Patent Application
20250000585
The present disclosure relates to identifying a vascular access site for inserting an interventional device in order to reach a target site in a vasculature. A computer-implemented method, a computer program product, and a system, are disclosed.
Interventional procedures in the vasculature can often be performed by choosing one of multiple vascular access sites to reach an intended target site. For example, percutaneous coronary interventions “PCI” procedures can often be performed with either radial access, or femoral access. The choice of the access site is typically made by a physician, and is based on multiple factors such as the ease of navigating an interventional device from the access site to the target site, and the expected outcome of performing a vascular intervention at the target site using the vascular access site.
By way of an example, a document by Jolly, S. et al., entitled “Radial versus femoral access for coronary angiography or intervention and the impact on major bleeding and ischemic events: A systematic review and meta-analysis of randomized trials,” American Heart Journal 157(1): 132-140 (2009), reports a meta-analysis of 23 randomized trials in which PCI procedures with radial access had a 73% lower chance of bleeding, a 0.4 day shorter hospital stay, and a trend toward reduced occurrence of stroke, as compared to those with femoral access. However, this meta-analysis also showed that procedures with radial access showed a trend toward increased inability to cross the target lesion with a wire, balloon, or stent.
Multiple factors may therefore influence a physician's choice of a vascular access site, including the location of the target site, the tortuosity of the vasculature leading to the target site, the presence of obstructions such as stenoses or calcification in the vasculature that may hamper navigation to the target site.
Due to the multitude of factors that affect the choice of a vascular access site to perform an interventional procedure at a target site, a physician may make an intuitive decision as to which access site to use. However, mentally resolving all the relevant information from various sources into an optimal choice of vascular access site is complex and time-consuming, while a poor choice of vascular access site may for example impact procedure time. This is because the target site may not be reachable and the procedure may have to be restarted from a new access site, lengthening the procedure time. Therefore, a need exists for a system to support a physician's choice of a vascular access site.
WO 2017/139894 A1 relates to systems and methods enabling a personalized solution for allowing more efficient access to the carotid artery (or vertebral arteries) in patients needing endovascular/neurointervention procedures. In particular, such system may comprise a scan data reader configured to determine the course of a vessel within a body based on a scan of the vessel; memory configured to store the physical properties of one or more vessel lines; a processor configured to determine: a route to reach a destination point within the vessel using a vessel line based on the determined vessel course, and whether it is possible to reach the destination point for each of the stored vessel lines. Accordingly, a user is supported in selecting a vessel line, for example a catheter system including a catheter and a guidewire, that is able to reach the destination point along the determined vessel course.
Still, there remains room to improve the task of choosing a vascular access site in order to perform an interventional procedure.
According to one aspect of the present disclosure, a computer-implemented method of identifying a vascular access site for inserting an interventional device in order to reach a target site in a vasculature, is provided. The method includes:
The result of these operations is to identify an optimal vascular access site for the target site. In particular, for an “optimal” vascular access site, not only is the target site reachable, but the risk of complications is as low as possible, as well. For example, if the vasculature is heavily calcified or has eccentric plaque or thrombus, the navigation of interventional devices risks complications such as bleeding, vascular injury including dissection, extravasation and rupture, or even embolus. Hence, a sub-optimal selection of a vascular access site can even harm the patient.
For this purpose, in the method of claim 1, success factors affecting the outcome of performing the vascular intervention, in particular factors that reduce the risk of complications, are taken into consideration and weighted. The success factors are determined for one or more potential access sites individually, and the resulting success metrics can be used as a basis to determine a ranking of the potential vascular access sites. The optimal vascular access site that has been identified accordingly may then be recommended to a physician. The optimal vascular access site is provided in a consistent and efficient manner, obviating some of the existing challenges in determining an optimal access site, while reducing the risk of complications during the vascular intervention
Further aspects, features, and advantages of the present disclosure will become apparent from the following description of examples, which is made with reference to the accompanying drawings.
Examples of the present disclosure are provided with reference to the following description and figures. In this description, for the purposes of explanation, numerous specific details of certain examples are set forth. Reference in the specification to “an example”, “an implementation” or similar language means that a feature, structure, or characteristic described in connection with the example is included in at least that one example. It is also to be appreciated that features described in relation to one example may also be used in another example, and that all features are not necessarily duplicated in each example for the sake of brevity. For instance, features described in relation to a computer implemented method, may be implemented in a system, in a corresponding manner.
In the following description, reference is made to computer-implemented methods that involve identifying a vascular access site for inserting an interventional device in order to reach a target site in a vasculature. Reference is made to an example of a PCI procedure in which femoral or radial access is used to reach the coronary arteries. However, it is to be appreciated that the disclosure is not limited to use in performing coronary interventional procedures, or to use with these example access sites. The target site, may therefore be anywhere in the vasculature.
A vascular access site is in general a superficial position on the body that provides access to an artery, or to a vein of the vasculature. Examples of vascular access sites include a femoral vascular access site, i.e. a superficial position on the body that provides access to the femoral artery, a radial vascular access site, i.e. a superficial position on the body that provides access to the radial artery, a brachial arterial access site, an axillary arterial access site, and a jugular access site. However, is to be appreciated that the disclosure is not limited to use with these examples of access sites, and that the access site may alternatively be at any suitable superficial position on the body that provides access to an artery, or to a vein of the vasculature.
Reference is made herein to a PCI procedure as an example of an interventional procedure that is performed at the target site. A PCI procedure is used to restore blood flow to a coronary artery, and typically involves the insertion of a balloon catheter from a femoral or radial access site, to a target site in the coronary arteries. The balloon is inflated inside a stent at the target site in order to open-up the artery. However, it is to be appreciated that the methods disclosed herein are not limited to this example PCI procedure, or to this example of an interventional device. Thus, the methods may be used with various types of interventional procedures, and as appropriate, with various types of interventional devices, such as, and without limitation: a guidewire, a catheter, a balloon catheter, an atherectomy device, an intravascular ultrasound “IVUS” imaging device, an Optical Coherence Tomography “OCT” imaging device, a blood pressure device and/or flow sensor device.
It is noted that the computer-implemented methods disclosed herein may be provided as a non-transitory computer-readable storage medium including computer-readable instructions stored thereon, which, when executed by at least one processor, cause the at least one processor to perform the method. In other words, the computer-implemented methods may be implemented in a computer program product. The computer program product can be provided by dedicated hardware, or hardware capable of running the software in association with appropriate software. When provided by a processor, the functions of the method features can be provided by a single dedicated processor, or by a single shared processor, or by a plurality of individual processors, some of which can be shared. The functions of one or more of the method features may for instance be provided by processors that are shared within a networked processing architecture such as a client/server architecture, a peer-to-peer architecture, the Internet, or the Cloud.
The explicit use of the terms “processor” or “controller” should not be interpreted as exclusively referring to hardware capable of running software, and can implicitly include, but is not limited to, digital signal processor “DSP” hardware, read only memory “ROM” for storing software, random access memory “RAM”, a non-volatile storage device, and the like. Furthermore, examples of the present disclosure can take the form of a computer program product accessible from a computer-usable storage medium, or a computer-readable storage medium, the computer program product providing program code for use by or in connection with a computer or any instruction execution system. For the purposes of this description, a computer-usable storage medium or a computer readable storage medium can be any apparatus that can comprise, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. The medium can be an electronic, magnetic, optical, electromagnetic, infrared, or a semiconductor system or device or propagation medium. Examples of computer-readable media include semiconductor or solid state memories, magnetic tape, removable computer disks, random access memory “RAM”, read-only memory “ROM”, rigid magnetic disks and optical disks. Current examples of optical disks include compact disk-read only memory “CD-ROM”, compact disk-read/write “CD-R/W”, Blu-Ray™ and DVD.
As mentioned above, multiple factors may influence a physician's choice of a vascular access site to perform an interventional procedure at a target site. Consequently, a physician may make an intuitive decision as to which access site to use. Alternatively, the physician may search through electronic health record data for the patients in order to make a more informed choice. However, this process is inefficient because extracting all the relevant information from various sources and mentally resolving the information into an optimal choice of vascular access site is complex and time-consuming.
The result of these operations is to identify an optimal vascular access site for the target site. In particular, for an “optimal” vascular access site, not only is the target site reachable, but the risk of complications is as low as possible, as well. The optimal vascular access site may then be recommended to a physician. The optimal vascular access site is also provided in a consistent and efficient manner, obviating some of the existing challenges in determining an optimal access site.
The method described above may also be implemented by the system 200 illustrated in
With reference to
In the operation S120 image data 140 representing at least a portion of the vasculature, is received. In this respect, the image data 140 may include one or more types of image data, such as for example: intravascular ultrasound “IVUS” image data, optical coherence tomography “OCT” image data, computed tomography angiography “CTA” image data, magnetic resonance “MR” image data, MR angiography “MRA” image data, digital subtraction angiography “DSA” image data, and positron emission tomography “PET” image data. The image data 140 may be received by the one or more processors 210 illustrated in
In the operation S130, a success metric 150 is computed for a plurality of potential vascular access sites 1101 . . . n, based on the image data 140. The potential vascular access sites may include sites such as a femoral access site, a radial access site, a brachial arterial access site, an axillary arterial access site, and a jugular access site, for example. The success metric represents an ease of navigating the interventional device 120 from the vascular access site 110 to the target site 130 via the vasculature.
By way of an example the interventional device 120 may be a balloon catheter. The success metric may be computed based on the value of one or more success factors that affect the ease of navigating the interventional device from the vascular access site 110 to the target site 130 via the vasculature. In this regard, the success factors may represent features such as:
A success factor representing a tortuosity of a portion of the vasculature between the vascular access site 110 and the target site 130, can be determined from received image data 140, such as CTA, MRA, and DSA image data, for example. Tortuous vessel segments may be detected in such image data using detection techniques such as computing overall curvature, computing curvature sign changes, or other techniques including machine learning techniques. Tortuous vessels may also be identified using a neural network. The neural network may be trained in an unsupervised manner. For example, a generative neural network such as a variational autoencoder “VAE” may be trained on a dataset including non-tortuous vessels. A VAE learns the distribution over the latent representation of the training data. Therefore, by training the VAE on non-tortuous vessels, the distribution learned by the VAE can provide a good representation of vasculature without tortuosity. At inference, when vasculature without tortuosity is inputted into the VAE, its latent representation will be an inlier in the learned distribution. However, if vasculature with higher tortuosity than represented in the training data, is inputted into the VAE, its latent representation will be an outlier in the learned distribution. Tortuous segments in the vasculature between the potential access sites and the target site can therefore be identified by detecting the outlier predictions of the VAE.
A tortuosity of a portion of the vasculature may be determined based on a count of the total number of tortuous segments and/or a magnitude of the tortuosity, between each potential vascular access site 1101 . . . n, and the target site 130. The tortuosity may then be used to determine a success factor affecting the ease of navigating the interventional device from the vascular access site 110 to the target site 130 via the vasculature. In this example, the success factor may increase with decreasing tortuosity, and so the success factor may be inversely related to the tortuosity of the portion of the vasculature.
A success factor representing a difficulty of passing a stenosis in the vasculature between the vascular access site 110 and the target site 130, may also be calculated based on image data. Navigating an interventional device past a stenosis can be complex, and may increase procedure time in view of the need to avoid dislodging stenosis material. The positions of stenoses may be identified in image data such as IVUS, OCT, CTA and MRA by detecting narrowed regions of a lumen in the image data.
A difficulty of passing a stenosis in the vasculature may be determined based on a count of the total number of stenoses, and/or a total length of the stenosis/stenoses, and/or a minimum diameter of the stenosis/stenoses between the vascular access site 110 and the target site 130. This difficulty of passing the stenosis may be used to determine a success factor affecting the ease of navigating the interventional device from the vascular access site 110 to the target site 130 via the vasculature. In this example, the success factor may decrease as a count of the total number of stenoses increases, and so the success factor may be inversely related to the difficulty of passing a stenosis in the vasculature.
A success factor representing a difficulty of passing an implanted device in the vasculature between the vascular access site 110 and the target site 130, may also be calculated based on image data. Navigating an interventional device past an implanted device can also be complex and can therefore increase procedure time. The presence of an implanted device may be detected in image data such as IVUS, OCT, CTA and MRA image data by detecting characteristic patterns of the implanted device in the image data, such as stent struts. A success factor representing a difficulty of passing an implanted device, can be calculated in a similar manner as for a stenosis, and likewise used to calculate the success factor.
A success factor representing a difficulty of passing a calcification in the vasculature between the vascular access site 110 and the target site 130, may also be calculated based on image data. Vascular calcifications constrict vessels, hampering the navigation of interventional devices and increasing the risk of complications for a patient. The navigation of interventional devices past calcifications can cause the calcification to break off, resulting in downstream embolus and vascular ischemia. The positions of calcifications may be identified in image data such as IVUS, OCT, CTA, PET, and MRA image data. Calcifications can be identified in such image data using segmentation techniques, such as a U-Net, region growing, and thresholding, or using feature detection techniques. Neural networks may be trained to detect calcifications in a supervised manner using image data from clinical investigations on a variety of patients and in which the ground truth is annotated.
In the operation S140 the vascular access site 110 is identified based on the computed success metrics. This may involve comparing the computed success metrics 150 for each of the potential vascular access sites 1101 . . . n and identifying the vascular access site 110 having the highest success metric. The result of these operations is therefore to identify an optimal vascular access site for the target site. The optimal vascular access site may then be recommended to a physician. The optimal vascular access site is provided in a consistent and efficient manner, thereby obviating some of the existing challenges in determining an optimal access site.
The optimal vascular access site 110 may be identified in the operation S140 in various ways. For example, the success metric for the optimal vascular access site 110, and/or the success metrics 150 for each of the potential vascular access sites 1101 . . . n, may be outputted to a display. The success metrics may be outputted to the display 230 illustrated in
The operation of identifying S140 the vascular access site 110 may include providing a ranking of the potential vascular access sites 1101 . . . n based on the computed success metrics 150. Providing such a ranking permits a physician to choose between a highest-ranked access site and a lower-ranked access site. The ranking may be provided in the form of a table, or a numerical order indicating the hierarchy in the ranking. The ranking may alternatively be indicated graphically by displaying the potential vascular access sites 1101 . . . n, on an anatomical image, and indicating their ranking on the anatomical image. In a related example, the operation of providing a ranking of the potential vascular access sites 1101 . . . n, includes:
Such a dominant feature may for example be a tortuous portion of the vasculature. The dominant feature may for example be identified by means of an attention map, or a bounding box. An example of such a bounding box is illustrated in
In a related example, the operation of identifying S140 the vascular access site 110 includes:
This example is illustrated with reference to
As mentioned above, the success metric for the potential vascular access sites 1101 . . . n may be based on one or more success factors. If the success metric is based on a single success factor, such as for example the tortuosity of the portion of the vasculature, the success metric may simply be equated with the success factor. The success metrics may then simply be compared for the vascular access sites, and the vascular access site 110 having the highest success metric may be outputted as the optimal, or recommended, vascular access site 110. However, if the success metrics for the potential vascular access sites 1101 . . . n are based on multiple different success factors 1501 . . . k, the individual success factors may be weighted to provide the success metric 150 for the potential vascular access site. The success metrics 150 may then be compared for the vascular access sites, and the vascular access site 110 having the highest success metric may be outputted as the optimal, or recommended, vascular access site 110.
In one example, the success factors may be weighted by setting the values of the weights of the individual success factors to fixed values. The values of the weights may be set such that the presence of calcifications and implanted devices result in a relatively lower success metric, whilst the absence of high tortuosity results in a relatively higher success metric, for example. In another example, the values of the weights may have variable values. For example, the values of the weights may be set by designing rules based on the values of each of the individual success factors 1501.k such that the resulting success metric 150 matches an expert assessment of the ease of navigating the interventional device from the vascular access site to the target site via the vasculature. A neural network may be trained to generate the values of the weights based on the values of each of the individual success factors 1501.k, or based on the image data from which the individual success factors 1501.k are generated, such that the resulting success metric 150 matches an expert assessment of the ease of navigating the interventional device from the vascular access site to the target site via the vasculature. A supervised learning approach may be used to train the neural network to generate the resulting success metric 150 based on how experts consider the various individual success factors 1501.k to influence the resulting success metric 150.
her success metric 150 that is computed in the operation S130 also represents an outcome of performing a vascular intervention at the target site 130 using the vascular access site 110. Accordingly, success factors that affect the outcome, such as for example the duration of a vascular intervention, the risk of complications such as bleeding and ischemic events, the likelihood of having to repeat the procedure, and other factors resulting from navigating the vasculature between the vascular access site 110 and the target site 130, may also be taken into account in computing the success metric 150. Values for each of these success factors may be obtained from clinical studies, such as reported in the document mentioned above by Jolly, S. et al., entitled “Radial versus femoral access for coronary angiography or intervention and the impact on major bleeding and ischemic events: A systematic review and meta-analysis of randomized trials,” American Heart Journal 157(1): 132-140 (2009). These success factors may be included in the success metric 150 by weighting one or more of these factors with weights, the values of which are determined as described for the success factors 1501 . . . k above. By incorporating the outcome of performing a vascular intervention at the target site 130 using the vascular access site 110, into the success metric in this manner, a more evidence-based recommendation of the optimal vascular access site may be provided.
In some examples, a neural network is used to a determine the success metric 150 and/or to identify the optimal vascular access site 110. In one example, the operation of computing S130 a success metric 150 and/or the identifying S140 the vascular access site 110, is determined by inputting the target site 130, and the image data 140 representing the at least a portion of the vasculature, into at least one neural network 170. The at least one neural network 170 is trained to predict the success metric 150 and/or to identify the vascular access site 110, based on the inputted target site 130 and the image data 140.
The predicted success factors 150 may then be ranked by the ranking controller illustrated in
In this example, a calcification was used to illustrate the type of features that the neural network may be trained to recognize. However, the single neural network may be trained to recognize multiple features that represent a difficulty of navigating through the vasculature between the vascular access site 110 and the target site 130, including calcifications, stenoses, tortuosity, and so forth. More generally, the advantage of using a single neural network to predict the success factors is that it implicitly learns the relevant features associated with the success metric 150. The neural network may therefore learn features associated with the success factor that were not anticipated at the training stage, and therefore not accounted for by training the neural network based on a single type of data, such as calcification data.
In the example illustrated in
Thus, in the example illustrated in
The neural networks described above with reference to
An example of a technique for generating confidence values associated with a neural network's predictions is disclosed in a document by Ramalho, T. et al., entitled “Density estimation in representation space to predict model uncertainty”, https://arxiv.org/pdf/1908.07235.pdf. The neural network may be trained in accordance with this technique to generate confidence values such that when the neural network is presented with an image that is very different from its training dataset, the neural network it is able to recognize this and the neural network outputs a low confidence value. The technique described in this document generates confidence values by estimating the training data density in representation space, and determining whether the trained network is expected to make a correct prediction for the input by measuring the distance in representation space between the input and its closest neighbors in the training set. Alternative techniques may also be used to generate confidence values associated with the predictions of the neural network. The dropout technique may be used, for example. The dropout technique involves iteratively inputting the same data into a neural network and determining the neural network's output whilst randomly excluding a proportion of the neurons from the neural network in each iteration. The outputs of the neural network are then analyzed to provide mean and variance values. The mean value represents the final output, and the magnitude of the variance indicates whether the neural network is consistent in its predictions, in which case the variance is small, or whether the neural network was inconsistent in its predictions, in which case the variance is larger.
The neural network(s) described above with reference to
EHR data may include information from historic procedures that can better inform the choice of access site, in particular information that has a negative impact regarding the risk of complications. For example, reports on the outcome of historic procedures may indicate previous instances of bleeding, or details about implanted devices.
Further, EHR data may include information relating to body mass index, smoking history, age, gender, and so forth. EHR data may also include information from clinical investigations on a patient, such as vessel characteristics, tortuosity, thickness, artifacts within vessels such as calcifications, the existence of implanted devices such as stents, and so forth. EHR data may include information that can help inform physicians about potential difficulties they might encounter during endovascular navigation. For example, smoking has been shown to constrict vessels and to be highly correlated with atherosclerotic changes in vessels, i.e. a build-up of plaque, as described in a document by Ambrose, J., et al., entitled “The pathophysiology of cigarette smoking and cardiovascular disease: an update”, J. Am. Coll. Cardiol. 43(10): 1731-1737 (2004). Doi: 10.1016/j.jacc.2003.12.047. Therefore, features such as vessel width and vessel tortuosity may present different levels of severity based on patients' smoking history.
Accordingly, in certain examples, relevant EHR data such as the patients' smoking history may cause a change in the computed values for one or more success factors affecting procedure outcome and/or their relative weighting. Thus, by inputting such EHR data into a neural network trained to compute the success metric, patient-specific past data can be used in computing the success metric thereby further improving the reliability of the identified vascular access sites being recommended to the user.
The EHR data may be stored on a database, such as a picture archiving and communication system “PACS” system. In this example, the EHR data may therefore be received from a database. The EHR data may be received by a processor, such as the one or more processors 210 illustrated in
In one example, the image data that is inputted into the neural network(s) 170 includes ultrasound image data that is generated prior to inserting the interventional device (120) into the identified vascular access site. Ultrasound image data is increasingly used immediately prior to the start of interventional procedures in order to visualize the vascular structure around an access site. As an example, for femoral access, an ultrasound probe may be used to find the common femoral artery. In this example, the additional ultrasound imaging information is also processed by the neural network(s) in order to re-evaluate the optimal vascular access site that is identified. For example, if, in assessing the difficulty of passing a calcification in the vasculature between the vascular access site 110 and the target site 130, i.e. “CI”, calcification or plaque is identified near this access site, this can lower the success metric 150 at the current potential access site 1101 . . . n, which may result in a different vascular access site being optimal. Alternatively, by observing changes in the success metric 150 as the ultrasound probe is moved around near the optimal access site and choosing the exact location with the highest success metric 150, a precise location for incision at the access site can be identified.
In a similar manner, in one example, image data that is generated during an interventional procedure, such as for example ultrasound, or X-ray image data, may be inputted into the one or more neural networks 170, and used to periodically update the predicted success factors for the potential vascular access sites. In so doing, additional information from this image data may be used to confirm that an interventional procedure is being carried out using the optimal vascular access site on the basis of the most up-to date information, or to otherwise recommend an alternative vascular access site. In this example, the suitability of an initially selected access point may be constantly reassessed whilst a user navigates the interventional device through the vasculature.
In one example, the neural network(s) are also trained to predict a simulated path for the interventional device. The simulated path is then used to calculate a complexity metric for navigating the interventional device. The complexity metric is then used to calculate the success metric. In this example, the at least one neural network 170 is further trained to predict a simulated path for the interventional device to reach the target site 130 in the vasculature from each of the potential vascular access sites 1101 . . . n; and the at least one neural network 170 is trained to compute the success metric 150 and/or to identify the vascular access site 110, based further on a complexity metric representing a difficulty of reaching the target site 130 in the vasculature from each of the potential vascular access sites 1101 . . . n with the interventional device.
Using the complexity metric in the calculation of the success metric takes account of the capabilities of navigating a particular interventional device along the paths between the potential vascular access sites, and the target site. The complexity metric may be based on factors that positively affect the success metric, such as taking a step toward the target site, and factors that negatively affect the success metric, such as a count of the number of times the interventional device touches the lumen walls, the total number of steps taken to reach the target site, and so forth. If the interventional device frequently touches the lumen walls, this indicates narrow or tortuous lumen. If a high number of steps are used to reach the target site, this might also indicate tortuous lumens and/or a larger path length between the vascular access site and target site.
The training of the neural network(s) 170 is now detailed below.
In general, the training of a neural network involves inputting a training dataset into the neural network, and iteratively adjusting the neural network's parameters until the trained neural network provides an accurate output. Training is often performed using a Graphics Processing Unit “GPU” or a dedicated neural processor such as a Neural Processing Unit “NPU” or a Tensor Processing Unit “TPU”. Training often employs a centralized approach wherein cloud-based or mainframe-based neural processors are used to train a neural network. Following its training with the training dataset, the trained neural network may be deployed to a device for analyzing new input data during inference. The processing requirements during inference are significantly less than those required during training, allowing the neural network to be deployed to a variety of systems such as laptop computers, tablets, mobile phones and so forth. Inference may for example be performed by a Central Processing Unit “CPU”, a GPU, an NPU, a TPU, on a server, or in the cloud.
The process of training the neural network(s) 170 described above therefore includes adjusting its parameters. The parameters, or more particularly the weights and biases, control the operation of activation functions in the neural network. In supervised learning, the training process automatically adjusts the weights and the biases, such that when presented with the input data, the neural network accurately provides the corresponding expected output data. In order to do this, the value of the loss functions, or errors, are computed based on a difference between predicted output data and the expected output data. The value of the loss function may be computed using functions such as the negative log-likelihood loss, the mean squared error, or the Huber loss, or the cross entropy loss. During training, the value of the loss function is typically minimized, and training is terminated when the value of the loss function satisfies a stopping criterion. Sometimes, training is terminated when the value of the loss function satisfies one or more of multiple criteria.
Various methods are known for solving the loss minimization problem such as gradient descent, Quasi-Newton methods, and so forth. Various algorithms have been developed to implement these methods and their variants including but not limited to Stochastic Gradient Descent “SGD”, batch gradient descent, mini-batch gradient descent, Gauss-Newton, Levenberg Marquardt, Momentum, Adam, Nadam, Adagrad, Adadelta, RMSProp, and Adamax “optimizers” These algorithms compute the derivative of the loss function with respect to the model parameters using the chain rule. This process is called backpropagation since derivatives are computed starting at the last layer or output layer, moving toward the first layer or input layer. These derivatives inform the algorithm how the model parameters must be adjusted in order to minimize the error function. That is, adjustments to model parameters are made starting from the output layer and working backwards in the network until the input layer is reached. In a first training iteration, the initial weights and biases are often randomized. The neural network then predicts the output data, which is likewise, random. Backpropagation is then used to adjust the weights and the biases. The training process is performed iteratively by making adjustments to the weights and biases in each iteration. Training is terminated when the error, or difference between the predicted output data and the expected output data, is within an acceptable range for the training data, or for some validation data. Subsequently the neural network may be deployed, and the trained neural network makes predictions on new input data using the trained values of its parameters. If the training process was successful, the trained neural network accurately predicts the expected output data from the new input data.
In one example, the at least one neural network 170 is trained to predict the success metric 150 and/or to identify the vascular access site 110, by:
This training method is illustrated in
Similarly, when multiple success factors are taken into account, a single neural network may be trained to predict a physician-assessed overall success metric that takes into account the multiple factors. The single neural network may also be trained in an unsupervised manner using a generative neural network like a variational autoencoder “VAE” with images of portions of vasculature containing no, or negligible, features associated with a difficulty of navigating an interventional device from the vascular access site 110 to the target site 130′, as assessed by an expert physician or physicians. Therefore, the VAE learns the distribution over the latent representation of the training data containing vasculature without tortuosity or calcifications or stenoses and so on. When vasculature without tortuosity or calcifications or stenoses and so forth are inputted into the trained VAE, the latent representations of these inputs will be inliers in the learned distribution. However, if vasculature with tortuosity or calcifications or stenoses and so on are inputted into the trained VAE, their latent representations will be outliers in the learned distribution. Further, since VAEs are trained to reconstruct the inputted data, the trained VAE will include reconstruction errors in regions of images containing features that were not present in the training data, allowing the identification of such regions of high reconstruction error. The more tortuous the vasculature (or, similarly, the larger the calcification or stenosis), the larger the reconstruction error. Quantification of these errors and/or number of regions including errors will allow a single neural network to predict overall success metrics that take multiple factors into account. Additionally, this allows the network to learn features associated with the success factor that were not anticipated.
When multiple neural networks are trained to predict the success metric, each individual neural network may be trained individually with the physician-assessed ground truth success factors for a particular feature, such as:
In some examples, the ground truth data may include annotations of image features in the form of binary masks or bounding boxes representing a difficulty of navigating an interventional device from the vascular access site 110 to the target site 130′ such as calcifications or stenoses and so on. In this example, the neural network 170 or networks 1701 . . . k may be trained to predict these annotations; and the adjusting parameters of the neural network 170 or networks 1701 . . . k, is repeated until a value of a loss function representing a difference between an annotation predicted by the neural network 170 or networks 1701 . . . k, and the ground truth annotation, meets a stopping criterion. In some examples, the ground truth data may include a ranking of multiple potential vascular access sites. For instance, identified factors that represent a difficulty of navigating an interventional device from the vascular access site 110 to the target site 130′ such as tortuosity or calcifications and so on may be presented to experts who can rank various access sites based on the influence of the identified features on navigation and/or procedure success. This ranking may also be used to train the neural network to predict the ranking of the vascular access sites. Thus, in one example, the ground truth data comprises a ranking of vascular access sites 110 for each of the target sites 130. In this example, the neural network 170 is trained to predict a ranking of the vascular access sites for each inputted target site 130′; and the adjusting parameters of the neural network 170, is repeated until a value of a loss function representing a difference between a ranking of the vascular access sites predicted by the neural network, and the ground truth ranking of the access sites, meets a stopping criterion. In some examples, the ground truth includes expert labels determining which images contain features representing a difficulty of navigating an interventional device from the vascular access site 110 to the target site 130′ such as tortuosity or calcifications and so on and which images do not contain such features. In this example, the neural network 170 or networks 1701 . . . k may be trained in an unsupervised manner with images that do not contain features representing a difficulty of navigating an interventional device from the vascular access site 110 to the target site 130′. For instance, a VAE may be trained to reconstruct the inputted training data by adjusting parameters of the neural network 170 or networks 1701 . . . k until a value of a loss function representing a difference between the reconstructed image and the inputted image meets a stopping criterion, and a value of a loss function representing a difference between the predicted components of a distribution over the latent representation of the inputted training data and the components of a standard distribution meets a stopping criterion. When images containing features representing a difficulty of navigating an interventional device from the vascular access site 110 to the target site 130′ are inputted into such a trained neural network, they are identified as outliers in the learned distribution over the latent representation of the training data.
In one example, a computer program product is provided. The computer program product includes instructions which when executed by one or more processors, cause the one or more processors to carry out a method according to any of the attached claims.
As mentioned above, features of the computer implemented method, may also be implemented in a system. Thus, in one example, a system 200 for identifying a vascular access site 110 for inserting an interventional device 120 in order to reach a target site 130 in a vasculature, is provided. The system includes one or more processors 210 configured to carry out the steps of a method according to any of the attached claims
An example of this system 200 is illustrated in
The above examples are to be understood as illustrative of the present disclosure, and not restrictive. Further examples are also contemplated. For instance, the examples described in relation to computer-implemented methods, may also be provided by the computer program product, or by the computer-readable storage medium, or by the system 200, in a corresponding manner. It is to be understood that a feature described in relation to any one example may be used alone, or in combination with other described features, and may be used in combination with one or more features of another of the examples, or a combination of other examples. Furthermore, equivalents and modifications not described above may also be employed without departing from the scope of the invention, which is defined in the accompanying claims. In the claims, the word “comprising” does not exclude other elements or operations, and the indefinite article “a” or “an” does not exclude a plurality. The mere fact that certain features are recited in mutually different dependent claims does not indicate that a combination of these features cannot be used to advantage. Any reference signs in the claims should not be construed as limiting their scope.
| Number | Date | Country | Kind |
|---|---|---|---|
| 21218158.0 | Dec 2021 | EP | regional |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/EP2022/079847 | 10/26/2022 | WO |
| Number | Date | Country | |
|---|---|---|---|
| 63272251 | Oct 2021 | US |
