Image processing for stroke characterization

Abstract

The present invention discloses an apparatus for stroke characterization. The apparatus includes a processor. The processor is configured to: receive image data representing a three-dimensional image of a subject's head; identify a region within the image data corresponding to a bone in the subject's head; apply an adjustment to the image data to compensate for effects caused by radiation scattered from the bone 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 prediction model to identify a region of interest in the two-dimensional image for stroke characterization.

Description

Technical Field

[0001] The present invention relates to stroke characterization, and more particularly to image processing for identifying regions of interest for stroke characterization. An apparatus, method, and computer program product for stroke characterization are disclosed. Background Art

[0002] Stroke is a medical condition in which poor blood flow to the human brain leads to cell death. Two main types of stroke may occur: ischemic (due to lack of blood flow) and hemorrhagic (due to bleeding). Patients who suffer from hemorrhagic stroke may also suffer from intracerebral hemorrhage (ICH), also known as cerebral hemorrhage, which is a type of intracranial hemorrhage that occurs in people's brain tissue or cerebral ventricles. The type of treatment provided to a person suffering from a stroke depends on the type of stroke, the cause of the stroke, and the affected brain part. Typically, medications (such as medications that prevent and dissolve blood clots or medications that reduce bleeding) will be used to treat the person suffering from a stroke. In some cases, surgery may also be required to treat the person suffering from a stroke.

[0003] Every minute that a stroke is not treated will cause the death of approximately 2 million neurons in the brain, causing part of the brain to malfunction. The faster a person receives treatment for a stroke, the less brain damage that may occur. Therefore, it is desirable to detect the occurrence of a stroke as quickly as possible. For example, it would be 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 when a patient encounters a medical facility (e.g., in an emergency room); by quickly diagnosing a stroke, the necessary treatment of the patient can be started quickly, thereby providing the patient with a more favorable medical outcome. Summary of the Invention

[0004] The inventors of the present invention have recognized the need to be able to provide rapid and reliable diagnoses regarding patients presenting with potential symptoms of a stroke. Accordingly, the present invention provides a mechanism by which three-dimensional image data can be used for stroke characterization. More specifically, three-dimensional image data (e.g., a three-dimensional scan of a patient's head) can be preprocessed so that it can be provided as input to a predictive model (e.g., a machine learning model) that is trained to identify regions of interest that may be relevant to and used for stroke detection and characterization. The preprocessing performed on the three-dimensional image data effectively cleans the image data and converts the three-dimensional image into a two-dimensional image while retaining many, if not all, of the details that are useful for detecting and characterizing a 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 head of an object; identify an area within the image data corresponding to a bone in the head of the object; apply adjustments to the image data to compensate for effects caused by radiation scattered from the bone 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 to identify a region of interest in the two-dimensional image for stroke characterization.

[0006] By processing the image data in this manner (i.e., by adjusting the image data to compensate for radiation scatter from bone), the prediction model can more efficiently analyze the image data and produce a more accurate output. Specifically, any increase in pixel intensity in the image data caused by radiation scatter can be compensated (reduced or removed) to reduce the likelihood that the prediction model will misinterpret or mischaracterize the pixel intensity as being associated with bleeding within the brain. Thus, the likelihood of a false positive characterization (e.g., a stroke diagnosis) is reduced.

[0007] By performing maximum intensity projection on the three-dimensional image data, a two-dimensional image can be generated that retains detail from the three-dimensional image that can be used to detect and / or characterize stroke in the image data. Furthermore, by processing the three-dimensional data before performing the maximum intensity projection, only the most relevant portions of the image data are analyzed, resulting in highly accurate predictions (e.g., identifying regions of interest). The combination of this processing and the maximum intensity projection means that any regions of interest (such as bleeding events) are clearly visible and identifiable in the generated two-dimensional image, allowing the output of the prediction model to be regularly reviewed and confirmed by medical professionals.

[0008] In some embodiments, the processor may be further configured to register the image data to the three-dimensional representation of the brain using a plurality of fiducial landmarks common to the image data and the three-dimensional representation.

[0009] In some embodiments, the processor can be configured to identify, for each of multiple consecutive slices of the image data, a portion of the image data corresponding to the boundary of the subject's brain by: applying a mask to remove the region in the image data corresponding to the bone; defining a boundary around the region in the image data corresponding to the bone; identifying a sub-region of the image data having a maximum number of contiguous pixels within the bounded region; and determining the identified sub-region in the image data corresponding to the subject's brain.

[0010] In some embodiments, the processor can be configured to identify the portion of the image data corresponding to the lower boundary of the subject's brain by: analyzing the image data as it progresses through successive slices of the head to determine a measure of the brain visible in each slice; and determining that the particular slice includes image data corresponding to the lower boundary of the subject's brain in response to determining that a change in the measure of the brain visible in the particular slice relative to a measure of the brain visible in a slice immediately preceding the particular slice is below a defined threshold.

[0011] In some embodiments, the processor may be 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 may be further configured to obtain an indication of the identified region of interest as an output of the predictive model; and generate a bounding box around the region of interest for presentation in the representation of the brain of the subject.

[0013] In some embodiments, the region of interest may include an area where bleeding has occurred.The processor may be further configured to (e.g., automatically) calculate a score indicating the severity of the bleeding based on the output of the prediction model.

[0014] In some embodiments, the image data may include data acquired using a non-contrast computed tomography imaging modality.

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

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

[0017] Wherein y is the pixel intensity, and wherein x is the distance from the surface of the bone in millimeters.

[0018] 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 head of an object; identifying an area within the image data corresponding to a bone in the head of the object; applying an adjustment 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 input to a predictive model to identify a region of interest in the two-dimensional image for stroke characterization.

[0019] According to a third aspect, the present invention provides a method of processing image data for use in stroke characterization, the method comprising preparing a training dataset for each of a plurality of subjects by: receiving three-dimensional image data representing the subject's head; and pre-processing the three-dimensional image data by: applying adjustments to the three-dimensional image data to compensate for effects caused by radiation scattered from the bone during acquisition of the three-dimensional image data; and generating a two-dimensional image based on the three-dimensional image data.

[0020] In some embodiments, the method may further include using the training dataset to train a prediction model to identify a region of interest for stroke characterization in the two-dimensional image.

[0021] According to a fourth aspect, the present invention provides a computer program product comprising a non-transitory computer-readable medium having computer-readable code embodied therein, the computer-readable code being configured such that, when executed by a suitable computer or processor, the computer or processor performs the steps of the method disclosed herein.

[0022] These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] For a better understanding of the invention and in order to show more clearly how it may be carried into effect, reference will now be made, by way of example only, to the accompanying drawings, in which:

[0024] Figure 1 is a schematic illustration of an example of an apparatus according to various embodiments disclosed herein;

[0025] Figure 2 is a schematic illustration of an example of a prediction model for use according to various embodiments disclosed herein;

[0026] Figure 3 is a flow chart of an example of a process performed according to various embodiments disclosed herein;

[0027] Figure 4 is a flow chart of another example of a process performed according to various embodiments disclosed herein;

[0028] Figure 5 is a diagram showing the effect of radiation scattering;

[0029] Figure 6 is a flow chart of an example of a method according to various embodiments disclosed herein;

[0030] Figure 7 is a flowchart of another example of a method according to various embodiments disclosed herein; and

[0031] Figure 8 is a schematic illustration of a computer-readable medium in communication with a processor. DETAILED DESCRIPTION

[0032] The examples described herein provide apparatus and methods that can be used to characterize stroke in medical images. A three-dimensional image of a subject's head can be acquired, for example, using known medical imaging techniques, and processed as described herein such that the processed image can be provided as input to a trained predictive model (e.g., a neural network or deep learning model) for analysis to detect and / or characterize physical signs indicating that the subject has suffered a stroke. By processing the three-dimensional data in the manner disclosed herein, the data provided to the trained predictive model is "cleaner" and is more likely to produce a highly accurate output. The ability to obtain accurate output is particularly important in the medical field, where rapid diagnosis means that appropriate measures can be taken in a timely manner. According to various embodiments disclosed herein, the present invention also provides a method of processing image data for use in stroke characterization, such as preparing a training data set that can be used to train a predictive model for stroke characterization.

[0033] Figure 1 FIG2 shows a block diagram of an apparatus 100 that can be used for stroke characterization according to an embodiment. For example, the apparatus 100 can be used to process data to be used for stroke identification or stroke characterization. Figure 1 , the apparatus 100 includes a processor 102 that controls the operation of the apparatus 100 and may implement the methods described herein.

[0034] The apparatus 100 may also include a memory 106 including instruction data representing a set of instructions. The memory 106 may be configured to store instruction data in the form of program code that can be executed by the processor 102 to perform the methods described herein. In some embodiments, the instruction data may include multiple software and / or hardware modules that are each configured to perform or be used to perform a single or multiple steps of the methods described herein. In some embodiments, the memory 106 may be part of a device that also includes 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 separate from the other components of the apparatus 100.

[0035] In some embodiments, the memory 106 may include multiple sub-memories, each of which is capable of storing a piece of instruction data. In some embodiments in which the memory 106 includes multiple sub-memories, the instruction data representing a set of instructions may be stored in a single sub-memory. In other embodiments in which the memory 106 includes multiple sub-memories, the instruction data representing a set of instructions may be stored at multiple sub-memories. For example, at least one sub-memory may store instruction data representing at least one instruction in the set of instructions, while at least one other sub-memory may store instruction data representing at least one other instruction in the set of instructions. Therefore, according to some embodiments, instruction data representing different instructions may be stored at one or more different locations in the device 100. In some embodiments, the memory 106 may be used to store information, data (e.g., image data), signals, and measurements collected or performed by the processor 102 of the device 100 or from any other component of the device 100.

[0036] The processor 102 of the device 100 can be configured to communicate with the memory 106 to execute a set of instructions. When executed by the processor 102, the set of instructions can cause the processor 102 to perform the methods described herein. The processor 102 may include 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 may include multiple (e.g., interoperable) 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 can be located in different locations and can perform different steps of the methods described herein and / or different parts of a single step.

[0037] Return to Figure 1In some embodiments, apparatus 100 may include at least one user interface 104. In some embodiments, user interface 104 may be part of a device that also includes one or more other components of apparatus 100 (e.g., processor 102, memory 106, and / or one or more other components of apparatus 100). In alternative embodiments, user interface 104 may be part of a device that is separate from the other components of apparatus 100.

[0038] The user interface 104 can be used to provide information generated by the method according to the embodiments of the present invention to the user of the device 100 (e.g., a medical professional such as a radiologist or any other user). When executed by the processor 102, the set of instructions can cause the processor 102 to control one or more user interfaces 104 to provide information generated by the method according to the embodiments of the present invention. Alternatively or additionally, the user interface 104 can be configured to receive user input. In other words, the user interface 104 can allow the user of the device 100 to manually input instructions, data or information. When executed by the processor 102, the set of instructions can cause the processor 102 to obtain user input from one or more user interfaces 104.

[0039] The user interface 104 may be any user interface that enables information, data, or signals to be presented (or output or displayed) to a user of the device 100. Alternatively or additionally, the user interface 104 may be any user interface that enables a user of the device 100 to provide user input, interact with the device 100, and / or control the device 100. For example, the user interface 104 may include one or more switches, one or more buttons, a keypad, a keyboard, a mouse, a mouse wheel, a touch screen or application (e.g., on a tablet computer or smartphone), a display screen, a graphical user interface (GUI) or other visual presentation component, one or more speakers, one or more microphones or any other audio component, one or more lights, a component for providing tactile feedback (e.g., a vibration function), or any other user interface, or combination of user interfaces.

[0040] In some embodiments, as Figure 1 As shown in , apparatus 100 may also include a communication interface (or circuitry) 108 for enabling apparatus 100 to communicate with an interface, memory, and / or device that is part of apparatus 100. Communication interface 108 may communicate with any interface, memory, and device wirelessly or via a wired connection.

[0041] It should be understood that Figure 1Only components necessary to illustrate this aspect of the present disclosure are shown, and in actual implementations, the device 100 may include additional components beyond those shown. For example, the device 100 may include a battery or other power source for powering the device 100, or means for connecting the device 100 to a mains power source.

[0042] The apparatus 100 can be used for stroke characterization. More specifically, the apparatus 100 can be used to process image data (e.g., medical image data) and can therefore be referred to as a medical image processing apparatus. According to embodiments disclosed herein, the processor 102 is configured to receive image data representing a three-dimensional image of a subject's head. While in some examples, the image data may represent the subject's entire head, in other examples, the image data may represent only a portion of the subject's head. Typically, for stroke characterization, the subject's brain (or at least a large portion thereof) is imaged so that events occurring within the brain that may indicate the occurrence of a stroke can be detected and analyzed.

[0043] The image data may, for example, include data acquired using a non-contrast computed tomography imaging modality. For example, the image data may include three-dimensional (e.g., volumetric) image data acquired from a non-contrast computed tomography (CT) scan. In one example, the image data may be acquired using a low-dose computed tomography (low-dose CT or LD CT) imaging modality. In other examples, other imaging modalities may be used to acquire the three-dimensional image data. Typically, an imaging modality involves directing electromagnetic radiation toward an object to be imaged and detecting the interaction of the electromagnetic radiation with the object. The type of radiation used generally depends on the imaging modality.

[0044] Processor 102 is configured to identify regions within the image data corresponding to bones in the subject's head.The regions within the image data corresponding to bones may include regions corresponding to the subject's skull or a portion of the skull.

[0045] Once processor 102 has identified regions in the image data corresponding to bone, processor 102 is configured to apply adjustments to the image data to compensate for effects caused by radiation scattered from the bone during acquisition of the image data. As described above, electromagnetic radiation is directed toward a subject during imaging. The behavior of the radiation during imaging (e.g., absorption by matter in the subject's head) enables the subject to be imaged. However, when the radiation encounters bone (e.g., the skull), it reflects and scatters from the bone's surface. The scattered radiation is detected by the imaging device and may be visible in the image data. The scattered radiation (or its visible effects) may obscure other objects in the subject's head, making it difficult to clearly see objects or events occurring within the subject's brain. Furthermore, the effects of the scattered radiation (e.g., its appearance in an image of the subject's brain) may resemble characteristic events indicative of a stroke. Therefore, in an image (e.g., a 3D scan) of the subject's brain, radiation scattered from the inner surface of the skull may appear as events associated with a stroke (e.g., a hemorrhagic event), leading to a misdiagnosis, such as an erroneous determination that the subject has suffered a stroke. Therefore, adjustments applied to the image data are intended to compensate for scattered radiation so that events (eg, an increase in intensity of a group of pixels in an image) are not misinterpreted as events characteristic of a stroke.

[0046] The processor 102 is further configured to generate a two-dimensional image based on the adjusted image data. In some embodiments, as discussed in more detail below, this may be accomplished by performing a maximum intensity projection of the adjusted three-dimensional image data.

[0047] The processor 102 is also configured to provide the generated two-dimensional image as an input to a prediction model to identify a region of interest in the two-dimensional image for stroke characterization. The prediction model may include a model to be executed using the processor 102. Alternatively, a different processor (e.g., a processor external to and / or remote from the device 100) may be used to execute the prediction model. The prediction model, which may include an artificial neural network or a classifier (discussed in more detail below), may be trained to analyze the two-dimensional input image so as to identify portions of the image related to stroke characterization (i.e., regions of interest) based on a set of features. For example, the prediction model may be trained to identify regions in the image that indicate hemorrhage. Such regions in the image may not be easily identified by humans (e.g., radiologists), but may be more recognizable using a trained prediction model.

[0048] In some examples, the predictive model may include an artificial neural network. Artificial neural networks, or simply neural networks, and other machine learning models will be familiar to those skilled in the art, but in short, 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 consists of layers, each layer including a plurality of neurons. Each neuron includes a mathematical operation. In the process of classifying a portion of the data, the mathematical operation of each neuron is performed on that portion of the data to produce a numerical output, and the output of each layer in the neural network is fed sequentially to the next layer. Typically, the mathematical operation associated with each neuron includes one or more weights that are tuned during the training process (e.g., the values of the weights are updated during the training process to tune the model to produce more accurate classifications).

[0049] For example, in a neural network model used to classify the content of an image, each neuron in the neural network may include a mathematical operation comprising a weighted linear sum of the pixel (or, in three dimensions, voxel) values in the image, followed by a nonlinear transformation. Examples of nonlinear transformations used in neural networks include the sigmoid function, the hyperbolic tangent function, and the rectified linear function. Neurons in each layer of the neural network typically comprise a differently weighted combination of a single type of transformation (e.g., the same type of transformation, sigmoid, etc., but with different weights). As those skilled in the art are familiar with, in some layers, each neuron may apply the same weight in the linear sum; this applies, for example, to convolutional layers. The weights associated with each neuron can make certain features more prominent (or conversely, less prominent) than other features during the classification process, and thus, adjusting the neuron weights during training trains the neural network to place increased importance on specific features when classifying images. In general, a neural network may have weights associated with neurons and / or weights between neurons (e.g., which modify the data values passed between neurons).

[0050] As briefly noted above, in some neural networks (such as convolutional neural networks), lower layers (such as input layers or hidden layers) in the neural network (i.e., layers toward the beginning of a series of layers in the neural network) are activated by (i.e., their outputs depend on) small features or patterns in the portion of the data being classified, while higher layers (i.e., layers toward the end of the series of layers in the neural network) are activated by increasingly larger features in the portion of the data being classified. As an example, where the data comprises images, lower layers in the neural network are activated by small features (e.g., such as edge patterns in the image), mid-level layers are activated by features in the image (e.g., larger shapes and forms), and layers closest to the output (e.g., upper layers) are activated by entire objects in the image.

[0051] Typically, the weights of the final layer of a neural network model (called the output layer) depend most strongly on the specific classification problem being solved by the neural network. For example, the weights of the outer layers may depend largely on whether the classification problem is a localization problem or a detection problem. The weights of the lower layers (e.g., the input layer and / or the hidden layers) tend to depend on the content (e.g., features) of the data being classified, and therefore it has been recognized herein that the weights in the input layer and hidden layers of neural networks processing the same type of data can, with sufficient training, converge toward the same values over time, even if the outer layers of the model are tuned to solve different classification problems.

[0052] In some examples, the prediction model may include a convolutional neural network that is trained to determine whether the identified region of interest indicates a bleeding event. Such a prediction model can be trained on features associated with bleeding so that the prediction model can identify areas where bleeding has occurred or areas of the brain that have been affected by bleeding, even though a human observer (e.g., a radiologist) may not immediately determine the occurrence of bleeding by simply viewing the image data. In one specific example, a convolutional neural network model called Inception-ResNet can be employed as the prediction model. The Inception-ResNet model includes approximately 780 layers and is Figure 2 A compressed representation of the model is shown schematically in . Figure 2 The illustrated model 200 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 a residual network layer 216. Other examples of models suitable for use with the methods disclosed herein may include fewer or additional layers and / or functions.

[0053] Now refer to Figure 3 and 4 Examples of image data processing methods that may be performed to aid in detecting and / or characterizing stroke in medical image data acquired about a subject are discussed. Figure 3 is a flow chart of an example workflow that may be performed with respect to three-dimensional image data acquired for a subject. The three-dimensional image data may include a three-dimensional CT volume of the subject's head. For example, the subject may be suffering from symptoms indicative of a stroke and, therefore, may be referred to a medical facility (e.g., a hospital) for further investigation. The subject's head may be scanned using an imaging modality such as 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 a plane (e.g., an axial plane, a coronal plane, and / or a sagittal plane), and processing of the three-dimensional image data typically includes processing each of the slices in turn. Thus, in Figure 3In the illustrated workflow, three-dimensional image data 300 includes multiple slices of image data. At block 302, the i-th slice of the multiple slices is read. At block 304, the i-th slice undergoes preprocessing, discussed in greater detail below, and once the i-th slice has been preprocessed, the workflow proceeds to block 306, where the next (i+1)th slice is obtained from the three-dimensional image data 300 for processing. Thus, after preprocessing of the slice of image data is complete, the workload involves acquiring the next slice, and the workflow returns to block 302.

[0054] The workflow continues until all slices containing image data relevant to the subject's brain have been preprocessed. An example of a stopping criterion is discussed below, which is considered satisfied once all image slices containing image data representing the brain have been processed. Once all relevant slices have been processed, the workflow proceeds to block 308, where a two-dimensional image is generated based on the preprocessed image data. In some examples, multiple two-dimensional images may be generated, as discussed below. The two-dimensional image(s) may then be provided as input to a trained prediction model for analysis, which may include identifying and / or characterizing a region of interest indicative of a stroke.

[0055] Figure 4 is a part of the pre-processing that can form each slice of three-dimensional image data (e.g., Figure 3 306 of the present invention. In general, pre-processing may include identifying (block 402) a region within a slice of image data that represents a subject's bone (e.g., a skull), identifying (blocks 406 and 408) a portion of a mask that corresponds to a portion of the bone that encapsulates the subject's brain, performing an adjustment of the image data (block 410) to compensate for the effects of radiation scattered from the bone when acquiring the three-dimensional image data, and extracting (block 412) a portion of the image data representing the user's brain from the image data. In some examples, only the scattered radiation compensation adjustment (block 410) may be performed during pre-processing of the image data. In other examples, one or more of the other processes discussed herein may also be performed.

[0056] To help identify bones within the image data, preprocessing of the image data can first include registering or aligning the three-dimensional image data with a reference representation (e.g., an image or volume) to achieve the expected orientation of the image data and ensure consistency between the orientations of the slices of the image data. Registration of the three-dimensional image data is performed with respect to a three-dimensional volume (including slices of all captured image data). In other words, registration is performed with respect to the volume of the image data as a whole rather than with respect to each individual in the slice. In some embodiments, registration of the image data can include, for example, non-rigid registration of the image data to a three-dimensional representation of the human brain using an anatomical atlas. Registration of image data to an anatomical atlas can be achieved using multiple fiducial landmarks, each of which includes features in the brain, such as a ventricle or cerebellum.

[0057] In some embodiments, machine learning techniques (such as trained prediction models) can be used to enable identification of fiducial landmarks in the image data used in the registration process. For example, an artificial neural network or a support vector machine can be used. The prediction model can be trained on one or more features (such as pixel intensity, entropy, gray-level co-occurrence matrix (GLCM), and texture features). For example, the prediction model can determine that a region of interest within the three-dimensional image data is likely (within a defined certainty threshold) to represent a particular anatomical structure within 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 can be labeled with an identifier, such as the name of the anatomical structure to which it is associated). A voxel can be considered to fall within a region of interest if it falls within a sphere of a defined radius formed around the center of the region of interest.

[0058] When classifying a particular voxel as forming part of a region of interest, a majority voting process can be performed such that once all voxels within a particular region of interest have been labeled as forming part of the region of interest, the majority vote label is assigned to the region of interest. Multiple labeled regions of interest are then aggregated to define a unique fiducial landmark (e.g., formed from a set of individual fiducial landmarks), which is used to register and align the image data with a reference representation (e.g., from an anatomical atlas).

[0059] The registration process discussed above can be performed using processor 102. Thus, in some embodiments, processor 102 can be configured to register the image data to the three-dimensional representation of the brain using a plurality of fiducial landmarks common to the image data and the three-dimensional representation. In some examples, the plurality of fiducial landmarks can be identified using a trained classifier or predictive model, as discussed above.

[0060] While registering the image data with a reference representation helps ensure that the image data is properly aligned for further processing, the registration process is not required and, therefore, may be omitted from the processing procedure.

[0061] As mentioned above, the registration of the image data is performed using the entire 3D volume, however, Figure 4 The processes in blocks 402 through 412 are performed with respect to individual slices of image data. Typically, processing of slices of image data is performed using axial slices (i.e., slices through an axial plane of the subject's head). However, it should be understood that in some examples, processing may be performed with respect to slices of image data through other planes (e.g., through a coronal or sagittal plane). Figure 4 The image shown in the box of is merely an example showing how the image may change when pre-processing is performed.

[0062] In order to remove those portions of the image data that are not used in the analysis process, various bone-skull stripping techniques familiar to those skilled in the art may be used. Figure 4 In block 402 of the flowchart in FIG, those regions within the slice of image data representing bone are identified. In one embodiment, machine learning techniques can be used to identify bone regions within the image data, such as a predictive model trained to identify portions of image data representing bone. In some embodiments (and as part of the machine learning techniques), bone regions in the image data can be identified by analyzing structural patterns of bones (e.g., skull bones) in the image data. Boundary parameters and / or intensity values of the image data can be considered to determine the portion of the image representing bone. In addition, one or more filters can be applied to identify and / or extract bone regions from the image. For example, a filter can be applied to remove portions of the image that fall outside a defined range on the Heinz scale. The Heinz scale is a quantitative scale for describing radiodensity. Thus, in one example, pixels corresponding to regions having a Heinz unit (HU) measurement that falls outside the range of 0 to 100 can be removed from the image data. It should be understood that regions within the range of 0 to 100 HU represent regions of blood and brain tissue that are to be retained in the image data.

[0063] Once the skeleton regions have been identified, a mask is generated for the identified regions ( Figure 4404 of the flowchart in FIG). A 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 general mesh is provided and deformed to align with the skull-brain interface; that is, the inner boundary of the bone region identified in the slice of image data. The mesh is then geometrically modified using an affine transformation by minimizing the sum of the squares of the distances between the mesh vertices and the bone regions in the image data. The mesh can then be deformed using a coarsely segmented image (that is, an image segmented using anatomical landmarks) and also using the original image data. The resulting modified mask is intended to capture all bones / skull contained within the slice of image data, as shown in the example in block 404.

[0064] As will be apparent from the example shown in block 404, the generated mask includes regions representing portions of the skull that enclose the subject's brain, as well as other regions representing other portions of the skull, such as the eye sockets and / or jaw. Figure 4 At blocks 406 and 408 of the flowchart in FIG. , preprocessing involves identifying portions of the mask corresponding to portions of the bone that encapsulate the subject's brain. At block 406, a contour line 406a is formed around the entire bone region. In some examples, the contour line 406a can be formed using a Bresenham line algorithm based on convex boundary points of the skull. As shown at 406b, the contour line surrounds the identified bone region in the slice of image data.

[0065] At box 408, the area within the mask corresponding to the subject's brain is extracted. To achieve this, those areas of the mask that correspond to the eye sockets and / or jaw of the subject's skull can be removed. In some embodiments, this can be achieved by "flood filling" (i.e., extracting) those areas that fall outside the contour line 406a, for example, as shown in image 408a. In order to remove smaller portions of the mask that do not represent the subject's brain area, connected component analysis can be used. In this regard, connected component analysis involves identifying the largest connected component within the mask; in other words, identifying the portion of the mask that has the largest number of connected (i.e., continuous) pixels. Those 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 area corresponding to the subject's brain and the area corresponding to the subject's eye sockets.

[0066] In order to remove those portions of the mask corresponding to the object's eye sockets, a morphological erosion and / or morphological dilation operation may be performed. In some embodiments, a morphological erosion operation (using the structuring element S e ), followed by a morphological dilation operation (using the structuring element S d )), where S d >S eAfter the morphological erosion and dilation operations, the original interior mask (i.e., the mask before any morphological operations have been performed) is used. mask A morphological "AND" operation is performed to preserve the boundaries of the brain in the mask. The "AND" operation can be defined as: As shown in image 408c, the resulting mask includes only the portion corresponding to the subject's brain and may be referred to as a brain mask.

[0067] Can be Figure 1 The processor 102 performs identification of portions of the mask corresponding to portions of the bone enclosing the subject's brain ( Figure 4 4 and 5. The above process may be considered to identify a boundary of the subject's brain (e.g., a boundary between the brain and the skull). Thus, in some embodiments, processor 102 may be configured to identify a portion of the image data corresponding to a boundary of the subject's brain by, for each of a plurality of successive slices (e.g., axial slices) of the image data: applying a mask to remove regions in the image data corresponding to bone; defining a boundary around the regions in the image data corresponding to bone; identifying a sub-region of the image data having a greatest number of contiguous pixels within the defined region; and determining that the identified sub-region of the image data corresponds to the subject's brain.

[0068] Once a brain mask has been created or generated, it can be used in Figure 4 Radiation scatter compensation is performed at box 410 of the flowchart of . As described above, during imaging (e.g., a CT scan), electromagnetic radiation (e.g., X-rays) is directed at the object being imaged (e.g., the head of the object). Bones (e.g., the skull) within the object can cause the electromagnetic radiation to be scattered, and this scattering can be visible in the resulting image data. Scattered radiation can manifest itself as relatively bright (e.g., high pixel intensity) patches in the image data, and in some cases, these bright patches can resemble hemorrhagic events. Therefore, bright patches in the image data caused by radiation scatter from the bone can be misinterpreted as hemorrhagic events, and therefore, the object can be misdiagnosed as having suffered a hemorrhage (i.e., a stroke), resulting in a false-positive diagnosis. Corrections or adjustments can be made in the image data to compensate for the effects caused by radiation scattered from the bone during the acquisition of the image data.

[0069] To compensate for radiation scattered from bone, a scatter profile can first be determined. The scatter profile defines the variation in pixel intensity in the image data as a function of distance from the bone surface from which radiation is scattered. In this example, the scatter profile is defined based on the variation in intensity as a function of distance from the inner surface of the subject's skull toward the brain (i.e., the intracranial space). Figure 5 An example of the scattering profile of radiation scattered from the inner surface of the skull is shown. Figure 5In , image 502 of a slice of image data includes an area within the inner surface of the skull, where an increase in pixel intensity has occurred due to radiation scattered from the skull during image acquisition. Due to the radiation scattering, the pixel intensity gradually decreases from high (in the skull area) to low (in the brain area). The increased pixel intensity is shown in a magnified image 504 of a portion of image 502. A graph 506 shows how the pixel intensity y varies as a function of the distance x (in mm) from the skull surface. Specifically, graph 506 shows the peak pixel intensity at a location corresponding to the skull, where the pixel intensity decreases logarithmically with distance from the skull. According to some examples, a logarithmic intensity compensation model can be used, which is based on the Maxwell Boltzmann distribution of an increasing entropy system. In such a model, the decay of pixel intensity is modeled with respect to increasing entropy. Increasing entropy can be considered equivalent to increasing distance from the skull surface to the intracranial space.

[0070] In one example, a model defining pixel intensity y as a function of distance x (in millimeters) from the skull surface can be expressed as:

[0071]

[0072] Once the additional pixel intensities resulting from the scattered radiation have been calculated, a correction may be applied to the image data to compensate for the increased pixel intensities. The adjustment or correction may be applied by processor 102. Thus, in some embodiments, the adjustment applied to the image data may include reducing pixel intensities in regions corresponding to regions of the subject's head into which radiation was scattered from bone during acquisition of the image data.

[0073] Once the radiation scatter compensation portion of the preprocessing has been performed (block 410), a brain mask (i.e., the portion of the skull mask corresponding to the region of the subject's brain) may be applied to the image data using known techniques. Any portion of the image data that falls outside the brain mask is extracted, removed, or ignored, leaving only the portion of the image data corresponding to the subject's brain. As described above, a reference image is performed for each slice of the image data. Figure 4 Thus, having completed the preprocessing process for a particular slice, the next slice in the three-dimensional image data is acquired and preprocessed in the same manner.

[0074] The three-dimensional image data may include image data slices that do not depict portions of the brain. The three-dimensional image data may also include image data corresponding to portions of the brain that are less likely to be affected by a stroke. For example, one or more slices may have been captured below the lower border of the brain. For the purpose of identifying regions of interest relevant to stroke characterization, only those slices of image data depicting the brain are preprocessed and analyzed. Therefore, according to some embodiments, preprocessing of image data slices may be stopped when a specific stopping criterion has been met. Thus, the processor 102 may be configured to identify a portion (e.g., a slice) of image data that corresponds to the lower border of the subject's brain. In one example, whether to preprocess the next image data slice in a stack / group of image data slices may be determined based on the proportion of the image data slices representing the subject's brain. For example, if the amount of "brain per slice" (e.g., area) decreases from one slice to the next by more than a defined threshold, it may be determined that no additional slices are to be preprocessed. In this case, it may be determined that the slices comprising the smaller amount of brain are slices captured near the base of the brain. This region of the brain is less likely to be affected by events associated with a stroke (e.g., a hemorrhagic event), and therefore, preprocessing slices below this region is unnecessary.

[0075] In one example, the stopping criterion discussed above may be defined as:

[0076] if Preprocessing is then stopped at the i-th slice. n is a defined threshold, and the value of n can be selected based on the expected accuracy. In some examples, n can be between 0.5 and 0.8. In some examples, n=0.7. In other words, if the amount of brain per slice in a particular slice (e.g., the area visible in the slice) is less than 70% of the amount of brain per slice in the slice immediately preceding the particular slice, then the particular slice is the last slice to be preprocessed; otherwise, the next slice in the stack of slices is processed. Any slice that is not to be preprocessed can be considered irrelevant for the purpose of stroke characterization and is therefore removed or ignored. For example, image data or slices that do not correspond to the brain or to areas of the brain that are unlikely to be affected by stroke-related events are discarded or ignored.

[0077] By using the above stopping criteria, the lower boundary of the subject's brain can be identified within the image data. This can be determined by Figure 1102. Thus, in some embodiments, processor 102 may be configured to identify a portion of the image data corresponding to a lower boundary of the subject's brain by analyzing successive slices of the image data progressing downward through the head to determine a measure (e.g., area or volume) of the brain visible in each slice; and in response to determining that a change in the measure of the brain visible in a particular slice relative to a measure of the brain visible in a slice immediately preceding the particular slice is below a defined threshold, determining that the particular slice includes image data corresponding to the lower boundary of the subject's brain. In some examples, successive axial slices of the image data progressing downward through the subject's head may be analyzed.

[0078] Once a complete set of relevant slices has been preprocessed (e.g., using the preprocessing techniques discussed above), a two-dimensional image can be generated based on the processed three-dimensional image data. In one example, a maximum intensity projection (MIP) of the image data can be performed to generate the two-dimensional image. Those skilled in the art of image analysis will be familiar with maximum intensity projection. Applying maximum intensity projection to a series of slices (e.g., slicing image data representing the brain) involves projecting those voxels with maximum intensity onto a projection plane, where the maximum intensity will intercept parallel rays traced from the viewpoint to the projection plane. The output of this process is a two-dimensional projection or image showing those objects that appear in the image data with maximum intensity. Thus, by applying the MIP technique, all relevant data from the original three-dimensional image data is retained and represented in the two-dimensional image. Notably, any information in the original three-dimensional image data that could be used to determine whether bleeding has occurred is not lost.

[0079] In some examples, multiple maximum intensity projections may be performed to obtain multiple two-dimensional images. For example, a first MIP may be performed in the axial plane, a second MIP may be performed in the coronal plane, and a third MIP may be performed in the sagittal plane. In the two-dimensional images obtained from the MIP in the coronal plane and the MIP in the sagittal plane, image interpolation may be performed to compensate for the thickness of the image data slices. For example, if the slice thickness is increased in the coronal and sagittal planes, discontinuities between slices may occur when performing the MIP operation. Therefore, interpolation is performed to ensure continuity between slices with smooth transitions.

[0080] Therefore, in some embodiments, the processor 102 ( Figure 1 ) can be configured to generate a two-dimensional image based on the image data by performing maximum intensity projection of the image data through at least one of the coronal plane, the axial plane, and the sagittal plane.

[0081] The generated two-dimensional image (or images) can be provided as input (or inputs) to a prediction model for analysis. As described above, the prediction model can be trained to identify regions of interest in the two-dimensional image that are associated with stroke characterization. In some embodiments, the processor 102 can be configured to obtain an indication of the identified regions of interest as an output of the prediction model. For example, the output of the prediction model can include an indication, such as a visual or textual indication, that a particular region in the image data is considered to be a region of interest associated with the identification and / or characterization of a stroke. In some embodiments, the processor 102 can be configured to provide the generated two-dimensional image as input to a prediction model (e.g., the same prediction model or a different prediction model) to determine whether the identified regions of interest indicate a hemorrhagic event. If it can be determined that the image data does not include a region of interest or that the region of interest does not indicate a hemorrhagic event, then it can be concluded that the subject is unlikely to have suffered a stroke and, therefore, does not require emergency treatment for a stroke. On the other hand, if it can be determined that the region of interest indicates a stroke-related event (such as a hemorrhage), then appropriate treatment can be administered to improve the subject's long-term health suspicion. Therefore, in some embodiments, the processor 102 can be configured such that, in response to determining that the identified region of interest indicates a hemorrhagic event, the processor generates an alarm signal. The alarm signal may include, for example, an audible alarm or a visual alarm presented on the user interface 104 .

[0082] In some embodiments, the processor 102 may be 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 may be drawn to be displayed around the relevant region of interest on a two-dimensional image or in some representation of the subject's brain representation, enabling a human observer (e.g., a medical professional such as a radiologist) to identify the location of the region of interest in the brain relative to other parts of the brain. This may also help provide confidence in the predictive model to the human observer, as the human observer can verify the accuracy of the output.

[0083] In some embodiments, where the three-dimensional image data has been registered to an anatomical atlas as described above, knowledge of the various anatomical objects within the brain can be used to provide richer information to a human observer. For example, the anatomical portion or object closest to the region of interest within the brain can be determined. Thus, in some embodiments, the processor 102 can be configured to provide an indication of the anatomical portion of the brain of the object corresponding to the identified region of interest for presentation. For example, if the region of interest is evidence of a hemorrhagic event that has occurred in the cerebellum, an indication can be provided (e.g., presented) on a representation of the subject's brain that indicates the cerebellum as the portion of the brain where the hemorrhage has occurred.

[0084] Bleeding that occurs in the brain is sometimes referred to as intracerebral hemorrhage (ICH). The severity of intracerebral hemorrhage can be measured and defined using an intracerebral hemorrhage score or ICH score. In some embodiments, the region of interest identified by the prediction model can include an area where bleeding has occurred. In such an embodiment, the processor 102 can be configured to calculate a score indicating the severity of the hemorrhage based on the output of the prediction model. The score can, for example, include an ICH score. The ICH score can be calculated by assigning the score to various factors, such as the Glasgow Coma Scale (GCS) score, the volume of intracerebral hemorrhage, the presence of intraventricular hemorrhage (IVH), the age of the subject, and the origin of the hemorrhage. The output of the prediction model can, for example, include an indication of the ICH volume, the presence of intraventricular hemorrhage, and the origin of the hemorrhage. Other information (e.g., GCS score and the age of the subject) can be provided by a human operator and obtained from a database or storage medium (e.g., memory 106) accessible to the processor. The ICH score can be automatically calculated by the processor 102 by mapping all required inputs to the model, thereby obtaining the ICH score. The processor 102 can, for example, display the ICH score on a display (such as a user interface 104).

[0085] According to another aspect of the present invention, a method is provided. Figure 66 is a flow chart of an example of a method 600, such as a method for stroke characterization in medical image data. The method 600 includes, at step 602, receiving image data representing a three-dimensional image of a subject's head. The image data may, for example, include image data acquired using non-contrast computed tomography (CT) imaging techniques. 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) obtaining the image data from a storage medium or database). At step 604, the method 600 includes identifying regions within the image data corresponding to bones in the subject's head. The identification of bones (e.g., skull bones) in the image data may be performed using the techniques disclosed above. The method 600 includes, at step 606, applying adjustments to the image data to compensate for effects caused by radiation scattered from the bones during acquisition of the image data. As described above, the adjustments to the image data may include reducing the pixel intensity of pixels in the image data that have increased due to electromagnetic radiation scattered from the bones during image acquisition. In some embodiments, the pixel intensities may be adjusted according to Equation 1 discussed above. At step 608, method 600 includes generating a two-dimensional image based on the adjusted image data. The two-dimensional image can be generated using a maximum intensity projection technique. In some embodiments, multiple two-dimensional images can be generated, for example, by performing maximum intensity projection through an axial plane, a coronal plane, and a sagittal plane. Method 600 includes, at step 610, providing the generated two-dimensional image (or multiple two-dimensional images) as input to a prediction model to identify a region of interest for stroke characterization in the two-dimensional image. As described above, in some examples, the prediction model can include a convolutional neural network. The prediction model can be trained to identify regions in the brain of the subject where bleeding has occurred. Therefore, the method can also include determining whether the identified region of interest indicates a hemorrhagic event. In some embodiments, the prediction model can provide an indication of the location of the region of interest (e.g., hemorrhage), and additionally other information about 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 (e.g., an ICH score) indicating the severity of the event (e.g., hemorrhage) that has occurred.

[0086] Method 600 can be performed using a processor, such as processor 102. Thus, method 600 may include one or more additional steps performed by processor 102, as discussed above. For example, the method may also include identifying a portion of the image data corresponding to a boundary of the subject's brain. In some examples, the method may also include providing a two-dimensional image and an indication of a region of interest in the two-dimensional image for display.

[0087] Thus far, the description has focused on using a predictive model to identify regions of interest associated with stroke. According to another aspect, the present invention is directed to training a predictive model to identify regions of interest. Figure 7 7 is a flow chart illustrating another example of a method 700, such as a method for processing image data for use in stroke characterization. Processing image data according to the steps of method 700 can be considered equivalent to preprocessing image data in the manner described above. Method 700 includes, at step 702, preparing a training dataset. The training dataset is prepared by receiving image data for each of a plurality of subjects (step 704) and preprocessing the image data (step 706). Each subject may, for example, comprise a patient who has undergone an imaging procedure to acquire three-dimensional image data. Thus, the training dataset is prepared (step 702) by receiving (step 704) three-dimensional image data representing the subject's head (for each of the plurality of subjects) and preprocessing (step 706) the three-dimensional image data. The three-dimensional image data is preprocessed (step 706) by applying (step 708) adjustments to the three-dimensional image data to compensate for effects caused by radiation scattered from bone during acquisition of the three-dimensional image data, and a two-dimensional image is generated (step 710) based on the three-dimensional image data. Thus, the three-dimensional image data is pre-processed as described above (e.g., in a slice-by-slice manner) and then a two-dimensional image is generated (e.g., using maximum intensity projection). The two-dimensional image generated for each of the plurality of objects can then be used to train the prediction model. As described above, in some embodiments, the plurality of two-dimensional images can be generated, for example, by performing a plurality of maximum intensity projections through three different planes. In such an example, the plurality of two-dimensional images can be used as input to train the prediction model.

[0088] Thus, method 700 may further include training a prediction model using a training data set to identify a region of interest for stroke characterization in a two-dimensional image. As will be appreciated, the prediction model may be trained on various features in order to perform the desired task. For example, the prediction 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 and / or weights used in the prediction model may be selected based on the desired output to be provided by the prediction model.

[0089] Example

[0090] A specific example is now discussed in which a prediction model is trained according to an embodiment of the present invention and used to provide an indication of a region of interest for stroke characterization according to other embodiments of the present invention.

[0091] In this example, three-dimensional image data of multiple subjects was collected in the form of a dataset named "CQ 500" obtained from the "qure.ai" website. The dataset is understood to include non-contrast head CT scan image data obtained from the Center for Advanced Research in Imaging, Neuroscience and Genomics (CARING) in 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. Based on evaluations performed by three radiologists, the scans were annotated as being associated with brain that had experienced bleeding (labeled "bleeding") or being associated with brain that had not experienced bleeding (labeled "no bleeding"). Each scan volume contained approximately 300 slices, with a slice thickness of 0.625 mm.

[0092] Each slice of the image data is preprocessed using the techniques described herein, and for each scan volume, a maximum intensity projection is performed to generate a two-dimensional image of each object. The collection of two-dimensional images forms a training dataset for training the prediction model. The two-dimensional images are provided as input to the prediction model with a size of 512×512 pixels and three color channels. In this example, the prediction model used is the one discussed above and in Figure 2 Figure 1 shows the Inception-ResNet convolutional neural network model schematically shown in Figure 2. Initially, a set of Imagenet weights was used in the neural network model. The neural network model was trained for 10 epochs, freezing the combined layers. The model was then compiled again, and the entire neural network model was trained for 50 epochs using the stochastic gradient descent (SGD) optimizer.

[0093] In this example, 379 images were used to train the neural network model. The images were rotated 90° and also flipped vertically to create a training dataset of 1137 images. The accuracy achieved by the neural network model on the training dataset was 0.97, with a sensitivity of 0.95 and a specificity of 0.98. Table 1 below shows the confusion matrix for the output of the neural network model using the training dataset of 1137 images.

[0094] Predicted no bleeding Predicted bleeding No actual bleeding 649 32 Actual bleeding 7 449

[0095] Table 1: Confusion matrix for the training dataset

[0096] The trained neural network model was tested on a validation dataset containing approximately 52 images (each from a different subject). The images in the validation dataset were preprocessed in a manner similar to the preprocessing performed on the training dataset images. Using the validation dataset, the output of the neural network model achieved 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 dataset.

[0097] Predicted no bleeding Predicted bleeding No actual bleeding 23 1 Actual bleeding 1 27

[0098] Table 2: Confusion matrix of validation dataset

[0099] From the above examples, it is clear that by pre-processing image data in the manner described herein and training a prediction model to identify regions of interest for stroke characterization, a particularly accurate prediction model can be achieved.

[0100] According to another aspect of the present invention, a computer program product is disclosed. Figure 8 8 is a schematic illustration of an example of a processor 802 in communication with a computer-readable medium 804. According to an embodiment of the present invention, a computer program product includes a non-transitory computer-readable medium 804 having computer-readable code embodied therein, the computer-readable code being configured to, when executed by a suitable computer or processor 802, cause the computer or processor to perform the steps of the methods disclosed herein. The processor 802 may include or be similar to the processor 102 of the apparatus 100 discussed above.

[0101] The processor 102, 802 may include 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 particular embodiments, the processor 102, 802 may include multiple software and / or hardware modules configured to perform or be used to perform individual or multiple steps of the methods described herein.

[0102] Thus, as disclosed herein, embodiments of the present invention provide a mechanism for preparing image data so that it can be used to train a predictive model to identify regions of interest associated with stroke characterization, or provided as input to a trained predictive model capable of identifying such regions of interest. By pre-processing the data in this manner, it has been found that providing the data as input to a predictive model provides particularly accurate results. Specifically, the likelihood that regions of increased pixel intensity in an image will be misinterpreted as hemorrhage is eliminated or substantially reduced by compensating for radiation scattered from the inner surface of the skull within the brain.

[0103] The term "module" as used herein is intended to include a hardware component (such as a processor or component of a processor configured to perform a specific function) or a software component (such as a set of instruction data that has a specific function when executed by a processor).

[0104] It will be appreciated that embodiments of the present invention also apply to computer programs, particularly computer programs on or in a carrier, suitable for putting the present invention into practice. The program may be in the form of source code, object code, a code intermediate source code and object code (such as in partially compiled form), or in any other form suitable for implementing a method according to an embodiment of the present invention. It will also be appreciated that such a program may have many different architectural designs. For example, the program code that implements the functions of a method or apparatus according to the present invention may be subdivided into one or more subroutines. Many different ways of distributing functions among these subroutines will be apparent to those skilled in the art. The subroutines may be stored together in an executable file to form a self-contained program. Such an executable file may include computer-executable instructions, such as processor instructions and / or interpreter instructions (e.g., Java interpreter instructions). Alternatively, one or more or all subroutines may be stored in at least one external library file and linked to a main program statically or dynamically (e.g., at runtime). The main program may include at least one call to at least one subroutine. The subroutines may also include function calls to each other. Embodiments relating to a computer program product include computer-executable instructions corresponding to each processing stage of at least one method set forth herein. These instructions may be subdivided into subroutines and / or stored in one or more files that may be linked statically or dynamically. Another embodiment of a computer program product comprises computer-executable instructions corresponding to each element of at least one of the apparatus and / or products set forth herein. These instructions may be subdivided into subroutines and / or stored in one or more files that may be linked statically or dynamically.

[0105] The carrier of a computer program may be any entity or device capable of carrying the program. For example, the carrier may include a storage medium, such as a ROM (e.g., a CDROM or semiconductor ROM), or a magnetic recording medium (e.g., a hard disk). In addition, the carrier may be a transmissible carrier such as an electrical or optical signal, which may be communicated via an electrical or optical cable or by radio or other means. When the program is embodied in such a signal, the carrier may be constituted by such a cable or other device or unit. Alternatively, the carrier may be an integrated circuit in which the program is embodied, the integrated circuit being suitable for executing the relevant method, or for the execution of the relevant method.

[0106] A person skilled in the art will be able to understand and implement variations of the disclosed embodiments when practicing the claimed invention by studying the drawings, the disclosure and the dependent claims. In the claims, the word "comprising" does not exclude other elements or steps, and the word "a" or "an" does not exclude a plurality. A single processor or other unit may implement the functions of several items recited in the claims. Although specific measures are recited in mutually different dependent claims, this does not indicate that a combination of these measures cannot be used to advantage. The computer program may be stored / distributed on a suitable medium, such as an optical storage medium or solid-state medium supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunications means. Any reference signs in the claims should not be interpreted as limiting the scope.

Claims

1. A device for characterizing stroke, comprising: A processor configured to: receiving image data representing a three-dimensional image of a subject's head; identifying regions within the image data corresponding to bones in the head of the subject; applying an adjustment to the image data to compensate for effects caused by radiation scattered from the bone during acquisition of the image data, wherein the adjustment applied to the image data comprises reducing pixel intensity in an area corresponding to an area of the subject's head into which radiation was scattered from the bone during acquisition of the image data, wherein the adjustment of the pixel intensity comprises adjusting the pixel intensity based on a distance of the pixel from a surface of the bone; generating a two-dimensional image based on the adjusted image data; and The generated two-dimensional image is provided as input to a prediction model to identify a region of interest for stroke characterization in the two-dimensional image.

2. The device according to claim 1, wherein The processor is further configured to ignore portions of the image data that do not correspond to the brain or correspond to regions of the brain that are unlikely to be affected by a stroke-related event.

3. The device according to claim 1, wherein The processor is further configured to: The image data is registered to the three-dimensional representation of the brain using a plurality of fiducial landmarks common to the image data and the three-dimensional representation.

4. The device according to any one of the preceding claims, wherein The processor is configured to: For each of a plurality of successive slices of the image data, identifying a portion of the image data corresponding to a boundary of the subject's brain by: applying a mask to remove the region of the image data corresponding to bone; defining a boundary around the region in the image data corresponding to the bone; identifying a sub-region of the image data having a maximum number of contiguous pixels within the defined region; and An identified subregion of the image data corresponding to the subject's brain is determined.

5. The device according to any one of claims 1 to 3, wherein: The processor is configured to: The portion of the image data corresponding to the lower boundary of the subject's brain is identified by: analyzing successive slices of image data progressing through the head to determine a measure of brain visible in each slice; and In response to determining that a change in the measure of the brain visible in a particular slice relative to a measure of the brain visible in a slice immediately preceding the particular slice is below a defined threshold, determining that the particular slice includes image data corresponding to a lower boundary of the subject's brain.

6. The device according to any one of claims 1 to 3, wherein: The processor is configured to generate a two-dimensional image based on the image data by: A maximum intensity projection of the image data through at least one of the coronal, axial, and sagittal planes is performed, wherein the image data has been processed to compensate for effects caused by radiation scattered from the bone during acquisition of the image data.

7. The device according to any one of claims 1 to 3, wherein: The processor is further configured to: obtaining an indication of the identified region of interest as an output of the predictive model; and A bounding box is generated around the region of interest for presentation in a representation of the brain of the subject.

8. The device according to any one of claims 1 to 3, wherein: The region of interest includes an area where bleeding has occurred, and wherein the processor is further configured to: Based on the output of the predictive model, a score is calculated that indicates the severity of the bleeding.

9. The device according to any one of claims 1 to 3, wherein: The image data includes data acquired using a non-contrast computed tomography imaging modality.

10. The device according to any one of claims 1 to 3, wherein: The predictive model includes a convolutional neural network trained to determine whether the identified region of interest is indicative of a bleeding event.

11. The device according to any one of claims 1 to 3, wherein: The adjusting includes adjusting the pixel intensity according to the following formula: Wherein y is the pixel intensity, and wherein x is the distance from the surface of the bone.

12. 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 within the image data corresponding to bones in the head of the subject; applying an adjustment to the image data to compensate for effects caused by radiation scattered from the bone during acquisition of the image data, wherein the adjustment applied to the image data comprises reducing pixel intensity in an area corresponding to an area of the subject's head into which radiation was scattered from the bone during acquisition of the image data, wherein the adjustment of the pixel intensity comprises adjusting the pixel intensity based on a distance of the pixel from a surface of the bone; generating a two-dimensional image based on the adjusted image data; and The generated two-dimensional image is provided as input to a prediction model to identify a region of interest for stroke characterization in the two-dimensional image.

13. A method of processing image data for use in stroke characterization, the method comprising: For each of the multiple objects, prepare the training dataset by doing the following: receiving three-dimensional image data representing a head of a subject; and The three-dimensional image data is preprocessed by the following operations: applying an adjustment to the three-dimensional image data to compensate for effects caused by radiation scattered from bone during acquisition of the three-dimensional image data, wherein the adjustment applied to the image data comprises reducing pixel intensity in an area corresponding to an area of the subject's head into which radiation was scattered from the bone during acquisition of the image data, wherein the adjustment of the pixel intensity comprises adjusting the pixel intensity based on a distance of the pixel from a surface of the bone; and A two-dimensional image is generated based on the three-dimensional image data.

14. The method according to claim 13, further comprising: A prediction model is trained using the training dataset to identify regions of interest for stroke characterization in the two-dimensional image.

15. A computer program product comprising a non-transitory computer readable medium having computer readable code embodied therein, the computer readable code being configured such that, when executed by a suitable computer or processor, the computer or processor is caused to perform the method according to claims 12 to 14.

Images