WO2023104559A1 - Mesure de traitement de thrombus - Google Patents

Mesure de traitement de thrombus Download PDF

Info

Publication number
WO2023104559A1
WO2023104559A1 PCT/EP2022/083396 EP2022083396W WO2023104559A1 WO 2023104559 A1 WO2023104559 A1 WO 2023104559A1 EP 2022083396 W EP2022083396 W EP 2022083396W WO 2023104559 A1 WO2023104559 A1 WO 2023104559A1
Authority
WO
WIPO (PCT)
Prior art keywords
angiographic
neural network
procedure
inputted
training
Prior art date
Application number
PCT/EP2022/083396
Other languages
English (en)
Inventor
Ayushi Sinha
Javad Fotouhi
Vipul Shrihari Pai Raikar
Leili SALEHI
Ramon Quido ERKAMP
Sean Joseph KYNE
Original Assignee
Koninklijke Philips N.V.
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority claimed from EP22156482.6A external-priority patent/EP4195215A1/fr
Application filed by Koninklijke Philips N.V. filed Critical Koninklijke Philips N.V.
Publication of WO2023104559A1 publication Critical patent/WO2023104559A1/fr

Links

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H30/00—ICT specially adapted for the handling or processing of medical images
    • G16H30/40—ICT specially adapted for the handling or processing of medical images for processing medical images, e.g. editing
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems

Definitions

  • the present disclosure relates to predicting a success metric that will be achieved by performing a treatment procedure on a thrombus.
  • a computer-implemented method, a computer program product, and a system, are disclosed.
  • a thrombus is a blockage in a blood vessel.
  • a thrombus may occur in a vein or in an artery.
  • blood becomes congested, leading to swelling and fluid congestion.
  • an arterial thrombus the supply of blood is restricted, leading to a condition known as ischemia, which risks damage to tissue supplied by the artery.
  • a portion of the thrombus can also break-away as an embolus. The embolus can become lodged elsewhere in the body and form an embolism that likewise blocks a blood vessel.
  • Thromboses may occur in various parts of the body, including in the heart and the brain, where their effects can be severe unless treated quickly. In the brain, for example, a thrombus, or an embolism, can lead to conditions such as (ischemic) stroke.
  • thromboses Various treatments are available for treating thromboses. These include pharmacological treatments in which thrombolytic drugs such as Alteplase are administered in order to break-up a thrombus by means of thrombolysis.
  • thrombolytic drugs such as Alteplase are administered in order to break-up a thrombus by means of thrombolysis.
  • treatment procedures are also available for treating thromboses. Such treatment procedures include the use of treatment devices such as mechanical thrombectomy devices, and which are used in so-called mechanical thrombectomy procedures.
  • Aspiration catheters typically include a delivery catheter that is used to deliver an irrigation fluid to the thrombus, and an extraction catheter that is used to extract the irrigation fluid, together with broken-up pieces of the thrombus.
  • an aspiration catheter is positioned close to the clot, and at which position the aspiration takes place, resulting in the broken pieces of the clot being extracted from the body.
  • Stent retrievers typically include an expandable wire mesh tube that is designed to remove the clot in one piece.
  • a stent retriever is positioned close to the thrombus using a delivery catheter.
  • the wire mesh tube is extended out of the delivery catheter, where it expands and captures the clot.
  • the stent retriever is then withdrawn into the delivery catheter and the stent retriever, together with the clot, is removed from the body.
  • mechanical thrombectomy devices have been shown to have a higher clinical efficacy in achieving re-perfusion of blood vessels than thrombolytic drugs.
  • the success of mechanical thrombectomy procedures in reducing long-term functional dependency and mortality has been found to be highly correlated with the “technical success” of the procedure.
  • Technical success is assessed based on several criteria, including the speed of the procedure and completeness of reperfusion.
  • the tortuosity of a vessel in the vicinity of the thrombus can indicate the likely success of a mechanical thrombectomy procedure.
  • unfavorable vascular anatomy such as tortuosity
  • tortuosity is combined with sub-optimal device selection and/or placement, this can result in a poor long-term outcome for a patient.
  • the location of a thrombus within tortuous intracerebral arteries can for example affect the success of both stent retriever and aspiration catheter based treatments, as described in the aforementioned document by Alveme, F., et al.
  • the location of the thrombus allows the force applied to a stent retriever during its withdrawal into the delivery catheter to be in the same direction, or in a similar direction, throughout the withdrawal, then the chance of successfully retrieving the thrombus is high.
  • the direction of the force applied to the stent retriever changes significantly during its withdrawal into the delivery catheter, for example due to the tortuosity of the vasculature, then the probability of successful retrieval are low. Therefore, changes in the angle of the force applied to the stent retriever during the withdrawal of the thrombus into the delivery catheter affect the success of stent retriever-based treatments.
  • an angle of interaction between the aspiration catheter and the thrombus of > 125.5 degrees has been associated with high chance of success, as described in the aforementioned document by Bemava, G., et al.
  • an angle of interaction of 180 degrees occurs when a distal end of the aspiration catheter and the thrombus he in a straight line.
  • a computer-implemented method of predicting a success metric achieved by performing a treatment procedure on a thrombus includes: receiving angiographic image data, including one or more angiographic images comprising the thrombus; inputting the angiographic image data into a neural network; and calculating the success metric based on the output of the neural network; and wherein the neural network is trained using training data comprising angiographic training images representing the treatment procedure, and corresponding ground truth procedure outcome data.
  • the result of these operations is to provide a reliable success metric indicating the success that may be achieved by the treatment procedure. Since the success metric is based on the angiographic images of a patient, it may account for factors such as the tortuosity of the vasculature in the vicinity of the thrombus.
  • Fig. 1 is a flowchart illustrating an example of a method of predicting a success metric achieved by performing a treatment procedure on a thrombus, in accordance with some aspects of the present disclosure.
  • Fig. 2 is a schematic diagram illustrating an example of a system 200 for predicting a success metric achieved by performing a treatment procedure on a thrombus, in accordance with some aspects of the present disclosure.
  • Fig. 3 is a schematic diagram illustrating A) a first example of the training of a neural network 140 to generate latent space representations z t of inputted angiographic images 130’, and B) the performance of inference with the trained first example of a neural network 140, in accordance with some aspects of the present disclosure.
  • Fig. 4 is a schematic diagram illustrating A) a second example of the training of a neural network 140 to generate latent space representations z t of inputted angiographic images 130’, and B) the performance of inference with the trained second example of a neural network 140, in accordance with some aspects of the present disclosure.
  • Fig. 5 is a schematic diagram illustrating A) a third example of the training of a neural network 140 to classify inputted angiographic images 130’ with an expected outcome 15Oo,i of the procedure, and B) the performance of inference with the trained third example of a neural network 140, in accordance with some aspects of the present disclosure.
  • Fig. 6 is a schematic diagram illustrating A) a fourth example of the training of a neural network 140 to generate latent space representations Zi for inputted angiographic images 130’ and to predict from the generated latent space representations Zi future angiographic images 130f U ture to the inputted angiographic images 130, and B) the performance of inference with the trained fourth example of a neural network 140, in accordance with some aspects of the present disclosure.
  • the success metric may be predicted prior to, or during a treatment procedure.
  • the success metric may be predicted for a planned treatment procedure, or alternatively it may be predicted for a current treatment procedure.
  • the thrombus may be located in a vein, or in an artery.
  • the thrombus may be a venous thrombus, or alternatively it may be an arterial thrombus.
  • the thrombus may be located in various parts of the body, including in the brain, the heart, the lungs, or in a limb such as the leg, for example.
  • the thrombus may be located in the brain and, the purpose of the treatment procedure may be for the treatment of (ischemic) stroke.
  • the treatment procedure is a mechanical thrombectomy procedure.
  • mechanical thrombectomy devices that are used in such procedures: aspiration catheters, and stent retrievers. Examples are described herein wherein the mechanical thrombectomy device is one of these two types of device.
  • the computer-implemented methods described herein may also be used with other types of mechanical thrombectomy devices, for example, coil retrievers.
  • the methods described herein may be used to predict a success metric achieved with other types of treatment procedures and in which other types of devices are used to treat a thrombus.
  • ultrasound thrombolysis devices for example, and wherein ultrasound energy is used to break-up a thrombus
  • rheolytic or rotational embolectomy devices wherein pressurized saline or a rotating device, respectively, are used to macerate or fragment the thrombus which is then aspirated by an aspiration catheter.
  • 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.
  • 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.
  • 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.
  • 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.
  • DSP digital signal processor
  • ROM read only memory
  • RAM random access memory
  • 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.
  • 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-RayTM and DVD.
  • Fig. 1 is a flowchart illustrating an example of a method of predicting a success metric achieved by performing a treatment procedure on a thrombus, in accordance with some aspects of the present disclosure.
  • the method includes: receiving SI 10 angiographic image data, including one or more angiographic images 130 comprising the thrombus 120; inputting S120 the angiographic image data into a neural network 140; and calculating S 130 the success metric 110 based on the output of the neural network 140; and wherein the neural network 140 is trained using training data comprising angiographic training images 130’ representing the treatment procedure, and corresponding ground truth procedure outcome data.
  • the result of these operations is to provide a reliable success metric indicating the success that may be achieved by the treatment procedure. Since the success metric is based on the angiographic images of a patient, it may account for factors such as the tortuosity of the vasculature in the vicinity of the thrombus.
  • Fig. 2 is a schematic diagram illustrating an example of a system 200 for predicting a success metric achieved by performing a treatment procedure on a thrombus, in accordance with some aspects of the present disclosure.
  • operations described in relation to the flowchart illustrated in Fig. 1 may also be performed by the system 200 illustrated in Fig. 2, and vice versa.
  • angiographic image data including one or more angiographic images 130 comprising the thrombus 120
  • the angiographic image data may include one or more 2D angiographic images, i.e. angiographic “projection” images.
  • the angiographic image data may include one or more 3D angiographic images, i.e. angiographic “volumetric” images.
  • the 2D angiographic images may be generated by a projection X-ray imaging system.
  • Projection X-ray imaging systems typically include a support arm such as a so-called “C-arm”, or an “0-arm”, that supports an X-ray source-detector arrangement.
  • Projection X-ray imaging systems may alternatively include a support arm with a different shape to these examples.
  • Projection X-ray imaging systems typically generate projection X-ray images with the support arm held in a static position with respect to an imaging region during the acquisition of image data.
  • the 2D angiographic images 130 may be fluoroscopic, i.e. live images.
  • the 2D angiographic images 130 may be generated using a digital subtraction angiography “DSA” technique, and wherein each image is generated by subtracting from the image the corresponding pixel intensities of a background image.
  • the 2D angiographic images may be generated by the projection X-ray imaging system 220 illustrated in Fig. 2, for example.
  • the 2D angiographic images may be generated by the Philips Azurion 7 X-ray imaging system marketed by Philips Healthcare, Best, The Netherlands.
  • the 3D angiographic images may be generated by a volumetric imaging system, such as for example a volumetric X-ray imaging system.
  • a volumetric X-ray imaging system typically generates image data whilst rotating, or stepping, an X-ray source-detector arrangement around an imaging region, and subsequently reconstructs the image data obtained from multiple rotational angles into a 3D, or volumetric image.
  • volumetric X-ray imaging systems include computed tomography “CT” imaging systems, cone beam CT “CBCT” imaging systems, and spectral CT imaging systems.
  • 3D angiographic images include CT angiography “CTA” and 3D rotational angiography “3DRA” images.
  • the 3D angiographic images 130 may be fluoroscopic, i.e. live images.
  • the 3D angiographic images 130 may be generated using a digital subtraction angiography “DSA” technique.
  • the 3D angiographic images may be generated by a magnetic resonance imaging “MRI” system.
  • magnetic resonance angiography “MRA” volumetric image data may be generated by injecting a contrast agent into the vasculature and using oscillating magnetic fields at specific resonance frequencies to generate images of various anatomical structures using an MRI imaging system.
  • a single image is received in the operation SI 10
  • multiple images are received in the operation SI 10.
  • the multiple images may form a temporal sequence of images, wherein the images are generated at regular time intervals.
  • the multiple images may be generated intermittently, i.e. on an ad-hoc basis.
  • the angiographic image data received in the operation SI 10 may be received from an imaging system, such as one of the imaging systems described above, or from a computer readable storage medium, or from the Internet or the Cloud, for example.
  • the angiographic image data may be received by the one or more processors 210 illustrated in Fig. 2.
  • the angiographic image data may be received via any form of data communication, including wired, optical, and wireless communication.
  • wired or optical communication when wired or optical communication is used, the communication may take place via signals transmitted on an electrical or optical cable, and when wireless communication is used, the communication may for example be via RF or optical signals.
  • the angiographic image data is inputted into a neural network 140.
  • the neural network 140 is trained using training data comprising angiographic training images 130’ representing the treatment procedure, and corresponding ground truth procedure outcome data.
  • the angiographic training images 130’ represent the treatment procedure on a thrombus.
  • the treatment procedure may for example be a mechanical thrombectomy procedure.
  • the mechanical thrombectomy procedure may be performed using a stent retriever device, or a catheter aspiration device, for example.
  • the angiographic training images 130’ may be 2D or 3D angiographic images, as described above for the images that are inputted into the neural network in the operation S120.
  • the ground truth procedure outcome data that corresponds to the angiographic training images 130’ may include one or more outcome factors.
  • the outcome factors may include a classified outcome, such as a success or a failure of the procedure.
  • the classified outcome may include an associated probability of its occurrence.
  • the success of the procedure may include a percentage probability of success.
  • the ground truth procedure outcome data may include outcome factors such as a speed of the procedure, a measure of completeness of re-perfusion being achieved by the procedure, a 90-day mortality, or a 90-day modified Rankin Scale (mRS), and whether the procedure needed to be repeated.
  • Such outcome factors may likewise include an associated probability of their occurrence.
  • a success metric 110 is calculated based on the output of the neural network 140.
  • the calculated success metric 110 may for example be a success or failure of the procedure.
  • the success metric may include an associated probability of its occurrence.
  • Various techniques for calculating the success metric are described in the examples below. In general, these include directly predicting the success metric using the neural network, and analyzing the ground truth procedure outcome data for the angiographic training images 130 in order to provide the success metric 110. As described in the examples below, in the former case, the ground truth procedure outcome data that is used to train the neural network includes one or more outcome factors that represent the success metric 110, and at inference the neural network predicts an expected outcome that similarly includes one or more outcome factors.
  • the one or more outcome factors of the expected outcome that is predicted by the neural network are used to calculate the success metric.
  • the ground truth procedure outcome data that is used to train the neural network includes one or more outcome factors representing the success metric
  • the one or more outcome factors of some of the ground truth procedure outcome data specifically the one or more outcome factors of ground truth procedure outcome data of angiographic training images having latent space representations within a predetermined distance of the latent space representation of the inputted angiographic image, are analysed to provide the success metric 110.
  • the success metric 110 is then outputted, and may thus be used to inform a user on the suitability of a particular treatment procedure for treating the thrombus.
  • the success metric may be outputted to a display, such as the display 240 illustrated in Fig. 2, for example.
  • the neural network 140 may include one or more architectures, such as for example a convolutional neural network “CNN”, an autoencoder network, or one its variants (e.g., variational autoencoder “VAE”, maximum mean discrepancy “MMD” VAE, etc.).
  • the encoder and decoder components of the autoencoder may include a convolutional neural network “CNN” architecture, or a recurrent neural network “RNN” or transformer architecture.
  • 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.
  • CPU Central Processing Unit
  • GPU GPU
  • NPU Neural Processing Unit
  • TPU Tensor Processing Unit
  • the process of training the neural network 140 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.
  • 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.
  • 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 absolute error (or El norm), the mean squared error, the root mean squared error (or L2 norm), the Huber loss, or the (binary) cross entropy loss.
  • 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.
  • a confidence value may also be calculated for each of the angiographic images 130 that are inputted into the trained neural network 140 in the operation S120.
  • the confidence values may then be outputted.
  • the confidence values may be outputted to a display, such as the display illustrated in Fig. 2, for example.
  • the confidence values represent a confidence of the predictions made by the neural network 140, and permit decisions to be made based on the success metric. For example, if the confidence is low, it might be decided not to rely upon the success metric 110. In this regard, a warning may be outputted if the confidence value is below a predetermined threshold value.
  • the neural network 140 may be trained in accordance with this technique to generate confidence values such that when the neural network 140 is presented with an image that is significantly different from its training dataset, the neural network 140 it is able to recognize this and the neural network 140 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 140.
  • the neural network is trained to reconstruct the angiographic images that are inputted into the neural network.
  • a confidence value may be calculated based on a difference, i.e. error, between the reconstructed angiographic image and the inputted angiographic image. If the error is small, the confidence value may be high, whereas if the error is large, the confidence value may be lower.
  • the dropout technique may be used to generate confidence values for the angiographic images 130 that are inputted into the trained neural network 140.
  • 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 and confidence high, or whether the neural network was inconsistent in its predictions, in which case the variance is larger and confidence low.
  • the neural network 140 is trained to generate latent space representations Zi representing the inputted angiographic images 130, and at inference, the success metric 110 is calculated by analyzing the ground truth procedure outcome data 150’GT of angiographic training images 130’ having similar latent space representations z t to the inputted angiographic images 130.
  • Fig. 3 is a schematic diagram illustrating A) a first example of the training of a neural network 140 to generate latent space representations z t of inputted angiographic images 130’, and B) the performance of inference with the trained first example of a neural network 140, in accordance with some aspects of the present disclosure.
  • the neural network 140 may be provided by an autoencoder, or one of its variants (e.g., variational autoencoder “VAE”, maximum mean discrepancy “MMD” VAE, etc.).
  • VAE variational autoencoder
  • MMD maximum mean discrepancy
  • the neural network 140 is trained by forcing the neural network 140 reconstruct the angiographic training images 130’ that are inputted into the neural network.
  • the encoder component of the autoencoder on the left-hand side of the neural network 140 compresses each of the inputted angiographic training images 130’, into a reduced dimension latent representation z t .
  • the decoder component of the autoencoder on the right-hand side of the neural network 140 decodes, or reconstructs, each inputted angiographic training image 130’ from the latent representation z t .
  • the goal is to learn a reduced dimension embedding space, and in which distances between embeddings are indicative of a similarity between the inputted angiographic training images.
  • the encoder and decoder components of the autoencoder may be provided by a convolutional neural network “CNN” architecture when the input into the neural network is a single image, or by a recurrent neural network “RNN” or transformer architecture when the input to the neural network includes multiple images, such as a temporal sequence of images.
  • the neural network 140 is trained to generate latent space representations Zi representing the inputted angiographic images 130, by : receiving angiographic training data, including a plurality of angiographic training images 130’, and wherein each training image comprises a thrombus 120’; inputting the angiographic training data into the neural network 140; and for each of a plurality of the inputted angiographic training images 130’: generating a latent space representation z t of the inputted angiographic training image, using the neural network 140; reconstructing the inputted angiographic training image 130’ from the latent space representation z t , using the neural network 140; and adjusting parameters of the neural network 140 based on a difference between the inputted angiographic training image and the reconstructed inputted angiographic training image; and repeating the generating, the reconstructing, and the adjusting, until a stopping criterion is met.
  • the angiographic training data that is used to train the neural network 140 may include 2D or 3D angiographic training images 130’, as described above for the images that are inputted into the neural network in the operation S120.
  • the adjusting of the parameters of the neural network 140 during training may be performed using backpropagation, as described above.
  • the difference between the inputted angiographic training image and the reconstructed inputted angiographic training image, and which is used to adjust the parameters of the neural network 140, is illustrated in Fig. 3 by the symbol D.
  • the value of this difference may be calculated using a loss function such as LI norm, L2 norm, binary cross entropy, and so forth.
  • the result of the training operation is that the neural network 140 is trained to provide the latent space representations z t for the training images, and from which the original training images can be accurately reconstructed.
  • the latent space representations z t of the training images are used, together with their corresponding ground truth procedure outcome data 150’GT, to calculate the success metric 110 for new inputted angiographic images 130.
  • the neural network 140 is thus trained to generate latent space representations z; representing the inputted angiographic images 130.
  • the method described with reference to Fig. 1 includes: generating, for each inputted angiographic image, a latent space representation z using the neural network 140; and wherein the calculating S 130 the success metric 110 based on the output of the neural network 140, comprises analyzing the ground truth procedure outcome data 150’GT of angiographic training images 130’ having latent space representations z t within a predetermined distance 160 of the latent space representation z; of the inputted angiographic image, and calculating the success metric 110 for the inputted angiographic image based on the analyzed ground truth procedure outcome data.
  • the trained neural network 140 At inference, the trained neural network 140 generates a latent space representation z; for each inputted angiographic image 130 (when using a CNN based implementation) or for a set of inputted angiographic images (when using an RNN based implementation).
  • the trained encoder component of the VAE neural network 140 on the left-hand side of Fig. 3B is used for this purpose. It is noted that the trained decoder component of the VAE neural network 140 on the right-hand side of Fig. 3B is not essential for inference, although it may be used to calculate confidence values, as described below.
  • the ground truth procedure outcome data 150’GT of angiographic training images 130’ having similar latent space representations z t to the latent space representation z; of the inputted angiographic image, are then analyzed in order to calculate the success metric for the inputted image.
  • the latent space representations z t of the angiographic training images 130’ are illustrated by the circular symbols in the central portion of Fig. 3.
  • the corresponding ground truth procedure outcome data 150’GT for each latent space representation z t is illustrated by an un-filled circular symbol for a successful procedure, and by a dark-fdled circular symbol for an unsuccessful procedure.
  • the latent space representation z; of an inputted angiographic image 130 that is inputted during inference, is illustrated by the square symbol amongst the circular symbols of the training angiographic images.
  • the relative positions of the symbols indicate the distances between their latent space representations: symbols that are relatively closer to one another have more-similar latent space representations than symbols that are further apart.
  • An insight exploited in this, and other, examples is that angiographic images that have similar latent space representations, i.e. representations with a relatively small distance between them, tend to have similar treatment procedure outcomes. This is illustrated by the upper half of the latent space in the central portion of Fig. 3 tending to have ground truth procedure outcome data 150’GT that represents successful outcomes, i.e.
  • ground truth procedure outcome data 150’GT that represent unsuccessful outcomes, i.e. dark-fdled circular symbols.
  • the ground truth procedure outcome data 150’GT of angiographic training images 130’ that have latent space representations z t within a predetermined distance 160 of the latent space representation z; of the inputted angiographic image, are identified. This distance may be measured by a function such as the Euclidean distance or the geodesic distance.
  • the success metric 110 for the inputted angiographic image is then calculated using the ground truth procedure outcome data for these identified angiographic training images 130’.
  • the ground truth procedure outcome data 150’GT of the angiographic training images 130’ having latent space representations within the predetermined distance 160 may be analyzed in various ways.
  • the ground truth procedure outcome data 150’GT may represent a binary classification of the success or failure of a treatment procedure.
  • the success metric 110 may be calculated from ground truth procedure outcome data 150’GT with such a binary classification by computing a ratio of the total number of successful outcomes in the ground truth procedure outcome data 150’GT, i.e. un-filled circular symbols, to all outcomes in the ground truth procedure outcome data 150’GT, i.e. un-filled circular symbols and dark-filled circular symbols, that are within the predetermined distance 160 of the latent space representation z; of the inputted angiographic image.
  • This ratio provides a probability of success, and may be used as the success metric 110.
  • This ratio may also be converted to a binary outcome, i.e. a success or a failure of the treatment procedure by applying a threshold, such as 50 percent, to the ratio.
  • This ratio may also be computed separately for each device type that may be used to perform the treatment.
  • the ground truth procedure outcome data may include one or more outcome factors.
  • the outcome factors may include a classified outcome, such as a success or a failure of the procedure.
  • the ground truth procedure outcome data may include one or more outcome factors such as a speed of the procedure, a measure of completeness of re-perfusion being achieved by the procedure, a 90-day mortality, or a 90-day modified Rankin Scale “mRS”, and whether the procedure needed to be repeated.
  • outcome factors may include an associated probability of their occurrence.
  • the success metric may be determined as described above for the binary classification of the success or failure of a treatment procedure or for the calculation of a probability of success of a treatment procedure.
  • the outcome values may be digitized to provide binary outcome values representing positive and negative outcomes by applying a threshold to the range of possible outcome values. A percentage measure of completeness of re-perfusion may be digitized in such a manner, for example.
  • the binary outcome values may then be used to determine a ratio of the total number of positive outcomes, to all outcomes, within the predetermined distance 160 of the latent space representation z; of the inputted angiographic image to provide a probability of success, which may be used as the success metric 110, as described above.
  • the un-digitized outcome values for ground truth procedure outcome data 150’GT having latent space representations z t within the predetermined distance 160 of the latent space representation z; of an inputted image may be combined in another manner.
  • these un-digitized values may be averaged to provide the success metric 110.
  • Outcome factors such as the speed of a procedure may be averaged in this manner, for example.
  • the outcome values of angiographic training images 130’ having latent space representations z t that are within a predetermined distance 160 of the latent space representation z; of the inputted angiographic image may for example be weighted with weighting values for each outcome factor, and summed to provide the success metric 110.
  • the predetermined distance 160 may be set to a predefined value. For example, if the trained neural network 140 is a VAE, then the predetermined distance 160 may be set based on the standard deviation, a. learned by the neural network. For instance, the predetermined distance 160 may be set to 0.25(7 around Zi. In other examples, the predetermined distance 160 may be set based on user input. In one example, the method described with reference to Fig.
  • 1 includes: receiving user input indicative of an extent of the predetermined distance 160, and the method further comprises outputting: a graphical representation of the latent space representation of the inputted angiographic image z;; a graphical representation of the latent space representations of at least some of the angiographic training images 130’ used to train the neural network z t ; and an indication of the predetermined distance 160.
  • the user input indicative of the predetermined distance may be received via a graphical user interface “GUI”, for example.
  • GUI graphical user interface
  • the graphical representations may also be outputted to the GUI.
  • the user may trade-off the accuracy of the neural network’s predictions against the confidence of its predictions. Reducing the predetermined distance 160 has the effect of increasing the accuracy of the predictions because only the outcomes of more-similar angiographic images are considered when determining the success metric 110. However, this also reduces the total number of outcomes that are used to calculate the success metric, and thus may ultimately decrease the confidence in the neural network’s predictions.
  • the graphical representation of the latent space representations of the angiographic training images 130’ may be generated using an algorithm such as t-distributed Stochastic Neighbor Embedding “t- SNE”, for example. This algorithm may be used to project high-dimensional latent representations to two, or three dimensions, or to another number of dimensions, to provide an intuitive visualization of the latent space.
  • the latent space representations of the angiographic training images that were used to train the neural network 140 may also be labelled with their ground truth procedure outcome data.
  • the labels may for example distinguish successful outcomes from unsuccessful outcomes, as illustrated by the un-filled and dark-filled circular symbols in the central portion of Fig. 3A.
  • the labels may also distinguish different types of treatment procedures that were used to train the neural network 140.
  • the labels may for example include colors, or shapes that distinguish the different outcomes, and different treatment procedures.
  • the labels may also be grouped together. For example, labels that indicate a successful outcome may be provided with a perimeter or a predefined level of transparency, in order to distinguish them from labels that indicate an unsuccessful outcome. In so doing, a user may visualize the group to which the angiographic images 130 inputted into the neural network 140 pertain.
  • Confidence values may also be calculated and outputted for each of the angiographic images 130 that are inputted into the neural network 140 during inference.
  • the confidence values may be calculated using the decoder component of the VAE on the right-hand side of the neural network 140.
  • the decoder component was trained to reconstruct the inputted angiographic image 130, as described above.
  • a difference, or error, between the inputted angiographic image 130, and the reconstructed angiographic image 130 that is generated by the decoder component may be calculated using a function such as the structural similarity loss, the LI norm, or the L2 norm, for example.
  • This error provides a measure of how well the network is able to reconstruct the input at inference time, and provides a confidence value for the latent representation
  • This error may be used to calculate the confidence value.
  • Low error values indicate high confidence, and vice versa. If the confidence value is below a threshold value, a user may be notified of this, for example by means of a display.
  • the predictions of the neural network may also be inhibited, or prevented from being outputted, if the confidence value is below the threshold value. In so doing, it is prevented that the user relies upon unreliable predictions.
  • a low confidence value may also indicate that the neural network 140 does not generalize to the current data and may need to be retrained with additional data.
  • confidence values may alternatively be computed using the neural network by using the dropout technique.
  • the method described with reference to Fig. 1 may be used to calculate the success metric 110 for a plurality of different types of treatment procedures.
  • the method may therefore be used prior to a treatment procedure commencing, i.e. in a planning stage, or immediately prior to a treatment procedure commencing, in order to select the most appropriate treatment procedure from multiple potential treatment procedures.
  • the method may also be used during a current treatment procedure, wherein the success metric may be provided for the current treatment procedure, as well as for one or more alternative treatment procedures. In so doing, a user may be re-assured of their chosen current treatment procedure, and also informed of an alternative treatment procedure(s) that may be selected instead in the event that the physician experiences difficulties with the current treatment procedure.
  • the success metric may be calculated for a mechanical thrombectomy procedure that includes a stent retriever device, and also for a catheter aspiration device, for example, as explained above.
  • the values of the success metrics may therefore assist a user in selecting an optimal treatment procedure.
  • different neural networks may be trained, each with angiographic training data for a particular type of treatment procedure, and the outputs of the different neural networks may be compared in order to determine the optimal procedure.
  • the neural network 140 may be trained using angiographic training data from multiple different types of procedures, and the success metric 110 may be calculated for each of the multiple procedures using the ground truth procedure outcome data 150’GT for the procedure.
  • the angiographic images 130 that are inputted into the neural network 130 at inference include the treatment device that is used to perform the treatment procedure.
  • the neural network may observe features in the inputted angiographic images 130, such as the tortuosity of the vasculature surrounding the thrombus, and uses these features to generate its output.
  • the success metric 110 may be calculated without the need to insert a treatment device into the vasculature.
  • the device used to perform the treatment procedure is present in the angiographic images 130 that are inputted into the neural network 140 at inference, then the predictions of the neural network may be more accurate.
  • parameters such as the angle between a distal end of the treatment device and the thrombus, might also be encoded in the latent space representations z t of the neural network, and this might also be used to calculate the success metric 110.
  • the angiographic images 130 also include a deployment catheter 170 for deploying a mechanical thrombectomy device to treat the thrombus 120.
  • the calculated success metric 110 is the success metric achieved by deploying the mechanical thrombectomy device from the deployment catheter.
  • the mechanical thrombectomy device may be a stent retriever, or an aspiration catheter type of treatment device, for example.
  • the angiographic training data that is used to train the neural network also includes angiographic training images 130’ that include a deployment catheter.
  • the neural network 140 may be trained to encode the position of the deployment catheter respective the thrombus in its latent space representations z t , and this may also be used to calculate the success metric 110.
  • the angiographic image data 130 that is inputted into the neural network at inference may include a temporal sequence of real-time angiographic images.
  • the success metric 110 may then be provided in real-time for each angiographic image.
  • the real-time success metric may be used to inform a user of the position at which the mechanical thrombectomy device may be deployed from the deployment catheter to perform the treatment procedure in order to achieve a desired level of success.
  • the real-time success metric may also be used to inform a user of the type of mechanical thrombectomy device (e.g., stent retriever or aspiration catheter) that may be deployed from the deployment catheter to perform the treatment procedure in order to achieve a desired level of success.
  • a mechanical thrombectomy device e.g., stent retriever or aspiration catheter
  • Fig. 4 is a schematic diagram illustrating A) a second example of the training of a neural network 140 to generate latent space representations z t of inputted angiographic images 130’, and B) the performance of inference with the trained second example of a neural network 140, in accordance with some aspects of the present disclosure.
  • the example illustrated in Fig. 4 shares with the example in Fig. 3 the principle that the neural network 140 is trained to generate latent space representations Zi representing the inputted angiographic images 130.
  • the example in Fig. 4 also shares with the example in Fig.
  • the neural network 140 illustrated in Fig. 4 may also have a similar architecture to that in the example in Fig. 3.
  • the neural network 140 is trained to generate latent space representations Zi representing the inputted angiographic images 130, by: receiving angiographic training data, including a plurality of angiographic training images 130’, and wherein each training image comprises a thrombus 120’; receiving ground truth procedure outcome data 150’GT corresponding to the angiographic training data, the ground truth procedure outcome data representing, for each angiographic training image, a success or a failure achieved by performing the treatment procedure on the thrombus; inputting the angiographic training data into the neural network; and for each of a plurality of the inputted angiographic training images: generating a latent space representation z t of the inputted angiographic training image, using the neural network 140; predicting a procedure outcome 15O’o,i achieved by performing the procedure on the thrombus from the latent space representation z t , using the neural network 140; and adjusting parameters of the neural network 140 based on a difference between the ground truth procedure outcome 150’GT and
  • the neural network 140 is trained to generate latent space representations Zi of inputted angiographic images 130, using the ground truth procedure outcome 150’GT of angiographic training images.
  • the ground truth procedure outcome 150’GT may for example be a binary value representing a success or failure of the procedure.
  • the ground truth procedure outcome 150’GT may be a percentage chance of success or failure of the procedure.
  • the adjusting of the parameters of the neural network 140 during training may be performed using backpropagation, as described above.
  • the difference between the ground truth procedure outcome 150’GT and the predicted procedure outcome 15O’o.i, for the inputted angiographic training image, and which is used to adjust the parameters of the neural network 140, is illustrated in Fig. 4 by the symbol D.
  • the value of this difference may be calculated using a loss function such as LI norm, L2 norm, binary cross entropy, and so forth.
  • Inference may be performed with the neural network 140 illustrated in Fig. 4 in the same manner as described for Fig. 3.
  • the neural network 140 is trained to generate latent space representations Zi representing the inputted angiographic images 130, and at inference, the method described with reference to Fig.
  • the calculating S 130 the success metric 110 based on the output of the neural network 140 comprises analyzing the ground truth procedure outcome data 150’GT of angiographic training images 130’ having latent space representations z t within a predetermined distance 160 of the latent space representation z; of the inputted angiographic image, and calculating the success metric 110 for the inputted angiographic image based on the analyzed ground truth procedure outcome data.
  • the neural network 140 is trained to predict the procedure outcome 15O’o,i that will be achieved by performing the procedure on the thrombus, as in the Fig. 4 example.
  • the predicted procedure outcome 15O’o,i is itself used to calculate the success metric 110 for an inputted image, rather than the success metric being calculated by analyzing the ground truth outcome data 150’GT of angiographic training images having similar latent space representations z t to the latent space representation z; representing the inputted image. This example is described with reference to Fig.
  • FIG. 5 is a schematic diagram illustrating A) a third example of the training of a neural network 140 to classify inputted angiographic images 130’ with an expected outcome 15Oo,i of the procedure, and B) the performance of inference with the trained third example of a neural network 140, in accordance with some aspects of the present disclosure.
  • the neural network 140 is trained to classify the inputted angiographic images 130 with an expected outcome 15Oo,i of the procedure, by: receiving angiographic training data, including a plurality of angiographic training images 130’, and wherein each training image comprises a thrombus 120; receiving ground truth procedure outcome data 150’GT corresponding to the angiographic training data, the ground truth procedure outcome data representing, for each angiographic training image a success or a failure achieved by performing the treatment procedure on the thrombus; inputting the angiographic training data into the neural network 140; and for each of a plurality of the angiographic training images 130’: predicting a procedure outcome 15O’o,i achieved by performing the procedure on the thrombus, using the neural network 140; and adjusting parameters of the neural network 140 based on a difference between the predicted procedure outcome 15O’o,i and the ground truth procedure outcome data 150’GT for the inputted angiographic training image; and repeating the predicting, and the adjusting, until
  • the neural network 140 in the example illustrated in Fig. 5 may have a similar architecture to that of the example in Fig. 3.
  • the ground truth procedure outcome 150’GT may for example be a binary value representing a success or failure of the procedure.
  • the ground truth procedure outcome 150’GT may be a percentage chance of success or failure of the procedure.
  • the adjusting of the parameters of the neural network 140 during training may be performed using backpropagation, as described above.
  • the difference between the predicted procedure outcome 15O’o,i and the ground truth procedure outcome data 150’GT for the inputted angiographic training image, and which is used to adjust the parameters of the neural network 140, is illustrated in Fig. 5 by the symbol D.
  • the value of this difference may be calculated using a loss function such as LI norm, L2 norm, binary cross entropy, and so forth.
  • inference may be performed by inputting angiographic images 130 into the trained neural network 140.
  • the ground truth procedure outcomes 150’GT of angiographic training images having similar latent space representations are analyzed to determine the success metric 110 for an inputted image
  • the success metric 110 is predicted by the neural network 140 itself.
  • the neural network 140 is trained to classify the inputted angiographic images 130 with an expected outcome 15Oo,i of the procedure.
  • the method described with reference to Fig. 1 comprises: classifying each inputted angiographic image 130 with an expected outcome 15Oo,i of the procedure, using the neural network 140; and wherein the calculating S 130 the success metric 110 based on the output of the neural network 140, comprises calculating the success metric 110 for the inputted angiographic image based on the classified expected outcome 15Oo,i of the procedure.
  • the success metric may be provided by the classified expected outcome 15Oo,i itself.
  • the classified expected outcome 15Oo,i directly provides the success metric 110.
  • the success metric is based on multiple outcome factors, the classified expected outcome 15Oo,i may be calculated for each of these outcome factors, and the results may be combined, for example by weighting and summing their values, to provide the success metric 110.
  • the neural network 140 is trained using temporal sequences of angiographic training images 130’, and each temporal sequence includes a thrombus 120’ and a deployment catheter 170’.
  • the neural network 140 is trained to perform a future prediction task.
  • the neural network 140 is trained to predict the future position of the deployment catheter 170’ in a future angiographic training image 130’.
  • the neural network is trained to generate latent space representations Zi of inputted angiographic training images, and the success metric 110 for an inputted image is determined by analyzing the ground truth procedure outcomes 150’GT of angiographic training images having similar latent space representations z t to that of the inputted image, as in the Fig. 3 and Fig. 4 examples.
  • FIG. 6 is a schematic diagram illustrating A) a fourth example of the training of a neural network 140 to generate latent space representations Zi for inputted angiographic images 130’ and to predict from the generated latent space representations Zi future angiographic images 130f u ture to the inputted angiographic images 130, and B) the performance of inference with the trained fourth example of a neural network 140, in accordance with some aspects of the present disclosure.
  • the neural network 140 is trained to generate latent space representations Zi representing the inputted angiographic images 130 and to predict, from the generated latent space representations z future angiographic images 130f U ture to the inputted angiographic images 130, the future angiographic images including predicted future positions of the deployment catheter 170, by: receiving angiographic training data, including a plurality of temporal sequences of angiographic training images 130’, and wherein each temporal sequence of training images comprises a thrombus 120’ and a deployment catheter 170’; inputting the angiographic training data into the neural network 140; and for each temporal sequence of angiographic training images 130’: generating, for each angiographic training image in the temporal sequence, a latent space representation z t of the inputted angiographic training image, using the neural network 140; predicting, from the generated latent space representation z t , a future angiographic image 130’f u ture to the inputted angiographic image
  • the neural network 140 in the example illustrated in Fig. 6 may have a similar architecture to that of the example in Fig. 3.
  • the ground truth procedure outcome 150’GT may for example be a binary value representing a success or failure of the procedure.
  • the ground truth procedure outcome 150’GT may be a percentage chance of success or failure of the procedure.
  • the angiographic training data includes a plurality of temporal sequences of angiographic training images 130’.
  • the temporal sequences represent treatment procedures that have been performed on one or more patients.
  • the temporal sequences include a thrombus and a deployment catheter 170’.
  • the deployment catheter is used in the treatment procedure to deploy a treatment device such as a mechanical thrombectomy device to treat the thrombus 120 and thereby achieve the corresponding ground truth procedure outcome 150’GT.
  • a temporal sequence may represent the deployment catheter in multiple positions during the treatment procedure. These temporal sequences of angiographic training images 130’ are used to train the neural network 140.
  • the ground truth procedure outcome data 150’GT that corresponds to the angiographic training images 130’ is analyzed in order to calculate the success metric 110.
  • the ground truth procedure outcome data 150’GT is not required to train the neural network 140.
  • the above operations represent a future prediction task, in which the neural network 140 is trained with the temporal sequences of angiographic training images 130’ to predict a future angiographic image 130’f u ture to each inputted angiographic image.
  • the future angiographic image 130’f u ture to the inputted angiographic image that is predicted in this example may be the next angiographic image in the temporal sequence.
  • the future angiographic image 130’f U ture may be a later image in the temporal sequence.
  • the adjusting of the parameters of the neural network 140 during training may be performed using backpropagation, as described above.
  • the value of this difference may be calculated using a loss function such as LI norm, L2 norm, structural similarity loss, and so forth.
  • Inference may be performed with the neural network 140 illustrated in Fig. 6 in a similar manner as described for Fig. 3, i.e. by inputting angiographic images 130, determining their latent space representations z and analyzing the ground truth procedure outcomes 150’GT of angiographic training images having similar latent space representations z t , in order to calculate the success metric 110 for an inputted image or for a predicted future image.
  • the received angiographic image data includes a plurality of angiographic images 130 comprising the thrombus 120, wherein the angiographic images 130 further include a deployment catheter 170 for deploying a mechanical thrombectomy device to treat the thrombus 120, and wherein the success metric is the success metric achieved by deploying the mechanical thrombectomy device from the deployment catheter; and wherein the neural network 140 is trained to generate latent space representations Zi representing the inputted angiographic images 130 and to predict, from the generated latent space representations z future angiographic images 130f U ture to the inputted angiographic images 130, the future angiographic images 130f u ture including predicted future positions of the deployment catheter 170.
  • the method described with reference to Fig. 1 comprises: generating, for each inputted angiographic image 130, a latent space representation z using the neural network 140; and wherein the calculating the success metric 110 based on the output of the neural network 140, comprises analyzing the ground truth procedure outcome data 150’GT of angiographic training images 130’ having latent space representations z t within a predetermined distance 160 of the latent space representation z; of the inputted angiographic image 130, and calculating the success metric 110 for a predicted future angiographic image 130f u ture to the inputted angiographic image 130, based on the analyzed ground truth procedure outcome data.
  • the calculating the success metric 110 based on the output of the neural network 140 comprises analyzing the ground truth procedure outcome data 150’GT of angiographic training images 130’ having latent space representations z t within a predetermined distance 160 of the latent space representation z; of the inputted angiographic image 130, and calculating the success metric 110 for
  • the success metric 110 is thus calculated for a predicted future angiographic image 130f U ture to the inputted angiographic image 130.
  • the success metric is provided for the expected position of the deployment catheter.
  • the expected position may also be outputted in order to guide a user in navigating the deployment catheter.
  • the predicted angiographic image 130f U ture illustrated in Fig. 6 may be outputted.
  • the image may be outputted to the display 240 illustrated in Fig. 2, for example.
  • inference may be performed with both the trained neural network 140 in the example illustrated in Fig. 3 and the trained neural network 140 in the example illustrated in Fig. 6. That is, the trained neural network 140 in the example illustrated in Fig. 3 is used to calculate the success metric for the inputted angiographic image showing, for instance, the current position of the deployment catheter, while the trained neural network 140 in the example illustrated in Fig. 6 is used to calculate the success metric for a predicted future angiographic image 130f U ture to the inputted angiographic image 130, for instance, showing the expected future position of the deployment catheter. Providing both success metrics may assist a user in deciding if additional time should be spent in the deployment catheter placement or if the treatment device should be deployed.
  • the user may decide to proceed with treatment device deployment rather than improving the placement of the deployment catheter.
  • the angiographic image data that is inputted into the neural network 140 at inference was described as including either 2D angiographic images, or 3D angiographic images.
  • both 2D and 3D angiographic images may be inputted into the neural network.
  • the angiographic image data includes one or more 2D angiographic images, and one or more corresponding 3D angiographic images; and the training data comprises a plurality of 2D angiographic training images, and a plurality of corresponding 3D angiographic training images.
  • treatment device data is used to improve the predictions of the neural network 140.
  • the method described with reference to Fig. 1 includes: receiving device data for a mechanical thrombectomy device to be used in the treatment procedure; and inputting the device data into the neural network 140; and wherein the neural network 140 is trained further using device data corresponding to the angiographic training images 130’;
  • the device data may include information such as the size of stent in a stent retriever device. At inference, this information may be used to select the ground truth procedure outcome data 150GT such that only the ground truth procedure outcome data for treatment procedures that have the same, or a similar stent size as the current procedure, is analyzed when calculating the success metric 110.
  • patient data is used to improve the accuracy of the neural network’s predictions.
  • the method described with reference to Fig. 1 includes: receiving patient data relating to the thrombus 120; and inputting the patient data into the neural network 140; and wherein the network 140 is trained further using patient data corresponding to the angiographic training images 130’.
  • the patient data may for example include electronic health record “EHR” data relating to a historic procedure on the vasculature, clinical findings relating to the vasculature, and so forth.
  • EHR electronic health record
  • an assessment of the amount of plaque in the vasculature of the current patient may be inputted into the neural network and used to calculate the success metric by analyzing only the ground truth procedure outcome data for treatment procedures that have the same, or a similar, amount of plaque.
  • a region of interest such as an extent of a thrombus, or the extent of its surrounding area may be identified in the angiographic images in the training data.
  • the difference between the inputted angiographic training image and the reconstructed inputted angiographic image is calculated by applying a higher weighting factor within the identified region of interest than outside the identified region of interest.
  • the region of interest may be identified using a bounding box, for example.
  • the neural network 140 is also trained to predict an angle between the delivery catheter and the thrombus.
  • the neural network may be trained to predict this angle in a supervised manner.
  • the angiographic training images may be annotated with the angle.
  • the angle may be labelled manually by experts. This information may assist a user in performing the treatment procedure.
  • a computer program product comprises instructions, which when executed by one or more processors, cause the one or more processors to carry out a method of predicting a success metric 110 achieved by performing a treatment procedure on a thrombus 120, the method comprising: receiving SI 10 angiographic image data, including one or more angiographic images 130 comprising the thrombus 120; inputting S120 the angiographic image data into a neural network 140; and calculating S 130 the success metric 110 based on the output of the neural network 140; and wherein the neural network 140 is trained using training data comprising angiographic training images 130’ representing the treatment procedure, and corresponding ground truth procedure outcome data.
  • a system 200 for predicting a success metric 110 achieved by performing a treatment procedure on a thrombus 120 comprises one or more processors 210 configured to: receive SI 10 angiographic image data, including one or more angiographic images 130 comprising the thrombus 120; input S120 the angiographic image data into a neural network 140; and calculate S 130 the success metric 110 based on the output of the neural network 140; and wherein the neural network 140 is trained using training data comprising angiographic training images 130’ representing the treatment procedure, and corresponding ground truth procedure outcome data.
  • the system 200 may also include one or more of: an imaging system for providing the angiographic image data, such as the example projection X-ray imaging system 220 illustrated in Fig. 2; a treatment device for performing a treatment procedure on a thrombus, such as the example mechanical thrombectomy device 230 illustrated in Fig. 2; a monitor 240 for displaying the calculated success metric 110, the angiographic image data, the GUI, other outputs generated by the system 200, and so forth; a patient bed 250; and a user input device configured to receive user input (not illustrated in Fig. 1) such as a keyboard, a mouse, a touchscreen, and so forth.
  • the system may also include a robotic device for manipulating the treatment device.
  • Example 1 A computer-implemented method of predicting a success metric (110) achieved by performing a treatment procedure on a thrombus (120), the method comprising: receiving (SI 10) angiographic image data, including one or more angiographic images (130) comprising the thrombus (120); inputting (S120) the angiographic image data into a neural network (140); and calculating (S130) the success metric (110) based on the output of the neural network (140); and wherein the neural network (140) is trained using training data comprising angiographic training images (130’) representing the treatment procedure, and corresponding ground truth procedure outcome data.
  • Example 2 The computer-implemented method according to Example 1, wherein the neural network (140) is trained to generate latent space representations (zi) for the inputted angiographic images (130), and wherein the method further comprises: generating, for each inputted angiographic image, a latent space representation (zi), using the neural network (140); and wherein the calculating (S130) the success metric (110) based on the output of the neural network (140), comprises analyzing the ground truth procedure outcome data (150’GT) of angiographic training images (130’) having latent space representations (z t ) within a predetermined distance (160) of the latent space representation (zi) of the inputted angiographic image, and calculating the success metric (110) for the inputted angiographic image based on the analyzed ground truth procedure outcome data.
  • the neural network (140) is trained to generate latent space representations (zi) for the inputted angiographic images (130)
  • the method further comprises: generating, for each inputted angi
  • Example 3 The computer-implemented method according to Example 1, wherein the neural network (140) is trained to classify the inputted angiographic images (130) with an expected outcome ( 15Oo,i) of the procedure, and wherein the method further comprises: classifying each inputted angiographic image (130) with an expected outcome ( 15Oo,i) of the procedure, using the neural network (140); and wherein the calculating (S130) the success metric (110) based on the output of the neural network (140), comprises calculating the success metric (110) for the inputted angiographic image based on the classified expected outcome ( 15Oo,i) of the procedure.
  • Example 4 The computer-implemented method according to Example 1, wherein the received angiographic image data includes a plurality of angiographic images (130) comprising the thrombus (120), wherein the angiographic images (130) further include a deployment catheter (170) for deploying a mechanical thrombectomy device to treat the thrombus (120), and wherein the success metric is the success metric achieved by deploying the mechanical thrombectomy device from the deployment catheter; and wherein the neural network (140) is trained to generate latent space representations (zi) for the inputted angiographic images (130) and to predict, from the generated latent space representations (zi), future angiographic images (130f U ture) to the inputted angiographic images (130), the future angiographic images (130f u ture) including predicted future positions of the deployment catheter (170), and wherein the method further comprises: generating, for each inputted angiographic image (130), a latent space representation (zi), using the neural network (
  • Example 5 The computer-implemented method according to any one of Examples 2 - 4, wherein: the angiographic image data includes one or more 2D angiographic images, and one or more corresponding 3D angiographic images; and wherein the training data comprises a plurality of 2D angiographic training images, and a plurality of corresponding 3D angiographic training images; and/or wherein the method further comprises: receiving device data for a mechanical thrombectomy device to be used in the treatment procedure; and inputting the device data into the neural network (140); and wherein the neural network (140) is trained further using device data corresponding to the angiographic training images (130’); and/or wherein the method further comprises: receiving patient data relating to the thrombus (120); and inputting the patient data into the neural network (140); and wherein the network (140) is trained further using patient data corresponding to the angiographic training images (130’).
  • Example 6 The computer-implemented method according to any previous Example, further comprising: calculating a confidence value for each of the inputted angiographic images (130); and outputting the confidence value.
  • Example 7 The computer-implemented method according to any previous Example, wherein the success metric (110) is calculated for a plurality of different types of treatment procedures.
  • Example 8 The computer-implemented method according to any previous Example, wherein the angiographic images (130) further comprise a deployment catheter (170) for deploying a mechanical thrombectomy device to treat the thrombus (120), and wherein the success metric (110) is the success metric achieved by deploying the mechanical thrombectomy device from the deployment catheter.
  • Example 9 The computer-implemented method according to Example 8 wherein the received angiographic image data includes a temporal sequence of real-time angiographic images, and wherein the success metric (110) is provided in real-time for each angiographic image.
  • Example 10 The computer-implemented method according to Example 9 when dependent on
  • Example 2 or Example 4 further comprising: receiving user input indicative of an extent of the predetermined distance (160), and wherein the method further comprises outputting: a graphical representation of the latent space representation of the inputted angiographic image (zi); a graphical representation of the latent space representations of at least some of the angiographic training images (130’) used to train the neural network (z t ); and an indication of the predetermined distance (160).
  • Example 11 The computer-implemented method according to Example 1, wherein: the success metric (110) represents a probability of success; and/or the ground truth procedure outcome data (150’GT) represents a binary classification of success or failure of the procedure; and/or the ground truth procedure outcome data (150’GT) is based on one or more of: a speed of the procedure, a measure of completeness of re-perfusion achieved by the procedure, a mortality rate subsequent to the procedure, and whether the procedure needed to be repeated.
  • the success metric (110) represents a probability of success
  • the ground truth procedure outcome data (150’GT) represents a binary classification of success or failure of the procedure
  • the ground truth procedure outcome data (150’GT) is based on one or more of: a speed of the procedure, a measure of completeness of re-perfusion achieved by the procedure, a mortality rate subsequent to the procedure, and whether the procedure needed to be repeated.
  • Example 12 The computer-implemented method according to Example 2, wherein the neural network (140) is trained to generate latent space representations (zi) for the inputted angiographic images (130), by: receiving angiographic training data, including a plurality of angiographic training images (130’), and wherein each training image comprises a thrombus (120’); inputting the angiographic training data into the neural network (140); and for each of a plurality of the inputted angiographic training images (130’): generating a latent space representation (z t ) of the inputted angiographic training image, using the neural network (140); reconstructing the inputted angiographic training image (130’) from the latent space representation (z t ), using the neural network (140); and adjusting parameters of the neural network (140) based on a difference between the inputted angiographic training image and the reconstructed inputted angiographic training image; and repeating the generating, the reconstructing, and the adjusting, until a stopping
  • Example 13 The computer-implemented method according to Example 2 or Example 3, wherein the neural network (140) is trained to generate latent space representations (zi) for the inputted angiographic images (130), by: receiving angiographic training data, including a plurality of angiographic training images (130’), and wherein each training image comprises a thrombus (120’); receiving ground truth procedure outcome data (150’GT) corresponding to the angiographic training data, the ground truth procedure outcome data representing, for each angiographic training image, a success or a failure achieved by performing the treatment procedure on the thrombus; inputting the angiographic training data into the neural network; and for each of a plurality of the inputted angiographic training images: generating a latent space representation (z t ) of the inputted angiographic training image, using the neural network (140); predicting a procedure outcome ( 15O’o,i) achieved by performing the procedure on the thrombus from the latent space representation (z t ), using the
  • Example 14 The computer-implemented method according to Example 3, wherein the neural network (140) is trained to classify the inputted angiographic images (130) with an expected outcome ( 15Oo,i) of the procedure, by: receiving angiographic training data, including a plurality of angiographic training images (130’), and wherein each training image comprises a thrombus (120); receiving ground truth procedure outcome data (150’GT) corresponding to the angiographic training data, the ground truth procedure outcome data representing, for each angiographic training image a success or a failure achieved by performing the treatment procedure on the thrombus; inputting the angiographic training data into the neural network (140); and for each of a plurality of the angiographic training images (130’): predicting a procedure outcome (15O’o,i) achieved by performing the procedure on the thrombus, using the neural network (140); and adjusting parameters of the neural network (140) based on a difference between the predicted procedure outcome (15O’o,i) and the ground truth
  • Example 15 The computer-implemented method according to Example 4, wherein the neural network (140) is trained to generate latent space representations (zi) for the inputted angiographic images (130) and to predict, from the generated latent space representations (zi), future angiographic images (130f u ture) to the inputted angiographic images (130), the future angiographic images including predicted future positions of the deployment catheter (170), by: receiving angiographic training data, including a plurality of temporal sequences of angiographic training images (130’), and wherein each temporal sequence of training images comprises a thrombus (120’) and a deployment catheter (170’); inputting the angiographic training data into the neural network (140); and for each temporal sequence of angiographic training images (130’): generating, for each angiographic training image in the temporal sequence, a latent space representation (z t ) of the inputted angiographic training image, using the neural network (140); predicting, from the generated latent

Landscapes

  • Health & Medical Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Medical Informatics (AREA)
  • Public Health (AREA)
  • Epidemiology (AREA)
  • General Health & Medical Sciences (AREA)
  • Primary Health Care (AREA)
  • Biomedical Technology (AREA)
  • Radiology & Medical Imaging (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Pathology (AREA)
  • Apparatus For Radiation Diagnosis (AREA)

Abstract

L'invention concerne un procédé, mis en œuvre par ordinateur, de prédiction d'une mesure de réussite (110) obtenue par la réalisation d'une procédure de traitement sur un thrombus (120). Le procédé consiste : à recevoir (S110) des données d'images angiographiques, comprenant une ou plusieurs images angiographiques (130) comportant le thrombus (120) ; à entrer (S120) les données d'images angiographiques dans un réseau neuronal (140) ; et à calculer (S130) la mesure de réussite (110) sur la base de la sortie du réseau neuronal (140).
PCT/EP2022/083396 2021-12-08 2022-11-28 Mesure de traitement de thrombus WO2023104559A1 (fr)

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
US202163287165P 2021-12-08 2021-12-08
US63/287,165 2021-12-08
EP22156482.6 2022-02-14
EP22156482.6A EP4195215A1 (fr) 2021-12-08 2022-02-14 Métrique de traitement du thrombus

Publications (1)

Publication Number Publication Date
WO2023104559A1 true WO2023104559A1 (fr) 2023-06-15

Family

ID=84488052

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/EP2022/083396 WO2023104559A1 (fr) 2021-12-08 2022-11-28 Mesure de traitement de thrombus

Country Status (1)

Country Link
WO (1) WO2023104559A1 (fr)

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200120233A1 (en) * 2018-10-10 2020-04-16 Onfido Ltd Image set alignment
US20200184660A1 (en) * 2018-12-11 2020-06-11 Siemens Healthcare Gmbh Unsupervised deformable registration for multi-modal images
US20210137384A1 (en) * 2017-12-13 2021-05-13 Washington University System and method for determining segments for ablation
US20210236080A1 (en) * 2020-01-30 2021-08-05 GE Precision Healthcare LLC Cta large vessel occlusion model
WO2021188843A1 (fr) * 2020-03-18 2021-09-23 Qualcomm Technologies, Inc. Gestion d'occlusion dans la poursuite de siamois à l'aide de pertes structurées
US20210334935A1 (en) * 2018-11-09 2021-10-28 Samsung Electronics Co., Ltd. Image resynthesis using forward warping, gap discriminators, and coordinate-based inpainting
US20210334965A1 (en) * 2020-01-07 2021-10-28 Cleerly, Inc. Systems, methods, and devices for medical image analysis, diagnosis, risk stratification, decision making and/or disease tracking
US20210343063A1 (en) * 2020-05-04 2021-11-04 Microsoft Technology Licensing, Llc Computing photorealistic versions of synthetic images

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210137384A1 (en) * 2017-12-13 2021-05-13 Washington University System and method for determining segments for ablation
US20200120233A1 (en) * 2018-10-10 2020-04-16 Onfido Ltd Image set alignment
US20210334935A1 (en) * 2018-11-09 2021-10-28 Samsung Electronics Co., Ltd. Image resynthesis using forward warping, gap discriminators, and coordinate-based inpainting
US20200184660A1 (en) * 2018-12-11 2020-06-11 Siemens Healthcare Gmbh Unsupervised deformable registration for multi-modal images
US20210334965A1 (en) * 2020-01-07 2021-10-28 Cleerly, Inc. Systems, methods, and devices for medical image analysis, diagnosis, risk stratification, decision making and/or disease tracking
US20210236080A1 (en) * 2020-01-30 2021-08-05 GE Precision Healthcare LLC Cta large vessel occlusion model
WO2021188843A1 (fr) * 2020-03-18 2021-09-23 Qualcomm Technologies, Inc. Gestion d'occlusion dans la poursuite de siamois à l'aide de pertes structurées
US20210343063A1 (en) * 2020-05-04 2021-11-04 Microsoft Technology Licensing, Llc Computing photorealistic versions of synthetic images

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
ALVERNE, F. ET AL.: "Unfavorable Vascular Anatomy during Endovascular Treatment of Stroke: Challenges and Bailout Strategies", JOURNAL OF STROKE, vol. 22, no. 2, 2020
BERNAVA, G. ET AL.: "Direct thromboaspiration efficacy for mechanical thrombectomy is related to the angle of interaction between the aspiration catheter and the clot", JOURNAL OFNEUROINTERVENTIONAL SURGERY, vol. 12, no. 4, 2020, pages 396 - 400

Similar Documents

Publication Publication Date Title
CN110998744A (zh) 针对弥漫性和串联性病变中冠状动脉疾病的非侵入性评估和治疗指导
WO2019025270A9 (fr) Évaluation non invasive et guidage de thérapie pour une coronaropathie dans des lésions diffuses et en tandem
US20230000364A1 (en) Methods and systems for assessing a vasculature
KR102217392B1 (ko) 관상동맥 진단영상 학습장치와 방법, 및 상기 학습장치와 방법을 통해 구축된 학습모델을 사용하는 유의한 관상동맥 협착 진단 장치 및 방법
EP4195215A1 (fr) Métrique de traitement du thrombus
US20230394654A1 (en) Method and system for assessing functionally significant vessel obstruction based on machine learning
WO2023104559A1 (fr) Mesure de traitement de thrombus
US20240020877A1 (en) Determining interventional device position
EP4252666A1 (fr) Détermination d'une valeur de propriété physique d'un thrombus
WO2023186514A1 (fr) Détermination d'une valeur d'une propriété physique d'un thrombus
EP4254428A1 (fr) Prédiction d'étape de procédé intravasculaire
EP4181058A1 (fr) Angiographie à résolution temporelle
EP4173585A1 (fr) Méthode pour l'identification d'un site d'accès
WO2023186610A1 (fr) Prédiction d'étape de procédure intravasculaire
US20230334686A1 (en) Determining interventional device shape
WO2023083700A1 (fr) Angiographie à résolution temporelle
EP4345839A1 (fr) Évaluation du comportement d'un opérateur pendant une procédure médicale
EP4353176A1 (fr) Fourniture de données de plaque biomécanique pour un système de simulation de dispositif d'intervention
WO2023072752A1 (fr) Notation d'image de projection de rayons x
WO2024110335A1 (fr) Fourniture d'images de projection
US20240029257A1 (en) Locating vascular constrictions
EP4311491A1 (fr) Détermination d'une valeur d'un paramètre de flux sanguin
WO2023072973A1 (fr) Identification de site d'accès vasculaire
EP4125033A1 (fr) Prédiction de l'état d'une procédure d'embolisation
CN118159216A (zh) 识别血管进入部位

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22822478

Country of ref document: EP

Kind code of ref document: A1