An intervertebral foramen under contrast positioning method of dorsal root ganglion based on deep learning

Abstract

The present application relates to the technical field of spine CT image processing, and particularly relates to a method for locating dorsal root ganglion under intervertebral foramen contrast based on deep learning. The present application inputs multiple spine CT images of different scanning angles into a neural network model by a user, so as 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; a de-overlapping processing is implemented based on a gray level activation difference of multiple pooling layers, so as to determine an actual outer boundary of the neural recognition area; according to a corresponding relationship between the image area of each spine CT image and an intervertebral space, a credibility label is marked on a position of the dorsal root ganglion in the actual outer boundary; and according to the credibility label result of the position of the dorsal root ganglion, the dorsal root ganglion of the user is visually displayed. The present application can reduce the technical problem that the peripheral morphologically similar tissues of the dorsal root ganglion affect the recognition accuracy.

Description

Technical Field

[0001] This invention relates to the technical field of spinal CT image processing, and specifically to a deep learning-based method for locating dorsal root ganglia during foraminal lumbar CT. Background Technology

[0002] The dorsal root ganglion (DRG), as the cluster of cell bodies of sensory neurons, is the essential pathway for the transmission of pain, touch, and other signals to the central nervous system. Abnormalities of the DRG (such as displacement or injury) are closely related to diseases such as lumbosacral 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 neuromodulation technologies. Therefore, accurate identification of the DRG is of great significance for neuroscience, clinical medicine, and pain management. However, traditional methods, which often use neural network models to determine the location of the DRG, suffer from insufficient accuracy due to various recognition interferences. Summary of the Invention

[0003] To reduce the impact of similar morphological tissues surrounding the dorsal root ganglion on its identification accuracy, the present invention aims to provide a deep learning-based method for locating the dorsal root ganglion under foraminal lumbar discectomy. The specific technical solution adopted is as follows:

[0004] This invention provides a deep learning-based method for locating dorsal root ganglia during foraminal lumbar ...

[0005] The user inputs multiple spinal CT images from different scanning angles into a neural network model to obtain the neural recognition area of ​​each spinal CT image, as well as the image area of ​​the dorsal root ganglion under the corresponding compression state at different scanning angles.

[0006] The neural recognition region is de-overlapped based on the differences in gray-level activation of multiple pooling layers in the neural network model to determine the actual outer boundary of the neural recognition region. The pooling window sizes of the multiple pooling layers are different.

[0007] Based on the correspondence between the image area of ​​each spinal CT image and the intervertebral space, the credibility of the location of the dorsal root ganglion within the actual outer boundary is marked.

[0008] Based on the credibility annotation results of the actual outer boundary of the neural recognition area and the location of the dorsal root ganglion, the user's dorsal root ganglion is visualized.

[0009] In one optional embodiment, the neural recognition region is de-overlapped based on the gray-level activation differences of multiple pooling layers in the neural network model to determine the actual outer boundary of the neural recognition region, including:

[0010] Based on the pooling window size corresponding to each pooling layer, max pooling is performed on the neural recognition region of each spinal CT image to obtain the activation value of the neural recognition region under different pooling windows.

[0011] The activation values ​​of the same region in each pooling layer on each spinal CT image are statistically analyzed and normalized to obtain the edge conformity of the statistical region.

[0012] When the edge conformity of the statistical region is greater than the preset conformity threshold, the corresponding statistical region is determined as the actual outer boundary of the neural recognition region.

[0013] In one 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 conformity of the statistically analyzed region, including:

[0014] Obtain the activation values ​​of the statistical region in each pooling layer's pooling windows, arranged from smallest to largest;

[0015] Calculate the difference in activation values ​​of the statistical region in adjacent pooling layers, and count the number of positive activation value differences;

[0016] The sum of activation values ​​of all pooling layers in the same statistical region is multiplied by the statistical quantity, and then normalized based on a preset range to obtain the edge conformity of the corresponding statistical region.

[0017] In one optional embodiment, the location of the dorsal root ganglion within the actual outer boundary is labeled with confidence level based on the correspondence between the image area and the intervertebral space of each spinal CT image, including:

[0018] Based on 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 under the uncompressed state is obtained;

[0019] The area conformity of the current dorsal root ganglion is obtained based on 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.

[0020] Based on the area conformity of all dorsal root ganglia, the reliability of the location of the dorsal root ganglia within the actual outer boundary is marked.

[0021] In one optional embodiment, the actual maximum area of ​​the current dorsal root ganglion under uncompressed state is obtained based on the correspondence between the image area of ​​each spinal CT image and the intervertebral space, including:

[0022] The image area of ​​adjacent dorsal root ganglia in each spinal CT image is analyzed by interval distance to obtain the first scanning angle of the upper and lower vertebral bodies of the intervertebral space where the dorsal root ganglia are located at the shortest distance.

[0023] The image area represented by the same dorsal root ganglion in each spinal CT image is compared to obtain the second scanning angle of the dorsal root ganglion at the maximum image area;

[0024] Based on the angle difference between the first scanning angle and the second scanning angle, and the shortest distance analyzed by the interval distance, the area restoration degree of the current dorsal root ganglion at the corresponding scanning angle is obtained;

[0025] Based on the maximum image area and the maximum area restoration degree compared with the image area, the actual maximum area of ​​the current dorsal root ganglion under the uncompressed state is obtained.

[0026] In one optional embodiment, the area reconstruction degree of the current dorsal root ganglion at the corresponding scanning angle is obtained based on the angular difference between the first scanning angle and the second scanning angle, and the shortest distance resolved by the interval distance, including:

[0027] The minimum target distance is obtained by finding the minimum extreme value of all shortest distances in the interval distance analysis.

[0028] The angle difference is negatively correlated and normalized to obtain the first analysis coefficient; the minimum target distance is negatively correlated and normalized to obtain the second analysis coefficient.

[0029] The area of ​​the dorsal root ganglion at the corresponding scanning angle is obtained by multiplying the first analysis coefficient and the second analysis coefficient.

[0030] In one optional embodiment, the area conformity of the current dorsal root ganglion is obtained based on 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, including:

[0031] The absolute value of the area difference of the current dorsal root ganglion is obtained by comparing the theoretical maximum area with the actual maximum area.

[0032] The area conformity of the current dorsal root ganglion is obtained by multiplying the third reciprocal of the absolute value of the area difference with the number of ganglion references.

[0033] In an alternative embodiment, before visualizing the user's dorsal root ganglia, the method further includes:

[0034] The reliability of the location of the dorsal root ganglia is updated based on the interference of contrast agents on the similar morphology of the vascular network presented on spinal CT images.

[0035] In one optional embodiment, the confidence level of the location of the dorsal root ganglia is updated based on the interference of contrast agent on the similar morphological appearance of the vascular network in the dorsal root ganglia on the spinal CT image, including:

[0036] Obtain the maximum and minimum gray values ​​of the current dorsal root ganglion, the first gray mean of the contrast-interference-affected region in all spinal CT images, and the second gray mean of the current dorsal root ganglion in the current spinal CT image.

[0037] The degree of recognition error of the current dorsal root ganglion is obtained based on the maximum gray value, the minimum gray value, the first gray value, the second gray value, and the area conformity, which represents the credibility of the current dorsal root ganglion's location.

[0038] The confidence level of the location of the corresponding dorsal root ganglion is determined by comparing the identification error level of each dorsal root ganglion with the preset error level threshold.

[0039] In one optional embodiment, the degree of recognition error of the current dorsal root ganglion is obtained based on the maximum gray value, the minimum gray value, the first gray average, the second gray average, and the area conformity, which characterizes the credibility of the current dorsal root ganglion's location, including:

[0040] The gray-level difference of the current dorsal root ganglion in each spinal CT image is obtained by calculating the absolute value of the difference between the first gray-level mean and the second gray-level mean.

[0041] The cumulative grayscale difference of the current dorsal root ganglion is obtained by summing the cumulative grayscale differences of the current dorsal root ganglion image in all spinal CT images.

[0042] The degree of recognition error for each dorsal root ganglion is obtained based on the maximum gray value, minimum gray value, cumulative gray difference, and area conformity of each dorsal root ganglion.

[0043] The present invention has the following beneficial effects:

[0044] This invention, when performing dorsal root ganglion localization under deep learning-based foraminal lumbar discectomy, inputs multiple spinal CT images of the user at different scanning angles into a neural network model to obtain the nerve recognition area of ​​each spinal CT image and the image area of ​​the dorsal root ganglion under the corresponding compression state at different scanning angles. Since the pooling window sizes of multiple pooling layers in the neural network model differ, the nerve recognition area can be de-overlapped based on the differences in gray-level activation of multiple pooling layers to determine the actual outer boundary of the nerve recognition area. Furthermore, based on the correspondence between the image area of ​​each spinal CT image and the intervertebral space, the location of the dorsal root ganglion within the actual outer boundary is labeled with confidence. Based on the actual outer boundary of the nerve recognition area and the confidence labeling results of the location of the dorsal root ganglion, the user's dorsal root ganglion is visualized. This technical solution, when performing dorsal root ganglion localization on spinal CT images, considers the morphological similarity of blood vessels and the deformation of the dorsal root ganglion due to compression, and employs corresponding technical means to eliminate interference, reducing the technical problem of the influence of morphologically similar tissues surrounding the dorsal root ganglion on its recognition accuracy. Attached Figure Description

[0045] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 A flowchart illustrating a deep learning-based method for locating dorsal root ganglia during foraminal lumbar discectomy, as provided in one embodiment of the present invention.

[0047] Figure 2 A schematic diagram of the dorsal root ganglion and surrounding tissues provided in an embodiment of the present invention;

[0048] Figure 3 This is an overall logic diagram of the dorsal root ganglion localization method provided in one embodiment of the present invention.

[0049] Figure labeling: 1-vertebra, 2-spinal nerve, 3-intervertebral space. Detailed Implementation

[0050] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a deep learning-based method for locating dorsal root ganglia under foraminal lumbar discectomy according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable manner.

[0051] 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 this invention pertains.

[0052] The following description, in conjunction with the accompanying drawings, details a specific scheme for a deep learning-based method for locating dorsal root ganglia during foraminal lumbar discectomy.

[0053] Please see Figure 1 This document illustrates a flowchart of a deep learning-based method for locating dorsal root ganglia during foraminal lumbar discectomy, according to an embodiment of the present invention. This location method can be run on a dorsal root ganglion angiography recognition terminal, which can be a computer or server capable of running the location method; no specific limitations are imposed here. The location method includes:

[0054] S11. Input multiple spinal CT images taken by the user at different scanning angles into the neural network model to obtain the neural recognition area of ​​each spinal CT image, as well as the image area of ​​the dorsal root ganglion under the corresponding compression state at different scanning angles.

[0055] Specifically, the user can be a patient requiring spinal diagnosis and treatment. The angle between the patient's back plane and the electric scanning bed is the scanning angle. Multiple scanning angles allow for a more accurate assessment of spinal tissue lesions, such as whether the dorsal root ganglia are abnormally positioned or damaged. Accurate identification of the dorsal root ganglia is not only crucial for basic research but also a key prerequisite for the development of precision medicine and neuromodulation technologies. The neural network model can be a convolutional neural network model trained on a training set of spinal CT images. The model is trained by labeling the locations of the intervertebral foramina, the upper and lower edges of the pedicles, and the dorsal root ganglia in the spinal CT images in the training set until training is complete. After inputting multiple spinal CT images from the user into the model, the nerve recognition area for each spinal CT image can be obtained. The nerve recognition area is the outer boundary of all nerves on the spinal CT image. Since different angles cause varying degrees of compression to the spinal nerves, the image area of ​​each spinal CT image can characterize the changes in the dorsal root ganglia under corresponding compression states at different scanning angles.

[0056] Please see Figure 2 , Figure 2 This is a schematic diagram of the dorsal root ganglion and surrounding tissues. When identifying nerve recognition areas based on multiple spinal CT images, multi-angle images from the spinal CT scan can be used as input data. The images must include the vertebral body, pedicles, and nerve root structures, such as... Figure 2 As shown, intervertebral spaces exist between adjacent vertebrae, and spinal nerves attach to the vertebrae via connections with blood vessels and other tissues. Posterior root ganglia are present on the spinal nerves. When identifying neural recognition regions based on a neural network model, Faster R-CNN can be used to locate bony structures such as the pedicles and intervertebral foramina, and a Region Proposal Network (RPN) can be used to generate candidate boxes. An integrated U-Net structure is used to segment neural tissues (such as DRGs), preserving multi-scale features through skip connections. By combining the deep feature extraction capabilities of ResNet-50 (solving the gradient vanishing problem) with the segmentation accuracy of U-Net, end-to-end localization is achieved, resulting in a neural network model capable of recognizing multiple tissue locations.

[0057] S12. Based on the differences in gray-level activation of multiple pooling layers in the neural network model, the neural recognition region is de-overlapped to determine the actual outer boundary of the neural recognition region. The pooling window sizes of the multiple pooling layers are different.

[0058] Specifically, due to scanning angle issues, the edges of some nerves displayed in the nerve recognition area of ​​a spinal CT image may overlap with those of another nerve recognition area, causing errors in nerve recognition and thus affecting the identification of the dorsal root ganglion location. During feature recognition in a neural network model, different pooling layers will result in varying activation values ​​at the nerve edges in the spinal CT image due to differences in pooling window size. Therefore, by observing the degree of change in activation values ​​at corresponding locations under different pooling layers, the location of hidden nerve edges can be determined, thus identifying the actual outer boundary of the nerve recognition area. Please continue reading... Figure 2 As shown in the figure, region A has a relatively complex tissue structure, and the actual outer boundary of the neural recognition area can be determined based on the above method.

[0059] For example, step S12 includes sub-steps S12-1 to S12-3, which are described in detail below:

[0060] S12-1. Based on the pooling window size corresponding to each pooling layer, perform max pooling on the neural recognition region of each spinal CT image to obtain the activation value of the neural recognition region under different pooling windows.

[0061] Specifically, the pooling window size for different pooling layers can be configured as 1×1, 2×1, 2×2, 2×3, 2×4, 2×5, 2×6, 2×7, 2×8, 2×9, 2×1 ... 3 Pooling operations, such as 4×4 and others, are performed on the neural recognition regions in spinal CT images, i.e., max pooling. When the number of pixels shared by pooling windows in different pooling layers is maximized, these two pooling windows correspond to the same CT image region. For the region p containing pooling window i in a 1×1 pooling layer, the activation values ​​of this region and its corresponding regions in other pooling layers k are obtained. .

[0062] S12-2. Statistical analysis and normalization are performed on the activation values ​​of the same region in each pooling layer on each spinal CT image to obtain the edge conformity of the statistically analyzed region. It can be understood that as the pooling window becomes larger, the edge region of a nerve may be divided into more ordinary regions, thus masking edge features more. Therefore, for the same nerve edge region, the activation values ​​in pooling layers with increasingly larger pooling windows will gradually decrease. Based on this characteristic, statistical analysis and normalization of the activation values ​​are performed to calculate the edge conformity of the statistically analyzed region. The edge conformity characterizes whether the region with the statistically analyzed activation values ​​represents the actual outer boundary of the neural recognition area.

[0063] For example, sub-step S12-2 is implemented in the following ways:

[0064] The first step is to obtain the activation values ​​of the statistical region within each pooling layer's pooling window, arranged in ascending order. The values ​​are sorted in ascending order based on the pooling window size of each pooling layer; for example, a 1×1 pooling window is ranked first, and a 2×1 pooling window is ranked second. The second position is assigned, and so on, until the arrangement is completed, and the activation values ​​of the neural recognition region under different pooling windows are obtained.

[0065] The second step is to calculate the difference in activation values ​​of the statistical region in adjacent pooling layers, and to count the number of positive differences. This allows us to calculate the magnitude of the difference in activation values ​​for the same region p in adjacent pooling layers "k-1, k". (Positive and negative values ​​are distinguished) and recorded as the difference in activation values. Statistics. Number of This refers to the statistical number of activation value differences that are positive.

[0066] The third step involves multiplying the sum of activation values ​​from all pooling layers within the same statistical region by the statistical count, and then normalizing the result based on a preset range to obtain the edge conformity of the corresponding statistical region. This calculates the sum of activation values ​​for the same region p across multiple pooling layers. When the total activation value The larger the value, and the greater the statistical quantity. The larger the value, the more the region p within the neural recognition area conforms to the characteristics of the neural masking boundary, and the more likely it is to be the outer boundary of the neural network. This allows us to obtain the statistical results for the region p containing pooling window i in a pooling layer with a pooling window size of 1×1. ,in, Let represent the number of all pooling layers, and norm denotes linear normalization. The statistical results are analyzed using the max-min normalization method. Normalization is performed to obtain the edge conformity. Its value range is configured as [0, 1].

[0067] S12-3. When the edge conformity of the statistical region is greater than the preset conformity threshold, the corresponding statistical region is determined as the actual outer boundary of the neural recognition area. The conformity threshold can be configured to 0.7, or it can be configured to other values ​​based on actual needs. When the boundary is not equal to or less than the boundary threshold, region p is determined as the actual outer boundary of the neural recognition area; otherwise, if the boundary conformity is less than or equal to the conformity threshold, the corresponding statistical region is determined as the non-outer boundary of the neural recognition area. The actual outer boundary of the neural recognition area is redefined using spinal CT images, and the dorsal root ganglion region is distinguished using a neural network model. The regions to which the dorsal root ganglia belong in the spinal CT images are then marked.

[0068] S13. Based on the correspondence between the image area of ​​each spinal CT image and the intervertebral space, the credibility of the location of the dorsal root ganglion within the actual outer boundary is marked.

[0069] Specifically, various factors such as spinal degeneration, trauma, disease, or biomechanical imbalance can lead to narrowing of the intervertebral space, thereby compressing and deforming the dorsal root ganglion. This deformation may result in misjudgment or omission of the dorsal root ganglion at a single location; therefore, the actual size of the dorsal root ganglion should be reconstructed based on the degree of compression. This step assesses the current degree of compression of the dorsal root ganglion based on changes in the intervertebral space, reconstructs the actual size of the dorsal root ganglion, and achieves reliable labeling of its location.

[0070] For example, step S13 includes sub-steps S13-1 to S13-3, which are described in detail below:

[0071] S13-1. Based on the correspondence between the image area and the intervertebral space of each spinal CT image, obtain the actual maximum area of ​​the current dorsal root ganglion under uncompressed conditions. Multiple dorsal root ganglia may exist on each spinal CT image; the current dorsal root ganglion is denoted as dorsal root ganglion g. The degree of compression received by the dorsal root ganglion varies at different scanning angles. The corresponding intervertebral space can be derived based on the change in image area. Analyzing the correspondence between the two yields the actual maximum area of ​​the current dorsal root ganglion g under uncompressed conditions.

[0072] Specifically, sub-step S13-1 includes:

[0073] The first step involves analyzing the image area of ​​adjacent dorsal root ganglia in each spinal CT image to obtain the first scanning angle of the superior and inferior vertebrae of the intervertebral space containing the dorsal root ganglion at the shortest distance. A neural network model is then used to identify vertebrae at different locations in the spinal CT images to obtain the shortest distance in the spinal CT image m of the superior and inferior vertebrae of the dorsal root ganglion g at different scanning angles. shortest distance The corresponding scanning angle is the first scanning angle.

[0074] The second step involves comparing the image areas represented by the same dorsal root ganglion in each spinal CT image to obtain the second scanning angle at which the dorsal root ganglion has the largest image area. By comparison, the maximum image area of ​​the dorsal root ganglion g in spinal CT image m at different scanning angles can be obtained, and this maximum image area is denoted as... , to maximize image area The angle between the CT scanner and the patient's back plane during a CT scan is recorded as the second scanning angle.

[0075] The third step involves obtaining the area restoration degree of the current dorsal root ganglion at the corresponding scanning angle based on the angle difference between the first and second scanning angles and the shortest distance resolved by the interval distance. When obtaining the maximum image area of ​​the dorsal root ganglion g, the smaller the difference between the patient's scanning angle and the patient's angle when obtaining the shortest distance, the greater the degree of compression of the current dorsal root ganglion g, and the greater the difference between the currently obtained maximum image area and the maximum area of ​​the dorsal root ganglion g when it is normal. Conversely, the smaller the shortest distance and the smaller the angle difference, the smaller the maximum image area of ​​the dorsal root ganglion g relative to the area of ​​the dorsal root ganglion g when it is not compressed. Based on the above relationship, a correspondence between the angle difference, the shortest distance, and the area restoration degree can be constructed, thereby obtaining the area restoration degree of each dorsal root ganglion at the corresponding scanning angle.

[0076] For example, the minimum extreme value is obtained by analyzing all the shortest distances in the interval distance analysis to obtain the minimum target distance; the angle difference is negatively correlated and normalized as the first analysis coefficient; the minimum target distance is negatively correlated and normalized as the second analysis coefficient; the area restoration degree of the current dorsal root ganglion at the corresponding scanning angle is obtained by multiplying the first analysis coefficient and the second analysis coefficient.

[0077] You can use either exp(-x) or norm(-x) for negative correlation normalization; there are no restrictions on which method you use.

[0078] All shortest distances After finding the minimum and extreme values, the minimum target distance can be obtained. This allows us to obtain the degree of area reduction of the dorsal root ganglion g. ,in, The difference in angles. The first analytical coefficient, To minimize the target distance, This is the second analytical coefficient. The degree of area reduction can also be assessed using the max-min normalization method. After normalization, we get Its range is (0, 1).

[0079] The fourth step involves comparing the maximum image area and the degree of area restoration to obtain the actual maximum area of ​​the dorsal root ganglion under uncompressed conditions. When the degree of restoration... The larger the dorsal root ganglion g is, the larger its maximum image area is when it is not compressed. Therefore, the actual maximum area of ​​the dorsal root ganglion g when it is not compressed can be obtained, and this actual maximum area is denoted as [missing information]. The following relation exists: , The degree of area restoration after normalization; denoted as the maximum image area of ​​the dorsal root ganglion g, corresponding to the spinal CT image m.

[0080] S13-2. Based on 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, obtain the area conformity of the current dorsal root ganglion. The reference number of ganglia is the number of dorsal root ganglia adjacent to the current dorsal root ganglion. For example, if there are 3 dorsal root ganglia above and below the current dorsal root ganglion g, the reference number of ganglia is 6. The theoretical maximum area of ​​the current dorsal root ganglion can be obtained based on an experimental database. Although there are individual differences in dorsal root ganglia, the area differences are not significant. Therefore, a reference calculation can be performed based on the theoretical maximum area of ​​the current dorsal root ganglion to obtain the area conformity of the current dorsal root ganglion. The area conformity is determined based on the actual maximum area to determine whether it has a reliable reference value. The area conformity can be denoted as... .

[0081] It is understandable that the size of the dorsal root ganglion varies significantly at different locations in the spine (i.e., different vertebral levels). These differences mainly stem from variations in vertebral segmental position. For example, in the lumbar spine, the size of the dorsal root ganglion increases progressively from L1 to L5. In spinal CT images, the presence of congestion can lead to misidentification of the dorsal root ganglion (DRG). In such cases, the size of the DRG at the corresponding vertebral segment can be used for differentiation. Therefore, the area accuracy of the current dorsal root ganglion can be determined based on the following method.

[0082] The first step is to obtain the absolute value of the area difference of the current dorsal root ganglion based on the difference between the theoretical maximum area and the actual maximum area. This will allow us to obtain the average maximum area of ​​the dorsal root ganglion g in the experimental data. Average maximum area size The maximum area of ​​the three dorsal root ganglia above and below the detected dorsal root ganglion g will be determined as the theoretical maximum area. The actual maximum area will be denoted as... The absolute value of the area difference is .

[0083] The second step is to obtain the area conformity of the current dorsal root ganglion by multiplying the third reciprocal of the absolute value of the area difference by the number of ganglion references. The number of ganglion references is denoted as... When the image area of ​​the dorsal root ganglion g is relatively consistent with the actual distribution of different locations in the spine, the number of reference dorsal root ganglia is considered. The larger the area difference, the greater the absolute value of the area difference. The smaller the area, the more accurate the neural network model's identification of the current dorsal root ganglion g. From this, the area conformity of the dorsal root ganglion g can be obtained. .

[0084] S13-3. Based on the area conformity of all dorsal root ganglia, the location of the dorsal root ganglia within the actual outer boundary is labeled with credibility. The area conformity characterizes whether the image area of ​​the corresponding dorsal root ganglion is credible, and thus credibility can be labeled accordingly. Credibility can be labeled with different levels or quantified as a percentage for easy reference by users or medical personnel.

[0085] In practical applications, because blood vessels have a similar distribution pattern to nerves, thrombi and other protrusions on them may be misidentified as dorsal root ganglia. In contrast to blood vessels, contrast agents diffuse spatially homogeneously within nerve tissue and do not exhibit clearing or washing-off over time. To eliminate interference from blood vessels, the changes in the presence of contrast agents in different regions can be analyzed. Based on this, the localization method of this invention further includes: updating the reliability of the location of the dorsal root ganglia based on the interference of contrast agents on the similar morphological appearance of the vascular network on spinal CT images of the dorsal root ganglia. Specifically, this includes:

[0086] S13-4. Obtain the maximum and minimum gray values ​​of the current dorsal root ganglion location, the first gray-mean value of the contrast-affected areas in all spinal CT images, and the second gray-mean value of the current dorsal root ganglion in the current spinal CT image. The image grayscale range after processing is 0-255. Based on the image grayscale of the current dorsal root ganglion location, the maximum and minimum gray values ​​can be resolved. Some areas are affected by contrast agent interference, such as areas with blood vessels. The image grayscale of these areas is resolved, and the average grayscale value is calculated to obtain the first gray-mean value. The first gray-mean value represents the average grayscale value of each location affected by contrast agent in the spinal CT image. The second gray-mean value is calculated based on the image grayscale of the region to which the current dorsal root ganglion belongs in the current spinal CT image.

[0087] S13-5. Based on the maximum gray value, minimum gray value, first gray mean, second gray mean, and area conformity (representing the credibility of the current dorsal root ganglion's location), the recognition error degree of the current dorsal root ganglion is obtained. A data processing model can be constructed based on the correlation between each value and the recognition error degree. By inputting all data into this data processing model, the recognition error degree of the current dorsal root ganglion g can be obtained.

[0088] Furthermore, the degree of identification error can be calculated based on the following methods, specifically including:

[0089] The first step is to obtain the gray-level difference of the current dorsal root ganglion in each spinal CT image by calculating the absolute value of the difference between the first gray-level mean and the second gray-level mean. The first gray-level mean is denoted as... The second grayscale mean is denoted as The grayscale difference is .

[0090] The second step involves summing the grayscale differences of the current dorsal root ganglion image across all spinal CT images to obtain the cumulative grayscale difference. The larger the difference between the maximum and minimum grayscale values ​​at position g of the dorsal root ganglion during the spinal CT image acquisition phase, and the greater the grayscale difference across multiple spinal CT images... , cumulative summation result ( The larger the number of spinal CT images containing the dorsal root ganglion (g position), the greater the area conformity. The smaller the size, the faster the contrast agent at the dorsal root ganglion g position identified by the neural network model dissipates, and the greater the difference in grayscale performance between it and other areas. The size is less consistent with the actual size and less consistent with the display of the dorsal root ganglion g, making the identification of this position more problematic.

[0091] The third step involves determining the recognition error level of each dorsal root ganglion based on its maximum gray value, minimum gray value, cumulative gray value difference, and area conformity. The maximum gray value is denoted as... The minimum grayscale value is denoted as Based on the above analysis, the degree of recognition error of the dorsal root ganglion g by the neural network model can be obtained, and the degree of recognition error is denoted as . , The degree of identification error can be assessed using the min-max normalization method. After normalization, we get Its range is (0, 1).

[0092] S13-6. Based on the comparison between the recognition error rate of each dorsal root ganglion and the preset error rate threshold, determine the reliability of the location of the corresponding dorsal root ganglion. The error rate threshold can be set based on the actual situation, for example, set to 0.8. If the dorsal root ganglion g identified by the neural network model does not match the actual location, it indicates a problem with the model's identification. When the error rate in identifying the dorsal root ganglion exceeds a preset error rate threshold, the reliability of the corresponding dorsal root ganglion's location is considered low; when the error rate is less than or equal to the preset error rate threshold, the reliability of the corresponding dorsal root ganglion's location is considered high.

[0093] S14. Based on the credibility labeling results of the actual outer boundary of the neural recognition area and the location of the dorsal root ganglion, visualize the user's dorsal root ganglion.

[0094] Specifically, the reliability of each dorsal root ganglion can be determined based on the above method, and image annotation can be implemented. The dorsal root ganglions can then be visualized for the user through a display terminal. Furthermore, the locations of the dorsal root ganglia identified by the neural network model can be judged using the above method, and the resulting error levels can be matched with the corresponding locations and stored in a database. SQL (Structured Query Language) can then be used to mark different dorsal root ganglion locations and locations of incorrect judgments in the image, and then visualized.

[0095] The following embodiments of the present invention will describe the localization process of the dorsal root ganglion. Please refer to... Figure 3 , Figure 3 A logical diagram of the dorsal root ganglion localization method. Specifically, it includes:

[0096] S301. The location of the intervertebral foramen, the upper and lower edges of the pedicle, and the dorsal root ganglion in the patient's spine is obtained through a deep learning model.

[0097] S302. Accurately identify the edge of the neural region based on the grayscale feature changes at certain locations in different pooling layers.

[0098] S303. Based on the changes in the intervertebral space, determine the degree of compression of the dorsal root ganglion and restore the actual size of the dorsal root ganglion.

[0099] S304. Further assessment of the dorsal root ganglion is made by observing the differences in size of the dorsal root ganglion at different positions within the spinal column.

[0100] S305. Based on the presence and changes of contrast agents, morphological similarity interferences such as vascular networks are eliminated.

[0101] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0102] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A deep learning-based method for locating dorsal root ganglia during foraminal lumbar discectomy, characterized in that, The method includes: The user inputs multiple spinal CT images from different scanning angles into a neural network model to obtain the neural recognition area of ​​each spinal CT image, as well as the image area of ​​the dorsal root ganglion under the corresponding compression state at different scanning angles. The neural recognition region is de-overlapped based on the gray-level activation differences of multiple pooling layers in the neural network model to determine the actual outer boundary of the neural recognition region, wherein the pooling window sizes of the multiple pooling layers are different; Based on the correspondence between the image area of ​​each spinal CT image and the intervertebral space, the location of the dorsal root ganglion within the actual outer boundary is marked with confidence. Based on the actual outer boundary of the neural recognition area and the confidence labeling results of the location of the dorsal root ganglion, the user's dorsal root ganglion is visualized. The method for determining the actual outer boundary of the neural recognition region includes: Based on the pooling window size corresponding to each pooling layer, max pooling is performed on the neural recognition region of each spinal CT image to obtain the activation value of the neural recognition region under different pooling windows. The activation values ​​of the same region in each pooling layer on each spinal CT image are statistically analyzed and normalized to obtain the edge conformity of the statistical region. When the edge conformity of the statistical region is greater than the preset conformity threshold, the corresponding statistical region is determined as the actual outer boundary of the neural recognition region; The methods for confidence labeling of the location of the dorsal root ganglion within the actual outer boundary include: Based on 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 under the uncompressed state is obtained; The area conformity of the current dorsal root ganglion is obtained based on 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. Based on the area conformity of all dorsal root ganglia, the reliability of the location of the dorsal root ganglia within the actual outer boundary is marked.

2. The method for locating the dorsal root ganglion under foraminal lumbar vertebral angiography according to claim 1, characterized in that, The step of statistically analyzing and normalizing the activation values ​​of the same region in each pooling layer on each spinal CT image to obtain the edge conformity of the statistically analyzed region includes: Obtain the activation values ​​of the statistical region in each pooling layer's pooling windows, arranged from smallest to largest; Calculate the activation value difference of the statistical region in adjacent pooling layers, and count the number of positive activation value differences; The normalized value of the sum of activation values ​​of all pooling layers in the same statistical region is multiplied by the statistical quantity, and then normalized based on a preset range to obtain the edge conformity of the corresponding statistical region.

3. The method for locating the dorsal root ganglion under foraminal lumbar vertebral angiography according to claim 1, characterized in that, The method of obtaining the actual maximum area of ​​the dorsal root ganglion under uncompressed conditions based on the correspondence between the image area of ​​each spinal CT image and the intervertebral space includes: The image area of ​​adjacent dorsal root ganglia in each spinal CT image is analyzed by interval distance to obtain the first scanning angle of the upper and lower vertebral bodies of the intervertebral space where the dorsal root ganglia are located at the shortest distance. The image area represented by the same dorsal root ganglion in each spinal CT image is compared to obtain the second scanning angle of the dorsal root ganglion at the maximum image area; Based on the angle difference between the first scanning angle and the second scanning angle, and the shortest distance analyzed by the interval distance, the area restoration degree of the current dorsal root ganglion at the corresponding scanning angle is obtained; Based on the maximum image area and the maximum area restoration degree compared with the image area, the actual maximum area of ​​the current dorsal root ganglion under the uncompressed state is obtained.

4. The method for locating the dorsal root ganglion under foraminal lumbar vertebral angiography according to claim 3, characterized in that, The step of obtaining the area reconstruction degree of the current dorsal root ganglion at the corresponding scanning angle based on the angle difference between the first scanning angle and the second scanning angle and the shortest distance analyzed by the interval distance includes: The minimum target distance is obtained by finding the minimum extreme value of all shortest distances in the interval distance analysis. The angle difference is negatively correlated and normalized to obtain the first analysis coefficient; the minimum target distance is negatively correlated and normalized to obtain the second analysis coefficient. The area restoration degree of the current dorsal root ganglion at the corresponding scanning angle is obtained by multiplying the first analysis coefficient and the second analysis coefficient.

5. The method for locating the dorsal root ganglion under foraminal lumbar vertebral angiography according to claim 1, characterized in that, The step of obtaining the area conformity of the current dorsal root ganglion based on 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: The absolute value of the area difference of the current dorsal root ganglion is obtained based on the difference between 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 by multiplying the third reciprocal of the absolute value of the area difference with the number of ganglion references.

6. The method for locating the dorsal root ganglion under foraminal lumbar vertebral angiography according to claim 1, characterized in that, Before visualizing the user's dorsal root ganglia, the method further includes: The reliability of the location of the dorsal root ganglia is updated based on the interference of contrast agent on the similar morphology of the vascular network presented on the spinal CT image.

7. The method for locating the dorsal root ganglion under foraminal lumbar vertebral angiography according to claim 6, characterized in that, The update of the reliability of the location of the dorsal root ganglia based on the interference of contrast agent on the similar morphology of the vascular network presented by the dorsal root ganglia on the spinal CT image includes: Obtain the maximum and minimum gray values ​​of the current dorsal root ganglion, the first gray mean of the contrast-interference-affected region in all spinal CT images, and the second gray mean of the current dorsal root ganglion in the current spinal CT image. The degree of recognition error of the current dorsal root ganglion is obtained based on the maximum gray value, the minimum gray value, the first gray average value, the second gray average value, and the area conformity that characterizes the credibility of the current dorsal root ganglion's location. The confidence level of the location of the corresponding dorsal root ganglion is determined by comparing the identification error level of each dorsal root ganglion with the preset error level threshold.

8. The method for locating the dorsal root ganglion under foraminal lumbar vertebral angiography according to claim 7, characterized in that, The step of obtaining the recognition error degree of the current dorsal root ganglion based on the maximum gray value, the minimum gray value, the first gray average value, the second gray average value, and the area conformity degree representing the credibility of the current dorsal root ganglion's location includes: The grayscale difference of the current dorsal root ganglion in each spinal CT image is obtained by calculating the absolute value of the difference between the first grayscale mean and the second grayscale mean. The cumulative grayscale difference of the current dorsal root ganglion is obtained by summing the cumulative grayscale differences of the current dorsal root ganglion image in all spinal CT images. The degree of recognition error for each dorsal root ganglion is obtained based on the maximum gray value, minimum gray value, cumulative gray difference, and area conformity of each dorsal root ganglion.

Images