Image processing for stroke characterization

Abstract

The present invention discloses an apparatus for stroke characterization, comprising a processor configured to receive image data representing a three-dimensional image of a subject's head, identify regions within the image data corresponding to bones of the subject's head, apply adjustments to the image data to compensate for effects caused by radiation scattered from the bones during acquisition of the image data, generate a two-dimensional image based on the adjusted image data, and provide the generated two-dimensional image as input to a predictive model for identifying regions of interest in the two-dimensional image for stroke characterization.

Description

[Technical field]

[0001] FIELD OF THE DISCLOSURE The present invention relates to stroke characterization, and more particularly to image processing for identifying regions of interest for stroke characterization.Apparatus as well as methods and computer program products for stroke characterization are disclosed. [Background technology]

[0002] Stroke is a medical condition in which insufficient blood flow to a person's brain leads to cell death. Two main types of strokes occur: ischemic, caused by lack of blood flow, and hemorrhagic, caused by bleeding. Patients who suffer from hemorrhagic strokes may also suffer from intracerebral hemorrhage (ICH), also known as cerebral hemorrhage, which is a type of intracranial bleeding that occurs within the tissue or ventricles of a person's brain. The type of treatment provided to a person who has suffered a stroke depends on the type of stroke, the cause of the stroke, and the part of the brain that is affected. Typically, a person who has suffered a stroke is treated with medications, such as drugs to prevent and dissolve blood clots, or drugs to reduce bleeding. In some instances, surgery may be required to treat a person who has suffered a stroke. Summary of the Invention [Problem to be solved by the invention]

[0003] Every minute that a stroke goes untreated can result in the death of approximately 2 million neurons in the brain, causing parts of the brain to not function properly. The sooner a person receives treatment for a stroke, the less damage to the brain will occur. Therefore, it is desirable to detect the onset of a stroke as early as possible. For example, it is desirable to be able to detect and characterize (e.g., understand the type of stroke, the location of the stroke, and the cause of the stroke) early on the patient's contact with a medical facility (e.g., in an emergency room), so that a stroke can be diagnosed quickly so that necessary treatment can be initiated quickly, resulting in a more favorable medical outcome for the patient. [Means for solving the problem]

[0004] The inventors of the present invention have recognized a need to be able to provide a rapid and reliable diagnosis in relation to patients exhibiting potential symptoms of stroke. Accordingly, the present invention provides a mechanism by which three-dimensional image data may be used for stroke characterization. More specifically, three-dimensional image data (e.g., a three-dimensional scan of a patient's head) may be pre-processed such that it may be provided as input to a predictive model (e.g., a machine learning model) trained to identify regions of interest associated with and used for stroke detection and characterization. The pre-processing performed in relation to the three-dimensional image data effectively cleans the image data, converting the three-dimensional image into a two-dimensional image, while preserving many, if not all, of the details used to detect stroke and characterize stroke.

[0005] According to a first aspect, the present invention provides an apparatus for stroke characterization, the apparatus comprising a processor configured to receive image data representing a three-dimensional image of a subject's head, identify regions in the image data corresponding to bones of the subject's head, apply adjustments to the image data to compensate for effects caused by radiation waves scattered from the bones during acquisition of the image data, generate a two-dimensional image based on the adjusted image data, and provide the generated two-dimensional image as input to a predictive model for identifying regions of interest for stroke characterization in the two-dimensional image.

[0006] By processing the image data in this manner (i.e., by adjusting the image data to compensate for radiation scattering from bone), the predictive model may analyze the image data more effectively and generate more accurate output. In particular, any increases in pixel intensity in the image data caused by radiation scattering may be compensated for (reduced or removed) to reduce the likelihood that the predictive model will misinterpret or erroneously feature pixel intensities as being associated with hemorrhage in the brain. As a result, the likelihood of false positive features (e.g., stroke diagnosis) is reduced.

[0007] By performing maximum intensity projection on the three-dimensional image data, it is possible to generate detail-preserving two-dimensional images from the three-dimensional images that are used to detect and / or characterize strokes in the image data. Furthermore, by processing the three-dimensional data before performing maximum intensity projection, only the most relevant parts of the image data are analyzed, resulting in highly accurate predictions (e.g. identifying regions of interest). The combination of processing and maximum intensity projection means that any regions of interest, e.g. bleeding events, are clearly visible and identifiable in the generated two-dimensional images, and therefore the output of the predictive model can be regularly checked and approved by medical experts.

[0008] In some embodiments, the processor is further configured to overlay the image data on the three-dimensional representation of the brain using a plurality of fiducial markers common to the image data and the three-dimensional representation.

[0009] In some embodiments, the processor is configured to identify a portion of the image data that corresponds to a boundary of the subject's brain by applying a mask to each of a plurality of consecutive slices of the image data to remove a region of the image data that corresponds to the bone, defining a boundary enclosing the region of the image data that corresponds to the bone, identifying a subregion of the image data within the boundary having a maximum number of contiguous pixels, and determining the identified subregion of the image data that corresponds to the subject's brain.

[0010] In some embodiments, the processor determines the brain area visible in each slice. Size analyzing successive slices of the image data progressing through the head to identify brain regions visible in the slice immediately preceding the particular slice; Size of brain tissue visible in a particular slice relative to Size and in response to determining that the change in is less than a specified threshold, determining that the particular slice contains image data corresponding to the lower boundary of the subject's brain, thereby identifying a portion of the image data corresponding to the lower boundary of the subject's brain.

[0011] The processor, in some embodiments, is configured to generate a two-dimensional image based on the image data by performing a maximum intensity projection of the image data through at least one of a coronal plane, an axial plane, and a sagittal plane.

[0012] In some embodiments, the processor is further configured to obtain an indication of the identified region of interest as an output of the predictive model, and to generate a bounding box for presentation around the region of interest in a representation of the subject's brain.

[0013] In some embodiments, the region of interest comprises a region in which bleeding occurs. The processor is further configured to calculate (e.g., automatically) a score indicative of the severity of the bleeding based on the output of the predictive model.

[0014] The image data, in some embodiments, includes data acquired using a non-contrast computed tomography imaging modality.

[0015] In some embodiments, the predictive model includes a convolutional neural network trained to determine whether an identified region of interest is indicative of a bleeding event.

[0016] In some embodiments, the adjustment applied to the image data includes reducing pixel intensities in regions corresponding to regions of the subject's head into which radiation is scattered from bone during acquisition of the image data. For example, the adjustment may include adjusting pixel intensities according to the following formula:

number

[0017] According to a second aspect, the present invention provides a method for stroke characterization in medical image data, the method comprising receiving image data representing a three-dimensional image of a subject's head, identifying regions in the image data corresponding to bones of the subject's head, applying adjustments to the image data to compensate for effects caused by radiation waves scattered from the bones during acquisition of the image data, generating a two-dimensional image based on the adjusted image data, and providing the generated two-dimensional image as input to a predictive model for identifying regions of interest for stroke characterization in the two-dimensional image.

[0018] According to a third aspect, the present invention provides a method of processing image data for use in stroke characterization comprising receiving three-dimensional image data representing a subject's head, pre-processing the three-dimensional image data by applying adjustments to the three-dimensional image data to compensate for effects caused by radiation waves scattered from bones during acquisition of the three-dimensional image data, and generating a two-dimensional image based on the three-dimensional image data, thereby preparing a training data set for each of a plurality of subjects.

[0019] In some embodiments, the method further comprises training a predictive model using the training data set to identify regions of interest in the two-dimensional image for stroke characterization.

[0020] According to a fourth aspect, the present invention provides a computer program product comprising a non-transitory computer readable medium including computer readable code embodied therein, the computer readable code being configured, when executed by a suitable computer or processor, to cause the computer or processor to perform the steps of the methods disclosed herein.

[0021] These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter.

[0022] For a better understanding of the invention, and to show more clearly how the same may be carried into effect, reference will now be made, by way of example only, to the accompanying drawings in which: [Brief description of the drawings]

[0023] [Figure 1] 1 is a schematic diagram of an example of an apparatus according to various embodiments disclosed herein. [Diagram 2] FIG. 1 is a schematic diagram of an example of a predictive model for use in accordance with various embodiments disclosed herein. [Diagram 3] 1 is a flowchart of an example process performed by various embodiments disclosed herein. [Figure 4] 11 is a flowchart of a further example of processing performed by various embodiments disclosed herein. [Diagram 5] FIG. 1 illustrates the effect of radial wave scattering. [Figure 6] 1 is a flowchart of an example method according to various embodiments disclosed herein. [Figure 7] 11 is a flowchart of a further example method according to various embodiments disclosed herein. [Figure 8]1 is a schematic diagram of a computer-readable medium in communication with a processor. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0024] The examples described herein provide apparatus and methods that can be used for stroke characterization in medical images. A three-dimensional image of a subject's head is acquired, for example, using known medical imaging techniques, and processed as described herein such that the processed image is provided as input to a trained predictive model (e.g., a neural network, or a deep learning model) for analysis to detect signs that indicate the subject has suffered a stroke and / or to characterize signs that indicate the subject has suffered a stroke. By processing the three-dimensional data with the techniques disclosed herein, the data provided to the trained predictive model is "cleaner" and is more likely to result in a highly accurate output. The ability to obtain accurate output is particularly important in the medical sector, where a quick diagnosis means that appropriate action can be taken in a timely manner. According to various embodiments disclosed herein, the present invention further provides a method for processing image data for use in stroke characterization, for example, preparing a training data set that can be used to train a predictive model for stroke characterization.

[0025] 1 shows a block diagram of an apparatus 100 according to an embodiment that may be used for stroke feature acquisition. For example, the apparatus 100 may be used to process data used for stroke identification or stroke feature acquisition. With reference to FIG. 1, the apparatus 100 comprises a processor 102 that controls the operation of the apparatus 100 and may implement the methods described herein.

[0026] The apparatus 100 further comprises a memory 106 including instruction data representing an instruction set. The memory 106 may be configured to store the instruction data in the form of program code that can be executed by the processor 102 to perform the methods described herein. In some implementations, the instruction data may include multiple software and / or hardware modules, each configured to perform or to perform an individual step or steps of the methods described herein. In some embodiments, the memory 106 may be part of a device that further comprises one or more other components of the apparatus 100 (e.g., the processor 102 and / or one or more other components of the apparatus 100). In alternative embodiments, the memory 106 may be part of a device that is separate from the other components of the apparatus 100.

[0027] In some embodiments, memory 106 comprises multiple sub-memories, each of which may store a portion of the instruction data. In some embodiments in which memory 106 comprises multiple sub-memories, instruction data representing an instruction set may be stored in one sub-memory. In other embodiments in which memory 106 comprises multiple sub-memories, instruction data representing an instruction set may be stored in multiple sub-memories. For example, at least one sub-memory stores instruction data representing at least one instruction of the instruction set, and at least one other sub-memory stores instruction data representing at least one other instruction of the instruction set. Thus, in some embodiments, instruction data representing different instructions may be stored in one or more different locations in device 100. In some embodiments, memory 106 is used to store information, data (e.g., image data), signals, and measurements obtained or created by processor 102 of device 100, or from any other component of device 100.

[0028] The processor 102 of the device 100 may be configured to communicate with the memory 106 to execute a set of instructions that, when executed by the processor 102, cause the processor 102 to perform the methods described herein. The processor 102 may comprise one or more processors, processing units, multi-core processors, and / or modules configured or programmed to control the device 100 in the manner described herein. In some embodiments, for example, the processor 102 comprises multiple (e.g., interoperated) processors, processing units, multi-core processors, and / or modules configured for distributed processing. It will be understood by those skilled in the art that such processors, processing units, multi-core processors, and / or modules may be located in different locations and may perform different steps of the methods described herein and / or different portions of a step.

[0029] 1 , in some embodiments, the device 100 includes at least one user interface 104. In some embodiments, the user interface 104 may be part of a device that further includes one or more other components of the device 100 (e.g., the processor 102, the memory 106, and / or one or more other components of the device 100). In alternative embodiments, the user interface 104 may be part of a device that is separate from the other components of the device 100.

[0030] The user interface 104 is for use in providing information resulting from a method according to an embodiment herein to a user of the device 100 (e.g., a medical professional, such as a radiologist, or any other user). The set of instructions, when executed by the processor 102, causes the processor 102 to control the one or more user interfaces 104 to provide information resulting from a method according to an embodiment herein. Alternatively, or in addition, the user interface 104 may be configured to receive user input. In other words, the user interface 104 may allow a user of the device 100 to manually input instructions, data, or information. The set of instructions, when executed by the processor 102, causes the processor 102 to obtain user input from the one or more user interfaces 104.

[0031] User interface 104 may be any user interface that allows for the rendering (or output or display) of information, data, or signals to a user of device 100. Alternatively, or additionally, user interface 104 is any user interface that allows a user of device 100 to provide user input, interact with, and / or control device 100. For example, user interface 104 may comprise one or more switches, one or more buttons, a keypad, a keyboard, a mouse, a mouse wheel, a touch screen (e.g., in a tablet or smartphone) or application, a display screen, a graphical user interface (GUI) or other visual rendering component, one or more speakers, one or more microphones or other audio components, one or more lights, a component for providing haptic feedback (e.g., a vibration function), or any other user interface or combination of user interfaces.

[0032] 1, the apparatus 100 further comprises a communication interface (or circuitry) 108 to enable the apparatus 100 to communicate with interfaces, memories, and / or devices that are part of the apparatus 100. The communication interface 108 communicates with any interfaces, memories, and devices wirelessly or via a wired connection.

[0033] 1 shows only the components required to illustrate this aspect of the disclosure, and it will be understood that in a practical implementation, device 100 may include additional components to those shown. For example, device 100 may include a battery or other power source for powering device 100, or a means for connecting device 100 to a mains power source.

[0034] The device 100 is used for stroke characterization. More specifically, the device 100 is used for processing image data (e.g., medical image data) and as such is referred to as a medical image processing device. According to embodiments disclosed herein, the processor 102 is configured to receive image data representing a three-dimensional image of a subject's head. In some examples, the image data represents the subject's entire head, and in other examples, the image data represents only a portion of the subject's head. Typically, the subject's brain (or at least a large portion thereof) is imaged for stroke characterization so that events occurring in the brain indicative of the onset of a stroke can be detected and analyzed.

[0035] The image data includes data acquired using, for example, a non-contrast computed tomography imaging modality. For example, the image data includes three-dimensional (e.g., volumetric) image data acquired from a non-contrast computed tomography (CT) scan. In one example, a low-dose computed tomography (low-dose CT or LD CT) imaging modality is used to acquire the image data. In another example, another imaging modality is used to acquire the three-dimensional image data. In general, imaging modalities involve directing electromagnetic radiation towards a subject being imaged and detecting the interaction of the electromagnetic radiation with the subject. The type of radiation used typically depends on the imaging modality.

[0036] The processor 102 is configured to identify regions within the image data that correspond to bones of the subject's head. Regions within the image data that correspond to bones include regions that correspond to the subject's skull or portions of the skull.

[0037] After the processor 102 identifies a region of the image data corresponding to the bone, the processor 102 is configured to apply an adjustment to the image data to compensate for the effects caused by radiation waves scattered from the bone during acquisition of the image data. As described above, electromagnetic radiation waves are directed at the subject during imaging. The behavior of the radiation waves during imaging (e.g., absorption by materials in the subject's head) allows the subject to be imaged. However, when the radiation waves strike a bone (e.g., the skull), the radiation waves reflect and scatter from the bone's surface. The scattered radiation waves are detected by the imaging device, and the scattered radiation waves are visible in the image data. The scattered radiation waves (or the visible effects of the scattered radiation waves) obscure other objects in the subject's head, making it difficult to clearly see objects or events occurring in the subject's brain. Furthermore, the effects of the scattered radiation waves (e.g., the appearance of the scattered radiation waves in an image of the subject's brain) are similar to characteristic events indicative of a stroke. Thus, in an image (e.g., a 3D scan) of a subject's brain, radiation scattered from the inner surface of the skull may appear to be a stroke-related event (e.g., a bleeding event), resulting in a misdiagnosis, such as a false identification that the subject has suffered a stroke. Thus, adjustments applied to the image data are intended to compensate for the scattered radiation, such that an event (e.g., an increase in intensity of a group of pixels in an image) is not misinterpreted as an event characteristic of a stroke.

[0038] The processor 102 is further configured to generate a two-dimensional image based on the adjusted image data, in some embodiments, this is done by performing a maximum intensity projection of the adjusted three-dimensional image data, as described in more detail below.

[0039] The processor 102 is further configured to provide the generated two-dimensional image as an input to a predictive model for identifying regions of interest for stroke characterization in the two-dimensional image. The predictive model includes a model executed using the processor 102. Alternatively, a different processor (e.g., a processor external and / or remote to the device 100) is used to execute the predictive model. The predictive model, including an artificial neural network or a classifier (described in more detail below), is trained to analyze the two-dimensional input image to identify portions of the image (i.e., regions of interest) relevant for stroke characterization based on a feature set. For example, the predictive model is trained to identify regions of the image that exhibit hemorrhage. Such regions in the image may not be easily identifiable by a human (e.g., a radiologist) but may be better identifiable using the trained predictive model.

[0040] In some examples, the predictive model includes an artificial neural network. Although artificial neural networks, or simply neural networks, and other machine learning models, are understood by those skilled in the art, in brief, a neural network is a type of model that can be used to classify data (e.g., classify or identify the content of image data). The structure of a neural network is inspired by the human brain. A neural network comprises layers, each layer including a number of neurons. Each neuron includes a mathematical operation. In the process of classifying a portion of data, the mathematical operation of each neuron is performed on the portion of data to generate a numerical output, and the output of each layer in the neural network is sequentially fed to the next layer. Generally, the mathematical operation associated with each neuron includes one or more weights that are adjusted during a training process (e.g., the value of the weights is updated during the training process to adjust the model to generate more accurate classifications).

[0041] For example, in a neural network model for classifying the content of an image, each neuron in the neural network includes a mathematical operation that includes a weighted linear sum of pixel (or voxel in three dimensions) values ​​in the image, followed by a nonlinear transformation. Examples of nonlinear transformations used in neural networks include sigmoid functions, hyperbolic tangent functions, and regularized linear functions. Neurons in each layer of a neural network generally include different weighted combinations of one type of transformation (e.g., the same type of transformation, sigmoid, etc., but with different weightings). As can be appreciated by those skilled in the art, in some layers the same weight may be applied by each neuron in the linear sum, which is the case, for example, in the case of a convolutional layer. The weights associated with each neuron make certain features more prominent (or conversely, less prominent) than other features in the classification process, and thus adjusting the weights of neurons in the training process trains the neural network to give higher significance to certain features when classifying images. Generally, a neural network has weights associated with neurons and / or weights between neurons (e.g., modifying the data values ​​passing between neurons).

[0042] As briefly mentioned above, in some neural networks, e.g., convolutional neural networks, lower layers in the neural network, e.g., input layers or hidden layers (i.e., layers toward the beginning of the sequence of layers in the neural network), are activated by small features or patterns in the portion of the data to be classified (i.e., the output of the lower layers depends on small features or patterns in the portion of the data to be classified), while higher layers (i.e., layers toward the end of the sequence of layers in the neural network) are activated by increasingly larger features in the portion of the data to be classified. As an example, if the data includes images, the lower layers in the neural network are activated by small features (e.g., edge patterns in the image, for example), mid-level layers are activated by features in the image, e.g., larger shapes and forms, and the layers closest to the output (e.g., top layers) are activated by entire objects in the image.

[0043] In general, the weights in the final layer (known as the output layer) of a neural network model are most dependent on the particular classification problem being solved by the neural network. For example, the weights in the outer layers will depend heavily on whether the classification problem is a localization or detection problem. It is recognized herein that the weights in lower layers (e.g., input and / or hidden layers) tend to depend on the content (e.g., features) of the data being classified, and thus, with sufficient training, the weights in the input and hidden layers of neural networks processing the same type of data will converge towards the same values ​​over time, even if the outer layers of the model are tuned to solve different classification problems.

[0044] In some examples, the predictive model includes a convolutional neural network trained to determine whether an identified region of interest indicates a hemorrhagic event. Because such a predictive model is trained on features associated with hemorrhage, the predictive model may identify the area where hemorrhage occurred or the area of ​​the brain affected by hemorrhage, even if a human observer (e.g., a radiologist) cannot immediately identify the onset of hemorrhage by simply looking at the image data. In one particular example, a convolutional neural network model known as Inception-ResNet is used as the predictive model. The Inception-ResNet model includes approximately 780 layers, and a concise representation of the model is shown diagrammatically in FIG. 2. The model 200 shown in FIG. 2 includes blocks representing a convolutional layer 202, a max pooling function 204, an average pooling function 206, a concatenation function 208, a dropout function 210, a fully connected layer 212, a softmax function 214, and the remaining network layers 216. Other examples of models suitable for use in the methods disclosed herein may include fewer or additional layers and / or functions.

[0045] Examples of image data processing methods implemented to aid in stroke detection and / or characterization in medical image data acquired in relation to a subject are described below with reference to Figures 3 and 4. Figure 3 is a flow chart of an exemplary workflow implemented in relation to three-dimensional image data acquired in relation to a subject. The three-dimensional image data includes a three-dimensional CT volume of the subject's head. For example, the subject suffers from symptoms indicative of a stroke and is therefore referred to a medical facility (e.g., a hospital) for further examination. The subject's head is scanned using an imaging modality, for example, a non-contrast CT. The three-dimensional image data (e.g., a 3D CT volume) typically includes a series of slices of image data acquired in planes (e.g., axial, coronal, and / or sagittal planes), and processing of the three-dimensional image data generally involves processing each of the slices in succession. Thus, in the workflow illustrated in Figure 3, the three-dimensional image data 300 includes a plurality of slices of image data. In block 302, an i-th slice of the plurality of slices is retrieved. In block 304, the i-th slice undergoes pre-processing, which will be described in more detail below, and after the i-th slice has been pre-processed, the workflow proceeds to block 306 where the next (i+1)-th slice is acquired from the three-dimensional image data 300 for processing. Thus, after completion of pre-processing of the slice of image data, the workload involves acquiring the next slice, and the workflow returns to block 302.

[0046] The workflow continues until all of the slices containing image data relevant to the subject's brain have been pre-processed. An example of a stopping criterion that is recognized as being met when all of the image slices containing image data representative of the brain have been processed is described below. After all of the relevant slices have been processed, the workflow proceeds to block 308 where a two-dimensional image is generated based on the pre-processed image data. In some examples, multiple two-dimensional images are generated as described below. The two-dimensional image is then provided as an input to a trained predictive model for analysis including identifying regions of interest indicative of a stroke and / or obtaining features of the regions of interest.

[0047] 4 is a flow chart illustrating examples of various operations that form part of the pre-processing of each slice of three-dimensional image data (e.g., in block 306 of FIG. 3). In general, the pre-processing includes identifying regions in the slice of image data that represent the bones (e.g., skull) of the subject (block 402), generating a mask for the identified bone regions (block 404), identifying portions of the mask that correspond to portions of the bones that encase the subject's brain (blocks 406 and 408), performing adjustments of the image data to compensate for the effects of radiation scattered from the bones when the three-dimensional image data is acquired (block 410), and extracting portions of the image data that represent the user's brain from the image data (block 412). In some examples, only the compensation adjustments for the scattered radiation (block 410) are performed during the pre-processing of the image data. In other examples, one or more other operations described herein may also be performed.

[0048] To aid in the identification of bones in the image data, pre-processing of the image data includes first registering or aligning the three-dimensional image data to a reference representation (e.g., an image or volume) to achieve the intended orientation of the image data and to ensure consistency between the orientations of slices of the image data. The registration of the three-dimensional image data is performed with respect to a three-dimensional volume that includes all of the captured slices of the image data. In other words, the registration is performed globally with respect to the volume of image data, rather than with respect to each of the slices individually. In some embodiments, the registration of the image data includes a non-rigid registration of the image data to a three-dimensional representation of a human brain, for example, using an anatomical atlas. The registration of the image data to the anatomical atlas is achieved using a plurality of fiducial markers, each of which includes a feature within the brain, for example, a ventricle or a cerebellum.

[0049] In some embodiments, the identification of fiducial markers in the image data used in the registration process is accomplished using machine learning techniques, such as trained predictive models. For example, artificial neural networks or support vector machines may be used. The predictive models are trained on one or more features, such as pixel intensity, entropy, gray-level co-occurrence matrices (GLCMs), and texture features. For example, the predictive models determine that a region of interest in the three-dimensional image data is likely (within a specified confidence threshold) to represent a particular anatomical structure in the brain. When a region of interest is detected in the image data, each voxel that falls within the region of interest is classified as falling within a particular anatomical category (e.g., each voxel within the region of interest is labeled with an identifier, such as the name of the anatomical structure with which it is associated). A voxel is considered to fall within the region of interest if it falls within a sphere of a specified radius formed around the center of the region of interest.

[0050] When classifying a particular voxel as forming part of a region of interest, a majority vote process is performed such that if all of the voxels within a particular region of interest are labeled as forming part of that region of interest, the majority label is assigned to the region of interest. A number of labeled regions of interest are then collected to define a unique fiducial landmark (e.g., formed from a collection of individual fiducial landmarks) that are used to overlay and align the image data to a reference representation (e.g., from an anatomical atlas).

[0051] The above-mentioned registration process is performed using the processor 102. Thus, in some embodiments, the processor 102 is configured to register the image data on the three-dimensional representation of the brain using a number of fiducial markers common to the image data and the three-dimensional representation. In some examples, the fiducial markers are identified using a trained classifier or predictive model, as described above.

[0052] Although registering the image data against the reference representation helps to ensure that the image data is properly aligned for further processing, the registration process is not essential and therefore may be omitted from the processing procedure.

[0053] As discussed above, the registration of image data is performed using the entire three-dimensional volume. However, the processing in blocks 402 to 412 of FIG. 4 is performed with respect to individual slices of image data. Typically, the processing of slices of image data is performed using axial slices (i.e., slices through the axial plane of the subject's head). However, it will be appreciated that in some instances, processing is performed with respect to slices of image data through other planes (e.g., through the coronal or sagittal planes). The images shown in the blocks of FIG. 4 are merely illustrative of how the images change as pre-processing is performed.

[0054] Various bone and skull stripping techniques, as understood by one skilled in the art, are used to remove portions of the image data that are not used in the analysis process. In block 402 of the flow chart in FIG. 4, regions within a slice of image data that represent bone are identified. In one embodiment, bone regions within the image data are identified using machine learning techniques, such as predictive models trained to identify portions of the image data that represent bone. In some embodiments (and as part of the machine learning techniques), bone regions of the image data are identified by analyzing bone (e.g., skull) structure patterns in the image data. To identify portions of the image that represent bone, boundary parameters and / or intensity values ​​of the image data are considered. In addition, one or more filters are applied to identify and / or extract bone regions from the image. For example, a filter is applied to remove portions of the image that are outside of a defined range on the Hounsfield scale. The Hounsfield scale is a quantitative measure used to describe radioactive density. Thus, in one example, pixels corresponding to regions with Hounsfield Units (HU) measurements outside the range of 0 to 100 are removed from the image data. It is understood that the regions within the range of 0 HU to 100 HU represent the regions of blood and brain tissue that are retained in the image data.

[0055] After the bone regions are identified, a mask for the identified regions is generated (block 404 of the flowchart in FIG. 4). The mask is generated based on a deformable model using a simplex mesh. The properties of simplex meshes make them suitable for a range of image segmentation tasks. A generic mesh is provided and deformed to align with the skull-brain interface, i.e., the inner boundary of the identified bone regions in the slice of image data. The mesh is then geometrically corrected using an affine transformation by minimizing the sum of the squared distances between the mesh vertices and the bone regions in the image data. The mesh is then deformed using the coarsely segmented image (i.e., the image segmented using anatomical landmarks) and further using the original image data. The resulting corrected mask is intended to capture all of the bones / skull contained within the slice of image data, as shown in the example in block 404.

[0056] As can be seen from the example shown in block 404, the generated mask includes regions representing portions of the skull that encase the subject's brain, along with other regions representing other portions of the skull, such as the eye sockets and / or jaw. Thus, in blocks 406 and 408 of the flowchart in FIG. 4, pre-processing involves identifying portions of the mask that correspond to portions of bone that encase the subject's brain. In block 406, a contour line 406a is formed around the entire bone region. In some examples, the contour line 406a is formed using a Bresenham line algorithm based on the convex boundary points of the skull. As shown in 406b, the contour line surrounds the identified bone region in the slice of image data.

[0057] At block 408, the region in the mask corresponding to the subject's brain is extracted. To accomplish this, regions of the mask corresponding to the orbits and / or jaw of the subject's skull are removed. In some embodiments, this is accomplished by "flood filling" (i.e., extracting) the regions outside of the contour line 406a, for example as shown in image 408a. To remove smaller portions of the mask that do not represent the subject's brain regions, connected component analysis is used. In this regard, connected component analysis involves identifying the largest connected component in the mask, in other words, the portion of the mask that contains the largest number of connected (i.e., contiguous) pixels is identified. Portions of the mask that do not form part of the largest connected component are removed or ignored. The largest connected component is shown in image 408b and includes the region corresponding to the subject's brain and the region corresponding to the subject's orbit.

[0058] A morphological erosion and / or dilation operation is performed to remove the portions of the mask corresponding to the subject's eye sockets. e ) is performed, then the morphological erosion operation (using the structuring element S d A morphological dilation operation (using d >S e After the morphological erosion and dilation operations, a morphological “AND” operation is performed on the original inner mask (i.e., the mask before any morphological operations are performed) I to maintain the brain boundary in the mask. mask The "AND" operation is defined as follows:

number

[0059] Identification of the portions of the mask corresponding to portions of bones enclosing the subject's brain (blocks 406 and 408 of FIG. 4) is performed by the processor 102 of FIG. 1. The above-described processes are considered to identify boundaries of the subject's brain (e.g., the boundary between the brain and the skull). Thus, the processor 102 is configured in some embodiments to identify the portions of the image data corresponding to boundaries of the subject's brain by applying a mask to each of a plurality of consecutive slices (e.g., axial slices) of the image data to remove regions of the image data corresponding to bone, defining a boundary that encloses the region of the image data corresponding to bone, identifying a subregion of the image data within the region enclosed by the boundary that has a maximum number of contiguous pixels, and locating the identified subregion of the image data that corresponds to the subject's brain.

[0060] After the brain mask is created or generated, radiation scattering compensation is performed in block 410 of the flowchart of FIG. 4. As described above, during imaging (e.g., CT scan), electromagnetic radiation (e.g., x-rays) are directed to an object to be imaged (e.g., a subject's head). Bones (e.g., skull) within the subject cause the electromagnetic radiation to be scattered, which is visible in the resulting image data. The scattered radiation appears itself as relatively bright (e.g., high pixel intensity) patches in the image data, and in some instances, these bright patches resemble bleeding events. Thus, bright patches in the image data caused by radiation scattering from bones may be mistaken for bleeding events, and thus the subject may be misdiagnosed as having a bleeding (i.e., a stroke), resulting in a false positive diagnosis. Corrections or adjustments are made to the image data to compensate for the effects caused by radiation scattered from bones during acquisition of the image data.

[0061] To compensate for radiation scattering from bone, a scattering profile is first determined. The scattering profile defines the change in pixel intensity in the image data depending on the distance from the bone surface where the radiation is scattered. In this example, the scattering profile is defined in terms of the change in intensity as a function of distance from the inner surface of the subject's skull towards the brain (i.e., intracranial space). FIG. 5 shows an example of a scattering profile for radiation scattered from the inner surface of the skull. In FIG. 5, an image 502 of a slice of image data includes a region within the inner surface of the skull where the increase in pixel intensity is due to radiation scattered from the skull during image acquisition. The pixel intensity gradually decreases from high (in the skull region) to low (in the brain region) due to radiation scattering. Higher pixel intensities are shown in a magnified image 504 of a portion of image 502. A graph 506 shows how the pixel intensity y changes as a function of distance x (in mm) from the surface of the skull. In particular, graph 506 shows peak pixel intensity at a location corresponding to the skull, with pixel intensity decreasing logarithmically with distance from the skull. In some examples, a logarithmic intensity compensation model based on the Maxwell-Boltzmann distribution of increasing entropy systems is used. In such models, the decay of pixel intensity is modeled in terms of increasing entropy, which is considered to be equal to increasing distance from the skull surface into the intracranial space.

[0062] In one example, the model defining pixel intensity y as a function of distance x (in millimeters) from the skull surface is expressed as follows:

number

[0063] After the additional pixel intensities caused by the scattered radiation are calculated, a modification is applied to the image data to compensate for the higher pixel intensities. The adjustment or modification is applied by the processor 102. Thus, in some embodiments, the adjustment applied to the image data includes a reduction in pixel intensities in areas corresponding to areas of the subject's head where radiation is scattered in from bone during acquisition of the image data.

[0064] After the radial scatter compensation portion of the pre-processing (block 410) is performed, a brain mask (i.e., the portion of the skull mask corresponding to the region of the subject's brain) is applied to the image data using known techniques. Any portion of the image data that is outside the brain mask is extracted, removed, or ignored, leaving only the portion of the image data that corresponds to the subject's brain. As discussed above, the pre-processing discussed with reference to FIG. 4 is performed for each slice of the image data. Thus, having completed the pre-processing procedure for a particular slice, the next slice in the three-dimensional image data is acquired and pre-processed in the same manner.

[0065] The three-dimensional image data includes slices of image data that do not represent parts of the brain. The three-dimensional image data further includes image data corresponding to parts of the brain that are unlikely to be affected by a stroke. For example, one or more slices may have been captured below the lower boundary of the brain. Only slices of image data that represent the brain are pre-processed and analyzed for the purpose of identifying regions of interest related to stroke characterization. Thus, according to some embodiments, pre-processing of slices of image data stops when a certain stopping criterion is met. Thus, the processor 102 is configured to identify a portion (e.g., slice) of the image data that corresponds to the lower boundary of the subject's brain. In one example, the determination of whether a next slice of image data in a stack / collection of image data slices is to be pre-processed is made based on the proportion of image data slices that represent the subject's brain. For example, if the amount (e.g., area) of "brain per slice" decreases from one slice to the next by more than a specified threshold, it is determined that further slices will not be pre-processed. In such a scenario, it is determined that the slices that include less brain are slices captured near the bottom of the brain. This region of the brain is unlikely to have been affected by stroke-related events (eg, hemorrhagic events) and therefore preprocessing slices below this region is unnecessary.

[0066] In one example, the above-mentioned stopping criteria are:

number

[0067] Using the above-mentioned stopping criteria, the lower boundary of the subject's brain is identified in the image data. This is performed by the processor 102 of FIG. 1. Thus, in some embodiments, the processor 102 determines the lower boundary of the brain visible in each slice. Size analyzing successive slices of the image data proceeding downward through the head to identify (e.g., an area or volume) of brain tissue visible in the slice immediately preceding a particular slice; Size of brain tissue visible in a given slice Size and in response to determining that the change in is less than a specified threshold, identifying a particular slice as including image data corresponding to an inferior boundary of the subject's brain. In some examples, successive axial slices of the image data are analyzed as they progress downward through the subject's head.

[0068] After the entire set of relevant slices has been pre-processed (e.g., using the pre-processing techniques described above), a two-dimensional image is generated based on the processed three-dimensional image data. In one example, a maximum intensity projection (MIP) of the image data is performed to generate the two-dimensional image. Maximum intensity projection is understandable to those skilled in the art of image analysis. Applying maximum intensity projection to a series of slices (e.g., slices of image data representing the brain) involves projecting voxels with maximum intensity that intercept a parallel line drawn from a viewpoint to a projection plane. The output of the processing is a two-dimensional projection or image that shows the object that appears in the image data with maximum intensity. Thus, by applying the MIP technique, all of the relevant data from the original three-dimensional image data is maintained and represented in the two-dimensional image. Importantly, any information in the original three-dimensional image data that is used to determine whether bleeding has occurred is not lost.

[0069] In some examples, multiple maximum intensity projections are performed to obtain multiple two-dimensional images. For example, a first MIP is performed through an axial plane, a second MIP is performed through a coronal plane, and a third MIP is performed through a sagittal plane. In the two-dimensional images obtained from the MIP through the coronal plane and the MIP through the sagittal plane, image interpolation is performed to compensate for the effect of image data slice thickness. For example, assuming that the slice thickness increases in the coronal and sagittal planes, discontinuities between slices will occur when MIP operations are performed. Therefore, interpolation is performed to ensure continuity between slices with smooth transitions.

[0070] Thus, in some embodiments, the processor 102 (FIG. 1) is configured to generate a two-dimensional image based on the image data by performing a maximum intensity projection of the image data through at least one of a coronal plane, an axial plane, and a sagittal plane.

[0071] The generated two-dimensional image (or multiple two-dimensional images) may be provided as an input (or multiple inputs) to a predictive model for analysis. As described above, the predictive model is trained to identify regions of interest in the two-dimensional image associated with stroke characterization. In some embodiments, the processor 102 is configured to obtain an indication of the identified regions of interest as an output of the predictive model. For example, the output of the predictive model includes an indication, e.g., a visual or textual indication, that a particular region in the image data is considered to be a region of interest associated with stroke identification and / or characterization. The processor 102 is configured in some embodiments to provide the generated two-dimensional image as an input to a predictive model (e.g., the same predictive model or a different predictive model) to determine whether the identified region of interest is indicative of a hemorrhagic event. If it can be identified that the image data does not include a region of interest or that the region of interest is not indicative of a hemorrhagic event, it is concluded that the subject is unlikely to have suffered a stroke and thus emergency treatment for stroke is not required. On the other hand, if the region of interest can be determined to be indicative of a stroke-related event, such as a hemorrhage, appropriate treatment can be administered to ameliorate the subject's long-term health concerns. Thus, in some embodiments, the processor 102 is configured such that in response to determining that an identified region of interest is indicative of a hemorrhage event, the processor generates an alarm signal. The alarm signal includes, for example, an audible alarm to a visual alarm presented at the user interface 104.

[0072] In some embodiments, the processor 102 is configured to generate a bounding box for presentation around the region of interest in the representation of the subject's brain. Thus, a bounding box is drawn around the relevant region of interest in the two-dimensional image or on a portion of the representation of the subject's brain, so that a human observer (e.g., a medical professional, e.g., a radiologist) can identify the location of the region of interest in the brain relative to other parts of the brain. This also serves to provide confidence in the predictive model to the human observer, as the human observer may verify the accuracy of the output.

[0073] In some embodiments, when the three-dimensional image data is overlaid on an anatomical atlas as described above, knowledge of various anatomical objects in the brain is used to provide richer information to a human observer. For example, it is possible to identify the anatomical part or object in the brain that is closest to the region of interest. Thus, the processor 102 in some embodiments is configured to provide for presentation an indication of the anatomical part in the subject's brain that corresponds to the identified region of interest. For example, if the region of interest is evidence of a hemorrhagic event that has occurred in the cerebellum, an indication is provided (e.g., presented) in a representation of the subject's brain that shows the cerebellum as the part of the brain where the hemorrhage has occurred.

[0074] Hemorrhage occurring in the brain is sometimes referred to as intracerebral hemorrhage (ICH). The severity of intracerebral hemorrhage is measured and defined using an intracerebral hemorrhage score or ICH score. In some embodiments, the region of interest identified by the predictive model includes the region where the hemorrhage occurred. In such embodiments, the processor 102 is configured to calculate a score indicative of the severity of the hemorrhage based on the output of the predictive model. The score includes, for example, an ICH score. The ICH score is calculated by assigning points to various factors, such as, for example, the Glasgow Coma Scale (GCS) score, the volume of the intracerebral hemorrhage, the presence of intraventricular hemorrhage (IVH), the subject's age, and the infratentorial origin of the hemorrhage. The output of the predictive model includes, for example, the ICH volume, an indication of whether an intraventricular hemorrhage is present, and an indication of the infratentorial origin of the hemorrhage. Other information (e.g., the GCS score and the subject's age) is provided by a human operator, as retrieved from a database or storage medium (e.g., memory 106) accessible by the processor. The ICH score is calculated automatically by the processor 102 by mapping all necessary inputs to the model, thus arriving at the ICH score. The processor 102 displays the ICH score on a display, for example, the user interface 104.

[0075] According to another aspect of the invention, a method is provided. Figure 6 is a flow chart of an example of a method 600, e.g., a method for stroke characterization in medical image data. The method 600 comprises, at step 602, receiving image data representing a three-dimensional image of a subject's head. The image data includes, e.g., image data acquired using a non-contrast computed tomography (CT) imaging technique. The image data may be provided manually, e.g., by a radiologist inputting the image data via a user interface, or automatically, e.g., by a processor (e.g., processor 102) retrieving the image data from a storage medium or database. At step 604, the method 600 comprises identifying regions in the image data corresponding to bones of the subject's head. Marking of the bones (e.g., skull) in the image data is performed using techniques disclosed hereinbefore. The method 600 comprises, at step 606, applying an adjustment to the image data to compensate for effects caused by radiation scattered from bones during acquisition of the image data. As described above, the adjustment made to the image data comprises lowering pixel intensities of pixels in the image data that have increased in intensity as a result of electromagnetic radiation scattering from bones during image acquisition. The pixel intensities are adjusted according to Equation 1 above in some embodiments. At step 608, the method 600 includes generating a two-dimensional image based on the adjusted image data. The two-dimensional image is generated using a maximum intensity projection technique. In some embodiments, a plurality of two-dimensional images are generated, for example, by performing maximum intensity projection through axial, coronal, and sagittal planes. At step 610, the method 600 includes providing the generated two-dimensional image (or images) as an input to a predictive model for identifying a region of interest for stroke characterization in the two-dimensional image. As mentioned above, the predictive model includes a convolutional neural network in some examples. The predictive model is trained to identify a region of the subject's brain where bleeding has occurred. Thus, the method further includes determining whether the identified region of interest is indicative of a bleeding event.In some embodiments, the predictive model provides an indication of the location of the region of interest (e.g., hemorrhage) and further other information regarding the region of interest, such as the volume of the region, an indication of the anatomical part of the brain affected, and / or a score indicative of the severity of the event that occurred (e.g., hemorrhage) (e.g., an ICH score).

[0076] Method 600 is implemented using a processor, such as processor 102. Method 600 thus includes one or more further steps implemented by processor 102 as described above. For example, the method further includes identifying a portion of the image data corresponding to a boundary of the subject's brain. In some examples, the method further includes providing for display a two-dimensional image and an indication of a region of interest in the two-dimensional image.

[0077] So far, the description has focused on the use of predictive models to identify regions of interest associated with stroke. According to another aspect, the invention relates to training predictive models to identify regions of interest. FIG. 7 is a flow chart of a further example of a method 700, e.g., a method for processing image data for use in stroke characterization. Processing image data according to the steps of the method 700 may be considered as equivalent to preprocessing the image data in the manner described above. The method 700 comprises, in step 702, preparing a training data set. The training data set is prepared by receiving image data (step 704) for each of a plurality of subjects and preprocessing the image data (step 706). Each subject comprises, for example, a patient who has undergone an imaging procedure to obtain three-dimensional image data. Thus, the training data set is prepared (step 702) by receiving three-dimensional image data representative of the subject's head (step 704) (for each of a plurality of subjects) and preprocessing the three-dimensional image data (step 706). The three-dimensional image data is preprocessed (step 706) by applying an adjustment to the three-dimensional image data to compensate for effects caused by radiation scattered from bones during acquisition of the three-dimensional image data (step 708) and generating a two-dimensional image based on the three-dimensional image data (step 710). Thus, the three-dimensional image data is preprocessed as described above (e.g., by a slice-by-slice approach) and then a two-dimensional image is generated (e.g., using maximum intensity projection). The generated two-dimensional images for each of the multiple subjects are then used to train a predictive model. As described above, in some embodiments, multiple two-dimensional images are generated, for example, by performing multiple maximum intensity projections through three different planes. In such an example, the predictive model is trained using the multiple two-dimensional images as input.

[0078] Thus, the method 700 further includes training a predictive model using the training data set to identify regions of interest in the two-dimensional image for stroke feature acquisition. As will be appreciated, the predictive model is trained on various features to perform the intended task. For example, the predictive model may identify the region of interest, locate the region of interest (i.e., provide an indication of the location of the region of interest), provide an indication of the size (e.g., volume) of the region of interest, provide an indication of the anatomical portion of the brain corresponding to the region of interest, and / or provide other information related to the region of interest. The features or weights used in the predictive model are selected depending on the intended output to be provided by the predictive model. EXAMPLES

[0079] A specific example in which a predictive model was trained in accordance with an embodiment of the present invention and used to provide indications of regions of interest for stroke characterization in accordance with other embodiments of the present invention is described below.

[0080] In this example, three-dimensional image data of multiple subjects was acquired in the form of a data set named "CQ 500" obtained from the "qure.ai" website. The data set is understood to include non-contrast head CT scan image data obtained from the Center for Advanced Research in Imaging, Neuroscience, and Genomics (CARING), New Delhi, India. The image data was acquired using one or more of the following CT scanner models: GE BrightSpeed, GE Discovery CT750 HD, GE LightSpeed, GE Optima CT660, Philips MX 16- slice, and Philips Access-32 CT. Scans were annotated as relating to brains in which bleeding had occurred (labeled "bleed") or brains in which bleeding had not occurred (labeled "non-bleed") based on evaluations made by three radiologists. Each scan volume included approximately 300 slices with a slice thickness of 0.625 mm.

[0081] Each slice of image data was preprocessed using the techniques described herein, and for each scanned volume, a maximum intensity projection was performed to generate a two-dimensional image for each subject. The collection of two-dimensional images formed the training data set used to train the predictive model. The two-dimensional images were provided as input to the predictive model with a size of 512×512 pixels with three color channels. In this example, the predictive model used was the Inception-ResNet convolutional neural network model described above, and is shown diagrammatically in FIG. 2. First, a collection of “Imagenet” weights was used in the neural network model. The neural network model was trained for 10 epochs, freezing the combination layer. The model was then recompiled, and the entire neural network model was trained for 50 epochs using a stochastic gradient descent (SGD) optimizer.

[0082] In this example, the neural network model was trained using 379 images. The images were rotated 90 degrees and flipped vertically to generate a training data set of 1137 images. The accuracy achieved by the neural network model on the training data set was 0.97, the sensitivity was 0.95, and the specificity was 0.98. Table 1 below shows the confusion matrix for the output of the neural network model using the training data set of 1137 images.

[0083] [Table 1]

[0084] The trained neural network model was tested on a validation data set containing approximately 52 images, each from a different subject. The images in the validation data set were preprocessed in a manner similar to the preprocessing performed in connection with the training data set images. Using the validation data set, the output of the neural network model has an accuracy of 0.96, a sensitivity of 0.96, and a specificity of 0.95. Table 2 below shows the confusion matrix for the output of the neural network model using the validation data set.

[0085] [Table 2]

[0086] From the above examples it is clear that by pre-processing the image data according to the techniques described herein and training the predictive model to identify regions of interest for stroke characterization, a particularly accurate predictive model is achieved.

[0087] According to a further aspect of the present invention, a computer program product is disclosed. Figure 8 is a schematic diagram of an example of a processor 802 in communication with a computer readable medium 804. According to an embodiment of the present invention, the computer program product comprises a non-transitory computer readable medium 804 having computer readable code embodied therein, which is configured, when executed by a suitable computer or processor 802, to cause the computer or processor to perform the steps of the methods disclosed herein. The processor 802 comprises or is similar to the processor 102 of the device 100 described above.

[0088] The processor 102, 802 may comprise one or more processors, processing units, multi-core processors, or modules configured or programmed to control the device 100 in the manner described herein. In certain embodiments, the processor 102, 802 may include multiple software and / or hardware modules, each configured to perform or for performing individual steps or steps of the methods described herein.

[0089] Thus, as disclosed herein, embodiments of the present invention provide a mechanism for preparing image data such that it can be used to train a predictive model to identify regions of interest relevant for stroke characterization, or provided as input to a trained predictive model capable of identifying such regions of interest. It has been found that pre-processing the data in this manner provides particularly accurate results when provided as input to a predictive model. In particular, compensating for radial wave scattering from the inner surface of the skull into the brain eliminates or greatly reduces the likelihood that regions of higher pixel intensity in the image will be mistaken for hemorrhage.

[0090] The term "module" as used herein is intended to encompass a hardware component, e.g., a processor or component of a processor, configured to perform a particular function, or a software component, e.g., a collection of instructions that have a particular function when executed by a processor.

[0091] It is understood that the embodiments of the present invention also apply to computer programs, in particular computer programs on or in a medium, adapted to carry out the present invention. The programs may be in the form of source code, object code, code intermediate source, object code, for example in partially compiled form, or in any other form suitable for use in implementing the method according to the embodiments of the present invention. It is further understood that such programs may adopt many different architectural designs. For example, the program code implementing the functions of the method or the apparatus according to the present invention may be subdivided into one or more subroutines. Many different ways of distributing functionality among these subroutines will be apparent to the skilled person. The subroutines may be stored together in one executable file to form a self-contained program. Such an executable file includes computer executable instructions, for example processor instructions and / or interpreter instructions (e.g. Java interpreter instructions). Alternatively, one or more or all of the subroutines may be stored in at least one external library file and linked statically or dynamically, for example at run-time, to the main program. The main program includes at least one call to at least one of the subroutines. The subroutines may further include function calls to each other. One embodiment of a computer program product includes computer-executable instructions corresponding to each processing stage of at least one of the methods described herein. These instructions may be subdivided into subroutines and / or stored in one or more files that may be statically or dynamically linked. Another embodiment of a computer program product includes computer-executable instructions corresponding to each means of at least one of the apparatuses and / or products described herein. These instructions may be subdivided into subroutines and / or stored in one or more files that may be statically or dynamically linked.

[0092] The medium of a computer program may be any entity or device capable of recording a program. For example, the medium encompasses a data storage, e.g. a ROM, e.g. a CD ROM or a semiconductor ROM, or a magnetic storage medium, e.g. a hard disk. Furthermore, the medium may be a transmissible medium, such as an electric or optical signal conveyed via an electric or optical cable or by radio or other means. If the program is embodied in such a signal, the medium may be constituted by such a cable or other device or means. Alternatively, the medium may be an integrated circuit in which the program is embedded, the integrated circuit being adapted to perform or used for the performance of the relevant method.

[0093] Variations of the disclosed embodiments can be understood and realized by those skilled in the art practicing the claimed invention, from a study of the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the singular does not exclude a plurality. A single processor or other unit may fulfill the functions of several items recited in the claims. The mere fact that certain means are recited in mutually different dependent claims does not indicate that a combination of these means cannot be used to advantage. A computer program may be stored / distributed on a suitable medium, such as an optical storage medium or a solid-state medium, provided integrally with or as part of other hardware, but may also be distributed in other forms, for example via the Internet or other wired or wireless telecommunication devices. Any reference signs in the claims should not be interpreted as limiting the scope of the claims.

Claims

1. 1. An apparatus for stroke characterization, said apparatus comprising: receiving image data representative of a three-dimensional image of a subject's head; identifying an area within the image data corresponding to a bone of the head of the subject; applying adjustments to the image data to compensate for effects caused by radiation scattered from the bone during acquisition of the image data; generating a two-dimensional image based on the adjusted image data; and providing the generated two-dimensional image as an input to a predictive model trained to identify regions of interest associated with stroke characterization in the two-dimensional image; obtaining an indication of the identified region of interest as an output of the predictive model; A processor for performing a predictive model that inputs the two-dimensional image and outputs an indication of the identified region of interest; Equipped with the adjustment applied to the image data includes reducing pixel intensities in areas of the subject's head corresponding to areas into which the radiation waves are scattered from the bone during acquisition of the image data, and the adjustment of the pixel intensities includes adjusting the pixel intensities based on a distance of the pixel from a surface of the bone. Device.

2. the processor further disregarding portions of the image data that do not correspond to a brain or that correspond to brain regions that are unlikely to be affected by a stroke-related event due to brain anatomical structure.

2. The apparatus of claim 1.

3. the processor further overlays the image data on the three-dimensional representation of the brain anatomy using a plurality of common locations in the image data and the three-dimensional representation as fiducial markers.

2. The apparatus of claim 1.

4. The processor, for each of a plurality of consecutive slices of the image data; applying a mask to remove the regions of the image data corresponding to bone; defining a boundary surrounding the region of the image data corresponding to a bone; identifying a sub-region of the image data having a maximum number of contiguous pixels within the bounded region; identifying the identified sub-region of the image data corresponding to a brain of the subject; identifying a portion of the image data corresponding to the boundary of the brain of the subject by An apparatus according to any one of claims 1 to 3.

5. The processor, analyzing successive slices of the image data progressing through the head to identify the size of the brain visible in each slice; identifying the particular slice containing image data corresponding to a lower boundary of the subject's brain in response to determining that a change in the size of the brain visible in the particular slice relative to the size of the brain visible in the slice immediately preceding the particular slice is less than a specified threshold; identifying a portion of the image data corresponding to the lower boundary of the brain of the subject by An apparatus according to any one of claims 1 to 4.

6. The processor, performing a maximum intensity projection of the image data through at least one of a coronal plane, an axial plane, and a sagittal plane, the image data being processed to compensate for effects caused by radiation scattered from the bone during acquisition of the image data; generating a two-dimensional image based on the image data; 6. Apparatus according to any one of claims 1 to 5.

7. The processor further comprises: generating a bounding box for presentation around the region of interest in a representation of the subject's brain; An apparatus according to any one of claims 1 to 6.

8. The region of interest includes a region where bleeding occurs, and the processor further comprises: calculating a score indicative of the severity of the bleeding based on the output of the predictive model; An apparatus according to any one of claims 1 to 7.

9. the image data comprises data acquired using a non-contrast computed tomography imaging modality; Apparatus according to any one of claims 1 to 8.

10. the predictive model comprises a convolutional neural network trained to determine whether the identified region of interest is indicative of a bleeding event; Apparatus according to any one of claims 1 to 9.

11. The adjustment, [0050] where y is the pixel intensity and x is the distance from the bone surface.

2. The apparatus of claim 1.

12. 1. A method for stroke characterization in medical image data, said method comprising: receiving image data representative of a three-dimensional image of a subject's head; identifying an area within the image data corresponding to a bone of the head of the subject; applying adjustments to the image data to compensate for effects caused by radiation scattered from the bone during acquisition of the image data; generating a two-dimensional image based on the adjusted image data; providing the generated two-dimensional image as an input to a predictive model trained to identify regions of interest associated with stroke characterization in the two-dimensional image, the predictive model inputting the two-dimensional image and outputting indications of regions of interest; obtaining an indication of the identified region of interest as an output of the predictive model; The method, wherein the adjustment applied to the image data includes reducing pixel intensity in areas of the subject's head corresponding to areas into which the radiation waves are scattered from the bone during acquisition of the image data, and the adjusting of the pixel intensity includes adjusting the pixel intensity based on a distance of the pixel from a surface of the bone.

13. 1. A method of processing image data to train a predictive model to identify regions of interest associated with stroke characterization, the method comprising: For each of the multiple subjects, receiving three-dimensional image data representative of a head of the subject; Pre-processing the three-dimensional image data, applying adjustments to the three-dimensional image data to compensate for effects caused by radiation scattered from bones during acquisition of the three-dimensional image data; and generating a two-dimensional image based on the three-dimensional image data; and preprocessing the image by preparing a training data set according to The method, wherein the adjustment applied to the three-dimensional image data includes a reduction in pixel intensity in areas of the subject's head corresponding to areas into which the radiation waves are scattered from the bone during acquisition of the three-dimensional image data, and the adjustment of the pixel intensity includes adjusting the pixel intensity based on a distance of the pixel from a surface of the bone.

14. training the predictive model using the training data set to identify regions of interest in the two-dimensional images relevant to stroke characterization; The method of claim 13 further comprising:

15. A non-transitory computer readable medium having computer readable code embodied therein, said computer readable code, when executed by a suitable computer or processor, causing said computer or processor to perform the method of any one of claims 12 to 14.

Images