Optimal path lookup based on spinal centerline extraction
By using three-dimensional voxel-level segmentation and programming path search technology in the centerline extraction of spinal canal, the image distortion problem caused by segmentation error is solved, and the accurate extraction and image formatting of spinal canal centerline is achieved, which is suitable for a variety of imaging system.
Patent Information
- Application Number
- CN202380026910.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2022-05-04
- Filing Date
- 2023-03-07
- Publication Date
- 2024-11-22
AI Technical Summary
The prior art is susceptible to segmentation errors when extracting the spinal canal centerline, resulting in unreal reformatted images, especially in emergencies that may ignore critical findings.
By receiving scan data, perform three-dimensional voxel-level spinal tube segmentation, generate an inverse density three-dimensional array, and perform program-based optimal spinal tube path search, gradually increase resolution and reduce size until the predetermined threshold is met, resulting in the final maximum The optimal spinal duct path serves as the center line of the spinal duct.
The centerline extraction of the spinal canal insensitive to segmentation errors is achieved to ensure the accuracy and authenticity of reformatted images, suitable for spinal images generated by various imaging systems.
Smart Images

Figure CN119013695A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of medical imaging, and in particular to a method and device for extracting the centerline of a spinal canal, which is insensitive to segmentation errors. Background Art
[0002] Medical imaging is often used to visualize a patient's anatomy, where the visualized data is used to diagnose disease or injury. In trauma settings, medical personnel often perform imaging scans, such as computed tomography (CT) scans, to diagnose fractures of the ribs and spine.
[0003] To help radiologists read CT spine images, the spine images are often reformatted into a straightened curved plane reformatted (CPR). The reformatted images can be two-dimensional or three-dimensional straight views based on the spinal canal curve. The first step in generating the reformatted image is to extract the centerline of the spinal canal. This is usually done by performing a three-dimensional (3D) spinal canal segmentation and then extracting the segmented centerline. If the segmentation results are good, extracting the centerline may not be difficult. However, there are often some errors in the segmentation. Sometimes, parts of the spinal canal are missed (gaps between slices). In other cases, there are also false segmentations, in which the predicted spinal canal does not reflect the actual spinal canal. These problems pose complex challenges to centerline extraction.
[0004] If the extracted spinal canal centerline is incorrect, the reformatted image will also be incorrect, resulting in unrealistic image distortions within the resulting image. Reading imaging scans, and more specifically trauma or emergency room scans, is a time-critical task that needs to be completed with a high level of concentration to avoid overlooking critical findings. Reformatting the image scan into a straightened view in 2D or 3D is performed according to the spinal canal curve to aid medical professionals in diagnosing severe rib and spinal fractures. A reformatted image with unrealistic image distortions defeats the purpose of the reformatted image.
[0005] Therefore, innovative techniques are needed to extract the center of the spinal canal that are insensitive to segmentation errors. Summary of the invention
[0006] The object of the present invention is to provide a technique for extracting the center of the spinal canal, which is insensitive to segmentation errors. The technique can be applied to almost any spinal image, reformatting it into a 2D or 3D straightened view based on the spinal canal curve. These spinal images can be generated by a variety of imaging systems, including CT, CT arms, single photon emission computed tomography CT (SPECT-CT), magnetic resonance CT (MR-CT), positron emission tomography CT (PET-CT) and magnetic resonance imaging (MRI) systems.
[0007] According to the first aspect of the present invention, a method for generating an optimal spinal canal is provided. The method comprises: receiving scan data as a three-dimensional array of a scanned spinal image; and performing three-dimensional voxel-level spinal canal segmentation according to the received scan data to generate a binary three-dimensional array corresponding to the scanned spinal image; generating an inverse density three-dimensional array corresponding to the generated binary three-dimensional array, the inverse density three-dimensional array having the same size as the generated binary three-dimensional array; downsampling the inverse density three-dimensional array to a predetermined size; performing programming-based optimal spinal canal path search to generate an optimal spinal canal path; sampling the inverse three-dimensional array around the generated optimal spinal canal path at a higher resolution and a limited size; performing programming-based optimal spinal canal path search to generate a revised optimal spinal canal path; repeating the sampling and the programming-based optimal spinal canal path search by gradually increasing the resolution and reducing the size; determining whether the accuracy of the optimal spinal canal path meets a predetermined threshold; and generating a final optimal spinal canal path as a spinal canal centerline, wherein a reformatted image is generated based on the final optimal spinal canal path as the spinal canal centerline.
[0008] According to the second aspect of the present invention, an optimal path finding device is provided. The device comprises: a memory configured to store computer executable instructions; and at least one processor configured to execute the computer executable instructions so that the optimal path finding device performs the following operations: receiving scan data as a three-dimensional array of a scanned spinal image; performing three-dimensional voxel-level spinal canal segmentation according to the received scan data to generate a binary three-dimensional array corresponding to the scanned spinal image; generating an inverse density three-dimensional array corresponding to the generated binary three-dimensional array, the inverse density three-dimensional array having the same size as the generated binary three-dimensional array; downsampling the inverse density three-dimensional array to a predetermined size; performing encoding-based segmentation. The invention relates to a method for performing a programming-based optimal spinal canal path search to generate an optimal spinal canal path; sampling an inverse three-dimensional array around the generated optimal spinal canal path at a higher resolution and a limited size; performing a programming-based optimal spinal canal path search to generate a revised optimal spinal canal path; repeating the sampling and the programming-based optimal spinal canal path search by gradually increasing the resolution and reducing the size; determining whether the accuracy of the optimal spinal canal path meets a predetermined threshold; and generating a final optimal spinal canal path as a spinal canal centerline, wherein a reformatted image is generated based on the final optimal spinal canal path as the spinal canal centerline.
[0009] In a third aspect of the present invention, a non-transitory computer-readable medium is provided on which instructions for causing a processing circuit to perform a process are stored. The process includes: receiving scan data as a three-dimensional array of a scanned spinal image; performing three-dimensional voxel-level spinal canal segmentation based on the received scan data to generate a binary three-dimensional array corresponding to the scanned spinal image; generating an inverse density three-dimensional array corresponding to the generated binary three-dimensional array, the inverse density three-dimensional array having the same size as the generated binary three-dimensional array; downsampling the inverse density three-dimensional array to a predetermined size; performing a programming-based optimal spinal canal path search to generate an optimal spinal canal path; sampling the inverse three-dimensional array around the generated optimal spinal canal path at a higher resolution and a limited size; performing a programming-based optimal spinal canal path search to generate a revised optimal spinal canal path; repeating the sampling and the programming-based optimal spinal canal path search by gradually increasing the resolution and reducing the size; determining whether the accuracy of the optimal spinal canal path meets a predetermined threshold; and generating a final optimal spinal canal path as a spinal canal centerline, wherein a reformatted image is generated based on the final optimal spinal canal path as the spinal canal centerline.
[0010] In an embodiment, spinal canal segmentation is performed using machine or deep learning techniques, wherein a value of 1 is assigned to each spinal canal voxel and a value of 0 is assigned to the background in the generated binary 3D array. A floating value is assigned to each voxel in the inverse density 3D array, wherein the higher the density value of the 1 voxel in the binary 3D array, the lower the floating value assigned to the corresponding voxel in the inverse density 3D array.
[0011] In another embodiment, a loss pointer array is created, the loss pointer array corresponding to the inverse density three-dimensional array, wherein each element in the loss pointer array includes a loss component and a pointer component, wherein each loss component in the loss pointer array corresponds to a voxel at the same position in the inverse density three-dimensional array, and wherein the pointer component is a three-dimensional vector pointing to a physical coordinate in the binary three-dimensional array.
[0012] In a preferred embodiment, the programming-based optimal spinal canal path finding is an optimal spinal canal path finding based on dynamic programming. Density estimation is applied to generate an inverse density three-dimensional array, wherein the density estimation is performed via a kernel density estimation technique, and wherein the kernel is at least one of the following: Gaussian, linear, cosine, uniform, Epanichnikov, exponential, and top hat. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] In the following figures:
[0014] Figure 1 is a schematic diagram of an exemplary imaging system;
[0015] Figure 2 Examples of a straightened CPR without errors in spine segmentation and centerline extraction, and a straightened spine with errors in spine segmentation and centerline extraction are illustrated.
[0016] Figure 3 is a flow chart representing a method of generating a combined rib and spine view according to some embodiments;
[0017] Figure 4 illustrates an example of a coarse-to-fine optimal path finding method according to some embodiments; and
[0018] Figure 5 Three-dimensional optimal path finding is illustrated according to some embodiments. DETAILED DESCRIPTION
[0019] Figure 1 An imaging system 100, such as a computed tomography (CT) imaging system, is illustrated. The imaging system 100 includes a substantially stationary gantry 102 and a rotating gantry 104. The rotating gantry 104 is rotatably supported by the stationary gantry 102 and rotates relative to the stationary gantry 102 about an examination region 106 about a longitudinal or z-axis.
[0020] A patient support 112, such as a couch, supports an object or examination subject, such as a human patient, in the examination region 106. The support 112 is configured to move the object or examination subject to load, scan, and unload the object or examination subject.
[0021] A radiation source 108, such as an X-ray tube, is rotatably supported by the rotating gantry 104. The radiation source 108 rotates together with the rotating gantry 104 and emits radiation that passes through the examination region 106.
[0022] A radiation sensitive detector array 110 is disposed across the examination region 106 in an angular arc relative to the radiation source 108. The detector array 110 includes one or more rows of detectors extending along the z-axis direction, detects radiation passing through the examination region 106, and generates projection data indicative thereof.
[0023] A general purpose computing system or computer is used as an operator console 114 and includes (one or more) input devices 116 (such as a mouse, keyboard, and / or other) and (one or more) output devices 122 (such as a display monitor, film camera, etc.). The console 114 allows the operator to control the operation of the system 100. This includes control of the imaging processing device 118. The imaging processing device includes an optimal path finding device 120. The imaging processing device 118 receives data representing a three-dimensional (3D) diagnostic image generated by the imaging system 100 and generates a reformatted image that shows a continuous and straightened visualization of the thorax and spine. The reformatted image can be output to the output device 122 to be read by medical personnel.
[0024] The optimal path finding device 120 can be configured as part of the imaging processing device 118, or configured outside the imaging processing device 118. It should be understood that the optimal path finding device 120 can be implemented by (one or more) microprocessors, which execute (one or more) computer-readable instructions encoded or embedded on a computer-readable storage medium (such as a physical memory and other non-transient media). Additionally or alternatively, (one or more) microprocessors can execute (one or more) computer-readable instructions carried by carrier waves, signals and other transient (or non-transient) media.
[0025] The optimal path finder 120 receives the scan data as a 3D array of scanned spinal images. The 3D array of scanned spinal images can be generated by the imaging processing device 118 of the imaging system 100. The optimal path finder 120 performs spinal canal segmentation by running a three-dimensional voxel-level spinal canal segmentation based on the received scan data to generate a binary three-dimensional array corresponding to the scanned spinal image. The optimal path finder 120 generates an inverse density 3D array corresponding to the generated binary 3D array, and the inverse density 3D array has the same size as the generated binary three-dimensional array. The basic principles for generating the inverse density 3D array will be explained in more detail later. The optimal path finder 120 performs downsampling of the inverse density 3D array to a predetermined size, and performs programming-based optimal spinal canal path search to generate an optimal spinal canal path.
[0026] A programming-based optimal spinal canal path search is then performed to generate a revised optimal spinal canal path. Sampling and programming-based optimal spinal canal path search are repeated until the accuracy of the optimal spinal path meets a predetermined threshold. Sampling is repeated around the 3D density image volume at a higher resolution and a limited size each time. For each repeated sampling at a higher resolution and a limited size, the programming-based optimal spinal canal path search is repeated, and the accuracy of the generated optimal path improves with each sampling cycle. This is an example of a coarse-to-fine strategy that is implemented to accelerate optimal path search. Once the optimal spinal path meets the predetermined threshold, the final optimal spinal canal path is generated as the spinal canal centerline. Then, the final optimal spinal canal path as the spinal canal centerline can be used to generate a reformatted image.
[0027] Figure 2 204 is an example of a straightened CPR without errors in spine segmentation and centerline extraction and a straightened spine with errors in spine segmentation and centerline extraction. A reformatted image with a correctly extracted spinal canal centerline is shown at 202. A reformatted image with errors in the extracted spinal canal centerline is shown at 204. The reformatted image at 204 illustrates a reformatted image with unrealistic image distortion and gaps between slices. Therefore, the reformatted image at 204 cannot help medical personnel perform a diagnosis.
[0028] Figure 3 is a flow chart representing a method for generating a combined rib and spine view according to one embodiment. Image scan data is received at 302. The image scan data is a 3D array, wherein each element in the 3D array represents a voxel. At 304, 3D voxel-level spinal canal segmentation is performed using the received scan data to generate a binary 3D array corresponding to the scanned spinal image. In the generated 3D binary array, a value of "1" indicates a spinal canal voxel and a value of "0" indicates background. Spinal canal segmentation may be performed by any deep learning or machine learning method or some other conventional algorithm.
[0029] At 306, an inverse density 3D array is generated. In the generated 3D binary array, a value of "1" indicates a spinal canal voxel, and a value of "0" indicates background. After the inverse density 3D array is generated, the higher the density of "1" voxels in the original volume, the lower the value assigned to the voxel in the inverse density volume.
[0030] If the optimal path finding is performed on the entire inverse density volume at full resolution, the process may be very demanding on the available resources and the process may be slow. Therefore, in an embodiment, a process of implementing a coarse-to-fine optimal path finding strategy is implemented to speed up the process. At 308, downsampling of the inverse density 3D array to a predetermined size is performed. In other words, the inverse density 3D array is downsampled to a low resolution array. At 310, a programming-based optimal spinal canal path finding is performed to generate an optimal spinal canal path. Then, at 312, the inverse density 3D array is sampled around the image volume of the optimal spinal canal path generated at 310. The sampling at 312 is performed at a limited size and a higher resolution.
[0031] At 314, a programming-based optimal spinal canal path search is performed on the sampled image volume 312 to generate a revised optimal spinal canal path. The sampling at 312 and the programming-based optimal spinal canal path search at 314 are repeated until the accuracy of the optimal spinal canal meets a predetermined threshold. When the accuracy of the optimal spinal canal meets the predetermined threshold at 316, a final optimal spinal canal path is generated at 318. The final optimal spinal canal path generated at 318 can then be used to generate a reformatted image.
[0032] Figure 4 An example of a coarse-to-fine optimal path finding method according to an embodiment is illustrated.
[0033] At 402, the inverse density 3D array is downsampled to a predetermined resolution. As illustrated at 402, the predetermined resolution is 25 mm x 25 mm x 25 mm. That is, sampling is performed every 25 mm along the z-axis, every 25 mm around the x-axis, and every 25 mm around the y-axis. The resolutions illustrated in 402, 406, and 410 are for ease of illustration and should not be construed as limiting the scope.
[0034] At 404, a programming-based optimal spinal canal path search is performed to generate a low-resolution optimal spinal canal path. At 406, an inverse density 3D array is downsampled in a finite size volume around the optimal spinal canal path generated at 404 at a resolution of 25mm x 5mm x 5mm. Here, sampling is performed every 25mm along the z-axis, every 5mm around the x-axis, and every 5mm along the y-axis. Thus, the resolution is increased.
[0035] At 408, a programming-based optimal spinal canal path search is performed to generate a modified optimal spinal canal path with a higher resolution (intermediate resolution). At 410, downsampling is performed in a limited size volume around the optimal spinal canal path generated at 408 with a resolution of 15mm x1mm x 1mm. Here, sampling is performed every 15mm along the z-axis, every 1mm around the x-axis, and every 1mm around the y-axis. Therefore, the resolution is further improved. At 412, a programming-based optimal spinal canal path search is performed to generate a final optimal spinal canal path with high resolution.
[0036] although Figure 4 The coarse-to-fine strategy illustrated in includes three cycles of downsampling and optimal path finding, but it should be understood that this is for ease of description and is not to be construed as limiting the scope.
[0037] Figure 5 Illustration of the three-dimensional optimal path finding based on loss analysis.
[0038] Typically, the spinal canal moves monotonically from the sacrum upward to the head. After segmentation, in the predicted spinal canal binary 3D array, if high loss is assigned to background voxels and low loss is assigned to predicted spinal canal voxels, then even in the presence of noise, as long as the actual spinal canal voxels dominate the foreground, the path from the bottom slice to the top slice with the lowest loss will be along the actual spinal canal. In addition, for the predicted spinal canal voxels, if we assign low loss to voxels with high density and high loss to voxels with low density, then the shortest path with the lowest loss will be along the centerline of the actual spinal canal, because the center of the spinal canal usually has a higher density.
[0039] Finding the optimal path from the sacrum to the head or from the head to the sacrum can use a dynamic programming method. Starting from the second slice of the 3D binary array, each voxel will calculate the loss from the neighboring voxels in the last slice to it. The neighboring voxels can cover a small area or all the voxels in the last slice. The loss consists of three parts. The first part is the physical distance from the voxel in the last slice to the current voxel in the current slice. The second part is the inverse density of the voxel selected in the last slice. The reason for using the inverse density is to find the optimal path with the lowest cumulative loss, but the path needs to pass voxels as densely as possible. Therefore, the inverse density will reduce the contribution of high density to the cumulative loss.
[0040] Density estimation can use kernel density estimation methods. Kernels can be Gaussian, linear, cosine, etc. Using kernel density can reduce the impact of outliers' contribution to the density of a given voxel. Given a voxel, the estimated density will take into account the number of voxels around it. Density can be 3D density or slice-level 2D density.
[0041] The third part is the loss that has been calculated in the last slice for the selected voxel. For each voxel in the current slice, after calculating the loss from the neighboring voxels in the last slice to it, a pointer is assigned to the current voxel, pointing to the neighboring voxel in the last slice with the lowest loss, and the lowest loss is set as the cumulative loss for the current voxel. After performing the calculation from the first slice to the last slice, each voxel in the last slice has a loss value and a pointer value pointing to the voxel in the second-to-last slice. Then starting from the last slice, the optimal path is found as follows. In the last slice, find the voxel with the lowest loss. This is the starting voxel. According to the pointer assigned to this voxel, find the second voxel in the second-to-last slice. Then according to the pointer assigned to the second voxel, we can find the third voxel in the third-to-last slice. Therefore, by linking all the pointers together slice by slice, an optimal path can be generated.
[0042] As reference Figure 3 As indicated, 3D voxel-level spinal canal segmentation is performed using the received scan data to generate a binary 3D array corresponding to the scanned spinal image. In the generated 3D binary array, the value "1" indicates a spinal canal voxel, and the value "0" indicates the background. An inverse density 3D array is generated. In the generated 3D binary array, the value "1" indicates a spinal canal voxel, and the value "0" indicates the background. After the inverse density 3D array is generated, the higher the density of "1" voxels in the original volume, the lower the value assigned to the voxel in the inverse density volume.
[0043] In a preferred embodiment, an optimal spinal canal path based on dynamic programming is performed. For example, assuming the estimated density of the voxel is d, then 1-d is the inverse density. However, other mathematical calculations can be used to invert the density.
[0044] Figure 5 The inverse density 3D array is shown as a loss pointer array. Based on the inverse density 3D array, another 3D array named loss pointer array is created. The loss pointer array has the same size as the inverse density 3D array. Each element in the loss pointer array corresponds to a voxel at the same position in the inverse density 3D array and contains two parameters: a loss value parameter and a pointer value parameter. The pointer parameter is a 3D vector pointing to the physical coordinates in the original spinal canal segmentation binary image.
[0045] For the elements in the first slice 502 of the loss pointer array, zero values are assigned to all loss values, and each pointer value is set to the coordinates of its own corresponding voxel. If the index of the current slice 504 is set to i+1, then the index of the adjacent slice below the current slice 506 is i. For any voxel V in the current slice 504 of the inverse density array i+1,j, the loss is calculated based on the neighboring voxels in the region defined in slice i 506. Any voxel V in slice i+1 i+1,j With the voxel V in the slice i,k The loss between is calculated according to equation (1).
[0046] Loss (V i+1,j ,V i,k )=Loss(V i,k )+w1·distance(V i+1,j ,V i,k )+w2·inverse_density i,k
[0047] According to equation (1), any voxel V in slice i+1 i+1,j Assign the minimum loss value to any voxel V in the loss pointer array. i+1,j The loss value is, where V i,K With voxel V i+1,j minimum loss.
[0048] Equation (2) describes the i+1,j The loss and pointer, where the coordinates (V i,K ) is the voxel V i,K The physical coordinates of . (2)
[0050]
[0051] This step is a forward step. After calculating the entire loss pointer array, move backward from the last slice 508 of the loss pointer array to find the optimal path in the following manner.
[0052] In the last slice 508 of the loss pointer array, find the element with the lowest loss value. Then, get the corresponding pointer. Assume that the pointer is P l .P l These are actually the physical coordinates of the voxel in the second-to-last slice.
[0053] In the penultimate slice, find the element at position Pl, and then get the corresponding pointer as P l-1 .P l-1 are the physical coordinates of the voxel in the third-to-last slice.
[0054] Slice by slice, we can get the coordinate list as {P l ,P l-1 ,P l-2 ,…,P l-(N-1)}, where N is the total number of slices in the loss pointer array. The coordinate list describes the optimal path.
[0055] While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive; the invention is not limited to the disclosed embodiments.
[0056] Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from a study of the drawings, the disclosure, and the appended claims.
[0057] In the claims, the word "comprising" does not exclude other elements or steps, and the word "a" or "an" does not exclude a plurality.
[0058] A single element or other unit may fulfill the functions of several items recited in the claims.The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.
[0059] The computer program may be stored / distributed on suitable media, such as optical storage media or solid-state media provided together with or as part of other hardware, but the computer program may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunications systems.
Claims
1. A method for generating an optimal spinal canal path, the method comprising: receiving scan data as a three-dimensional array of images of a scanned spine; performing three-dimensional voxel-level spinal canal segmentation based on the received scan data to generate a binary three-dimensional array corresponding to the scanned spinal image; generating an inverse density three-dimensional array corresponding to the generated two-element three-dimensional array, wherein the inverse density three-dimensional array has the same size as the generated two-element three-dimensional array; downsampling the inverse density three-dimensional array to a predetermined size; performing programming-based optimal spinal canal path search to generate an optimal spinal canal path; sampling an inverse three-dimensional array around the generated optimal spinal canal path at a higher resolution and limited size; performing programming-based optimal spinal canal path search to generate a revised optimal spinal canal path; Repeating the sampling and the programming-based optimal spinal canal path finding by gradually increasing the resolution and decreasing the size; determining whether the accuracy of the optimal spinal canal path meets a predetermined threshold; and A final optimal spinal canal path is generated as a spinal canal centerline, wherein a reformatted image is generated according to the final optimal spinal canal path as the spinal canal centerline.
2. The method for generating an optimal spinal canal path according to claim 1, wherein: Spinal canal segmentation is performed using machine or deep learning techniques, wherein a value of 1 is assigned to each spinal canal voxel and a value of 0 is assigned to the background in the generated binary three-dimensional array.
3. The method for generating an optimal spinal canal path according to claim 2, wherein: Generating the inverse density three-dimensional array includes assigning a floating value to each voxel, wherein the higher the density value of a voxel in the binary three-dimensional array, the lower the floating value assigned to the corresponding voxel in the inverse density three-dimensional array.
4. The method for generating an optimal spinal canal path according to claim 1, wherein: The programming-based optimal spinal canal path search includes: Create a loss pointer array, which is a three-dimensional array corresponding to the inverse density three-dimensional array, wherein each element in the loss pointer array includes a loss component and a pointer component, wherein each loss component in the loss pointer array corresponds to a voxel at the same position in the inverse density three-dimensional array, and wherein the pointer component is a three-dimensional vector pointing to a physical coordinate in the binary three-dimensional array.
5. The method for generating an optimal spinal canal path according to claim 1, wherein: The optimal spinal canal path search based on programming is an optimal spinal canal path search based on dynamic programming.
6. The method for generating an optimal spinal canal path according to claim 1, wherein: Density estimation is applied to generate the inverse density three-dimensional array.
7. The method for generating an optimal spinal canal path according to claim 6, wherein: The density estimation is performed via a kernel density estimation technique, and wherein the kernel is at least one of: Gaussian, linear, cosine, uniform, Epanichinnikov, exponential, and top hat.
8. An optimal path finding device, comprising: a memory configured to store computer-executable instructions; as well as At least one processor is configured to execute the computer executable instructions so that the optimal path finding device performs the following operations: receiving scan data as a three-dimensional array of images of a scanned spine; performing three-dimensional voxel-level spinal canal segmentation based on the received scan data to generate a binary three-dimensional array corresponding to the scanned spinal image; generating an inverse density three-dimensional array corresponding to the generated two-element three-dimensional array, wherein the inverse density three-dimensional array has the same size as the generated two-element three-dimensional array; downsampling the inverse density three-dimensional array to a predetermined size; performing programming-based optimal spinal canal path search to generate an optimal spinal canal path; sampling an inverse three-dimensional array around the generated optimal spinal canal path at a higher resolution and limited size; performing programming-based optimal spinal canal path search to generate a revised optimal spinal canal path; Repeating the sampling and the programming-based optimal spinal canal path finding by gradually increasing the resolution and decreasing the size; determining whether the accuracy of the optimal spinal canal path meets a predetermined threshold; and A final optimal spinal canal path is generated as a spinal canal centerline, wherein a reformatted image is generated according to the final optimal spinal canal path as the spinal canal centerline.
9. The optimal path finding device according to claim 8, wherein: Spinal canal segmentation is performed using machine or deep learning techniques, wherein a value of 1 is assigned to each spinal canal voxel and a value of 0 is assigned to the background in the generated binary three-dimensional array.
10. The optimal path finding device according to claim 9, wherein: The processor is configured to assign a floating value to each voxel in the generated density three-dimensional array, wherein the higher the density value of the voxel in the binary three-dimensional array, the lower the floating value assigned to the corresponding voxel in the inverse density three-dimensional array.
11. The optimal path finding device according to claim 8, wherein: The processor is also configured to create a loss pointer array, which is a three-dimensional array corresponding to the inverse density three-dimensional array, wherein each element in the loss pointer array includes a loss component and a pointer component, wherein each loss component in the loss pointer array corresponds to a voxel at the same position in the inverse density three-dimensional array, and wherein the pointer component is a three-dimensional vector pointing to a physical coordinate in the binary three-dimensional array.
12. The optimal path finding device according to claim 8, wherein: The processor is configured to perform the programming-based optimal spinal canal path finding via dynamic programming techniques.
13. The optimal path finding device according to claim 8, wherein: The processor is configured to apply density estimation to generate the inverse density three-dimensional array.
14. The optimal path finding device according to claim 13, wherein: The density estimation is performed via a kernel density estimation technique, and wherein the kernel is at least one of: Gaussian, linear, cosine, uniform, Epanichinnikov, exponential, and top hat.
15. A non-transitory computer-readable medium having stored thereon instructions for causing a processing circuit to perform a process, the process comprising: receiving scan data as a three-dimensional array of images of a scanned spine; performing three-dimensional voxel-level spinal canal segmentation based on the received scan data to generate a binary three-dimensional array corresponding to the scanned spinal image; generating an inverse density three-dimensional array corresponding to the generated two-element three-dimensional array, wherein the inverse density three-dimensional array has the same size as the generated two-element three-dimensional array; downsampling the inverse density three-dimensional array to a predetermined size; performing programming-based optimal spinal canal path search to generate an optimal spinal canal path; sampling an inverse three-dimensional array around the generated optimal spinal canal path at a higher resolution and limited size; performing programming-based optimal spinal canal path search to generate a revised optimal spinal canal path; Repeating the sampling and the programming-based optimal spinal canal path finding by gradually increasing the resolution and decreasing the size; determining whether the accuracy of the optimal spinal canal path meets a predetermined threshold; and A final optimal spinal canal path is generated as a spinal canal centerline, wherein a reformatted image is generated according to the final optimal spinal canal path as the spinal canal centerline.
