a background. The technical effect of using a trained neural network to classify a plurality of pixels / voxels comprising an MR image / map as foreground or background, determining an intensity threshold based on the MR image / map, and applying the intensity threshold to background pixels / voxels rather than foreground pixels / voxels is that a greater degree of background noise reduction can be achieved while mitigating the loss of anatomical regions in the foreground by forming "holes". Another technical effect of the above-described method is that a trained neural network to which the subject application is applied can be used to segment MR images of substantially any anatomical region, and thus the above-described method can be generally applied to MR images / maps of anatomical structures without significantly changing the method workflow or the parameters of the neural network.

[0137] As used herein, an element or step listed in the singular and beginning with the word "one" or "a kind of" should be understood as not excluding a plurality of said elements or steps, unless such exclusion is explicitly stated. In addition, reference to "an embodiment" of the present disclosure is not intended to be interpreted as excluding the existence of additional embodiments that also include the described features. In addition, unless explicitly stated to the contrary, an embodiment that "includes," "includes," or "has" an element or multiple elements with a particular characteristic may include additional such elements that do not have the characteristic. The terms "including" and "in which" are used as the concise language equivalents of the corresponding terms "comprising" and "wherein." In addition, the terms "first," "second," and "third," etc. are used only as marks, and are not intended to impose numerical requirements or specific positional order on their objects.

[0138] This written description uses examples to disclose the invention, including the best mode, and also to enable a person skilled in the relevant 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 a person 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 visualizing a magnetic resonance (MR) image or an MR parameter map, the method comprising: segmenting the MR image or parametric map into a foreground and a background, the foreground including a region of the anatomical structure of interest and the background outside the region of the anatomical structure of interest; applying an intensity threshold to the background of the MR image or parametric map and not applying the intensity threshold to the foreground of the MR image or parametric map to produce a noise-reduced MR image or noise-reduced parametric map, wherein the intensity threshold is selectable by a user; as well as The noise-reduced MR image or the noise-reduced parametric map is displayed via a display device. 2 . The method of claim 1 , wherein segmenting the MR image or parametric map into foreground and background comprises classifying each pixel or voxel of the MR image or parametric map as foreground or background using a trained neural network.

3. The method according to claim 1, further comprising: receiving a user-selected intensity threshold; applying the user-selected intensity threshold to the background and not applying the user-selected intensity threshold to the foreground to produce a second noise-reduced MR image or a second noise-reduced parametric map; as well as The second noise reduced MR image or the second noise reduced parametric map is displayed via the display device.

4. The method of claim 1 , wherein applying the intensity threshold to the background comprises: comparing the intensity of each pixel or voxel of the background to the intensity threshold; as well as In response to the intensity being less than the intensity threshold, the intensity of the pixel or voxel is set to zero.

5. The method according to claim 1, further comprising: calculating an Otsu threshold based on the MR image or parametric map; as well as The intensity threshold is set to the Otsu threshold.

6. The method according to claim 1, further comprising: determining a maximum intensity of the background; as well as The intensity threshold is set to the maximum intensity of the background.

7. The method according to claim 6, further comprising: receiving a user-selected intensity range; applying the user-selected range to the noise-reduced MR image or noise-reduced parametric map; as well as A subset of the foreground within the user-selected intensity range is displayed. 8 . The method of claim 1 , wherein the MR image is one of a functional MR image, a DWI image, or a PWI image, and the parametric map is an ADC map.

9. A method for visualizing an MR parameter map, the method comprising: segmenting an MR image corresponding to the MR parameter map into a foreground and a background, the foreground including a region of an anatomical structure of interest and the background outside the region of the anatomical structure of interest; applying an intensity threshold to the background of the MR image and not applying the intensity threshold to the foreground of the MR image to generate a mask, wherein the intensity threshold is selectable by a user; applying the mask to the MR parametric map to produce a noise-reduced parametric map; and The noise reduced parametric map is displayed via a display device.

10. The method of claim 9, wherein applying the intensity threshold to the background of the MR image and not applying the intensity threshold to the foreground of the MR image to generate the mask comprises: comparing the intensity of each pixel or voxel of the background to the intensity threshold; as well as In response to the pixel intensity being less than the intensity threshold, the value of the corresponding pixel or voxel in the mask is set to zero. 11 . The method of claim 9 , wherein the MR image is a DWI image, and the MR parameter map is an ADC map corresponding to the DWI image.

12. A magnetic resonance (MR) image processing system, comprising: Image processing system; and a display device coupled to the image processing system; The image processing system is configured as follows: segmenting the MR image or parametric map into a foreground and a background, the foreground including a region of the anatomical structure of interest and the background outside the region of the anatomical structure of interest; applying an intensity threshold to the background of the MR image or parametric map but not applying the intensity threshold to the foreground of the MR image or parametric map to produce a noise-reduced MR image or noise-reduced parametric map, wherein the intensity threshold is selectable by a user; and The display device is caused to display the noise-reduced MR image or the noise-reduced parametric map.

13. The MR image processing system according to claim 12, further comprising a user input device, wherein the image processing system is further configured to: receiving, via the user input device, a user-selected intensity threshold; applying the user-selected intensity threshold to the background of the MR image or parametric map and not applying the user-selected intensity threshold to the foreground of the MR image or parametric map to produce a second noise-reduced MR image or a second noise-reduced parametric map; and The second noise reduced MR image or the second noise reduced parametric map is displayed via the display device.

14. The MR image processing system of claim 12, wherein applying the intensity threshold to the background but not applying the intensity threshold to the foreground to produce the noise-reduced MR image or noise-reduced parametric map comprises: comparing the intensity of each pixel or voxel of the background to the intensity threshold; as well as Responding to the intensity of the pixel or voxel being less than the intensity threshold is done by setting the intensity of the pixel or voxel to zero.

15. The MR image processing system of claim 12, further comprising a storage device storing a trained neural network, wherein segmenting the MR image or parametric map into foreground and background comprises using the trained neural network Each pixel or voxel of the MR image or parametric map is classified as foreground or background.

16. The MR image processing system of claim 15, wherein the trained neural network is a convolutional neural network.

17. The MR image processing system according to claim 12, further comprising a threshold module, wherein the threshold module is configured to: calculating an Otsu threshold based on the MR image or parametric map; and The intensity threshold is set to the Otsu threshold.

18. The MR image processing system according to claim 12, further comprising a threshold module, wherein the threshold module is configured to: determining a maximum intensity of the background; and The intensity threshold is set to the maximum intensity of the background.

19. The MR image processing system of claim 12, further comprising a user input device, wherein the image processing system is further configured to: receiving, via the user input device, a user-selected intensity range; applying the user-selected range to the noise-reduced MR image or noise-reduced parametric map; and A subset of the foreground within the user-selected intensity range is displayed via the display device. 20 . The MR image processing system of claim 12 , wherein the MR image is one of a functional MR image, a DWI image, or a PWI image, and the parametric map is an ADC map.

Images