Magnetic resonance imaging apparatus, image processing apparatus, and noise reduction method for images

Abstract

The present application provides a magnetic resonance imaging apparatus, an image processing apparatus, and a noise reduction method of an image. In an image noise reduction process using a CNN, effective noise reduction is performed regardless of the signal level and noise level of an image. Noise information and signal level information are calculated from an image input to the CNN, the calculated information is used to determine a normalization factor suitable for the CNN, and normalization of the input image is performed. In the case where the input image is an MRI image, the noise information is estimated from the size of the background noise of the input image. The signal level information can be calculated in such a manner that, for an image obtained by dividing the input image by the size of the background noise, the pixel values of the subject region are used to calculate, for example, as an average value.

Description

TECHNICAL FIELD

[0001] The present application relates to a noise reduction technique for medical images using a CNN (Convolutional Neural Network), and particularly relates to a noise reduction technique for images obtained in a magnetic resonance imaging apparatus (hereinafter, referred to as an MRI apparatus). BACKGROUND

[0002] In medical images obtained in a medical imaging apparatus, noise is mixed due to characteristics of a modality, imaging conditions, and the like. Although a low-noise image can be obtained by putting effort on imaging conditions, an image taken in a short time for example in order to shorten the imaging time becomes high noise, and there is a limit to noise reduction. As a technique for reducing noise that is inevitably mixed through post-processing, various methods using a CNN have been proposed (for example, Patent Literature 1). In this method, a noise-rich image (noise image) and a noise-poor image (correct answer image) are used as learning images, a CNN that has been learned so that the noise image becomes the correct answer image is used, and noise is reduced by applying this to a noise-rich image actually obtained.

[0003] When the CNN is applied after learning, due to a difference in signal level between the image used in learning (noise image) and the image input when the CNN is applied (input image), sometimes a sufficient noise reduction effect is not obtained. In order to prevent such insufficient de-noising due to a difference in signal level, in noise reduction using a CNN, an image is generally normalized by the maximum value of the signal value (Patent Literature 2). Normalization is performed on both the image for learning and the input image of the CNN.

[0004] PRIOR ART DOCUMENTS

[0005] PATENT LITERATURE

[0006] Patent Literature 1: U.S. Patent Application Publication No. 2010-0183217

[0007] Patent Literature 2: Chinese Patent Application Publication No. 113344951

[0008] In the case where the pixel value of the image to be reduced in noise (input image) is normalized by the maximum value as in the prior art, if the pixel value of the region of interest deviates from the range of the pixel value of the region of interest of the learning image (noise image and correct answer image), there is a case where the noise reduction effect of the CNN obtained by learning is not obtained. SUMMARY

[0009] The present application aims to reduce noise of an image surely and effectively by performing standardization that brings a pixel value of a region of interest of an input image into a range of pixel values of a region of interest of a learning image.

[0010] In the present application, the standardization is performed by taking into account noise, thereby solving the above-described problem.

[0011] That is, the MRI apparatus of the present application includes a measurement unit that measures a nuclear magnetic resonance signal generated from an object, and a calculation unit that generates an image using the nuclear magnetic resonance signal, the calculation unit including a noise reduction unit that reduces noise of an input image of a CNN using the CNN that has been learned using a noise image that has been standardized and a correct answer image that has less noise than the noise image. The noise reduction unit has a standardization unit that standardizes the input image using noise information and signal level information of the input image with a standardization factor suitable for the CNN, and the noise reduction unit performs noise reduction processing using the CNN on the input image that has been standardized.

[0012] The function of the calculation unit of the above-described MRI apparatus can be realized as an image processing apparatus that is independent of the MRI apparatus, and the present application includes such an image processing apparatus.

[0013] Further, the noise reduction method of the present application reduces noise of an input image of a CNN using the CNN that has been learned using a noise image that has been standardized and a correct answer image that has less noise than the noise image, and includes a step of standardizing the input image using noise information and signal level information of the input image with a standardization factor suitable for the CNN, and a step of performing noise reduction processing using the CNN on the input image that has been standardized.

[0014] The noise reduction method of the present application is preferably applied to an image (MR image) acquired in an MRI apparatus.

[0015] In the present specification, the "standardization factor" refers to a value used for standardizing an image, and in particular, in the case where standardization is performed by multiplying a given pixel value by a coefficient, the coefficient is referred to as a "standardization coefficient".

[0016] Effects of the Invention

[0017] According to the present application, the input image is standardized by a signal level estimated taking into account noise information, and thus the input image can be processed without deviating the pixel value of a region of interest of the input image from the pixel value of a learning image of the CNN, and noise reduction can be effectively performed on images of various noise levels. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1is a diagram showing an outline of the MRI apparatus as a whole.

[0019] Figure 2 is a functional block diagram of a computer.

[0020] Figure 3 is a diagram showing an outline of the processing of noise reduction.

[0021] Figure 4A is a functional block diagram of the noise reduction section of Embodiment 1.

[0022] Figure 4B is a functional block diagram of the CNN learning section of Embodiment 1.

[0023] Figure 5A is a diagram showing the processing steps of the CNN learning section of Embodiment 1.

[0024] Figure 5B is a flowchart of the noise reduction processing of Embodiment 1.

[0025] Figure 6 is a diagram showing an example of the structure of a CNN.

[0026] Figure 7 is a diagram explaining estimation of background noise and region extraction.

[0027] Figure 8 is a diagram explaining subject region extraction and pixel calculation.

[0028] Figure 9 is a diagram showing an example of a GUI related to denoising strength.

[0029] Figure 10 is a functional block diagram of the noise reduction section of Embodiment 2.

[0030] Figure 11A is a diagram showing the processing steps of CNN learning of Embodiment 2.

[0031] Figure 11B is a diagram showing the processing steps of target noise setting of Embodiment 2.

[0032] Figure 12A is a diagram explaining details of target noise setting.

[0033] Figure 12B is a diagram explaining details of target noise setting.

[0034] Figure 13 is a flowchart of the noise reduction processing of Embodiment 2.

[0035] Figure 14 is a functional block diagram of the noise reduction section of Embodiment 3.

[0036] Figure 15 is a graph showing the steps of the CNN learning and the approximation formula calculation of Embodiment 3.

[0037] Figure 16 is a graph showing the relationship between the SNR and the normalization coefficient of Embodiment 3.

[0038] Figure 17 is a graph explaining the noise reduction processing of Embodiment 3.

[0039] Explanation of Reference Numerals

[0040] 1: MRI apparatus, 20: computer, 210: image reconstructing section, 230: noise reducing section, 231: noise estimating section, 2311: background noise estimating section, 2312: noise level calculating section, 232: pixel value calculating section, 233: normalizing section, 234: CNN selecting section, 235: CNN, 236: normalization coefficient calculating section, 237: target noise calculating section, 238: SNR estimating section, 239: approximation formula calculating section, 250: CNN learning section DETAILED DESCRIPTION

[0041] Embodiments of the noise reduction method of the present application will be explained below. The noise reduction method of the present application is not limited and can be suitably applied to an image (MR image) obtained in MRI, in which case, the MRI apparatus or an image processing apparatus that processes the MR image becomes a structure provided with a function to realize the noise reduction method of the present application.

[0042] An embodiment of an MRI apparatus will be explained below as an embodiment of the present application. Referring to Figure 1 the overall structure of an MRI apparatus 1 to which the present application is first applied will be explained.

[0043] Structure of MRI Apparatus

[0044] The MRI apparatus 1 of the present embodiment has, as shown in Figure 1 Fig. 1, a static magnetic field generating section such as a static magnetic field coil 11 that generates a static magnetic field in a space where a subject is placed, a transmitting high frequency coil 12 (hereinafter simply referred to as a transmitting coil) and a transmitter 16 that transmit a high frequency magnetic field pulse to a measurement region of the subject, a receiving high frequency coil 13 (hereinafter simply referred to as a receiving coil) and a receiver 17 that receive a nuclear magnetic resonance signal generated from the subject, a gradient magnetic field coil 14 that gives a magnetic field gradient to the static magnetic field generated by the static magnetic field coil 11 and a gradient magnetic field power supply 15 that is a driving power source thereof, a sequence control apparatus 18, and a computer 20. Each section of the MRI apparatus 1 except for the computer 20 is collectively referred to as a measurement section 10.

[0045] In the MRI apparatus 1, the static magnetic field coil 11 adopts various forms according to the direction of the generated static magnetic field, the perpendicular magnetic field method, the horizontal magnetic field method. The gradient magnetic field coil 14 is constituted by a combination of a plurality of coils that generate gradient magnetic fields in three-axis directions (x-direction, y-direction, z-direction) orthogonal to each other, and is driven by the gradient magnetic field power supply 15, respectively. By applying the gradient magnetic fields, position information can be added to the nuclear magnetic resonance signals generated from the subject.

[0046] In addition, in the illustrated example, a case where the transmission coil 12 and the reception coil 13 are separated is shown, but there is also a case where one coil having both the functions of the transmission coil 12 and the reception coil 13 is used. The high-frequency magnetic field irradiated by the transmission coil 12 is generated by the transmitter 16. The nuclear magnetic resonance signals detected by the reception coil 13 are sent to the computer 20 via the receiver 17.

[0047] The sequence control device 18 controls the operation of the gradient magnetic field power supply 15, the transmitter 16, and the receiver 17, controls the timing of the application of the gradient magnetic fields and the high-frequency magnetic field and the reception of the nuclear magnetic resonance signals, and performs measurement. The time chart of the control is called an imaging sequence, and is set in advance in correspondence with the measurement and stored in a storage device or the like provided in the computer 20 described later.

[0048] The computer 20 is an information processing device provided with a CPU, a memory, a storage device, and the like, controls the operation of each part of the MRI apparatus via the sequence control device 18, and performs arithmetic processing on the received echo signals to obtain an image of a predetermined imaging region. The functions implemented by the computer 20 will be described later, and the functions can be implemented as the computer 20 included in the MRI apparatus 1, and can also be implemented by a computer, a workstation, or the like independent of the MRI apparatus. That is, it can be an image processing device including part or all of the functions of the computer 20.

[0049] The computer 20 is connected to a display device 30, an input device (input section) 40, an external storage device 50, and the like. The display device 30 is an interface that displays the results obtained in the arithmetic processing and the like to an operator. The input device 40 is an interface for the operator to input conditions, parameters, and the like required for the measurement and the arithmetic processing implemented in the present embodiment. The user can input, for example, the number of measured echoes, the echo time TE, the echo interval, and the like, which are measurement parameters, via the input device 40. The external storage device 50 holds, together with the storage device inside the computer 20, data used in various arithmetic processes performed by the computer 20, data obtained by the arithmetic processes, input conditions, parameters, and the like.

[0050] The computer 20 controls the measurement unit 10 of the MRI apparatus and processes the signals measured by the measurement unit 10 as described above. However, the MRI signals measured by the measurement unit 10 are obtained as signals from the subject being imaged, which are superimposed with a given amount of noise depending on the characteristics of the apparatus, imaging conditions, etc., which causes image quality degradation. The computer 20 of this embodiment has a function to reduce the noise superimposed on the image as part of the signal processing.

[0051] In this embodiment, a CNN is used for noise reduction. For this purpose, computer 20, such as... Figure 2 As shown, the system includes an image reconstruction unit 210, a noise reduction unit 230, and a CNN learning unit 250. The noise reduction unit 230 further includes: a noise estimation unit 231 that calculates information related to noise in the image being processed; a normalization unit 233 that normalizes the image to a given pixel value based on the estimated noise information; and a learned CNN 235. The processing method in the noise reduction unit 230 will be described in detail in a later embodiment.

[0052] Furthermore, although not shown in the figure, the computer 20 may include a measurement control unit that controls the various elements of the measurement unit 10, an image generation unit that generates an image for display on the display device, etc.

[0053] In addition, Figure 2 In this paper, the CNN learning unit 250 is shown as a function included in the computer 20. However, the function of the CNN learning unit 250 can also be implemented by the computer of the MRI apparatus 1 that performs noise reduction processing on the actual measurement images, or by a computer different from the image processing apparatus. In this case, the CNN learning unit 250 of the computer 20 is omitted. Figure 2 As shown by the dotted line, a CNN learning unit 250 with the same function is set up in another computer 20A, and the CNN learned in this CNN learning unit 250 is imported into the computer 20 of the MRI device 1. Therefore, in the following description, the CNN learning performed in the CNN learning unit 250 includes both the case where it is performed in a device equipped with a learned CNN (e.g., the computer 20 of the MRI device) and the case where it is performed in a device other than that (computer 20A).

[0054] The CNN learning unit 250 uses a large number of pairs of images containing noise (noisy images) with images without noise or with less noise than the noisy images (positive resolution images) to learn the CNN, so that for input of noisy images, the output is a positive resolution image. At this time, the learning images are normalized according to the noise intensity of the noise images used, and the learned CNN is obtained.

[0055] The noise reduction section 230 performs noise reduction processing using the learned CNN 235. The noise reduction section 230 performs standardization suitable for the learned CNN with respect to the noise of an input image before applying the learned CNN 235 to the input image as a target of noise reduction. To this end, as shown in Figure 3 (S1), the standardization section 233 performs standardization of the input image according to the noise level estimated by the noise estimation section 231 so that the pixel values of the input image become an appropriate range with respect to the pixel value range of the learning image (the learning image subjected to standardization) of the learned CNN 235 to which the input image is applied (S2). The noise reduction section 230 inputs the input image subjected to standardization to the CNN 235 and performs noise reduction processing (S3). The output image of the CNN is subjected to processing for restoration to the pixel values before standardization as needed, and a noise-reduced image is output (S4).

[0056] As described above, according to the present embodiment, the noise reduction section 230 uses the CNN learned by the learning image subjected to standardization according to the noise level, and in the application of the learned CNN, the input image is standardized according to the noise level of the input image so as to become the pixel value range of the learned CNN, and noise reduction is performed, whereby the problem that the pixel values of the region of interest of the input image deviate from the pixel values of the learning image of the CNN and the noise reduction effect becomes small can be solved, and noise reduction can be surely performed on images of various noise levels.

[0057] A specific embodiment of a method of standardizing an input image to perform noise reduction will be described below.

[0058] <Embodiment 1>

[0059] The present embodiment prepares a plurality of CNNs having different noise levels, selects a CNN having the same noise level as the noise level of an input image from the plurality of CNNs, and performs noise reduction after performing the same standardization as the selected CNN on the input image. Estimation of the noise level of the image is performed in order to perform processing corresponding to the noise level.

[0060] In the estimation of the noise level, the present embodiment uses the characteristic of an MR image that the noise level of air is proportional to the noise of the image itself. That is, the noise level of air, that is, the background noise, is estimated, and the estimated background noise is used to determine a standardization coefficient. Standardization is performed on the image input to the CNN in both the learning stage and the application stage of the CNN.

[0061] To this end, the noise reduction section 230 of the present embodiment performs standardization of the input image according to the noise level of the input image as shown in Figure 4AAs shown, there are provided a background noise estimation section 2311 that estimates the size of the background noise of an image (input image) that is the subject of noise reduction processing, a pixel value calculation section 232 that extracts a subject region and calculates the pixel values of the region based on the estimated size of the background noise, and a noise level calculation section 2312 that calculates the noise level using the pixel values of the image of the subject region. Further, the noise reduction section 230 is provided with a CNN selection section 234 that selects a CNN to be used in the processing of the input image from among a plurality of CNNs 235 that have been learned for each noise level. In addition, there are provided functional sections (251, 252, 253, etc.) that correspond to the functions of the noise estimation section 231, the pixel value calculation section 232, and the normalization section 233, which are the functions of the noise reduction section 230. Figure 2 Elements having the same functions as those shown in the drawing are denoted by the same reference numerals.

[0062] Further, in the CNN learning section 250, as shown in Figure 4B , there are provided functional sections (251, 252, 253, etc.) that correspond to the functions of the noise estimation section 231, the pixel value calculation section 232, and the normalization section 233, which are the functions of the noise reduction section 230.

[0063] The noise reduction processing of the present embodiment will be described below with reference to the flow of the processing shown in Figure 5A , Figure 5B . Figure 5A is the flow of the learning phase of the CNN, Figure 5B is the processing flow of the noise reduction.

[0064] First, the learning of the CNN by the CNN learning section 250 will be described. In the following description, when distinguishing between the CNN before learning and the CNN after learning, the former is denoted as CNN235B and the latter is denoted as CNN235A, and when no distinction is needed, they are collectively denoted as CNN.

[0065] The CNN learning section 250 performs a given normalization on a learning image input to the CNN235B, performs repeated operations using the learning image that has been normalized, and performs learning so that the CNN outputs an image in which the noise has been reduced.

[0066] The CNN235 has a configuration in which a plurality of layers (convolution layers and pooling layers) are overlaid between an input layer and an output layer, as shown in Figure 6 , and can be obtained as a publicly known software in which the number of layers and the kernel size of each layer are set, and in addition, a configuration in which the number of layers, the kernel size, and the like are appropriately customized can also be used.

[0067] CNN235B determines the coefficients (bias, kernel, etc.) used in the CNN by repeatedly inputting and outputting a large number of training images, resulting in a final output image with reduced noise, which becomes the trained CNN235A. In this embodiment, the training images, for example in MRI, are a set of images with high SNR (signal-to-noise ratio) obtained by repeatedly taking and summing the same images (forward images) and images with noise added to the forward images (called noisy images). The added noise can be, for example, Gaussian noise or other noise with a given distribution. Alternatively, the training images can also be a set of forward images with low SNR (with noise added to high SNR images) and noisy images (with further noise added to low SNR forward images). By using the low SNR forward images in the training, excessive noise reduction can be suppressed, resulting in an effect where the granularity of the noise in the noise-reduced image can be naturally observed.

[0068] Furthermore, in MRI, multiple contrast images (T2-emphasis images, T1-emphasis images, etc.) can be acquired, and there are situations where these multiple contrast images are mixed together for learning. In this case, the SNR of the forward image may differ for each contrast image. Therefore, by adjusting the noise added to the high SNR forward image for each contrast image, a low SNR forward image with a uniform SNR across all contrast images can be created. By adding the same noise to the created low SNR forward images across contrast images to create noisy images, the ratio of the SNR of the forward image to the SNR of the noisy image can be uniform across contrast images.

[0069] Before learning, the CNN learning unit 250 changes the noise level and creates a noise image for learning at each noise level. Figure 5A (S11) and normalizes the learning image using a normalization factor (normalization factor) determined by noise information. To this end, firstly, the CNN learning unit 250, if paired with a learning image of a given noise level, estimates the magnitude N of the background (air) noise outside the region of interest of the noisy image (S12). As a method of noise estimation, for example from... Figure 7 In the histogram of the noise image shown, the peak with the lowest signal value and the highest frequency is determined as the background noise, and this signal value is set as the noise magnitude N (hereinafter referred to as background noise N).

[0070] Next, the subject region is extracted (S13). The entire noisy image (each pixel value) is divided by the background noise N, and pixels whose values ​​are greater than a given threshold T are extracted. This is used to extract the subject region. Figure 7The lower side of FIG. 8 shows a relationship between a histogram of the image divided by the background noise N and the threshold value. An image having a pixel value equal to or higher than the threshold value T is extracted as the subject region. The image before and after the extraction is shown in FIG. 9. The pixel value calculating section 232 calculates the average value M of the pixel values of the extracted region of the image divided by the background noise N. Further, 1 / M (=L) is calculated, and L is set as the noise level of the noise image (S14). Figure 8 The average value M of the pixel values of the extracted region of the image divided by the background noise N is calculated by the pixel value calculating section 232. Further, 1 / M (=L) is calculated, and L is set as the noise level of the noise image (S14).

[0071] The CNN learning section 250 (normalization section) divides the noise image as a whole (the image divided by the noise N) by the average value M of the pixel values obtained for the extracted image. Thus, the noise image is normalized by the average value M of the pixels of the extracted region (S15). The correct answer image is normalized by multiplying (1 / N) / M (S16). Thus, the correct answer image is normalized in the same manner as the noise image.

[0072] The learning of the CNN 235B is performed using the pair of the correct answer image and the noise image which are normalized in this manner, and a learned CNN is obtained (S17). The size of the noise added to the correct answer image is changed, and the above-described processing S11 to S17 is repeated (S18), and finally, a plurality of noise levels of the CNN 235A for each noise level are created.

[0073] Further, as the region extraction method in the processing S13, the histogram-based method used in S11 is simple as described above, but a method of extracting a region by tracing the outline of the subject region, or a method of extracting the subject region by a region expansion method and extracting the remaining region other than the subject region as a background region, and calculating the noise from the pixel values thereof, or the like can be used.

[0074] Further, as the image from which the pixel values of the subject region are calculated, an image from which noise is reduced in advance can be used. Such an image can be obtained by performing a known noise removal filter process, or changing the matrix size of the image (for example, adding the pixel values of adjacent pixels of a 256 x 256 image to reduce to 128 x 128, or the like).

[0075] Next, the flow of the processing of the noise reduction section 230 using the CNN 235B for each noise level of a plurality of noise levels created by the CNN learning section 250 will be described.

[0076] In the noise reduction unit 230, the standardization process for the input image is similar to that in the CNN learning process. First, the background noise estimation unit 2311 extracts the region (background) other than the subject from the input image and estimates the background noise N (S21). Next, the pixel value calculation unit 232 divides the input image by the noise N, extracts the region that is above a given threshold as the subject region, and calculates the average pixel value M of the extracted region (S22). The standardization unit 233 standardizes the input image by dividing the entire input image (the input image after dividing by N) by the average value M (S23). Furthermore, the noise level calculation unit 2312 calculates the noise level L (=1 / M) of the input image. In this way, the noise level and pixel value (signal level) are used as standardization factors to standardize the input image.

[0077] During the CNN learning process, the noise reduction unit 230 selects the CNN235A whose noise level L is the same as or closest to the noise level calculated on the input image from among the multiple CNN235A (S24), inputs the selected CNN235A with the input image, and obtains an image with reduced noise as its output image (S25).

[0078] As explained above, the CNN is prepared for each noise level and the CNN235A corresponding to the noise level of the input image is selected. The same standardization is performed on the learning image using noise level and signal level as standard factors and the CNN235A is applied. This allows for effective noise reduction of images with various signal and noise levels without loss of noise reduction effect.

[0079] Furthermore, this embodiment describes the case where a given CNN is selected from multiple CNNs according to the noise level, but the CNN can also be selected according to the degree of denoising desired by the user. For example, it can be... Figure 9 The GUI shown is displayed on the display device 30, allowing the user to select any denoising intensity. In this case, when setting the CNN with the noise level closest to the input image's noise level (Linput) as NLmedium, the CNN with the low noise level as NLlight, and the CNN with the high noise level as NLheavy, if the denoising intensity is selected as "weak," NLlight is selected as the CNN to be used; if the denoising intensity is selected as "strong," NLheavy is selected, and so on.

[0080] By setting up such a GUI, users can have more freedom in choosing the level of noise reduction.

[0081] Further, in the present embodiment, a case where one CNN 235A corresponding to one noise level is used for input image selection is described, but two or more CNNs 235A can also be used. For example, two CNNs 235A close to the noise level of the input image can also be selected and used, and the outputs of the two can be combined.

[0082] Embodiment 2

[0083] In Embodiment 1, an embodiment in which a plurality of CNNs 235A different in noise level are prepared, one or more are selected in correspondence with the noise level of the input image, and the input image is normalized so as to become the same pixel value range as the learning image of the selected CNN 235A, and noise reduction is performed, is described. In the present embodiment, one CNN 235A is created, and when the one CNN 235A is used, the input image is normalized using different coefficients (normalization coefficients) in correspondence with the noise level so as to be normalized to the optimum pixel value range. When the prepared CNN is used, the noise level that becomes the target (target noise) is set, and the normalization coefficients are set to the value (range) that becomes the target noise for each noise level.

[0084] The configuration of the noise reduction section 230 of the present embodiment is shown in Figure 10 . In Figure 10 , elements having the same functions as Figure 2 and FIG. 4 are shown with the same reference numerals, and overlapping descriptions are omitted. Further, the function of the CNN learning section 250 can be realized by a device different from the MRI device or the image processing device that performs noise reduction, as in Embodiment 1.

[0085] As shown in Figure 10 , the noise reduction section 230 of the present embodiment adds a normalization coefficient calculation section 236 and a target noise calculation section 237 to the noise reduction section 230 of Embodiment 1. The normalization coefficient calculation section 236 calculates the normalization coefficient that minimizes the difference between the noise image and the correct answer image. The target noise calculation section 237 uses the CNN 235A created by the CNN learning section 250 to calculate the noise (optimum background noise) after normalizing various noise images different in noise level using the normalization coefficient calculated by the normalization coefficient calculation section 236, as the target noise. The normalization section 233 uses the target noise to perform normalization of the input image. Further, although not shown, the CNN learning section 250 has a function corresponding to the noise reduction 230, as in Embodiment 1.

[0086] The processing of CNN learning and noise reduction in the present embodiment is described below with reference to FIGS. 11 to Figure 13 . Figure 11A , 11B andFigure 12A , 12B is a flow and an explanatory diagram showing the processing of CNN learning and target noise setting, Figure 13 is a diagram showing the noise reduction processing.

[0087] The learning of the CNN in the CNN learning section 250 (normalization of the image for the previous learning) is first explained.

[0088] The pairs of the correct answer image and the noise image are prepared as in Embodiment 1 (S31). The noise used in the noise image can use the noise of the given noise level. Next, the background (air) noise N of the subject region of the correct answer image is estimated (noise estimation function) (S32). The entire correct answer image is divided by the background noise N (S33), and the region of the correct answer image obtained by dividing by the noise N is extracted in which the pixel value is larger than the threshold T (T > 1) (that is, the subject region).

[0089] In Embodiment 1, in the processing S12, S13 corresponding to the above processing S32, S33, the estimation of the noise N and the extraction of the subject region are performed with the noise image as the object, but in this embodiment, since the normalization that is different for each noise level is not performed, the noise image is not required to be used, and it is set to use the correct answer image. But the noise image can be used as in Embodiment 1.

[0090] The average M of the pixels of the extracted subject region is calculated, and the correct answer image is divided by the calculated average M. Thereby, the correct answer image is normalized by the pixel value (average) of the extracted region. Next, the noise image is normalized by multiplying (1 / N) / M on the noise image (S34). The learning of the CNN is performed using the pairs of the correct answer image and the noise image normalized in S34 (S35), and this is the same as in Embodiment 1, but here, the repetition for each noise level (S18) is not required to be performed, and a single CNN 235A is made. Figure 5A : S18), but a single CNN 235A is made.

[0091] Next, the target noise is set. In order to set the target noise, first, the normalization coefficient calculation section 236 calculates the normalization coefficient that makes the difference between the input image and the correct answer image the minimum in the case where the CNN 235A made by the CNN learning section 250 is applied to various images different in noise intensity. Next, the target noise calculation section 237 normalizes the normalization coefficient calculated by the normalization coefficient calculation section 236 to obtain the background noise of the image before the noise reduction is performed, and determines the optimum range of the background noise as the target noise.

[0092] Specifically, as in the processing at the time of learning of the CNN, noise is added to the correct answer image to create a noise image (S41). The noise image is multiplied by a plurality of coefficients (Kmin to Kmax) to obtain a number of noise images equal to the number of coefficients (S42), and noise reduction using the learned CNN is performed on these noise images (S43). After dividing the output of the learned CNN, that is, the image after noise reduction, by each of the coefficients (Kmin to Kmax), the root mean square error (RMSE) with the correct answer image is calculated, and the coefficient (Kj) at which this value becomes the minimum is found (S44). Next, the background noise N' of the noise image before noise reduction after multiplication by the coefficient (Kj) is estimated (S45). The estimation method of the background noise N' can use the same method as that used to estimate the background noise from the correct answer image in the learning of the CNN.

[0093] The processing S41 to S45 is performed with a change in noise intensity (S46), and the coefficient Kj at which the RMSE with the correct answer image becomes the minimum and the background noise N' of the noise image before noise reduction after multiplication by the coefficient Kj are found for each noise image with a different noise intensity. Thereafter, the allowable range of the background noise N' of each noise image (optimum background noise) is determined as the target noise (S47).

[0094] The details of the calculation of the target noise will be further described using FIG. 12. Figure 12A , Figure 12B ). Figure 12A is a graph showing the relationship between the coefficient and the RMSE of each image in the case where noise reduction is performed with a change in the coefficient for four noise images with different noise intensities, and the coefficient at the position at which the RMSE becomes the minimum (the position shown by a circle in the graph) is Kj. The noise intensity is represented by "noise SD", and four cases where the noise SD is changed to 1, 2, 3, and 5 are shown in FIG. 12. In the graph, in the input image with noise SD = 1 (the minimum noise intensity), the difference (RMSE) from the correct answer image becomes the minimum at Kj = 0.6, in the input image with noise SD = 2, the difference (RMSE) from the correct answer image becomes the minimum at Kj = 1, and the higher the noise intensity, the larger the value of Kj at which the RMSE becomes the minimum.

[0095] Figure 12B is a graph showing the relationship between the coefficient and the background noise N' for four noise images with different noise intensities, and in the graph of Figure 12A , the position of Kj at which the RMSE becomes the minimum is also shown by a circle. The target noise calculation section 237 calculates the value within the range of the background noise covered by Kj for each noise intensity as the target noise. In Figure 12BIn the graph, the background noise N' (N'(SNR1), N'(SNR2), N'(SNR3), N'(SNR4)) becomes approximately 0.18 to 0.22 for Kj (4 points) obtained for each of the 4 noise images. The target noise calculating section 237, for example, calculates the average value and the median value of the 4 background noises N' to set the target noise.

[0096] In the above processing, the creation of the learned CNN and the setting of the target noise are completed. The noise reduction section 230 uses these learned CNN and the target noise to perform processing of the input image. The noise reduction processing will be described below with reference to Figure 13

[0097] The background (air) noise N outside the region of interest of the input image is initially estimated (S51), and the entire input image is divided by the noise N (S52). Next, the input image (image obtained by dividing by the noise N) is multiplied by the target noise (S53). Thus, the input image is standardized so that the background noise becomes an optimal range. Although the standardized image can be used as it is as the input image of the learned CNN to perform the noise reduction processing, in the present embodiment, after the adjusted coefficient F that adjusts the denoising strength is multiplied on the standardized noise image (S54), it is input to the learned CNN to obtain an output image in which the noise is reduced (S55).

[0098] The adjusted coefficient F is a value determined in advance in correspondence with the denoising strength, and is, for example, F = 2.0 if the strength is weak (Light), F = 1.0 if it is medium (Medium), and F = 0.5 if it is strong (Heavy). By multiplying such a coefficient F, the input image is standardized by the background noise x coefficient F, and it is possible to weaken or strengthen the denoising strength. Which adjusted coefficient to use can be accepted by the user via the GUI shown in FIG. 8 described in Embodiment 1, or can be set to a prescribed value in correspondence with the noise level. Figure 10

[0099] According to the present embodiment, by deciding the target noise (optimal background noise) for using the learned CNN, and standardizing the input image with the target noise, it is possible to perform the noise reduction processing with one learned CNN for input images having different noise levels. Further, since the adjusted coefficient F corresponding to the denoising strength is set in addition to the target noise, it is possible to improve the versatility of the processing.

[0100] <Embodiment 3>

[0101] ​​This embodiment also uses a fully trained CNN on input images (noise reduction processed images) of various noise levels, just like in embodiment 2. However, in embodiment 2, when using the fully trained CNN, the input image is normalized to the optimal target noise, but in this embodiment, the normalization coefficient is calculated as a function of the signal-to-noise ratio (SNR), and the SNR of the input image is estimated. The normalization coefficient is then obtained based on the SNR, thereby normalizing the input image.

[0102] The structure of the noise reduction unit 230 in this embodiment is illustrated in... Figure 14 As shown. In Figure 14 In the middle, and with Figure 4 and Figure 10 The same elements are shown with the same reference numerals, and repeated descriptions are omitted. Figure 14 As shown, the noise reduction unit 230 of this embodiment includes an SNR estimation unit 238 as a form of the noise estimation unit 231. Furthermore, the normalization unit 233 includes an approximation calculation unit 239 that calculates an approximation of the normalization coefficients. Also, although not shown, the CNN learning unit 250 has the same functionality as the noise reduction unit 230, and the CNN learning unit 250 can be located in either the computer 20 or any of the other computers 20A, as in other embodiments.

[0103] The following is for reference Figures 15-17 This embodiment will be used to explain the noise reduction processing that includes the processing contents of the above-mentioned parts.

[0104] The initial explanation focused on the processing of the CNN learning department 250.

[0105] CNN Learning Department 250 Figure 15 As shown, a low-noise image (forward image) and a noisy image with added noise are prepared (S61). Next, the median of the regions in the forward image where pixel values ​​are not zero (non-zero regions) is calculated. Both the forward image and the noisy image are then normalized by dividing by the calculated median (S62). Alternatively, the median may be the median of the noise image (its non-zero regions) instead of the median of the forward image. The normalized forward image and the noise image are then paired for CNN learning, resulting in a learned CNN235A (S63).

[0106] Next, the approximation calculation unit 239 calculates the approximation formula for the standardized coefficients. To do this, firstly, similar to the processes S61 and S62 described above, a positive solution image and a noisy image are created (S71), and the noise image is standardized by dividing the entire noise image by the median value Mnoise of the non-zero region of the noise image (S72). Then, the standardized noise image is multiplied by multiple (N) coefficients K (K1 to K2). NAfter a plurality of noise images (N images) are created from the input image (S71), noise reduction is performed on each noise image using the completed CNN 235A obtained in S63 (S73). After the noise-reduced image is multiplied by the "median value of the noise image Mnoise" / "coefficient to be multiplied by the noise image" to restore the original value, the root mean square error (RMSE) with the correct answer image is calculated, and the coefficient at which this value becomes the minimum (optimal coefficient) Kopt is found (S74). The background noise N" of the noise image before noise reduction, which is multiplied by the optimal coefficient Kopt on the original noise image, is estimated (S75). The SNR estimation unit 238 finds the median value M'noise of the non-zero region of the noise image multiplied by the optimal coefficient Kopt, and calculates the SNR of this image as M'noise / N" (S76).

[0107] The processing of S71 to S76 is repeated by changing the noise intensity of the noise image created in S71 (S77), and the combination of the optimal coefficient Kopt and the SNR is calculated for each noise image having a different noise intensity (S78). In Figure 16 the relationship between the SNR and the optimal coefficient Kopt obtained by the above-described repetition is shown in FIG. 8. As illustrated, the relationship between the two can be approximated by a linear expression, for example.

[0108] Ki = a • SNR + b (1)

[0109] The linear regression slope a and the linear regression intercept b of the approximation expression are approximated from the relationship between Kopt and SNR obtained for each noise image having a different noise intensity. However, the approximation expression is not limited to a linear expression.

[0110] Through the above processing, the creation of the completed CNN and the calculation of the approximation expression representing the relationship between the standardization coefficient and the noise are completed. The noise reduction unit 230 uses the completed CNN and the approximation expression to perform the standardization and noise reduction processing of the input image. The noise reduction processing will be described below with reference to Figure 17 .

[0111] First, for the input image, the noise estimation unit 231 (background noise estimation unit 2311) finds the background noise N (S81) as in Embodiment 1, and the pixel value calculation unit 232 calculates the median value Minput of the non-zero region of the input image. The SNR estimation unit 238 calculates Minput / N, and takes this value as the SNR of the input image (S82).

[0112] The standardization unit 233 imports this SNR into the approximation expression calculated by the approximation expression calculation unit 239 to find the optimal coefficient K opt (S83).

[0113] Kopt = a · (SNR / f) + b (2)

[0114] In the formula, f is an adjustment coefficient that determines the strength of de-noising, and has the same meaning as the adjustment coefficient F explained in Embodiment 2. That is, by increasing the value of f, the strength of de-noising is enhanced, and by decreasing the value of f, the strength of de-noising is weakened. In Embodiment 2, the case where f takes a discrete value as F was explained, but in the present embodiment, the standardization coefficient becomes a function of SNR, and therefore, with respect to the adjustment coefficient f, an intermediate value can also be taken. The value of f can be fixed at 1, which is a moderate strength of de-noising, or can be made to correspond to user designation or the like, and for example, a value between 0.5 and 2.0 can be adjusted as appropriate. In the case where f = 1 is set, it becomes the aforementioned formula (1).

[0115] The standardization section 233 multiplies the optimal coefficient K obtained as described above by the input image divided by the median Minput opt (the following formula (3)). Thus, the input image is standardized (S84).

[0116] I n = (1 / Minput(I0)) · K opt · I0

[0117] = (1 / Minput(I0)) · (a · (SNR / f) + b) · I0

[0118] In the formula, I0 is the input image before standardization, and I n is the input image after standardization.

[0119] The input image thus standardized is input to the learned CNN, and an output image in which noise is reduced is obtained (S85).

[0120] According to the present embodiment, the relationship between the index SNR that represents noise and the optimal standardization coefficient is approximately obtained, and in the use of a CNN that has been learned using a learning image that has been standardized according to noise, the input image is standardized using the optimal standardization coefficient obtained according to the approximately obtained relationship between SNR and the optimal standardization coefficient, according to the estimated SNR of the input image, and thereby noise reduction corresponding to various input images that differ in noise level and signal level can be achieved. Further, according to the present embodiment, by using the relationship formula of SNR and the standardization coefficient, the learned CNN can be prepared as a single CNN, and noise reduction with good precision can be performed.

[0121] Furthermore, according to the present embodiment, the strength of de-noising can be adjusted in the relationship formula, and the degree of freedom in adjusting the strength of de-noising by the user is improved.

[0122] The above describes the embodiments of the noise reduction method of the present application by way of processing in an MRI apparatus, but the present application is not limited to an MRI apparatus and can be implemented in a device (an image processing apparatus, a computer, a workstation, or the like) that can implement a general image processing program. Furthermore, in the embodiments, the noise reduction processing of an MR image using the characteristic of the MR image that the background noise is proportional to the noise of the entire image is described, but in the noise reduction processing of an image using a CNN in the present application, the main point is to optimize the standardization so that the pixel values of the input image are included in the pixel value range of the standardized image used in the CNN, and in particular, the standardization is performed so that the noise information of the image is reflected, and this can be performed for any image from which noise information can be estimated, and is not limited to an MR image. Furthermore, the noise information that can be used and the method of reflecting it in the standardization are not limited to the method of the embodiments.

Claims

1. A magnetic resonance imaging device, characterized in that, have: The measuring unit measures the nuclear magnetic resonance signal generated from the subject; and The computing unit uses the nuclear magnetic resonance signal to generate an image. The arithmetic unit includes: The noise reduction unit utilizes a CNN learned by pairing a normalized noisy image with a corresponding normalized positive image (without noise or with less noise) to reduce the noise in the input image of the CNN, thereby generating a noise-reduced image. The noise reduction unit has: The standardization department standardizes the input image using a standardization factor. The noise reduction unit performs noise reduction processing on the standardized input image using the CNN. The noise reduction unit also includes: An SNR estimation unit estimates the SNR of the input image; and The approximation calculation unit calculates the relationship between the SNR of the noise image and the corresponding optimal normalization factor by repeatedly pairing noise learning images with different SNRs and corresponding forward learning images based on different normalization factors used in the normalization unit. The standardization unit determines the optimal standardization factor for the input image based on the SNR estimated by the SNR estimation unit and the calculated approximation relationship, so as to standardize the input image with the determined optimal standardization factor.

2. The magnetic resonance imaging device according to claim 1, characterized in that, The approximation calculation unit multiplies the SNR of the approximation by an adjustment coefficient for adjusting the noise reduction intensity to produce an approximation.

3. A noise reduction method, comprising using a CNN learned by pairing a standardized noisy image containing noise with a corresponding standardized positive solution image with no noise or less noise than the noisy image, to reduce the noise of the input image of the CNN in order to generate a noise-reduced image, characterized in that, It includes the following steps: The input image is standardized using a standardization factor. The noise reduction method applies the CNN to the standardized input image for noise reduction processing. It also includes the following steps: Estimate the SNR of the input image; and By repeatedly pairing noisy learning images with different SNRs and their corresponding forward learning images based on different sets of normalization factors used in the normalization step, the difference between the noise-reduced image generated from the corresponding noise learning image and the corresponding forward learning image in the noise reduction process is minimized, thereby approximating the relationship between the SNR of the noisy image and the corresponding optimal normalization factor. Based on the estimated SNR of the input image, the optimal normalization factor is determined for the input image according to the calculated approximation relationship, and the input image is normalized with the determined optimal normalization factor.

4. The noise reduction method according to claim 3, characterized in that, The input image is an image obtained by a magnetic resonance imaging device.

5. An image processing apparatus comprising a CNN trained using pairs of a normalized noisy image containing noise and corresponding normalized positive resolution images with no noise or less noise than the noisy image, for reducing noise in the input image of the CNN to generate a noise-reduced image, characterized in that it includes a noise reduction unit. The noise reduction unit has: The standardization department standardizes the input image using a standardization factor. The noise reduction unit performs noise reduction processing on the standardized input image using the CNN. The noise reduction unit also includes: An SNR estimation unit estimates the SNR of the input image; and The approximation calculation unit calculates the relationship between the SNR of the noise image and the corresponding optimal normalization factor by repeatedly pairing noise learning images with different SNRs and corresponding forward learning images based on different normalization factors used in the normalization unit. The standardization unit determines the optimal standardization factor for the input image based on the SNR estimated by the SNR estimation unit and the calculated approximation relationship, so as to standardize the input image with the determined optimal standardization factor.

Images