Denoising Medical Images Using Machine Learning

Abstract

A method for denoising medical images. The medical image is rectified or normalised using a noise map that defines estimates of one or more statistical parameters for each pixel of the medical image. The rectified medical image is then processed using machine learning methods to produce a denoised medical image.

Description

[Technical field]

[0001] The present invention relates to the field of medical imaging, and in particular to denoising medical images. [Background technology]

[0002] Medical imaging is of increasing interest among medical professionals to (non-invasively) assist in the evaluation and / or diagnosis of the condition of a subject or patient under examination. Various forms of medical imaging methods or modalities are known in the art and employ invasive or non-invasive imaging techniques. Examples include computed tomography (CT) or x-ray imaging, magnetic resonance (MR) imaging, (intravenous) ultrasound imaging, positron emission tomography (PET) imaging, optical coherence tomography, transesophageal echocardiography, etc.

[0003] A continuing concern with medical images is noise and artifacts. When assessing a subject's condition, noise and artifacts can interfere with the identification of potentially important features of the subject (e.g., by obscuring them or making them difficult to identify) and can be mistaken for diagnostically relevant features. Thus, there is a continuing desire to reduce the amount of artifacts in medical images.

[0004] One recently developed technique for accurate denoising of medical images is to use appropriately trained machine learning methods to perform the denoising. However, there are many parameters that can be changed during the medical imaging process. For example, in a CT scan, different forms of reconstruction filters can be used to generate a CT medical image. Machine learning methods tend to overfit to the training data and therefore often fail to generalize to a wide range of real-world parameter settings (which are not necessarily sampled in the training). Summary of the Invention [Problem to be solved by the invention]

[0005] Therefore, there is a need for improved techniques for medical image denoising. [Means for solving the problem]

[0006] The invention is defined by the claims.

[0007] According to an embodiment of an aspect of the present invention, a computer-implemented method for denoising a medical image to generate a denoised medical image is provided.

[0008] The computer-implemented method includes obtaining a medical image formed from a plurality of pixels, obtaining a noise map including estimated measurements of statistical parameters for each pixel of the medical image, correcting the medical image using the noise map to generate a corrected medical image, and processing the corrected medical image using a machine learning method to generate a denoised medical image.

[0009] The proposed invention proposes a technique whereby a global (i.e. image-wide) noise level normalization / equalization of medical images is performed based on statistical parameters of the noise in the medical images. The normalized / equalized medical images are then processed using machine learning methods, for example for further noise reduction.

[0010] This allows the machine learning method to consistently process medical images that have already been normalized or equalized based on local / global conditions. This means that the images provided to the machine learning method are at a consistent noise level and / or have been equalized (e.g., to decorrelate the noise). This allows the machine learning method to be trained on normalized / equalized medical images, which means that the problem of overfitting is avoided.

[0011] Thereby, the proposed approach provides an improved and more accurate way to denoise medical images.

[0012] Correcting the medical image may include dividing the medical image by the noise map, which may be a pixel-by-pixel or point-by-point division.

[0013] In some embodiments, processing the corrected medical image includes processing the corrected medical image using a machine learning method to generate a predicted noise image, the predicted noise image representing the amount of noise predicted at each pixel of the corrected medical image; multiplying the corrected medical image by a noise map to generate a calibrated predicted noise image; and subtracting the calibrated predicted noise image or a scaled version of the calibrated predicted noise image from the medical image to generate a denoised medical image.

[0014] In some examples, processing the modified medical image includes inputting the modified medical image into a machine learning method and receiving a denoised medical image as an output from the machine learning method. The denoised medical image output by the machine learning method may not be calibrated (e.g., the medical image is divided by a noise map to modify the medical image). Thus, the denoised medical image output by the machine learning method may be recalibrated using the noise map. In particular, the inverse (or reverse) of the modification made to the medical image to generate the modified medical image may be applied to the denoised medical image to recalibrate the denoised medical image. This may include, for example, multiplying the denoised medical image by the noise map (e.g., pixel by pixel) to generate a recalibrated denoised medical image.

[0015] The machine learning method may include or be a neural network, which provides an accurate and reliable approach for processing the corrected medical images to perform the denoising process.

[0016] The machine learning method is preferably a residual (output) machine learning method, such as a residual neural network. In the context of the present disclosure, a residual machine learning method is a method that processes a corrected medical image to provide a predicted noise image that indicates the amount of noise expected in each pixel of the corrected medical image.

[0017] The use of residual machine learning methods is advantageous because it makes the output of the neural network more reliable. Specifically, the predicted noisy image can be assumed to be within the range of the output the machine learning method was trained on (because noise only has a limited range of estimates). Non-residual or direct machine learning methods (e.g., methods that directly infer the denoised medical image) are more likely to lead to predicted denoised images outside the range the machine learning method was trained on.

[0018] The noise map may provide an estimate of the standard deviation or variance of the noise for each pixel of the medical image.

[0019] In another embodiment, the noise map provides, for each pixel in the medical image, an estimated correlation between the noise of the pixel and the noise of one or more neighboring pixels.

[0020] In another embodiment, the medical image is one of a plurality of medical images produced by a multi-channel imaging process and representing the same scene, and the noise map provides, for each pixel of the medical image, an estimated measure of covariance or correlation between the noise of the pixel and the noise of a corresponding pixel of another medical image of the plurality of medical images.

[0021] This approach allows crosstalk between channels of a multi-channel imaging process to be more effectively reduced or taken into account when denoising a medical image (of a multi-channel imaging process). It will be appreciated that the medical image of each channel, i.e., each of the multiple images, may be processed separately using the methods described herein.

[0022] In some examples, acquiring the medical image includes acquiring a first medical image, processing the first medical image using a frequency filter to acquire a filtered medical image having values ​​within a predetermined frequency range, and setting the first filtered medical image as the medical image.

[0023] In at least one embodiment, the method further includes processing the first medical image to obtain a second filtered medical image having values ​​within a second, different, predetermined frequency range (e.g., not including any of the first frequency ranges), and combining the second filtered medical image and the denoised medical image to generate a denoised first medical image.

[0024] In a preferred embodiment, the medical images are medical images reconstructed from raw data using a first reconstruction algorithm, and the machine learning method is a machine learning method trained using a training data set including one or more training images reconstructed from raw data using a second, different reconstruction algorithm. Such an embodiment recognizes that different reconstruction filters have different noise characteristics. By equalizing the medical images processed using the machine learning method, the machine learning method trained using medical images generated with different reconstruction filters can still be used with high confidence.

[0025] In some embodiments, the machine learning method is preferably trained using a training dataset that includes one or more training images, each of which is modified with a corresponding noise map in the same manner as the medical images that are to be processed using the machine learning method, which improves the relevance and reliability of the machine learning method.

[0026] The medical images may be computed tomography medical images. It has been recognized that the proposed approach is particularly useful for medical images that may be generated using different reconstruction filters, and is therefore particularly useful for use with computed tomography images where a variety of reconstruction filters may be used.

[0027] Also proposed is a computer program product comprising computer program code means which, when executed on a computing device having a processing system, causes the processing system to perform all the steps of any of the methods described herein.

[0028] Also proposed is a processing system configured to denoise a medical image to generate a denoised medical image, the processing system being configured to obtain a medical image formed from a plurality of pixels, obtain a noise map comprising estimated measurements of statistical parameters for each pixel of the medical image, modify the medical image using the noise map to generate a modified medical image, and process the modified medical image using machine learning methods to generate the denoised medical image.

[0029] There is also provided a system comprising the processing system described above and a medical imaging system configured to generate medical images and provide the generated medical images to the processing system.

[0030] These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter.

[0031] For a better understanding of the present invention and to show more clearly how it may be carried into effect, reference will now be made, by way of example only, to the accompanying drawings in which: [Brief description of the drawings]

[0032] [Figure 1] 1 is a flow chart illustrating a method according to an embodiment. [Diagram 2]1 is a flow chart illustrating a process used in the method. [Diagram 3] 1 is a flow chart illustrating a method according to an embodiment. [Figure 4] The operation of the proposed method is shown. [Diagram 5] 1 illustrates a processing system according to an embodiment. [Figure 6] 1 illustrates a system according to an embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0033] The present invention will now be described with reference to the drawings.

[0034] It should be understood that the detailed description and specific examples, while indicating exemplary embodiments of the devices, systems, and methods, are for purposes of illustration only and are not intended to limit the scope of the invention. These and other features, aspects, and advantages of the devices, systems, and methods of the present invention will become better understood from the following description, the appended claims, and the accompanying drawings. It should be understood that the figures are schematic representations only and are not drawn to scale. It should also be understood that the same reference numerals are used throughout the figures to indicate the same or similar parts.

[0035] The present invention provides a technique for denoising medical images. A noise map that defines estimates of one or more statistical parameters for each pixel of the medical image is used to correct or normalise the medical image. The corrected medical image is then processed using machine learning methods to produce a denoised medical image.

[0036] The embodiment is based on the recognition that the reconstruction filters or algorithms used to generate medical images result in differently reconstructed medical image noise with different statistical characteristics. This reduces the efficiency of the denoising machine learning method, since it may not have been trained on images generated using the same reconstruction filter. By correcting the medical images using estimates of statistical parameters, each medical image can have a virtually normalized statistical level of noise. This greatly improves the consistency, and therefore the accuracy, of the denoising machine learning method.

[0037] The proposed concept can be used in any medical imaging system that can be used in a wide range of clinical environments.

[0038] In the context of this disclosure, a medical image is an image obtained using a medical imaging modality such as an X-ray image, a CT (Computed Tomography) image, a PET (Positron Emission Tomography) image, a MR (Magnetic Resonance) image, or an ultrasound image. Other forms of medical images will be apparent to those skilled in the art.

[0039] FIG. 1 is a flow chart illustrating a method 100 for denoising a medical image 105 according to an embodiment.

[0040] Method 100 may be performed, for example, on a single image. In another example, method 100 may be performed on one or more (e.g., each) of a plurality of images. The plurality of medical images may be generated by a multi-channel imaging process, such that each image represents a different channel and may represent the same scene. In other words, each of the plurality of images may represent the same anatomical region.

[0041] The method 100 includes acquiring 110 a medical image to be denoised. The medical image is formed of a number of pixels and may be two-dimensional (2D) or three-dimensional (3D). The pixels of a 3D image may be labelled as voxels.

[0042] Step 110 may itself involve generating a medical image, for example using a suitably configured medical imaging device. In other embodiments, step 110 involves retrieving an already generated medical image, for example from a medical imaging device and / or a memory or storage unit.

[0043] The method 100 further comprises a step 120 of obtaining a noise map 107 of the medical image. The noise map comprises estimated measurements of statistical parameters of the noise for each pixel of the medical image. In other words, the noise map defines, for each pixel of the medical image, an estimated value of the statistical parameters of the noise for that pixel. It is emphasized that the noise map provides an estimated statistical measurement of the noise, not a specific intensity at a pixel of the medical image.

[0044] The noise map may include a dedicated estimated measurement for each pixel of the medical image (e.g., formed of as many pixels as there are in the medical image). In another embodiment, the noise map may include estimated measurements for groups of two or more pixels, such that a single estimated measurement may represent a statistical parameter of the noise for multiple pixels. Thus, a single estimated measurement may represent a statistical parameter of the noise for each region of the medical image.

[0045] In some embodiments, the noise map may include an estimate of the noise variance and / or noise standard deviation for each pixel of the medical image. Techniques for generating such noise maps are established in the art. Some examples are disclosed in U.S. Patent having Patent No. 10,984,564(B2). Other examples include U.S. Patent No. 9,1591,22(B2) filed November 12, 2012, entitled "Image domain de-noising," U.S. Patent Application Publication No. 2016 / 0140725(A1), filed June 26, 2014, entitled "Methods of utilizing image noise information," and U.S. Patent No. 8,938,110(B2), filed October 29, 2015, entitled "Enhanced image data / dose reduction."

[0046] In another example, the noise map may contain (for each pixel / region) an estimated measure of the correlation between the noise of that pixel / region and the noise of its neighboring pixels / regions. Any (coarse) method for noise removal known in the art can be used to generate a (coarse) estimate of the noise of an image or image region. From such noise estimates, it is possible to directly estimate the local noise correlation by performing standard correlation analysis.

[0047] As yet another example, the noise map may include a noise probability density function (noise PDF) for each pixel of the medical image or for each region of the medical image, which may be used to determine or predict, for example, the noise variance and / or standard deviation of the pixel / region.

[0048] As yet another example, the noise map may include the covariance or correlation of the noise of a pixel with the noise of a corresponding pixel of another medical image, which may represent the same scene and form part of a set of medical images generated, for example, by a multi-channel image process.

[0049] Step 120 may itself involve generating a noise map, for example using any of the techniques described above, or step 120 may involve retrieving an already generated noise map of the medical image, for example from a memory or storage unit.

[0050] The method 100 also includes a step 130 of correcting the medical image using the noise map to generate a corrected medical image.

[0051] Step 130 may include normalizing the medical image, for example using a statistical measure of the estimated noise, with the aim of effectively normalizing or normalizing this statistical measure across the entire medical image.

[0052] As an example, step 130 involves dividing the medical image by the noise map. Specifically, the value of each pixel of the medical image is divided by an estimated measurement of the statistical parameter of the noise for that pixel provided by the noise map. In this way, a point-by-point or pixel-by-pixel division of the medical image by the estimated value of the statistical parameter of the noise can be performed.

[0053] As another example, step 130 may include processing the noise map to obtain statistical information about the noise map and / or the medical image. This statistical information may then be used to modify the medical image.

[0054] For example, the noise map can be processed to determine the average estimated deviation of the measurements of the statistical parameters (e.g., for the entire image or for different sections of the image). The medical image can then be divided by the average estimated deviation, e.g., by section for an average that represents the average of a particular section, or by the entire image if the average represents the average of the entire image.

[0055] As yet another example, the noise map can be processed to determine or predict the shape of the noise in the frequency domain, which can then be used to frequency domain filter the medical image, for example to normalize the frequency of the noise in the medical image.

[0056] As another example, step 130 includes performing a decorrelation process on the noise of the medical image, i.e., a process of decorrelating the noise of the medical image. Thus, step 130 includes (spatially) decorrelating the noise of the medical image to generate a corrected medical image.

[0057] This can be done using a noise map that indicates a measure of correlation between different pixels and / or regions. As another example, this can be done by processing the image using one or more means of approximate deconvolution or transformation of the PDF by passing the values ​​through some non-linear function, using the noise map to provide a noise PDF for each pixel and / or region.

[0058] Those skilled in the art will readily recognize various other techniques for correcting the medical image in step 130. More generally, step 130 is a step of correcting the medical image such that the noise response across the corrected medical image is more uniform than across the (original) medical image.

[0059] The output of step 130 is a corrected medical image 135 .

[0060] Next, the method 100 performs a process 140 of processing the corrected medical image using machine learning methods to generate a denoised medical image.

[0061] FIG. 1 illustrates one embodiment for performing the process 140.

[0062] In this example, the process 140 involves directly predicting or inferring the denoised medical image 145 from the rectified medical image 135. Thus, the machine learning method receives as input the rectified medical image and provides as output the denoised medical image. Thus, the machine learning method can be trained to generate a clean or denoised image from a (noisy) medical image.

[0063] In some embodiments, the machine learning method outputs only an uncalibrated denoised medical image, which is (re)calibrated using the noise map, and in particular, the uncalibrated denoised medical image is modified using the uncalibrated denoised medical image and the noise map to perform the reverse of the procedure performed in step 130 to generate a denoised medical image.

[0064] As an example, if step 130 involves performing a pixel-by-pixel division of the medical image by the noise map, then the uncalibrated denoised medical image is subjected to a pixel-by-pixel multiplication with the noise map.

[0065] (Re)calibration may not be necessary in some situations or embodiments, such as when step 130 involves decorrelating noise in the medical image.

[0066] FIG. 2 illustrates another embodiment for performing process 140, which is labeled as process 240 for distinction.

[0067] The process 240 includes a step 241 of inputting the corrected medical image 135 into a machine learning method configured to generate a predicted noise image 245. The predicted noise image is an image that contains the same number of pixels as the medical image and indicates (for each pixel of the medical image) a measure of the estimated / predicted noise of that pixel.

[0068] Next, the process 240 performs step 242 of multiplying the predicted noise image 245 by the noise map, i.e., renormalizing the estimated noise image 245 to generate a calibrated predicted noise image 247.

[0069] Of course, if the medical image was not divided by the noise map in step 130, then step 242 is modified to reverse the procedure performed in step 130 using the estimated noise image instead of the corrected noise image (i.e., the reverse of procedure 130 is performed in step 242 using the estimated noise image 245 and the noise map 107 as inputs). Thus, the inverse (or reverse) of the modifications made to the medical image to generate the corrected medical image can be applied to the estimated noise image to calibrate the estimated noise image.

[0070] The process 240 then subtracts the calibrated predicted noise image from the medical image 105 in step 243 to generate a denoised medical image 145 .

[0071] In some embodiments, the calibrated predicted noise image is weighted (e.g., scaled down) before being subtracted from the medical image. This approach recognizes that some clinicians prefer to retain some (non-zero) level of noise in the medical image to reduce the artificial appearance of the denoised medical image (which may be distracting to the clinician if not addressed).

[0072] Thus, step 243 involves subtracting a scaled version of the calibrated predicted noise image from the medical image. The scaled version may be calculated by multiplying the value of each pixel of the calibrated predicted noise image by a predetermined value, where the predetermined value is between 0 and 1, for example between 0.25 and 0.75.

[0073] Thus, in process 240, the machine learning method is configured to receive as input the modified medical image and provide as output a predicted noise image that indicates, for each pixel, the amount of noise that is predicted or estimated at that pixel, which may be on a pixel-by-pixel basis.

[0074] In the proposed approach, each machine learning method processes medical images that have already been normalized or equalized (e.g., decorrelated) based on statistical information about the noise. This means that the images provided as input to the machine learning method are already at a consistent and / or decorrelated noise level. This allows the machine learning method to be trained with normalized medical images, reducing the risk of overfitting (e.g., to a particular reconstruction filter or noise level).

[0075] 1, the method 100 further includes a step 150 of controlling a user interface to provide a visual representation of the denoised medical image 145 output by the process 140. The user interface is a display, such as a monitor.

[0076] In some embodiments, the method 100 includes storing 155 the denoised medical image, for example in a memory or storage unit. Step 155 may include storing the denoised medical image in an electronic medical record of a subject of the denoised medical image.

[0077] FIG. 3 illustrates a method 300 according to another embodiment.

[0078] The method 300 differs from the above method 100 in that the step 110 of acquiring a medical image includes a step 311 of acquiring a first medical image 305, a step 312 of processing the first medical image using a frequency filter (filtering according to a predetermined frequency range) to acquire a first filtered medical image 315, and a step 313 of setting the first filtered medical image as the medical image. The frequency filter is, for example, a high-pass filter or a band-pass filter.

[0079] In this manner, the medical image that is denoised may be a frequency filtered portion of the medical image.

[0080] The frequency filtered portion is preferably a high frequency portion of the medical image, i.e. a portion of the medical image having a frequency higher than a predetermined frequency value, recognizing that noise in medical images may typically be high frequency and that denoising only the high frequency portions of the medical image may provide more efficient and improved denoising.

[0081] In some embodiments, step 110 further includes step 314 of processing the first medical image using another frequency filter to obtain a second filtered medical image 316. The second filtered medical image has a different frequency range than the first filtered medical image. In one example, the second filtered medical image is a portion of the (original) medical image that is not in a predetermined frequency range.

[0082] For example, if the frequency filter is a high-pass filter, the second filtered medical image is a low-pass filtered portion of the medical image. That is, the medical image is processed using a low-pass filter. The low-pass filter and the high-pass filter can have the same cut-off frequency. Thus, the medical image is effectively split (by steps 312 and 314) into a high-frequency medical image and a low-frequency medical image to form the first filtered medical image and the second filtered medical image, respectively.

[0083] An alternative to the optional step 314 to generate the second filtered medical image 316 is to subtract the first filtered medical image 315 from the first medical image 305 .

[0084] Similarly, if step 314 includes processing the first medical image using a filter, such as a low pass filter, step 312 may be modified to include subtracting the second filtered medical image 316 from the first medical image 305 to generate the first filtered medical image 315.

[0085] The method 300 may further include a step 360 of combining the second filtered medical image 316 with the denoised medical image 145 output by the process 140 to recreate a denoised version 349 of the first medical image. Step 360 may include simply summing the second filtered medical image with the denoised medical image output by the process 140.

[0086] Similar to method 100, method 300 may include step 150 of controlling a user interface to provide a visual representation of the denoised medical image 145 output by process 140 and / or step 155 of saving the denoised medical image, for example in a memory or storage unit. Method 300 is adapted accordingly.

[0087] The proposed approach has been found to be particularly effective when the machine learning method is a residual learning method, which produces as output a predicted noise image containing the same number of pixels as the medical image and provides (for each pixel) a measure of the noise expected for that pixel.

[0088] Testing and analysis of the proposed technique have shown that it generalizes well across different reconstruction filters and also to higher noise levels (e.g., resulting from the use of low radiation dose levels) than those seen during training of machine learning methods, thus improving the robustness of machine learning in practice.

[0089] Some methods for generating noise maps recognize that the actual computed noise map may not reflect the statistical variation in noise levels introduced by different reconstruction filters, but rather may reflect differences in noise levels due to other factors, such as widespread systematic errors, the use of low radiation doses, etc.

[0090] In this case, in order to generalize the proposed approach to different systems, the noise map obtained in step 120 of the above method can simply be scaled by a global scaling factor before being used to process the medical image.

[0091] For example, in the field of CT imaging, it is well known that reconstruction filters usually contain two parts or functions: a ramp function and an additional Modulation Transfer Function (MTF). A more complete understanding is given in Thorsten M. Buzug (2008) in "Computed Tomography" (Springer-Verlag, Berlin, Heidelberg). In general, the various reconstruction filters used in CT imaging differ only by their MTF modulation part.

[0092] This global scaling factor is the (Ramp*MTF) of the reconstruction filter used.2 It can be calculated from the area under the curve. MTF is the modulation transfer function. In fact, the noise variance is (Ramp*MTF) 2 It is believed to scale linearly with the area underneath. The (square root) ratio of the area of ​​the reconstruction filter used to reconstruct the medical image being processed to the corresponding area of ​​the reconstruction filter used in training the machine learning method provides an appropriate scaling factor for the noise map.

[0093] For example, if the generation of the noise map is image-based or takes into account reconstruction filters (used to generate medical images), scaling of the noise map is not necessary.

[0094] In summary, this technique allows the use of a pre-trained machine learning method for application to an unknown reconstruction filter by scaling the corresponding standard deviation noise map by the above-mentioned scaling factor, which is merely a function of the two filters and thus incurs virtually no additional computational cost or overhead and therefore does not require retraining of the machine learning method.

[0095] The action of the proposed concept on medical image denoising is shown in Fig. 4. Two denoised CT head images that are partially occluded (outside the white circled areas in each image) are provided in Fig. 4. Each CT image is generated by processing the same medical CT image acquired at a dose level of 25%.

[0096] The first denoised CT head image 410 was generated using conventional denoising machine learning methods (in particular convolutional neural networks) without normalising or correcting the medical image using a noise map as proposed in the present invention.

[0097] A second denoised CT head image 420 was generated using the proposed denoising technique, in particular a denoising technique that normalised the statistical parameters of the noise before being processed using machine learning methods.

[0098] In both cases, the machine learning method was trained using CT images acquired at a dose level of 25% and reconstructed using a first reconstruction filter. The medical CT images that were later denoised (to produce head image 420) were generated using a second, different reconstruction filter with less noise suppression. Different reconstruction filters have different noise characteristics but also provide other benefits, such as providing sharper or smoother images and / or highlighting different anatomical features. Thus, an operator can select a noisier or less noisy reconstruction filter depending on clinical preference.

[0099] It is evident that the proposed method produces denoised images with less noise, which makes the method more robust against reconstruction filters with higher noise levels than when training the machine learning method used in the denoising method.

[0100] Note that the noise map obtained during the generation of the second denoised CT head image was scaled by a factor of 1.4 to more accurately reflect the higher noise level in the image as a result of the different reconstruction filters, in accordance with the technique described above. The value of 1.4 was chosen based on the mechanism for determining the scaling factor described above.

[0101] The proposed embodiment uses machine learning methods, which are any self-training algorithms that process input data to generate or predict output data, where the input data comprises modified medical images and the output data comprises either denoised medical images or predicted noisy images.

[0102] Machine learning methods suitable for use in the present invention will be apparent to those skilled in the art. Examples of suitable machine learning methods include decision tree algorithms and artificial neural networks. Other machine learning methods such as logistic regression, support vector machines, or naive Bayes models are suitable alternatives.

[0103] The structure of an artificial neural network (or simply a neural network) is inspired by the human brain. A neural network is made up of layers, each layer containing multiple neurons. Each neuron contains a mathematical operation. In particular, each neuron may contain a different weighted combination of a single type of transformation (e.g. the same type of transformation, such as sigmoid, but with different weightings). In the process of processing input data, the mathematical operation of each neuron is performed on the input data to generate a numerical output, and the output of each layer of the neural network is fed in turn to the next layer. The final layer provides the output.

[0104] In the present disclosure, embodiments are particularly advantageous when using residual learning machine learning methods such as residual neural networks, which differ from traditional neural networks in that they can use skip connections, such that, for example, the output of a layer can skip one or more layers (i.e., all outputs of any given layer need not be provided consecutively as inputs to the next layer).

[0105] Training methods for machine learning methods are well known. Typically, such methods include obtaining a training data set that includes training input data entries and corresponding training output data entries. An initialized machine learning method is applied to each input data entry to generate a predicted output data entry. The error between the predicted output data entry and the corresponding training output data entry is used to correct the machine learning method. This process is repeated until the error converges, i.e., until the predicted output data entry is sufficiently similar to the training output data entry (e.g., ±1%). This is commonly known as a supervised learning technique.

[0106] For example, when a machine learning method is formed from a neural network, the (weightings) of the mathematical operations of each neuron are modified until the error converges. Known methods for modifying neural networks include the gradient descent algorithm, the backpropagation algorithm, etc.

[0107] The training input data entries correspond to modified medical image examples. In particular, each training input data entry should include a medical image that has been pre-processed or modified using a (medical image) noise map, e.g., a medical image that has been subjected to step 130 as described with reference to method 100. This improves the reliability and accuracy of the machine learning method used in method 100.

[0108] The training output data entries correspond to denoised medical images and / or example noisy images. Information regarding the reconstruction filters used to generate the example medical images for the training input data entries may be stored, for example, to facilitate scaling of the noise map, as described above.

[0109] For example, the training input data entries are medical images with artificial noise added (modified) and the training output data entries are medical images without artificial noise added (modified).

[0110] Any of the methods proposed herein can be performed by the imaging system itself (i.e., the system that produces the medical images), by a processing system on the same premises as the imaging system, by a mobile device (smartphone, tablet, laptop, etc.), or using a distributed processing system, i.e., the "cloud."

[0111] Those skilled in the art will be able to readily develop a processing system to perform the methods described herein, and each step of the flowchart thus represents a different action performed by the processing system and may be executed by a corresponding module of the processing system.

[0112] Thus, the embodiments employ a processing system. The processing system can be implemented in a variety of ways using software and / or hardware to perform the various functions required. The processor is one example of a processing system that employs one or more microprocessors that are programmed using software (e.g., microcode) to perform the required functions. However, the processing system may be implemented without regard to the employment of a processor, and may be implemented as a combination of dedicated hardware to perform some functions and processors (e.g., one or more programmed microprocessors and associated circuitry) to perform other functions.

[0113] Examples of processing system components that may be employed in various embodiments of the present disclosure include, but are not limited to, conventional microprocessors, application specific integrated circuits (ASICs), and field programmable gate arrays (FPGAs).

[0114] In various implementations, a processor or processing system may be associated with one or more storage media, such as volatile and non-volatile computer memory, such as RAM, PROM, EPROM, and EEPROM. The storage media may be encoded with one or more programs that, when executed on the one or more processors and / or processing systems, perform the necessary functions. The various storage media may be fixed within the processor or processing system or may be transportable such that one or more programs stored thereon may be loaded into the processor.

[0115] As a further example, Figure 5 illustrates an example processing system 500 in which one or more portions of the embodiments may be used. The various operations described above utilize the functionality of the processing system 500. For example, one or more portions of the system for denoising medical images may be incorporated into any of the elements, modules, applications, and / or components described herein. In this regard, it should be understood that the functional blocks of the system may operate on a single computer or may be distributed across several computers or locations (e.g., connected via the Internet).

[0116] The processing system 500 may include, but is not limited to, a PC, a workstation, a laptop, a PDA, a palm device, a server, storage, etc. In general, with respect to a hardware architecture, the processing system 500 includes one or more processors 501, a memory 502, and one or more I / O devices 507 communicatively coupled via a local interface (not shown). The local interface may be, for example, but not limited to, one or more buses or other wired or wireless connections as known in the art. The local interface may have controllers, buffers (caches), drivers, repeaters, and receivers to enable communication. Additionally, the local interface may include address, control, and / or data connections to enable appropriate communication between the aforementioned components.

[0117] Processor 501 is a hardware device that executes software that may be stored in memory 502. Processor 501 may be virtually any custom or commercially available processor, central processing unit (CPU), digital signal processor (DSP), or coprocessor of multiple processors associated with processing system 500, and processor 501 may be a semiconductor-based microprocessor (in the form of a microchip) or microprocessor.

[0118] The memory 502 may include any one or combination of volatile memory elements (e.g., random access memory (RAM), such as dynamic random access memory (DRAM), static random access memory (SRAM), etc.) and non-volatile memory elements (e.g., ROM, erasable programmable read only memory (EPROM), electronically erasable programmable read only memory EEPROM), programmable read only memory (PROM), tape, compact disk read only memory (CD-ROM), disk, diskette, cartridge, cassette, etc.). Additionally, the memory 502 may incorporate electronic, magnetic, optical, and / or other types of storage media. It is noted that the memory 502 may have a distributed architecture where various components are remote from each other but accessible by the processor 501.

[0119] The software in memory 502 may include one or more separate programs, each including an ordered list of executable instructions for implementing logical functions. The software in memory 502 includes a suitable operating system (O / S) 505, a compiler 504, source code 503, and one or more applications 506, according to an exemplary embodiment. As shown, the applications 506 include a number of functional components for implementing the features and operations of the exemplary embodiments. The applications 506 of the processing system 500 may represent various applications, computational units, logic, functional units, processes, operations, virtual entities, and / or modules, according to an exemplary embodiment, although the applications 506 are not intended to be limiting.

[0120] Operating system 505 controls the execution of other computer programs and provides scheduling, input / output control, file and data management, memory management, and communication control and related services. The inventors contemplate that application 506 for implementing the exemplary embodiment is applicable to any commercially available operating system.

[0121] The application 506 may be a source program, an executable program (object code), a script, or any other entity that includes a set of instructions to be executed. In the case of a source program, the program is typically translated via a compiler (such as compiler 504), assembler, interpreter, etc., which may or may not be included in memory 502, so that the program operates properly in conjunction with the O / S 505. Additionally, the application 506 may be written as an object-oriented programming language with classes of data and methods, or a procedural programming language (such as, but not limited to, C, C++, C#, Pascal, BASIC, API calls, HTML, XHTML, XML, ASP script, JavaScript, FORTRAN, COBOL, Perl, Java, ADA, .NET, etc.) with routines, subroutines, and / or functions.

[0122] The I / O devices 507 include input devices such as, but not limited to, a mouse, keyboard, scanner, microphone, camera, etc. Additionally, the I / O devices 507 also include output devices such as, but not limited to, a printer, display, etc. Finally, the I / O devices 507 also include devices that communicate both input and output such as, but not limited to, a NIC or modulator / demodulator (for accessing remote devices, other files, devices, systems, or networks), radio frequency (RF) or other transceivers, telephone interfaces, bridges, routers, etc. The I / O devices 507 also include components for communicating over various networks such as the Internet or an intranet.

[0123] If the processing system 500 is a PC, workstation, intelligent device, etc., the software in memory 502 may further include a Basic Input / Output System (BIOS) (omitted for simplicity). The BIOS is a set of essential software routines that initializes and tests the hardware on power-up, starts the O / S 505, and supports data transfers between hardware devices. The BIOS is stored in some type of read-only memory, such as ROM, PROM, EPROM, EEPROM, etc., so that it is executed when the processing system 500 is powered up.

[0124] During operation of the processing system 500, the processor 501 is configured to execute software stored in the memory 502, to transfer data to and from the memory 502, and to generally control the operation of the processing system 500 in accordance with the software. Applications 506 and O / S 505, in whole or in part, are read by the processor 501 and, in some cases, buffered within the processor 501 before being executed.

[0125] It should be noted that if the application 506 is implemented in software, the application 506 may be stored on virtually any computer-readable medium for use with or in connection with any computer-related system or method. In the context of this document, a computer-readable medium may be an electronic, magnetic, optical, or other physical device or means that can store or retain a computer program for use with or in connection with a computer-related system or method.

[0126] The application 506 may be embodied in any computer-readable medium for use with or in connection with an instruction execution system, apparatus, or device (such as a computer-based system, processor-based system, or other system that can fetch instructions from and execute instructions from an instruction execution system, apparatus, or device). In the context of this document, a "computer-readable medium" may be any means that can store, communicate, propagate, or transfer a program for use with or in connection with an instruction execution system, apparatus, or device. Computer-readable media include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, devices, or propagation media.

[0127] 6 shows a schematic diagram of a system 600 including an imaging system 602 and a processing system 612. Here, the imaging system is a CT scanner configured for spectral (multi-energy) imaging. However, other forms of imaging systems can be used. The processing system 612 may be embodied as the processing system 500 described above.

[0128] The illustrated imaging system 602 generally includes a stationary gantry 604 and a rotating gantry 606. The rotating gantry is rotatably supported by the stationary gantry 604 and rotates about an examination region 608 about the z-axis. A subject support 610, such as a couch, supports an object or subject within the examination region 608.

[0129] A radiation source 612, such as an x-ray tube, is rotatably supported by and rotates with the rotating gantry 606 to emit radiation that traverses the examination region 608. In one example, the radiation source 612 includes a single broad spectrum x-ray tube. In another example, the radiation source 612 includes a single x-ray tube configured to switch between at least two different emission voltages (e.g., 80 kVp and 640 kVp) during a scan. In yet another example, the radiation source 612 includes two or more x-ray tubes configured to emit radiation having different average spectra. In yet another example, the radiation source 612 includes a combination thereof.

[0130] A radiation sensitive detector array 614 subtends an arc for an angle on opposite sides of the examination region 608 of the radiation source 612. The radiation sensitive detector array 614 detects radiation traversing the examination region 608 and generates electrical signals (projection data) indicative thereof. Whereas the radiation source 612 includes a single broad spectrum x-ray tube, the radiation sensitive detector array 612 includes energy resolving detectors (e.g., direct conversion photon counting detectors, at least two sets of scintillators (multilayers) with different spectral sensitivities, etc.). In kVp switching and multi-tube configurations, the detector array 614 can include single layer detectors, direct conversion photon counting detectors, and / or multilayer detectors. Direct conversion photon counting detectors include a conversion material such as CdTe, CdZnTe, Si, Ge, GaAs, or other direct conversion materials. An example of a multi-layer detector is a double-decker detector, such as that described in U.S. Patent No. 7,968,853 (B2), filed April 60, 2006, and entitled "Double Decker Detector for Spectral CT," which is incorporated herein by reference in its entirety.

[0131] A reconstructor 616 receives the spectral projection data from the detector array 614 and reconstructs spectral volumetric image data, such as sCCTA image data, high energy images, low energy images, photoelectric images, Compton scatter images, iodine images, calcium images, virtual non-contrast images, bone images, soft tissue images, and / or other underlying material images. The reconstructor 616 can also reconstruct non-spectral volumetric image data, such as by combining the spectral projection data and / or the spectral volumetric image data. Typically, the spectral projection data and / or the spectral volumetric image data includes data for at least two different energies and / or energy ranges.

[0132] In this manner, the reconstructor 616 generates or reconstructs a medical image.

[0133] Here, the processing system 618 serves as an operator console. The console 618 includes a human readable output device such as a monitor and input devices such as a keyboard, mouse, etc. Software resident on the console 618 allows an operator to interact with and operate the scanner 602, such as via a graphical user interface (GUI). The console 618 further includes a processor 620 (e.g., a microprocessor, controller, central processing unit, etc.) and a computer readable storage medium 622 (excluding non-transitory media), including transitory media such as physical memory devices. The computer readable storage medium 622 includes instructions 624 for denoising the generated medical images, i.e., includes a medical image denoiser 625. The processor 620 is configured to execute the instructions 624. The processor 620 may additionally be configured to execute one or more computer readable instructions carried by a carrier wave, signal, and / or other transitory media. In an alternative embodiment, the processor 620 and the computer readable storage medium 622 are part of a separate processing system, separate from the processing system 618 .

[0134] It will be understood that the disclosed methods are preferably computer-implemented methods. Therefore, the concept of a computer program is also proposed, which includes code means for carrying out any of the described methods when executed on a processing system such as a computer. Thus, different parts, lines or blocks of code of a computer program according to an embodiment may be executed by a processing system or a computer to carry out any of the methods described herein. In some alternative implementations, the functions shown in the block diagrams or flow charts may occur out of the order shown in the figures. For example, two blocks shown in succession may in fact be executed substantially simultaneously, or the blocks may be executed in the reverse order depending on the functionality involved.

[0135] Variations of the disclosed embodiments can be understood and effected by those skilled in the art in implementing the claimed invention, from a study of the drawings, the disclosure and the appended claims. In the claims, the term "comprising" does not exclude other elements or steps, and the singular elements do not exclude a plurality. A single processor or other unit may fulfill the functions of several items recited in the claims. The mere fact that certain means are recited in mutually different dependent claims does not indicate that a combination of these means cannot be used to advantage. In the case where a computer program is described above, it can be stored / distributed on any suitable medium, such as an optical storage medium or a solid-state medium, provided together with or as part of other hardware, but can also be distributed in other forms, such as via the Internet or other wired or wireless communication systems. It should be noted that when the term "adapted to" is used in the claims or description, the term "adapted to" is intended to be equivalent to the term "configured to". Any reference signs in the claims should not be interpreted as limiting the scope.

Claims

1. A computer-implemented method for removing noise from a medical image to generate a denoised medical image, comprising: obtaining the medical image formed from a plurality of pixels; obtaining a noise map including estimated measured values of statistical parameters of each pixel of the medical image; correcting the medical image using the noise map to generate a corrected medical image; processing the corrected medical image using a machine learning method to generate the denoised medical image; wherein the step of correcting the medical image includes dividing the medical image by the noise map; a computer-implemented method.

2. The step of processing the corrected medical image includes: processing the corrected medical image using the machine learning method to generate a predicted noise image, wherein the predicted noise image represents the amount of noise predicted at each pixel of the corrected medical image; multiplying the corrected medical image by the noise map to generate a calibrated predicted noise image; subtracting the calibrated predicted noise image or a scaled version of the calibrated predicted noise image from the medical image to generate the denoised medical image; The computer-implemented method according to claim 1.

3. The step of processing the corrected medical image includes inputting the corrected medical image into the machine learning method and receiving, as an output from the machine learning method, the denoised medical image. The computer-implemented method according to claim 1.

4. The machine learning method includes a neural network. The computer-implemented method according to claim 1.

5. The noise map provides an estimate of the standard deviation or variance of the noise of each pixel of the medical image. The computer-implemented method according to claim 1.

6. The noise map provides an estimated correlation between the noise of each pixel of the medical image and the noise of one or more adjacent pixels for each pixel of the medical image. The computer-implemented method according to claim 1.

7. The medical image is generated by a multi-channel imaging process and is one of a plurality of medical images representing the same scene. The noise map provides an estimated measurement value of the covariance or correlation between the noise of each pixel of the medical image and the noise of the corresponding pixel of another medical image among the plurality of medical images, according to the computer-implemented method of claim 1.

8. The step of acquiring the medical image includes a step of acquiring a first medical image, a step of processing the first medical image using a frequency filter to acquire a first filtered medical image having values within a predetermined frequency range, a step of setting the first filtered medical image as the medical image according to the computer-implemented method of claim 1.

9. A step of processing the first medical image to acquire a second filtered medical image having values within a second different predetermined frequency range, and a step of combining the second filtered medical image and the noise-reduced medical image to generate a noise-reduced first medical image further included in the computer-implemented method of claim 8.

10. The medical image is a medical image reconstructed from raw data using a first reconstruction algorithm, The machine learning method is a machine learning method trained using a training data set including one or more training images reconstructed from raw data using a second different reconstruction algorithm, according to the computer-implemented method of claim 1.

11. The computer-implemented method of claim 1, wherein the medical image is a computed tomography medical image.

12. A computer program including computer program code means which, when executed on a computing device having a processing system, causes the processing system to perform all steps of the computer-implemented method of claim 1.

13. A processing system for removing noise from a medical image to generate a noise-reduced medical image, the processing system acquires the medical image formed from a plurality of pixels, acquires a noise map including estimated measurement values of statistical parameters of each pixel of the medical image, modifies the medical image using the noise map to generate a modified medical image To generate the noise-removed medical image, the corrected medical image is processed using a machine learning method configured to The step of correcting the medical image includes dividing the medical image by the noise map processing system

14. The processing system according to claim 13, a medical imaging system that generates the medical image and provides the generated medical image to the processing system A system comprising

Images