Automated Coronary Angiography Analysis

Abstract

It is an object of the present invention to provide a method and apparatus that allows for the acquisition and analysis of diagnostic image data in a reliable, repeatable, objective and efficient manner. A method and apparatus for analyzing diagnostic image data is provided in which a plurality of acquired images of a blood vessel of interest acquired with a predetermined acquisition method are received at a trained classifier and classified by the classifier to extract at least one quantitative feature of the blood vessel of interest from at least one acquired image of the plurality of acquired images, where the at least one quantitative feature is output in association with the at least one acquired image while acquisition of the diagnostic image data is still in progress, and one or more adjustable image acquisition settings are adjusted based on the at least one quantitative feature to optimize acquisition of the diagnostic image data.

Description

[Technical Field]

[0001] The present invention relates to a method, a corresponding device, and a respective computer program for analyzing diagnostic image data, in particular X-ray angiography image data. In particular, the present invention relates to an improved method and device that allows automatically deriving quantitative feature information from diagnostic image data acquired using predefined acquisition settings and using the quantitative feature information thus derived to adjust the predetermined acquisition settings accordingly, thereby improving the quality of the acquisition process. [Background technology]

[0002] Today, coronary artery disease is a widespread disease in many societies. Various types of coronary artery disease are known. In order to develop an appropriate treatment plan for each coronary artery disease, it is very important that the disease is correctly evaluated, i.e., that the medical data by which the coronary artery disease can be evaluated have high quality and high reliability.

[0003] One well-established method used in clinical practice for the evaluation of coronary artery disease is coronary X-ray angiography. Coronary angiography is typically performed by injecting a contrast agent into a blood vessel and then irradiating the contrast-filled coronary vessels with X-ray radiation to obtain a sequence of angiographic images in which these vessels, and therefore the coronary vasculature, are clearly visible. This means that the number and orientation of these angiographic image sequences, the dose of contrast agent, and the respective analysis of the image data may vary from patient to patient, making objective analyses that are comparable across different patients extremely difficult.

[0004] To overcome these problems, improved acquisition approaches have been developed in which predetermined acquisition settings are used to acquire angiographic image sequences. The use of these predetermined acquisition settings reduces the variability of the acquired data because the specific acquisition settings are known for each patient.

[0005] One such acquisition approach is the Xper Swing acquisition, in which angiographic image data is acquired at different orientations along a predetermined, repeatable trajectory with a predetermined dose of contrast agent. The Xper Swing acquisition thereby provides angiographic image data as a single image sequence that can be analyzed for evaluation of specific coronary arteries. However, the quality of angiographic image data acquired using Xper Swing still varies due to patient-dependent optimization of specific acquisition settings, patient-to-patient variations in anatomy, and residual variability in acquisition settings.

[0006] As a result, automating data analysis is difficult: it requires complex calculations that take into account all the variability that may occur in the data for different patients.

[0007] To date, no approach has been developed that allows for reliable, repeatable, objective, and efficient acquisition and analysis of diagnostic image data, such as angiographic image data, particularly diagnostic image data acquired using specific, predetermined acquisition settings. Summary of the Invention [Problem to be solved by the invention]

[0008] It is therefore an object of the present invention to provide a method and apparatus that allows for the acquisition and analysis of diagnostic image data in a reliable, repeatable, objective and efficient manner.

[0009] More specifically, it is an object to provide a method and apparatus for efficiently acquiring diagnostic image data using a medical imaging modality, such as X-ray angiography, at predetermined acquisition settings to obtain reliable results that allow for objective and patient-independent assessment of coronary artery disease. Even more specifically, it is an object of the present invention to provide a method and apparatus that allows for a reduction in the radiation dose and contrast agent dose delivered to a patient during image-based coronary artery disease assessment. [Means for solving the problem]

[0010] This object is achieved by a method for analyzing diagnostic image data, comprising the steps of: receiving diagnostic image data having a plurality of acquired images of a blood vessel of interest in a trained classifier, the diagnostic image data being acquired using a predetermined acquisition method; classifying the diagnostic image data to extract at least one quantitative feature of the blood vessel of interest from at least one acquired image of the plurality of acquired images; outputting the at least one quantitative feature of the blood vessel of interest associated with the at least one acquired image while acquisition of the diagnostic image data is still in progress; and adjusting one or more adjustable image acquisition settings based on the at least one quantitative feature to optimize acquisition of the diagnostic image data.

[0011] That is, this object is solved by a method that uses a trained classifier, such as a convolutional neural network, to automatically analyze diagnostic image data already during the acquisition of said diagnostic image data in order to adjust a set of adjustable acquisition settings, such as specific acquisition parameters, during said ongoing acquisition in order to optimize data acquisition for specific vascular characteristics, i.e. for a specific patient.

[0012] Here, the term diagnostic image data may refer to a set of acquired images representing a patient's vasculature. Here, the term vasculature may refer to a vascular tree or a single blood vessel. Here, the term vasculature may refer specifically to one or more blood vessels and / or segments thereof of interest. Here, the term vessel of interest may refer to a patient's blood vessel that is to be evaluated for potential lesions and / or other diseases using diagnostic image data. In some embodiments, the acquired images of the diagnostic image data may each represent a blood vessel of interest in the coronary vasculature.

[0013] The diagnostic image data may in particular comprise one or more acquired images of said one or more vessels of interest, where the term acquired image may be understood to typically refer to a single image acquired of a vessel of interest, whereby multiple acquired images may be included in the diagnostic image data.

[0014] The plurality of acquired images may be acquired by a medical imaging modality such as computed tomography (CT), ultrasound (US) imaging or magnetic resonance (MR) imaging, among others. In some embodiments, the medical imaging modality may correspond to X-ray angiography, among others, and more particularly to X-ray angiography performed with a set of predetermined acquisition settings, such as a predetermined imaging trajectory and a predetermined dose of contrast agent used.

[0015] In some embodiments, the medical imaging modality may be gated, where a gated medical imaging modality may typically use gated reconstruction, where the acquisition of acquired images is performed in parallel with the acquisition of data providing information across the cardiac cycle, such as electrocardiogram (ECG) or photoplethysmography (PPG) data, so that this data can be used to gate image acquisition and reconstruction with each selected phase point of the cardiac cycle.

[0016] The diagnostic image data is received at a trained classifier. The term classifier may particularly refer to a classifier or classification unit integrated within a respective device for analyzing the diagnostic image data. In some embodiments, the term classifier may also refer to a classifier provided separately from the device. In some embodiments, the classifier may particularly be implemented as a convolutional neural network (CNN).

[0017] The classifier is a trained classifier, i.e., the classifier has been previously trained using a training dataset showing correlations between diagnostic image data and one or more quantitative features, such as vessel length, vessel location, lesion severity, etc. Specifically, training is performed using a training dataset that includes diagnostic image data equivalent to the classification target annotated with each quantitative feature, whereby the annotations may be obtained by manual annotating of the diagnostic image data by a clinical expert, or may be essentially known ground truth if a simulated training dataset is used.

[0018] The weights and parameters of the classifier, particularly the convolutional neural network, are optimized during the training process so that for a training data set input, the resulting neural network output is numerically close to the corresponding annotated feature values. That is, the neural network optimization minimizes the difference between the neural network output and the annotated feature values ​​for all training data sets. The comparison of the neural network output to the annotated feature values ​​may be achieved by various types of appropriate metrics, such as the L2 norm or generalized Dice loss. In some examples, the optimization may use the Adam optimizer.

[0019] During training of the classifier, typically known forms of data augmentation, such as image scaling, translation, or contrast modification, may be used. An exemplary network structure for such a task may be an encoder-decoder neural network architecture.

[0020] The classifier is used to classify the diagnostic image data to extract at least one quantitative feature from the diagnostic image data. That is, based on training, the classifier is enabled to derive values ​​of at least one quantitative feature of a vessel of interest for one or more of the acquired images in the diagnostic image data. In some embodiments, a corresponding value for one particular quantitative feature may be derived for each acquired image. Thus, multiple values ​​for a particular quantitative feature may be derived for multiple acquired images.

[0021] In some embodiments, the quantitative features may correspond specifically to features such as vessel length, vessel location, vessel diameter, lesion severity, myocardial enhancement value, visibility score value, etc. for lesions and / or vessels in individual acquired images, i.e., features that can be derived for each image.

[0022] In some embodiments, the quantitative features may alternatively or additionally include values ​​indicative of fluid dynamics through the vessel of interest, such as a fractional flow reserve (FFR) value, an instantaneous fractional flow reserve (iFR) value, or a coronary flow reserve (CFR) value. Conventionally, these parameters may be derived from a fluid dynamics model capable of modeling fluid dynamics through the vessel of interest, as described, for example, in International Applications WO 2016 / 087396, WO 2020 / 053099 A1, and WO 2019 / 101630 A1. In this embodiment, it may be possible to derive values ​​of these parameters directly from the classifier. That is, by training using a training dataset, the classifier may be enabled to implicitly learn the fluid properties of the vessel(s) of interest and, therefore, the fluid parameters associated therewith, without having to simulate or model the fluid flow through the vessel(s) of interest. This makes it possible to avoid the use of a fluid dynamics model, but rather to obtain the fluid parameters directly from the trained classifier.

[0023] On the other hand, quantitative features may correspond to features related to the diagnostic image data as a whole, such as a completeness score indicating whether sufficient angular information about the vessel of interest is available to obtain a reliable analysis, a baseline deviation index indicating whether the visible vasculature resembles the patient's average baseline, or an occlusion score indicating whether future tomographic reconstructions are likely to show strong artifacts if the current trajectory is continued. Thus, the occlusion score may be used in particular when the implantation of a particular external device is within the field of view. That is, if a particular trajectory results in a device obstructing the field of view in future projections of the planned trajectory, it may be beneficial to modify the trajectory to avoid such obstruction.

[0024] In this case, the quantitative features are output in association with one acquired image, i.e. the values ​​of the quantitative features derived on the basis of a corresponding acquired image are associated with said acquired image and then output for further evaluation and / or further processing, in particular while the image acquisition by the medical imaging modality is still in progress.

[0025] In this case, based on the output, a computing unit or other processing device evaluates at least one quantitative feature, each its value, associated with each acquired image to determine whether the current acquisition settings used render sufficient image quality. In some embodiments, the computing unit may thereby use quantitative features such as visibility scores, completeness scores, etc. If the evaluation indicates that the current acquisition settings do not produce sufficient acquired images, one or more of the adjustable acquisition settings are adjusted. Hereby, the adjustment may be performed automatically, particularly based on the previous classification.

[0026] Here, the term adjustable acquisition settings may refer in particular to acquisition settings used that are not predefined for the medical imaging modality used. In the present context, therefore, a distinction is made between predefined acquisition settings that are not changed, i.e., remain the same, to reduce variability, and adjustable acquisition settings that can be changed according to the individual requirements of each patient.

[0027] By keeping certain acquisition settings constant and predefined while adjusting other acquisition settings based on automatic analysis of previously acquired diagnostic image data, it is possible to acquire diagnostic image data in a reliable, efficient, and reproducible manner, while reducing variability in different data sets acquired at different measurement times for different patients.

[0028] More specifically, during diagnostic image data acquisition, "live" adjustments of acquisition settings can be performed by analyzing a subset of already acquired diagnostic image data, which allows these acquisition settings to be individually optimized for each person, thereby allowing diagnostic image data to be acquired in the most efficient manner. This reduces the amount of radiation and contrast agent delivered to the patient, as sufficient completeness can be achieved more quickly for some patients than for others.

[0029] In some embodiments, the method may be implemented to perform out-of-distribution detection. That is, the method may be implemented to determine whether diagnostic image data input to a computing unit or other processing device falls within an expected distribution based on training of the classifier. This may make it possible to detect whether diagnostic image data is input to the computing unit or other processing device that cannot be associated with the type of diagnostic data the classifier was trained on.

[0030] If so, an indication may be output to the user that the acquired diagnostic image data cannot be properly evaluated because it is not related to the type of diagnostic image data expected to be evaluated. This indication may be a simple warning that the diagnostic image data cannot be evaluated or can only be evaluated improperly. Alternatively or additionally, the indication may comprise a suggestion to perform new or additional diagnostic image data acquisition. In some embodiments, the method may alternatively or additionally be implemented to nevertheless perform an evaluation of the diagnostic image data, whereby output of the evaluation may be made with a respective large error bar.

[0031] In some embodiments, adjusting the one or more adjustable image acquisition settings includes prematurely terminating acquisition of diagnostic image data if an already acquired portion of the diagnostic image data is determined to meet at least one predetermined reliability criterion.

[0032] In some embodiments, adjusting the adjustable acquisition settings may include terminating the acquisition before its scheduled end, particularly if it is determined that sufficient diagnostic information has already been acquired. That is, the acquired diagnostic image data is differentiated into two or more subsets of diagnostic image data, whereby a first subset is evaluated while a second subset is currently being acquired. The size of each subset may be highly dependent on a given medical imaging modality and acquisition quality. A single acquired image may form a subset. In other embodiments, more acquired images may form a subset of diagnostic image data.

[0033] The first subset is evaluated to determine whether the diagnostic information derived therefrom meets a predetermined reliability criterion, i.e., whether sufficient angular information is already present to provide a reliable assessment of the vessel of interest. In some embodiments, the reliability criterion may be specifically quantified with respect to the completeness score. That is, a threshold may be determined for the completeness score, and as soon as the completeness score is higher than the threshold, it is determined that sufficient angular information is available for a reliable diagnosis. Alternatively or additionally, the reliability criterion may include further scores and / or criteria.

[0034] If it is determined that sufficient information is available, the acquisition setting that is adjusted may in particular be the acquisition end time. More specifically, the acquisition end time may be set so that the acquisition is immediately terminated, for example, using a termination signal. By terminating the acquisition as soon as sufficient information is available, it is possible to keep the radiation dose received by the patient as low as possible. On the other hand, if it is determined that the reliability criterion is not met, i.e., sufficient information is not yet available, the measurement may continue, i.e., no adjustment of the adjustable acquisition settings is performed. This feedback loop that allows adjustment of the acquisition time may be repeated frequently until the acquisition is stopped because sufficient information is available.

[0035] In some embodiments, adjusting the one or more adjustable acquisition settings includes adjusting an image acquisition trajectory to improve visibility of the vessel of interest in the diagnostic image data. In some modifications, adjusting the one or more adjustable acquisition settings includes adjusting a contrast injection rate into the vessel of interest during image acquisition.

[0036] In some embodiments, adjusting the adjustable image acquisition settings may additionally or alternatively include adjusting the imaging trajectory used for image acquisition. In this case, a visibility score for the vessel of interest and / or lesions therein is determined for a first subset of diagnostic image data. For this purpose, the entire diagnostic image data is considered, not just individual acquired images. The visibility score may be compared to a preset reference value or threshold, such that sufficient visibility is considered if the score is above (or below) that value, and insufficient visibility is considered if the score is below (or above) that value. If the visibility score indicates insufficient visibility, i.e., insufficient visibility is given, adjusting the adjustable imaging settings may include, in particular, adjusting the image acquisition trajectory used to acquire the acquired images. This allows for improved image quality, meaning that fewer acquired images are needed to obtain sufficient diagnostic information. This effectively reduces the radiation dose delivered to the patient. The adjusted trajectory also prevents diagnoses from having to be made on images with less-than-ideal visibility.

[0037] In some embodiments, adjusting one or more adjustable acquisition settings may include adjusting a contrast injection rate into the vessel of interest. That is, the contrast of the vessel of interest may be determined for a first subset of diagnostic image data using a classifier. By reviewing the contrast, it may be determined whether sufficient contrast has been injected into the vessel of interest. This may involve varying the amount of contrast depending on the patient, as patients with narrower vessels may require less contrast than patients with wider vessels to achieve similar visibility. Therefore, based on the contrast of the vessel of interest, it may be evaluated whether there is sufficient contrast in the vessel of interest, and therefore whether the contrast injection rate is sufficient, or whether the contrast injection rate should be adjusted because too little or too much contrast is currently being injected into the vessel of interest.

[0038] In this case, adjusting the adjustable acquisition settings includes adjusting the contrast injection rate based on the characteristics of the vessel of interest, which may be used to optimize the contrast dose delivered to each patient.

[0039] These examples may also be used in a feedback loop, whereby individual subsets of diagnostic image data may be evaluated frequently during acquisition to perform live adaptation of acquisition settings.

[0040] In some embodiments, the method further comprises acquiring training image data of the vessel of interest according to a predetermined acquisition method; extracting at least one quantitative feature from the training image data; generating at least one training dataset for the classifier, the training dataset comprising training image data associated with the at least one quantitative feature; and training the classifier using the at least one training dataset.

[0041] The classifiers may be trained using respective training data sets, which in some embodiments may be derived based on training image data, where the term training image data may refer in particular to a plurality of training image data acquired in a clinical setting, i.e., measured data, or a plurality of training images generated by simulation.

[0042] One or more quantitative features may then be extracted from each individual training image as well as the training image data as a whole, where whether the individual images or data are used as a whole depends on their respective quantitative values. Feature extraction here may be performed manually by one or more users, automatically by a respective algorithm, or may correspond to quantitative features readily available from simulation of the data.

[0043] Using the extracted quantitative features and the training image data, respective training data sets are generated, i.e., the quantitative feature values ​​are associated with respective training images and / or training image data to derive correlations between the quantitative feature values ​​and the respective image data. The training data sets thus generated may be used to train a classifier.

[0044] In yet another embodiment, the training image data comprises simulated training image data generated by simulating image acquisition according to a predetermined acquisition method, where simulating comprises acquiring at least one three-dimensional geometric model of a vessel of interest, acquiring at least one two-dimensional background image of the vessel of interest, and simulating contrast agent fluid dynamics through the patient's vasculature based on the at least one contrast agent fluid parameter. In some modifications, simulating further comprises obtaining transformation translation and rotation data and augmenting the simulated training image data based on the translation and rotation data. In some modifications, generating the at least one training data set further comprises receiving additional patient data and adjusting the at least one training data set according to the additional patient data.

[0045] In some embodiments, the training image data is generated using a simulation. To this end, at least one three-dimensional geometric model of the patient's vasculature, including the vessel of interest, is obtained. Thereby, the geometric model may be obtained from medical images, which may be acquired by any medical imaging modality that allows acquiring three-dimensional medical images.

[0046] In some embodiments, the medical imaging modality may correspond to the medical imaging modality for which live adaptation is to be performed. In some embodiments, the medical imaging modality may be a different imaging modality. The geometric model may also be purely virtual and defined by common anatomical knowledge. Furthermore, at least one two-dimensional background image of the patient's vasculature, including the vessel of interest, is acquired. Using the background image, it is possible to distinguish between the background and the vasculature in the medical image in order to properly perform vessel identification of the vessels in the vasculature.

[0047] The background image may provide a realistic appearance to the simulated data, whereby the two-dimensional background image may be derived from an actual clinical acquisition and / or constructed from a forward projection of a three-dimensional medical image and / or may be a virtual image designed to mimic a typical background found in the diagnostic data to be simulated.

[0048] Additionally, the three-dimensional medical images and / or the two-dimensional background image may be used to generate a fluid dynamics model that represents the fluid dynamics through the patient's vasculature. In some embodiments, the fluid dynamics model may comprise, among other things, a lumped parameter model.

[0049] The term lumped parameter model may refer, inter alia, to a model in which the fluid dynamics of a vessel are approximated by a topology of distinct entities. As an example, a vascular system, such as a vascular tree, may be represented by a topology of resistive elements, each with a specific resistance. Thus, the outlet at the distal end of the vessel is also represented by a specific resistive element. In this case, this resistive element is connected to ground to represent the termination of the vessel. Similarly, each resistive element may be connected to a series of resistive elements representing the vessel of interest, e.g., to represent the outflow from the vessel of interest at a specific branch point. These resistive elements may typically be connected to ground.

[0050] These lumped parameter models reduce the number of dimensions compared to other approaches, such as Navier-Stokes. Therefore, using lumped parameter models may allow for simplified calculation of fluid dynamics inside blood vessels, which may ultimately result in reduced processing time. The use of such lumped parameter models is described, for example, in International Application WO2016 / 087396.

[0051] The fluid dynamics model thus generated may then be used to simulate the flow of contrast fluid through a patient's vasculature, particularly one or more blood vessels of interest. This allows for the generation of training image data representative of the vasculature and the corresponding fluid dynamics therethrough. In some embodiments, deformation translation and rotation data may be added to the simulation as additional information to enhance the training image data. The training image data thus generated may then be provided to a classifier for training.

[0052] In some embodiments, the at least one quantitative feature comprises one or more of a vascular label of a vessel in the patient's vasculature, and / or a vascular length of a vessel in the patient's vasculature, and / or a severity of a lesion in a vessel in the patient's vasculature, and / or a vascular diameter of a vessel in the patient's vasculature, and / or a visibility score of a lesion and / or vessel in the patient's vasculature, and / or an integrity score for at least one of the plurality of acquired images and / or myocardial staining values.

[0053] In some embodiments, additional patient information, such as ECG data, aortic pressure values, or historical data for a particular patient, may also be added to the training dataset and / or classification. This may have the added advantage that additional patient abnormalities, such as excessively elevated aortic pressure, may be detected in these cases, such that the injection of contrast agent must be adjusted as well.

[0054] According to some embodiments, outputting the at least one quantitative feature for further evaluation comprises displaying the at least one quantitative feature to a user and / or outputting the at least one quantitative feature in a predetermined format for automatic reporting to a reporting entity. In some embodiments, a user may input additional data in response to the output, whereby the additional data may be further used to train a classifier and / or to evaluate the diagnostic image data.

[0055] According to a further aspect, an apparatus for analyzing diagnostic image data is provided, the apparatus comprising: a trained classifier configured to receive diagnostic image data having a plurality of acquired images of a blood vessel of interest, where the diagnostic image data is acquired using a predetermined acquisition method; classify the diagnostic image data to extract at least one quantitative feature of the blood vessel of interest from at least one acquired image of the plurality of acquired images; and output the at least one quantitative feature of the blood vessel of interest associated with the at least one acquired image while acquisition of the diagnostic image data is still ongoing; and a computing unit configured to adjust one or more adjustable image acquisition settings based on the at least one quantitative feature to optimize acquisition of the diagnostic image data.

[0056] In some embodiments, the apparatus further comprises an input unit configured to acquire training image data of the vessel of interest according to a predetermined acquisition method, and a training dataset generation unit configured to extract at least one quantitative feature of the vessel of interest from the training image data and generate at least one training dataset for the classifier, wherein the training dataset comprises the training image data associated with the at least one quantitative feature, and to provide the at least one training dataset to the classifier for training. In some embodiments, the apparatus may also comprise a display unit configured to generate a graphical representation of at least one acquired image of the plurality of acquired images and / or the at least one quantitative feature, and a user interface configured to receive user input in response to the graphical representation.

[0057] In a further aspect, a computer program for controlling an apparatus according to the invention is provided, which computer program is arranged to perform the method steps according to the invention when executed by a processing unit. In yet another aspect, a computer readable medium having stored thereon the above-cited computer program is provided.

[0058] It is to be understood that the method of claim 1, the apparatus of claim 11, the computer program of claim 14 and the computer-readable medium of claim 15 have similar and / or identical preferred embodiments, in particular as defined in the dependent claims.

[0059] It is to be understood that a preferred embodiment of the invention can be any combination of the dependent claims or the above embodiments with the respective independent claim.

[0060] These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter. [Brief explanation of the drawings]

[0061] [Figure 1] 1 illustrates a schematic diagram of an apparatus for analyzing diagnostic image data blood vessels according to one embodiment. [Figure 2] 1 shows a flowchart of a method for analyzing diagnostic image data according to one embodiment. [Figure 3] 1 shows a flowchart of a method for generating training data according to one embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0062] The figures in the drawings are schematic and in different drawings similar or identical elements are provided with the same reference numbers.

[0063] 1 illustrates schematically an exemplary embodiment of an apparatus 1 for analyzing diagnostic image data. The apparatus 1 comprises an input unit 100, a training dataset generation unit 200, a classification unit 300, a calculation unit 400, and a display unit 500. Furthermore, the classification unit 300 and the communication unit 400 are communicatively coupled to a medical imaging modality 2 in a feedback loop 600.

[0064] The input unit 100 is configured to receive training image data 10 of a patient's vasculature. The training image data 10 may correspond to or include image data previously acquired using, in particular, a predetermined acquisition method, i.e., an acquisition method performed with one or more predetermined (known) acquisition settings, such as a known contrast agent dose and acquisition trajectory. In a particular embodiment according to Fig. 1, the training image data 10 may correspond in particular to clinical data acquired using X-ray angiography using a C-arm. That is, in the particular embodiment of Fig. 1, the training image data 10 is derived from actual measurement data.

[0065] However, it should be understood that, alternatively or additionally, the training image data 10 may be generated using simulations, etc. For simulated training image data 10, three-dimensional medical images typically acquired using CT and / or MR imaging modalities may be acquired and used to generate a three-dimensional model of the vessel of interest, and combined with two-dimensional background data showing cardiac images without contrast filling of the arteries. The training image data and corresponding training dataset are generated based on the contrast injection parameters used for a given acquisition method and the corresponding fluid dynamics model. Thus, transformation translations and rotations may be applied to the three-dimensional representation of the vessel of interest and the two-dimensional background projection to achieve data augmentation. In this case, the entire range of acquisition trajectories is typically covered by the cardiac motion model.

[0066] The input unit 100 provides the training image data 10 to the training dataset generation unit 200. The training dataset generation unit 200 is configured to extract one or more quantitative features of the patient's vasculature, in particular the vessels of interest, from the training image data 10. In some embodiments, these quantitative features may relate to, among others, a vessel label of a vessel within the vasculature, a vessel number, a vessel location and / or a vessel length, a lesion or lesion severity in one or more vessels of interest, a myocardial enhancement value, a vessel diameter of the vessel of interest, a visibility score of a lesion in the vessel of interest for each individual training image of the training image data, a completeness score indicating whether sufficient angle information for a given vessel is available to enable a reliable analysis, a norm deviation index indicating the similarity of the visible vasculature to a norm, etc.

[0067] Using the extracted quantitative features, the training dataset generation unit 200 is configured to generate at least one training dataset having the training image data 10 and respective predetermined features associated with one or more of the training images in the training image data 10. Thus, the training dataset generation unit 200 obtains correlations between the training image data 10 and the extracted predetermined features and generates a corresponding dataset having correlation information, which is then provided to the classification unit or classifier 300 as the training dataset 20.

[0068] The classification unit 300 has an input port 301 configured to receive a training dataset from the training dataset generation unit 200. The classifier 300 uses the training dataset 20, or optionally multiple training datasets 20, to train relationships between quantitative features and training images in the training image data 10. In the exemplary embodiment according to Fig. 1, the classification unit comprises or corresponds to a convolutional neural network, in some embodiments a deep convolutional neural network. That is, the classification unit 400 implements multiple convolutional layers in combination with pooling layers.

[0069] The training data set 20 input to the classification unit 300 according to the particular embodiment of Fig. 1 corresponds to a plurality of training images acquired using X-ray angiography. In particular, in the particular embodiment of Fig. 1, 10,000 individual angiographic images are used as respective training images. These training images are provided with respective feature data in terms of a pixel mask provided for each individual angiographic image, whereby each pixel is classified as belonging to either the left anterior descending artery (LAD), the left circumflex artery (LCX), the obtuse marginal artery (OM), the right coronary artery, etc., or as belonging to the background. Alternatively or additionally, the training images may be provided with feature data having a single value for each angiographic image indicating the minimum diameter of the artery and / or indicating that (part of) the artery is not visible.

[0070] During training, in the particular embodiment of the classifier 300 according to FIG. 1, the weights and parameters of the neural network are optimized for the input training data set 20 so that the resulting neural network output is numerically close to the corresponding annotated feature values. That is, the neural network optimization minimizes the difference between the neural network output and the annotated feature values ​​for all training data sets. Thereby, the comparison of the neural network output with the annotated feature values ​​can be realized by various types of suitable metrics, such as the L2 norm or the generalized Dice loss. In the particular embodiment according to FIG. 1, the optimization may use, in particular, the Adam optimizer.

[0071] During training, typical known forms of data augmentation may be used, such as image scaling, translation, or contrast modification. An exemplary network structure for such a task may be an encoder-decoder neural network architecture.

[0072] Upon completion of training using the training dataset 20, the classification unit 300 is configured to receive, via input port 302, a first subset of diagnostic image data 30 obtained for a particular patient from the medical imaging modality 2. The first subset of diagnostic image data 30 may in particular comprise a plurality of acquired images 31 acquired using a predetermined acquisition method, whereby the predetermined acquisition method corresponds to the predetermined acquisition method for the training image data, to ensure that the classification unit 300 is trained with an appropriate training dataset to accurately classify the diagnostic image data 30.

[0073] 1, the input to the classifier corresponds to multiple acquired images 31 in diagnostic image data 30, with each acquired image 31 corresponding to a single two-dimensional X-ray angiogram image. Alternatively, the multiple acquired images 31 may correspond to a time-series stack of multiple two-dimensional angiogram images, such as respective C-arm angulations. That is, the input to the classifier corresponds to the same diagnostic image data 30 that is presented to a user, such as a physician, for visual review.

[0074] After classification of the diagnostic image data, at least one quantitative feature suitable for analyzing the diagnostic image data 30 is extracted from the diagnostic image data 30 .

[0075] The extracted quantitative feature values ​​and the first subset of diagnostic image data 30 comprising one or more acquired images 31 are then provided to the computing unit 400 for further processing. It will be appreciated that the first subset of diagnostic image data 30 is provided to the computing unit 400 for further processing while the acquisition of the second subset of diagnostic image data 30 is still in progress. This allows the evaluation by the computing unit 400 to be used to adjust the image acquisition where possible and / or necessary.

[0076] In other words, the computation unit 400 determines, based on the first subset of diagnostic image data and the extracted quantitative features, whether adjustment of acquisition parameters for the image acquisition may be beneficial. In the particular example of Figure 1, the computation unit 400 derives, for that purpose, a reliability criterion for the diagnostic information to be derived from the diagnostic image data 30 and the quantitative features.

[0077] Furthermore, the computation unit 400 processes the first subset of diagnostic image data 30 and the quantitative features derived therefrom to determine whether a reliability criterion is met. In certain embodiments, this is achieved by comparing the diagnostic information that can be derived from the first subset of diagnostic image data 30 and the quantitative features to a threshold that indicates sufficiency of the diagnostic information.

[0078] The reliability criterion is considered to be met when sufficient diagnostic information is obtained. In this case, the computing unit 400 is configured to adjust the adjustable image acquisition settings by outputting a corresponding termination signal to the medical imaging modality 2, i.e., by adjusting the acquisition settings such that the acquisition is terminated early, i.e., before the originally set end point. In response to said termination signal, the medical imaging modality 2 terminates further image acquisition, thereby avoiding unnecessary radiation and contrast agent doses being delivered to the patient.

[0079] On the other hand, if the calculation unit 400 determines that the reliability criterion is not met, i.e., sufficient information is not yet available, the calculation unit 400 does not output any termination signal to the medical imaging modality 2, and the medical imaging modality 2 continues to acquire a second subset of diagnostic image data.

[0080] It should be appreciated that the above-described evaluation process may be repeated for the second subset (and any subsequent subsets) of diagnostic image data 30 until the reliability criteria are met, thereby allowing computing unit 400 to terminate the acquisition procedure as soon as it is determined that sufficient diagnostic information is available.

[0081] In some embodiments, adjusting the adjustable image acquisition settings may additionally or alternatively include adjusting the imaging trajectory used for image acquisition. In this case, evaluating the first subset of diagnostic image data 30 may include determining a visibility score for the vessel of interest in each acquired image. If the computing unit 400 registers insufficient visibility, the computing unit 400 may be configured to automatically adjust the imaging trajectory to improve the visibility of the vessel of interest. By adjusting the imaging trajectory to improve visibility, fewer acquired images 31 are required to obtain sufficient diagnostic information, thereby optimizing the radiation dose delivered to the patient.

[0082] In another embodiment, the computing unit 400 may evaluate the first subset of diagnostic image data 30 together with the extracted quantitative features to determine the contrast of the vessels of interest. This allows for determining whether sufficient contrast agent has been injected into the vessels of interest. The amount of contrast agent required to provide sufficient visibility of the vessels of interest may vary from patient to patient. Thus, patients with narrower vessels may require less contrast agent, while patients with wider vessels may require more contrast agent to achieve similar visibility. Therefore, as a further adjustable acquisition setting, the computing unit 400 may be configured to adjust the contrast agent injection rate based on the characteristics of the vessels of interest, such that a lower rate is used for patients with narrower vessels (i.e., requiring less contrast agent) and a higher rate is used for patients with wider vessels (i.e., requiring more contrast agent). Using this adjustment, the contrast agent dosage delivered to each patient may be optimized.

[0083] Similarly, in these examples, it will be appreciated that the above-described evaluation process may be repeated for the second subset (and any subsequent subsets) of diagnostic image data 30 until the reliability criteria are met, i.e., until sufficient diagnostic information is available.

[0084] It should further be understood that different adjustment procedures may be combined with one another, so that the computing unit 400 may be configured to adjust the contrast injection rate according to the patient's respective vascular characteristics and further to terminate the acquisition procedure as soon as it is determined that sufficient diagnostic information is available.

[0085] With this configuration, a feedback loop is realized that allows live adaptation of acquisition parameters to optimize diagnostic image data acquisition.

[0086] 1, the diagnostic image data 30, together with the extracted features, are further provided to a display unit 500. The display unit 500 may have, among other things, a screen 501 for displaying information graphically and a user interface 502, such as a keyboard, touchpad, mouse, touchscreen, etc., configured to allow a user to provide input and generally operate the device.

[0087] The display unit 500 is configured to generate a graphical representation of the image data 30 and the extracted quantitative features and present this information to a user on a screen 501. The user may then review the presented information and provide respective input via a user interface 502. The user input may then be used for further evaluation of the data. In some embodiments, the user input may also be used to feed back to the trained classification unit 300 and used by the classification unit for further training.

[0088] Figure 2 shows a flowchart of a method 1000 for analysing diagnostic image data using the apparatus 1 according to Figure 1. In step S101, the input unit 100 receives training image data 10, which may be generated as described in relation to Figure 3. Alternatively or additionally, the training image data may be generated by different means.

[0089] In step S102, the input unit 100 provides training image data 10 to the training dataset generation unit 200. In step S201, the training dataset generation unit 200 receives the training image data 10 and in step S202 extracts one or more quantitative features of the patient's vasculature, in particular one or more blood vessels of interest in the patient's vasculature, from the training image data 10. Optionally, in step S203, the training dataset generation unit 200 associates the one or more quantitative features with the training image data 10.

[0090] In step S204, the training dataset generation unit 200 generates at least one training dataset 20. This at least one training dataset 20 is provided to the classification unit 300 in step S205.

[0091] In step S301, the classification unit 300 receives the training dataset 20 from the training dataset generation unit 200. In step S302, the classification unit 300 uses the training dataset 20 for training, in this case as described in relation to FIG.

[0092] After this, the classification unit 300 receives in step S303 a first subset of diagnostic image data 30 acquired by the medical imaging modality 2. In step S304, the classification unit 300 classifies a plurality of acquired images in the first subset of diagnostic image data 30 in order to extract at least one quantitative feature, in particular at least one value for the at least one quantitative feature, from the at least one acquired image 31 of the diagnostic image data 30.

[0093] Thereby, classification is performed based on previous training using the training dataset 20. In step S305, the classification unit 300 provides the diagnostic image data 30 together with the extracted features to the computation unit 400 for further processing. Here again, it will be appreciated that the first subset of diagnostic image data 30 is provided to the computation unit 400 for further processing while acquisition of the second subset of diagnostic image data 30 is still in progress, thereby enabling the evaluation by the computation unit 400 to be used to adjust image acquisition.

[0094] To this end, in step S401, the calculation unit 400 receives a first subset of diagnostic image data 30 together with the extracted quantitative features and, in step S402, evaluates the received data to determine whether adjustment of one or more adjustable acquisition settings may be necessary. In the specific embodiment of Figure 2, this means that the calculation unit 400 determines, based on the first subset of diagnostic images 30, a visibility score for the vessel of interest in each of the individual acquired images 31 within the diagnostic images 30. In step S403, the calculation unit 400 then compares the visibility score with a reference value to determine whether the visibility of the vessel of interest is sufficient or needs to be improved.

[0095] In the former case ("Y"), i.e., if the visibility is sufficient, the method proceeds to step S404. In the latter case ("N"), i.e., if the visibility needs to be improved in step S403, the method proceeds to step S404'. In step S404', the computing unit 400 determines an optimized imaging trajectory to improve the visibility, and in step S407 generates a corresponding adjustment signal and provides the adjustment signal to the medical imaging modality 2 to automatically adjust the imaging trajectory. In response to the adjustment signal, the medical imaging modality adjusts the imaging trajectory used to acquire a second subset of diagnostic image data 30. In this case, the method is repeated in a loop starting from step S303 using the second subset of diagnostic image data 30.

[0096] As indicated herein above, if it is determined in step S403 that the visibility is sufficient (“Y”), the method proceeds to step S404. In step S404, the calculation unit 400 evaluates the first subset of diagnostic image data 30 to determine whether sufficient diagnostic information can be derived from the accumulation of acquired images 31 in the first subset of diagnostic image data 30. If so (“Y”), the calculation unit 400 generates a termination signal in step S545 and provides said termination signal to the medical imaging modality 2. This results in the termination of image acquisition in step S406.

[0097] Otherwise ("N"), i.e., if it is determined in step S404 that there is insufficient diagnostic information that can be derived from the accumulation of acquired images 31 in the first subset of diagnostic image data 30, the method proceeds to step S405'. In step S405', no termination signal is generated and the method proceeds to steps S303-S404 where a second (or subsequent) subset of diagnostic image data 30 is received and processed. With this configuration, a feedback loop is realized that allows live adaptation of acquisition parameters to optimize diagnostic image data acquisition.

[0098] 3 shows a flowchart of a method 2000 for generating training image data according to one embodiment. In step S2001, a simulation unit acquires at least one medical image of a patient and generates therefrom a three-dimensional geometric model of the patient's vasculature. The medical image may in particular have been obtained by a medical imaging modality. The medical imaging modality may correspond to medical imaging modality 2 or may be a different imaging modality.

[0099] In step S2002, the simulation unit further acquires at least one two-dimensional perspective image of the patient's vasculature to accurately distinguish the background from the vasculature.

[0100] In step S2003, the simulation unit then performs vessel identification to identify vessels within the vasculature. Furthermore, in step S2004, the simulation unit uses the three-dimensional medical image and / or the two-dimensional background image to generate a fluid dynamics model of blood flow through the patient's vasculature. In particular embodiments according to Figure 3, the fluid dynamics model may comprise or correspond to a lumped parameter model, i.e., a model in which the fluid dynamics of the vessels are approximated by the topology of separate entities.

[0101] This model is used to simulate the flow of contrast fluid through the patient's vasculature in step S2005. In step S2006, the simulation unit may optionally receive deformation translation and rotation data as additional information. In step S2007, the simulation unit may then use the additional information to augment the training image data. In step S2008, the training image data is output to be provided to classification unit 400.

[0102] In the above embodiment, the training data is generated based on simulations using a fluid dynamics model, but it should be understood that the training data may be derived from historical clinical data of one or more patients.

[0103] In the above embodiments, adjustments to adjustable acquisition parameters relate to changes in the imaging trajectory and termination of the acquisition process, but it should be understood that other types of adjustments may be made automatically based on classification of already received diagnostic image data, such as adjusting the radiation dose delivered to the target region and / or adjusting the injection rate into the blood vessel of interest.

[0104] Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from a study of the drawings, the disclosure, and the appended claims.

[0105] In the claims, the word "comprise" does not exclude other elements or steps and the indefinite article "a" or "an" does not exclude a plurality.

[0106] A single unit or device may fulfill the functions of several items recited in the claims. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

[0107] Also, procedures that may be described as being performed by a single unit, such as generating a training data set, training a classifier, classifying image data, and simulating training image data to generate training image data, may be performed by multiple units, and certain steps may be performed by the same unit rather than separate units.

[0108] The computer program may be stored / distributed on a suitable medium, such as an optical storage medium or a solid-state medium, supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunications systems.

[0109] Any reference signs in the claims should not be construed as limiting the scope.

[0110] The present invention relates to a method for analyzing diagnostic image data, comprising the steps of receiving diagnostic image data having a plurality of acquired images of a blood vessel of interest in a trained classifier, the diagnostic image data being acquired using a predetermined acquisition method; classifying the diagnostic image data to extract at least one quantitative feature of the blood vessel of interest from at least one acquired image of the plurality of acquired images; outputting the at least one quantitative feature of the blood vessel of interest associated with the at least one acquired image while acquisition of the diagnostic image data is still in progress; and adjusting one or more adjustable image acquisition settings based on the at least one quantitative feature to optimize acquisition of the diagnostic image data.

Claims

1. A method of operating an apparatus for analyzing diagnostic image data, the apparatus having a trained classifier and a computing unit, the method comprising: receiving, by the trained classifier, diagnostic image data having a plurality of acquired images of a vessel of interest, the diagnostic image data having been acquired using a predetermined acquisition method; classifying the diagnostic image data such that the trained classifier extracts at least one quantitative feature of the vessel of interest from at least one acquired image of the plurality of acquired images; the trained classifier outputting the at least one quantitative feature of the vessel of interest associated with the at least one acquired image while the acquisition of the diagnostic image data is still in progress; the computing unit adjusting one or more adjustable image acquisition settings based on the at least one quantitative feature to optimize acquisition of the diagnostic image data; 1. A method of analyzing diagnostic image data having: the predetermined acquisition method is such that the diagnostic image data is acquired at different orientations along a predetermined, repeatable image acquisition trajectory; adjusting the one or more adjustable acquisition settings comprises adjusting an image acquisition trajectory to improve visibility of the vessel of interest in the diagnostic image data. method.

2. adjusting the one or more adjustable image acquisition settings early terminating acquisition of the diagnostic image data if it is determined that an already acquired portion of the diagnostic image data meets at least one predetermined reliability criterion; 2. The method of claim 1, comprising:

3. adjusting the one or more adjustable acquisition settings adjusting the rate of contrast injection into the vessel of interest during image acquisition; 2. The method of claim 1, comprising:

4. The apparatus has an input unit and a training dataset generation unit, and the method comprises: the input unit acquiring training image data of the vessel of interest according to the predetermined acquisition method; the training dataset generation unit extracting the at least one quantitative feature from the training image data; the training dataset generation unit generating at least one training dataset for the classifier, the training dataset comprising the training image data associated with the at least one quantitative feature; using the at least one training data set by the classifier to train the classifier; The method of claim 1 further comprising:

5. The training image data comprises simulated training image data generated by simulating image acquisition according to the predetermined acquisition method, the simulation comprising: obtaining at least one three-dimensional geometric model of the vessel of interest; acquiring at least one two-dimensional background image for the vessel of interest; simulating contrast agent fluid dynamics through the vessel of interest based on at least one contrast agent fluid parameter; 5. The method of claim 4, comprising:

6. The simulating step includes: obtaining deformation translation and rotation data; augmenting the simulated training image data based on the translation and rotation data; The method of claim 5 further comprising:

7. The step of generating at least one training data set comprises: receiving additional patient data; adjusting the at least one training data set according to the additional patient data; The method of claim 4 further comprising:

8. 2. The method of claim 1, wherein the at least one quantitative feature comprises one or more of a vessel label of the vessel of interest, a vessel length of the vessel of interest, a severity of a lesion in the vessel of interest, a vessel diameter of the vessel of interest, a visibility score of a lesion and / or a vessel in the vessel of interest, and / or an integrity score for at least one of the plurality of acquired images and / or myocardial staining values.

9. outputting the at least one quantitative feature, displaying said at least one quantitative feature to a user; and / or outputting said at least one quantitative characteristic in a predetermined format for automatic reporting to a reporting entity; 2. The method of claim 1, comprising:

10. receiving diagnostic image data including a plurality of acquired images of a vessel of interest, said diagnostic image data being acquired using a predetermined acquisition method; classifying the diagnostic image data to extract at least one quantitative feature of the vessel of interest from at least one acquired image of the plurality of acquired images; outputting at least one quantitative feature of the vessel of interest associated with the at least one acquired image while the acquisition of the diagnostic image data is still in progress; a trained classifier configured to a computing unit configured to adjust one or more adjustable image acquisition settings based on the at least one quantitative feature to optimize acquisition of the diagnostic image data; and 1. An apparatus for analyzing diagnostic image data, comprising: the predetermined acquisition method is such that the diagnostic image data is acquired at different orientations along a predetermined, repeatable image acquisition trajectory; adjusting the one or more adjustable acquisition settings comprises adjusting an image acquisition trajectory to improve visibility of the vessel of interest in the diagnostic image data. Device.

11. an input unit configured to acquire training image data of the vessel of interest according to the predetermined acquisition method; a training dataset generation unit configured to extract the at least one quantitative feature of the vessel of interest from the training image data and generate at least one training dataset for the classifier, the training dataset comprising the training image data associated with the at least one quantitative feature, and to provide the at least one training dataset to the classifier for training; The apparatus of claim 10 further comprising:

12. a display unit configured to generate a graphical representation of at least one acquired image and / or the at least one quantitative feature of the plurality of acquired images; a user interface configured to receive user input in response to the graphical representation; The apparatus of claim 10 further comprising:

13. A computer program for controlling an apparatus according to any one of claims 10 to 12, arranged to carry out a method according to any one of claims 1 to 9 when executed by a processing unit.

14. A computer readable medium storing the computer program of claim 13.

Images