Vertebral fracture detection

Abstract

Three-dimensional image data including at least a portion of a spine of a subject is received. In the three-dimensional image data, the spine of the subject is identified and a spine curve approximating a local curvature along the spine is defined. A plurality of volumes of interest (VOLs of Interest) are defined, each volume of interest including at least a portion of a vertebra of the spine of the subject. Each VOL of Interest is defined relative to an adjacent segment of the spine. A fracture in at least one of the VOLs of Interest is identified.

Description

[Technical Field]

[0001] The present invention relates generally to systems and methods for spinal fracture detection using machine learning, and more particularly to spinal fracture detection using convolutional neural networks (CNNs) on spinal canal-aligned volumes of interest (VOIs). [Background technology]

[0002] Emergency department care, such as in the context of triage for trauma, often requires rapid assessment of all injuries as well as overall patient condition. Ideally, such assessment is as complete and rapid as possible. The use of 3D computed tomography (CT) scans of the head to pelvis or head to toes has been widely established to aid in such rapid assessment. Summary of the Invention [Problem to be solved by the invention]

[0003] Among potential injuries, identifying spinal injuries, such as fractures, is particularly important. Unstable fractures often require prompt surgical intervention and result in subsequent injuries during hospital care due to patient transport and handling. Therefore, cervical spine fractures, especially in the setting of unclear polytrauma, must be identified or ruled out as soon as possible.

[0004] Existing methods do not readily use CT scans to detect spinal fractures. Because manual evaluation of the spine is time-consuming, some automated methods based on images such as CT scans use a first-aid system to classify individual vertebrae as either fractured or not. However, it is not always possible to determine what led to the fracture, and therefore it is not always possible to confirm the presence of such a fracture.

[0005] For example, in some fracture detection systems, a technician may be informed whether a given vertebra is fractured. However, such systems are typically vertebral fracture classification systems, and information regarding the location of the fracture may not be provided, or may only be provided in the form of a compressed preview of the network decision map. Thus, a positive indication from an automated fracture detection method may be difficult to confirm as either a fracture or a false positive.

[0006] As mentioned above, manual detection of spinal fractures is a lengthy procedure, and missed fractures can have serious health consequences for patients. Failure to detect unstable fractures can lead to spinal cord injury. Therefore, a false negative identified by an automated fracture detection method, or an insufficiently thorough manual examination by a physician, can have serious consequences. Similarly, incorrectly identifying the output of an automated method as a false positive can have serious consequences.

[0007] Therefore, there is a need for a spinal fracture detection system and method that can provide more accurate fracture detection, where such fracture detection can highlight the specific location of a suspected fracture. There is also a need for such a system that can indicate a confidence level, such as a confidence score, associated with an identified fracture. [Means for solving the problem]

[0008] A method for detecting spinal fractures is provided, according to one embodiment of the present invention, comprising receiving three-dimensional image data comprising at least a portion of a spine of a subject.

[0009] According to one embodiment, the method proceeds to identify the subject's spine within the three-dimensional image data and define a spine that approximates the local curvature along the subject's spine.

[0010] According to one embodiment, the method then proceeds to define a plurality of volumes of interest (VOIs), each VOI including at least a portion of a vertebra of the spine of interest, each VOI being defined relative to adjacent segments of the spine.

[0011] According to one embodiment, the method then identifies a fracture in at least one of the VOIs.

[0012] In some embodiments, the identification of vertebrae is by a convolutional neural network (CNN) trained to segment vertebrae, hi some such embodiments, the CNN is a foveated net applied to CT images.

[0013] In some embodiments, each VOI of the multiple VOIs has a corresponding center point sampled at a defined location relative to the spine. The center points of the multiple VOIs are then sampled at regular intervals along the spine. In some such embodiments, the center points of adjacent VOIs are positioned to generate overlapping VOIs.

[0014] In some embodiments, each VOI is formed around a center point and oriented based on the tangent of the vertebra adjacent to the corresponding center point. In some such embodiments, after defining each VOI of the plurality of VOIs, each VOI is extracted and resampled from the three-dimensional image data to a target resolution.

[0015] In some embodiments, identifying fractures is by applying a CNN to each VOI of the plurality of VOIs, in some such embodiments, the output of the CNN, when applied to a VOI of the plurality of VOIs, is a probability map that identifies the likelihood of fracture within the corresponding VOI.

[0016] In some such embodiments, multiple VOIs are defined such that adjacent VOIs overlap, generating multiple predictions for at least some equivalent voxels occurring in the multiple VOIs, and all predictions corresponding to a particular location are then aggregated into a final probability map.

[0017] In some embodiments, the method further includes generating a final probability map from the probability maps associated with the individual VOIs. The final probability map includes matching the VOI probability maps to a coherent representation of the three-dimensional image data. In some such embodiments, the method further includes generating a binary prediction based on the final probability map and filtering fracture candidates based on a relationship between the candidate locations and the vertebrae of interest.

[0018] In some embodiments, each VOI of the plurality of VOIs includes a portion of the spine in the corresponding volume.

[0019] In some embodiments, the vertebrae approximate the centerline of the spinal canal of the spine.

[0020] In some embodiments, a size of a first VOI of the plurality of VOIs is selected based on a first adjacent location along the spine, and then a size of a second VOI of the plurality of VOIs is selected based on a second adjacent location along the spine, the size of the first VOI being different from the size of the second VOI.

[0021] In some such embodiments, the size of the first VOI is based on the size of the expected vertebral structure at a first location along the spine.

[0022] In some embodiments, the method further comprises locating the identified fracture in a representation of the three-dimensional image data and displaying the identified fracture in relation to the subject's spine. [Brief explanation of the drawings]

[0023] [Figure 1]1 is a schematic diagram of a system according to one embodiment of the present invention. [Figure 2] 1 illustrates an exemplary imaging device according to one embodiment of the present invention. [Figure 3A] 1 illustrates segmentation of a spine in image data according to the disclosed method. [Figure 3B] 1 illustrates segmentation of a spine in image data according to the disclosed method. [Figure 3C] 1 illustrates segmentation of a spine in image data according to the disclosed method. [Figure 3D] 1 illustrates segmentation of a spine in image data according to the disclosed method. [Figure 3E] 1 illustrates segmentation of a spine in image data according to the disclosed method. [Figure 4] 1 illustrates the creation of a spine mask according to the disclosed method. [Figure 5] 1 illustrates the identification of a volume of interest (VOI) according to the disclosed method. [Figure 6] Figure 5 shows the identified and extracted VOIs. [Figure 7] 1 is a flowchart illustrating a method for vertebral fracture detection according to an embodiment of the present invention. [Figure 8A] 1 illustrates fractures identified in the context of image data. [Figure 8B] 1 illustrates fractures identified in the context of image data. DETAILED DESCRIPTION OF THE INVENTION

[0024] The description of illustrative embodiments according to the principles of the present invention is intended to be read in connection with the accompanying drawings, which are to be considered part of the entire written description. In describing the embodiments of the present invention disclosed herein, any reference to direction or orientation is intended for convenience of description only and is in no way intended to limit the scope of the present disclosure. Relative meanings such as "bottom," "top," "horizontal," "vertical," "up," "bottom," "above," "bottom," "up," "top," "bottom," and derivatives thereof (e.g., "horizontal," "bottom," "above," etc.) should be interpreted to refer to the orientation as then described or as shown in the drawings under discussion. These relative meanings are for convenience of description only and do not require that the device be constructed or operated in a particular orientation unless explicitly so indicated. "Mounted," "attached," "connected," "coupled," "interconnected," and similar terms refer to a relationship in which structures are fixed or attached to one another, both directly or indirectly through intervening structures, and through movable or rigid attachments or relationships, unless explicitly stated otherwise. Furthermore, the features and advantages of the present invention are illustrated by reference to the illustrated embodiments. Accordingly, the present disclosure should not be expressly limited to such exemplary embodiments, which illustrate some possible non-limiting combinations of features that may exist alone or in combination of features, and the scope of the present invention is defined by the claims appended hereto.

[0025] This disclosure describes the best mode of carrying out the present invention currently contemplated. This description is not intended to be understood in a limiting sense, but provides examples of the disclosure presented solely for illustrative purposes, with reference to the accompanying drawings, to advise those skilled in the art of the advantages and structure of the present disclosure. In the various views of the drawings, like reference numerals indicate like or similar parts.

[0026] It is important to note that the disclosed embodiments are merely examples of the many advantageous uses of the innovative teachings herein. In general, statements made in the specification of this application do not necessarily limit any of the various claimed embodiments. Moreover, some statements may apply to some inventive features, but not to others. In general, unless otherwise specified, singular elements may be in the plural and vice versa, without loss of generality.

[0027] 1 is a schematic diagram of a system 100 according to one embodiment of the present invention. As shown, the system 100 typically includes a processing device 110 and an imaging device 120.

[0028] Processing device 110 may apply processing routines to images (image data) and projections (projection data) received from imaging device 120. Processing device 110 may include memory 113 and processor circuitry 111. Memory 113 may store a plurality of instructions. Processor circuitry 111 may be coupled to memory 113 and configured to execute the instructions. The instructions stored in memory 113 may comprise processing routines and data associated with the processing routines, such as machine learning algorithms and various filters for processing images.

[0029] Processing device 110 may further include an input 115 and an output 117. Input 115 may receive information, such as three-dimensional image data or projection data, from imaging device 120. Output 117 may output information, such as a filtered image or a transformed two-dimensional image, to a user or user interface device. The output may include a monitor or display.

[0030] In some embodiments, processing device 110 may be directly associated with imaging device 120. In alternative embodiments, processing device 110 may be separate from imaging device 120, such that processing device 110 receives image or projection data for processing at input 115 via a network or other interface.

[0031] In some embodiments, the imaging device 120 may include a data processing device and a spectral CT scanning unit for generating CT spectral data when scanning an object (e.g., a patient). In some embodiments, the imaging device 120 may be a conventional CT scanning unit configured to generate the scan.

[0032] 2 illustrates an exemplary imaging device 200 according to one embodiment of the present disclosure. Although imaging device 200 is illustrated and the following description is generally in the context of CT images, it will be understood that similar methods may be applied in the context of other imaging devices, and that images to which these methods may be applied may be acquired in a wide variety of ways.

[0033] In the imaging device 200 according to an embodiment of the present invention, the CT scanning unit may be adapted to perform axial and / or helical scans of the object to generate CT projection data. In the imaging device 200 according to an embodiment of the present invention, the CT scanning unit may comprise an energy-resolving photon counting detector or a spectral dual-layer image detector. Spectral content may also be acquired using other detector setups. The CT scanning unit may include a radiation source that emits radiation to traverse the object when acquiring the projection data.

[0034] 2, a CT scanning unit 200, e.g., a computed tomography (CT) scanner, may include a stationary gantry 202 and a rotating gantry 204 that may be rotatably supported by the stationary gantry 202. The rotating gantry 204 may rotate about a longitudinal axis around an examination region 206 of the object when acquiring projection data. The CT scanning unit 200 may include a support 207 for supporting a patient in the examination region 206 and is configured to move the patient through the examination region during the imaging process.

[0035] The CT scanning unit 200 may include a radiation source 208, such as an X-ray tube, that may be supported by and configured to rotate with a rotating gantry 204. The radiation source 208 may include an anode and a cathode. A power supply voltage applied between the anode and the cathode may accelerate electrons from the cathode to the anode. The electron flow may provide a current flow from the cathode to the anode to generate radiation for traversing the examination region 206.

[0036] The CT scanning unit 200 may include a detector 210. The detector 210 may subtend an angled arc on an opposite side of the examination region 206 relative to the radiation source 208. The detector 210 may include a one-dimensional or two-dimensional array of pixels, such as direct or indirect conversion detector pixels. The detector 210 may be adapted to detect radiation traversing the examination region 206 and generate a signal indicative of its energy.

[0037] The CT scanning unit 200 may include generators 211 and 213. The generator 211 may generate tomographic projection data 209 based on signals from the detector 210. The generator 213 receives the projection data 209 and, in some embodiments, may generate three-dimensional CT image data 311 of the object based on the projection data 209. In some embodiments, the projection data 209 may be provided to the input 115 of the processing unit 110, while in other embodiments, the three-dimensional CT image data 311 is provided to the input of the processing unit.

[0038] In the context of 3D CT images 311, a spinal fracture detection algorithm can formulate fracture detection as a segmentation task rather than a classification task. Thus, the method can sample a volume of interest (VOI), and for each voxel in a given VOI, the method can attempt to determine whether the corresponding voxel in the vertebra is part of a fracture. This approach further enables the generation of renderings with per-voxel highlighting or color indication of the exact fracture location, as shown below in FIGS. 8A-8B. By highlighting specific locations, the method allows the clinician to efficiently check whether they match the algorithm. Such an approach also builds confidence, as the clinician can better understand the basis for the fracture / non-fracture classification decisions made by the system.

[0039] To evaluate the spine in the context of a segmentation task, the spine is segmented and then the spinal canal centerline is extracted. Once the centerline is extracted, VOIs can be sampled along the centerline, and these VOIs can be oriented based on the local angle of the centerline. Once the VOIs are sampled, they can be evaluated using a learning algorithm, such as a CNN, with respect to the underlying voxels.

[0040] Figures 3A-E illustrate segmentation of the spine in image data according to embodiments disclosed herein. Figure 4 illustrates creation of a spine mask according to embodiments disclosed herein. Figure 5 illustrates identification of a volume of interest (VOI) 500 according to embodiments disclosed herein. Figure 6 illustrates the extracted VOI 500 identified in Figure 5. Figure 7 is a flowchart illustrating spinal fracture detection according to an embodiment of the present invention.

[0041] As shown in Figure 7, a method for spinal fracture detection typically begins by retrieving data 700. The retrieved data is projection data acquired from multiple angles about a central axis.

[0042] 2, the object may be a patient on a support 207, and the central axis may be an axis passing through the examination region. The rotating gantry 204 rotates about the object's central axis to acquire projection data from various angles.

[0043] Once acquired, the projection data 311 is reconstructed 710 as three-dimensional image data 300 for further processing in the image domain.

[0044] The data (retrieved in 700) includes at least a portion of the subject's spine being analyzed. Next, processing (720) involves identifying the subject's spine in the three-dimensional image. While such identification can be done in several ways, one approach is shown in Figures 3A-3E.

[0045] This approach relies on a segmentation process to segment the vertebrae within the three-dimensional image data. Therefore, the method can first apply course segmentation (730) to globally separate the vertebrae as a single class from the rest of the image. This course segmentation is shown in Figure 3A. Such course segmentation can be performed by applying a learning algorithm, such as a convolutional neural network (CNN), to the entire field of view (FOV) contained in the image data. The CNN can be, for example, a foveated neural network (FNET) applied to CT images. The FNET can be applied to CT images using an isotropic resampling of 5.0 mm.

[0046] As shown in Figures 3B, 3C, and 3D, following course segmentation of the vertebrae across the entire FOV, multiple bounding boxes can be placed around the vertebrae to locally refine the segmentation. Within each bounding box, further learning algorithms, such as CNN, can be applied locally. In the illustrated embodiment, UNET can be applied within each bounding box using 2.0 mm isotropic resampling. This finer segmentation can then produce a multi-class output, resulting in segmentation of intermediate vertebrae classes, spinal canal classes, and landmark classes representing the centroids of each vertebral body, along with several key vertebral structures.

[0047] Figure 3B shows patch sampling of the initial course segmentation to enable further segmentation. Figure 3C then shows further segmentation into different classes, allowing labeling of multiple key vertebral structures, such as C2, T1, T-Last, and L-Last. Once these landmarks have been identified and labeled, the method can proceed to count the sorted landmark classes, allowing calculation of all body-centered labeled landmarks visible in the field of view (740), as shown in Figure 3D.

[0048] Once all landmarks have been identified, a centerline is traced 750 based on the landmarks previously identified in 740. The centerline typically corresponds to the spinal canal class of the refined segmentation and includes a set of points along the centerline.

[0049] As shown in FIG. 4, once a set of points 410 have been traced, vertebrae 400 are defined (760) to approximate the local curvature along the spine of interest. In the illustrated embodiment, such vertebrae are based on points 410 defined as centerlines (750), and therefore correspond directly to the identified spinal canal. Thus, in the spine mask shown in FIG. 4, the spinal canal centerlines 400 are shown as the generated vertebrae, and the sampled points 410 are shown linked by the vertebrae. However, it will be understood that the vertebrae may be defined in other ways, so long as they approximate the local curvature.

[0050] Once the spine 400 is defined, multiple volumes of interest (VOIs) 500 (770) are defined, as shown in Figure 5. Each such VOI includes at least a portion of a vertebra of the spine of interest, and each VOI is defined relative to adjacent segments of the spine 400.

[0051] In the illustrated embodiment, each VOI 500 is centered at a point 410 that is sampled as part of the definition of the centerline (750). The points 410 may then be sampled at regular intervals along the spine 400. In some embodiments, the sampled points 410 used to define the spine are used as shown in FIG. 5. In an alternative embodiment, a set of points may be defined along the spine to actively locate the center point of each VOI.

[0052] As shown in FIG. 5, when defining each VOI 500, the corresponding VOI may be formed around a corresponding center point 410 and oriented based on the tangent of the vertebra 400 adjacent to the corresponding center point. In this way, the orientation of each VOI can correspond to the local curvature of the spine, which can improve the performance of algorithms applied to the VOI. The orientation step therefore reduces the geometric variation that classification networks, such as the CNN described below, need to learn. This increases robustness and reduces the size of the required training set.

[0053] In the illustrated embodiment, the VOIs 500 may be defined so that they overlap. If points 410 are sampled at regular intervals along the spine 400, such points may be positioned closer together than the width of the VOI. For example, in the embodiment shown in FIG. 5, the center points 410 of each VOI 500 are sampled at a distance of 20 mm on the spine 400. Each VOI may have a size of, for example, 128 x 128 x 128 voxels at a processing resolution of 7 x 7 x 7. Thus, the sampling distance is selected to generate overlapping VOIs 500. In this way, when the fracture detection algorithm is applied to each VOI, it can provide multiple separate results associated with each VOI for overlapping sections of the VOI. Each vertebra is then processed in multiple views, both in training and inference. The VOIs are positioned so that the complete sampled set covers the entire spine, but no more, which can accelerate future processing steps.

[0054] In some embodiments, each VOI may be intentionally defined to include a portion of the spine within its corresponding volume. Such an approach aids subsequent application of algorithms by allowing spinal cord segments to be used for directional purposes. In some embodiments, VOIs may be utilized without duplicating adjacent spinal cord segments.

[0055] In some embodiments, each VOI 500 may be the same size. This is shown, for example, in FIG. 5 and described above. However, in some embodiments, different VOIs may be different sizes. For example, the size of a first VOI may be selected based on an adjacent first location along the spine, and the size of a second VOI may be selected based on an adjacent second location along the spine. Thus, the size of the first VOI may be different from the size of the second VOI. As described above, spine segmentation can provide a categorical segmentation, as shown in FIG. 3C. Such categorical segmentation can locate each VOI relative to different types of vertebrae, and thus, each VOI may be sized based on the size of the local vertebra. For example, lumbar vertebrae are larger than cervical vertebrae and therefore require a larger VOI. Thus, the size of the first VOI may be based on the size of the expected vertebral structure at the first location, while the size of the second VOI may be based on the size of the expected vertebral structure at the second location. Similarly, in some embodiments, the distance between sampled center points 410 may be varied based on the size of the local vertebral structure.

[0056] In some embodiments, after defining each VOI 500, each VOI is extracted and resampled (780) from the three-dimensional image to a target resolution. The target resolution may be the resolution at which an algorithm designed for fracture detection operates. This may be different from the initial target resolution used for spine segmentation (730). As mentioned above, the orientation of the VOI may be based on the local curvature of the spine 400. Therefore, when resampling the VOI, the volume may be rotated so that all processing may be better aligned with the local tangent vector of the spine 400.

[0057] Figure 6 shows the extracted VOIs 500 identified in Figure 5. As shown, each VOI 500 has been resampled and oriented so that the center of the VOI is aligned with the tangent to the spine 410. Typically, no further reformatting is applied so as not to introduce any distortions other than the resampling.

[0058] Following resampling (780) of each VOI, fractures are identified (790) in at least one of the VOIs 500. Fracture identification can be performed by applying a learning algorithm, such as a convolutional neural network (CNN), to each VOI. The results of the CNN (800) are output so that any identified fractures can be evaluated. The CNN architecture may be UNET, used for segmentation, although landmark detection or bounding box regression are also contemplated.

[0059] As mentioned above, a VOI may be defined to include the spine as a supporting structure, and thus the network may be trained to simultaneously segment the spine and fractures. Including the spine as a supporting structure in training and processing may achieve higher performance than segmenting fractures alone. Therefore, some embodiments include intentionally including the spine in the VOI. During network training, case weighting based on fracture occurrence and type, as well as adapted sampling to ensure an appropriate ratio between spine samples and fracture samples, may be applied.

[0060] The output of the CNN (800), when applied to a VOI 500 of the multiple VOIs, can be a probability map that identifies possible fractures within the corresponding VOI. Such a probability map can then be resampled back to the original image space, such as a three-dimensional image. As described above, the VOIs 500 can be defined such that adjacent VOIs overlap. This generates multiple predictions for at least some equivalent voxels contained in the multiple VOIs. All predictions corresponding to a particular location are then aggregated into a final probability map. Thus, in some embodiments, following the output of the CNN (800), a final probability map is generated from the probability maps associated with the individual VOIs. The final probability map includes matching the VOI probability map to a coherent representation of the three-dimensional image data. Such matching can be, for example, by maximum motion.

[0061] Once the probability map is complete, post-processing may be required to extract the final fracture candidates from the map. A binary prediction (810) is generated based on the final probability map, such as by using an argmax operation or by thresholding the final fracture probability map. The fracture candidates can then be filtered based on the relationship between the candidate locations and the vertebrae of interest, for example, to eliminate potential false positives outside the vertebrae. The resulting prediction can be masked with an extended vertebra mask, as shown in FIG. 3E.

[0062] Candidates can then be identified through connected component analysis of the binary predictions. Components can be filtered by size to exclude small fragments. Finally, a confidence score can be calculated for each candidate based on the fracture probability.

[0063] Once predictions are made (810), such predicted fractures may be identified and placed within a representation of the three-dimensional image data, such that an image containing the predicted fractures may be presented to a user in the context of the subject's spine (820).

[0064] 8A-B show fractures identified in the context of image data. These diagrams show how a fracture candidate 900 may be visualized in the original data. As shown, the fracture candidate 900 may be highlighted in the context of other original images.

[0065] Embodiments of the present invention can be implemented in a variety of contexts, including on a CT console associated with a particular CT scanner, as well as on a medical viewing workstation, a PACS system, or a cloud-based system.

[0066] A receiver operating characteristic (ROC) curve can be generated to validate the described fracture detection. In some embodiments, such a curve may be a ROC curve in a four-fold cross-validation setting using 200 training cases, with several fracture candidates identified per case. A fracture detection sensitivity of 87% could be achieved by allowing four false positives per case. In any case, a confidence score can be utilized to indicate the confidence in any particular fracture candidate. Such a confidence score can be used as a threshold metric, which can then be used to select the sensitivity of the embodiments described herein. A more sensitive implementation may result in more false positives being identified.

[0067] Embodiments according to the present disclosure may be implemented on a computer as a computer-implemented method, or on dedicated hardware, or a combination of both. Executable code for a method according to the present disclosure may be stored on a computer program product. Examples of computer program products include memory devices, optical storage devices, integrated circuits, servers, online software, etc. Preferably, the computer program product may include non-transitory program code stored on a computer-readable medium for performing a method according to the present disclosure when the program product is run on a computer. In one embodiment, a computer program may include computer program code adapted to perform all steps of a method according to the present disclosure when the computer program is run on a computer. The computer program may be embodied on a computer-readable medium.

[0068] While the present invention has been described at some length and with some particularity with respect to several embodiments, it should not be limited to any such details or embodiments or to any particular embodiment, but should be construed with reference to the appended claims so as to provide the broadest possible interpretation of such claims in view of the prior art, and therefore so as to effectively encompass the intended scope of the present disclosure.

[0069] All examples and conditional language recited herein are for educational purposes to aid the reader in understanding the principles of the present disclosure and concepts contributed by the inventors to further the art, and should not be construed as being limited to such specifically recited examples and conditions. Moreover, all statements herein reciting principles, aspects, and embodiments of the present disclosure, as well as specific examples thereof, are intended to encompass both structural and functional equivalents thereof. Furthermore, such equivalents are intended to include both currently known equivalents and equivalents developed in the future, i.e., any elements developed that perform the same function, regardless of structure.

Claims

1. 1. A computer-implemented method for detecting vertebral fractures, comprising: receiving three-dimensional image data including at least a portion of a spine of a subject; identifying the subject's vertebrae in the three-dimensional image data; defining a spine that approximates a local curvature along the subject's spine; defining a plurality of volumes of interest, each volume of interest including at least a portion of a vertebra of the subject's spine, each volume of interest being defined relative to adjacent segments of the spine; identifying a fracture in at least one of the volumes of interest; A method comprising:

2. 10. The method of claim 1, wherein the vertebrae are identified by a convolutional neural network trained to segment the vertebrae.

3. The method of claim 1 , wherein for each volume of interest, a corresponding center point is sampled at a defined location relative to the spine, and the center points of the volumes of interest are sampled at regular intervals along the spine.

4. The method of claim 3 , wherein center points of adjacent volumes of interest are positioned to create overlapping volumes of interest.

5. The method of claim 3 , wherein each volume of interest is formed around the center point and oriented based on a tangent to a vertebra adjacent to the corresponding center point.

6. The method of claim 5 , wherein after defining each volume of interest, each volume of interest is extracted and resampled to a target resolution from the three-dimensional image data.

7. 10. The method of claim 1, wherein the fractures are identified by applying a convolutional neural network to each volume of interest.

8. 8. The method of claim 7, wherein the output of the convolutional neural network is a probability map that, when applied to each volume of interest, identifies likely fractures within the corresponding volume of interest.

9. 9. The method of claim 8, wherein the volumes of interest are defined such that adjacent volumes of interest overlap to generate multiple predictions for at least some equivalent voxels occurring in the multiple volumes of interest, and all predictions corresponding to a particular location are aggregated into a final probability map.

10. 9. The method of claim 8, further comprising generating a final probability map from the probability maps associated with individual volumes of interest, the final probability map comprising matching the volume of interest probability maps to a coherent representation of the three-dimensional image data.

11. 11. The method of claim 10, further comprising generating a binary prediction based on the final probability map and filtering fracture candidates based on a relationship between candidate locations and the subject's vertebrae.

12. The method of claim 1 , wherein the vertebrae approximate a centerline of the spinal canal of the spine.

13. 10. The method of claim 1, wherein a size for a first volume of interest is selected based on an adjacent first location along the spine, and a size for a second volume of interest is selected based on an adjacent second location along the spine.

14. The method of claim 1 , further comprising: locating the identified fractures in the representation of the three-dimensional image data; and displaying the identified fractures in the context of the subject's spine.

15. a memory for storing a plurality of instructions; a processor, coupled to the memory, configured to execute the plurality of instructions, the plurality of instructions comprising: receiving three-dimensional image data including at least a portion of a spine of a subject; identifying the subject's vertebrae in the three-dimensional image data; defining a spine that approximates a local curvature along the subject's spine; defining a plurality of volumes of interest, each volume of interest including at least a portion of a vertebra of the subject's spine, each volume of interest being defined relative to adjacent segments of the spine; identifying a fracture in at least one of the volumes of interest; a processor, An apparatus having:

Images