METHOD FOR PROCESSING PARAMETERS OF A MACHINE LEARNING PROCEDURE AND RECONSTRUCTION PROCEDURES

Description

[0001] The invention relates to a method for processing parameters of a machine learning method for providing a correction data set for motion correction of a CT image data set of an object moving during its acquisition. Furthermore, the invention relates to a reconstruction method that uses the processed machine learning method and to a device.

[0002] In imaging using a computed tomography device (CT scanner), an X-ray source and / or a detector move along a predefined path, particularly a circular or helical path, relative to an object under examination. A (CT) image dataset is acquired from different (acquisition) angles. Particularly in the field of medical diagnostics, in which a patient's organ, such as the heart, is examined, the acquisition time for the CT image dataset is comparatively long for the entire angular range to be acquired, so that the organ, or at least part of the organ, moves during the acquisition of the CT image dataset.

[0003] As a result, motion artifacts of the moving object appear in a (reconstruction) image dataset reconstructed from the CT image dataset. Such motion artifacts, however, make a reliable diagnosis difficult. Motion artifacts represent a source of diagnostic error. For example, the structure of cardiac arteries is not clearly recognizable in a reconstruction image dataset presented as an image representation (image) due to the movement of the heart during acquisition of the CT image dataset. In particular, they are displayed in a blurred, distorted, and / or deformed manner.

[0004] Hardware-related solutions are known, such as using a second detector and a second X-ray source, or increasing the rotation speed of a gantry supporting the X-ray source and detector. However, these solutions are comparatively expensive.

[0005] Furthermore, methods are known that determine a correction data set for the acquired CT image data set in order to reduce motion artifacts. For example, a method for reducing motion artifacts in a reconstruction of coronary vessels is described in "Motion compensation in the region of the coronary arteries based on partial angle reconstruction from short scan data" by J. Hahn et al. in "Medical Physics", 2017 November, Vol. 44(11), pp. 5795-5813. In this method, the motion along a centerline of the corresponding coronary vessel is modeled using a motion vector field whose parameters are estimated using a cost function.

[0006] For example, in "Nonrigid registration-based coronary artery motion correction for cardiac computed tomography" by R. Bhagalia et al. in "Medical Physics", 2012 July, Vol. 39(7), pp. 4245-4254, it is further disclosed that motion compensation of coronary vessels is realized by means of a three-dimensional deformation ("warping") of a series of partial reconstructions based on estimated motion vector fields.

[0007] Furthermore, a method is known from "Cardiac motion corrected iterative conebeam CT reconstruction using a semi-automatic minimum cost path-based coronary centerline extraction" by AA Isola et al. in "Computerized medical imaging and graphics", 2012 April, Vol. 36(3), pp. 215-226, which uses motion vector fields determined from a number of centerlines of the corresponding coronary vessels to generate motion artifact-free reconstructions of coronary vessels.

[0008] What these methods have in common is that a motion vector field is determined using an algorithm as a correction data set for the object's movement during acquisition of the CT image data set. However, these methods are comparatively time-consuming, particularly due to the determination of the correction data set using the respective algorithm, so that a comparatively long time elapses between the acquisition of the CT image data set and the provision of the reconstruction image data set, motion-reduced using the correction data set, for diagnosis.

[0009] DE 10 2009 007 236 A1 discloses a method for scanning a moving examination subject with a CT system. Data is acquired during a rotating movement of a transmitter / receiver pair around the examination subject. Furthermore, cross-sectional images of the examination subject are obtained from the data using an iterative algorithm. The iterative algorithm takes into account motion information relating to the movement of the examination subject during data acquisition.

[0010] From the publication DE 10 2011 007 529 A1, a method, a radiotherapy system and a combination of CT system and radiotherapy system for determining a movement profile of an object moving within an examination object with a radiation source-detector system that is displaceable relative to the examination object is known, wherein the following method steps are carried out: Scanning the object under investigation in the area of ​​the moving object during a relative displacement of the emitter-detector system to the object under investigation and generating a pixel data set with attenuation values ​​over time, removing stationary structures from the pixel data set, determining an attenuation value induced by the moving object in each detector row at a plurality of successive times of the scanning and forming a 3D data set from the values ​​of the attenuation maximum of the detector rows across the detector rows and the readout times of the scanning, and determining at least one of the values ​​from the following list from the result data set: frequency and / or phase and / or amplitude of the movement of the object, location area of ​​the object during the scanning, position of the object at a predetermined phase of the movement.

[0011] From the document DE 10 2011 083 647 A1 a method for generating a motion-compensated CT image data set is known, wherein: a projection data set of a CT system is acquired from a predetermined movement phase and a projection angle range which allows the immediate reconstruction of a CT image data set, the movement field is determined iteratively by: multiple reconstruction of the one CT image data set with a first image resolution using a motion-compensating reconstruction method using a first analytical reconstruction algorithm and different movement fields each from a plurality of location-specific movement vectors, and determination of the movement field using at least one predetermined constraint, and reconstruction of a final CT image data set with a second image resolution using a motion-compensating reconstruction method on the basis of a second reconstruction algorithm and the determined movement field.

[0012] It is also known from WO 2017 / 223560 A1 to use a machine learning method, for example "deep learning" for an artificial neural network, for image reconstruction.

[0013] In "Deep Learning Computed Tomography" by Würfl, T. et al. in "International Conference on Medical Image Computing and Computer-Assisted Intervention," Springer, Cham, 2016, it is revealed that image reconstruction can be achieved using a neural network. In particular, it is shown that the neural network learns filters and weights to reduce reconstruction errors in CT images with a limited angular range.

[0014] US 2016 / 000383 A1 discloses a method for detecting movement during medical imaging. First, the acquisition of image data is initiated and physiological signals from a patient are recorded. Based on these signals, it is determined whether movement of the patient is likely and, if necessary, the acquisition of the image data is modified.

[0015] The invention is based on the object of providing a method that enables the comparatively rapid provision of a correction data set for the corresponding CT image data set. Furthermore, a reconstruction method for generating a reconstruction image data set is to be provided. Furthermore, a device for implementing one or both of the methods is to be provided.

[0016] This object is achieved according to the invention with regard to the method by the features of claim 1. With regard to the reconstruction method, the object is achieved according to the invention by the features of claim 5 and with regard to the device, the object is achieved according to the invention by the features of claim 6. Advantageous further developments and embodiments are the subject of the subclaims.

[0017] For this purpose, parameters of a machine learning method are processed in a process. This machine learning method serves to provide a correction data set for motion correction of a CT image data set of a moving object during its acquisition. The processing of the parameters of the machine learning method is also referred to as training. The parameters are initialized before training begins, for example, with random numbers. The CT image data set represents an intensity curve of the X-rays incident on a detector from different (acquisition) angles. The CT image data set is, in particular, a so-called sinogram.

[0018] According to the method, a training dataset with a number of reference image datasets is first provided and introduced for initiation (feeding) into the machine learning process. In other words, the training dataset is fed into the machine learning process as input dataset.

[0019] A reference image dataset is a previously known CT image dataset that is provided for processing the parameters of the machine learning process. For example, clinical images of an object such as a heart or images of a so-called phantom are used for this purpose. Alternatively, such reference image datasets are generated, for example, through a calculation or by simulating a moving object.

[0020] According to the method, an iteration with at least one iteration step is further performed. In each of the iteration steps, a correction data set is determined for each of the reference image data sets using the machine learning method, and a result of a cost function is determined. The cost function depends on the correction data sets determined in this iteration step.

[0021] A correction data set is designed in particular as a motion vector field which represents a movement of the object in an image data set.

[0022] The cost function is a mathematical function that represents the deviation (difference) between quantities or data determined using the machine learning method (the forecast) and the corresponding target values ​​or actual values. In other words, the cost function quantifies the error in the forecast of the machine learning method. For example, the cost function is the mean square deviation between the determined quantities and the corresponding target values.

[0023] Depending on the result of the cost function, the iteration is terminated or a parameter of the machine learning process is changed. In particular, a threshold is specified for this purpose. The threshold thus represents the maximum permissible error of the prediction made using the machine learning process. If the result of the cost function falls below the threshold, the iteration for the respective parameter is terminated; otherwise, another iteration step is performed with the changed parameter. After the iteration is terminated, the iteration is performed for the next reference image dataset.

[0024] Particularly preferably, the parameter is changed in such a way that the result of the cost function is reduced. In particular, the parameter change is carried out according to the so-called "stochastic gradient descent" method. Consequently, the error of a forecast determined using the machine learning method with the changed parameter is reduced.

[0025] For example, the training dataset can be divided into sub-training datasets. Then, an iteration is performed for each of the sub-training datasets. In each iteration step of the respective iteration, the parameters are processed analogously depending on the result of the cost function, or the iteration of the respective sub-training dataset is terminated. This is particularly advantageous when a device executing the machine learning method has a relatively small memory, so that not all reference image datasets can be considered in a single iteration or this would result in a comparatively long processing time.

[0026] The machine learning process, also referred to as machine learning, is implemented, for example, as a learning or prediction algorithm, which is executed by a controller. This machine learning process advantageously recognizes patterns and regularities after completing a training phase with previously known training data sets. Thus, the previously known reference image data sets of the input training data set and the correction data sets intended for them are not memorized, stored, or saved, but rather generalizations and / or correlations are determined. Thus, an unknown data set can be assessed or evaluated using the machine learning process, which is specifically designed as a prediction algorithm.After the training phase has ended, it is thus possible to evaluate an input data set relatively quickly using the trained machine learning method and to output the corresponding results of this evaluation—in this case, a correction data set for motion correction of an input CT image data set. Using such a machine learning method to provide a correction data set as the result of evaluating a CT image data set fed to it is advantageously done in a relatively short time. Consequently, CT image data sets can be corrected relatively quickly and made available for diagnosis.

[0027] According to the invention, in a first variant of the method, as an alternative to the second variant described below, the method comprises using differences between the determined correction data sets and previously known training correction data sets to determine the result of the cost function. The training correction data sets are stored in the training data set and assigned to the corresponding reference image data set. In particular, a mean value of the squared differences is used as the result of the cost function; alternatively, however, the differences are used as arguments of a mathematical function. In this first variant, the training data set therefore has previously known training correction data sets, each of which is assigned to the corresponding reference data set.These training correction data sets were determined, for example, from the corresponding reference data set using one of the methods mentioned above as state of the art, or alternatively were determined by means of a simulation of the reference data set and the associated training correction data set.

[0028] Alternatively, in a second variant of the method, the method according to the invention comprises that, in order to determine the result of the cost function, each of the reference image data sets is first deformed (adapted) according to the assigned and determined correction data set; in particular, the intensity profile (intensity distribution) of the reference image data set is changed according to the determined correction data set. For example, rigid or non-rigid methods known from the prior art are used for this purpose. A reconstruction image data set is then created from each of the deformed reference image data sets. In particular, an image representation (image) of the object is created as a reconstruction image data set using a known reconstruction method, for example so-called filtered back-projection. Alternatively, the reconstruction image data set is, for example, a deformed sinogram.

[0029] For example, the cost function can be determined using the reconstruction image datasets. For this purpose, these are compared, for example, with training reconstruction image datasets stored in the training dataset and assigned to the corresponding reference image datasets. Analogously to the first variant, these training reconstruction image datasets are determined from the corresponding reference dataset, for example, using one of the methods mentioned above as state of the art. Analogously, differences between the determined reconstruction datasets and previously known training reconstruction datasets are used to determine the result of the cost function.

[0030] Alternatively, according to an advantageous development, a variable dependent on the object's movement is determined from the reconstruction image data sets and used to determine the result of the cost function. For example, an average of the variables determined from the reconstruction image data sets is used as the result. Therefore, only reference data sets are used, but no other data sets associated with these reference data sets. As an alternative to this variable, a restriction to positive values ​​(positivity constraint) of the reference data sets deformed using the correction data sets is used as the cost function.

[0031] In a suitable further development, an entropy is determined from the corresponding reconstruction image dataset or a selected portion thereof and used as the quantity determined from the reconstruction image dataset. If the reconstruction image dataset is configured as an image, the region of interest that is more relevant for diagnosis is expediently selected as part of the image.

[0032] Alternatively or in addition to entropy, a gradient of entropy and / or the restriction to positive values ​​described above are used as the quantity determined from the reconstruction image data set.

[0033] According to a particularly suitable embodiment of the method, an (artificial) neural network, in particular a convolutional neural network (CNN), is used as the machine learning method. A neural network has an input layer into which data sets are fed, at least one intermediate layer (hidden layer), and an output layer. The network or layers each comprise a number of so-called neurons that are connected to one another. Each of the connections further has a weighting that determines the influence of a respective transmitting neuron on the corresponding target neuron (receiving neuron), whereby the individual neurons are each defined, for example, as a mathematical function. The parameters that are processed by means of the method are, in particular, the weightings.During parameter editing, several weightings are changed in a single step, with the parameters being appropriately modified in such a way that the forecast error or evaluation error is reduced, thus reducing the result of the cost function due to the parameter change. For example, the weightings are changed using the so-called "gradient descent" method or the so-called "stochastic gradient descent" method.

[0034] A convolutional neural network (CNN) is particularly suitable for analyzing image datasets. Alternatively, a neural network with fully connected neural layers (FCN) is used. However, a convolutional neural network has comparatively few parameters compared to a neural network with fully connected neural layers, so the computational effort of a convolutional neural network is comparatively low.

[0035] The previously known reference image datasets are integrated into the neural network based on training. The advantage of a neural network is that it can learn and operate without instruction. The previously known reference image datasets obtained through training and the correction datasets corresponding to the object's movements are integrated into the topology of the neural network—that is, the structure of the neural network and the connections between the neurons—so that when a dataset to be evaluated is input, it is examined for previously known features.

[0036] In a reconstruction method for generating a motion-corrected reconstruction image data set for a CT image data set of an object that is moving during its acquisition, a CT image data set of a moving object is acquired and provided in a first step using a computer tomography device.

[0037] In a subsequent second step, the CT image data set is evaluated using a machine learning method that was trained using a method in one of the variants described above.

[0038] As a result of the evaluation, a correction data set corresponding to the object's movement is provided. Using this data set and the CT image data set, a motion-corrected reconstruction image data set is generated. For this purpose, the CT image data set is deformed according to the correction data set. Alternatively, the correction data set is used as an initial data set for an iterative determination using a method known from the prior art, for example, one of those presented above, so that the duration of these methods is shortened. In summary, due to the use of a machine learning method, in particular one designed as a neural network, the motion-corrected reconstruction data set is provided comparatively quickly.

[0039] For example, the machine learning method can be further trained using this CT image data set and, if necessary, the correction data set.

[0040] According to a suitable embodiment, a device comprises a controller that is provided and configured to carry out the method in one of the variants described above and / or to carry out the reconstruction method. The device also expediently comprises a computer-readable medium on which the machine learning method is stored.

[0041] An embodiment of the invention is explained in more detail below with reference to a drawing. In the drawings: FIG 1 shows a schematic flow diagram of a method for processing parameters of a machine learning method in which an iteration is carried out, wherein in each of its iteration steps a result of a cost function is determined, FIG 2 shows a schematic flow diagram of an alternative embodiment of the method according to FIG 1 with an alternative determination of the result of the cost function, FIG 3 in a schematic flow diagram a reconstruction method for generating a motion-corrected reconstruction image data set for a CT image data set, FIG 4 a neural network for carrying out a machine learning method, and FIG 5 in a schematic front view a computed tomography device which is connected to a device designed as a computer, the device having a controller.

[0042] Corresponding parts are provided with the same reference numerals in all figures.

[0043] In FIG 1 A method for processing parameters P of a machine learning method L for providing a correction data set KD for a movement correction of a CT image data set CB of an object 2 moving during its recording is schematically shown using a flow chart.

[0044] In a first step, initiation 100, a training data set T with a number of reference image data sets RD is provided according to the method and introduced into the machine learning process L. In a subsequent second step, an iteration 110 is performed. The iteration 110 comprises at least one iteration step.

[0045] In each of the iteration steps of iteration 110, a correction data set KD is first determined for each of the reference image data sets RD using the machine learning method L, which is designed as an (artificial) neural network (step 111). The reference image data sets RD are then deformed using the respective correction data sets KD in a step 112, and a (motion-) corrected reconstruction image data set RK is created from each of the deformed reference image data sets RD.

[0046] Furthermore, in a further step 113, a result E of a cost function CF is determined. This depends on the correction data sets KD determined in this iteration step. For this purpose, a value G is determined from each of the reconstruction image data sets RK created using the correction data sets KD and the reference image data sets RD and is used to determine the result E of the cost function CF. The value G is an entropy of an image determined from the corresponding reconstruction image data set RK, for example the entropy of an image determined from the deformed reference image data set RD using a filtered back-projection. An average of the values ​​G determined from the reconstruction image data sets RK is used here as the result E of the cost function CF.

[0047] In a comparison step 114, this result E is compared with a, in particular, predetermined, threshold value S. The threshold value S serves to limit the maximum permissible error or deviation of a prediction made using the machine learning method L. If the result E is greater than the threshold value S, i.e., the deviation or error is too large, parameters P, in this case weights of the neural network, are changed (step 115). The weights P of the neural network L are changed using the so-called "stochastic gradient descent" method. However, if the result E falls below the threshold value S, iteration 110 is terminated (step 116). In this way, the parameters P of the machine learning method L and a topology of the neural network are processed.

[0048] The FIG 2 shows an alternative design of the procedure. This alternative design differs from the one in FIG 1 The variant shown is that the training data set TD, in addition to the reference image data sets RD, has a training correction data set TKD assigned to the corresponding reference image data set RD. In step 113, a difference between the corresponding data sets KD and TKD is determined from the correction data sets KD determined in step 111 and the assigned training correction data set TKD and is used to determine the result E of the cost function. Here, a mean of the squared differences is used as the result E of the cost function CF.

[0049] Step 112 takes place in the variant according to FIG 2 therefore does not take place. The other steps of the method, in particular the termination of iteration 110 when the result E of the cost function CF falls below the threshold value S (step 116), or otherwise the change of the parameters P (step 115), are carried out analogously.

[0050] In the FIG 3 A reconstruction method for generating a motion-corrected reconstruction image data set RK for a CT image data set CB of an object 2 that is moving during its acquisition is shown. In a first step 300, the CT image data set CB of the moving object 2 is acquired and provided using a computed tomography device 4.

[0051] In a subsequent second step 310, the CT image data set CB is evaluated using the machine learning method L ( FIG 4 ). This is in accordance with one of the procedures of the FIG 1 or the FIG 2 been trained.

[0052] As a result of the evaluation, a correction data set KD corresponding to the movement of object 2 is provided (step 320). Using this data set and the CT image data set CB, a motion-corrected reconstruction image data set RK is generated in a step 330. For this purpose, the CT image data set CB is deformed according to the correction data set KD, i.e., the CT image data set CB is changed according to the determined correction data set KD. In this way, the movement of object 2 is corrected during the acquisition of the CT image data set CB, and motion artifacts in the reconstruction image data set RK formed as an image are avoided. As a result, an error in an interpretation or diagnosis based on the now motion-corrected image is avoided.

[0053] In FIG 4 A neural network L is shown schematically and in a highly simplified manner, which is used in the methods of FIG 1 bis FIG 3 is used. The neural network L has an input layer 6 into which data records are fed, at least one hidden layer 8 and an output layer 10. Layers 6, 8 and 10 each comprise a number of unspecified neurons which are connected to one another and are each defined as a mathematical function. For the purpose of improved clarity, only nine connections to neurons are shown as an example in FIG. 10, but in this embodiment the neurons of input layer 6 are each connected to all neurons of intermediate layer 8 and the neurons of intermediate layer 8 are each connected to all neurons of output layer 10. Each of these connections has a weighting P which determines the influence of a respective sending neuron on the corresponding target neuron.

[0054] Input layer 6 is determined from the CT image dataset CB or, during the training phase, from the reference image datasets RD. For example, each pixel or a matrix of pixels of the image dataset fed in as a sinogram corresponds to a neuron of input layer 6.

[0055] The features of a movement of object 2 in the reference image data RD are integrated into mathematical functions and / or weights of the neurons of the neural network. In other words, the motion features of the reference image data RD are trained into the neural network and incorporated into its topology ( FIG 4 ) is integrated. In this way, the at least one intermediate layer (hidden layer) 8 extracts and recognizes characteristic movement features of the input data set. When evaluating an input data set, a movement is thus detected, and the corresponding correction data set KD for correcting these movement features is output as result E to the (output) neurons of the output layer 10.

[0056] IN FIG 5 1 shows a computed tomography device 4. This device has a holding frame 12 for an annular rotating ring 14, which is also referred to as a gantry. The rotating ring 14 is rotatably mounted on the holding frame 12. The rotating ring 14 carries an X-ray system 16 with an X-ray source 18 for generating X-rays and with an (X-ray) detector 20 for the X-rays arranged opposite the source. A patient bed 24 can be moved into or into the tunnel-shaped recording area 22 formed by the annular rotating ring 14. To record a CT image data set, the object 2 to be examined, in particular a patient, is placed on the patient bed 22 and moved into the recording area 22. When the rotating ring 14 rotates, the X-ray system 10 moves around the object 2 located on the patient bed 24 in a circle.Consequently, a CT image data set CB of the object 2 is acquired from different (spatial, recording) angles, whereby the CT image data set CB represents an intensity profile of the incident X-ray radiation according to the different (recording) angles.

[0057] To acquire the CT image data set CB recorded (acquired) by the detector 20, the detector 20 is connected for signal transmission purposes to a device 26 designed as a reconstruction computer. This device has a controller 28 for carrying out the FIG 1 or FIG 2 procedure shown and the one in FIG 3 shown reconstruction method, as well as a computer-readable medium 30 designed as a hard disk, on which the machine learning method L designed as a neural network of the FIG 1 bis FIG 4 is deposited.

[0058] The invention is not limited to the exemplary embodiments described above. Rather, other variants of the invention can also be derived therefrom by those skilled in the art without departing from the subject matter of the invention. In particular, all individual features described in connection with the exemplary embodiments can also be combined with one another in other ways without departing from the subject matter of the invention.

Claims

1. Computer-implemented method for processing parameters (P) of a machine learning method (L) for providing a correction data record (KD) for a movement correction of a CT image data record (CB) of an object (2) moving during its capture, - wherein a training data record (T) with a number of reference image data records (RD) is provided (100) for introduction into the machine learning method (L), - wherein an iteration (110) is carried out, wherein a correction data record (KD) is determined (111) by means of the machine learning method (L) in each of their iteration steps for each of the reference image data records (RD) and a result (E) of a cost function (CF) dependent on the correction data records (KD) determined in this iteration step is determined (113), and - wherein the iteration (110) is terminated (116) or a parameter (P) of the machine learning method (L) is modified (115) as a function (114) of the result (E) of the cost function (CF), wherein in order to determine the result (E) of the cost function (CF), differences from the determined correction data records (KD) and previously known training correction data records (TKD) are used, wherein the training correction data records (TKD) are stored in the training data record (T) and assigned to the corresponding reference image data records (RD) or wherein in order to determine the result (E) of the cost function (CF), each of the reference image data records (RD) is deformed (112) according to the assigned correction data record (KD) and a reconstruction image data record (RK) is created from the deformed reference image data record (RD).

2. Method according to claim 1, wherein in order to determine the result (E) of the cost function (CF), each of the reference image data records (RD) is deformed according to the assigned correction data record (KD) and a reconstruction image data record (RK) is created from the deformed reference image data record (RD), characterised in that a variable (G) dependent on the movement of the object (2) is determined from the reconstruction image data records (RK) and used to determine the result (E) of the cost function (CF).

3. Method according to claim 2, wherein in order to determine the result (E) of the cost function (CF), each of the reference image data records (RD) is deformed according to the assigned correction data record (KD) and a reconstruction image data record (RK) is created from the deformed reference image data record (RD), characterised in that an entropy is determined from the reconstruction image data record (RK) or a selected part thereof and used as the variable (G) determined from the reconstruction image data record (RK).

4. Method according to one of claims 1 to 3, characterised in that a neural network, in particular a convolutional neural network, is used as a machine learning method (L).

5. Computer-implemented reconstruction method for generating a movement-corrected reconstruction image data record (RK) for a CT image data record (CB) of an object (2) moving during its capture, - wherein the CT image data record (CB) of the moving object (2) is captured and provided by means of a computed tomography device (4), - wherein the CT image data record (CB) is evaluated by means of a machine learning method (L) trained according to one of claims 1 to 4, - wherein a correction data record (KD) corresponding to the movement of the object (2) is provided as a result of the evaluation, and - wherein a movement-corrected reconstruction image data record (RK) is generated by means of the correction data record (KD) and the CT image data record (CB).

6. Device (26) with a controller (28) for implementing the method according to one of claims 1 to 4, and / or for carrying out the reconstruction method according to claim 5.