Spinal fracture detection

Abstract

Three-dimensional image data including at least a portion of a spine of a subject is received. The spine of the subject is identified in the three-dimensional image data and a spline is defined that approximates a local curvature along the spine. A plurality of volumes of interest (VOIs) are defined, each VOI containing at least a portion of a vertebra of the spine of the subject. Each VOI is defined relative to an adjacent segment of the spline. A fracture in at least one of the VOIs is identified.

Description

Technical Field

[0001] The present invention generally relates to systems and methods for spinal fracture detection using machine learning. In particular, the present invention relates to spinal fracture detection using a convolutional neural network (CNN) on a volume of interest (VOI) aligned with the spinal canal. Background Art

[0002] In the emergency department setting (e.g., in the context of trauma triage), there is often a need to quickly assess all injuries as well as the patient's overall condition. Ideally, this assessment should be as complete and rapid as possible. To support this rapid assessment, 3D computed tomography (CT) from head to pelvis (or even head to toes) is widely used.

[0003] Among potential injuries, identification of spinal injuries (e.g., fractures) is particularly important. Unstable fractures often require rapid surgical intervention and can result in subsequent injury during hospital care due to transport and handling of the patient. Therefore, cervical spine fractures, in particular, must be identified or excluded as soon as possible in the setting of unexplained polytrauma.

[0004] In existing methods, such CT scans are not easily used to detect spinal fractures. Because manual assessment of the spine is time consuming, in some automated imaging-based methods (e.g., CT scans), emergency systems classify individual vertebrae as either fractured or unfractured. However, it is not always possible to determine what caused the method to indicate the presence of a fracture, and as such, it is not always possible to confirm the presence of such a fracture.

[0005] For example, in some fracture detection systems, technicians can be informed whether a given vertebra is fractured. However, such systems are typically full vertebra fracture classification systems and may not provide information about the location of the fracture, or may only provide information about the location of the fracture in the form of a compressed preview of a network decision graph. Therefore, a positive indication from an automated fracture detection method may be difficult to confirm whether it is a fracture or a false positive.

[0006] As mentioned above, manual detection of spinal fractures is a lengthy process, and overlooked fractures can have serious health consequences for patients. Failure to detect unstable fractures can lead to spinal cord injury. Therefore, false negatives identified by automated fracture detection methods, or inadequate manual inspection by physicians, can have serious consequences. Similarly, erroneously identifying the output of automated methods as false positives can also have serious consequences.

[0007] Therefore, there is a need for a spinal fracture detection system and method, wherein more accurate fracture detection can be provided, and wherein such fracture detection can highlight the specific location of a presumed fracture. There is also a need for such a system, which can indicate a confidence level (e.g., a confidence score) associated with an identified fracture. Summary of the invention

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

[0009] According to one embodiment, the method continues by identifying the spine of the object in the three-dimensional image data and defining a spline that approximates a local curvature along the spine of the object.

[0010] According to one embodiment, the method then proceeds to define a plurality of volumes of interest (VOIs), each VOI containing at least part of a vertebra of the spine of the subject.Each VOI is defined relative to an adjacent segment of the spline.

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

[0012] In some embodiments, the spine is identified by a convolutional neural network (CNN) trained to segment the spine. In some such embodiments, the CNN is a fovea network applied to CT images.

[0013] In some embodiments, each of the plurality of VOIs has a corresponding center point sampled at a position relative to the spline definition. The center points of the plurality of VOIs are then sampled at regular intervals along the spline. In some such embodiments, the center points of adjacent VOIs are located so as to generate overlapping VOIs.

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

[0015] In some embodiments, a fracture is identified by applying a CNN to each of the plurality of VOIs. In some such embodiments, when the CNN is applied to a VOI of the plurality of VOIs, an output of the CNN is a probability map identifying a possible fracture within the corresponding VOI.

[0016] In some such embodiments, the plurality of VOIs are defined such that adjacent VOIs overlap in order to generate multiple predictions for at least some equivalent voxels occurring in the plurality of VOIs.All predictions corresponding to a particular location are aggregated into a final probability map.

[0017] In some embodiments, the method further comprises generating a final probability map based on the probability maps associated with individual VOIs, the final probability map then comprising collating the VOI probability maps into a coherent representation of the three-dimensional image data. In some such embodiments, the method further comprises generating a binary prediction based on the final probability map, and screening fracture candidates based on a relationship between candidate locations and the spine of the subject.

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

[0019] In some embodiments, the spline approximates a centerline of the spinal canal of the spine.

[0020] In some embodiments, the size of a first VOI in the plurality of VOIs is selected based on a proximal first position along the spine.The size of a second VOI in the plurality of VOIs is then selected based on a proximal second position along the spine, and the size of the first VOI is different from the size of the second VOI.

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

[0022] In some embodiments, the method further comprises locating a fracture identified in the representation of the three-dimensional image data and displaying the identified fracture in the context of the spine of the subject. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a schematic diagram of a system according to an embodiment of the present invention.

[0024] Figure 2 An exemplary imaging device according to an embodiment of the present invention is illustrated.

[0025] Figures 3A-3E The segmentation of the spine in image data according to the disclosed method is illustrated.

[0026] Figure 4 Creation of a spine mask according to the disclosed method is illustrated.

[0027] Figure 5Identifying a volume of interest (VOI) according to the disclosed method is illustrated.

[0028] Figure 6 The diagram shows Figure 5 The extracted VOI is identified in .

[0029] Figure 7 is a flow chart illustrating a method for spinal fracture detection according to one embodiment of the present invention.

[0030] Figures 8A-8B The fracture is illustrated identified in the context of the image data. DETAILED DESCRIPTION

[0031] The description of the illustrative embodiments according to the principles of the present invention is intended to be read in conjunction with the accompanying drawings, which are considered to be part of the entire written description. In the description of the embodiments of the invention disclosed herein, any reference to direction or orientation is only for the convenience of description and is not intended to limit the scope of the present disclosure in any way. Relevant terms such as "below", "above", "horizontally", "vertically", "above", "below", "above", "below", "top" and "bottom" and their derivatives (e.g., "horizontally", "downward", "upward", etc.) should be interpreted as referring to the orientation described later or shown in the drawings discussed. These relevant terms are only for the convenience of description, and unless explicitly stated, it is not required to construct or operate the device in a specific orientation. Terms such as "attach", "attach", "connect", "couple", "interconnect" refer to the relationship in which structures are directly or indirectly fastened or attached to each other through intermediate structures and movable or rigid attachments or relationships, unless otherwise explicitly described. In addition, the features and benefits of the present invention are explained by reference to exemplary embodiments. Therefore, the present disclosure should not be explicitly limited to exemplary embodiments that illustrate certain possible non-limiting feature combinations, which may exist alone or in the form of other feature combinations; the scope of the present invention is defined by the appended claims.

[0032] The present disclosure describes one or more currently contemplated best modes for practicing the present invention. The description is not intended to be understood in a limiting sense, but rather to provide examples of the present disclosure presented for illustrative purposes only, with reference to the accompanying drawings, to inform those of ordinary skill in the art of the advantages and configurations of the present disclosure. In the various views of the accompanying drawings, the same reference numerals represent the same or similar parts.

[0033] It is important to note that the disclosed embodiments are merely examples of many advantageous uses of the innovative teachings herein. In general, the statements in the specification of the present application do not necessarily limit any of the various claimed embodiments. In addition, some statements may apply to some inventive features but not to other inventive features. In general, unless otherwise stated, singular elements may be plural, and vice versa, without loss of generality.

[0034] Figure 1 FIG. 1 is a schematic diagram of a system 100 according to an embodiment of the present invention. As shown in the figure, the system 100 generally includes a processing device 110 and an imaging device 120 .

[0035] The processing device 110 may apply a processing routine to an image (image data) or projection (projection data) received from the imaging device 120. The processing device 110 may include a memory 113 and a processor circuit 111. The memory 113 may store a plurality of instructions. The processor circuit 111 may be coupled to the memory 113 and may be configured to execute instructions. The instructions stored in the memory 113 may include a processing routine and data associated with the processing routine (e.g., a machine learning algorithm and various filters for processing images).

[0036] The processing device 110 may also include an input 115 and an output 117. The input 115 may receive information (e.g., three-dimensional image data or projection data) from the imaging device 120. The output 117 may output information (e.g., a filtered image or a converted two-dimensional image) to a user or a user interface device. The output may include a monitor or display.

[0037] In some embodiments, processing device 110 may be directly associated with imaging device 120. In alternative embodiments, processing device 110 may be distinct from imaging device 120, such that processing device 110 receives image or projection data at input 115 through a network or other interface for processing.

[0038] 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 a target (eg, a patient). In some embodiments, the imaging device 120 may be a conventional CT scanning unit configured to generate scans.

[0039] Figure 2 An exemplary imaging device 200 is illustrated in accordance with one embodiment of the present disclosure. It should be understood that although a CT imaging device 200 is illustrated and the following discussion is generally conducted in the context of CT images, similar methods may also be applied in the context of other imaging devices, and images to which these methods may be applied may be acquired in a variety of ways.

[0040] In the imaging device 200 according to an embodiment of the present invention, the CT scanning unit may be adapted to perform an axial scan and / or a spiral scan of the target to generate CT projection data. In the imaging device 200 according to an embodiment of the present invention, the CT scanning unit may include an energy-resolved photon counting detector or a dual-layer spectral image detector. Other detector settings may also be used to collect spectral content. The CT scanning unit may include a radiation source that emits radiation for passing through the target when collecting projection data.

[0041] exist Figure 2 In the illustrated example, a CT scanning unit 200 (e.g., a computed tomography (CT) scanner) may include a fixed gantry 202 and a rotating gantry 204, which may be rotatably supported by the fixed gantry 202. The rotating gantry 204 may rotate about a longitudinal axis around an examination region 206 for a subject when acquiring projection data. The CT scanning unit 200 may include a support 207 for supporting a patient in the examination region 206 and configured to allow the patient to pass through the examination region during an imaging procedure.

[0042] The CT scanning unit 200 may include a radiation source 208 (e.g., an X-ray tube), which may be supported by the rotating gantry 204 and configured to rotate with the rotating gantry 204. The radiation source 208 may include an anode and a cathode. A source voltage applied between the anode and the cathode may accelerate electrons from the cathode to the anode. The electron flow may provide an electric current from the cathode to the anode to generate radiation that passes through the examination region 206.

[0043] The CT scanning unit 200 may include a detector 210. The detector 210 may be opposite the examination region 206 at an angular arc relative to the radiation source 208. The detector 210 may include a one-dimensional or two-dimensional array of pixels (e.g., direct or indirect conversion detector pixels). The detector 210 may be adapted to detect radiation that passes through the examination region 206 and generate a signal indicative of the energy of the radiation.

[0044] The CT scanning unit 200 may include generators 211 and 213. The generator 211 may generate tomographic projection data 209 based on the signal from the detector 210. The generator 213 may receive the projection data 209, and in some embodiments, the generator 213 generates 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 device 110, while in other embodiments, the three-dimensional CT image data 311 is provided to the input of the processing device.

[0045] In the context of a three-dimensional CT image 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 of a given VOI, the method can attempt to determine whether the corresponding voxel of a vertebra is part of a fracture. This approach also allows for the generation of a rendering with a voxel-by-voxel highlighted or colored indication of the exact fracture location, as shown below Figures 8A-8B By highlighting specific locations, this approach allows clinicians to effectively check whether they agree with the algorithm. This approach also builds trust because clinicians are able to better understand the basis for the fracture / non-fracture classification decision made by the system.

[0046] To evaluate the spine in the context of a segmentation task, the spine is segmented and the spinal canal centerline is extracted. Once the centerline is extracted, a VOI can be sampled along the centerline and such a VOI can be oriented based on the local angle of the centerline. Once the VOI is sampled, a learning algorithm (e.g., CNN) can be used to evaluate the VOI in terms of underlying voxels.

[0047] Figures 3A-3E Illustrated is segmentation of the spine in image data according to embodiments disclosed herein. Figure 4 Illustrated is the creation of a spine mask according to embodiments disclosed herein. Figure 5 Identifying a volume of interest (VOI) 500 is illustrated according to embodiments disclosed herein. Figure 6 The diagram shows Figure 5 The extracted VOI 500 is identified in FIG. Figure 7 is a flow chart illustrating spinal fracture detection according to one embodiment of the present invention.

[0048] like Figure 7 As shown, the method for spinal fracture detection generally first retrieves data (700). The retrieved data is projection data collected from multiple angles around a central axis.

[0049] Therefore, in Figure 2 In the context of the imaging device 200 of FIG. 2 , the object may be a patient on a support 207 and the central axis may be an axis passing through an examination region. The rotating gantry 204 rotates around the central axis of the object to acquire projection data from various angles.

[0050] Once the projection data 311 is acquired, the projection data is reconstructed 710 into three-dimensional image data 300 in preparation for further processing in the image domain.

[0051] The data (retrieved at 700) includes at least a portion of the spine of the subject being analyzed. The processing (at 720) then includes identifying the subject's spine in the three-dimensional image. This identification can be accomplished in a variety of ways, but is Figures 3A-3E Only one method is shown.

[0052] The method relies on a segmentation process for segmenting the spine in the three-dimensional image data. Therefore, the method may first apply route segmentation (730) to separate the spine as a single class in its entirety from the rest of the image. Figure 3A Such route segmentation is shown in . Such route segmentation can be achieved by applying a learning algorithm (e.g., a convolutional neural network (CNN)) to the entire field of view (FOV) contained in the imaging data. The CNN can be, for example, a foveal network (FNET) applied to CT images. The FNET can be applied to CT images using an isotropic resampling of 5.0 mm.

[0053] like Figure 3B , Figure 3C and Figure 3D As shown, after segmenting the spine's route throughout the FOV, multiple bounding boxes can be placed around the spine to locally refine the segmentation. Within each bounding box, further learning algorithms (e.g., CNN) can be applied locally. In the illustrated embodiment, a UNET with 2.0 mm isotropic resampling can be applied within each bounding box. This finer segmentation can then generate a multi-class output, thereby segmenting out intermediate vertebrae categories, spinal canal categories, and landmark categories representing the center of gravity within each vertebral body, as well as multiple key vertebral structures.

[0054] Figure 3B A sampling of patches of a preliminary route segmentation is shown to allow for further segmentation. Figure 3C Further segmentation into different categories is then shown to allow for labeling of multiple key vertebral structures (e.g., C2, T1, T-Last, and L-Last). Once these landmarks are identified and labeled, the method can proceed to counting by the classified landmark categories, thereby allowing for the counting of labeled landmarks for all vertebral centers visible in the field of view (740), as shown in FIG. Figure 3D shown.

[0055] Once all landmarks are identified, a centerline is outlined 750 based on the landmarks previously identified at 740. The centerline generally corresponds to a finely segmented spinal canal class and includes a set of points along the centerline.

[0056] like Figure 4As shown, once the set of points 410 is outlined, a spline 400 is defined (760) that approximates the local curvature along the subject's spine. In the embodiment shown, such a spline is based on the points 410 defined as the centerline (750) and therefore corresponds directly to the identified spinal canal. Figure 4 In the illustrated spine mask, the spinal canal centerline 400 is shown as a generated spline, and the sampling points 410 are shown as being connected by the spline. However, it should be understood that the spline may be defined in other ways as long as it approximates the local curvature.

[0057] Once the spline 400 is defined, a plurality of volumes of interest (VOI) 500 (770) are defined, such as Figure 5 Each such VOI includes at least a portion of a vertebra of the subject's spine, and each VOI is defined relative to an adjacent segment of the spline 400 .

[0058] In the embodiment shown in the figures, each VOI 500 is centered at a point 410 sampled as part of the definition of the centerline (750). Points 410 may then be sampled at regular intervals along the spline 400. In some embodiments, as Figure 5 Sampling points 410 are used to define the spline as shown. In an alternative embodiment, a set of points may be defined along the spline to prospectively locate the center point of each VOI.

[0059] like Figure 5 As shown, when defining each VOI 500, the corresponding VOI is formed around the corresponding center point 410, and the corresponding VOI can be oriented based on the tangent of the spline 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 the algorithm applied to the VOI. Therefore, the orientation step reduces the geometric variability that the classification network (e.g., CNN discussed below) needs to learn. This improves robustness and reduces the size of the required training set.

[0060] In the embodiment shown in the figures, the VOIs 500 may be defined so that they overlap. Where the points 410 are sampled at regular intervals along the spline 400, the points may be positioned so that the distance between them is smaller than the width of the VOI. For example, Figure 5In the illustrated embodiment, the center point 410 of each VOI 500 is sampled at a distance of 20 mm on the spline 400. With a processing resolution of 0.7×0.7×0.7, the size of each VOI can be, for example, 128×128×128 voxels. Therefore, the sampling distance is chosen to generate overlapping VOIs 500. In this way, once the fracture detection algorithm is applied to each VOI, multiple different results associated with each VOI can be provided for the overlapping portions of the VOI. Each vertebra is then processed in multiple views in training and inference. The VOIs are positioned so that the complete sampling set covers the entire spine, but no more, which can speed up future processing steps.

[0061] In some embodiments, each VOI may be intentionally defined so that each VOI includes the portion of the spine within the corresponding volume. This approach facilitates subsequent application of the algorithm by allowing the spinal segment to be used for orientation purposes. In some embodiments, the VOI may be used without duplicating adjacent spinal cord segments.

[0062] In some embodiments, each VOI 500 may be exactly the same size. Figure 5 , and discussed above. However, in some embodiments, the sizes of different VOIs may be different. For example, the size of a first VOI may be selected based on a proximity to a first location along the spine, and the size of a second VOI may be selected based on a proximity to a second location along the spine. The size of the first VOI may be different from the size of the second VOI. As described above, spine segmentation may provide class segmentation, such as Figure 3C As shown. This category segmentation can position each VOI relative to a different type of vertebra, and as such, the size of each VOI can be determined based on the size of the local vertebra. For example, the lumbar spine is larger than the cervical spine, and therefore requires a larger VOI. Therefore, the size of the first VOI can be based on the size of the vertebral structure expected at the first location, while the size of the second VOI can be based on the size of the vertebral structure expected at the second location. Similarly, in some embodiments, the distance between the center points 410 of the samples can vary based on the size of the local vertebral structure.

[0063] In some embodiments, after each VOI 500 is defined, each VOI is then extracted from the three-dimensional image and each VOI is resampled (780) to a target resolution. The target resolution may be the resolution at which the algorithm designed for fracture detection operates. This may be different from the initial target resolution used for spine segmentation (730). As described above, the orientation of the VOI may be based on the local curvature of the spline 400. Thus, when the VOI is resampled, the volume may be rotated so that all processing may be better aligned with the local tangent vector of the spline 400.

[0064] Figure 6 Shown in Figure 5 4. As shown, each VOI 500 has now been resampled and oriented so that they are aligned with the tangent of the spline 410 at the center of the VOI. Typically, no further reformatting is applied beyond resampling to avoid introducing any distortion.

[0065] After each VOI has been resampled (780), a fracture is identified in at least one of the VOIs 500 (790). The fracture may be identified by applying a learning algorithm (e.g., a convolutional neural network (CNN)) to each VOI. The results of the CNN are output (800) so that any identified fractures can be evaluated. The CNN architecture may be a UNET for segmentation, but landmark detection or bounding box regression is also envisioned.

[0066] As described above, the VOI can be defined to include the spine as a supporting structure, and as such, the network can be trained to simultaneously segment the spine and fractures. Including the spine as a supporting structure in training and processing can achieve higher performance than segmenting only fractures. As such, some embodiments include intentionally including the spine in the VOI. During network training, case weighting based on fracture incidence and type as well as adjusted sampling can be applied to ensure an appropriate ratio between spine and fracture samples.

[0067] When the CNN is applied to a VOI 500 among multiple VOIs, the output (800) of the CNN may be a probability map identifying possible fractures within the corresponding VOI. Such a probability map may then be resampled back to the original image space (e.g., a three-dimensional image). As described above, the VOI 500 may be defined so 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. Therefore, in some embodiments, after the CNN (800) is output, a final probability map is generated based on the probability map associated with the individual VOI. The final probability map then includes organizing the VOI probability map into a clear and understandable representation of the three-dimensional image data. This organization may be performed, for example, by a maximum value operation.

[0068] Once the probability map is complete, post-processing may be required to extract the final fracture candidates from the probability map. A binary prediction is generated based on the final probability map (e.g., by using an argmax operation or by thresholding the final fracture probability map) (810). The fracture candidates may then be screened based on the relationship between the candidate locations and the subject's spine, e.g., to exclude potential false positives outside the spine. Figure 3EThe dilated spine mask is shown to mask the resulting predictions.

[0069] Candidates can then be identified via connected component analysis of the binary predictions. Components can be filtered based on their size to exclude small fragments. Finally, a confidence score can be calculated for each candidate based on fracture probability.

[0070] Once a prediction is made (810), such predicted fracture may be identified and located in a representation of the three-dimensional image data. Thus, an image including the predicted fracture may be presented (820) to a user in the context of the subject's spine.

[0071] Figures 8A-8B Identified fractures are illustrated in the background of image data. These figures illustrate how fracture candidates 900 can be visualized in the raw data. As shown, fracture candidates 900 can be highlighted in the background of other raw images.

[0072] Embodiments of the present invention can be implemented in a variety of contexts, including a CT console associated with a particular CT scanner, as well as a medical viewing workstation, a PACS system, or a cloud-based system.

[0073] A receiving operating characteristic (ROC) curve can be generated to verify the described fracture detection. In some embodiments, such a curve can be a ROC curve in a 4-fold cross-validation setting using 200 training cases, wherein each case identifies several fracture candidates. When considering that each case has 4 false positives, the fracture detection sensitivity can reach 87%. In any case, a confidence score can be used to indicate the confidence of any specific fracture candidate. This confidence score can be used as a thresholding metric, which in turn can be used to select the sensitivity of the embodiments described herein. More sensitive implementations may result in identifying additional false positives.

[0074] According to the embodiments of the present disclosure, it can be implemented on a computer as a computer-implemented method, or implemented in dedicated hardware, or implemented in a combination of the two. The executable code for the method according to the present disclosure can be stored on a computer program product. Examples of computer program products include storage devices, optical storage devices, integrated circuits, servers, online software, etc. Preferably, the computer program product may include a non-transient program code stored on a computer-readable medium, and the non-transient program code is used to perform the method according to the present disclosure when the program product is run on a computer. In one embodiment, the computer program may include a computer program code, and the computer program code is suitable for performing all steps of the 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.

[0075] While the present invention has been described at length and with a certain particularity with respect to several embodiments, it is not intended that the present invention should be limited to any such details or embodiments or to any particular embodiment, but rather the present invention should be construed with reference to the appended claims in order to provide the broadest possible interpretation of such claims in light of the prior art, and thereby effectively encompass the intended scope of the present disclosure.

[0076] All examples and conditional language recorded in this article are intended to be used for teaching purposes, to help readers understand the principles of the present disclosure and the conception contributed by the inventor to further develop this area, and should be interpreted as not being limited to such specific examples and conditions of recording. In addition, all statements and specific examples thereof of the principles, aspects and embodiments of the present disclosure recorded in this article are intended to cover their structure and function equivalents. In addition, it is intended that such equivalents include currently known equivalents and equivalents developed in the future (that is, any element of the performance of the same function developed, regardless of structure).

Claims

1. A computer-implemented method for detecting spinal fractures, comprising: receiving three-dimensional image data including at least a portion of a spine of a subject; identifying the spine of the subject in the three-dimensional image data; defining a spline that approximates a local curvature along the spine of the object; defining a plurality of volumes of interest (VOIs), each VOI containing at least a portion of a vertebra of the spine of the subject, wherein each VOI is defined relative to an adjacent segment of the spline; as well as A fracture is identified in at least one of the VOIs.

2. The method according to claim 1, wherein: The spine is identified by a convolutional neural network (CNN) trained to segment the spine.

3. The method according to claim 1, wherein: For each VOI, a corresponding center point is sampled at a position defined relative to the spline, wherein the center points of the VOI are sampled at regular intervals along the spline.

4. The method according to claim 3, wherein: The center points of adjacent VOIs are located to generate overlapping VOIs.

5. The method according to claim 3, wherein: Each VOI is formed around the center point and is oriented based on a tangent line of the spline adjacent to the corresponding center point.

6. The method according to claim 5, wherein: After each VOI is defined, each VOI is extracted from the three-dimensional image data and resampled to a target resolution.

7. The method according to claim 1, wherein: The fractures were identified by applying a convolutional neural network (CNN) to each VOI.

8. The method according to claim 7, wherein: When the CNN is applied to each VOI, the output of the CNN is a probability map identifying possible fractures within the corresponding VOI.

9. The method according to claim 8, wherein: The VOIs are defined such that adjacent VOIs overlap in order to generate multiple predictions for at least some equivalent voxels present in the multiple VOIs, wherein all predictions corresponding to a particular location are aggregated into a final probability map.

10. The method of claim 8, further comprising generating a final probability map from the probability maps associated with individual VOIs, the final probability map comprising collating the VOI probability maps into a legible representation of the three-dimensional image data.

11. The method according to claim 10, further comprising: A binary prediction is generated based on the final probability map, and fracture candidates are screened based on a relationship between candidate locations and the spine of the subject.

12. The method according to claim 1, wherein: The spline approximates a centerline of the spinal canal of the spine.

13. The method according to claim 1, wherein: A size of the first VOI is selected based on a proximate first location along the spine, and wherein a size of the second VOI is selected based on a proximate second location along the spine.

14. The method according to claim 1, further comprising: A fracture identified in the representation of the three-dimensional image data is located and the identified fracture is displayed in the context of the spine of the subject.

15. An apparatus comprising: a memory storing a plurality of instructions; as well as a processor coupled to the memory and configured to execute the plurality of instructions to: receiving three-dimensional image data including at least a portion of a spine of a subject; identifying the spine of the subject in the three-dimensional image data; defining a spline that approximates a local curvature along the spine of the object; defining a plurality of volumes of interest (VOIs), each VOI containing at least a portion of a vertebra of the spine of the subject, wherein each VOI is defined relative to an adjacent segment of the spline; and A fracture is identified in at least one of the VOIs.

Images