Deep learning-based dorsal root ganglion positioning method under foramen intervertebrale radiography

Abstract

The invention relates to the technical field of spine CT image processing, in particular to a deep learning-based dorsal root ganglion positioning method under foramen intervertebrale radiography. The method comprises the following steps: inputting a plurality of spine CT images of a user at different scanning angles into a neural network model to obtain a neural recognition region of each spine CT image and image areas of dorsal root ganglions in corresponding compression states at different scanning angles; performing de-overlapping processing based on gray activation differences of the plurality of pooling layers to determine an actual outer boundary of the neural recognition region; according to the corresponding relation between the image area of each spine CT image and the intervertebral space, the position of the actual outer boundary inner dorsal root ganglion is subjected to credibility labeling; according to the credibility labeling result of the position where the dorsal root ganglion is located, the dorsal root ganglion of the user is visually displayed. According to the invention, the technical problem that the identification accuracy is influenced by similar tissues around the dorsal root ganglion can be reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of spine CT image processing, and in particular to a dorsal root ganglion positioning method under intervertebral foramen contrast based on deep learning. BACKGROUND

[0002] The dorsal root ganglion (DRG) is a cell body aggregation area of sensory neurons, and is a necessary path for transmitting pain and touch signals to the central nervous system. Abnormalities of the dorsal root ganglion, such as ectopic location and damage, are closely related to lumbar and sacral nerve root canal stenosis and chronic pain. Accurate identification is not only crucial for basic research, but also a key prerequisite for the development of precision medicine and neural regulation technology. Therefore, accurate identification of the dorsal root ganglion is of great significance to neuroscience, clinical medicine and pain management. However, traditional methods often use neural network models to determine the location of the dorsal root ganglion, which has the problem of insufficient recognition accuracy due to various recognition interferences. SUMMARY

[0003] In order to reduce the influence of the morphologically similar tissues around the dorsal root ganglion on its recognition accuracy, the present application aims to provide a dorsal root ganglion positioning method under intervertebral foramen contrast based on deep learning, and the technical solution adopted is as follows: The present application provides a dorsal root ganglion positioning method under intervertebral foramen contrast based on deep learning, which comprises the following steps: Inputting multiple spine CT images of different scanning angles of a user into a neural network model to obtain a neural recognition area of each spine CT image and an image area of the dorsal root ganglion under different scanning angles and corresponding compression states; According to the gray level activation difference of multiple pooling layers in the neural network model, the neural recognition area is de-overlapped to determine the actual outer boundary of the neural recognition area, wherein the sizes of the pooling windows of the multiple pooling layers are different; According to the corresponding relationship between the image area of each spine CT image and the intervertebral space, the position of the dorsal root ganglion within the actual outer boundary is marked with a confidence level; According to the actual outer boundary of the neural recognition area and the confidence level marking result of the position of the dorsal root ganglion, the dorsal root ganglion of the user is visually displayed.

[0004] In an alternative embodiment, according to the gray level activation difference of multiple pooling layers in the neural network model, the neural recognition area is de-overlapped to determine the actual outer boundary of the neural recognition area, which comprises: According to the size of the pooling window corresponding to each pooling layer, the neural recognition area of each spine CT image is maximum-pooled to obtain the activation value of the neural recognition area under different pooling windows; The activation values ​​of the same area on each spinal CT image under each pooling layer are statistically analyzed and normalized to obtain the edge consistency of the statistical area; When the edge conformity of the statistically counted area is greater than a preset conformity threshold, the corresponding statistically counted area is determined to be the actual outer boundary of the neural recognition area.

[0005] In an optional embodiment, the activation values ​​of the same region on each spinal CT image under each pooling layer are statistically analyzed and normalized to obtain the edge conformance of the statistically analyzed region, including: Get the activation value of the statistical area in each pooling window arranged from small to large; Calculate the activation value difference of the statistical area in adjacent pooling layers, and count the number of positive activation value differences; The sum of all pooling layer activation values ​​in the same statistical area is multiplied by the statistical quantity and normalized based on the preset range to obtain the edge conformity of the corresponding statistical area.

[0006] In an optional embodiment, the reliability of the location of the dorsal root ganglion within the actual outer boundary is marked based on the correspondence between the image area of ​​each spinal CT image and the intervertebral space, including: According to the correspondence between the image area of ​​each spinal CT image and the intervertebral space, the actual maximum area of ​​the current dorsal root ganglion in the uncompressed state is obtained; According to the reference number of ganglia adjacent to the current dorsal root ganglion, the theoretical maximum area and the actual maximum area of ​​the current dorsal root ganglion, the area conformity of the current dorsal root ganglion is obtained; According to the area consistency of all dorsal root ganglia, the locations of the dorsal root ganglia within the actual outer boundary are labeled with confidence.

[0007] In an optional embodiment, obtaining the actual maximum area of ​​the current dorsal root ganglion in an uncompressed state based on the correspondence between the image area of ​​each spinal CT image and the intervertebral space includes: The image areas of adjacent dorsal root ganglia in each spinal CT image were analyzed for interval distance to obtain the first scanning angle of the superior and inferior vertebrae at the shortest distance in the intervertebral space where the dorsal root ganglia were located. The image areas represented by the same dorsal root ganglion in each spinal CT image are compared to obtain a second scanning angle at which the dorsal root ganglion has the largest image area; According to the angle difference between the first scanning angle and the second scanning angle and the shortest distance of the interval distance analysis, the area restoration degree of the current dorsal root ganglion at the corresponding scanning angle is obtained; According to the maximum image area and the maximum area reduction degree compared by the image area, the actual maximum area of the current dorsal root ganglion in the non-compression state is obtained.

[0008] In an optional embodiment, according to the angle difference between the first scanning angle and the second scanning angle, the shortest distance parsed by the interval distance, the area reduction degree of the current dorsal root ganglion under the corresponding scanning angle is obtained, comprising: The minimum extreme value of all the shortest distances parsed by the interval distance is obtained to obtain the minimum target distance; The normalization processing of the angle difference is negatively correlated as the first analysis coefficient, and the normalization processing of the minimum target distance is negatively correlated as the second analysis coefficient; According to the product of the first analysis coefficient and the second analysis coefficient, the area reduction degree of the current dorsal root ganglion under the corresponding scanning angle is obtained.

[0009] In an optional embodiment, according to the reference number of the ganglion adjacent to the current dorsal root ganglion, the theoretical maximum area and the actual maximum area of the current dorsal root ganglion, the area compliance of the current dorsal root ganglion is obtained, comprising: According to the area difference between the theoretical maximum area and the actual maximum area of the current dorsal root ganglion, the area difference absolute value of the current dorsal root ganglion is obtained; According to the product of the third reciprocal of the area difference absolute value and the reference number of the ganglion, the area compliance of the current dorsal root ganglion is obtained.

[0010] In an optional embodiment, before the dorsal root ganglion of the user is visually displayed, the method further comprises: According to the similar morphological interference of the contrast agent on the presentation of the vascular network on the dorsal root ganglion on the spine CT image, the reliability of the position of the dorsal root ganglion is updated.

[0011] In an optional embodiment, according to the similar morphological interference of the contrast agent on the presentation of the vascular network on the dorsal root ganglion on the spine CT image, the reliability of the position of the dorsal root ganglion is updated, comprising: The maximum gray value and the minimum gray value of the position of the current dorsal root ganglion, the first gray mean value of the contrast agent interference area in all spine CT images, and the second gray mean value of the current dorsal root ganglion in the current spine CT image are obtained; According to the maximum gray value, the minimum gray value, the first gray mean value, the second gray mean value, and the area compliance representing the reliability of the position of the current dorsal root ganglion, the recognition failure degree of the current dorsal root ganglion is obtained; According to the comparison result of the recognition failure degree of each dorsal root ganglion and the preset failure degree threshold, the reliability of the position of the corresponding dorsal root ganglion is determined.

[0012] In an optional embodiment, the identification error degree of the current dorsal root ganglion is obtained according to the maximum gray value, the minimum gray value, the first gray mean value, the second gray mean value and the area coincidence degree representing the reliability of the position of the current dorsal root ganglion, comprising: The gray difference value of the current dorsal root ganglion in each spine CT image is obtained according to the absolute value of the difference between the first gray mean value and the second gray mean value; The cumulative gray difference of the current dorsal root ganglion is obtained according to the cumulative summation result of the gray difference values of the images containing the current dorsal root ganglion in all spine CT images; The identification error degree of the corresponding dorsal root ganglion is obtained according to the maximum gray value, the minimum gray value, the cumulative gray difference and the area coincidence degree of each dorsal root ganglion.

[0013] The present application has the following beneficial effects: When the dorsal root ganglion is positioned under the intervertebral foramen radiography based on deep learning, the present application inputs multiple spine CT images at different scanning angles into a neural network model to obtain a neural recognition area of each spine CT image and an image area of the dorsal root ganglion under a corresponding compression state at different scanning angles. Since the sizes of the pooling windows of the multiple pooling layers in the neural network model are different, the neural recognition area can be de-overlapped based on the gray activation difference of the multiple pooling layers to determine the actual outer boundary of the neural recognition area. Further, according to the corresponding relationship between the image area of each spine CT image and the intervertebral space, the position of the dorsal root ganglion in the actual outer boundary is marked with reliability. According to the actual outer boundary of the neural recognition area and the reliability marking result of the position of the dorsal root ganglion, the dorsal root ganglion of the user is visually displayed. When the dorsal root ganglion of the spine CT image is positioned, the present application considers the shape similarity of the blood vessels and the deformation of the dorsal root ganglion under compression, and adopts corresponding technical means to exclude interference, thereby reducing the technical problem that the surrounding shape similar tissues of the dorsal root ganglion affect the identification accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, and the advantages thereof, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without any creative effort.

[0015] Figure 1 A flowchart of a dorsal root ganglion positioning method under intervertebral foramen radiography based on deep learning provided by an embodiment of the present application; Figure 2Fig. 1 is a schematic diagram of a dorsal root ganglion and peripheral tissue according to an embodiment of the present application; Figure 3 Fig. 4 is a logic diagram of a method for locating a dorsal root ganglion according to an embodiment of the present application.

[0016] Fig. 1 is a schematic diagram of a dorsal root ganglion and peripheral tissue according to an embodiment of the present application; DETAILED DESCRIPTION

[0017] In order to further clarify the technical means and effects taken by the present application to achieve the predetermined object, the following describes in detail the specific implementation, structure, features and effects of a method for locating a dorsal root ganglion under fluoroscopy of an intervertebral foramen based on deep learning according to the present application, in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0019] The following specifically describes a specific scheme of a method for locating a dorsal root ganglion under fluoroscopy of an intervertebral foramen based on deep learning according to the present application, in combination with the accompanying drawings.

[0020] Please refer to Figure 1 Fig. 1 is a schematic diagram of a dorsal root ganglion and peripheral tissue according to an embodiment of the present application; S11, input multiple CT images of the user's spine at different scanning angles into a neural network model to obtain a neural recognition area of each CT image of the spine, and an image area of the dorsal root ganglion under different scanning angles in a corresponding compression state.

[0021] Specifically, the user can be a patient who needs to implement a spine diagnosis and treatment, and the included angle between the back plane of the patient and the electric scanning bed is the scanning angle. Through multiple scanning angles, the pathological condition of the spinal tissue, such as whether the dorsal root ganglion exists position abnormality, damage, etc., can be more accurately judged. The accurate identification of the dorsal root ganglion is not only crucial to basic research, but also a key prerequisite for the development of precision medicine and neural regulation technology. The neural network model can be a convolutional neural network model trained based on a training set of spine CT images. The positions of the intervertebral foramen, the upper and lower edges of the pedicle, the dorsal root ganglion, etc. in the training set of spine CT images are labeled to implement model training until the model training is completed. After inputting multiple spine CT images of the user into the model, the neural recognition area of each spine CT image can be obtained. The neural recognition area is the outer boundary of all nerves on the spine CT image. Since there is different degree of compression of the spinal nerves under different angles, the image area of each spine CT image can represent the change characteristics of the dorsal root ganglion under the corresponding compression state at different scanning angles.

[0022] Please refer to Figure 2 , Figure 2 is a schematic view of the dorsal root ganglion and the surrounding tissue. When identifying the neural recognition area based on multiple spine CT images, the multi-angle photographed images of the spine CT images can be used as input data. The images need to include the vertebral body, the pedicle and the nerve root structure, as shown in Figure 2 , there is an intervertebral space between adjacent vertebrae, the spinal nerve is attached to the vertebrae under the junction of blood vessels and other tissues, and the dorsal root ganglion exists on the spinal nerve. When identifying the neural recognition area based on the neural network model, the Faster R-CNN can be used to locate the bony structures such as the pedicle and the intervertebral foramen, and the region proposal network (RPN) can be used to generate candidate boxes. The U-Net structure is integrated to segment the neural tissue (such as DRG), and the multi-scale features are retained through the jump connection. Combining the deep feature extraction ability of ResNet-50 (solving gradient disappearance) and the segmentation accuracy of U-Net, an end-to-end positioning neural network model that can identify the positions of multiple tissues is realized.

[0023] S12, according to the gray activation difference of the plurality of pooling layers in the neural network model, the neural recognition area is de-overlapped to determine the actual outer boundary of the neural recognition area, wherein the sizes of the pooling windows of the plurality of pooling layers are different.

[0024] Specifically, due to the problem of scanning angle, the part of nerve edge displayed in the nerve recognition area in the spine CT image may cause the recognition error of the nerve recognition area due to the position overlap with another nerve recognition area, thereby affecting the identification of the dorsal root ganglion position. When the neural network model performs feature recognition, in different pooling layers, due to the different sizes of the pooling window, different activation values of the nerve edge in the spine CT image are caused. Thus, the actual outer boundary of the nerve recognition area can be determined by determining the hidden nerve edge position according to the activation value change degree of the corresponding position in different pooling layers. Please continue to refer to Figure 2 , the region A shown in the figure has a relatively complex tissue structure, and the actual outer boundary of the nerve recognition area can be determined based on the above-mentioned manner.

[0025] For example, step S12 includes sub-steps S12-1 to S12-3, which are specifically described as follows: S12-1, according to the size of the pooling window corresponding to each pooling layer, the maximum pooling processing is performed on the nerve recognition area of each spine CT image to obtain the activation value of the nerve recognition area under different pooling windows.

[0026] Specifically, the sizes of the pooling windows of different pooling layers can be configured as 1x1, 2 3 , 4x4, etc., in pixels, which are respectively used for pooling operation, i.e. maximum pooling processing, on the nerve recognition area in the spine CT image. When the number of pixel points commonly contained in the pooling windows in different pooling layers is the most, the two pooling windows correspond to the same CT image region. For the region p where the pooling window i in the pooling layer with the size of 1x1 is located, the activation values of this region and the same region in other pooling layers k are obtained .

[0027] S12-2, the activation values of the same region under each pooling layer on each spine CT image are counted and normalized to obtain the edge coincidence degree of the counted region. It can be understood that when the size of the pooling window is larger, the edge region of the nerve may be divided into more normal regions, thereby more hiding the edge features. Therefore, the activation values of the same nerve edge region in the pooling layers with the size of the pooling window from small to large will gradually decrease, and based on this feature, the activation values are counted and normalized, so as to calculate the edge coincidence degree of the counted region. The edge coincidence degree represents whether the counted activation value region is the actual outer boundary of the nerve recognition area.

[0028] For example, sub-step S12-2 specifically includes the following implementation: The first step is to obtain the activation values of the statistical region in the pooling window of each pooling layer arranged from small to large. Based on the ascending order of the pooling window size of each pooling layer, for example, the pooling window is 1x1, the first is arranged, the pooling window is 2x2, the second is arranged, and so on, until the arrangement is completed, and the activation values of the neural recognition region under different pooling windows are obtained.

[0029] The second step is to calculate the difference of the activation values of the statistical region in the adjacent pooling layer, and count the number of positive activation value difference. The difference between the activation values of the same region p in the adjacent pooling layer "k-1, k" can be calculated (different positive and negative), which is recorded as the activation value difference. The number of the number of positive activation value difference is counted.

[0030] The third step is to multiply the sum of the activation values of all pooling layers in the same statistical region by the number of statistics, and to normalize the sum based on the preset range to obtain the edge compliance of the corresponding statistical region. The sum of the activation values of the same region p in multiple pooling layers is calculated . When the sum of the activation values is larger, and the number of statistics is larger, the region p in the neural recognition region is more in line with the characteristics of the neural mask boundary, and it is more likely to be the outer boundary of the nerve. Thus, the statistical result of the region p in which the pooling window i of the pooling layer with the pooling window size of 1x1 is obtained , wherein is the number of all pooling layers, and norm represents linear normalization processing. The statistical result is normalized by using the maximum and minimum normalization method to obtain the edge compliance , and the value range is configured as [0, 1].

[0031] S12-3, when the edge compliance of the statistical region is greater than the preset compliance threshold, it is determined that the corresponding statistical region is the actual outer boundary of the neural recognition region. The compliance threshold can be configured as 0.7, or other values based on actual requirements. When the edge compliance , it is determined that the region p is the actual outer boundary of the neural recognition region; otherwise, when the edge compliance is less than or equal to the compliance threshold, it is determined that the corresponding statistical region is the non-outer boundary of the neural recognition region. The actual outer boundary of the neural recognition region is re-determined for the spine CT image, the dorsal root ganglion region is distinguished by the neural network model, and the region to which the dorsal root ganglion belongs in the spine CT image is marked.

[0032] S13, according to the corresponding relationship between the image area of each spine CT image and the intervertebral space, the position of the dorsal root ganglion in the actual outer boundary is marked with credibility.​​

[0033] Specifically, various factors such as spinal degeneration, trauma, disease or biomechanical imbalance can cause the intervertebral space to narrow, thereby causing compression deformation of the dorsal root ganglion. The occurrence of deformation can cause misjudgment or omission of the dorsal root ganglion at a single location, so the actual size of the dorsal root ganglion should be restored according to the compression condition of the dorsal root ganglion. This step can judge the compression degree of the current dorsal root ganglion according to the change of the intervertebral space, restore the actual size of the dorsal root ganglion, and realize the credibility labeling of the location of the dorsal root ganglion.

[0034] For example, step S13 includes sub-steps S13-1 to S13-3, which are described as follows: S13-1, according to the corresponding relationship between the image area of each spinal CT image and the intervertebral space, the actual maximum area of the current dorsal root ganglion in the un-compressed state is obtained. There can be multiple dorsal root ganglions on each spinal CT image, and the current dorsal root ganglion is denoted as dorsal root ganglion g. The compression degree received by the dorsal root ganglion is different at different scanning angles, and based on the change of the image area, the corresponding intervertebral space can be obtained, and the corresponding relationship between the two can be analyzed to obtain the actual maximum area of the current dorsal root ganglion g in the un-compressed state.

[0035] Specifically, sub-step S13-1 includes: First, the image areas of adjacent dorsal root ganglions in each spinal CT image are analyzed at an interval distance to obtain the first scanning angle of the upper and lower vertebral bodies of the intervertebral space where the dorsal root ganglion is located at the shortest distance. The vertebral bones at different positions in the spinal CT image are recognized by the neural network model to obtain the upper and lower vertebral bodies where the dorsal root ganglion g is located, and the shortest distance between the two vertebral bodies in the spinal CT image m at different scanning angles The shortest distance The corresponding scanning angle is the first scanning angle.

[0036] Second, the image areas represented by the same dorsal root ganglion in each spinal CT image are compared to obtain the second scanning angle of the dorsal root ganglion at the maximum image area. By comparison, the maximum image area of the dorsal root ganglion g in the spinal CT image m at different scanning angles can be obtained, which is denoted as The maximum image area The angle between the CT machine and the posterior plane of the patient during CT scanning is denoted as the second scanning angle.

[0037] Thirdly, according to the angle difference between the first scanning angle and the second scanning angle and the shortest distance parsed from the interval distance, the area reduction degree of the current dorsal root ganglion under the corresponding scanning angle is obtained. When the maximum image area of the dorsal root ganglion g is obtained, the smaller the angle difference between the scanning angle of the patient and the angle when the shortest distance is obtained, the greater the degree of compression of the current dorsal root ganglion g, and the greater the difference between the maximum image area obtained at present and the maximum area of the dorsal root ganglion g under normal circumstances. Conversely, when the shortest distance is smaller and the angle difference is smaller, the maximum image area of the dorsal root ganglion g is smaller relative to the area of the dorsal root ganglion g without compression. Based on the above relationship, a corresponding relationship between the angle difference, the shortest distance and the area reduction degree can be established, and then the area reduction degree of each dorsal root ganglion under the corresponding scanning angle can be obtained.

[0038] For example, the minimum extreme value of all the shortest distances parsed from the interval distance is obtained to obtain the minimum target distance; the normalization processing of the angle difference in negative correlation is taken as the first analysis coefficient; the normalization processing of the minimum target distance in negative correlation is taken as the second analysis coefficient; and the product of the first analysis coefficient and the second analysis coefficient is used to obtain the area reduction degree of the current dorsal root ganglion under the corresponding scanning angle.

[0039] The normalization processing in negative correlation can be specifically exp(-x) or norm(-x), and no limitation is made thereto.

[0040] All the shortest distances After the minimum extreme value is obtained, the minimum target distance can be obtained , and thus the area reduction degree of the dorsal root ganglion g can be obtained , wherein is the angle difference, is the first analysis coefficient, is the minimum target distance, is the second analysis coefficient. The maximum-minimum normalization method can also be used to normalize the area reduction degree to obtain , the value range of which is (0, 1).

[0041] Fourthly, according to the maximum image area and the maximum area reduction degree obtained by comparing the image areas, the actual maximum area of the current dorsal root ganglion under the non-compression state is obtained. When the area reduction degree is greater, the maximum image area of the dorsal root ganglion g is greater when it is not compressed, and thus the actual maximum area of the dorsal root ganglion g when it is not compressed can be obtained, which is denoted as . The following relationship exists: , is the area reduction degree after normalization processing. The maximum image area of the dorsal root ganglion g corresponds to the CT image m of the spine.

[0042] S13-2, according to the number of ganglion references adjacent to the current dorsal root ganglion, the theoretical maximum area and the actual maximum area of the current dorsal root ganglion, the area compliance of the current dorsal root ganglion is obtained. The number of ganglion references is a plurality of dorsal root ganglions adjacent to the current dorsal root ganglion, for example, 3 dorsal root ganglions above and below the current dorsal root ganglion g, and the number of ganglion references is 6. The theoretical maximum area of the current dorsal root ganglion can be obtained based on the experimental database. Although there are individual differences in dorsal root ganglions, the area difference is not too large, so the reference calculation can be performed based on the theoretical maximum area of the current dorsal root ganglion to obtain the area compliance of the current dorsal root ganglion. The area compliance is determined based on the actual maximum area whether it has a reliable reference quantity, and the area compliance can be recorded as .

[0043] It can be understood that the size of the dorsal root ganglion varies significantly at different positions of the spine (i.e., different vertebral levels). These differences are mainly due to the change of the vertebral segment position. For example, in the lumbar segment, the size of the dorsal root ganglion increases from L1 to L5. In the CT image of the spine, the recognition of the dorsal root ganglion (DRG) can be wrong due to the presence of blood stasis, etc., at this time, the size of the dorsal root ganglion below the position of the vertebral segment can be distinguished by detection. Therefore, the area compliance of the current dorsal root ganglion can be obtained based on the following way.

[0044] First, according to the area difference between the theoretical maximum area and the actual maximum area of the current dorsal root ganglion, the absolute value of the area difference of the current dorsal root ganglion is obtained. The average maximum area size of the dorsal root ganglion g in the experimental data is obtained , the average maximum area size of the three dorsal root ganglions above and below the detected dorsal root ganglion g is determined as the theoretical maximum area. The actual maximum area is recorded as , and the absolute value of the area difference is .

[0045] Second, according to the product of the third inverse of the absolute value of the area difference and the number of ganglion references, the area compliance of the current dorsal root ganglion is obtained. The number of ganglion references is recorded as When the image area size of the dorsal root ganglion g is more consistent with the reference dorsal root ganglion of the actual spine at different positions, the number of reference dorsal root ganglions is larger, and the absolute value of the area difference is smaller, the recognition of the current dorsal root ganglion g by the neural network model is more accurate. Thus, the area compliance of the dorsal root ganglion g is obtained .

[0046] S13-3. Based on the area conformance of all dorsal root ganglia, the locations of the dorsal root ganglia within the actual outer boundary are labeled with confidence. The area conformance indicates whether the image area of ​​the corresponding dorsal root ganglia is reliable, and thus credibility labeling can be implemented based on this. The credibility can be labeled with different levels or quantified as a percentage for easy reference by users or medical personnel.

[0047] In actual applications, since blood vessels are similar to nerves in distribution morphology, protrusions such as thrombi on them may be mistakenly identified as dorsal root ganglia. Compared with blood vessels, contrast agents diffuse spatially homogenously in nerve tissue, and there is no clearance or washout over time. In order to eliminate the interference caused by blood vessels, the presence and changes of contrast agents in different regions can be analyzed. Based on this, the positioning method of the present invention also includes: updating the credibility of the position of the dorsal root ganglia based on the similar morphological interference of the contrast agent on the vascular network presented on the dorsal root ganglia on the spinal CT image. Specifically including: S13-4. Obtain the maximum grayscale value and minimum grayscale value of the current dorsal root ganglion location, the first grayscale mean of the contrast agent interference area in all spinal CT images, and the second grayscale mean of the current dorsal root ganglion in the current spinal CT image. The range of the image grayscale after processing is 0-255. Based on the image grayscale of the current dorsal root ganglion location, the maximum grayscale value and the minimum grayscale value can be parsed. Some areas are interfered by contrast agents, such as areas with blood vessels. The image grayscale of the area is parsed and the average grayscale value is calculated to obtain the first grayscale mean. The first grayscale mean characterizes the grayscale mean of each position affected by the contrast agent in the spinal CT image. The second grayscale mean is calculated based on the image grayscale of the area to which the current dorsal root ganglion belongs in the current spinal CT image.

[0048] S13-5. Obtain the degree of error in identifying the current dorsal root ganglion based on the maximum grayscale value, the minimum grayscale value, the first grayscale mean, the second grayscale mean, and the area consistency indicating the reliability of the current dorsal root ganglion location. A data processing model can be constructed based on the correlation between each value and the degree of error in identifying the current dorsal root ganglion g. All data can be input into the data processing model to determine the degree of error in identifying the current dorsal root ganglion g.

[0049] Furthermore, the degree of recognition error can be calculated based on the following methods, including: The first step is to obtain the grayscale difference of the current dorsal root ganglion in each spinal CT image based on the absolute value of the difference between the first grayscale mean and the second grayscale mean. The first grayscale mean is recorded as , the second grayscale mean is recorded as , the grayscale difference is .

[0050] Second, according to the cumulative sum of the gray value difference of the current dorsal root ganglion image in all the spine CT images, the cumulative gray value difference of the current dorsal root ganglion is obtained. The greater the difference between the maximum gray value and the minimum gray value of the dorsal root ganglion g position in the stage of obtaining the spine CT image, and the greater the gray value difference in the multiple spine CT images , the cumulative sum result ( is the number of spine CT images containing the dorsal root ganglion g position), the smaller the area coincidence , the faster the contrast agent in the dorsal root ganglion g position recognized by the neural network model dissipates, and the greater the difference in gray value from other areas, the less the size matches the actual, the less it matches the display of the dorsal root ganglion g, and the more problems there are in the identification of this position.

[0051] Third, according to the maximum gray value, the minimum gray value, the cumulative gray value difference and the area coincidence of each dorsal root ganglion, the identification error degree of the corresponding dorsal root ganglion is obtained. The maximum gray value is denoted as , the minimum gray value is denoted as , based on the above analysis, the identification error degree of the dorsal root ganglion g recognized by the neural network model is obtained, and the identification error degree is denoted as , ; the identification error degree can be normalized by using the maximum and minimum normalization method to obtain , whose value range is (0, 1).

[0052] S13-6, according to the comparison result of the identification error degree of each dorsal root ganglion and the preset error degree threshold, the credibility of the position where the corresponding dorsal root ganglion is located is determined. The error degree threshold can be set based on actual conditions, for example, it is set to 0.8, when , it is determined that the dorsal root ganglion g recognized by the neural network model does not match the actual one, indicating that the identification of the neural network model has problems. When the identification error degree of the dorsal root ganglion is greater than the preset error degree threshold, it is concluded that the credibility of the position where the corresponding dorsal root ganglion is located is low; when the identification error degree is less than or equal to the preset error degree threshold, it is concluded that the credibility of the position where the corresponding dorsal root ganglion is located is high.

[0053] S14, according to the actual outer boundary of the neural recognition area and the credibility labeling result of the position where the dorsal root ganglion is located, the dorsal root ganglion of the user is visually displayed.

[0054] Specifically, the reliability of each dorsal root ganglion can be obtained based on the above method, and image labeling is implemented, and the dorsal root ganglion of the user is visually displayed through a display terminal. Further, the position of the dorsal root ganglion identified by the neural network model can also be judged by the above method, and the obtained error degree is matched with the corresponding position and stored in the database. Different dorsal root ganglion positions and positions with judgment errors are marked in the image using SQL (Structured Query Language), and visual display is performed.

[0055] The positioning process of the dorsal root ganglion will be described in the following embodiments of the present application. Please refer to Figure 3 , Figure 3 The overall logic diagram of the dorsal root ganglion positioning method. Specifically, it includes: S301, the intervertebral foramen, the upper and lower edges of the pedicle of the patient's spine, and the position of the dorsal root ganglion are obtained by a deep learning model.

[0056] S302, the edges of the neural region are accurately identified according to the gray scale feature changes of part of the positions in different pooling layers.

[0057] S303, the compression degree of the current dorsal root ganglion is judged according to the size of the change of the intervertebral space, and the actual size of the dorsal root ganglion is restored.

[0058] S304, the dorsal root ganglion is further judged according to the size difference when the dorsal root ganglion is in different positions in the spine.

[0059] S305, the morphological similar interference such as blood vessel network is excluded according to the change of the existence of the contrast agent.

[0060] It should be noted that the above-mentioned embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or may be advantageous.

[0061] Each embodiment in the specification is described in a progressive manner, and the same or similar parts between each embodiment can be referred to each other, and each embodiment mainly describes the difference from other embodiments.

Claims

1. A deep learning-based method for localizing dorsal root ganglia during intervertebral foraminagraphy, characterized in that: The method comprises: Multiple spinal CT images from different scanning angles are input into the neural network model to obtain the nerve identification area of ​​each spinal CT image and the image area of ​​the dorsal root ganglion in the corresponding compression state at different scanning angles; De-overlapping the neural identification region according to grayscale activation differences of multiple pooling layers in the neural network model to determine an actual outer boundary of the neural identification region, wherein the multiple pooling layers have different pooling window sizes; Performing reliability marking on the location of the dorsal root ganglion within the actual outer boundary according to the correspondence between the image area of ​​each spinal CT image and the intervertebral space; The dorsal root ganglion of the user is visualized according to the credibility labeling results of the actual outer boundary of the nerve identification area and the position of the dorsal root ganglion.

2. The method for locating dorsal root ganglion under intervertebral foraminagraphy according to claim 1, characterized in that: The de-overlapping process of the neural identification area according to the grayscale activation differences of multiple pooling layers in the neural network model to determine the actual outer boundary of the neural identification area includes: According to the pooling window size corresponding to each pooling layer, the maximum pooling process is performed on the neural recognition area of ​​each spinal CT image to obtain the activation value of the neural recognition area under different pooling windows; The activation values ​​of the same area on each spinal CT image under each pooling layer are statistically analyzed and normalized to obtain the edge consistency of the statistical area; When the edge conformity of the statistically counted area is greater than a preset conformity threshold, the corresponding statistically counted area is determined to be the actual outer boundary of the neural recognition area.

3. The method for locating dorsal root ganglion under intervertebral foraminagraphy according to claim 2, characterized in that: The activation values ​​of the same area on each spinal CT image under each pooling layer are statistically analyzed and normalized to obtain the edge conformity of the statistical area, including: Get the activation value of the statistical area in each pooling window arranged from small to large; Calculating the activation value difference of the statistical area in adjacent pooling layers, and counting the number of positive activation value differences; The normalized value of the sum of activation values ​​of all pooling layers in the same statistical area is multiplied by the statistical quantity, and normalized based on a preset range to obtain the edge conformity of the corresponding statistical area.

4. The method for locating dorsal root ganglion under intervertebral foraminagraphy according to claim 1, characterized in that: The reliability marking of the location of the dorsal root ganglion within the actual outer boundary according to the correspondence between the image area of ​​each spinal CT image and the intervertebral space includes: According to the correspondence between the image area of ​​each spinal CT image and the intervertebral space, the actual maximum area of ​​the current dorsal root ganglion in the uncompressed state is obtained; Obtaining an area conformity of the current dorsal root ganglion according to a reference number of ganglia adjacent to the current dorsal root ganglion, a theoretical maximum area of ​​the current dorsal root ganglion, and the actual maximum area; According to the area consistency of all dorsal root ganglia, the locations of the dorsal root ganglia within the actual outer boundary are labeled with credibility.

5. The method for locating dorsal root ganglion under intervertebral foraminagraphy according to claim 4, characterized in that: The method of obtaining the actual maximum area of ​​the current dorsal root ganglion in an uncompressed state based on the correspondence between the image area of ​​each spinal CT image and the intervertebral space includes: The image areas of adjacent dorsal root ganglia in each spinal CT image were analyzed for interval distance to obtain the first scanning angle of the superior and inferior vertebrae at the shortest distance in the intervertebral space where the dorsal root ganglia were located. The image areas represented by the same dorsal root ganglion in each spinal CT image are compared to obtain a second scanning angle at which the dorsal root ganglion has the largest image area; Obtaining the area restoration degree of the current dorsal root ganglion at the corresponding scanning angle according to the angle difference between the first scanning angle and the second scanning angle and the shortest distance of the interval distance analysis; According to the maximum image area and the maximum area restoration degree compared with the image areas, the actual maximum area of ​​the current dorsal root ganglion in the non-compressed state is obtained.

6. The method for locating dorsal root ganglion under intervertebral foraminagraphy according to claim 5, characterized in that: The obtaining, based on the angle difference between the first scanning angle and the second scanning angle and the shortest distance resolved by the interval distance, the area restoration degree of the current dorsal root ganglion at the corresponding scanning angle includes: Calculate the minimum extreme value of all the shortest distances of the interval distance analysis to obtain the minimum target distance; Performing negative correlation normalization processing on the corresponding angle difference as a first analysis coefficient; performing negative correlation normalization processing on the minimum target distance as a second analysis coefficient; The area restoration degree of the current dorsal root ganglion at the corresponding scanning angle is obtained according to the product of the first analysis coefficient and the second analysis coefficient.

7. The method for locating dorsal root ganglion under intervertebral foraminagraphy according to claim 4, characterized in that: The obtaining of the area conformity of the current dorsal root ganglion according to the reference number of ganglia adjacent to the current dorsal root ganglion, the theoretical maximum area of ​​the current dorsal root ganglion, and the actual maximum area includes: Obtaining an absolute value of the area difference of the current dorsal root ganglion according to the difference between the theoretical maximum area of ​​the current dorsal root ganglion and the actual maximum area; The area conformity of the current dorsal root ganglion is obtained according to the product of the third reciprocal of the absolute value of the area difference and the reference number of ganglia.

8. The method for locating dorsal root ganglion under intervertebral foraminagraphy according to claim 1, characterized in that: Before visually displaying the dorsal root ganglion of the user, the method further includes: According to the contrast agent's interference with the vascular network in the similar morphology of the dorsal root ganglion on the spine CT image, the reliability of the position of the dorsal root ganglion is updated.

9. The method for locating dorsal root ganglion under intervertebral foraminagraphy according to claim 8, characterized in that: The updating of the credibility of the position of the dorsal root ganglion according to the interference of the contrast agent on the vascular network with the similar morphology of the dorsal root ganglion on the spinal CT image comprises: Obtain the maximum grayscale value and the minimum grayscale value of the current dorsal root ganglion location, the first grayscale mean value of the contrast agent-interfered area in all spine CT images, and the second grayscale mean value of the current dorsal root ganglion in the current spine CT image; Obtaining a degree of error in identifying the current dorsal root ganglion according to the maximum grayscale value, the minimum grayscale value, the first grayscale mean, the second grayscale mean, and an area consistency indicating the reliability of the position of the current dorsal root ganglion; The reliability of the position of the corresponding dorsal root ganglion is determined based on the comparison result of the recognition error degree of each dorsal root ganglion and the preset error degree threshold.

10. The method for locating dorsal root ganglion under intervertebral foraminagraphy according to claim 9, characterized in that: Obtaining the degree of error in identifying the current dorsal root ganglion according to the maximum grayscale value, the minimum grayscale value, the first grayscale mean, the second grayscale mean, and the area consistency representing the credibility of the current dorsal root ganglion position includes: Obtaining a grayscale difference value of the current dorsal root ganglion in each spinal CT image according to an absolute value of a difference between the first grayscale mean and the second grayscale mean; Obtaining the cumulative grayscale difference of the current dorsal root ganglion according to the cumulative summation result of the grayscale difference values ​​of all spinal CT images including the current dorsal root ganglion image; According to the maximum gray value, minimum gray value, cumulative gray difference and area coincidence of each dorsal root ganglion, the degree of identification error of the corresponding dorsal root ganglion is obtained.

Images