Estimation of the risk of unexpected movement of at least one device during a vascular intervention

Abstract

To estimate a risk of an unexpected movement of at least one device during a vascular intervention, an image sequence (7) of successive images is obtained, which depict a vascular structure and at least one device moving in the vascular structure, wherein the at least one medical device is inserted into a body of a patient (5) at a body opening. A risk characteristic value (9) is determined for whether a non-linear relationship exists or is imminent between a causal movement of the at least one medical device at an end of the at least one medical device that is proximal to the body opening and a resulting movement of the at least one medical device at an end of the at least one medical device that is distal to the body opening.Determining the risk index (9) involves applying a trained machine learning model, MLM, (8) to input data (6) containing the image sequence (7).

Description

[0001] The present invention relates to a computer-implemented method for estimating a risk of unexpected movement of at least one medical device during a vascular intervention and to a computer-implemented training method for providing a trained machine learning model for use in such a computer-implemented method. The invention further relates to a data processing system for carrying out such a computer-implemented method or training method, as well as to a corresponding computer program product.

[0002] During vascular interventions, medical devices, also referred to as medical tools, such as vascular catheters and / or guidewires, are inserted into the blood vessels for minimally invasive therapy via an artificial body opening, for example in the patient's hip. It is also possible, particularly in neurovascular interventions, for several such medical devices to be inserted, which are also referred to as a stack or combination of devices. During the intervention, the at least one medical device is mechanically pushed further into the body by a person performing the intervention or by a robot or a robotic manipulator at an end of the at least one medical device that is proximal to the body opening, which leads to a movement of the at least one medical device at an end of the at least one medical device that is distal to the body opening.Pulling and rotating movements are also possible. In the case of a stack of medical devices, these devices may be moved together, or one or more devices in the stack may be moved relative to one or more other devices in the stack.

[0003] When the at least one medical device moves in the vascular structure, tensile and compressive forces are exerted on the at least one medical device and energy builds up in the at least one medical device.

[0004] When the built-up energy dissipates, the at least one medical device may slip, which in turn may change the state of the at least one medical device and thus lead to further movements. These movements occur particularly unintentionally and uncontrolled within the vascular structure and are also referred to as unexpected movements. These can lead to a delay in the intervention and can also lead to perforation, rupture or even dissection of a vessel. Similar effects can occur if, for example, a guidewire or catheter is removed from the stack. The stability or support of the guidewire or catheter is then removed from the remaining stack. This can lead to the stack no longer remaining in place or maintaining its shape but moving unexpectedly.

[0005] The energy buildup may occur in parts of the at least one medical device that are not in the immediate field of vision of the person performing the procedure or are overlooked. Experienced persons intuitively use the haptic feedback at the proximal end of the at least one medical device and / or the live fluoroscopic images, when the part of the stack subject to the energy buildup is visible, to estimate how much play has built up in the system consisting of the medical device or stack and vessel and whether unexpected movement of the at least one medical device is present or imminent. However, this is unreliable, especially for less experienced persons or robot-guided, for example, remote-controlled, procedures.

[0006] Methods for tracking a medical device are known, e.g., from the publication MEI, Ziyang, et al. Real-time detection and tracking of guide wire / catheter for interventional embolization robot based on deep learning. In: 2023 IEEE International Conference on Mechatronics and Automation (ICMA). IEEE, 2023. pp. 778-783 or the publication EP 3 441 977 A1.

[0007] Methods for robotic control of a medical device are known, e.g., from the publication CHEN, Alvin I., et al. Deep learning robotic guidance for autonomous vascular access. Nature Machine Intelligence, 2020, Vol. 2, No. 2, pp. 104-115.

[0008] In the publication J. Donahue et al.: “Long-term Recurrent Convolutional Networks for Visual Recognition and Description” (arXiv:1411.4389) a recurrent convolutional architecture is described for artificial neural networks (ANNs) that is suitable for extensive visual learning and can be trained consistently.

[0009] In the paper X. Shi et al.: "Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting" (arXiv:1506.04214), nowcasting is considered a spatiotemporal sequence prediction problem where both the input and the prediction are spatiotemporal sequences. This is achieved by extending the fully connected LSTM (long short-term memory) to a convolutional LSTM (ConvLSTM).

[0010] In the publication TN Sainath et al.: “Convolutional, Long Short-Term Memory, fully connected Deep Neural Networks”, 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), South Brisbane, QLD, Australia, 2015, pp. 4580-4584, convolutional neural networks (CNNs), LSTMs and deep neural networks (DNNs) are combined in a unified architecture.

[0011] The publication K. Zhu et al.: “LSTM enhanced by dual-attention-based encoder-decoder for daily peak load forecasting”, Electric Power Systems Research, Volume 208, 2022, 107860 describes a forecasting model based on the LSTM and enhanced by a dual attention-based encoder-decoder.

[0012] It is an object of the present invention to more reliably estimate the risk of unexpected movement of at least one medical device during a vascular intervention.

[0013] This object is achieved by the respective subject matter of the independent claims. Advantageous further developments and preferred embodiments are the subject matter of the dependent claims.

[0014] The invention is based on the idea of using a trained machine learning model, MLM, to estimate from a sequence of consecutive images a risk index for the risk that an unexpected connection between a causal movement and a resulting movement of the at least one medical device exists or is imminent.

[0015] According to one aspect of the invention, a computer-implemented method for estimating a risk of unexpected movement of at least one medical device during a vascular intervention is provided. An image sequence of consecutive images is obtained. The image sequence, in particular the consecutive images, depict a vascular structure, in particular of a patient, as well as at least one device moving within the vascular structure. The at least one medical device is inserted into the patient's body through a body opening, in particular an artificial body opening.A risk index is determined for the existence or imminent occurrence of an unexpected connection between a causal movement of the at least one medical device at a proximal end of the at least one medical device with respect to the body opening and a movement of the at least one medical device resulting from the causal movement at a distal end of the at least one medical device with respect to the body opening. Determining the risk index involves applying a trained machine learning model (MLM) to input data containing the image sequence. In particular, the risk index is determined depending on the image sequence.

[0016] Unless otherwise stated, all steps of the computer-implemented method can be carried out by a data processing system that includes at least one data processing device. In particular, the at least one data processing device is set up or adapted to carry out the steps of the computer-implemented method. For this purpose, the at least one data processing device can, for example, store a computer program that contains instructions that, when executed by the at least one data processing device, cause the at least one data processing device to carry out the computer-implemented method. The computer-implemented method can also be implemented wholly or partially in hardware. The terms “data processing system” and “at least one data processing device” can be used interchangeably here and below. This also applies to corresponding expressions derived therefrom.

[0017] In the event that the at least one data processing device includes two or more data processing devices, certain steps performed by the at least one data processing device can also be understood as different data processing devices performing different steps or different parts of a step. In particular, it is not necessary for each data processing device to perform the steps. In other words, the execution of the steps can be distributed among the two or more data processing devices.

[0018] Each embodiment of the computer-implemented method results in a corresponding embodiment of a method for estimating a risk for an unexpected movement of at least one medical device during a vascular intervention, which is not purely computer-implemented, by including corresponding steps for generating the image sequence.

[0019] The consecutive images of the image sequence can, for example, be X-ray images, i.e., images generated using an X-ray imaging system. However, the consecutive images of the image sequence can also be ultrasound images generated using an ultrasound imaging system or images generated using another imaging modality.

[0020] Generally speaking, a trained MLM can replicate cognitive functions that humans associate with a different human mind. Specifically, by training based on training data, the MLM can be able to adapt to new circumstances and detect and extrapolate patterns. Another term for a trained MLM is "trained function."

[0021] In general, the parameters of an MLM can be adjusted or updated through training. In particular, supervised training, semi-supervised training, unsupervised training, reinforcement learning, and / or active learning can be used. Representation learning, also known as feature learning, can also be used. In particular, the parameters of MLMs can be adjusted iteratively through multiple training steps. In particular, a certain loss function, also known as the cost function, can be minimized during training. When training an artificial neural network (ANN), the backpropagation algorithm can be used, in particular.

[0022] In particular, an MLM may include an ANN, a support vector machine, a decision tree, and / or a Bayesian network, and / or the MLM may be based on k-means clustering, Q-learning, genetic algorithms, and / or association rules. In particular, an ANN may be or include a deep neural network, a convolutional neural network, a CNN, or a convolutional deep neural network. Furthermore, an ANN may be an adversarial network, a deep adversarial network, and / or a generative adversarial network.

[0023] The MLM used in the present computer-implemented method according to the invention is characterized in that it can process input data in the form of an image sequence.

[0024] The MLM therefore uses not only the information contained in the individual images, but also the chronological order of the images within the image sequence. This can be implemented in particular by feeding the images of the image sequence to the MLM one after the other and / or by the input data for each of the images in the image sequence also containing an order within the image sequence.

[0025] If the MLM is designed as an ANN, it can be configured as a recurrent neural network (RNN), particularly as an LSTM (long short-term memory), such as a recurrent convolutional neural network (RCNN). Alternatives to the LSTM include ANNs with SRUs (single recurrent units), Jordan networks, or GRUs (gated recurrent units). Apart from an ANN, other possible MLMs include Bayesian hierarchical temporal models, support vector regression models, and so on.

[0026] The at least one medical device can also be referred to as at least one medical tool. The at least one medical device can in particular include one or more guidewires and / or one or more catheters, in particular vascular catheters. It is also possible for the at least one medical device to include a stent or a vascular prosthesis or the like. Generally, when the images of the image sequence are recorded, part of the at least one medical device is located inside the patient's body and another part of the at least one medical device is located outside the patient's body.

[0027] The causal movement is performed, in particular, by a human or a robot at the proximal end of the at least one medical device. The proximal end is located, in particular, outside the patient's body. The causal movement results in the resulting movement of the at least one medical device at the distal end. The proximal end is located, in particular, inside the patient's body.

[0028] The causal movement can, for example, involve pushing the at least one medical device further into the patient's body or pulling the at least one medical device partially out of the patient's body or rotating the at least one medical device. Accordingly, the resulting movement is a forward movement, a backward movement or a rotational movement of the at least one medical device. There is a nominal relationship, which can also be referred to as an expected relationship, between the causal movement and the resulting movement. For example, the nominal relationship can be a linear relationship. This means that according to the nominal relationship, when the at least one medical device is advanced by a distance S as a causal movement, the distal end of the at least one medical device also moves forward by the distance S.The same applies to pulling or rotating movements. However, the nominal relationship does not necessarily have to be linear. In particular, the nominal relationship can be supralinear, so that according to the nominal relationship, for example, in the case of a feed by a distance S as the causal movement, the distal end of the at least one medical device also moves forward by a distance S' < S.

[0029] An unexpected relationship can be defined as a relationship that deviates significantly from the nominal relationship. What constitutes a significant deviation can be determined application-specifically by defining one or more tolerance ranges for one or more variables that specify the relationship.

[0030] In some embodiments, a mathematical definition of the unexpected correlation or the significant deviation is not required. In particular, the MLM can be trained in a supervised manner, with the training image sequences being annotated with the help of human experts. The human experts have, for example, themselves carried out the movement of the at least one medical device during the generation of the training image sequences. For example, the human experts can assign each training image sequence a subjectively perceived risk that the unexpected correlation exists. The assignment can be binary ("yes" versus "no") or according to more than two values. For example, two to four values for the risk parameter seem reasonable.However, it is also possible to determine the annotations additionally or alternatively based on physical measurements captured during the movement of the at least one medical device during the generation of the training image sequences. For example, the movement amplitude or the force applied to perform the causal movement could be measured.

[0031] Once the MLM has been trained in this way, it can predict the risk index based on the image sequence. The risk index indicates how high the risk is that the unexpected connection exists or is imminent. If such an unexpected connection exists or is imminent, the risk of an unexpected movement of the at least one medical device, for example a slipping or an uncontrolled propulsion of the at least one medical device, is also increased. The risk index can be displayed or transmitted directly or in a modified or processed form to the person making the causal movement or to the corresponding robot and / or third parties. In response, the person or robot can react accordingly to prevent a sudden slipping or the like of the at least one medical device.Accordingly, with the aid of the present invention, the risk of unexpected movement of the at least one medical device can be monitored automatically and live during a vascular intervention, thereby preventing such unexpected movement. Consequently, the risk to the patient's health and / or the risk of delaying the vascular intervention due, for example, to necessary corrections to the positioning of the at least one medical device or the like is reduced.

[0032] According to at least one embodiment, the input data includes metadata. In particular, the risk indicator is determined based on the metadata.

[0033] The metadata contains, for example, information that describes circumstances or boundary conditions during the generation of the images of the image sequence. In such embodiments, the MLM is trained to determine the risk index based at least on the image sequence and the metadata. One advantage of this is that fewer training data sets with corresponding training image sequences need to be processed during training of the MLM, since the additional training metadata provides the MLM with the context for the image sequence, thus enabling more efficient training. It should be noted that although the aforementioned advantage of more efficient training becomes apparent during the training phase, the manner in which training is carried out also influences the application phase of the MLM, namely in this case in the way that the corresponding metadata must also be provided in the application phase.

[0034] According to at least one embodiment, the metadata contains patient characteristics of the patient. The metadata is, for example, in numerical form.

[0035] This provides the MLM with important context when generating the image sequence, enabling particularly efficient training. Patient characteristics can include age, height, weight, geometric dimensions of the patient, and so on, as well as disease diagnoses or other anatomical characteristics of the patient. Disease diagnoses can include, in particular, vascular diseases such as ICAD (intracranial atherosclerotic disease) or other arteriosclerotic diseases. Other relevant anatomical characteristics can include, for example, geometric properties of the vascular structure, the type of aortic arch, and so on.

[0036] According to at least one embodiment, the metadata includes device properties of the at least one medical device.

[0037] This provides the MLM with important context when generating the image sequence, enabling particularly efficient training. The device properties can include, for example, a type or kind of the at least one medical device and / or an elasticity or rigidity of the at least one medical device and / or a surface quality, such as a surface structuring or surface roughness, and / or a diameter of the at least one medical device and / or a working length of the at least one medical device and / or shape parameters of the at least one medical device, and so on.

[0038] In the case of two or more medical devices, the device properties may include properties of the two or more medical devices as a whole and / or properties of the individual medical devices.

[0039] According to at least one embodiment, the at least one medical device includes a first medical device and a second medical device, and the device properties include data relating to an arrangement of the first medical device and the second medical device with respect to one another.

[0040] In corresponding embodiments, this can also be extended analogously to more than two medical devices, so that the device properties include, for example, data relating to a respective arrangement of the individual medical devices.

[0041] According to at least one embodiment, the metadata includes intervention data relating to a previous course of the vascular intervention.

[0042] This provides the MLM with important context when generating the image sequence, enabling particularly efficient training. The previous course of the vascular intervention includes, in particular, a period during which the image sequence was generated. For example, the previous course ends with the generation of a final image from the consecutive images of the image sequence.

[0043] According to at least one embodiment, the intervention data includes a length of a part of the at least one medical device located in the body when the image sequence is generated.

[0044] This length has a significant impact on the expected behavior of at least one medical device and on the interpretation of the image sequence. Therefore, including this length as part of the metadata can enable particularly efficient training.

[0045] According to at least one embodiment, the intervention data includes data relating to a feed force or tensile force applied to the at least one medical device during the generation of the image sequence or a torque applied to the at least one medical device during the generation of the image sequence.

[0046] The feed force or tensile force, or the torque, have a significant influence on the expected behavior of the at least one medical device and on the interpretation of the image sequence. Therefore, incorporating this data as part of the metadata can enable particularly efficient training. The feed force or tensile force, or the torque, can be determined, for example, using appropriate measuring devices, even during a movement of the at least one medical device at the proximal end. Integrated sensors can be used in the case of a robot or the like performing the movement.

[0047] According to at least one embodiment, the MLM includes a recurrent convolutional neural network, RCNN.

[0048] The RCNN can be understood as an ANN that contains one or more convolutional layers as well as one or more recurrent units. This allows the images of the image sequence to be processed efficiently, while also leveraging the advantages of an RNN. In principle, the MLM could process the images of the image sequence without the information regarding the order of the images in the image sequence. In this case, CNNs or transformer networks, for example, could also be used, which are not recurrent but rather pure feed-forward networks. One advantage of using an RCNN is that it can take the temporal sequence of the images in the image sequence into account, which ultimately makes it possible to provide the risk index in real time during the vascular intervention.

[0049] According to at least one embodiment, the RCNN comprises a convolution module adapted to convert the image sequence into a sequence of feature sets by applying the convolution module to the image sequence.

[0050] The convolution module includes, in particular, at least one convolutional layer and / or is configured as a CNN. By applying the convolution module to the image sequence, the convolution module generates, in particular, for each image of the image sequence, exactly one feature set of the sequence of feature sets. Converting the image sequence into the sequence of feature sets can also be referred to as feature extraction from the images of the image sequence or as encoding the images of the image sequence.

[0051] According to at least one embodiment, the RCNN comprises a recurrence module adapted to generate a common feature set by applying the recurrence module to the sequence of feature sets.

[0052] The recurrence module is, in particular, an RNN, for example, an LSTM. The convolution module and the recurrence module together can also be considered, for example, the encoder module or feature encoder of the RCNN.

[0053] According to at least one embodiment, the RCNN comprises a prediction module adapted to predict the risk index depending on the common feature set.

[0054] The prediction module can be a decoder module. The prediction module can be an ANN for classification. In this case, the risk index can be determined using binary classification or multiclass classification. Alternatively, the risk index can be predicted using regression if the prediction module is an ANN for regression.

[0055] In embodiments where the input data includes metadata, an augmented feature set may, for example, be generated by combining the common feature set and the metadata, for example, concatenating or otherwise fusing them. The prediction module is adapted to predict the risk metric by applying the prediction module to the augmented feature set. Alternatively, the feature sets of the sequence of feature sets may each be combined with the metadata, and the common feature set may be generated based on the combinations. The prediction module may then predict the risk metric by applying the prediction module to the common feature set.

[0056] For example, in embodiments where the input data does not include the metadata, the prediction module may be applied to the common feature set to predict the risk metric.

[0057] According to at least one embodiment, a nominal relationship between the causal movement and the resulting movement is a linear relationship. The unexpected relationship between the causal movement and the resulting movement corresponds to a relationship that deviates from the nominal relationship by more than a predetermined tolerance.

[0058] To define such a tolerance, the relationship between the causal movement and the resulting movement can be approximated linearly, for example, and a difference between the relationship between the causal movement and its linear approximation can be determined. The relationship is then an unexpected relationship if the difference is greater than a specified threshold. However, other methods can also be used to quantify the deviation of the relationship between the causal movement and the resulting movement from the nominal relationship and compare it with the specified tolerance.

[0059] According to at least one embodiment, the causal movement and the resulting movement are each feed movements or each pulling movements or each rotational movements.

[0060] According to at least one embodiment, the at least one medical device includes one or more vascular catheters and / or one or more guidewires.

[0061] According to at least one embodiment, a user output is generated and output depending on the risk characteristic.

[0062] The output can, for example, directly reflect the risk index or reflect a category of the risk index (e.g., "high risk," "medium risk," "low risk," or the like). The output can, for example, be provided visually and / or acoustically and / or haptically via a corresponding output device, in particular a display device, a loudspeaker, and / or a haptic output device. This allows the person performing the vascular intervention and / or a person monitoring or observing the vascular intervention to be immediately alerted to a potentially increased risk of unexpected movement and to react accordingly.

[0063] The computer-implemented method according to the invention neither involves the introduction of the at least one medical device into the patient's body, nor the creation of the artificial body opening, for example, for introducing the at least one medical device into the body, nor the execution of the causal movement, nor any other interactions with the patient's body. This generally also applies to non-purely computer-implemented embodiments of the method according to the invention.

[0064] According to a further aspect of the invention, however, a method is also provided which, in addition to the steps of an embodiment of a computer-implemented method according to the invention for estimating the risk of unexpected movement of the at least one medical device during vascular intervention, includes one or more of the following steps: - creating the artificial body opening for introducing the at least one medical device into the patient's body; and / or - introducing the at least one medical device into the patient's body and into the vascular structure; and / or - moving the at least one medical device in the vascular structure by causing a corresponding causal movement at the proximal end of the at least one medical device during the generation of the image sequence.

[0065] According to a further aspect of the invention, a computer-implemented training method is provided for providing a trained MLM for use in a computer-implemented method according to the invention for estimating a risk of unexpected movement of at least one medical device during a vascular intervention. The MLM is obtained in an untrained or partially trained state. A training image sequence of successive training images is obtained, which depict or simulate a vascular structure and at least one device moving in the vascular structure, wherein the at least one medical device is inserted into a patient's body at a body opening.A training risk index is determined for whether an unexpected connection exists or is imminent between a causal movement of the at least one medical device at an end of the at least one medical device that is proximal to the body opening and a resulting movement at an end of the at least one medical device that is distal to the body opening. Determining the training risk index involves applying the MLM to training input data that contains the image sequence. A predetermined loss function is evaluated depending on the training risk index and a predetermined ground truth value for the training image sequence, in particular depending on a difference between the training risk index and the ground truth value. The MLM is updated depending on a result of the evaluation of the loss function.

[0066] If the MLM is an ANN, updating the MLM includes, in particular, updating weights of the MLM, for example, using a backpropagation algorithm.

[0067] The training images can be obtained, for example, by simulating a vascular intervention and / or from actual interventions on humans or animals and / or by vascular interventions on phantom objects.

[0068] Common regression losses or classification losses can be used as the loss function, as explained in the publications mentioned above.

[0069] Unless otherwise stated, all steps of the computer-implemented training method can be performed by a further data processing system that includes at least one further data processing device. In particular, the at least one further data processing device is configured or adapted to carry out the steps of the computer-implemented training method. For this purpose, the at least one further data processing device can, for example, store a further computer program that includes instructions that, when executed by the at least one further data processing device, cause the at least one further data processing device to carry out the computer-implemented training method.

[0070] In the event that the at least one further data processing device includes two or more further data processing devices, certain steps performed by the at least one further data processing device can also be understood as different further data processing devices performing different steps or different parts of a step. In particular, it is not necessary for each further data processing device to perform the steps. In other words, the execution of the steps can be distributed among the two or more further data processing devices.

[0071] Each embodiment of the computer-implemented training method results in a corresponding embodiment of a training method that is not purely computer-implemented by including corresponding steps for generating the training image sequence and / or the ground truth value.

[0072] According to at least one embodiment of the computer-implemented training method, the training input data includes training metadata which includes patient characteristics and / or device characteristics of the at least one medical device and / or intervention data.

[0073] In this regard, reference is made to the explanations of the metadata relating to the computer-implemented procedure for estimating a risk of unexpected movement of at least one medical device.

[0074] According to at least one embodiment of the computer-implemented method for estimating a risk for an unexpected movement of at least one medical device, the MLM has been or is being trained using a computer-implemented training method according to the invention.

[0075] According to a further aspect of the invention, a data processing system is provided which is adapted to carry out a computer-implemented method according to the invention for estimating a risk for an unexpected movement of at least one medical device.

[0076] According to a further aspect of the invention, an X-ray imaging system is provided, which includes an X-ray source and an X-ray detector, as well as a data processing system according to the invention. The X-ray imaging system can, for example, be an X-ray angiography system.

[0077] Further embodiments of the X-ray imaging system according to the invention follow directly from the various embodiments of the methods according to the invention, and vice versa. In particular, individual features and corresponding explanations as well as advantages relating to the various embodiments of the methods according to the invention can be transferred analogously to corresponding embodiments of the X-ray imaging system according to the invention. In particular, the X-ray imaging system according to the invention is designed or programmed to carry out a method according to the invention for estimating a risk of an unexpected movement of at least one medical device during a vascular intervention, or it carries out such a method.

[0078] According to a further aspect of the invention, a further data processing system is provided which is adapted to carry out a computer-implemented training method according to the invention.

[0079] According to a further aspect of the invention, a computer program with instructions is provided. When executed by a data processing system, the instructions cause the data processing system to perform a method according to the invention for estimating a risk of an unexpected movement of at least one medical device.

[0080] The instructions can be provided, for example, as program code. The program code can be provided, for example, as binary code or assembly code and / or as source code of a programming language, for example, C, and / or as a program script, for example, Python.

[0081] According to a further aspect of the invention, a further computer program with further instructions is provided. When the further instructions are executed by a further data processing system, the further instructions cause the further data processing system to perform a training method according to the invention.

[0082] The additional instructions can be provided, for example, as program code. The program code can be provided, for example, as binary code or assembly code and / or as source code of a programming language, for example, C, and / or as a program script, for example, Python.

[0083] According to a further aspect of the invention, a computer-readable storage medium is provided which stores a computer program according to the invention and / or a further computer program according to the invention.

[0084] The computer program, the further computer program and the computer-readable storage medium are each computer program products with the instructions or the further instructions, respectively.

[0085] Above and below, the inventive solution is described both with respect to the claimed systems and with respect to the claimed methods. Features, advantages, or alternative embodiments may be assigned to the other claimed subject matter, and vice versa. In other words, the claims and embodiments for the systems may be enhanced by features described or claimed in connection with the respective methods. In this case, the functional features of the method are implemented by physical units of the system.

[0086] Furthermore, the inventive solution is described above and below with respect to methods and systems for estimating a risk of an unexpected movement of at least one medical device, as well as with respect to methods and systems for providing a trained MLM. Features, advantages, or alternative embodiments can be assigned to the other claimed subject matter and vice versa. In other words, claims and embodiments for providing a trained MLM can be improved with features described or claimed in connection with the estimation of the risk of the unexpected movement.In particular, the data sets used in the methods and systems may have the same properties and characteristics as the corresponding data sets used in the methods and systems for providing a trained MLM, and the trained MLMs provided by the respective methods and systems may be used in the methods and systems for estimating the risk for the unexpected movement.

[0087] Further features and combinations of features of the invention will become apparent from the figures and their description, as well as from the claims. In particular, further embodiments of the invention do not necessarily have to contain all features of one of the claims. Further embodiments of the invention may have features or combinations of features that are not mentioned in the claims.

[0088] The invention is explained in more detail below with reference to specific embodiments and associated schematic drawings. In the figures, identical or functionally equivalent elements may be provided with the same reference numerals. The description of identical or functionally equivalent elements may not necessarily be repeated for different figures.

[0089] The figures show Fig. 1 is a schematic representation of an exemplary embodiment of an X-ray imaging system according to the invention; Fig. 2 shows a schematic flow diagram of an exemplary embodiment of a computer-implemented method according to the invention for estimating a risk of unexpected movement of at least one medical device during a vascular intervention; Fig. 3 shows a schematic flow diagram of a further exemplary embodiment of a computer-implemented method according to the invention for estimating a risk of an unexpected movement of at least one medical device during a vascular intervention; Fig. 4 shows a schematic flow diagram of a further exemplary embodiment of a computer-implemented method according to the invention for estimating a risk of an unexpected movement of at least one medical device during a vascular intervention; Fig. 5 a schematic representation of an ANN; Fig. 6 a schematic representation of a CNN; and Fig. 7 a schematic representation of an RNN.

[0090] In Fig. Figure 1 schematically illustrates an exemplary embodiment of an X-ray imaging system 1 according to the invention. The X-ray imaging system 1 comprises an X-ray source 4 and an X-ray detector 3, by means of which X-ray images 7 of a patient 5 can be generated.

[0091] The X-ray imaging system 1 can be used, for example, in a vascular intervention on a vascular structure of the patient 5, in which at least one medical device, for example at least one guide wire and / or at least one vascular catheter for minimally invasive interventions, which is introduced into the body of the patient 5 via a body opening, is moved in the vascular structure of the patient 5 in order to bring the at least one medical device to a target position in the vascular structure.The vascular intervention is carried out in particular with X-ray support, wherein corresponding X-ray images 7 can be generated in respective successive frames and displayed on a display device of the X-ray imaging system 1, so that, for example, a treating person can, among other things, observe and track the movement of the at least one medical device at an end of the at least one medical device that is distal with respect to the body orifice, which is also referred to as fluoroscopy. To do this, the person moves the at least one medical device at an end of the at least one medical device that is proximal with respect to the body orifice, which is also referred to as the causal movement of the at least one medical device. The causal movement then leads to a resulting movement of the at least one medical device at an end of the at least one medical device that is distal with respect to the body orifice.

[0092] For example, in neurovascular interventions, stacks of medical devices, particularly catheters for minimally invasive therapy, are sometimes used for possible treatments such as aneurysm coiling, stent placement, embolization delivery, and so on. To maneuver the catheters to the target vessel, a guidewire can first be inserted and slowly advanced into the target vessel, avoiding small vascular branches known as perforators and, for example, an aneurysm dome, as this could lead to perforation or rupture and cause bleeding.

[0093] After the guidewire, catheters, such as guiding catheters, intermediate catheters, and / or microcatheters, can be introduced at an offset to reach the target vessel. Two catheters are referred to as a biaxial stack, while three catheters are referred to as a triaxial stack. The number of medical devices, especially catheters, depends, for example, on the anatomy and the location of the anomaly to be treated.

[0094] If energy builds up in the system due to friction or resistance in the vascular structure and / or between the medical devices, this can, for example, lead to slippage of the at least one medical device or to other undesirable and unexpected movements of the at least one medical device.

[0095] For example, in X-ray imaging systems 1, the treating person relies on detecting the impending unexpected movement on the current X-ray image 7 and / or based on the resistance offered by at least one medical device to further advancement. This requires considerable expertise and experience on the part of the treating person.

[0096] The X-ray imaging system 1 according to the invention, on the other hand, has a data processing system 2 which is adapted to carry out a computer-implemented method for estimating a risk of an unexpected movement of at least one medical device during a vascular intervention. Fig. 2 shows a schematic flow diagram of such a method.

[0097] The data processing system 2 receives an image sequence 7 of consecutive images, for example x-ray images, which depict the vascular structure as well as at least one medical device moving within the vascular structure. In some embodiments, the images can also be biplanar x-ray images. The at least one medical device is inserted into the body of the patient 5 at the body opening. The data processing system 2 determines a risk characteristic 9 for the existence or imminence of an unexpected connection between the causal movement of the at least one medical device at the proximal end of the at least one medical device and the resulting movement of the at least one medical device at the distal end of the at least one medical device. To this end, the data processing system 2 applies a trained MLM 8 to input data 6 containing the image sequence 7.

[0098] Depending on the risk parameter 9, the data processing system 2 can, for example, control an output device to generate and output a user output. Based on this, the treating person can adjust the movement of the at least one medical device accordingly to reduce or eliminate sagging, for example, by rotating, retracting, or pushing the at least one medical device further forward.

[0099] For example, the unexpected relationship between the causal motion and the resulting motion corresponds to a relationship that deviates from a nominal relationship by more than a specified tolerance. The nominal relationship can, for example, be a linear relationship.

[0100] Fig. 3 shows a schematic flow diagram of a further exemplary embodiment of a computer-implemented method according to the invention for estimating a risk for an unexpected movement of at least one medical device during a vascular intervention, which is based, for example, on the Fig. 2 illustrated embodiment.

[0101] Here, the MLM 8 is designed as an RCNN, which contains an encoder module 11 and a prediction module 18. The encoder module 11 includes a convolution module 12 designed as a CNN and a subsequent recurrence module 14, which can be designed, for example, as an LSTM.

[0102] The convolution module 12 is applied to the image sequence 7 and thereby converts the image sequence 7 into a corresponding sequence of feature sets 13. These serve as input data for the prediction module 18, which generates a common feature set 16 as output. The prediction module 18, in turn, is applied to this set, which predicts the risk index, for example, through classification or regression.

[0103] Fig. 4 shows a schematic flow diagram of a further exemplary embodiment of a computer-implemented method according to the invention for estimating a risk for an unexpected movement of at least one medical device during a vascular intervention, which is based on the Fig. 3 illustrated embodiment.

[0104] The encoder module 11 is constructed as Fig. 3. However, in addition to the image sequence 7, in this embodiment 10, the input data 6 includes metadata in numerical form, which includes, for example, patient characteristics of the patient 5 and / or device characteristics of the at least one medical device and / or intervention data relating to a previous course of the vascular intervention. These are fused, for example, with the common feature set 16, for example by concatenation, resulting in a supplemented feature set 15. The prediction module 18 is then applied to the supplemented feature set 15 to predict the risk parameter.

[0105] Fig. 5 shows an embodiment of an artificial neural network, ANN, 800. The ANN 800 includes nodes 820, ..., 832 and edges 840, ..., 842, where each edge 840, ..., 842 is a directed connection from a first node 820, ..., 832 to a second node 820, ..., 832. In general, the first node 820, ..., 832 and the second node 820, ..., 832 are different nodes 820, ..., 832. However, it is also possible that the first node 820, ..., 832 and the second node 820, ..., 832 are identical. In Fig. 5, for example, edge 840 is a directed connection from node 820 to node 823, and edge 842 is a directed connection from node 830 to node 832. An edge 840, ..., 842 from a first node 820, ..., 832 to a second node 820, ..., 832 is also referred to as an incoming edge for the second node 820, ..., 832 and an outgoing edge for the first node 820, ..., 832.

[0106] In this example, the nodes 820, ..., 832 of the ANN 800 may be arranged in layers 810, ..., 813, where the layers may have an intrinsic order introduced by the edges 840, ..., 842 between the nodes 820, ..., 832. In particular, the edges 840, ..., 842 may only exist between adjacent layers of nodes. In the example shown, there is an input layer 810 consisting only of the nodes 820, ..., 822 with no incoming edges, an output layer 813 consisting only of the nodes 831, 832 with no outgoing edges, and hidden layers 811, 812 between the input layer 810 and the output layer 813. In general, the number of hidden layers 811, 812 can be chosen arbitrarily. For a multilayer perceptron (MLP), this number is at least one. The number of nodes 820, ..., 822 within the input layer 810 typically refers to the number of input values of the artificial neural network 800, and the number of nodes 831, 832 within the output layer 813 typically refers to the number of output values of the artificial neural network 800.

[0107] In particular, each node 820, ..., 832 of the artificial neural network 800 can be assigned a real number as a value. Here, x denotes (n) ithe value of the i-th node 820, ..., 832 of the n-th layer 810, ..., 813. The values of the nodes 820, ..., 822 of the input layer 810 correspond to the input values of the artificial neural network 800. The values of the nodes 831, 832 of the output layer 813 correspond to the output value of the artificial neural network 800. In addition, each edge 840, ..., 842 can have a weight that is a real number. In particular, the weight is a real number within the interval [-1, 1] or within the interval [0, 1]. Here, w denotes (m,n) i,j the weight of the edge between the i-th node 820, ..., 832 of the m-th layer 810, ..., 813 and the j-th node 820, ..., 832 of the n-th layer 810, ..., 813. In addition, the abbreviation w (n) i,j for the weight w (n,n+1) i,jdefined. In particular, to calculate the output values of the neural network 800, the input values are propagated through the neural network 800. In particular, the values of the nodes 820, ..., 832 of the (n+1)-th layer 810, ..., 813 can be calculated based on the values of the nodes 820, ..., 832 of the n-th layer 810, ..., 813 as xj(n+1)=f(∑ixi(n)wi,j(n)).

[0108] Therein, the function f is referred to as the transfer function or activation function. Well-known transfer functions are step functions, sigmoid functions, for example, the logistic function, the generalized logistic function, the hyperbolic tangent, the arctangent function, the error function, the smoothstep function, or rectifier functions. The transfer function is mainly used for normalization. In particular, the values are propagated layer by layer through the neural network 800, where the values of the input layer 810 are given by the input of the neural network 800, where the values of the first hidden layer 811 can be calculated based on the values of the input layer 810 of the neural network 800, where the values of the second hidden layer 812 can be calculated based on the values of the first hidden layer 811, and so on.

[0109] To get the values w (m,n)i,j for the edges, the neural network 800 must be trained with training data. The training data includes, in particular, training input data and training output data (denoted as t i ). In a training step, the neural network 800 is applied to the training input data to generate calculated output data. In particular, the training data and the calculated output data comprise a number of values that corresponds to the number of nodes of the output layer. In particular, a comparison between the calculated output data and the training data is used to recursively adjust the weights within the neural network 800 (backpropagation algorithm). In particular, the weights are changed according to the following formula wi,j'(n)=wi,j(n)−γδj(n)xi(n), where γ is a predefined learning rate, and the numbers δ (n) j can be calculated recursively as δj(n)=(∑kδk(n+1)wj,k(n+1))f'(xi(n)wi,j(n)) based on δ (n+1) j , if the (n+1)-th layer is not the output layer 813, and δj(n)=(xj(n+1)+tj(n+1))f'(xi(n)wi,j(n)), if the (n+1)-th layer is the output layer 813, where f is the first derivative of the activation function, and t (n+1) j is the comparison training value for the j-th node of the output layer 813.

[0110] A convolutional neural network (CNN) is an ANN that uses a convolution operation instead of general matrix multiplication in at least one of its layers. These layers are called convolutional layers. Specifically, a convolutional layer performs a dot product of one or more convolution kernels on the input data of the convolutional layer, where the entries of the one or more convolution kernels are parameters or weights that can be adjusted through training. In particular, one can use the Frobenius inner product and the ReLU activation function. A convolutional neural network may include additional layers, such as pooling layers, fully connected layers, and / or normalization layers.

[0111] Using convolutional neural networks allows for very efficient input processing, as a convolution operation based on different kernels can extract different image features. Adjusting the weights of the convolution kernel allows for the relevant image features to be determined during training. Furthermore, due to the shared weights in the convolution kernels, fewer parameters need to be trained, preventing overfitting during the training phase and allowing for faster training or more layers in the network, thus improving network performance.

[0112] Fig. 6 shows an exemplary embodiment of a convolutional neural network 700. In the illustrated embodiment, the convolutional neural network 700 includes an input node layer 710, a convolutional layer 711, a pooling layer 713, a fully connected layer 714, and an output node layer 716, as well as hidden node layers 712, 714. Alternatively, the convolutional neural network 700 may also include multiple convolutional layers 711, multiple pooling layers 713, and / or multiple fully connected layers 715, as well as other types of layers. The order of the layers can be chosen arbitrarily; typically, fully connected layers 715 are used as the last layers before the output layer 716.

[0113] In particular, in a convolutional neural network 700, the nodes 720, 722, 724 of a node layer 710, 712, 714 can be viewed as a d-dimensional matrix or as a d-dimensional image. In particular, in the two-dimensional case, the value of the node 720, 722, 724 indexed by i and j in the nth node layer 710, 712, 714 can be denoted as x(n)[i, j]. However, the arrangement of the nodes 720, 722, 724 of a node layer 710, 712, 714 as such has no influence on the calculations performed within the convolutional neural network 700, since these are determined solely by the structure and weights of the edges.

[0114] A convolutional layer 711 is a connecting layer between a front node layer 710 with node values x(n-1) and a back node layer 712 with node values x(n). A convolutional layer 711 is particularly characterized by the structure and weights of the incoming edges, which form a convolution operation based on a certain number of kernels. In particular, the structure and weights of the edges of the convolutional layer 711 are chosen such that the values x(n) of the nodes 722 of the back node layer 712 are calculated as a convolution x(n) = K * x(n-1) based on the values x(n-1) of the nodes 720 of the front node layer 710, where the convolution * is defined in the two-dimensional case as x(n)[i,j]=(K∗x(n−1))[i,j]=∑i'∑j'K[i',j']⋅x(n−1)[i−i',j−j'].

[0115] The kernel K is a d-dimensional matrix, in the present example a two-dimensional matrix, which is usually small compared to the number of nodes 720, 722, for example, a 3x3 matrix or a 5x5 matrix. This means, in particular, that the weights of the edges in the convolutional layer 711 are not independent, but are chosen to yield the aforementioned convolution equation. In particular, for a kernel that is a 3x3 matrix, there are only 9 independent weights, with each entry of the kernel matrix corresponding to an independent weight, regardless of the number of nodes 720, 722 in the front node layer 710 and the back node layer 712. In general, convolutional neural networks 700 use node layers 710, 712, 714 with a plurality of channels, in particular due to the use of a plurality of kernels in the convolutional layers 711.In these cases, the node layers can be considered as (d+1)-dimensional matrices, where the first dimension indexes the channels. The effect of a convolutional layer 711 is then defined in a two-dimensional example as: xb(n)[i,j]=∑a(Ka,b∗xa(n−1)[i,j])=∑a∑i'∑j'Ka,b[i',j']⋅aa(n−1)[i−i',j−j'], where xa(n) corresponds to the a-th channel of the previous layer 710, xb(n) corresponds to the b-th channel of the subsequent node layer 712 and K a,b corresponds to one of the kernels. If a convolutional layer 711 acts on a preceding node layer 710 with A channels and outputs a subsequent node layer 712 with B channels, there are A, B-independent d-dimensional kernels K a,b .

[0116] In general, activation functions can be used in convolutional neural networks 700. In this embodiment, ReLU (rectified linear unit) is used, with R(z) = max(0, z), so that the effect of the convolutional layer 711 in the two-dimensional example xb(n)[i,j]=R(∑a(Ka,b∗xa(n−1)[i,j])=R(∑a∑i'∑j'Ka,b[i',j']⋅xa(n−1)[i−i',j−j']) It is also possible to use other activation functions, such as ELU (Exponential Linear Unit), LeakyReLU, Sigmoid, Tanh, or Softmax.

[0117] In the illustrated embodiment, the input layer 710 includes 36 nodes 720 arranged in a two-dimensional 6x6 matrix. The first hidden node layer 712 includes 72 nodes 722 arranged as two-dimensional 6x6 matrices, each of which is the result of convolving the input layer values with a 3x3 kernel within the convolutional layer 711. Equivalently, the nodes 722 of the first hidden node layer 712 can be interpreted as a three-dimensional 2x6x6 matrix, where the first dimension corresponds to the channel dimension.

[0118] An advantage of using convolutional layers 711 is that a spatially local correlation of the input data can be exploited by enforcing a local connectivity pattern between the nodes of adjacent layers, in particular by having each node only connected to a small range of the nodes of the previous layer.

[0119] A pooling layer 713 is a connecting layer between a preceding node layer 712 with node values x(n-1) and a subsequent node layer 714 with node values x(n). A pooling layer 713 can be characterized in particular by the structure and weights of the edges and the activation function, which form a pooling operation based on a nonlinear pooling function f. For example, in the two-dimensional case, the values x(n) of the nodes 724 of the subsequent node layer 714 can be calculated based on the values x(n-1) of the nodes 722 of the anterior node layer 712 as follows: xb(n)[i,j]=f(xb(n−1)[id1,jd2],…, xb(n−1)[(i+1)d1−1,(j+1)d2−1]).

[0120] In other words, by using a pooling layer 713, the number of nodes 722, 724 can be reduced by replacing a number d1-d2 of neighboring nodes 722 in the preceding node layer 712 with a single node 722 in the subsequent node layer 714, which is calculated as a function of the values of said number of neighboring nodes. The pooling function f can, in particular, be the max function, the mean, or the L2 norm. In particular, in a pooling layer 713, the weights of the incoming edges are fixed and are not changed by training.

[0121] The advantage of using a pooling layer 713 is that the number of nodes 722, 724 and the number of parameters are reduced. This leads to a reduction in the computational effort in the network and a control of overfitting.

[0122] In the illustrated embodiment, pooling layer 713 is a max-pooling layer, where four neighboring nodes are replaced by only one node, with the value being 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-dimensional matrices, reducing the number of nodes from 72 to 18.

[0123] In general, the last layers of a convolutional neural network 700 may be fully connected layers 715. A fully connected layer 715 is a connecting layer between a preceding node layer 714 and a succeeding node layer 716. A fully connected layer 713 may be characterized in that a majority, in particular all, of the edges are present between the nodes 714 of the preceding node layer 714 and the nodes 716 of the succeeding node layer, and wherein the weight of each of these edges can be individually adjusted.

[0124] In this embodiment, the nodes 724 of the front node layer 714 of the fully connected layer 715 are represented both as two-dimensional matrices and additionally as non-contiguous nodes displayed as a line of nodes, with the number of nodes reduced for clarity. This process is also referred to as flattening. In this embodiment, the number of nodes 726 in the subsequent node layer 716 of the fully connected layer 715 is smaller than the number of nodes 724 in the previous node layer 714. Alternatively, the number of nodes 726 may be equal to or greater.

[0125] Additionally, in this embodiment, the softmax activation function is used within the fully connected layer 715. By applying the softmax function, the sum of the values of all nodes 726 of the output layer 716 is 1, and all values of all nodes 726 of the output layer 716 are real numbers between 0 and 1. In particular, when using the convolutional neural network 700 to categorize input data, the values of the output layer 716 can be interpreted as the probability that the input data falls into one of the various categories.

[0126] In particular, convolutional neural networks 700 can be trained based on the backpropagation algorithm. To prevent overfitting, regularization techniques can be used, such as omitting nodes 720, ..., 724, stochastic pooling, the use of artificial data, weight decay based on the L1 or L2 norm, or max-norm constraints.

[0127] Fig. Figure 7 shows a recurrent MLM (F), specifically a recurrent neural network (RNN). A recurrent MLM (F) is a machine learning model whose output depends not only on the current input value and the MLM parameters adjusted through the training process, but also on a hidden state vector. The hidden state vector is based on previous inputs used for the recurrent MLM (F). In particular, the recurrent MLM (F) may include additional memory states or additional structures that incorporate time delays or include feedback loops.

[0128] An RNN can be described as an ANN in which the connections between the nodes form a directed graph along a temporal sequence. In particular, an RNN can be interpreted as a directed acyclic graph. In particular, the RNN can be a finite impulse RNN (FIR) or an infinite impulse RNN (INFINITE IMPULSE RNN). A FIR can be unrolled and replaced by a pure feedforward ANN, whereas an FIR cannot be unrolled and replaced by a pure feedforward ANN. In particular, the training of an RNN can be based on the BPTT algorithm (backpropagation through time algorithm), the RTRL algorithm (real-time recurrent learning), and / or genetic algorithms.

[0129] In the Fig.Figure 7 shows the schematic structure of a recurrent MLM F in a recurrent representation on the left and an unfolded representation on the right. The recurrent MLM F receives as input several input data sets x1,...,x N and generates a corresponding number of output data sets y1, ...,y N . In addition, the output depends on a so-called hidden vector h1,..., h N , which implicitly includes information about input data sets previously used as input to the recurrent MLM F. By using these hidden vectors h1, ..., h N the sequentiality of the input data sets can be exploited. In a single processing step, the recurrent MLM F receives as input the hidden vector h created in the previous step n-1 and an input data set x n In this step, the recurrent MLM F produces as output an updated hidden vector h nand an output data set y n In other words, a processing step calculates (yn,hn)=F(xn,hn−1). or by dividing the recurrent MLM F into a part F y , which calculates the output data, and F h , which calculates the hidden vector, a processing step calculates yn=Fy(xn,hn−1), hn=Fh(xn,hn−1).

[0130] For the initial processing step n = 1, h n-1 = h0 can be randomly chosen, or all entries can be set to zero. The parameters of the recurrent MLM F, previously trained based on training datasets, do not change between the different processing steps. In particular, the output data and the hidden vector of a processing step can depend on all previous input datasets used in the preceding steps: yn=Fy(xn,Fh(xn−1,hn−1)). hn=Fh(xn,Fh(xn−1,hn−1)).

[0131] Regardless of the grammatical gender of a particular term, persons with male, female or other gender identity are included.

Claims

[1] Computer-implemented method for estimating a risk of unexpected movement of at least one medical device during a vascular intervention, wherein - an image sequence (7) of successive images depicting a vascular structure and at least one medical device moving in the vascular structure is obtained, wherein the at least one medical device is inserted into a body of a patient (5) at a body opening; - a risk index (9) is determined for the existence or imminence of an unexpected connection between a causal movement of the at least one medical device at a proximal end of the at least one medical device with respect to the body opening and a resulting movement of the at least one medical device at a distal end of the at least one medical device with respect to the body opening; and - determining the risk index (9) involves applying a trained machine learning model, MLM, (8) to input data (6) containing the image sequence (7). [2] Computer-implemented method according to claim 1, wherein the input data (6) include metadata (10) which include patient characteristics of the patient (5) and / or device characteristics of the at least one medical device and / or intervention data relating to a previous course of the vascular intervention. [3] Computer-implemented method according to claim 2, wherein the intervention data - a length of a part of the at least one medical device located in the body during generation of the image sequence (7); and / or - data relating to a feed force or tensile force applied to the at least one medical device during the generation of the image sequence (7) or a torque applied to the at least one medical device during the generation of the image sequence (7). [4] Computer-implemented method according to one of claims 2 or 3, wherein - the device properties include elasticity and / or stiffness and / or a surface quality of the at least one medical device; and / or - wherein the at least one medical device includes a first medical device and a second medical device, and the device properties include data relating to an arrangement of the first medical device and the second medical device with respect to one another. [5] A computer-implemented method according to any one of the preceding claims, wherein the MLM (8) includes a recurrent convolutional neural network, RCNN. [6] A computer-implemented method according to claim 5, wherein - the image sequence (7) is converted into a sequence of feature sets (13) by applying a convolution module (12) of the RCNN to the image sequence (7); - a common feature set (16) is generated by applying a recurrence module (14) of the RCNN to the sequence of feature sets (13); and - depending on the common feature set (16), the risk characteristic value (9) is predicted by a prediction module (18) of the RCNN. [7] Computer-implemented method according to claim 6 and one of claims 2 to 4, wherein an augmented feature set is generated by combining the common feature set (16) and the metadata (10), and the risk characteristic value (9) is predicted by applying the prediction module (18) to the augmented feature set (15). [8] Computer-implemented method according to one of the preceding claims, wherein - a nominal relationship between the causal movement and the resulting movement is a linear relationship; and - the unexpected relationship between the causal movement and the resulting movement corresponds to a relationship that deviates from the nominal relationship by more than a specified tolerance. [9] A computer-implemented method according to any one of the preceding claims, wherein the at least one medical device includes one or more vascular catheters and / or one or more guidewires. [10] Computer-implemented method according to one of the preceding claims, wherein a user output is generated and output depending on the risk characteristic value (9). [11] Computer-implemented training method for providing a trained MLM (8) for use in a computer-implemented method according to any one of the preceding claims, wherein - the MLM (8) is received in an untrained or partially trained state; - a training image sequence of successive training images is obtained, which depict or simulate a vascular structure and at least one device moving in the vascular structure, wherein the at least one medical device is inserted into a body of a patient (5) at a body opening; - a training risk indicator is determined that an unexpected connection exists or is imminent between a causal movement of the at least one medical device at an end of the at least one medical device that is proximal to the body opening and a resulting movement at an end of the at least one medical device that is distal to the body opening; - determining the training risk metric involves applying the MLM (8) to training input data containing the image sequence (7); - a given loss function is evaluated depending on the training risk value and a given ground truth value for the training image sequence; and - the MLM (8) is updated depending on a result of the evaluation of the loss function. [12] The computer-implemented training method of claim 11, wherein the training input data includes training metadata including patient characteristics and / or device characteristics of the at least one medical device and / or intervention data. [13] A computer-implemented method according to any one of claims 1 to 10, wherein the MLM (8) is trained using a computer-implemented training method according to any one of claims 11 or 12. [14] Data processing system (2) adapted to carry out a computer-implemented method according to any one of claims 1 to 10 or 13 and / or a computer-implemented training method according to any one of claims 11 or 12. [15] Computer program product comprising - instructions which, when executed by a data processing system (2), cause the data processing system (2) to carry out a computer-implemented method according to one of claims 1 to 10 or 13; and / or - further instructions which, when executed by a further data processing system (2), cause the further data processing system (2) to carry out a computer-implemented training method according to one of claims 11 or 12.

Images