DE102021210804A1 - Method and apparatus for providing planning information in preparation for a breast biopsy - Google Patents

Abstract

A method for providing planning information (33) for preparing a breast biopsy, comprising the steps of: - receiving medical image data (34) relating to a breast (35) to be biopsied, - determining the planning information (33) using a processing algorithm (36), which processes the image data (34) as input data and is trained by machine learning, the planning information (33) containing a presumably suitable parameter value for at least one biopsy parameter (37) of the breast biopsy and/or for at least one imaging parameter (38) of a medical imaging method for monitoring the breast biopsy specifies,- providing the planning information (33).

Description

The invention relates to a method for providing planning information for preparing a breast biopsy. In addition, the invention relates to a device, a computer program and a computer-readable data carrier.

During a breast biopsy, tissue is removed from the breast using a needle inserted into the breast. This serves in particular to be able to make a robust decision on the basis of the tissue sample, for example whether a tumor is present, when a conspicuous tissue change or lesion is detected, for example as part of an imaging examination method.

It is essential here that a sufficient amount of tissue is removed from the correct position, while at the same time the consequences of the intervention, such as scarring, should be avoided as far as possible. To achieve this, breast biopsies are often monitored using a medical imaging technique. In this case, the monitoring is frequently carried out by means of X-ray images, with stereoscopic X-ray images, tomosynthesis or computed tomography being able to be used in order to obtain depth information. Alternatively, e.g. ultrasound imaging can be performed.

In principle, the goals stated above can be achieved through suitable monitoring in combination with a suitable selection of the biopsy parameters, i.e. in particular the needle thickness, which is also referred to as “gauge”, and the puncture geometry. However, the problem here is that, for example, a doctor who performs such an intervention requires a relatively large amount of experience and training in order to reliably select optimal parameters for the biopsy, even if optimal prior information and medical image data are available to him to prepare for the intervention . The parameters that are perceived as being optimal are also subjective to a degree, resulting in a certain variability in the outcome of the intervention. As a result, the goal of avoiding tissue injuries or scarring as far as possible cannot always be optimally achieved.

In order to achieve optimal monitoring of the intervention using medical imaging, it is typically necessary to adapt the imaging parameters or the evaluation of the image data for the respective breast biopsy, for example to avoid metal artifacts caused by the biopsy needle in the image. This also requires a lot of experience or extensive training of the operating personnel and a subjective variability of the selected parameters, which can lead to poorer imaging overall, can hardly be avoided. If optimal image quality is not achieved during the monitoring of the intervention, this can in turn lead to a deterioration in the outcome of the intervention.

The above points result in the problem that medical personnel with sufficient experience and training cannot be made available everywhere in order to achieve an optimal biopsy result. In addition, a very high level of training is generally required to train medical personnel for optimal breast biopsies.

The invention is therefore based on the object of specifying a possibility of supporting the preparation of a breast biopsy and thereby being able to achieve better biopsy results with the person performing the procedure having the same level of training.

• - Empfangen von medizinischen Bilddaten, die eine zu biopsierende Brust betreffen,
• - Ermitteln der Planungsinformation durch einen Verarbeitungsalgorithmus, der die Bilddaten als Eingangsdaten verarbeitet und durch Maschinenlernen trainiert ist, wobei die Planungsinformation einen voraussichtlich geeigneten Parameterwert für wenigstens einen Biopsieparameter der Brustbiopsie und/oder für wenigstens einen Bildgebungsparameter eines medizinisches Bildgebungsverfahrens zur Überwachung der Brustbiopsie vorgibt,
• - Bereitstellen der Planungsinformation.

The object is achieved according to the invention by a method for providing planning information for preparing a breast biopsy, which comprises the following steps:

• - receiving medical image data relating to a breast to be biopsied,
• - Determination of the planning information by a processing algorithm that processes the image data as input data and is trained by machine learning, the planning information specifying a presumably suitable parameter value for at least one biopsy parameter of the breast biopsy and/or for at least one imaging parameter of a medical imaging method for monitoring the breast biopsy,
• - Providing the planning information.

It is therefore proposed to automatically determine suitable parameter values for biopsy parameters or a medical imaging method used for monitoring on the basis of medical image data of the breast to be biopsied, which is typically available anyway, in particular if medical imaging is to be used to monitor the intervention. These parameters can, for example, be output to a user, for example a doctor who performed the procedure, who can implement this suggestion or adapt it based on his own assessment. In this case, the processing algorithm acts as a kind of expert system.

In addition or as an alternative, the planning information can also be passed on to another facility, for example For example, it can be used to monitor the intervention or the positioning and orientation of the needle before the intervention and issue appropriate information, and/or to configure the imaging device. The presumably suitable parameter values are preferably checked by appropriately trained medical personnel and, if necessary, adjusted according to their own assessment.

The method according to the invention thus makes it possible to avoid misjudgments of a biopsy situation, which under certain circumstances could lead to an incorrect choice of parameter values. Quality assurance is achieved and even if the proposed parameter values are modified, treatment planning is typically significantly accelerated since at least a robust starting point for parameter selection is provided.

The method is preferably implemented as a computer-implemented method, which can be implemented, for example, on a workstation computer, a computing device used in the context of medical imaging, a server, or also decentrally, for example as a cloud solution.

The planning information can be provided to a user, for example via an output device, in particular graphically. For example, a model of the breast to be biopsied with a puncture geometry of the biopsy needle shown here can be displayed via the output device in order to visualize the proposed biopsy parameters for a user. Alternatively or additionally, the planning information can also be provided via one or more interfaces, for example in order to suitably configure a medical imaging device or another monitoring device. In principle, provision via a software interface to another program is also possible.

The method according to the invention serves to prepare for the breast biopsy. The method thus ends in particular before the start of the surgical intervention, ie in particular before a biopsy needle is inserted into the breast. In principle it is possible that at least parts of the planning information are also used within the scope of the biopsy itself, for example within the scope of imaging or image data processing within the scope of the biopsy. However, the method according to the invention itself manages without any steps that represent or include a surgical or therapeutic treatment of the human body or a diagnostic method on the human body.

The processing algorithm can in particular be or include a neural network. In this case, the image data are processed in particular by a convolutional neural network or by a plurality of convolution layers, with the processing result being further processed in particular by at least one fully connected layer, which provides the planning information or the parameter values. Network architectures known for the neural network, such as a U-Net or DenseNet, can be used here.

As will be explained in more detail below with regard to the training method, the processing algorithm can be trained in particular by monitored learning. For example, experts can specify the respective suitable parameter values on the basis of medical image data, which then form a training data set together with the respective medical image data. If a sufficient number of training data sets are available, training methods known per se, for example error feedback, can then be used to train the processing algorithm.

The planning information can specify the presumably suitable parameter value for a presumably suitable needle size and/or a presumably suitable puncture position and/or a presumably suitable puncture angle of a biopsy needle as a biopsy parameter. With regard to the needle size, in particular its diameter can be specified. This can be specified as a "gauge", for example, with this size typically being between 11 and 18 for the biopsy needle used in the context of a breast biopsy. In addition or as an alternative, however, the length of the biopsy needle can also be specified or the like. By specifying a suitable needle size, it can be achieved that a sufficient amount of tissue can be removed without unnecessarily large puncture sites and thus unnecessarily severe tissue injuries or scarring resulting.

By specifying a suitable puncture position and a suitable puncture angle, it is ensured that the relevant area is reached by the needle, whereby joint variations of these parameters can on the one hand avoid having to pierce relatively dense tissue, which could potentially result in a deflection of the needle, and on the other hand metal artefacts can be avoided or reduced in treatment-accompanying imaging. the one stitch angle is specified in particular as a solid angle.

In addition or as an alternative to the parameters mentioned above, it may also be expedient under certain circumstances to propose a biopsy method to be used as part of the planning information. For example, you can choose between a fine needle biopsy, a fine needle aspiration cytology and a vacuum biopsy.

The planning information can additionally or alternatively specify an expected suitable acceleration voltage and/or x-ray dose for at least one planned x-ray exposure and/or an expected suitable processing parameter for artifact reduction and/or an expected suitable imaging geometry as an imaging parameter. This can be used in particular to reduce artifacts, in particular metal artifacts, which can be generated, for example, by compression plates used to compress the breast and/or the biopsy needle itself, through suitable parameterization of the imaging and/or through appropriately parameterized post-processing. For example, in the context of image processing, the segmentation of metal objects can be significantly improved by specifying appropriate processing parameters, since, for example, the position of compression plates typically remains unchanged and, especially when combined with the provision of the biopsy parameters explained above, the dimensions and puncture geometry of the biopsy needle used are at least approximately are known.

In this case, it is possible to specify imaging parameters, for example the acceleration voltage and/or the x-ray dose, for all recordings used in the context of imaging. However, it can be advantageous to specify, for example, different acceleration voltages and/or X-ray doses for different projections within the framework of imaging, for example in tomosynthesis, in order to open up a larger parameter space for optimizing the recording quality.

The processing algorithm can process the output data from at least one preprocessing algorithm or from at least two preprocessing algorithms as additional input data in order to determine the planning information. In this context, use can be made of the fact that algorithms are already known which can extract specific information from medical image data. For example, it is known to identify and segment tissue areas in which certain features occur with sufficient probability. For example, a probability can be determined for specific segments that they belong to a specific class, that is, for example, that a lesion or a suspicious calcification is present there. For example, such an algorithm is used in the Transpara® program from ScreenPoint Medical B.V. provided.

The use of a separate algorithm for pre-processing and for determining the corresponding output data makes it possible to use proven algorithms that have been validated in daily use in order to feed the processing algorithm with information of known importance that was extracted from the image data.

Since the preprocessing algorithm or the preprocessing algorithms can thus be trained and validated separately from the processing algorithm or are already available in the trained state, the processing algorithm can be formed, for example, by a neural network with considerably fewer neurons than if a corresponding functionality were to be implemented within the processing algorithm itself . However, this also reduces the number of free parameters of the preprocessing algorithm, which must be determined during the training. Since, moreover, a relatively clear number of suitable parameter values is typically to be output as the processing result, a good training result can be achieved with a relatively small number of training data sets without overfitting or overtraining the network.

In cases in which the processing algorithm includes convolution layers for processing the medical image data and at least one fully linked layer that follows, it is particularly possible that at least parts of the output data of at least one of the preprocessing algorithms are fed directly to one of the fully linked layers and thus do not go through the convolution layers. This can further reduce the complexity of the neural network, as a result of which the robustness of the training explained above can also be further improved if the number of training data sets is not too large.

The pre-processing algorithm or at least one of the pre-processing algorithms can be trained by a method of machine learning. In particular, the respective pre-processing algorithm can be formed by a neural network. In this case, architectures are preferably used that have one or more convolution layers and at least one downstream fully ver include bonded layer. For example, the network architectures already mentioned with regard to the processing algorithm can be used. The pre-processing algorithm is preferably also trained by monitored training, for example by error feedback.

The pre-processing algorithm or a respective one of the pre-processing algorithms can provide the processing algorithm with output data relating to breast density information relating to the tissue density of the breast to be biopsied and/or area information relating to a distinguished region recognized in the image data. The breast density information can, for example, describe the tissue density for each point or voxel in the medical image data. However, it is also possible for the breast to be subdivided into segments of different tissue density, with a corresponding segmentation thus being able to be provided breast density information.

The respective marked area can on the one hand be an area from which tissue is expected to be removed as part of the biopsy, or on the other hand be an area whose injury is to be avoided. A distinguished area can be identified, for example, by an algorithm that detects potential lesions, as explained above. The area information can describe the location of the area, for example as a segmentation. Additionally or alternatively, for example, a dimension and/or a probable class or morphology of the area can be output as part of the area information.

With regard to the further processing of the output data by the processing algorithm, it can be expedient to process information that relates to individual points or voxels or that describes a segmentation, together with the medical image data, using convolution layers of the processing algorithm. The processing result of the convolution layers can then be further processed by fully linked layers of the processing algorithm, optionally together with output data from at least one preprocessing algorithm, which relate to individual parameters, for example diameters or classes of marked areas.

The planning information can be provided to a user, after which a user input from this user is recorded, which confirms the planning information and/or relates to a change in the planning information provided, after which an additional training of the processing algorithm is carried out depending on the user input and the image data.

In some cases, for example, a doctor performing the biopsy may change the proposed biopsy parameters and/or imaging parameters. This may be the case if such a change is expected to further improve the biopsy. A change can also be desired if different parameterizations are expected to lead to at least a similarly good biopsy result, with a doctor or a group of doctors, for example in a specific country or a specific hospital chain, preferring a different parameterization than that suggested by the processing algorithm becomes.

In this case, these change requests can be recorded and used for further training, for example by forming a further training data record using the image data and the correspondingly adapted parameter values. This training data set can then be used together with other corresponding training data sets and/or training data sets that were already used in the previous training of the processing algorithm, in order to carry out further training iterations and thus on the one hand to further improve the robustness of the processing algorithm and on the other hand to adapt the processing algorithm to the wishes and custom adapt to specific user groups.

It is particularly possible here for the further training to take place only for a specific group of users, for example to adapt to their preferences. For example, a fully trained processing algorithm can be provided by a provider of the method or a device performing the method or a corresponding computer program, which can be further trained by the user, for example by the operator of a hospital chain, in order to adapt it to specific preferences.

A position and/or an orientation of a biopsy needle with respect to the breast to be biopsied can be detected by sensors before the biopsy needle pierces the breast, with auxiliary information being output for the user depending on the position and/or the orientation and the planning information provided . The sensory detection can be carried out, for example, by one or more cameras, but other systems for position and/or orientation detection can also be used, many of which are well known in the field of medical technology.

The auxiliary information can indicate, for example, a deviation in the position and/or orientation of the biopsy needle from the corresponding parameters according to the planning information. In particular, if certain limit values are exceeded by the deviation, a warning signal can be output optically and/or acoustically and/or a color of a displayed piece of information can be adjusted depending on the deviation of the position or orientation from the target position or orientation specified by the planning information, for example, so that a user can immediately recognize strong deviations.

In principle, it is possible to continue monitoring the position and/or orientation and the output of corresponding auxiliary information even after the biopsy needle has been inserted. However, since the method according to the invention preferably ends before the start of the actual intervention, such further monitoring after the insertion of the needle would preferably take place outside of the method according to the invention.

The processing algorithm can evaluate at least one additional parameter provided independently of the image data. The additional parameter can be taken from a patient file or database, for example. The additional parameter can include, for example, the age of the patient, information on previous interventions or diagnoses that have already been made, texts of findings and the like. For example, it is possible to evaluate findings texts beforehand using a further algorithm trained by machine learning in order to extract relevant information from them, which can then be provided as additional parameters.

The processing algorithm can determine the planning data as a function of further medical image data provided as additional input data, the further medical image data also relating to the breast to be biopsied, but recorded at a different point in time and/or by a different medical imaging device. A recording at a different point in time is understood to mean, in particular, a recording as part of separate medical imaging. In particular, the medical image data and the further medical image data can each be recorded during a chest compression, with the chest compression being released between the recordings. The further medical image data can be produced, for example, as part of a check-up before a change in the breast tissue, with the biopsy being used to clarify this change. In this case, the additional medical image data can make it easier in particular to locate the biopsy-relevant tissue area and thus, for example, to segment it as a distinguished area, as explained above.

X-ray data and/or three-dimensional image data or stereoscopic image data can be used as image data. Three-dimensional X-ray image data could be provided by a computer tomograph, for example. However, since the image data can be recorded immediately before the breast biopsy when the breast is already compressed, it is often not possible to carry out a complete computed tomography. For example, tomosynthesis can therefore be used to provide three-dimensional image data, or it can also be sufficient to use stereoscopic X-ray images. In principle, the method according to the invention can also be carried out with two-dimensional image data, but a considerable improvement in robustness and accuracy can be achieved by using at least stereoscopic image data.

As an alternative or in addition, the medical image data can also be recorded, for example, by means of ultrasound imaging or a magnetic resonance tomograph.

In addition to the method according to the invention for providing planning information, the invention relates to a method for training a method algorithm which is used to determine planning information for preparing a breast biopsy on the basis of medical image data provided as input data relating to a respective breast to be biopsied, the Planning information specifies a parameter value that is expected to be suitable for at least one biopsy parameter of the breast biopsy and/or for at least one imaging parameter of a medical imaging method for monitoring the breast biopsy, with monitored machine learning taking place on the basis of training data sets, the respective training data set containing at least the medical image data and a respective target value for the at least one biopsy parameter and/or the at least one imaging parameter.

Details on the structure of the processing algorithm have already been explained above. The training data sets can be provided, for example, by using the target values for the at least one bio psieparameter or the at least one imaging parameter are determined by human experts. However, other investigative methods can also be used. The training can take place, for example, by error feedback. Corresponding training approaches are already known from other application areas and will therefore not be explained in detail. Details on possible further input data of the processing algorithm, which can in particular also be present in the training data records, and on parameters to be specifically output have already been explained above.

For the case already explained above, in which the processing algorithm processes the output data of at least one preprocessing algorithm as additional input data, several training approaches are possible. On the one hand it is possible that the respective training data record already includes the corresponding initial data. On the other hand, however, it is also possible to determine these initial data from the medical image data in each case within the scope of the training by using the corresponding pre-processing algorithm. These approaches can also be mixed so that, for example, output data from at least one preprocessing algorithm is already contained in the training data set and output data from at least one further preprocessing algorithm is only determined during training by applying the respective preprocessing algorithm to the medical image data.

The method for training the processing algorithm can be used, on the one hand, for initial training of the processing algorithm, through which, for example, the suitability of the processing algorithm for the stated purpose is first achieved. Additionally or alternatively, the method for training the processing algorithm can also be used to retrain a processing algorithm that has already been trained, in order to further improve the quality and robustness of the planning information provided on the one hand due to an increasing amount of training data and/or on the other hand by additional training of the processing algorithm adapt provided planning information, for example, to the preferences of a specific user group, as already explained above.

The training dataset may depend on user input from a user that is captured after planning information determined by the processing algorithm in a pre-training training state is provided to that user, with the trained processing algorithm depending on that user input. The processing algorithm can thus be further improved or adapted to the specific preferences of user groups, e.g. on the basis of user inputs made during actual operation to adapt the planning information. For example, a user can change at least one parameter value for the biopsy parameter or the imaging parameter and a training data record can be generated that includes the parameter values adjusted according to the user's wishes. In particular, after several of these training datasets have been generated, they can be used for additional iterations of the training process, either on their own or together with other training datasets that were already used in the original training.

In addition or as an alternative, the user input can be used, for example, to evaluate the planning information provided by the processing algorithm and/or to describe the strength of the changes made to the parameter values by the user. Such information that is additionally stored in the training data set can be used, for example, to adjust the weighting of errors as part of error feedback depending on such an evaluation or correction strength.

Additionally or alternatively, at least one of the training datasets can include an optical image of the breast affected by the medical image data and/or a processing result determined as a function of this image, the trained processing algorithm depending on this optical image and/or this processing result.

The optical imaging can show the breast in a state that is present after a breast biopsy that was performed according to the planning information stored in the training data set or with at least one parameter value stored there for the biopsy parameter and/or the imaging parameter. For example, it can be an optical image that was recorded as part of a follow-up examination. On the basis of the optical image or as a processing result, it can be taken into account in particular whether or to what extent the optical image shows scarring. This means that not only can existing expert knowledge be used during the training, but actual results from previous interventions can also be taken into account. In particular, the preceding intervention is not part of the claimed training method, but the corresponding one Feature only affects the use of training datasets with specific properties.

The invention also relates to a device with means that are set up to execute the method according to the invention for providing planning information and/or the method according to the invention for training a processing algorithm.

The means can in particular be a programmable processing device, for example a processor, an FPGA, a GPU or the like, with an associated memory, with the processing device executing a computer program that implements the respective method. However, such a program can also be executed by a number of distributed processing devices, for example in a cloud.

The device can be a local workstation computer, part of a medical imaging device, a server, or implemented as a decentralized system, for example as a cloud solution. In principle, it is also possible to use a hard-wired processing device, for example an ASIC, instead of a programmed processing device.

The invention also relates to a computer program which, when executed on a data processing device, implements the method according to the invention for providing planning information and/or the method according to the invention for training a processing algorithm. In particular, a data processing device known per se can form a device according to the invention by executing the computer program according to the invention.

The invention also relates to a computer-readable data carrier which comprises instructions which, when executed on a data processing device, implement the method according to the invention for providing planning information and/or the method according to the invention for training a processing algorithm.

• 1 ein Ablaufdiagramm eines Ausführungsbeispiels des erfindungsgemäßen Verfahrens zur Bereitstellung einer Planungsinformation,
• 2 Datenstrukturen und Algorithmen, die im Rahmen des mit Bezug auf 1 erläuterten Ausführungsbeispiels nutzbar sind,
• 3 relevante Algorithmen und Datenstrukturen im Rahmen des erfindungsgemäßen Verfahrens zum Trainieren eines Verarbeitungsalgorithmus,
• 4 eine mögliche Systemarchitektur zur Durchführung der mit Bezug auf die 1-3 erläuterten Verfahren einschließlich einer Vorrichtung zur Durchführung dieser Verfahren, und
• 5/6 vereinfachte Darstellungen von neuronalen Netzen, die in den erfindungsgemäßen Verfahren als Verarbeitungsalgorithmus nutzbar sind.

Further advantages and details of the invention result from the following exemplary embodiments and the associated drawings. Here show schematically:

• 1 a flowchart of an exemplary embodiment of the method according to the invention for providing planning information,
• 2 Data structures and algorithms referred to in the context of 1 explained embodiment can be used,
• 3 relevant algorithms and data structures within the scope of the method according to the invention for training a processing algorithm,
• 4 a possible system architecture for carrying out with reference to the 1-3 described methods including an apparatus for carrying out these methods, and
• 5 /6 simplified representations of neural networks that can be used as a processing algorithm in the method according to the invention.

1 shows a flowchart of a method for providing planning information 33 for preparing a breast biopsy. In the simplest case, it would be possible here for only steps S1, S5 and S6 to be carried out. The further steps serve on the one hand to further improve the result of the method, i.e. the planning information 33, and on the other hand to ensure that the processing algorithm 36 used in step S5 can be trained robustly with less effort and in particular a smaller number of training data sets. In the following, however, the simplest embodiment of the method, which exclusively includes steps S1, S5 and S6, will first be explained.

In step S1, medical image data 34 relating to a breast 35 to be biopsied are first received. An example of corresponding image data is in 2 shown schematically, only a two-dimensional image representation being shown there to improve clarity. In contrast, the image data 34 can preferably be stereoscopic image data or three-dimensional image data, which were recorded on the compressed breast using a tomosynthesis method, for example.

In step S5, a processing algorithm 36, which processes the image data 34 as input data, determines the planning information 33, which can then be made available in step S6, for example to a user or an external device.

The processing algorithm 36 is trained by machine learning, a possible approach to the training being discussed later with reference to FIG 3 will be explained. In the example, the planning information 33 includes parameter values for biopsy parameters 37 of the breast biopsy and imaging parameters 38 of a medical imaging method for monitoring the breast biopsy. in the in 2 example shown here are parameter values for an expected suitable needle size 39, a presumably suitable puncture position 40 and a presumably suitable puncture angle 41 of a biopsy needle are used as biopsy parameters 37. In the same example, an acceleration voltage 42 and an X-ray dose 43 for planned X-ray exposures, a presumably suitable processing parameter 44 for artifact reduction, for example an image area in which metal artifacts are expected to be expected, and a presumably suitable imaging geometry 45 are specified as imaging parameters 38. It is also possible that only parts of the parameters mentioned are provided as planning information 33 . In some cases it can also be advantageous to specify parts of the imaging parameters 38, such as the presumably suitable X-ray dose 43, separately for a number of separate X-ray recordings.

By using the additional in 1 illustrated steps S2 to S4 and S7 to S12 or by using only parts of these additional steps, the robustness of the method can be improved and the number of training data sets that are required for robust training of the processing algorithm 36 can potentially be significantly reduced. In addition, subsequent training can be used, for example, to adapt the processing algorithm 36 to the preferences of specific user groups.

In step S2 further medical image data 57 can be provided as additional input data, depending on which the processing algorithm 36 determines the planning data 33 . The further medical image data 57 also relate to the breast 35 to be biopsied, but are recorded at a different point in time and/or with a different medical imaging device. For example, the additional medical image data 57 can be image data that was captured using a different detection method, for example ultrasound data, if the medical image data 34 is X-ray data. In this way, under certain circumstances, additional information can be obtained that is relevant for planning the biopsy.

However, the further medical image data 57 is particularly preferably recorded in the same way as the medical image data 34, but at a previous point in time, for example at a point in time at which a tissue change to be biopsied was not yet present. This can make it much easier to identify areas of the breast 35 that are relevant for a biopsy.

Additional parameters 58 can also be provided in step S3, which are processed by the processing algorithm 36. For example, this can be information from a patient database or file. As already discussed in detail in the general part, it can be relevant, for example, to take into account the age of the patient or previous interventions, findings and/or diagnoses.

In step S4, the image data 34 can be pre-processed by various pre-processing algorithms 47, 48 in order to provide the respective output data 46 to the processing algorithm 36. In this context, use can be made of the fact that algorithms are already known and tested in everyday clinical practice which, for example, can identify and classify areas which could be lesions and/or which can determine local tissue properties, for example breast density. The pre-processing in step S3 makes it possible to avoid having to determine relevant information by the processing algorithm 36 itself, which reduces its complexity and as a result robust training can already be possible with a relatively small number of training data sets.

As output data, the processing algorithms 47, 48 can, for example, relate to breast density information 49 relating to the tissue density of the breast 35 to be biopsied and at least one area information item 50, which relates to marked regions 51, 52 recognized in the image data.

the inside 2 The distinguished area 51 shown can, for example, be an area of particularly high tissue density, which should be avoided during the biopsy in order to avoid deflecting the biopsy needle. The marked area 52 can be, for example, that area which is expected to be biopsied or in which a lesion is expected to be present. As already discussed in the general part, applications or algorithms are known which, for example, can indicate a probability that a lesion is present in certain tissue areas.

As in 2 is shown schematically, the output data 46 of the pre-processing algorithms 47, 48 can be fed into different layers of a processing algorithm 36 designed as a neural network. The processing algorithm 36 can, for example, initially comprise a plurality of convolution layers 70, the output data of which are supplied to at least one fully connected layer 71, the planning in formation 33 provides as output data. For example, the pre-processing algorithm 47 can be used to identify, measure and/or classify areas suspected of having a lesion. For example, the position of the area 52 can be described in that this area is segmented in the image data 34 . Since the segmentation can take place on the image data 34, it can be advantageous to also process this by the convolution layers 70.

On the other hand, specific characteristic variables, for example a diameter or a morphology or a class of the area 52 can be determined and, since this information is not assigned to specific image areas, can be fed directly to the at least one fully connected layer 71 .

The pre-processing algorithm 48 can, for example, determine the breast density information 49, in which case, for example, each pixel or voxel is assigned a respective density value, or, as in 2 is shown schematically, at least one area 51 is segmented with a specific breast density. Since the breast density information 49 is typically different in different image areas, it is typically appropriate to process it through the convolution layers 70. FIG.

The planning information 33 provided by the processing algorithm 36 is checked and possibly adjusted by a user, for example a radiologist, before the biopsy is carried out. In step S7, after the planning information 33 has been provided to the user, a user input 53 from this user can be recorded, which, for example, relates to an evaluation of the planning information 33 provided or, in particular, to the user's change requests. The planning information 33 changed according to the user input 53 or the original planning information 33 and the evaluation information can then be stored in step S12 together with the image data 34 as a new training data record 60, which can be used at a later point in time for further training of the processing algorithm 36, such as in 1 is represented by the dashed arrow.

While it may be sufficient to provide the planning information 33 to a user in step S6, for example by graphically visualizing a proposed puncture geometry on a breast model, it may also be advantageous to support the user immediately during the breast biopsy or during the more immediate preparation of the breast biopsy before the to support the insertion of the biopsy needle.

For this purpose, in particular before the biopsy needle is punctured, the position 54 and/or orientation 55 of the biopsy needle can be detected in step S8 by sensors, for example with the aid of video cameras or other position detectors known per se. In step S9, the detected position 54 and orientation 55 can be compared with the puncture position 40 or the puncture angle 41 of the biopsy needle provided as part of the planning information 33, and depending on this comparison, in step S9, auxiliary information 56 can be output to the user as required.

For example, the auxiliary information 56 can only be output if the deviation from the puncture position 40 or the puncture angle 41 exceeds a limit value. However, it is also possible to always output this auxiliary information 56, which indicates, for example, the deviation from the target puncture position or the target puncture angle. It is also possible to color-code corresponding deviations, for example, in order to indicate major deviations.

In step S10, the biopsy itself is performed. Since this is no longer a step in the method for providing planning information in preparation for the breast biopsy, this step is separated from the rest of the method by a dashed arrow. Nevertheless, it seemed appropriate to carry out step S10 in the flow chart in 1 in order to clarify the possibility presented below of generating further training data 60 for further training of the processing algorithm 36 .

Step S11 is carried out after the breast biopsy carried out in step S10 has been completed, for example as part of a follow-up appointment. Here, an optical of the breast 35, which serves in particular to document possible scarring or other indications that point to a non-optimal biopsy. The optical can be used to supplement a training data record 60 that includes the image data 34 and the planning information 33 determined and optionally changed by the user input 53 . This makes it possible, for example, to take into account during the training whether scarring or other disadvantages occurred when certain planning information 33 was used with given medical image data 34 . If this is the case, the corresponding training data set can be discarded which are weighted less or even negatively within the framework of the training.

3 shows schematically the interaction of the relevant algorithms and data structures in a method for training a processing algorithm 36, which is used to determine planning information 33 for preparing a breast biopsy on the basis of medical image data 34 provided as input data. In particular, by referring to 3 the method described in the reference to 1 and 2 The processing algorithm 36 used in the methods explained can be trained.

First, multiple training datasets 60 are provided, each of which includes at least the image data 34 and the target values 31-36 for at least one biopsy parameter 37 or imaging parameter 38 . Since, as explained above, at least one preprocessing algorithm 47 is to be used within the framework of determining the planning information, the output data of which is to be processed by the processing algorithm 36, the image data 34 of the respective training data set 60 are first processed by this preprocessing algorithm 47 in order to provide the output data 46 . These are processed together with the image data 34 by the initially untrained processing algorithm 36 .

The processing algorithm 36 has numerous parameters 72 that can be determined by the training and, for example, can initially be set to random values in the untrained state. The parameters 72 can be input weights of the neurons of a neural network to be trained, for example.

The planning information 33 determined by the processing algorithm 36 includes actual values 73-75 for the biopsy or imaging parameters, which, depending on the training state of the processing algorithm 36, deviate from the target values 61-63 provided with the respective training data record 60. Based on this deviation, for example, an error feedback 76 known per se from the prior art can be used to adapt the parameter values of the parameters 72 of the processing algorithm 36 in such a way that an error sum determined on the basis of the actual values 73-75 and the setpoint values 61-63 is minimized.

4 shows a possible system architecture within which the methods described above can be implemented. The method for providing the planning information or the method for training the processing algorithm can both be implemented by a common device 65 . Alternatively, it would also be possible to use separate devices for these two methods or to implement just one of these methods.

The device 65 has means for implementing the method, the means 66 being a programmable processing device, for example a microprocessor, and the means 67 being a memory which comprises a computer program 68 or a respective computer program 68, by which the method or the respective method to be implemented. The device 65 is thus a data processing device 69 known per se, which is set up by appropriate programming to carry out the method.

The medical image data 34 can be provided by a medical imaging device 59, with the provision being made indirectly via a database 78 in the example, so that the determination of the planning information would potentially also be possible significantly after the imaging itself. The database 78 can also store the additional medical image data 57 already explained above and/or additional parameters 58, which can be evaluated as part of the determination of the planning information.

In the simplest case, the planning information can be output to a user via an output device 81, for example a screen. Additionally or alternatively, the planning information can be used to provide imaging parameters suggested or adapted based on a user input at the user interface 79 directly to the imaging device 59 or to provide biopsy parameters to a monitoring device 80 . Together with a sensor system 77, which can be part of the imaging device 59, for example, the monitoring device 80 can provide the user with auxiliary information 56, as already mentioned above with reference to steps S8 and S9 in 1 was explained.

In particular, as discussed above, the processing algorithm 36 may be a neural network, preferably utilizing both convolutional layers and fully connected layers. The basic principle of neural networks is referred to below 5 are explained, with further special features of convolutional neural networks or convolutional layers then with reference to 6 be explained.

5 1 shows an exemplary embodiment of an artificial neural network 1. English expressions for the artificial neural network 1 are “artificial neural network”, “neural network”, “artificial neural net” or “neural net”.

The artificial neural network 1 comprises nodes 6 to 18 (nodes) and edges 19 to 21 (edges), each edge 19 to 21 being a directed connection from a first node 6 to 18 to a second node 6 to 18. In general, the first node 6-18 and the second node 6-18 are different nodes 6-18, but it is also conceivable that the first node 6-18 and the second node 6-18 are identical. For example, in 5 edge 19 is a directed connection from node 6 to node 9; and edge 21 is a directed connection from node 16 to node 18. An edge 19 to 21 from a first node 6 to 18 to a second node 6 to 18 is referred to as an incoming edge for the second node 6 to 18 and as an outgoing edge for the first node 6 to 18.

In this exemplary embodiment, the nodes 6 to 18 of the artificial neural network 1 can be arranged in layers 2 to 5 (layers), where the layers can have an intrinsic order introduced by the edges 19 to 21 between the nodes 6 to 18. In particular, edges 19-21 can only be provided between adjacent layers of nodes 6-18. In the exemplary embodiment shown, there is an input layer 2 which only has the nodes 6, 7, 8, each without an incoming edge. The output layer 5 comprises only the nodes 17, 18, each without outgoing edges, with hidden layers 3 and 4 also lying between the input layer 2 and the output layer 5. In the general case, the number of hidden layers 3, 4 can be chosen arbitrarily. The number of nodes 6, 7, 8 of the input layer 2 usually corresponds to the number of input values in the neural network 1, and the number of nodes 17, 18 in the output layer 5 usually corresponds to the number of output values of the neural network 1.

für das Gewicht ${\text{w}}_{\text{i}\text{,j}}^{\left(\text{n}\text{,n}+1\right)}$

definiert.In particular, a (real) number can be assigned to the nodes 6 to 18 of the neural network 1. Where x ⁽ⁿ⁾ i denotes the value of the i-th node 6 to 18 of the n-th layer 2 to 5. The values of the nodes 6, 7, 8 of the input layer 2 are equivalent to the input values of the neural network 1, while the Values of the nodes 17, 18 of the output layer 5 are equivalent to the output values of the neural network 1. In addition, each edge 19, 20, 21 can be assigned a weight in the form of a real number. In particular, the weight is a real number in the interval [-1, 1] or in the interval [0, 1,] . Where w ^(m ' ⁿ⁾ i,j denotes the weight of the edge between the i-th node 6 to 18 of the m-th layer 2 to 5 and the j-th node 6 to 18 of the n-th layer 2 to 5. Furthermore, the abbreviation ${\text{w}}_{\text{i}\text{,j}}^{\left(\text{n}\right)}$

for the weight ${\text{w}}_{\text{i}\text{,j}}^{\left(\text{n}\text{,n}+1\right)}$

Are defined.

In order to calculate output values of the neural network 1, the input values are propagated through the neural network 1. In particular, the values of the nodes 6 to 18 of the (n+1)th layer 2 to 5 can be calculated based on the values of the nodes 6 to 18 of the nth layer 2 to 5 by $${\text{x}}_{\text{j}}^{\left(\text{n}+1\right)}=\text{f}\left({\displaystyle \sum {}_{\text{i}}{\text{x}}_{\text{i}}^{\left(\text{n}\right)}}\cdot {\text{w}}_{\text{i},\text{j}}^{\left(\text{n}\right)}\right)$$

Here f is a transfer function, which can also be referred to as an activation function. Known transfer functions are step functions, sigmoid functions (for example the logistic function, the generalized logistic function, the hyperbolic tangent, the arctangent, the error function, the smoothstep function) or rectifier functions (rectifier). The transfer function is mainly used for normalization purposes.

In particular, the values are propagated layer by layer through the neural network 1 , with values of the input layer 2 being given by the input data of the neural network 1 . Values of the first hidden layer 3 can be calculated based on the values of the input layer 2 of the neural network 1, values of the second hidden layer 4 can be calculated based on the values in the first hidden layer 3, etc.

für die Kanten 19 bis 21 festlegen zu können, muss das neuronale Netz 1 unter Verwendung von Trainingsdaten trainiert werden. Insbesondere umfassen Trainingsdaten Trainingseingangsdaten und Trainingsausgangsdaten, die im Folgenden als t_i bezeichnet werden. Für einen Trainingsschritt wird das neuronale Netzwerk 1 auf die Trainingseingangsdaten angewendet, um berechnete Ausgangsdaten zu ermitteln. Insbesondere umfassen die Trainingsausgangsdaten und die berechneten Ausgangsdaten eine Zahl von Werten, wobei sich die Zahl als die Zahl der Knoten 17, 18 der Ausgangsschicht 5 bestimmt.About the values ${\text{w}}_{\text{i}\text{,j}}^{\left(\text{n}\right)}$

for the edges 19 to 21, the neural network 1 must be trained using training data. In particular, training data includes training input data and training output data, which are referred to as t _i in the following. For a training step, the neural network 1 is applied to the training input data in order to determine calculated output data. In particular, the training output data and the calculated output data include a number of values, the number being determined as the number of nodes 17, 18 of the output layer 5.

geändert werden, wobei γ eine Lernrate ist und die Zahlen ${\mathrm{\delta }}_{\text{j}}^{\left(\text{n}\right)}$

rekursiv berechnet werden können als $${\mathrm{\delta }}_{\text{j}}^{\left(\text{n}\right)}=\left({\displaystyle \sum {}_{\text{k}}{\mathrm{\delta }}_{\text{k}}^{\left(\text{n}+1\right)}}\cdot {\text{w}}_{\text{j}\text{,k}}^{\left(\text{n}+1\right)}\right)\cdot \text{f'}\left({\displaystyle \sum {}_{\text{i}}{\text{x}}_{\text{i}}^{\left(\text{n}\right)}\cdot {\text{w}}_{\text{i}\text{,j}}^{\left(\text{n}\right)}}\right)$$

basierend auf ${\mathrm{\delta }}_{\text{j}}^{\left(\text{n}+1\right)},$

wenn die (n+1)-te Schicht nicht die Ausgangsschicht 5 ist, und $${\mathrm{\delta }}_{\text{j}}^{\left(\text{n}\right)}=\left({\text{x}}_{\text{k}}^{\left(\text{n}+1\right)}-{\text{t}}_j^{\left(\text{n}+1\right)}\right)\cdot \text{f'}\left({\displaystyle \sum {}_{\text{i}}{\text{x}}_{\text{i}}^{\left(\text{n}\right)}\cdot {\text{w}}_{\text{i}\text{,j}}^{\left(\text{n}\right)}}\right)$$

falls die (n+1)-te Schicht die Ausgangsschicht 5 ist, wobei f' die erste Ableitung der Aktivierungsfunktion ist und ${\text{y}}_{\text{j}}^{\left(\text{n}+1\right)}$

der Vergleichstrainingswert für den j-ten Knoten 17, 18 der Ausgangsschicht 5 ist.In particular, a comparison between the calculated output data and the training output data is used to recursively adjust the weights within the neural network 1 (back propagation algorithm - "back propagation algorithms"). In particular, the weights accordingly $${\text{w'}}_{\text{i}\text{,j}}^{\left(\text{n}\right)}={\text{w}}_{\text{i}\text{,j}}^{\left(\text{n}\right)}-\mathrm{g}\cdot {\mathrm{\delta }}_{\text{j}}^{\left(\text{n}\right)}\cdot {\text{x}}_{\text{i}}^{\left(\text{n}\right)}$$

be changed, where γ is a learning rate and the numbers ${\mathrm{\delta }}_{\text{j}}^{\left(\text{n}\right)}$

can be computed recursively as $${\mathrm{\delta }}_{\text{j}}^{\left(\text{n}\right)}=\left({\displaystyle \sum {}_{\text{k}}{\mathrm{\delta }}_{\text{k}}^{\left(\text{n}+1\right)}}\cdot {\text{w}}_{\text{j}\text{,k}}^{\left(\text{n}+1\right)}\right)\cdot \text{f'}\left({\displaystyle \sum {}_{\text{i}}{\text{x}}_{\text{i}}^{\left(\text{n}\right)}\cdot {\text{w}}_{\text{i}\text{,j}}^{\left(\text{n}\right)}}\right)$$

based on ${\mathrm{\delta }}_{\text{j}}^{\left(\text{n}+1\right)},$

when the (n+1)th layer is not the output layer 5, and $${\mathrm{\delta }}_{\text{j}}^{\left(\text{n}\right)}=\left({\text{x}}_{\text{k}}^{\left(\text{n}+1\right)}-{\text{t}}_j^{\left(\text{n}+1\right)}\right)\cdot \text{f'}\left({\displaystyle \sum {}_{\text{i}}{\text{x}}_{\text{i}}^{\left(\text{n}\right)}\cdot {\text{w}}_{\text{i}\text{,j}}^{\left(\text{n}\right)}}\right)$$

if the (n+1)th layer is the output layer 5, where f' is the first derivative of the activation function and ${\text{y}}_{\text{j}}^{\left(\text{n}+1\right)}$

is the comparison training value for the j-th node 17, 18 of the output layer 5.

The following is with regard to 6 also given an example of a Convolutional Neural Network (CNN). It should be noted that the term "layer" is used there in a slightly different way than for classical neural networks. For a classical neural network, the term "layer" only refers to the set of nodes that form a layer, hence a specific generation of nodes. For a convolutional neural network, the term "layer" is often used as an object that actively changes data, in other words a set of nodes of the same generation and either the set of incoming or outgoing edges.

6 shows an embodiment of a convolutional neural network 22. In the illustrated embodiment, the convolutional neural network 22 includes an input layer 23, a convolutional layer 24 (convolutional layer), a pooling layer 25, a fully connected layer 26 and an output layer 27. In alternative embodiments For example, the convolutional neural network 22 may include multiple convolutional layers 24, multiple pooling layers 25, and multiple fully connected layers 26, just like other types of layers. The order of the layers can be selected in any order, with layers 26 that are usually completely connected forming the last layers before the starting layer 27 .

In particular, within a convolutional neural network 22, the nodes 28 to 32 of one of the layers 23 to 27 can be understood as being arranged in a d-dimensional matrix or as a d-dimensional image. In particular, in the two-dimensional case, the value of a node 28 to 32 with indices i,j in the nth layer 23 to 27 can be denoted as x ⁽ⁿ⁾ [i,j]. It should be noted that the arrangement of the nodes 28 to 31 of a layer 23 to 27 has no effect whatsoever on the calculations within the convolutional neural network 22 as such, since these effects are only given by the structure and the weights of the edges.

der Knoten 29 der Faltungsschicht 24 als eine Faltung ${\text{x}}_{\text{k}}^{\left(\text{n}\right)}={\text{K}}_{\text{k}}*{\text{x}}^{\left(\text{n}-1\right)}$

basierend auf den Werten x^(n-1) der Knoten 28 der vorangehenden Schicht 23 ermittelt werden, wobei die Faltung * im zweidimensionalen Fall definiert werden kann als $${\text{x}}_{\text{k}}^{\left(\text{n}\right)}\left[\text{i}\text{,j}\right]=\left({\text{K}}_{\text{k}}*{\text{x}}^{\left(\text{n}-1\right)}\right)\left[\text{i}\text{,j}\right]={\displaystyle \sum {}_{\text{i}}},{\displaystyle \sum {}_{\text{j}},{\text{ K}}_{\text{k}}\left[\text{i'}\text{,j'}\right]}\cdot {\text{x}}^{\left(\text{n}-1\right)}\left[\text{i}-\text{i'}\text{, j}-\text{j'}\right].$$

A convolution layer 24 is particularly distinguished in that the structure and the weights of the incoming edges form a convolution operation based on a certain number of kernels. In particular, the structure and weights of the incoming edges can be chosen such that the values ${\text{x}}_{\text{k}}^{\left(\text{n}\right)}$

the node 29 of the convolution layer 24 as a convolution ${\text{x}}_{\text{k}}^{\left(\text{n}\right)}={\text{K}}_{\text{k}}*{\text{x}}^{\left(\text{n}-1\right)}$

can be determined based on the values x ^(n-1) of the nodes 28 of the previous layer 23, where the convolution * in the two-dimensional case can be defined as $${\text{x}}_{\text{k}}^{\left(\text{n}\right)}\left[\text{i}\text{,j}\right]=\left({\text{K}}_{\text{k}}*{\text{x}}^{\left(\text{n}-1\right)}\right)\left[\text{i}\text{,j}\right]={\displaystyle \sum {}_{\text{i}}},{\displaystyle \sum {}_{\text{j}},{\text{K}}_{\text{k}}\left[\text{i'}\text{'y'}\right]}\cdot {\text{x}}^{\left(\text{n}-1\right)}\left[\text{i}-\text{i'}\text{, j}-\text{j'}\right].$$

Therein, the k-th kernel K _k is a d-dimensional matrix, in this exemplary embodiment a two-dimensional matrix, which is usually small compared to the number of nodes 28 to 32, for example a 3x3 matrix or a 5x5 matrix. In particular, this implies that the weights of the incoming edges are not independent, but are chosen such that they generate the convolution equation above. In the example of a kernel forming a 3x3 matrix, there are only nine independent weights (each entry of the kernel matrix corresponding to an independent weight), regardless of the number of nodes 28 through 32 in the corresponding layer 23 through 27. In particular, for a convolutional layer 24, the number of nodes 29 in the convolutional layer 24 is equivalent to the number of nodes 28 in the previous layer 23 multiplied by the number of convolutional kernels.

If the nodes 28 of the preceding layer 23 are arranged as a d-dimensional matrix, the use of the plurality of cores can be understood as adding another dimension, also referred to as depth dimension, so that the nodes 29 of the convolutional layer 24 as one (d+1)-dimensional matrix are arranged. If the nodes 28 of the preceding layer 23 are already arranged as a (d+1)-dimensional matrix with a depth dimension, the use of a plurality of convolution kernels can be understood as an expansion along the depth dimension, so that the nodes 29 of the convolution layer 24 are equally as a (d+1)-dimensional mat rix are arranged, the size of the (d+1)-dimensional matrix in the depth dimension being larger than in the preceding layer 23 by the factor formed by the number of nuclei.

The advantage of using convolutional layers 24 is that the spatially local correlation of the input data can be exploited by creating a local connection pattern between nodes of adjacent layers, in particular by each node having connections to only a small portion of the nodes of the previous layer.

In the illustrated embodiment, the input layer 23 includes thirty-six nodes 28 arranged as a two-dimensional 6x6 matrix. The convolution layer 24 comprises seventy-two nodes 29 arranged as two 6x6 two-dimensional matrices, each of the two matrices being the result of a convolution of the values of the input layer 23 with a convolution kernel. Likewise, the nodes 29 of the convolution layer 24 can be understood as being arranged in a 6x6x2 three-dimensional matrix, the latter dimension being the depth dimension.

berechnet werden. In anderen Worten kann durch die Verwendung einer Pooling-Schicht 25 die Zahl der Knoten 29, 30 reduziert werden, indem eine Anzahl von d₁ x d₂ benachbarter Knoten 29 in der vorangehenden Schicht 24 durch einen einzelnen Knoten 30 ersetzt werden, der als eine Funktion der Werte der genannten Anzahl benachbarter Knoten 29 berechnet wird. Insbesondere kann die Pooling-Funktion f eine Maximumsfunktion, eine Durchschnittsbildung oder die L2-Norm sein. Insbesondere können für eine Pooling-Schicht 25 die Gewichte der eingehenden Kanten festgelegt sein und nicht durch Training modifiziert sein.A pooling layer 25 is characterized in that the structure and the weights of the incoming edges and the activation function of its nodes 30 define a pooling operation based on a non-linear pooling function f. For example, in the two-dimensional case, the values x ⁽ⁿ⁾ of the nodes 30 of the pooling layer 25 can be calculated based on the values x ⁽ⁿ⁺¹⁾ of the nodes 29 of the previous layer 24 as $${\text{x}}^{\left(\text{n}\right)}\left[\text{i}\text{,j}\right]=\text{f}\left({\text{x}}^{\left(\text{n}-1\right)}\left[{\text{i.e}}_1,{\text{sb}}_2\right],...,{\text{x}}^{\left(\text{n}-1\right)}\left[{\text{i.e}}_1+{\text{i.e}}_1-\mathrm{1,}{\text{<figure-callout id="2" label="sb" filenames="DE102021210804A1\_0018.png" state="{{state}}">sb</figure-callout>}}_2+{\text{i.e}}_2-1\right]\right)$$

be calculated. In other words, by using a pooling layer 25, the number of nodes 29, 30 can be reduced by replacing a number of d ₁ xd ₂ neighboring nodes 29 in the preceding layer 24 with a single node 30 acting as a function of the values of said number of neighboring nodes 29 is calculated. In particular, the pooling function f can be a maximum function, an averaging, or the L2 norm. In particular, for a pooling layer 25, the weights of the incoming edges can be fixed and not modified by training.

The advantage of using a pooling layer 25 is that the number of nodes 29, 30 and the number of parameters is reduced. This leads to a reduction in the necessary amount of calculations within the convolutional neural network 22 and thus to control of the overfitting.

In the illustrated embodiment, the pooling layer 25 is a max-pooling layer in which four neighboring nodes are replaced with only a single node whose value is the maximum of the values of the four neighboring nodes. Max pooling is applied to each d-dimensional matrix of the previous layer; in this embodiment, max-pooling is applied to each of the two two-dimensional matrices, reducing the number of nodes from seventy-two to eighteen.

A fully connected layer 26 is characterized in that there are a plurality, in particular all, edges between the nodes 30 of the previous layer 25 and the nodes 31 of the fully connected layer 26, the weight of each of the edges being individually adjustable. In this embodiment, the nodes 30 of the previous layer 25 and the fully connected layer 26 are shown both as two-dimensional matrices and as non-contiguous nodes (represented as a row of nodes with the number of nodes reduced for clarity). In this embodiment, the number of nodes 31 in the fully connected layer 26 is equal to the number of nodes 30 in the previous layer 25. In alternative embodiments, the number of nodes 30, 31 may be different.

Furthermore, in this embodiment, the values of the nodes 32 of the output layer 27 are determined by applying the soft max function to the values of the nodes 31 of the previous layer 26. By applying the softmax function, the sum of the values of all nodes 32 of the output layer 27 is one and all values of all nodes 32 of the output layer are real numbers between 0 and 1. When the convolutional neural network 22 is used to classify input data, in particular the Values of the output layer 27 are interpreted as the probability that the input data fall into one of the different classes.

A convolutional neural network 22 can also have a ReLU layer, where ReLU is an acronym for "rectified linear units". In particular, the number of nodes and the structure of the nodes within a ReLU layer is equivalent to the number of nodes and the structure of the nodes of the previous layer. In particular, the value of each node in the ReLU layer can be calculated by applying a rectification function (rectifier function) on the value of the corresponding node of the previous layer. Examples of rectifier functions are f(x)=max(0,x), the hyperbolic tangent or the sigmoid function.

In particular, convolutional neural networks 22 can be trained based on the backpropagation algorithm. In order to avoid overfitting, regularization methods can be used, for example dropout of individual nodes 28 to 32, stochastic pooling, use of artificial data, weight decay based on the L1 or the L2 norm or maximum norm constraints.

Although the invention has been illustrated and described in detail by the preferred embodiment, the invention is not limited by the disclosed examples and other variations can be derived therefrom by those skilled in the art without departing from the scope of the invention.

Claims (15)

A method for providing planning information (33) to prepare for a breast biopsy, comprising the steps of: - receiving medical image data (34) relating to a breast (35) to be biopsied, - Determination of the planning information (33) by a processing algorithm (36), which processes the image data (34) as input data and is trained by machine learning, the planning information (33) having a presumably suitable parameter value for at least one biopsy parameter (37) of the breast biopsy and/or or for at least one imaging parameter (38) of a medical imaging method for monitoring the breast biopsy, - Providing the planning information (33). procedure after claim 1 , characterized in that the planning information (33) on the one hand the presumably suitable parameter value for a presumably suitable needle size (39) and/or a presumably suitable puncture position (40) and/or a presumably suitable puncture angle (41) of a biopsy needle as a biopsy parameter and/or on the other hand, a presumably suitable acceleration voltage (42) and/or x-ray dose (43) for at least one planned x-ray exposure and/or a presumably suitable processing parameter (44) for artifact reduction and/or a presumably suitable imaging geometry (45) as imaging parameters. procedure after claim 1 or 2 , characterized in that the processing algorithm (36) processes the output data (46) from at least one preprocessing algorithm (47, 48) or from at least two preprocessing algorithms (47, 48) as additional input data in order to determine the planning information (33). procedure after claim 3 , characterized in that the pre-processing algorithm (47, 48) or at least one of the pre-processing algorithms (47, 48) is trained by a machine learning method. procedure after claim 3 or 4 , characterized in that the pre-processing algorithm (47, 48) or a respective one of the pre-processing algorithms (47, 48) provides the processing algorithm (36) with output data (46) which contains breast density information (49) relating to the tissue density of the breast (35) to be biopsied and/or relate to area information (50) relating to a distinguished area (51, 52) recognized in the image data. Method according to one of the preceding claims, characterized in that the planning information (33) is provided to a user, after which a user input (53) from this user is recorded, which confirms the planning information (33) and/or a change in the planning information (33 ) concerns, after which an additional training of the processing algorithm (36) depending on the user input (53) and the image data (34) is carried out. Method according to one of the preceding claims, characterized in that a position (54) and / or an orientation (55) of a biopsy needle with respect to the breast (35) to be biopsied before the biopsy needle pierces the breast (35) are detected by sensors, wherein depending on the position (54) and/or the orientation (55) and the planning information (33) provided, auxiliary information (56) is output for the or a user. Method according to one of the preceding claims, characterized in that the processing algorithm (36) evaluates at least one additional parameter (58) provided independently of the image data (34). Method according to one of the preceding claims, characterized in that the processing algorithm (36) determines the planning information (33) as a function of further medical image data (57) provided as additional input data, the further medical image data (57) also determining the breast to be biopsied ( 35) concern, but at a different time point and/or recorded by another medical imaging device (59). Method according to one of the preceding claims, characterized in that X-ray data on the one hand and/or three-dimensional image data (34) or stereoscopic image data (34) on the other hand are used as image data (34). Method for training a processing algorithm (36), which is used to determine planning information (33) for preparing a breast biopsy on the basis of medical image data (34) provided as input data, which relate to a respective breast (35) to be biopsied, wherein the Planning information (33) specifies a parameter value that is expected to be suitable for at least one biopsy parameter (37) of the breast biopsy and/or for at least one imaging parameter (38) of a medical imaging method for monitoring the breast biopsy, with monitored machine learning taking place on the basis of training data sets (60), wherein the respective training data record (60) comprises at least the medical image data (34) and a respective target value (61-63) for the at least one biopsy parameter (37) and/or the at least one imaging parameter (38). procedure after claim 11 , characterized in that at least one of the training data sets - on the one hand depends on a user input (53) of a user that is recorded after planning information (33) that is determined by the processing algorithm (36) in a training state that is present before the training is made available to this user, and/or - on the other hand, comprises an optical image (64) of the breast (35) affected by the medical image data (34) and/or a processing result determined as a function of this image (64), the trained processing algorithm (36 ) depends on this user input (53) and/or this optical image (64) and/or this processing result. Device with means (66, 67) arranged to carry out the method according to one of the preceding claims. Computer program, characterized in that when it is carried out on a data processing device (69), a method according to one of Claims 1 until 12 implemented. Computer-readable data carrier, characterized in that it comprises instructions which, when executed on a data processing device (69), use a method according to one of Claims 1 until 12 to implement.

Images