SYSTEM AND METHOD FOR REDUCED CONTRAST AND / OR X-RAY ANGIOGRAPHY USING MACHINE LEARNING

Abstract

Methods, systems, and computer readable media for reduced contrast and / or x-ray angiography using machine learning (e.g., deep learning) are provided. Briefly, the techniques described herein can use neural networks trained to preserve / maintain angiographic image quality while reducing the amount of potentially harmful chemical contrast and / or x-ray radiation during angiography. As a result, angiographic anatomy can be extracted from images using deep learning neural networks in reduced contrast and / or x-ray angiography. Reduction of chemical contrast and x-ray dose can be achieved based on operations performed before, during, and / or after angiography imaging.

Description

[Technical field]

[0001] [CROSS REFERENCE TO RELATED APPLICATIONS] This application claims priority to U.S. Provisional Application No. 63 / 324,867, filed March 29, 2022, the contents of which are incorporated herein in their entirety.

[0002] The present disclosure relates generally to angiography, and more specifically to systems and methods for performing angiography with reduced contrast and / or x-rays using machine learning. [Background technology]

[0003] The heart pumps blood through the body in a series of pumps through the arteries. Blood travels through capillaries to the veins and back to the heart. The presence and movement of blood within blood vessels (generally branching tubular structures) can be dynamically imaged using a technique called angiography.

[0004] In fluoroscopic angiography, a chemical contrast agent is injected as a bolus into the vascular system (e.g., bloodstream) and a series of x-ray images are acquired. The chemical contrast agent may comprise one or more chemical substances in a liquid. The chemical contrast agent is denser than blood or tissue, and therefore attenuates the passage of x-rays more than blood or tissue. The denser the chemical contrast agent, the greater the image contrast imparted to the blood vessels in which it is contained during fluoroscopic angiography angiography. Some particularly dense chemical contrast agents contain iodine as ions, and such agents are referred to as "iodinated contrast agents." One example of an iodinated contrast agent is iohexol. Other chemical contrast agents do not contain iodine and are not iodized.

[0005] The chemical contrast agent passes through the vasculature and blocks the passage of X-rays at a given frame rate, forming a mold of the vascular structure that contains the contrast agent. The resulting spatiotemporal X-ray attenuation pattern creates a sequence of X-ray images that are obtained fluoroscopically on the X-ray sensor. This sequence is called an angiogram, which is, for example, a series of images that track the passage of a bolus of contrast agent. Angiographic images are typically two-dimensional in space and one-dimensional in time.

[0006] Injection of chemical contrast agents has toxic side effects to internal organs such as the kidneys. However, reducing the dose of contrast agent to reduce the risk of such toxic side effects may produce unsatisfactory images with poor signal-to-noise ratios, resulting in incomplete angiographic examinations in which the vascular anatomy is not adequately imaged. Also, when using less contrast agent, it may be necessary to advance the injection catheter further into the ductus arteriosus to concentrate the injected contrast agent within the anatomically relevant region. The need to advance the injection catheter further increases the risk of complications caused by the catheter injuring small vessels distal to the vascular tree. Summary of the Invention

[0007] An embodiment of the present invention is a method, system, and computer readable medium for angiography with reduced contrast and / or x-ray using machine learning (e.g., deep learning). Briefly, the techniques described herein can use machine learning to preserve / maintain angiographic image quality while reducing the amount of potentially harmful chemical contrast and / or x-ray radiation during angiography. As a result, angiographic anatomy can be extracted from images with reduced contrast and / or x-ray. Reduction of chemical contrast and x-ray can be achieved based on operations performed before, during and / or after angiography imaging.

[0008] Various mechanisms are provided for training machine learning models (e.g., deep learning neural networks) to generate higher quality angiographic images with reduced chemical contrast and / or x-ray radiation. These mechanisms may include collecting training data acquired with reduced chemical contrast and / or x-ray radiation ("reduced training data") interspersed with segmented standard full volume data ("full volume segmented data"). That is, a training data set may be generated using full volume segmented data acquired with full (standard) amounts of chemical contrast and x-ray radiation as known outputs and unsegmented reduced training data, and a neural network may be trained on the full volume segmented data. A training data set may be generated using full (standard) and physically reduced amounts in animal (e.g., non-human) angiograms, full (standard) and physically reduced amounts in realistic artificial organ models, and / or computer-simulated reduction amounts in full volume human angiogram data. Thus, in the neural network training step, multiple angiographic images acquired at full (standard) volume over a time interval may be employed to segment the vessels, and the neural network (e.g., a convolutional network) may obtain standard volume segmentation results using reduced volume image data. Thus, the neural network may be trained on training data having ground truth data of high quality angiographic structures and real or simulated degraded angiographic images taken with reduced chemical contrast agents and / or x-ray radiation.

[0009] In a post-training deployment mode of the neural network, the neural network can take angiographic data obtained with reduced chemical contrast agents and / or X-ray radiation and generate a segmentation of the image based on training against full dose segmentation. In one example, the low-quality angiographic image data from physically reduced chemical contrast agents and / or X-ray radiation may be fed into a deep neural network to estimate vascular structure according to training of the low-quality angiographic images against higher quality ground truth training data. In this way, the deep learning neural network can generate a segmentation of blood vessels for images obtained with physically reduced chemical contrast agents and / or X-ray radiation.

[0010] Neural networks can be applied to spans or sequences of angiographic images to identify vascular structures without the need for augmenting angiographic contrast and / or x-rays. For example, a deep learning neural network can infer vascular structures within an angiographic target image using an angiographic target image and one or more temporally adjacent or consecutive angiographic images. In this manner, the neural network can be tuned to detect spatiotemporal characteristics of blood vessels containing contrast agents.

[0011] Alternative neural network configurations may be designed to optimize offline and / or real-time computations. For example, the neural network may be modified for real-time use where multiple angiogram images for data input include multiple temporally preceding angiogram images to generate a segmentation of the most recent angiogram image.

[0012] A method according to a first aspect of the invention includes providing an angiographic target image obtained with a first amount of chemical contrast agent and / or x-ray radiation as an input to a machine learning model via a processor, the machine learning model being trained with (a) as input a second angiographic image obtained with a second amount of chemical contrast agent and / or x-ray radiation, and (b) as known output a third angiographic image obtained with a third amount of chemical contrast agent and / or x-ray radiation, the third amount being greater than the first and second amounts. The method further includes obtaining an output from the machine learning model via the processor, the angiographic image being a processed version of the angiographic target image. Processing a first angiographic image obtained with a first amount of chemical contrast agent and / or x-ray radiation with a machine learning model trained using a second angiographic image obtained with a second amount of chemical contrast agent and / or x-ray radiation as training data, and a third angiographic image obtained with a third amount of chemical contrast agent and / or x-ray radiation (wherein the third amount is greater than the first amount) as known output, can enable obtaining diagnostically useful angiographic images from angiographic images obtained with reduced chemical contrast agent and / or x-ray radiation, thereby improving the safety of angiographic examinations.

[0013] In one embodiment, the target image is one of a plurality of angiographic images in a first subsequence of angiographic images obtained using a first amount of chemical contrast agent and / or X-ray radiation and provided as input to the machine learning model, and the second angiographic image is one of a plurality of angiographic images in a second subsequence of angiographic images obtained using a second amount of chemical contrast agent and / or X-ray radiation and used to train the machine learning model. By providing the subsequence of angiographic images including the target image obtained using the first amount of chemical contrast agent and / or X-ray radiation to a machine learning model trained using the subsequence of angiographic images obtained using the second amount of chemical contrast agent and / or X-ray radiation, the embodiment can use spatiotemporal information from other images in the subsequence of images including the target image to improve the accuracy of the output.

[0014] In one embodiment, the first subsequence of angiographic images is provided to the machine learning model as a single vector. By providing the first subsequence of angiographic images to the machine learning model as a single vector, this embodiment may reduce processing time.

[0015] In one embodiment, the first and second subsequences of angiographic images each consist of odd-numbered angiographic images. Using odd-numbered angiographic images in each subsequence may improve the accuracy of the output because the target image is taken from the center of the first subsequence of angiographic images and may provide spatiotemporal information from before and after the target image.

[0016] In one embodiment, the first subsequence of angiographic images is composed of the same number of angiographic images as the second subsequence of angiographic images. By ensuring that the input in the unfolding mode matches the input in the training mode, this embodiment can reduce the time required to train the machine learning model.

[0017] In one embodiment, the first and second subsequences each consist of five angiographic images, with the advantage that the target image is located in the center of the subsequence and spatiotemporal information before and after the target image is available, and this embodiment has been found to provide diagnostically useful output with less processing than subsequences with more than five images.

[0018] In one embodiment, the first subsequence of angiographic images is one of a plurality of input subsequences, each of the plurality of input subsequences being extracted from a different portion of the sequence of angiographic images obtained using a first amount of chemical contrast agent and / or x-ray radiation. The plurality of input subsequences are provided as inputs to the machine learning model, which, via at least one processor, obtains an output including angiographic images that are processed versions of respective target images in each of the plurality of input subsequences. This embodiment allows the machine learning model to be used to process multiple target images in an image sequence, which may provide a diagnostic advantage.

[0019] In one embodiment, the second subsequence of angiographic images is one of a plurality of subsequences used to train the machine learning model, each of the plurality of subsequences being extracted from a different portion of the sequence of angiographic images obtained with a second amount of chemical contrast agent and / or x-ray radiation. This embodiment allows the machine learning model to be trained to process multiple target images in the image sequence, which may be diagnostically advantageous.

[0020] In one embodiment, each angiographic image in a first subsequence of angiographic images is temporally contiguous with at least one other angiographic image in the first subsequence, and each angiographic image in a second subsequence of angiographic images is temporally contiguous with at least one other angiographic image in the second subsequence. The use of temporally contiguous images can improve accuracy of the output by providing spatiotemporal information from the images to the target image.

[0021] In an embodiment, the first and second amounts of chemical contrast agent and / or X-ray radiation are the same. In this embodiment, the accuracy of the output may be improved.

[0022] In one embodiment, the first and second amounts are less than the amount necessary to obtain a diagnostically useful image from an x-ray imaging device, and the third amount is sufficient to obtain a diagnostically useful image from an x-ray imaging device. This embodiment may improve the safety of angiographic examinations by allowing angiographic images to be obtained with reduced amounts of chemical contrast agent and / or x-ray radiation.

[0023] In one embodiment, the angiographic images of the second subsequence are angiographic images of a non-human vasculature. This embodiment can reduce the training cost of the machine learning model.

[0024] In one embodiment, the angiographic images of the second subsequence are angiographic images of an artificial vascular structure. This embodiment can reduce the training cost of the machine learning model.

[0025] In an embodiment, the second amount of chemical contrast agent and / or x-ray radiation is greater than the first amount of chemical contrast agent and / or x-ray radiation, and the angiographic images in the second subsequence are angiographic images modified to simulate having been obtained using a lesser amount of chemical contrast agent and / or x-ray radiation. This embodiment allows for more control over the quality of the training data and can reduce the cost of training the machine learning model.

[0026] In an embodiment, the second amount of chemical contrast agent and / or x-ray radiation is greater than the first amount of chemical contrast agent and / or x-ray radiation, and the second angiographic image is modified to simulate being obtained using a lesser amount of chemical contrast agent and / or x-ray radiation. This embodiment allows for more control over the quality of the training data and can reduce the cost of training the machine learning model.

[0027] In one embodiment, the second angiographic image is modified by adding randomly generated noise to at least some pixels of the second angiographic image, which allows for more control over the quality of the training data and can reduce the training cost of the machine learning model.

[0028] In one embodiment, randomly generated noise is added only to pixels determined to correspond to blood vessels in the second angiographic image, which helps to reduce the training cost of the machine learning model and avoid registration errors caused by the movement of organs in the angiographic field before and after the chemical contrast agent is administered.

[0029] In one embodiment, the third angiographic image is a segmented angiographic image, and the output from the machine learning model is an angiographic image that is a segmented version of the first angiographic image. This embodiment may eliminate the need for a separate manual, semi-automated, or automated segmentation step, thereby reducing processing time.

[0030] In one embodiment, a processed version of the first angiogram image may be displayed on a display, whereby a medical professional may view and analyze the output to provide a medical opinion, such as a diagnosis.

[0031] Further aspects include systems and computer program products utilizing techniques substantially similar to those described above.

[0032] Other objects and advantages of these techniques will become apparent from the specification and drawings. [Brief description of the drawings]

[0033] [Figure 1A] FIG. 1 is a side view illustrating an example of a rotational X-ray system that may be used with embodiments of the present disclosure to obtain angiography data. [Figure 1B] FIG. 1 is a partial schematic diagram illustrating an example of a rotational X-ray system that may be used with embodiments of the present disclosure to obtain angiography data. [Diagram 2] FIG. 1 is a schematic diagram of a computer system or information processing device that may be used with embodiments of the present disclosure. [Figure 3A] 1 shows a flowchart of a method for training neural networks based on respective training data, according to an example embodiment. [Figure 3B] 1 shows a flowchart of a method for training neural networks based on respective training data, according to an example embodiment. [Figure 3C] 1 shows a flowchart of a method for training neural networks based on respective training data, according to an example embodiment. [Figure 4] 1 illustrates a system configured to generate images of vasculature simulated as obtained with reduced chemical contrast and / or X-ray radiation, according to an example embodiment. [Diagram 5] 1 illustrates an exemplary method for segmenting vasculature objects in a single image from angiographic images obtained with reduced chemical contrast and / or X-ray radiation, according to an example embodiment. [Figure 6] We demonstrate the use of machine learning in accordance with example embodiments to generate higher quality angiographic images from angiographic target images located at different positions within a subsequence of angiographic images. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0034] Fluoroscopic angiography using chemical contrast is a commonly employed medical imaging technique. Briefly, a physician percutaneously accesses a blood vessel (usually the femoral artery), guides a catheter to a root artery of the organ of interest, and injects a chemical contrast agent in time-coordinated communication with fluoroscopic x-ray image sequences.

[0035] Fluoroscopic angiography can be used to diagnose blockages in the coronary arteries of the heart and brain. Diagnosis of arterial blockages through angiography may pave the way for treatment, for example by placing a stent in the calcified plaque in the coronary arteries of the heart or applying a thrombolytic drug directly to a blood clot in a cerebral artery.

[0036] Angiographic imaging studies may contain foreground objects such as blood vessels. The greater the sharpness, detail, and clarity with which the foreground objects in an imaging study are displayed against the background, the greater the signal-to-noise ratio of the imaging study, and therefore the greater its diagnostic value. In some cases, however, fluoroscopic angiography may not provide sufficient sharpness of vascular structures in the angiographic image, even with the injection of chemical contrast agents in an attempt to improve the sharpness of the vessels in the anatomical region being fluoroscopically viewed.

[0037] In standard clinical practice, image quality in angiographic examinations can be improved by increasing the dose of injected chemical contrast agent and / or increasing the dose of fluoroscopic x-ray radiation. However, the use of chemical contrast agents and x-ray radiation in angiographic imaging involves risks. For example, the procedure of placing a catheter into a blood vessel and navigating the catheter to the target organ to inject the chemical contrast agent involves risks. Furthermore, chemical contrast agents and x-ray radiation can be harmful (e.g., cause toxic side effects) to human and animal subjects.

[0038] In general, chemical contrast agents may exhibit a certain degree of toxicity when injected into the vascular circulation. Once injected, chemical contrast agents are eliminated by biochemical and physiological processes through excretory organs such as the kidneys and liver. Chemical contrast agents may be toxic to excretory organs at certain doses.

[0039] Chemical contrast agents also impose a significant mass load within the vasculature, causing stress on vascular structures such as the heart and potentially inducing or exacerbating heart failure in patients with reduced pumping ability of the heart.Contrast agents may also impose stress on the excretory organs.

[0040] Additionally, some humans develop immune reactions to certain molecular structures contained in various chemical contrast agents, particularly iodinated contrast agents. Histiocyte-mediated immune reactions to injected contrast agents can quickly become severe and can be fatal if not recognized and treated promptly. For at least these reasons, increasing the amount of chemical contrast agent used to obtain better quality angiograms also increases the risk of harm to the subject.

[0041] Furthermore, X-ray doses can be harmful to irradiated tissues, especially those that are particularly sensitive to radiation, such as the thyroid gland and reproductive organs. Depending on the radiation dose, it can induce chronic inflammation or damage the constituent biomolecules of tissues, with consequences ranging from relatively minor skin inflammation along the X-ray path to the formation of cancer in the irradiated tissue. In particular, high doses of X-rays can directly damage the biomolecules of tissues, including the deoxyribonucleic acid (DNA) of the cell nucleus organelle, causing organ malformations and posing a risk of tumor formation.

[0042] Therefore, techniques are described herein for reducing the amount of chemical contrast agent and / or X-ray radiation given during angiography procedures. In particular, angiography image quality is preserved by clearly and accurately displaying the imaged vasculature in the foreground, even with reduced amounts of both chemical contrast agent and X-ray radiation. As described in more detail below, this can be achieved using machine learning (e.g., deep learning) techniques.

[0043] 1A, 1B, and 2 illustrate exemplary systems or devices that may be employed to implement embodiments of the present invention. It should be understood that such systems and devices are merely illustrative of representative systems and devices, and that other hardware and software configurations may be used with embodiments of the present invention. Thus, it is not intended that the embodiments be limited to the particular systems and devices illustrated herein, and it should be recognized that other suitable systems and devices may be employed without departing from the spirit and scope of the subject matter provided herein.

[0044] 1A and 1B, a rotating X-ray system 28 is illustrated that may be employed to obtain angiograms in fluoroscopic angiography. In obtaining an angiogram, a chemical contrast agent is injected into a patient positioned between an X-ray source and a detector, and the X-ray projections are captured by the X-ray detector as two-dimensional images (i.e., angiographic images or image frames). A sequence of such image frames constitutes an angiographic examination. In an embodiment, the sequence of angiographic image frames may be acquired at a rate faster than the subject's heart rate. For example, the subject's heart rate may be measured (e.g., using an electrocardiogram machine, etc.), and the sequence of angiographic image frames may be acquired at a rate faster than the measured heart rate.

[0045] As shown in FIG. 1A, one example of an angiographic imaging system is shown in the form of a rotating X-ray system 28 including a base having a C-arm 30 carrying an X-ray source assembly 32 at one end and an X-ray detector array assembly 34 at the other end. The base allows the X-ray source assembly 32 and the X-ray detector array assembly 34 to be oriented at various positions and angles around a patient positioned on a table 36 while providing a physician with access to the patient. The base includes a pedestal 38 having a horizontal leg 40 extending below the table 36 and a vertical leg 42 extending upwardly from an end of the horizontal leg 40 spaced apart from the table 36. A support arm 44 is rotatably secured to an upper end of the vertical leg 42 for rotation about a horizontal pivot axis 46.

[0046] The horizontal pivot axis 46 coincides with the centerline of the table 36, and the support arm 44 extends radially outward from the horizontal pivot axis 46 and supports at its outer end a C-arm drive assembly 47. The C-arm 30 is slidably fixed to the C-arm drive assembly 47 and is connected to a drive motor (not shown) that slides the C-arm 30 to rotate about a C-axis 48 as indicated by arrow 50. The horizontal pivot axis 46 and the C-axis 48 intersect at a system isocenter 56 located above the table 36 and are perpendicular to each other.

[0047] An x-ray source assembly 32 is mounted at one end of the C-arm 30 and an x-ray detector array assembly 34 is mounted at the other end. The x-ray source assembly 32 emits a beam of x-rays, which is directed at the x-ray detector array assembly 34. Both assemblies 32, 34 extend radially inwardly towards a horizontal pivot axis 46 such that a central ray of the beam passes through a system isocenter 56. Thus, the central ray of the beam can be rotated about either the horizontal pivot axis 46, the C-axis 48, or both, about the system isocenter during acquisition of x-ray attenuation data from a subject positioned on the table 36.

[0048] The x-ray source assembly 32 includes an x-ray source that emits a beam of x-rays when energized. The central ray passes through a system isocenter 56 and impinges on a two-dimensional flat panel digital detector 58 housed in an x-ray detector array assembly 34. The two-dimensional flat panel digital detector 58 may be, for example, a two-dimensional array of 2048 by 2048 detector elements. Each element generates an electrical signal representative of the intensity of the impinging x-rays and therefore the attenuation of the x-rays as they pass through the patient. During a scan, the x-ray source assembly 32 and detector array assembly 34 are rotated about the system isocenter 56 to acquire x-ray attenuation projection data from various angles. In some embodiments, the detector array is capable of acquiring 50 projections or image frames per second, which is the limiting factor that determines the number of image frames that can be acquired for a given scan path and speed.

[0049] 1B, rotation of assemblies 32, 34 and operation of the x-ray source are governed by an x-ray system control mechanism 60. Control mechanism 60 includes an x-ray controller 62 that provides power and timing signals to x-ray source assembly 32. A data acquisition system (DAS) 64 within control mechanism 60 samples data from the detector elements and passes the data to an image reconstructor (or module) 65. Image reconstructor 65 receives digitized x-ray data from DAS 64 and performs high speed image reconstruction according to the methods of the present disclosure. The reconstructed image is used as an input to computer 66, which stores the image in mass storage device 69 or further processes the image. Image reconstructor 65 may be a stand-alone computer or may be integrated with computer 66.

[0050] The control mechanism 60 also includes a base motor controller 67 and a C-axis motor controller 68. In response to motion commands from the computer 66, the motor controllers 67, 68 provide power to motors within the x-ray system that cause rotation about the horizontal pivot axis 46 and the C-axis 48, respectively. The computer 66 also receives commands and scan parameters from an operator via a console 70 having a keyboard and other manually operable controls. An associated display 72 allows the operator to observe reconstructed image frames and other data from the computer 66. The operator supplied commands are used by the computer 66 under the direction of a stored program to provide control signals and information to the DAS 64, the x-ray controller 62, and the motor controllers 67, 68. The computer 66 further operates a table motor controller 74 which controls the motorized table 36 to position the patient relative to the system isocenter 56.

[0051] Referring now to FIG. 2, there is shown a block diagram of a computer system or information processing device 80 (e.g., image reconstructor 65 and / or computer 66 of FIG. 1B) that may be incorporated into an angiographic imaging system, such as rotating X-ray system 28 of FIG. 1A and FIG. 1B, used as a stand-alone device for angiographic contrast agent and / or X-ray reduction by deep learning according to an embodiment of the present invention. Information processing device 80 may be local or remote to rotating X-ray system 28. In one example, the functions performed by information processing device 80 may be provided as a Software-as-a-Service (SaaS) option. SaaS refers to a software application that is stored on one or more remote servers (e.g., in the cloud) and provides one or more services (e.g., angiographic image processing) to a remote user. Angiographic images may be obtained directly from an angiographic imaging system, such as rotating X-ray system 28 of FIG. 1A and FIG. 1B, or may be obtained from other sources, such as a physical storage medium configured to store data.

[0052] In one embodiment, computer system 80 includes a monitor or display 82, a computer system 84 (including a processor 86, a bus subsystem 88, a memory subsystem 90, and a disk subsystem 92), a user output device 94, a user input device 96, and a communication interface 98. Monitor 82 may include hardware and / or software elements configured to generate a visual representation or display of information. Some examples of monitor 82 may include well-known display devices such as a television monitor, a cathode ray tube (CRT), a liquid crystal display (LCD), a light emitting diode (LED) display, etc. In some embodiments, monitor 82 may provide an input interface, such as incorporating touch screen technology.

[0053] The computer system 84 may include well-known computer components such as one or more central processing units (CPUs), memory or storage devices, graphics processing units (GPUs), communication systems, interface cards, and the like. As shown in FIG. 2, the computer system 84 may include one or more processors 86 that communicate with a number of peripheral devices via a bus subsystem 88. The processor 86 may include a commercially available central processing unit or the like. The bus subsystem 88 may include mechanisms for allowing the various components and subsystems of the computer system 84 to communicate with each other as intended. Although the bus subsystem 88 is shown diagrammatically as a single bus, alternative embodiments of the bus subsystem may utilize multiple bus subsystems. The peripheral devices in communication with the processor 86 may include a memory subsystem 90, a disk subsystem 92, a user output device 94, a user input device 96, a communication interface 98, and the like.

[0054] Processor 86 may be implemented using one or more analog and / or digital electrical or electronic components, and may include a microprocessor, a microcontroller, an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), programmable logic, and / or other analog and / or digital circuit elements configured to perform various functions described herein, such as by executing instructions stored in memory subsystem 90 and / or disk subsystem 92 or another computer program product.

[0055] The memory subsystem 90 and the disk subsystem 92 are examples of physical storage media configured to store data such as instructions executable by the one or more processors 86 to perform the operations described herein. The memory subsystem 90 may include a number of memories or memory devices including random access memory (RAM) that volatilely stores program code, instructions, and data during program execution, and read-only memory (ROM) where fixed program code, instructions, and data are stored. The disk subsystem 92 may include a number of file storage systems that provide persistent (non-volatile) storage for programs and data. Other types of physical storage media include floppy disks, removable hard disks, compact disk read-only memory (CD-ROM), digital video disks (DVD), optical storage media such as bar codes, semiconductor memories such as flash memory, read-only memory (ROM), battery-backed volatile memory, networked storage devices, and the like. The memory subsystem 90 and the disk subsystem 92 may be configured to store programming and data configurations that provide the functionality or features of the technology described herein. Software code modules and / or processor instructions that implement or provide functionality when executed by processor 86 may be stored in memory subsystem 90 and disk subsystem 92. Memory subsystem 90 may be a non-transitory computer-readable storage medium.

[0056] User input devices 96 may include hardware and / or software elements configured to receive input from a user for processing by components of computer system 80. User input devices may include all possible types of devices and mechanisms for inputting information to computer system 84. These may include keyboards, keypads, touch screens, touch interfaces integrated into displays, audio input devices such as microphones and voice recognition systems, and / or other types of input devices. In various embodiments, user input devices 96 may include computer mice, trackballs, trackpads, joysticks, wireless remotes, drawing tablets, voice command systems, eye tracking systems, and the like. In some embodiments, user input devices 96 are configured to allow a user to select or interact with objects, icons, text, and the like, which may be displayed on monitor 82 via commands, actions, or gestures, such as, for example, clicking a button.

[0057] User output devices 94 may include hardware and / or software elements configured to output information to a user from components of computer system 80. User output devices may include all possible types of devices and mechanisms for outputting information from computer system 84. These may include displays (e.g., monitor 82), printers, touch or force feedback devices, audio output devices, etc.

[0058] Communications interface 98 may include hardware and / or software elements configured to provide one-way or two-way communications with other devices. For example, communications interface 98 may provide an interface between computer system 84 and other communications networks and devices, such as via an Internet connection.

[0059] The techniques described herein may allow for a reduction in the amount of angiography (e.g., the dose of chemical contrast agent and / or the dose of X-ray radiation) required to obtain a diagnostically useful angiographic image. In particular, the amount of angiographic contrast agent and / or X-rays may be reduced compared to the amount of standard / conventional angiographic contrast agent and / or X-rays that would be required to obtain a diagnostically useful angiographic image in the absence of the techniques described herein (e.g., in the absence of the machine learning models such as the deep learning neural networks described herein). In one example, a "diagnostically useful" or "high quality" angiographic image is one that provides data of sufficient quality to provide meaningful clinical information and / or to allow treatment decisions to be made. For example, a diagnostically useful angiographic image is one that has sufficient clarity to allow a medical professional to visually identify and segment blood vessels in the image. For example, if one has chest pain due to insufficient blood flow in the coronary arteries that supply blood to the heart muscle, a diagnostically useful angiographic image of the coronary arteries is one that accurately displays the segments of the coronary arteries of the heart that have stenosis that impairs circulation to the heart muscle.

[0060] The x-ray dose required to produce a diagnostically useful image in angiography also varies with the physical characteristics of the patient / subject and the nature of the angiographic procedure. Methods for calculating x-ray doses are well known in the art. Typically, the total / standard dose of x-ray exposure for interventional cardiac procedures in a fluoroscopy machine ranges from 8 to 10 millisieverts of radiation. A sievert corresponds to one joule of energy per kilogram of mass. Due to the ionizing nature of x-rays, it is always desirable to minimize the x-ray exposure of the subject (and the medical staff involved in the angiographic procedure) while still providing useful visualization of the target tissue.

[0061] The main type of contrast agent used in angiography is the family of iodine contrast agents, which are ionic or, preferably, non-ionic iodine contrast agents. Such agents are well known in the art and include: iohexol (Omnipaque™, GE Healthcare), (iopromide (Ultravist™, Bayer Healthcare), iodixanol (Visipaque™, GE Healthcare), ioxaglate (Hexabrix™, Mallinckrodt Imaging), iothalamic acid (Cysto-Conray II™, Mallinckrodt Imaging) and iopamidol (Osbu (Etem, Bracco Imaging). See also, Lusick and Granstaff, "X-Ray Computed Tomography Contrast Agents," Chemical Reviews 13, 1641-1666 (2013). Other agents include gadolinium-based agents; see Ohse et al., "'Gadolinium' as an Alternative to Iodinated Contrast Medium for X-Ray Angiography in Severely Allergic Patients," Circulation Journal 2005, 69, 507-509 (2005).

[0062] The total (standard) amount of such contrast agents varies depending on the nature of the agent, the physical characteristics of the patient / subject, and the nature of the angiography procedure. In general, however, the total (standard) amount of contrast agent should improve visualization of the target tissue by increasing the absolute computerized tomography (CT) attenuation difference between the target tissue and the surrounding tissues and fluids. In fluoroscopic angiography, the injected chemical contrast agent typically increases the vascular CT attenuation value by 2-10 times the baseline level without the chemical contrast agent. In fluoroscopic angiography, the injection catheter is guided close to the target organ, resulting in a high local amount but a low systemic amount. In CT with intravenous injection of contrast agent, the contrast agent is diluted throughout the vascular system. Therefore, the local concentration of contrast agent in a given vessel may be low. The increase in CT attenuation value due to intravenous injection of such chemical contrast agents is typically 1.2-4 times the baseline CT attenuation value without the chemical contrast agent. It has been found that a CT attenuation value of at least 1.2-fold increase from baseline for CT angiography and at least 2-fold increase from baseline for fluoroscopic angiography results in a "diagnostically useful" or "high quality" angiographic image. The imaging medium should contain a high mole percent of X-ray attenuating atoms per agent (molecule, macromolecule, or particle) to reduce the dosage and concentration required for imaging. The tissue retention time of the contrast agent should also be long enough (e.g., 2-4 hours) to complete the CT scan and schedule instrument time in the diagnostic setting. In addition, the contrast agent should preferably: (a) localize or target the tissue of interest and have a favorable biodistribution and pharmacokinetic profile; (b) dissolve readily or form a stable suspension in aqueous physiological conditions (appropriate pH and osmolarity) with low viscosity; (c) be nontoxic; and (d) be cleared from the body in a reasonably short time, usually within a few hours (less than 24 hours).

[0063] In some embodiments, the amount of iodine angiographic contrast agent used to obtain diagnostically useful angiographic images may be reduced by about 25% from the total amount of contrast agent typically injected (standard amount). Further reductions may be achieved by combining the techniques described herein with the spatiotemporal reconstruction techniques described in U.S. Patent Application No. 16 / 784,073, filed February 6, 2020, which is incorporated herein by reference in its entirety. The spatiotemporal reconstruction of the image may be input to a machine learning model, or the output of the machine learning model may be processed using the spatiotemporal reconstruction.

[0064] To improve the sharpness and clarity of vasculature imaged in reduced chemical contrast and x-ray exposure, a deep learning neural network (e.g., having an input layer, an output layer, and three or more layers between the input layer and the output layer) is provided that has properties that promote the performance of detecting vasculature in reduced chemical contrast and / or x-ray. Although the description herein focuses on deep learning neural networks, it should be understood that these techniques may be utilized with any suitable machine learning model, and deep learning neural networks are merely an example. The machine learning model may be implemented by any suitable machine learning technique (e.g., mathematical / statistical, classifier, feed-forward, recurrent, convolutional, or other neural network, etc.). For example, a neural network including an input layer, one or more intermediate layers (e.g., including any hidden layer), and an output layer may be used. Each layer may include one or more nodes or neurons, and the neurons of the input layer may receive inputs (e.g., image data, image feature vectors, etc.) and associate them with weight values. For example, each node in the input layer may receive or be encoded with the relative brightness of one pixel of an angiogram image as input, where the relative brightness may be a floating point number between 0 and 1. Neurons in the hidden and output layers are connected to one or more neurons in the preceding layer and receive as input the output of the neurons to which they are connected. Each connection is associated with a weight value, and each neuron produces an output based on the combination of weights of the inputs to that neuron. The output of a neuron may further be based on bias values ​​for a particular type of neural net (e.g., a recurrent type of neural net). The weight (and bias) values ​​may be adjusted based on various training techniques.For example, machine learning of a neural network may be performed using a training set of contrast and / or x-ray reduced angiographic images as inputs and corresponding full-volume angiographic images of high quality as known outputs, where the neural network attempts to generate the known outputs by adjusting weight (and bias) values ​​(e.g., with training techniques such as backpropagation) using errors from the outputs generated by the neural network (e.g., the difference between the generated output and the known output). In an exemplary embodiment, the known output at each node may be a pixel value (e.g., a floating point number between 0 and 1) representing brightness or intensity, or a probability (p-value) that the pixel is a vessel (e.g., a floating point number between 0 and 1). A neural network trained using the latter technique (i.e., segmented angiographic images where the known outputs have p-values ​​between 0 and 1 for each pixel) may be advantageously used to generate segmented angiographic images from contrast and / or x-ray reduced angiographic images.

[0065] More specifically, in the training phase, a data conditioning system may be provided for training the deep learning neural network to perform well in a reduced chemical contrast and / or x-ray radiation setting, such that the training phase may enable the deep learning neural network in the deployment phase to output high quality angiographic images based on angiographic images obtained with reduced chemical contrast and / or x-ray radiation.

[0066] In an embodiment, a deep learning network data training system is provided that facilitates the ability of a convolutional spatio-temporal network to detect pixels corresponding to blood vessels even with reduced chemical contrast agents and / or fluoroscopic x-ray radiation. The deep learning neural network can be trained with angiographic training data. The angiographic training data can include (1) a first set of angiographic images obtained with conventional / standard amounts of chemical contrast agents and x-ray radiation, and (2) a second set of comparative angiographic images with reduced chemical contrast agents and / or x-ray radiation. In an embodiment, the comparative angiographic images with reduced chemical contrast agents and / or x-ray radiation can be used as inputs to the training set, and the angiographic images obtained with conventional / standard amounts of chemical contrast agents and x-ray radiation can be used as known outputs. In an embodiment, a feature vector can be extracted from the images and used as inputs along with the corresponding known outputs of the training set. The feature vector can include any suitable features (e.g., pixel intensity, etc.).

[0067] In an embodiment, the second set of comparative angiographic images may be identical or nearly identical to the corresponding images in the first set, except for differences in the amount of angiographic contrast and / or x-rays. For example, the same object may be imaged in both angiographic images and may be of substantially the same size, position, and orientation in both images. The first set of angiographic images and the second set of corresponding images are similar enough to provide useful training data for a deep learning neural network, e.g., the data may be useful for training a deep learning neural network to generate diagnostically useful angiographic images based on diagnostically unuseful angiographic images obtained in angiograms with reduced contrast and / or x-rays.

[0068] The training data can be obtained based on any suitable angiographic training images. For example, the training data can be obtained (1) in a laboratory setting using animals undergoing approved angiographic procedures, (2) in a laboratory setting using a physical model with fluid mechanically pumped into a synthetic organ to be angiographically imaged, and / or (3) from full-quality human clinical angiographic data that has been computationally modified to simulate reduced chemical contrast agents and / or x-rays. The data training system can use these training options alone or in any suitable combination to train the deep learning neural network.

[0069] FIG. 3A illustrates a flowchart of an exemplary method 100 for training a deep learning neural network based on vasculature, such as a non-human vasculature (e.g., training data option (1) training data obtained in a laboratory setting using animals undergoing approved angiography). In this embodiment, in step 102, a first set of angiographic images of one or more non-human vasculature are imaged using standard / conventional amounts of chemical contrast (e.g., for coronary arteries, approximately 10 ml of 300 mg of iodine / ml) and x-ray radiation (e.g., for angiography, approximately 400 dGy×cm). 2 In step 104, a second set of angiographic images of one or more non-human vascular structures is acquired with reduced chemical contrast (e.g., chemical contrast is ¼ the amount of standard iodine contrast) and x-ray radiation (e.g., x-ray radiation is ½ the amount of standard radiation). Then, in step 106, a deep learning neural network is trained based on the first and second sets.

[0070] In one embodiment, angiograms may be obtained using laboratory animals with conventional chemical contrast and x-ray radiation doses to provide a "gold standard" reference set of angiographic images, after which angiographic studies can be obtained with reduced chemical contrast and / or x-ray doses without repositioning the animal or the fluoroscopic imaging platform.

[0071] Angiographic images obtained based on conventional contrast agents and / or x-ray doses may be manually employed by a human technician to segment blood vessels. These segmentations may serve as a guide for segmenting blood vessels that are less visible in stress angiograms that are similar or identical in all respects except that they were obtained with reduced chemical contrast agents and x-rays. As used herein, the term "segmentation" may refer to the manual, semi-automatic, or fully automated identification and representation (e.g., display) of blood vessel elements in an angiographic image. "Segmenting" blood vessels may refer to processing an image such that blood vessel structures are represented as distinct from noise and other structures in the imaging field of view. For example, blood vessel structures in a "segmented" angiographic image may be represented as white objects against a black background. Angiographic images that are subject to segmentation may be referred to herein as "segmented angiographic images." In an embodiment, segmented angiographic images obtained with conventional chemical contrast agents and x-ray radiation doses may be used as training data (e.g., known output) for angiographic images obtained with reduced contrast agents and / or x-rays.

[0072] FIG. 3B shows a flowchart of an exemplary method 200 for training a deep learning neural network based on an artificial vascular structure (e.g., training option (2) training data obtained in a laboratory environment using a physical model having fluid mechanically pumped into the artificial organ to be angiographically imaged). In this example, in step 202, a first set of angiographic images of one or more artificial vascular structures is obtained with standard / conventional amounts of chemical contrast agent and x-ray radiation. In step 204, a second set of angiographic images of one or more artificial vascular structures is obtained with reduced amounts of chemical contrast agent and x-ray radiation. Then, in step 206, a deep learning neural network is trained based on the first and second sets.

[0073] The artificial vascular structure may be an artificial solid vascular organ phantom or organoid that includes a mechanical fluid pump to simulate pulsatile arterial flow. Fluids can be pumped through the artificial vascular structure in a network of hollow tubular structures that share the size, shape, and branching pattern of blood vessels. One manufacturer of artificial vascular structures is Heartroid® from JMC Corporation (Yokohama, Japan). The artificial vascular structure is suitable for producing images with systematic, standard, and reduced chemical contrast agents and X-ray radiation.

[0074] In one specific example, the organoid is placed in an angiography machine, a fluid containing a chemical contrast agent is injected, and fluoroscopic X-ray images are acquired to generate a series of angiography images of the chemical contrast agent flowing through the vascular channels of the organoid. A reduced interaction range of the chemical contrast agent and X-ray radiation can then be employed to obtain a large number of angiography images with varying amounts of chemical contrast agent and X-ray without changing the position of the organoid or the position of the imaging platform. The angiography images obtained with a large amount of contrast agent and / or X-ray serve as training data for the angiography images obtained with a small amount of contrast agent and / or X-ray. Thus, the neural network can be fed with the data with reduced chemical contrast agent and X-ray radiation and trained with the vascular data obtained from angiography with a large amount of contrast agent and / or X-ray.

[0075] Images obtained with large amounts of contrast agents and / or X-rays can be improved by artificial editing or mathematical processing, for example using the techniques discussed in AF Frangi, WJ Niessen, KL Vincken, and MA Viergever, "Multiscale vessel enhancement filtering," Lecture Notes in Computer Science (including sub-series Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 1496, p. 130-137, 1998, which is incorporated herein by reference in its entirety.

[0076] FIG. 3C shows a flowchart of an exemplary method 300 for training a deep learning network based on human vasculature (e.g., training option (3) training data obtained from full quality human clinical angiography data computationally modified to simulate reduced chemical contrast and / or x-rays). In this example, in step 302, a first set of angiography images of one or more human vasculature are obtained with standard amounts of chemical contrast and x-ray radiation. In step 304, a second set of angiography images of one or more human vasculature are obtained simulating reduced chemical contrast and / or x-ray radiation. Then, in step 306, a deep learning neural network is trained based on the first and second sets.

[0077] In one embodiment, anonymous human angiography data can be obtained with standard chemical contrast agents and X-ray doses currently used to generate high-quality angiography images. Vascular structures are extracted from these images by artificial editing or mathematical vascular structure filters. The angiography images can then be computationally modified to simulate reduced chemical contrast agents and reduced X-ray radiation, for example, based on the computational methods described in M. Elhamiasl and J. Nuyts, "Low-dose X-ray CT simulation from an available higher-dose scan." Phys Med Biol, vol. 65, no. 13, p. 135010, 2020 07 08, which is incorporated herein by reference in its entirety.

[0078] A simulation of reduced x-ray radiation can be generated by adding a mixture of Poisson and Gaussian noise to the data collected by the x-ray detector. The more noise added, the lower the simulated x-ray dose. To simulate reduced chemical contrast, the angiographic image is characterized by a histogram of pixel values. If the angiographic image is treated as normalized, the greater the attenuation of x-rays by the imaged tissue, the brighter it will be. Then, when a bolus of contrast passes through the imaged vasculature, more pixels will have bright values ​​due to the contrast in the imaged blood vessels, and the pixel histogram will shift to the right. Thus, reduced chemical contrast can be simulated by adding Poisson and Gaussian noise to the pixels that the bolus of contrast passes through and shifts to the right on the pixel histogram. For example, a pre-contrast image of an angiographic field can be obtained without a chemical contrast agent and with a full dose of X-ray radiation (e.g., before a chemical contrast agent is administered), and a post-contrast image of the same angiographic field can be obtained with a full dose of chemical contrast agent and X-ray radiation (e.g., after administration of the chemical contrast agent). For corresponding pixels in the two images, a brightness or intensity difference between the two corresponding pixels in the images can be determined (e.g., by subtracting the values ​​of the respective pixels). The brightness difference between the two corresponding pixels can be stored in a first matrix (e.g., as a floating point number between 0 and 1) and a second matrix of random numbers (e.g., floating point numbers between 0 and 1 randomly generated using a Poisson or Gaussian distribution) can be generated with a size corresponding to the first matrix. For each non-zero cell in the first matrix, the random number in the corresponding cell of the second matrix can be subtracted from the non-zero cell. The results are stored in a third matrix, which can be added to the pixel values ​​of the pre-contrast image to obtain an angiographic image with simulated reduced contrast and / or x-rays.

[0079] To avoid registration errors due to the movement of blood vessels in the angiographic field before and after the administration of chemical contrast agents, pixels representing blood vessels may be identified and a large amount of random noise may be added to these pixels to simulate reduced chemical contrast agents. In particular, the noise may have a negative mean value to simulate a drop in signal due to reduced contrast agents and a high standard deviation to simulate a drop in signal noise. Using DICOM pixel values ​​of 0 to 32,000 as an example (instead of floating point numbers between 0 and 1), the average pixel value of the image before the chemical contrast agent arrives is taken to be a first number (e.g., 50±10). When the chemical contrast agent arrives, some pixels become brighter representing blood vessels containing the chemical contrast agent. Thus, the average pixel value increases overall (e.g., to 60±15). In such a case, pixel values ​​greater than the pre-contrast mean value plus two standard deviations (e.g., 50+20) may be assumed to be blood vessel pixels (i.e., pixels representing blood vessels). Therefore, to simulate a reduced chemical contrast agent, a greater amount of random noise may be added, especially to the blood vessel pixels.

[0080] Any suitable data preparation operations may be applied to computationally transform the simulated contrast agent and / or X-ray reduction. In one example, one or more data augmentation operations may be performed to reduce overfitting in the deep learning neural network. Examples of such data augmentation operations include arbitrary translation and rotation of the training data. In one example, the transformation of the simulated contrast agent and / or X-ray reduction may be incorporated into the data augmentation step.

[0081] The simulated reduced contrast and / or x-ray images, preferably with a worse signal-to-noise ratio than the unmodified images, can be provided to the neural network. The neural network can then use the vascular structures obtained from images taken with standard chemical contrast and x-ray doses as reference vascular data (i.e., known output) that the neural network is trained to detect and generate. Thus, the simulated reduced contrast and / or x-ray data can be used as input to train the deep neural network to generate angiographic images comparable to data of angiographic images obtained from full chemical contrast and x-ray doses. In an embodiment, the reference vascular data can include segmented angiographic images obtained with full chemical contrast and x-ray radiation doses to train the deep neural network to generate high quality segmented angiographic images from corresponding non-segmented angiographic images obtained or simulated to be obtained with reduced chemical contrast and / or x-ray radiation doses.

[0082] FIG. 4 illustrates an exemplary system 400 configured to generate images of simulated vascular structures with reduced chemical contrast and x-ray radiation, according to an exemplary embodiment. The exemplary system 400 includes a computer 402, a display 404 connected to the computer, a pointing device 406, such as a mouse or trackpad, connected to the computer, and a keyboard 408 connected to the computer. The computer 402 may include a processor in communication with a memory and / or other non-transient data storage device including instructions executable by the processor for segmenting angiographic images. The computer 402 may also store instructions executable by the processor for operating a neural network in a training mode and / or a deployment mode. The computer 402 may also have a communication interface configured to transmit and receive angiographic data over a communication network, such as a local area network or a wide area network. The system 400 may be utilized for multi-frame deep learning. In an embodiment, a human coronary angiogram is obtained. In this example, an angiogram of the human heart is employed, but these techniques may be applied to angiograms of other organs and animals. In the example of FIG. 4, the neural network may be operating in a training mode.

[0083] The angiographic images stored in the computer 402 can be displayed on a display 404 of the computer system 400. An exemplary coronary angiographic image 410 is displayed on the display 404. The computer system 400 can obtain the image 410 in any conventional manner. For example, the image 410 may be obtained directly from an angiographic imaging system, such as the system shown in FIGS. 1A and 1B, via a wired connection, a wireless connection, or a communication network. In another embodiment, the computer system 400 can obtain the image 410 from a remote source via a local area network, a wide area network, or other type of communication network. In yet another example, the computer system 400 can upload the image from a portable data storage device, such as a USB thumb drive, a DVD, or the like. In one embodiment, a human analyst inspects the exemplary image 410 and interacts with the system using a graphical user interface device, such as a mouse 406, to select (i.e., paint) pixels that represent blood vessels in the segmented coronary angiographic image 412. In one embodiment, the segmentation can be performed exclusively by the human analyst by painting on the angiographic image. In other examples, a mathematical algorithm or deep learning segmentation algorithm may make an initial guess at the segmentation. In yet other embodiments, the mathematical algorithm or deep learning segmentation algorithm may perform the painting automatically.

[0084] 4 further illustrates a simulated reduced coronary angiogram image 414, which simulates a reduced X-ray dose compared to the angiogram image 410. As described above, the angiogram image 410 may be used to generate the simulated reduced coronary angiogram image 414 with any suitable technique for simulating a reduced X-ray radiation dose. For example, a computer system 400 may be used to generate the simulated reduced coronary angiogram image. Once generated, the simulated reduced coronary angiogram image 414 may be employed for training a deep learning neural network by referencing a segmentation of the same angiogram image obtained with a full X-ray dose in a stored segmented coronary angiogram image 416 (which may be the same image as the segmented image 412).

[0085] 4, the angiogram image 410 and the segmented image 412 are displayed side-by-side, but in other embodiments, the painting may be performed on the same image on the display 404. Thus, the painting may be overlaid on the angiogram image 410 to produce the segmented image 412.

[0086] In training mode, the neural network may obtain as training data a plurality of temporally consecutive, adjacent images (e.g., 5 images) acquired at a standard volume and a plurality of corresponding images acquired (or simulated to be acquired) at a reduced volume. The images may be selected from a Digital Imaging and Communications in Medicine (DICOM) file, for example, containing a total of about 80 images. Each image may be 512 x 512 pixels in size, and each pixel may be represented in the DICOM format as an integer (e.g., from about 0 to about 16,000).

[0087] In an embodiment, at least one of a plurality of temporally consecutive images (e.g., five images) acquired with a standard dose may be used as a known output of the neural network, and at least one of a plurality of corresponding images acquired (or simulated to be acquired) with a reduced dose of contrast agent and / or x-rays may be used to encode an input layer of the neural network. The neural network may form connections between individual neurons based on training data, each connection having one or more associated weights (e.g., floating point numbers) to generate a known output from an input. In an embodiment, if the known output consists of a segmented angiogram image, the neural network may be trained to assign to each pixel in the output image a floating point number between 0 and 1 representing the probability that the pixel is part of a blood vessel. In another embodiment, the neural network may assign to each pixel a plurality of floating point numbers between 0 and 1 representing the probability that the pixel is part of a particular feature (e.g., a blood vessel, a catheter, and / or a branch point). For example, three floating point numbers between 0 and 1 may be assigned, where one floating point number represents the probability that the pixel is part of a blood vessel, another floating point number represents the probability that the pixel is part of a catheter, and another floating point number represents the probability that the pixel is part of a branch point (e.g., where one blood vessel branches into two or more vessels).

[0088] A checkpoint file storing weights for a neural network may be used in the deployment / prediction mode. For example, training with images obtained with standard contrast and / or x-ray angiographic doses may result in a first checkpoint file with one set of weights, and training with images obtained (or simulated to be obtained) with reduced contrast and / or x-ray angiographic doses may result in a second checkpoint file with a different set of weights. The first checkpoint file may be applied to an input of one or more images obtained with standard contrast and / or x-ray angiographic doses to obtain one or more segmented images, and the second checkpoint file may be applied to an input of one or more images obtained (or simulated to be obtained) with reduced contrast and / or x-ray angiographic doses to obtain one or more segmented images. The second checkpoint file may be applied in a clinical setting where reduced contrast and / or x-ray angiographic doses are administered to preserve angiographic image quality but avoid the side effects of high contrast and / or x-ray angiographic doses on the patient.

[0089] 5 illustrates an exemplary method 500 for segmenting vasculature objects in a single image from angiogram images acquired with reduced amounts of chemical contrast or x-rays. In this example, a convolutional network draws information from multiple angiogram images to perform the segmentation. In the example of FIG. 5, the neural network may be operating in an unfolded mode.

[0090] Analyzing (e.g., performing calculations on) an entire sequence of angiographic images may exceed the practical limits of computer memory and computation speed. Thus, as described in FIG. 5, the techniques described herein allow for the generation of high-quality images based on a subsequence of images within the constraints of the computer memory and computation speed. As used herein, the term "sequence" may refer to an entire set of angiographic images, e.g., images acquired fluoroscopically over the travel of a bolus of injected contrast agent. The term "subsequence" may refer to a subset of a series of images provided to a deep learning neural network system to estimate vascular structure. The angiographic images in a "subsequence" may preferably be consecutive in time or consecutive but separated by intervening images.

[0091] FIG. 5 shows an example of vessel segmentation of a single angiogram image from a subsequence of noisy angiogram images acquired with reduced chemical contrast or reduced x-ray radiation dose. In this example, the angiogram image to be segmented is drawn from a subsequence of five temporally adjacent angiogram images 502(a)-(e) drawn from physically reduced chemical contrast and x-ray angiogram images. In this example, the middle (third) image 502(c) in the subsequence of five angiogram images is the target image. The target image may be an image for which the neural network is to generate a higher quality version.

[0092] The subsequence of five angiographic images 502(a)-(e) may be combined and fed as a single input to a convolutional neural network 504. For example, if each image consists of a 512×512 vector of pixel values, the five angiographic images may be input or encoded into the neural network as a 5×512×512 vector of pixel values, and the neural network may be trained to generate only one high quality and / or segmented angiographic image corresponding to one of the five angiographic images (e.g., the middle image 502(c)).

[0093] The convolutional neural network 504 may be pre-trained with images in which noisy contrast and / or x-ray reduction is simulated for vessel segmentation extracted from a noise-free standard volume image (e.g., as described above in connection with FIG. 4). For example, a subsequence of low quality images including a target image may be obtained to train the neural network to generate only one high quality and / or segmented angiogram image from five temporally consecutive low quality images. The target image is preferably located in the center of the subsequence such that there are one or more images before the target image and one or more images after the target image. In this way, the neural network can be trained to use spatiotemporal information in providing an output. In a subsequence, the number of images before the target image may be different from the number of images after the target image. However, in a preferred embodiment, the target image is located in the center in the subsequence such that there are an equal number of images on either side of the target image. For example, in the case of five temporally consecutive images, the target image is preferably located between two temporally consecutive images on either side of the target image. In an embodiment, the neural network may be trained with multiple subsequences. For example, if there are 10 images available from the first angiogram and 12 from the second angiogram, the training set can be assembled with up to 14 unique subsequences, with each subsequence consisting of 5 temporally consecutive images. That is, 6 unique subsequences of 5 temporally consecutive images can be assembled from the first angiogram (because the 3rd through 8th images are each in the middle of the 5 temporally consecutive images), and 8 subsequences of 5 temporally consecutive images can be assembled from the second angiogram (because the 3rd through 10th images are each in the middle of the 5 temporally consecutive images). The first, second, penultimate, and last images in each series are not targets in this example, since they are not in the middle of the 5 temporally consecutive images.While the use of subsequences of five temporally consecutive images has been found to be advantageous in that it allows for the generation of high quality and / or segmented images with less processing and with relatively few images at either end of the sequence that are not themselves targets, current technology can also be adapted to use subsequences containing less than five temporally consecutive images or more than five temporally consecutive images. In general, increasing the number of images in a subsequence tends to increase the signal-to-noise ratio of the images generated by the neural network. At the same time, increasing the number of images in a subsequence increases the number of images that are dropped at either end.

[0094] While it is advantageous to use an odd number of temporally consecutive images as a subsequence for training a neural network to generate high quality and / or segmented images corresponding to a target image that is in the center of the subsequence (e.g., because it is based on an equal amount of spatiotemporal information before and after the target image), the technique can be adapted to generate high quality and / or segmented images corresponding to target images that are not in the center of the subsequence (e.g., the first, second, fourth, or fifth image of a subsequence of five images).

[0095] The technique can also be adapted to use subsequences of even images that are consecutive in time. The advantage of using subsequences of even images that are consecutive in time is that it provides more robustness with respect to the location of the target image and the size of the subsequence (e.g., allowing the first or last image of an image sequence to be targeted using only subsequences of only two images that are consecutive in time). It should also be appreciated that the technique can be adapted to use a combination of odd and even subsequences to generate high quality and / or segmented images corresponding to corresponding targets located anywhere in the sequence. In another embodiment, multiple neural networks can be deployed, each trained with a different number of input images, from one to as many as processing resources allow. Using multiple neural networks in this way allows the generation of high quality and / or segmented images from all images of an angiogram.

[0096] In an embodiment, the images in the training set may be subjected to a data augmentation step, such as translating and / or rotating the images, before being input to a neural network to train and generalize well. It should be appreciated that simulations with reduced chemical contrast agents and / or x-ray radiation may be incorporated as part of the data augmentation step.

[0097] Because the trained convolutional neural network 504 is operating on the subsequence of five noisy angiographic images 502(a)-(e), it can estimate a single segmented image 506 that has a greater signal-to-noise ratio than if the trained convolutional neural network were operating on only a single noisy angiographic image. The segmented image 506 may represent a deep learning neural network estimation of the vascular structure of the central (third) image in the subsequence of five noisy angiographic images 502(a)-(e).

[0098] More specifically, the neural network 504 may load floating point numbers from a checkpoint file generated during training mode and output a segmented image 506 from the five noisy angiogram images 5(a)-(e) based on the floating point numbers. The neural network may generate one or more of three images as output: one image based on the probability that the pixel is part of a blood vessel, another image based on the probability that the pixel is part of a catheter, and another image based on the probability that the pixel is part of a branch point. These three images may be stacked and overlaid to generate a segmented image 506 showing the blood vessels, catheters, and / or branch points in high quality, for example, as if the target image was obtained with an angiogram with standard contrast and / or x-ray.

[0099] In one example, the convolutional neural network 504 used for angiogram segmentation is based on the "U-net architecture," as described, for example, in O. Ronneberger, P. Fischer, and T. Brox, "U-net: Convolutional networks for biomedical image segmentation," in Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015, N. Navab, J. Hornegger, WM Wells, and AF Frangi, Eds. Cham: Springer International Publishing, 2015, pp. 234-241, which is incorporated herein by reference in its entirety.

[0100] In a specific example, the source code of the deep learning neural network may be written in Python language and generated as a U-net architecture in the Pytorch machine learning software library (see https: / / pytorch.org, which is incorporated herein by reference in its entirety). The U-net architecture may be three-dimensional in the sense that each image has two spatial dimensions, and the vasculature in an image may be estimated from both the image and an angiographic image that is adjacent in time. In a specific example, the Pytorch library for neural network machine learning may include a Python base class called nn.Module, as found at https: / / pytorch.org / docs / stable / generated / torch.nn.Module.html, which is incorporated herein by reference in its entirety. All neural network modules are composed of the subclass nn.Module. A Python class called UNet may inherit nn.Module. The Python class UNet may include a convolutional neural net structure with spatiotemporal properties. It is understood that the techniques described herein may be implemented using any suitable machine learning mechanism.

[0101] It should be understood that any suitable programming language, library, toolset, and / or other mechanism may be used to deploy a deep learning neural network in accordance with the examples provided herein. For example, the U-net structure may be extended to a full three-dimensional structure that simultaneously estimates vascular structures in multiple temporally adjacent angiographic images.

[0102] The convolutional neural network 504 may have an encoder-decoder structure. The convolutional neural network 504 may also include jump connections between equal-sized layers of the encoder and decoder. These jump connections may enable the output of a segmented image 506 of the same granularity as the angiogram subsequence input (e.g., a subsequence of five angiogram images 502(a)-(e)). The loss function used to train this architecture may be a linear combination of a classification loss function, cross-entropy, and Dice loss function for sharp boundary detection.

[0103] To prevent vascular motion between angiographic image frames from interfering with the angiographic data, the structure of the convolutional neural network 504 may have a high spatiotemporal convolution density. That is, an organ with high motion, such as the heart, may be imaged at 15 Hz and a neighborhood of five images may be used to determine the vasculature present in the center image. An organ with low motion, such as the brain, may be imaged at, for example, 6 Hz and a neighborhood of five images may be used to determine the vasculature present in the center image. In this way, the convolutional neural network 504 can account for both low motion (such as the brain, which is enclosed in a rigid container called the skull and thus limited in motion) and high motion (such as the heart, which is a muscular organ that beats constantly to pump blood into the arterial system).

[0104] In the deep learning network training mode (as opposed to the deployment mode illustrated in FIG. 5), the roles of data sources (inputs) and products (outputs) may be changed. For example, image 506 may represent a ground truth representation of vasculature structures, such as obtained from an angiogram with full volume chemical contrast and x-ray radiation, and one or more images of subsequences 502(a)-(e) may be obtained from empirically contrast and / or x-ray reduced studies from animals or physical organoid model angiograms, or from computer simulated reduced images derived from the full volume ground truth image 506. In the training mode, training may be performed based on both one or more images of subsequences 502(a)-(e) and segmented image 506.

[0105] Although FIG. 5 illustrates subsequence 502(a)-(e) as having five angiographic images, it should be understood that the number five is merely exemplary. In some circumstances, fewer or more than five images may be employed from the same angiographic examination. For example, five or more angiographic images may be used to estimate a segmentation of intermediate images of an angiographic sequence that includes dozens of individual images. This allows for an increased signal-to-noise ratio in the angiographic image of interest (albeit at the expense of increased computational resources) compared to using only five angiographic images, thereby further reducing the amount of chemical contrast agent and / or x-ray radiation.

[0106] In this arrangement, the vessel segmentation of an angiographic image may be estimated based on fewer surrounding images toward the beginning or end of the sequence. For example, the sixth image may be segmented based on its location in the middle of five images (e.g., 4, 5, 6, 7, 8), but the fifth image may not be in the middle of the five-image subsequence, but may be treated as being in the middle of a three-image subsequence (e.g., 4, 5, 6).

[0107] In addition, there may be situations where the angiographic image of interest (e.g., the target image) is not in the middle of the subsequence but is in another location (e.g., near the beginning or end of the subsequence). For example, during an angiographic catheter positioning operation, the angiographer may choose to perform intermittent real-time angiographic injections while acquiring fluoroscopic images. The angiographic image that one wishes to see cannot be the central image in the context of the real-time angiographic images because future angiographic images have not yet been acquired. Instead, the deep learning network may perform signal-to-noise enhancement based on the target image and multiple images preceding it in time.

[0108] 6 is a diagram illustrating subsetting a sequence of angiographic images into multiple overlapping subsequences, each estimating vascular structure for one angiographic image. More specifically, FIG. 6 illustrates a first exemplary method for estimating a high-quality image when a target image is in the center of the subsequence and a second exemplary method for estimating a high-quality image when a target image is not in the center of the subsequence. The first method may be applied to offline fluoroscopic angiography, and the second method may be applied to real-time fluoroscopic angiography.

[0109] A first method 600 is shown in the top panel of FIG. 6. As shown, a sequence 602 of angiogram images of total length n is generated. Within the sequence 602 of angiogram images, there is a subsequence 604 of five angiogram images, with a central image of interest for segmentation by the deep learning network. The neural network may generate a segmented image 606 corresponding to the central (third) angiogram image of the subsequence. This process may be repeated image by image for the sequence 602 of angiogram images until all images except the first two and last two are images of interest for segmentation by deep learning. This may generate a sequence of segmented images 608.

[0110] A second method 700 is illustrated in the lower panel of FIG. 6. Here, the angiographic image of interest may be the first (most recent) angiographic image 704(a) of a subsequence of angiographic images 704(a)-(c). The neural network applies deep learning computations to improve the signal-to-noise ratio of the target image 704(a) and generate a high-quality segmented version 706(a) of the target image, thereby preserving image quality while reducing the dose of chemical contrast agent or x-ray dose. For example, the neural network may use the target image and one or more subsequent temporally consecutive images as inputs for the deep learning network computations.

[0111] The present invention may include methods, systems, apparatus, and / or computer program products at any level of technical detail in which they may be integrated. The computer program product may include a computer readable storage medium having computer readable program instructions for causing a processor to carry out aspects of the present invention.

[0112] The computer-readable program instructions described herein may be downloaded from a computer-readable storage medium to each computing / processing device or to an external computer or storage device over a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network may include conductive transmission cables, optical transmission fibers, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device may receive the computer-readable program instructions from the network and forward the computer-readable program instructions for storage in a computer-readable storage medium in each computing / processing device.

[0113] Aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer readable program instructions.

[0114] These computer-readable program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing device to produce a machine, whereby the instructions executed by the processor of the computer or other programmable data processing device create means for performing the functions / operations specified in the flowchart and / or block diagram blocks. These computer-readable program instructions may be stored on a non-transitory computer-readable storage medium capable of directing a computer, programmable data processing device, and / or other device to function in a particular manner. Thus, the computer-readable storage medium on which the instructions are stored constitutes an article of manufacture including instructions implementing aspects of the functions / operations specified in the flowchart and / or block diagram blocks.

[0115] The computer readable program instructions may be loaded into a computer, other programmable data processing apparatus, or other device and executed on the computer, other programmable apparatus, or other device to generate a sequence of operational steps to generate a computer-implemented process, such that the instructions executing on the computer, other programmable apparatus, or other device perform the functions / operations specified in the flowchart and / or block diagram blocks.

[0116] The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowcharts or block diagrams may represent a module, segment, or portion of instructions, which may include one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions noted in the blocks may occur out of the order noted in the figures. For example, two blocks shown in succession may in fact be executed substantially simultaneously, or the blocks may be executed in the reverse order, depending on the functionality involved. It should also be noted that each block of the block diagrams and / or flowchart diagrams, as well as combinations of blocks in the block diagrams and / or flowchart diagrams, may be implemented by a special purpose hardware-based system that executes the specified functions or operations, or executes a combination of special purpose hardware and computer instructions.

[0117] The above description is intended to teach those skilled in the art how to practice the subject matter of this application, and is not intended to detail all obvious modifications and variations that may become apparent to those skilled in the art upon reading the description. However, all such obvious modifications and variations are intended to be included within the scope of the invention as defined by the claims. The claims are intended to encompass the elements and steps in any order effective to fulfill the purpose intended therein, unless the context specifically indicates to the contrary.

Claims

1. 1. A method of providing an angiographic target image obtained using a first amount of a chemical contrast agent and / or x-ray radiation as an input to a machine learning model via a processor; the machine learning model is trained using (a) as input, a second angiographic image obtained with a second amount of chemical contrast agent and / or x-ray radiation, and (b) as known output, a third angiographic image obtained with a third amount of chemical contrast agent and / or x-ray radiation, the third amount being greater than the first and second amounts; obtaining an output from the machine learning model via the processor, the output including an angiogram image, the angiogram image being a processed version of the target image.

2. 2. The method of claim 1, wherein the target image is one of a plurality of angiographic images in a first subsequence of angiographic images obtained using the first amount of chemical contrast agent and / or X-ray radiation and provided as input to the machine learning model, and the second angiographic image is one of a plurality of angiographic images in a second subsequence of angiographic images obtained using the second amount of chemical contrast agent and / or X-ray radiation and used to train the machine learning model.

3. The method of claim 2 , wherein the first subsequence of angiographic images is provided to the machine learning model as a single vector.

4. The method of claim 2 , wherein the first and second subsequences of angiographic images each consist of an odd number of angiographic images.

5. The method of claim 4 , wherein the target image is located at the center of the first subsequence of angiographic images.

6. The method of claim 2 , wherein the first subsequence of angiographic images consists of the same number of angiographic images as the second subsequence of angiographic images.

7. The method of claim 2 , wherein the first and second subsequences each consist of five angiographic images.

8. the first subsequence of angiographic images is one of a plurality of input subsequences, each of the plurality of input subsequences extracted from a different portion of a sequence of angiographic images acquired using the first amount of chemical contrast agent and / or X-ray radiation; 3. The method of claim 2, further comprising providing the plurality of input subsequences as inputs to the machine learning model via at least one of the processors, and obtaining outputs from the machine learning model via at least one of the processors, the outputs including angiogram images that are processed versions of respective target images in each of the plurality of input subsequences.

9. 3. The method of claim 2, wherein the second subsequence of angiographic images is one of a plurality of subsequences used to train the machine learning model, each of the plurality of subsequences extracted from a different portion of a sequence of angiographic images obtained with the second amount of chemical contrast agent and / or X-ray radiation.

10. 3. The method of claim 2, wherein each angiographic image in the first subsequence of angiographic images is temporally contiguous with at least one other angiographic image in the first subsequence, and each angiographic image in the second subsequence of angiographic images is temporally contiguous with at least one other angiographic image in the second subsequence.

11. The method of claim 1 , wherein the first and second amounts are the same.

12. 2. The method of claim 1, wherein the first and second amounts are less than an amount necessary to obtain a diagnostically useful image from an x-ray imaging device, and the third amount is an amount sufficient to obtain a diagnostically useful image from the x-ray imaging device.

13. The method of claim 2 , wherein the angiographic images of the second subsequence are angiographic images of a non-human vasculature.

14. The method of claim 2 , wherein the angiographic images of the second subsequence are angiographic images of an artificial vasculature.

15. 3. The method of claim 2, wherein the second amount of chemical contrast agent and / or X-ray radiation is greater than the first amount of chemical contrast agent and / or X-ray radiation, and the angiographic images in the second subsequence are angiographic images modified to simulate having been obtained using an amount of chemical contrast agent and / or X-ray radiation less than the second amount.

16. 2. The method of claim 1, wherein the second amount of chemical contrast agent and / or x-ray radiation is greater than the first amount of chemical contrast agent and / or x-ray radiation, and the second angiographic image is modified to simulate having been obtained using an amount of chemical contrast agent and / or x-ray radiation less than the second amount.

17. 17. The method of claim 16, wherein the second angiographic image is modified by adding randomly generated noise to at least some pixels of the second angiographic image.

18. 18. The method of claim 17, wherein the randomly generated noise is added only to pixels determined to correspond to blood vessels in the second angiographic image.

19. 2. The method of claim 1, wherein the third angiographic image is a segmented angiographic image and the output from the machine learning model is an angiographic image that is a segmented version of the first angiographic image.

20. The method of claim 1 , further comprising displaying the processed version of the first angiographic image on a display via the processor.

21. 1. A system comprising: one or more memory devices; at least one processor coupled to one or more of said memory devices, said at least one processor comprising: providing an angiographic target image obtained with a first amount of chemical contrast agent and / or x-ray radiation as input to a machine learning model, said machine learning model being trained with (a) as input, a second angiographic image obtained with a second amount of chemical contrast agent and / or x-ray radiation, and (b) as known output, a third angiographic image obtained with a third amount of chemical contrast agent and / or x-ray radiation, said third amount being greater than said first amount; The system obtains an output from the machine learning model, the output including an angiogram image, the angiogram image being a processed version of the target image.

22. 22. The system of claim 21, wherein the target image is one of a plurality of angiographic images in a first subsequence of angiographic images obtained using the first amount of chemical contrast agent and / or X-ray radiation, and the second angiographic image is one of a plurality of angiographic images in a second subsequence of angiographic images obtained using the second amount of chemical contrast agent and / or X-ray radiation and used to train the machine learning model, and wherein at least one of the processors is configured to provide a plurality of the angiographic images in the first subsequence as the input to the machine learning model.

23. The system of claim 22 , wherein the target image is located at the center of the first subsequence of angiographic images.

24. 23. The system of claim 22, wherein the first subsequence of angiographic images is one of a plurality of input subsequences of angiographic images provided to the machine learning model, each of the plurality of input subsequences extracted from a different portion of a sequence of angiographic images obtained using the first amount of chemical contrast agent and / or X-ray radiation, and the processor is configured to provide the plurality of input subsequences as inputs to the machine learning model and to obtain output from the machine learning model including angiographic images that are processed versions of respective target images in each of the plurality of input subsequences.

25. 23. The system of claim 22, wherein the first subsequence of angiographic images is comprised of the same number of angiographic images as the second subsequence of angiographic images.

26. 22. The system of claim 21, wherein the third amount of chemical contrast agent and / or x-ray radiation is greater than the second amount of chemical contrast agent and / or x-ray radiation.

27. 22. The system of claim 21, wherein the second amount of chemical contrast agent and / or x-ray radiation is greater than the first amount of chemical contrast agent and / or x-ray radiation, and the second angiographic image is modified to simulate having been obtained using an amount of chemical contrast agent and / or x-ray radiation less than the second amount.

28. 22. The system of claim 21, wherein the first and second amounts are less than an amount necessary to obtain a diagnostically useful image from an x-ray imaging device, and the third amount is an amount sufficient to obtain a diagnostically useful image from the x-ray imaging device.

29. 22. The system of claim 21, wherein the third angiographic image is a segmented angiographic image and the output from the machine learning model is an angiographic image that is a segmented version of the first angiographic image.

30. 22. The system of claim 21, further comprising a display device, wherein the processor is configured to display the processed version of the first angiographic image on the display device.

31. A computer program product comprising one or more non-transitory computer-readable storage media having instructions stored thereon, the instructions being executable by at least one processor to cause the at least one processor to: providing an angiographic target image obtained with a first amount of chemical contrast agent and / or x-ray radiation as input to a machine learning model, said machine learning model being trained with (a) as input, a second angiographic image obtained with a second amount of chemical contrast agent and / or x-ray radiation, and (b) as known output, a third angiographic image obtained with a third amount of chemical contrast agent and / or x-ray radiation, said third amount being greater than said first amount; and obtaining an output from the machine learning model comprising an angiographic image, the angiographic image being a processed version of the target image.

32. 32. The computer program product of claim 31 , wherein the target image is one of a plurality of angiographic images in a first subsequence of angiographic images obtained using the first amount of chemical contrast agent and / or X-ray radiation, and the second angiographic image is one of a plurality of angiographic images in a second subsequence of angiographic images obtained using the second amount of chemical contrast agent and / or X-ray radiation and used to train the machine learning model, and the instructions are executable by at least one of the processors, and the instructions cause the at least one processor to provide a plurality of the angiographic images in the first subsequence as the input to the machine learning model.

33. 33. The computer program product of claim 32, wherein the target image is located at the center of the first subsequence of angiographic images.

34. 33. The computer program product of claim 32, wherein the first subsequence of angiographic images is one of a plurality of input subsequences of angiographic images provided to the machine learning model, each of the plurality of input subsequences extracted from a different portion of a sequence of angiographic images obtained using the first amount of chemical contrast agent and / or X-ray radiation, and wherein the instructions are executable by at least one of the processors to provide the plurality of input subsequences as inputs to the machine learning model and obtain output from the machine learning model including angiographic images that are processed versions of respective target images in each of the plurality of input subsequences.

35. 33. The computer program product of claim 32, wherein the first subsequence of angiographic images consists of the same number of angiographic images as the second subsequence of angiographic images.

36. 32. The computer program product of claim 31, wherein the third amount of chemical imaging agent and / or X-ray radiation is greater than the second amount of chemical imaging agent and / or X-ray radiation.

37. 32. The computer program product of claim 31 , wherein the second amount of chemical contrast agent and / or x-ray radiation is greater than the first amount of chemical contrast agent and / or x-ray radiation, and the second angiographic image is modified to simulate having been obtained using an amount of chemical contrast agent and / or x-ray radiation less than the second amount.

38. 32. The computer program product of claim 31 , wherein the first and second amounts are less than an amount necessary to obtain a diagnostically useful image from an x-ray imaging device, and the third amount is an amount sufficient to obtain a diagnostically useful image from the x-ray imaging device.

39. 32. The computer program product of claim 31 , wherein the third angiographic image is a segmented angiographic image and the output from the machine learning model is an angiographic image that is a segmented version of the first angiographic image.

40. 32. The computer program product of claim 31, wherein the instructions are executable by at least one of the processors, the instructions causing the at least one of the processors to display the processed version of the first angiographic image on a display device.

Images