Deep Learning for Perfusion in Medical Imaging

Abstract

The present invention relates to deep learning for perfusion in medical imaging. For decision support based on perfusion in medical imaging, a machine-learned model (such as a model trained with deep learning) generates perfusion examination information based on a CT scan of a patient. Other information, such as patient-specific information, can be used together with the CT scan to generate the perfusion examination information. Due to the use of a machine-learned model, the perfusion examination information can be estimated based on a number of scan shots or an amount of CT dose that is sparse in space and / or time. The results of perfusion imaging can be provided in a manner that is less than a current, standard or typical radiation dose.

Description

Background Art

[0001] The present embodiment relates to perfusion examinations in medical imaging. In cardiac imaging, perfusion imaging simultaneously images the detailed anatomy and function of the heart. Perfusion imaging based on computed tomography (CT) can provide advantages. For example, CT perfusion imaging can reduce health care costs by eliminating some hospital stays, cutting the number of tests, and / or speeding up patient diagnosis. These advantages are offset by the risks of patient exposure to x-ray radiation in CT imaging. Because perfusion imaging measures the concentration of contrast agent over time, patients are exposed to more x-ray radiation due to repeated scans. Summary of the invention

[0002] By way of introduction, the preferred embodiments described below include methods, systems, instructions, and non-transitory computer-readable media for decision support based on perfusion in medical imaging. A machine-learned model (such as a model trained with deep learning) generates perfusion examination information based on a CT scan of a patient. Other information (such as patient-specific information) can be used together with the CT scan to generate the perfusion examination information. Because a machine-learned model is used, the perfusion examination information can be estimated based on a number of scan shots or an amount of CT dose that is sparse in space and / or time. The results of the perfusion imaging can be provided in a manner that is less than a current, standard, or typical radiation dose.

[0003] In a first aspect, a method for estimating myocardial perfusion by a medical imager is provided. A computed tomography (CT) system scans a patient. The scan provides coronary CT data representing contrast agent in the patient's heart over time. In addition to the coronary CT data, information specific to the patient is acquired. Myocardial perfusion is estimated by a machine-learned model trained by deep learning. The machine-learned model outputs myocardial perfusion in response to input of the coronary CT data and the information. Myocardial perfusion is displayed, such as a volume or an image.

[0004] In one embodiment, less than five shots are used in the scan. The estimating is performed in response to input of the coronary CT data from no more than the five shots. In other embodiments, the scan uses less than 1 / 2 of the effective dose given a scan protocol. The estimating is performed in response to input of the coronary CT data corresponding to no more than the effective dose.

[0005] Various types of patient information can be obtained for estimation. For example, patient attributes are obtained for the patient. Patient attributes are clinical information such as weight, height, body mass index, gender, age, or performance measures (e.g., cardiac output). As another example, a CT angiogram image of the patient is obtained. In still another example, arterial enhancement curve information is obtained for the patient. Because blockage or other disease changes the flow of contrast agent in the artery, the arterial enhancement curve may be different for different patients, where the difference in flow results in different timing and / or amount of contrast agent for myocardial perfusion. As still another example of patient information, static myocardial measurements are obtained for the patient. Peak perfusion or other perfusion measurements at a given time or at a given event can be calculated and input.

[0006] The machine-learned model is estimated by inputting different information, such as the patient information, the coronary CT data, and injection protocol parameters. The injection protocol parameters characterize the injection of contrast media for the patient, such as the type, rate, volume, and / or location of contrast media injection.

[0007] Various types of outputs can be generated. For example, the myocardial perfusion is estimated as a parametric image. Any perfusion quantity, such as the time to peak perfusion, a data point value, or other perfusion curve characterization is determined for each spatial location, and an image is generated for that parameter. As another example, the myocardial perfusion is estimated as a perfusion quantity, such as peak, time to peak, blood flow, blood volume, mean transit time, or a combination thereof. In other embodiments, the image processor determines a diagnosis or treatment based on the estimated myocardial perfusion and / or based on the application of the machine-learned model. A diagnosis or treatment conclusion or recommendation can be displayed.

[0008] Various types of machine-learned models can be used. For models learned through deep learning, the machine-learned model can be a convolutional neural network, a generative network trained adversarially, or a recurrent neural network.

[0009] In a second aspect, a system for decision support based on perfusion in medical imaging is provided. A computed tomography (CT) scanner is provided for scanning a patient while a contrast agent is in the patient. The CT scanner is configured to output CT data for the patient. The CT data is from a sequence of six or fewer scan shots and represents the contrast agent. A memory is configured to store a machine-learned model. An image processor is configured to estimate perfusion information by inputting the CT data and patient-specific information into the machine-learned model. The perfusion information is provided using no more than six or fewer CT data scan shots representing the contrast agent. A display is configured to display the perfusion information.

[0010] In one embodiment, the CT scanner is configured to scan the patient at an effective dose that is less than 1 / 2 of a dose used for a scan given a scan protocol for perfusion of the six or fewer scans. The image processor is configured to estimate the perfusion information based on no more than the effective dose.

[0011] In an embodiment, the image processor is configured to perform the estimation by inputting the CT data, the patient-specific information, and injection protocol information for injecting the contrast agent into the patient. In still other embodiments, the patient-specific information is patient attributes, angiographic images, arterial enhancement curves, and / or quantification from the CT data. The perfusion information generated by the machine-learned model can be a perfusion image, a quantitative perfusion value, a diagnosis, and / or a treatment.

[0012] In a third aspect, a method for decision support based on perfusion measured by a medical imager is provided. A computed tomography (CT) system scans a patient. The scan provides CT data representing contrast agent in the patient over time, acquired at a CT dose index of 80 or less. Perfusion examination information is generated by a machine-learned model applied to the CT data. The perfusion examination information can be a perfusion image, a perfusion quantification, or a final result of an examination (e.g., a diagnosis and / or treatment). The perfusion examination information is output.

[0013] In one embodiment, diagnosis and / or treatment is provided as said perfusion examination information. In another embodiment, said machine-learned model is a deep-learned neural network applied to said CT data and personalized patient information.

[0014] The present invention is defined by the appended claims, and the content of this section should not be construed as a limitation on those claims. Features or aspects for one type of claim (e.g., method or system) may be used in another type of claim. Additional aspects and advantages of the present invention are discussed below in conjunction with preferred embodiments, and may be claimed later, independently or in combination. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The components and the figures are not necessarily to scale, emphasis instead being placed upon illustrating the principles of the present invention. Moreover, in the figures, like reference numerals designate corresponding parts throughout the different views.

[0016] Figure 1 is a flow chart illustration of one embodiment of a method for decision support based on perfusion measured by a medical imager;

[0017] Figure 2 illustrates an embodiment of generating perfusion examination information based on inputting various types of information into a machine-learned model; and

[0018] Figure 3 is one embodiment of a system for perfusion-based decision support. DETAILED DESCRIPTION

[0019] Deep learning-based myocardial perfusion decision support is provided. Deep learning-based support potentially addresses limitations of current CT perfusion technology by reducing the necessary number of scan shots and images from a patient. Artificial intelligence (AI) provides perfusion information (e.g., dynamic myocardial perfusion) with the same performance as conventional perfusion testing and reduced radiation dose and scan time. Such an approach can result in dose and / or time reduction. For example, tissue perfusion (e.g., myocardial perfusion) is determined with reduced CT radiation dose and scan time because sparse perfusion measurements in time and / or space can be acquired compared to standard perfusion CT imaging, and the model can determine missing information and / or perfusion results without complete sampling.

[0020] An example application used herein is myocardial CT perfusion examination. AI-enhanced perfusion examinations can be used in other modalities besides CT, such as in MR. AI-enhanced perfusion examinations can be used for non-cardiac perfusion imaging, such as neural perfusion, or perfusion studies of the liver or other organs or tissues.

[0021] Figure 1 is a flow chart of one embodiment of a method for decision support based on perfusion measured by a medical imager. In one embodiment, the method is for estimating myocardial perfusion. In other embodiments, the method provides other perfusion study information, such as for diagnosis or treatment recommendations in an end-to-end manner. A machine-learned model is used to generate estimates or information, thereby allowing perfusion study decision support using shorter scan times and / or less dose.

[0022] The medical scanner performs these actions. Alternatively or additionally, a workstation, server, or other computer performs one or more of these actions. In one embodiment, Figure 3 The system performs these actions. For example, the CT scanner performs action 10. The image processor performs actions 12, 14, and 16. Other devices may perform any of these actions or be used in the performance of any of these actions.

[0023] The actions are performed in the order shown (e.g., from top to bottom or in numerical order) or in other orders. For example, actions 10 and 12 are performed in any order.

[0024] Additional, different, or fewer actions may be provided. For example, the method may be performed without actions 12 and / or 16. As another example, an action for configuring a medical scanner is provided. In still another example, an action for using the generated output is provided.

[0025] In action 10, a sequence or data set of medical images is acquired. A medical image is a frame of data representing a patient. The data may be in any format. Although the terms "image" and "imaging" are used, the image or imaging data may be in a format before the image is actually displayed. For example, a medical image may be a plurality of scalar values ​​representing different positions in a Cartesian or polar coordinate format that is different from the display format (i.e., scan or voxel data). As another example, a medical image may be a plurality of red, green, blue (e.g., RGB) values ​​output to a display for generating an image in a display format. The medical image may not yet be a displayed image, may be a currently displayed image, or may be a previously displayed image in a display or other format. An image or imaging is a data set that can be used for anatomical imaging, such as scan data representing the spatial distribution of a patient's anatomical structure (e.g., coronary arteries). Medical data, such as image data or other data from a scan, is acquired.

[0026] The medical data is obtained by loading from a memory and / or transmitting via a computer network. For example, previously acquired data is accessed from a memory or a database. As another example, after being acquired from a scanned patient, the scanned data is transmitted over a network. In other embodiments, the medical image or scanned data is obtained by scanning a patient.

[0027] Any type of medical data can be used. In one embodiment, a CT scanner or system acquires CT data representing a patient and a contrast agent in the patient. The CT data is acquired by rotating an x-ray source and a relative detector around the patient. Any range of travel and / or path can be used, such as rotating along a spiral travel path. During the scan, several scan shots are acquired. Each scan shot is a cross-sectional image acquisition each time the CT scanner x-ray tube is turned on. In order to collect iodine distribution information at different time points during the first path of the contrast agent, multiple scan shots (e.g., 10) are acquired.

[0028] A C-arm or another x-ray imaging may alternatively be used.Computed tomography is then used to determine the two-dimensional or three-dimensional distribution of the x-ray attenuation from the projections detected by the detector.

[0029] Any number of scan shots or readings from the detector may be used. To obtain a sequence of CT data representing the patient and contrast agent at different times, the scan is repeated. Each scan repetition performed to obtain data at a given time is a scan shot.

[0030] In other embodiments, other types of scan data or medical images are obtained, such as magnetic resonance, x-ray, ultrasound, positron emission tomography (PET), single photon emission computed tomography (SPECT), or other medical imaging modalities.

[0031] In one embodiment, a CT scan is performed to acquire coronary or other CT data representing the heart or coronary artery region of the patient. Other coronary CT data may be acquired, such as dual energy or photon counting data.

[0032] Based on the received detected intensities, a three-dimensional representation of the patient and contrast agent (i.e., density or absorption in terms of voxels or locations) is generated by computed tomography processing. Alternatively, the scan data represents a two-dimensional cross-section of the patient. Data representing an internal region of the patient is obtained. Some of the data from each scan shot is a response to the contrast agent in the patient. The data frame represents a two-dimensional or three-dimensional region of the patient and contrast agent. Values ​​are provided for each of a plurality of locations distributed in two or three dimensions. By repetition, a sequence of data frames or images is generated to show perfusion or contrast agent at different times. CT perfusion sequence images are obtained.

[0033] Although any number of scan shots and / or doses can be used to acquire a sequence of CT images representing perfusion, a reduced number may be desirable. Rather than using a standard, current, or higher number of scan shots or doses, lower spatial and / or temporal sampling is performed. For example, less than 10, 8, 5, or other number of scan shots are used over a time period of 5, 10, 20, 30, or other seconds. Rather than using 10 to 15 scan shots and corresponding images, fewer scan shots and corresponding images are used. Given any standard or common number of scan shots for a protocol, fewer scan shots can be used. The same or even greater number of scan shots can be used with less spatial sampling. For example, the field of view is divided into multiple portions, and different groups of scan shots are performed on different portions, rather than performing complete spatial sampling on each scan shot. As another example, a reduced number of slices are scanned.

[0034] The temporal and / or spatial sparsity allowed by the use of AI-based examinations allows for a reduced CT dose index. Rather than the typical CT dose index of 200 or more for myocardial perfusion, the dose index is 100, 80, 70, 50 or other values. For any standard, guideline or scanning protocol in perfusion imaging, new technologies (e.g., detector or tube current modulation) can reduce the radiation dose. AI can be used to reduce further dose reductions compared to the standard, guideline or scanning protocol that does not use machine learning-based perfusion information, such as a 50% to 80% reduction. The effective dose for CT perfusion varies depending on the specific acquisition protocol, such as 2.5 to 21mSv, with an average radiation exposure of 12mSv. A combined dynamic rest and stress CT perfusion study can use a radiation exposure of about 18mSv. Using AI, only 9mSv can be used with a dose of, for example, 100mGy, instead of 18mSv, which represents a 50% reduction in effective dose. Given an acquisition protocol and common doses, using AI to predict a certain perfusion may result in a 40% to 60% or other reduction in effective dose. By using a lower radiation dose without reducing the intensity of the x-rays, the patient is exposed to less radiation. Because the machine-learned model can output information even with sparse sampling, the reduction is provided without a corresponding reduction in diagnostic value.

[0035] Other information may be collected. For example, injection protocol parameters may be obtained. The type, rate, volume, injection location, or other characteristics of the contrast medium injection may be obtained, such as from the medical record or radiology database for the patient. Information regarding contrast medium and agent injections used for the patient may be obtained.

[0036] In action 12, in addition to the coronary artery CT data, the image processor also obtains information specific to the patient. The patient-specific information is obtained by mining, loading, inputting, or another source. For example, the patient-specific information is loaded from a computerized medical record for the patient, and / or from soliciting input from a user.

[0037] By obtaining patient-specific information, personalized information can be used.Such personalization enables perfusion examination information to be estimated or generated based on similarities to other patients.

[0038] Various types of patient-specific information may be obtained. For example, patient attributes are obtained for the patient being scanned. Size, age, gender, cardiac output, other clinical measurements, other characteristics of the patient's body, medical history, family history, and / or genetics are obtained as patient attributes. Non-image information is obtained.

[0039] As another example, an angiographic image of a patient is acquired. A CT scanner may acquire the angiographic image, or the image may come from another source (e.g., a stored image taken at a different time and / or with a different modality). The angiographic image itself may be used as input to the machine-learned model. Alternatively, features, such as anatomical and plaque-related features, are extracted from the angiographic image. The extracted features are then input.

[0040] In yet another example, arterial enhancement curve information is obtained for a patient. A CT scanner or other imager (e.g., ultrasound) can scan one or more arteries. Rate, volume, volume flow, velocity, changes over time, and / or other information from arterial enhancement obtained by an injection bolus of contrast agent in the patient is determined. The arterial enhancement curve information can indicate restrictions, diseases, and / or other patient-specific information that affects perfusion.

[0041] As another example, static myocardial perfusion measurements are obtained for a patient. CT data over time show dynamic perfusion. To scan at rest, CT angiography images can be used to obtain CT perfusion. For stress, a CT scan is used. Perfusion at a given time (such as a set time after the arrival of contrast agent) or at peak perfusion can be used to determine the amount of static perfusion. For example, a peak value by position or average, time to peak, or other measurement or perfusion quantification for a data set is determined based on the CT data sequence. Static perfusion can be from peak enhancement. Static measurements can be used as patient-specific information to input into a machine-learned model.

[0042] Other sources of patient-specific information may be used. For example, a physics or other anatomical model may be adapted for use with scanned data representing the patient's anatomical structure. The personalized or adapted model is then used to model the operation of the anatomical structure, such as to calculate flow characteristics in the heart or blood vessels. Functional and / or physiological models may be used. Calculated values ​​from the modeling (such as verified estimates based on physics) may be used as input to a machine-learned model.

[0043] In act 14, the image processor generates perfusion examination information. CT data representing contrast agent in the patient over time, with or without other information (e.g., patient-specific information and / or values ​​for injection protocol parameters), is input to a machine-learned model, which outputs perfusion examination information in response to the input. In the case where deep learning is used to train the machine-learned model, the image or scan data is directly input. In other embodiments, one or more features are derived from the image or scan data and the input.

[0044] Machine-learned models are clinical decision support systems. Machine learning uses training data of hundreds or thousands of input samples, labeled with one or more ground truth outputs for each sample. The training data is used as knowledge of past cases to train a classifier to generate outputs for unseen cases. Training associates features of the input vectors with clinical decisions or other outputs in perfusion examinations.

[0045] Any machine learning or training can be used. Probability boosting trees, support vector machines, neural networks, sparse autoencoding classifiers, Bayesian networks, regression models, or other machine learning now known or developed later can be used. Any semi-supervised learning, supervised learning, or unsupervised learning can be used. Hierarchical methods or other methods can be used. In one embodiment, the classification is a machine-learned classifier that has been learned by using deep learning. As part of the learning features that distinguish different results, the classifier is also machine trained to output based on the learned features.

[0046] In one embodiment, deep learning is to train a machine-learned model to estimate myocardial perfusion or other perfusion examination information. For example, each sample input vector includes a CT perfusion sequence image, patient-specific information, and one or more values ​​for one or more injection protocol parameters. The ground truth provided for each sample in the training data includes: a perfusion parameter image, a color map image, a quantitative value (such as peak value, time to peak, cerebral blood flow (CBF) and / or cerebral blood volume (CBV)), and / or a cardiologist or radiologist's decision (e.g., diagnosis and / or treatment).

[0047] Any deep learning method or architecture can be used. For example, a convolutional neural network is used. The network can include convolutional layers, downsampling (e.g., max pooling) layers, fully connected layers, flexible maximum transfer functions (softmax), and / or other types of layers. By using convolutions, the number of possible features to be tested is limited. The fully connected layer operates to fully connect the features limited by the convolutional layer after max pooling. Other features can be added to the fully connected layer, such as non-imaging information or clinical information. Any combination of layers can be provided. A hierarchical structure is used for learning features or representations, or for classification or regression. Computer-based decision support systems use machine learning algorithms to make automated decisions.

[0048] In one embodiment, a recurrent neural network (RNN) is used. Since the perfusion data is sequential or temporal, an RNN model with a long short-term memory (LSTM) architecture can be used. This particular model can use prior information to predict enhancement or concentration values.

[0049] In another embodiment, a generative adversarial network (GAN) is used, such as an image-to-image or UNet generator and a CNN discriminator. Due to the sequential nature of the perfusion information from sequential slices, a deep learning-based GAN is trained to generate images from a smaller number of actual perfusion images compared to conventional CT perfusion. For example, spatial and / or temporal sparsity is provided in the scan. Instead of acquiring 10 scan shots for 10 real slices, a machine-learned model is trained by deep learning to generate perfusion for 10 slices based on input of only 4 temporal and / or spatial slices. The "synthetic" images produced by such a generation method may or may not be used for visual analysis. For example, in a favorable embodiment, the synthetic images generated by the trained GAN are used only for the purpose of image analysis and subsequent quantification.

[0050] A machine-learned model, either with or without deep learning, is trained to associate a class label (outputting perfusion examination information) with an extracted value of one or more features from the input. Machine learning uses training data together with ground truth to learn to make predictions based on the input vector. The resulting machine-learned model defines how to extract feature values ​​from the input and / or other features and output perfusion examination information.

[0051] In addition to CT data representing contrast agent in a patient's body over time, the input feature vector may also include personalized or patient-specific information. To ensure that any information displayed in the actual CT perfusion measurement is not compromised for a new patient being examined, one or more values ​​of the patient-specific information are provided in the training data and used in the application of the machine-learned model. In application, the machine-learned model is generated in response to the input of the CT data and the patient-specific or personalized information. One or more values ​​of one or more corresponding injection protocol parameters may be used in training and application, in conjunction with or in place of the patient-specific information.

[0052] Due to the predictive nature of the machine-learned model, the machine-learned model can operate on sparse CT data that exhibit perfusion. For example, the output is estimated from coronary CT data input from no more than five scans. As another example, the output is estimated from CT data corresponding to no more than 100 CT dose indices.

[0053] The machine-learned model outputs any perfusion study information. The model provides individual measurements or information from a conventional CT perfusion study. The machine-learned model can estimate myocardial perfusion as a parametric image. Generate a parametric perfusion image. Provide quantities such as data point values ​​(e.g., intensity return from contrast agent) or perfusion curve values ​​of perfusion over time for each location. A grayscale or color-mapped image representing the value of a parameter (e.g., time to peak) as a function of location is generated by the machine-learned model.

[0054] The machine-learned model can estimate myocardial perfusion as one or more quantities. For example, output peak value, time to peak, blood flow, blood volume, mean transit time, or a combination thereof for a location, region, or field of perfusion. Other quantitative values ​​may be used.

[0055] The machine-learned model can generate a diagnosis and / or treatment as perfusion exam information. The image processor uses the model to determine the diagnosis or treatment. The model is trained to generate a final diagnosis and / or decision as a "task-to-task" approach. Alternatively, the model is used to estimate myocardial perfusion, and the image processor applies a threshold or lookup from the perfusion information to determine the diagnosis or treatment.

[0056] The diagnosis can be binary, such as whether the patient has a given disease. The diagnosis can be the type of disease and / or the level or extent of the disease, such as the level of tissue death. The treatment can be the type of treatment, such as a recommendation for revascularization, stenting, bypass, or drug use. The treatment can be a setting for a given type, such as recommending a specific stent or type of stent. Machine learning relates inputs to output diagnoses and / or treatments. Any cardiologist or radiologist decision can be output.

[0057] The same or different machine-learned models can output multiple types of perfusion examination information. For example, output perfusion images and / or quantification and clinical decisions (e.g., treatment and / or diagnosis). Machine-learned models generate various types of information as outputs in an "end-to-end" evaluation. Figure 2 An example is shown. A perfusion image sequence (e.g., CT data of contrast agent in a patient over time) 20, injection protocol information 21, patient-specific characteristics 22, CT angiography images 23 or values ​​extracted therefrom, arterial enhancement curve information 24, and / or static perfusion information 25 are input to a machine-learned model 26. In response, the machine-learned model 26 outputs one or more perfusion images 27, perfusion magnitude values ​​28, and / or decision support (e.g., diagnosis or treatment) 29.

[0058] In act 16, the perfusion examination information is transmitted. The transmission is to a display such as a monitor, a workstation, a printer, a handheld device, or a computer. Alternatively or additionally, the transmission is to a memory (such as a database of patient records), or a network (such as a computer network). Other outputs of the perfusion examination information may be provided.

[0059] This transmission provides information for physician decision making. The clinical decision support system can provide an initial or preliminary recommendation, which can be approved or disapproved by the attending physician, and / or the clinical decision support system can provide perfusion information (e.g., perfusion images and / or volumes) to aid in diagnosis or treatment. The radiologist can use the perfusion exam information to determine what measurements to perform or how long any measurement will take.

[0060] In one embodiment, perfusion examination information such as myocardial perfusion is visualized as text or graphically (e.g., overlaid on a medical image) and presented to a clinician. One or more perfusion images generated by a machine-learned model may be displayed. Various quantities for perfusion may be displayed. Decision support information such as therapy, risk, guidelines, or other treatment or diagnostic information may be output.

[0061] The machine-learned model is implemented locally. One advantage of the machine-learned model is that online predictions are fast, with output results being almost instantaneous, or within seconds given access to the input information. Thus, the machine-learned model can be run directly on a medical scanner or workstation located at a clinic or hospital. The system can be run on a portable device. Alternatively, the decision support system can be available as a service and / or operated remotely. The hospital or clinician accesses a server or other remote processor to obtain the perfusion exam information output.

[0062] Figure 3 A system for decision support based on perfusion in medical imaging is shown. The system implements Figure 1 methods, Figure 2 Another method of arranging or outputting perfusion examination information. Given CT data that is sparse relative to full sampling or standard perfusion imaging without a machine-learned model, a machine-learned model 36 is used to generate an output.

[0063] The system includes a CT scanner 30, an image processor 32, a memory 34, a display 38, and one or more machine-learned models 36. Additional, different, or fewer components may be provided. For example, a network or network connection is provided, such as for networking with a medical imaging network or data archive system, or networking between the CT scanner 30 and the image processor 32. In another example, user input is provided. As another example, a server for implementing the image processor 32 and / or the machine-learned model 36 remotely from the CT scanner 30 is provided.

[0064] The image processor 32, memory 34, display 38, and / or machine-learned model 36 are part of the CT scanner 30. Alternatively, the image processor 32, memory 34, display 38, and / or machine-learned model 36 are part of an archive and / or image processing system separate from the CT scanner 30, such as associated with a medical record database workstation or server. In other embodiments, the image processor 32, memory 34, display 38, and machine-learned model 36 are a personal computer such as a desktop or laptop computer, a workstation, a server, or a combination thereof.

[0065] The CT scanner 30 is a medical diagnostic imaging CT system. The gantry supports the x-ray source and detector on opposite sides of the patient examination space. The gantry moves the source and detector around the patient to perform a CT perfusion scan. In a scanning shot, the detector acquires various x-ray projections from different positions relative to the patient. Computed tomography resolves the two-dimensional or three-dimensional distribution of the response from the projection. Ultrasound, x-ray, fluoroscopy, positron emission tomography, single photon emission computed tomography, and / or magnetic resonance systems may be used in addition or alternatively.

[0066] The CT scanner 30 is configured to perform a scan to achieve a perfusion study. The scan is performed while the contrast agent is in the patient. The CT scanner 30 is configured to output CT data for the patient as a sequence of images or data frames representing perfusion or contrast agent over time. In order to reduce radiation dose, the CT scanner 30 can be configured to sample perfusion sparsely, such as using a sequence of a total of six or fewer scans to image the contrast agent during a given study or injection. The patient can be subjected to a CT dose index of 100 or less. A larger dose and / or number of scans can be used.

[0067] The memory 34 may be graphics processing memory, video random access memory, random access memory, system memory, cache memory, a hard drive, optical media, magnetic media, a flash drive, a buffer, a database, a combination thereof, or other memory devices now known or later developed for storing data. The memory 34 may be part of the CT scanner 30, part of a computer associated with the image processor 32, part of a database, part of another system, picture archive storage, or a separate device.

[0068] The memory 34 is configured to store data, such as stored in a computerized patient record. Any data discussed herein may be stored, such as CT data, patient-specific information, and / or injection protocol parameter values. The memory 34 alternatively or additionally stores one or more machine-learned models 36. The memory 34 alternatively or additionally stores data during processing, such as storing information discussed herein or links related thereto.

[0069] The memory 34 or other memory is alternatively or additionally a non-transitory computer-readable storage medium storing data, the data representing instructions executable by a programmed image processor 32 or a processor implementing a clinical decision support and / or machine-learned model 36. Instructions for implementing the processes, methods and / or techniques discussed herein are provided on a non-transitory computer-readable storage medium or memory such as a cache, a buffer, a RAM, a removable medium, a hard drive or other computer-readable storage medium. Non-transitory computer-readable storage media include various types of volatile and non-volatile storage media. In response to one or more instruction sets stored in or on a computer-readable storage medium, the functions, actions or tasks illustrated in the figure or described herein are performed. The functions, actions or tasks are independent of a specific type of instruction set, storage medium, processor or processing strategy, and can be performed by software, hardware, integrated circuits, firmware, microcode, etc. operating alone or in combination. Similarly, processing strategies can include multi-processing, multi-tasking, parallel processing, etc.

[0070] In one embodiment, the instructions are stored on a removable media device for reading by a local or remote system. In other embodiments, the instructions are stored in a remote location for transmission over a computer network or on a telephone line. In still other embodiments, the instructions are stored within a given computer, CPU, GPU, or system.

[0071] The image processor 32 is a general purpose processor, a central processing unit, a control processor, a graphics processor, a digital signal processor, a three-dimensional rendering processor, an application specific integrated circuit, a field programmable gate array, a digital circuit, an analog circuit, a combination thereof, or other devices now known or later developed for applying a clinical decision predictor. The image processor 32 is a single device, or multiple devices operating in series, in parallel, or separately. The image processor 32 can be the main processor of a computer such as a laptop or desktop computer, or can be a processor for handling some tasks in a larger system such as in a CT scanner 30. The image processor 32 is configured by instructions, design, hardware, and / or software to perform the actions discussed herein.

[0072] The image processor 32 is configured to generate perfusion examination information, such as estimated perfusion information. In response to inputting the CT data, with or without patient-specific information and / or injection protocol parameter values, into the machine-learned model 36, the perfusion examination information is output.

[0073] The machine-learned model 36 is implemented by the image processor 32, or other processor that has access to the model 36 stored in the memory 34 or other memory. The machine-learned model 36 is defined as input channels (i.e., input vectors), weights, relationships between weighted inputs or other layers, connections, filter kernels, and output channels.

[0074] Using the machine-learned model 36, the image processor 32 is configured to estimate perfusion information by inputting CT data, patient-specific information, and / or injection protocol information for injecting contrast agent into the patient. The patient-specific information can be patient attributes, angiographic images, arterial enhancement curves, and / or quantification from CT data. The output perfusion information can be a perfusion image, a quantitative perfusion value, a diagnosis, and / or a treatment. In response to the input of CT data representing contrast agent based on six or fewer scan shots and / or a CT dose index of 100 or less, an output is generated.

[0075] Display 38 is a monitor, LCD, projector, plasma display, CRT, printer, or other now known or later developed device for outputting visual information. Display 38 receives graphic images, text, volume, spatial distribution of anatomical structure or perfusion, or other information from image processor 32, memory 34, CT scanner 30, or machine-learned model 36.

[0076] One or more images are displayed. The images may or may not include anatomical representations or imaging, such as anatomical images from CT data. The images include perfusion examination information, such as perfusion images, volumes, and / or clinical decisions. The images include indications, such as text, graphics, or coloring, of patient classification for perfusion studies or examinations.

[0077] Although the present invention has been described above with reference to various embodiments, it will be appreciated that many changes and modifications may be made without departing from the scope of the present invention. Therefore, it is intended that the foregoing detailed description be considered illustrative rather than restrictive, and it will be appreciated that the appended claims, including all equivalents, are intended to define the spirit and scope of the present invention.

Claims

1. A method for estimating myocardial perfusion by a medical imager, the method comprising: scanning a patient with a computed tomography CT system using one or more injection protocol parameters that provide an effective dose corresponding to no more than 100 CT dose indices, the scan providing sparse coronary artery CT data representing contrast agent in a heart of the patient over time; In addition to the coronary artery CT data, information specific to the patient is also acquired; estimating myocardial perfusion by using a machine-learned model trained by deep learning using training data including sequential CT images, relevant injection protocol parameters, and relevant patient-specific information, the machine-learned model outputting myocardial perfusion in response to inputs of the coronary CT data, one or more injection protocols, and the information; as well as The myocardial perfusion is shown.

2. The method according to claim 1, wherein: Scanning includes scanning in fewer than five shots, and wherein estimating includes estimating in response to input of coronary artery CT data from the fewer than five shots.

3. The method according to claim 1, wherein: Scanning includes scanning at an effective dose that is less than 1 / 2 of a dose used for scanning given a scanning protocol for perfusion, and wherein estimating includes estimating in response to input of the coronary artery CT data corresponding to the effective dose.

4. The method according to claim 1, wherein: Obtaining information includes obtaining patient attributes for the patient.

5. The method according to claim 1, wherein: Acquiring information includes acquiring a computed tomography angiography image of the patient.

6. The method according to claim 1, wherein: Acquiring information includes acquiring arterial enhancement curve information for the patient.

7. The method according to claim 1, wherein: Acquiring information includes acquiring static myocardial measurements for the patient.

8. The method according to claim 1, wherein: Estimating includes estimating said myocardial perfusion as a parametric image.

9. The method according to claim 1, wherein: Estimating includes estimating the myocardial perfusion as peak value, time to peak value, blood flow, blood volume, mean transit time, or a combination thereof.

10. The method of claim 1, further comprising determining, by an image processor, a diagnosis or treatment based on the estimated myocardial perfusion, and wherein, Displaying includes displaying the diagnosis or treatment.

11. The method according to claim 1, wherein: The estimation includes estimating using a machine-learned model, wherein the machine-learned model includes a convolutional neural network, an adversarially trained generative network, or a recurrent neural network.

12. A system for decision support based on perfusion in medical imaging, the system comprising: a computed tomography (CT) scanner for scanning a patient while a contrast agent is in the patient using one or more injection protocol parameters that provide an effective dose corresponding to no more than 100 CT dose indices, the CT scanner being configured to output sparse CT data for the patient; a memory configured to store a machine-learned model trained using training data including CT images and related patient-specific information; an image processor configured to estimate perfusion information provided by no more than six or fewer sparse CT data scan shots representing the contrast agent by inputting the CT data, one or more injection protocol parameters, and patient-specific information into the machine-learned model; and A display is configured to display the perfusion information.

13. The system according to claim 12, wherein: The CT scanner is configured to scan the patient with an effective dose that is less than 1 / 2 of a dose used for scanning given a scanning protocol for perfusion of the six or fewer scans captured, and wherein the image processor is configured to estimate the perfusion information based on no more than the effective dose.

14. The system according to claim 12, wherein: The patient-specific information includes patient attributes, angiographic images, arterial enhancement curves, and / or quantifications from the CT data.

15. The system of claim 12, wherein: The perfusion information includes perfusion images, quantitative perfusion values, diagnosis, and / or treatment.

16. A method for decision support based on perfusion measured by a medical imager, the method comprising: scanning the patient with a computed tomography (CT) system, the scan providing CT data representative of contrast agent in the patient over time, the CT data being acquired at a CT dose index of 80 or less; generating perfusion examination information via a machine-learned model applied to the CT data; as well as The perfusion inspection information is output.

17. The method according to claim 16, wherein: Generating includes generating a diagnosis and / or a treatment as the perfusion examination information.

18. The method according to claim 16, wherein: Generating includes generating by the machine-learned model, wherein the machine-learned model includes a deep-learned neural network applied to the CT data and personalized patient information.

Images