Computer-aided analysis system for spinal infection MRI image
Through a computer-aided analysis system, the LSTM model and DBSCAN clustering algorithm are used to screen the ROI area, build an evolution model, and correct the grayscale value of the spinal MRI image, solving the problem of pixel grayscale value deviation and improving the accuracy and efficiency of spinal infection diagnosis.
Patent Information
- Application Number
- CN202510760962.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-09
- Publication Date
- 2025-09-16
AI Technical Summary
In the prior art, there is a deviation between the grayscale value of a pixel in a spinal MRI image and the true value, which affects the accuracy of the diagnosis result.
A computer-aided analysis system was used to screen the ROI area through the LSTM model and DBSCAN clustering algorithm, and an evolution model was constructed. The current MRI image was corrected in combination with the grayscale value of the predicted scan image to obtain the target area image.
It improves the accuracy and efficiency of spinal infection diagnosis and provides a more reliable basis for diagnosis.
Smart Images
Figure CN120655607A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and in particular to a computer-aided analysis system for spinal infection MRI images. Background Art
[0002] Spinal infection refers to a series of infectious diseases caused by specific pathogenic microorganisms that can affect the vertebral body, intervertebral disc, spinal canal, paravertebral tissue, etc. Common ones include Brucella spondylitis, suppurative spondylitis, tuberculous spondylitis, etc. Currently, the differential diagnosis of spinal infection mainly relies on clinical manifestations, laboratory tests and imaging analysis. The specificity of clinical manifestations and laboratory tests is not high and often relies on the doctor's clinical experience. Imaging is currently the simplest and most commonly used method to distinguish different spinal infections. Compared with X-rays and CT images, magnetic resonance imaging (MRI) has the imaging characteristics of multiple sequences, multiple directions, no radiation, and strong repeatability. It is currently the most widely used differential diagnosis method in clinical practice.
[0003] MRI examinations rely primarily on the subjective experience of radiologists over many years. Different tissue contrast images are generated based on equipment from different manufacturers. Tissue abnormalities are observed visually, and lesion structure and morphological characteristics are used for diagnosis and identification. However, there may be a certain deviation between the grayscale distribution of pixels in the spinal MRI images captured by the equipment and the true value. This deviation directly affects subsequent analysis and judgment results. Therefore, how to accurately correct the grayscale values of pixels in the captured images to make them closer to the actual situation is an urgent problem that needs to be solved. Summary of the Invention
[0004] In order to solve the problem that there is a certain deviation between the grayscale value of the pixel points in the spinal MRI images collected by the existing method and the true value, the purpose of the present invention is to provide a computer-aided analysis system for spinal infection MRI images. The technical solution adopted is as follows:
[0005] The present invention provides a computer-aided analysis system for spinal infection MRI images, comprising a memory and a processor, wherein the processor executes a computer program stored in the memory to implement the following steps:
[0006] Obtaining historical MRI images of the spine of a reference person, historical MRI images of the spine of the person to be tested, and current MRI images;
[0007] Obtaining a predicted scan image of each person and cell state values corresponding to pixels in the reference image based on a reference image of the spine of each person, the reference image comprising a historical MRI image of the spine of the reference person and a historical MRI image of the spine of the person to be tested; and screening a ROI region in the reference image based on changes in the cell state values corresponding to the same person;
[0008] All reference images are divided into multiple categories based on the difference characteristics of the position distribution and shape distribution of the ROI regions in the reference images. The corresponding evolution model is constructed based on the position distribution of the ROI regions in each category of reference images.
[0009] The target area image of the person to be tested is obtained by integrating the position of the previous historical MRI image of the current MRI image of the spine of the person to be tested in the evolution model and the grayscale value of the pixel point in the predicted scan image.
[0010] Preferably, the predicted scan image includes:
[0011] For any person: all reference images of the said person are input into the LSTM model to obtain the predicted scan image of the said person.
[0012] Preferably, the cell state value corresponding to the pixel point in the reference image includes: obtaining the cell state value of each pixel point in each reference image through the LSTM model.
[0013] Preferably, screening the ROI region in the reference image according to the change in the cell state value corresponding to the same person includes:
[0014] For any person:
[0015] For any pixel point in the candidate image: obtaining a forward difference sequence of the cell state value of the any pixel point; taking the normalized average value of the absolute values of the elements in the forward difference sequence as the first probability value that the any pixel point belongs to the ROI area;
[0016] Based on the first likelihood values of all pixels in the candidate image belonging to the ROI region, clustering all pixels in the candidate image to obtain clusters;
[0017] Screening the ROI region in the candidate image according to the distribution of the first likelihood values of the pixels in each cluster;
[0018] The candidate image is any other reference image of the person except the first reference image.
[0019] Preferably, screening the ROI region in the candidate image according to the distribution of the first likelihood values of the pixels in each cluster includes:
[0020] A cluster in which the average value of the first likelihood values of all pixels in the cluster is greater than the average value of the first likelihood values of all pixels in the candidate image is determined as the ROI region in the candidate image.
[0021] Preferably, all reference images are divided into multiple categories according to the difference characteristics of the position distribution and the difference characteristics of the shape distribution of the ROI regions in the pairwise reference images, including:
[0022] For any reference image, each ROI region in the reference image and the ROI region in each reference image other than the reference image having the smallest Euclidean distance with each ROI region in the reference image form a matching region pair;
[0023] For any ROI region, determining a variation factor of the ROI region according to the area of the ROI region; determining a morphological coefficient of the ROI region according to the shape distribution of the ROI region; the variation factor and the morphological coefficient constitute a feature group of the ROI region;
[0024] For any two reference images: using the Euclidean distance between the feature groups of two ROI regions in the matching region pair of the any two reference images as the feature distance of the corresponding matching region pair; obtaining a second likelihood value that the any two reference images belong to the same class based on the position difference and feature distance between the two ROI regions in all matching region pairs of the any two reference images, wherein the position difference and the feature distance are both negatively correlated with the second likelihood value;
[0025] All reference images are clustered based on the second likelihood values of the pairwise reference images, and all reference images in the same cluster are regarded as one class.
[0026] Preferably, determining the morphological coefficient of any ROI region according to the shape distribution of any ROI region includes:
[0027] Extract the binary mask of any ROI area and obtain the lesion contour point set through the findContours function of OpenCV;
[0028] Uniformly sample the contour points in the contour point set to construct a shape context descriptor; based on the shape context descriptor, obtain the first morphological vector, the second morphological vector, and the third morphological vector of any ROI region:
[0029] The morphological coefficient of any ROI region is obtained by integrating the first morphological vector, the second morphological vector and the third morphological vector.
[0030] Preferably, the step of constructing a corresponding evolution model based on the position distribution of the ROI region in each type of reference image includes:
[0031] For any type of reference image:
[0032] Calculating the total area of all ROI regions in each reference image of any one type of reference images respectively; sorting all reference images in any one type of reference images in ascending order of total area to obtain an image sequence corresponding to any one type of reference images;
[0033] Mapping all reference images in the image sequence into a three-dimensional rectangular coordinate system, wherein the X axis of the three-dimensional rectangular coordinate system represents the order of the corresponding reference images in the image sequence, and all reference images in the image sequence are mapped in the YOZ plane of the three-dimensional rectangular coordinate system;
[0034] In a three-dimensional rectangular coordinate system, the SIFT algorithm is used to match the pixel points of any ROI area in each reference image in the three-dimensional rectangular coordinate system with the pixel points at the corresponding position in the next adjacent image to obtain multiple matching point sequences;
[0035] The least squares method is used to fit all the pixels in each matching point sequence into straight lines, and the fitting lines corresponding to all the edge points of the same region of interest are spatially fitted into quadratic surfaces; the closed area enclosed by all quadratic surfaces is the evolution model of this type of reference image.
[0036] Preferably, the step of synthesizing the positions of the previous historical MRI images of the current MRI image of the spine of the person to be tested in the evolution model and the grayscale values of the pixels in the predicted scan image to obtain the target area image of the person to be tested includes:
[0037] Mapping the current MRI image to the position of the previous historical MRI image in the evolution model where the previous historical MRI image is located, and taking the region in the current MRI image mapping result that is at the same position as the region of the previous historical MRI image in the evolution model as the first region in the current MRI image;
[0038] The grayscale values of the pixels in the first area of the current MRI image are corrected using the grayscale values of the pixels at positions corresponding to the first area in the predicted scan image of the person to be tested, to obtain the target area image.
[0039] Preferably, the step of correcting the grayscale values of the pixels in the first area of the current MRI image using the grayscale values of the pixels at positions corresponding to the first area in the predicted scan image of the person to be tested to obtain the target area image includes:
[0040] For any pixel point in the first area: calculating a first difference between the grayscale value of the pixel point and the grayscale value of the pixel point in the predicted scan image of the person to be tested, which is located at the same position as the pixel point; calculating the sum of a constant 1 and a normalized value of the first difference; and multiplying the grayscale value of the pixel point in the predicted scan image of the person to be tested, which is located at the same position as the pixel point, by the sum, to record the product as the grayscale feature value of the pixel point;
[0041] The grayscale eigenvalues of all pixels in the first region are mapped to the interval [0, 255], the original grayscale values are replaced by the mapping results of the grayscale eigenvalues, and the replaced first region is used as the target region image.
[0042] The present invention has at least the following beneficial effects:
[0043] The present invention first obtains historical MRI images of the spines of multiple reference persons, historical MRI images of the spines of the persons to be tested, and current MRI images, and uses the historical MRI images of the spines of the reference persons and the historical MRI images of the spines of the persons to be tested as reference images, and screens ROI regions based on the reference images of each person's spine. The ROI regions are regions that need to be focused on, and then all reference images are divided into multiple categories based on the difference characteristics of the position distribution and the difference characteristics of the shape distribution of the ROI regions in the reference images of each pair, and an evolution model corresponding to each category of reference images is constructed in combination with the position distribution of the ROI regions in each category of reference images. The evolution model can reflect the changes in the ROI regions in the reference images of the corresponding category. Since the time interval between the current MRI image of the spine of the person to be tested and the previous adjacent historical MRI image is short in acquisition time, the similarity of the features in the two MRI images is high. Therefore, the current MRI image is put into the evolution model where the previous historical MRI image is located, and combined with the grayscale value of the pixel points in the predicted scan image, the target area image of the person to be tested is determined. The target area image is the image of the area of interest of the spine of the person to be tested after the grayscale is corrected. When the doctor subsequently analyzes whether the person to be tested has a spinal infection and the degree of infection, he can directly analyze the target area image, which saves detection time and can effectively improve the accuracy of the doctor's diagnosis of the spinal infection of the person to be tested. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0045] Figure 1 A flowchart of a method performed by a computer-aided analysis system for spinal infection MRI images provided by an embodiment of the present invention;
[0046] Figure 2 This is the unit structure diagram of the LSTM model. DETAILED DESCRIPTION
[0047] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the computer-aided analysis system for spinal infection MRI images proposed by the present invention is described in detail below in combination with the accompanying drawings and preferred embodiments.
[0048] Unless defined otherwise, 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 belongs.
[0049] The specific scheme of the computer-aided analysis system for spinal infection MRI images provided by the present invention will be described in detail below with reference to the accompanying drawings.
[0050] Example of a computer-aided analysis system for spinal infection MRI images:
[0051] The specific scenario targeted by this embodiment is: when judging the spinal infection status of the person to be tested, it is necessary to first collect MRI images of the spine of the person to be tested. The collected images can roughly show the more abnormal areas, that is, the areas of interest. When subsequently analyzing whether the person to be tested has a spinal infection, it is necessary to pay higher attention to these areas and focus on detecting these areas. However, since the collected images are affected by the collection equipment or environmental factors, the grayscale values of the pixels in the area of interest are different from the actual situation, and the subsequent judgment results of the images are different. Therefore, in order to improve the accuracy of the subsequent detection results, the grayscale values of the pixels in these areas need to be corrected so that the grayscale characteristics presented in the area of interest can reflect the actual situation, providing a reliable reference basis for the doctor's subsequent detection.
[0052] This embodiment proposes a computer-aided analysis system for spinal infection MRI images. The system includes a memory and a processor. The processor executes a computer program stored in the memory to implement the following. Figure 1 The specific steps are as follows:
[0053] Step S1, obtaining a historical MRI image of the spine of a reference person, a historical MRI image of the spine of a person to be tested, and a current MRI image.
[0054] First, all spinal MRI images of multiple people with spinal infection during the detection process within a historical time period are obtained from the hospital database. These people are recorded as reference people, and the images obtained are recorded as historical MRI images of the spine of the reference people. The reference people are people with spinal infection, excluding patients with spinal tumors, spinal fractures, and spinal osteoporosis. In this embodiment, the number of reference people is 1000. In specific applications, the implementer can set it according to specific circumstances. At the same time, all spinal MRI images of the person to be tested during the detection process within a historical time period are obtained from the hospital database. The historical time period is a set of all historical moments whose time interval with the current moment is less than or equal to the preset time length. In this embodiment, the preset time length is half a year. In specific applications, the implementer can set it according to specific circumstances. And collect the MRI image of the spine of the person to be tested at the current moment, and record the MRI image as the current MRI image of the person to be tested.
[0055] It should be noted that all MRI images in this example were acquired using the same scanner, such as a Simens Aavanto 1.5T scanner, and the same contrast agent dose of gadoterate dimeglumine: 0.1 mmol / kg. All MRI images are rectangular and of equal size. All image data in this example was acquired with permission.
[0056] So far, this embodiment has acquired a plurality of historical MRI images of the spine of the reference persons, historical MRI images of the spine of the person to be tested, and current MRI images.
[0057] Step S2: obtaining a predicted scan image of each person and cell state values corresponding to pixel points in the reference image based on a reference image of each person's spine, wherein the reference image includes a historical MRI image of the reference person's spine and a historical MRI image of the spine of the person to be tested; and screening the ROI area in the reference image according to the change in the cell state value corresponding to the same person.
[0058] Examination of spinal infection is a long-term and continuous process, so it is extremely important for the prevention and treatment of lesions. Predicting changes in lesions can not only assist doctors in making treatment arrangements in advance, but also detect sudden or abnormal changes in patients' symptoms as early as possible. The change pattern of the patient's own historical images can be used to make preliminary predictions about subsequent images, but the infection does not change linearly, and relying solely on the patient's personal data lacks a certain degree of accuracy. The MRI image data of the reference persons obtained in the hospital database includes a large number of imaging changes in patients with spinal infections, and there are also some medical records similar to the persons to be tested. The development of these medical records also reflects the changes in the spinal column of the persons to be tested. Therefore, combining the data of the persons to be tested with the data of similar patients can more accurately predict the changes in the lesions of the persons to be tested.
[0059] For the same person, image changes are also possible changes in lesions, which are the areas of interest that doctors focus on. The changes in the images are the grayscale changes of the pixels therein. Therefore, by analyzing the grayscale changes of the pixels at the corresponding spinal position in each imaging examination, the possibility of the pixels being the disease area can be obtained. The Long Short-Term Memory (LSTM) network model can analyze the long-term changes in time series data. Another advantage of the LSTM model is that it can combine historical changes in the data to predict future changes, thereby being able to obtain image changes for subsequent imaging examinations of any person. Therefore, this embodiment uses the LSTM model to extract features from MRI images to predict the MRI image of each person at the next moment.
[0060] All historical MRI images of the spine of the reference person and all historical MRI images of the spine of the person to be tested that have been collected are recorded as reference images.
[0061] For any individual, all reference images of that individual are fed into the LSTM model in the order in which they were collected within that individual's historical time period. The output of the LSTM model is the next MRI image, which is recorded as the predicted scan image for that individual. Since all reference images are of the same size, there is a one-to-one correspondence between pixel positions within the reference images. Assuming that there are n reference images for that individual, there are n grayscale values for any position, and the n+1th grayscale value for that position is obtained. During the LSTM model training process, a binary cross-entropy loss function is used, and continuous iteration is performed to reduce the loss. The LSTM model training process is conventional and will not be further described here. The LSTM model can be used to obtain predicted scan images for each reference individual and the predicted scan image for the individual being tested.
[0062] like Figure 2 As shown in the figure, the unit structure of the LSTM model, t is the time point, x is the input, C t is the cell state value at time point t, C t-1 is the cell state value at time point t-1, h t is the state information at time point t, h t-1 is the state information at time point t-1, σ and Tanh are activation functions; C forms the long-term memory of grayscale changes, and h forms the short-term memory of grayscale changes. During the training process, if the grayscale changes of the pixel are stable, the forget gate f tWhen the grayscale of a pixel changes suddenly, the cell state is largely unchanged. Therefore, the cell state can reflect the long-term changes in the pixel. Within a region of interest, there is a sequence of cell state changes: pixels closer to the edge change later, and the cell state can remember earlier grayscale changes in intermediate pixels. Using the LSTM model, we can obtain the cell state value corresponding to each pixel in each reference image.
[0063] For any person:
[0064] Next, a reference image of the person is used as an example for description. Other images of the person and reference images of other persons can be processed using the method provided in this embodiment.
[0065] Specifically, any other reference image of the person except the first reference image is recorded as a candidate image.
[0066] For any pixel point in the candidate image: obtain the forward difference sequence of the cell state value of the pixel point; it should be noted that: the forward sequence of the pixel point is obtained based on the grayscale value of the pixel point in the candidate image and the grayscale value of the pixel point at the corresponding position in all historical MRI images before the candidate image of the person. The normalized result of the average value of the absolute value of the elements in the forward difference sequence is used as the first possibility value that the pixel point belongs to the ROI area; using this method, the first possibility value that each pixel point in the candidate image belongs to the ROI area can be obtained. There are many methods for normalizing data. In this embodiment, the maximum and minimum value normalization method is used to process the average value. This method is a prior art and will not be described in detail here. This embodiment does not extract the ROI area for the first reference image of each person.
[0067] Based on the first probability value that all pixels in the candidate image belong to the ROI area, the DBSCAN clustering algorithm is used to cluster all pixels in the candidate image to obtain multiple clusters; the DBSCAN clustering algorithm is an existing technology and will not be described in detail here.
[0068] Next, the ROI region in the candidate image is screened based on the distribution of the first likelihood values of the pixels within each cluster. Specifically, the average first likelihood value of all pixels in the candidate image is calculated, and the average first likelihood value of all pixels within each cluster in the candidate image is calculated separately. Each cluster in the candidate image has an average first likelihood value. Clusters whose average first likelihood value of all pixels within the cluster is greater than the average first likelihood value of all pixels in the candidate image are determined as ROI regions in the candidate image.
[0069] By using the above method, the ROI areas in all reference images can be screened out.
[0070] Step S3: Divide all reference images into multiple categories according to the difference characteristics of the position distribution and the shape distribution of the ROI regions in the reference images; and construct a corresponding evolution model based on the position distribution of the ROI regions in each category of reference images.
[0071] Because the progression of a patient's condition is not linear and may include sudden changes or new infections, identifying a reference patient with similar spinal characteristics and combining these with subsequent changes in the lesion area of similar patients can more accurately predict the changing characteristics of the patient's ROI. Furthermore, in the process of identifying reference patients with similar changing characteristics, the images of patients with spinal infections are integrated and classified, allowing for a model of the evolution of the changing characteristics of the same reference images from the beginning to the end of the disease course.
[0072] The more the regions of interest of the two reference images overlap, the more similar the change characteristics of the ROI regions in the images of the corresponding persons are. Therefore, this embodiment will evaluate the likelihood that each of the two reference images belongs to the same class and obtain a corresponding second likelihood value.
[0073] For any ROI region in any reference image, a matching region pair is formed between the ROI region in the reference image and the ROI region in each reference image other than the reference image that has the smallest Euclidean distance to the ROI region in the reference image. Using this method, matching regions of all ROI regions in the reference image are screened to obtain multiple matching region pairs. It should be noted that the Euclidean distance between two ROI regions in different reference images is obtained by calculating the Euclidean distance between the coordinates of the center points of the two ROI regions and using this Euclidean distance as the Euclidean distance between the corresponding two ROI regions. For each reference image, when obtaining the coordinates of each pixel point in the reference image, a two-dimensional rectangular coordinate system is constructed with the lower left corner vertex of each reference image as the coordinate origin, the lower edge line of each reference image as the x-axis, and the lower edge line of each reference image as the y-axis to obtain the coordinates of each pixel point in each reference image. In specific applications, the implementer can set the origin, x-axis, and y-axis of the two-dimensional rectangular coordinate system according to the specific situation.
[0074] For any ROI region in the reference image, a change factor for the ROI region is then determined based on the area of the ROI region. Specifically, the time interval between the acquisition time of the reference image containing the ROI region within the historical time period and the acquisition time of the first historical MRI image of the person corresponding to the reference image containing the ROI region within the historical time period is recorded as the first time interval, and the ratio between the area of the ROI region and the first time interval is recorded as the change factor of the ROI region. Using this method, the change factor of each ROI region in the reference image can be obtained.
[0075] For any ROI region in the reference image, the morphological coefficient of the ROI region is determined based on the shape distribution of the ROI region. Specifically, first, the binary mask of the ROI region is extracted using the U-Net segmentation result, and the lesion contour point set is obtained through the findContours function of OpenCV. Then, the contour points in the contour point set are uniformly sampled to construct a shape context descriptor. In this embodiment, the number of uniformly sampled points is set to 100. In specific applications, the implementer can set it according to the specific situation. With each contour point as the center, a histogram in the logarithmic polar coordinate system is constructed (5 angle partitions × 12 distance partitions). The histogram counts the density of the surrounding contour points in different areas to form a 60-dimensional feature vector (5×12). Based on the contour point set, 7 Hu invariant moments are calculated. Since the Hu moment is insensitive to translation, rotation, and scaling, it is used to reflect the moment characteristics of the overall shape of the matching ROI region, such as symmetry and ductility. Finally, the shape context (60 dimensions) and Hu moment (7 dimensions) are concatenated into a 67-dimensional hybrid feature vector. Principal component analysis (PCA) is used to reduce the dimensionality to 3 dimensions, retaining more than 95% of the feature information. The first morphological vector PC1, the second morphological vector PC2, and the third morphological vector PC3 are obtained. Next, the morphological coefficient of the ROI region is obtained by combining the first morphological vector, the second morphological vector, and the third morphological vector. The morphological coefficient of the ROI region can be expressed as:
[0076]
[0077] Wherein, β represents the morphological coefficient of the ROI area, L1 represents the modulus of the first morphological vector, L2 represents the modulus of the second morphological vector, and L3 represents the modulus of the third morphological vector.
[0078] This embodiment determines the morphological coefficient of the ROI region by integrating the first morphological vector, the second morphological vector, and the third morphological vector.
[0079] In computer-assisted analysis of spinal infection MRI images, the morphological characteristics of lesions (such as contour, edge regularity, and internal structure) are key to distinguishing infection types (such as purulent and tuberculous), determining the stage of the lesion, and predicting treatment efficacy. Typically, the morphology of purulent and tuberculous lesions may show significant differences in MRI images. Purulent infections are characterized by blurred lesion boundaries, irregular morphology, and possible abscess spread. Tuberculous infections, on the other hand, are characterized by lesions that are mostly circular or lobed, with relatively clear boundaries of the central necrotic zone. Therefore, by extracting the geometric distribution characteristics of the lesion contour, the regularity of the lesion can be quantitatively described. The morphological coefficient reflects the morphological characteristics of the ROI region. The more similar the morphological characteristics of the ROI regions in the image, the greater the likelihood that the corresponding individuals have the same type of infection.
[0080] For any ROI region: a pair consisting of the change factor of the ROI region and the morphological coefficient of the ROI region is used as the feature group of the ROI region. Using the above method, the feature group of each ROI region can be obtained.
[0081] For any two reference images, the Euclidean distance between the feature groups of the two ROIs in each matching region pair in the two reference images is calculated. This distance is used as the feature distance for the corresponding matching region pair. Each matching region pair has a corresponding feature distance. Based on the positional differences and feature distances between the two ROIs in all matching region pairs in the two reference images, a second likelihood value is obtained that the two reference images belong to the same class. The positional differences and feature distances are both negatively correlated with the second likelihood value.
[0082] Among them, the negative correlation relationship means that the dependent variable will decrease as the independent variable increases, and the dependent variable will increase as the independent variable decreases. It can be a subtraction relationship, a division relationship, etc., which is determined by actual application.
[0083] In this embodiment, a specific calculation formula for the second likelihood value is given. The second likelihood value that the two reference images belong to the same class can be expressed as:
[0084]
[0085] Where G represents the second probability value that the two reference images belong to the same class, W represents the number of matching region pairs in the two reference images, and l w represents the Euclidean distance between the center points of the two ROI regions in the w-th matching region pair in the two reference images, d w represents the characteristic distance of the w-th matching region pair in the two reference images, and exp[] represents an exponential function with a natural constant as the base.
[0086] The Euclidean distance between the center points of the two ROI regions in the w-th matching region pair represents the position difference between the two ROI regions in the w-th matching region pair in the image. The larger the distance, the greater the position difference between the two ROI regions in the image. The characteristic distance of the w-th matching region pair is used to reflect the difference between the shape features of the two ROI regions in the w-th matching region pair. This embodiment combines the position difference and the shape difference to determine the second probability value that the two reference images belong to the same class. When the position difference between the two regions in the matching region pair in the two reference images is smaller and the shape difference is smaller, it means that the two reference images are more similar, the two images are more likely to belong to the same class, and the second probability value that they belong to the same class is larger.
[0087] Using this method, we can obtain the second likelihood value that each pair of reference images belongs to the same class. Next, we use the DBSCAN clustering algorithm to cluster all reference images based on this second likelihood value, obtaining multiple clusters. All reference images within the same cluster are considered a single class. When using the DBSCAN clustering algorithm for clustering, the distance threshold is set to 0.2, and the sample number threshold is set to 2. In specific applications, the user can adjust these settings based on their specific circumstances. The DBSCAN clustering algorithm is currently available and will not be further elaborated here.
[0088] For any type of reference image:
[0089] Calculate the total area of all ROI regions in each reference image of this type of reference image respectively; sort all reference images in this type of reference image in ascending order of total area to obtain an image sequence corresponding to this type of reference image; map all reference images in the image sequence to a three-dimensional rectangular coordinate system. In this embodiment, the X-axis of the three-dimensional rectangular coordinate system represents the order of the corresponding reference images in the image sequence. All reference images in the image sequence are mapped in the YOZ plane of the three-dimensional rectangular coordinate system, that is, the reference images are parallel to the YOZ plane in the three-dimensional rectangular coordinate system.
[0090] In a three-dimensional rectangular coordinate system, the SIFT algorithm is used to match the pixels of any ROI region in each reference image with the corresponding pixels in the next adjacent image in the three-dimensional rectangular coordinate system to obtain multiple matching point sequences. The SIFT algorithm is an existing technology and will not be described in detail here.
[0091] The least squares method is used to fit all the pixels in each matching point sequence into straight lines, and the fitting lines corresponding to all the edge points of the same region of interest are spatially fitted into quadratic surfaces; the closed area enclosed by all quadratic surfaces is the evolution model of this type of reference image.
[0092] By using the above method, the evolution model corresponding to each type of reference image can be obtained.
[0093] Step S4, synthesizing the position of the previous historical MRI image of the current MRI image of the spine of the person to be tested in the evolution model and the grayscale value of the pixel point in the predicted scan image to obtain the target area image of the person to be tested.
[0094] Since the time interval between the current MRI image of the spine of the person to be tested and its adjacent previous historical MRI image is short in acquisition time, the similarity of the features in the two MRI images is high. Therefore, the evolution model of the previous historical MRI image of the current MRI image is regarded as the evolution model of the current MRI image.
[0095] The evolution model of the previous historical MRI image of the current MRI image of the person to be tested is recorded as the target evolution model, and the current MRI image is mapped to the position of the previous historical MRI image of the current MRI image in the target evolution model. The area in the current MRI image mapping result that is at the same position as the area of the previous historical MRI image of the current MRI image in the evolution model is taken as the first area in the current MRI image, that is, the area of interest in the current MRI image is obtained. In order to provide an accurate reference basis for the doctor's subsequent evaluation, the grayscale value of the pixel points in this area needs to be corrected to make it more consistent with the actual situation.
[0096] When the grayscale difference between the predicted scan image of the person to be tested and the first area in the current MRI image is small, it indicates that the prediction result based on the historical MRI image data of the person to be tested is relatively accurate; if the difference is large, it means that the person to be tested may currently have more special changes, so more reference should be made to the grayscale distribution characteristics of the pixel points in the first area in the current MRI image.
[0097] Based on this, this embodiment uses the grayscale values of the pixels at positions corresponding to the first region in the current MRI image in the predicted scan image of the person to be tested to correct the grayscale values of the pixels in the first region in the current MRI image.
[0098] Specifically, for any pixel point in the first area: calculate the first difference between the grayscale value of the pixel point and the grayscale value of the pixel point at the same position as the pixel point in the predicted scan image of the person to be tested; calculate the sum of the constant 1 and the normalized value of the first difference; and record the product of the grayscale value of the pixel point at the same position as the pixel point in the predicted scan image of the person to be tested and the sum as the grayscale feature value of the pixel point.
[0099] In this embodiment, a calculation formula for the grayscale eigenvalue is given. The grayscale eigenvalue of the r-th pixel in the first region can be expressed as:
[0100] q″ r =q r ×[1+tanh(q′ r -q r )]
[0101] Among them, q r represents the grayscale eigenvalue of the rth pixel in the first region, q r represents the grayscale value of the pixel at the same position as the rth pixel in the first area in the predicted scan image of the person to be tested, q' r represents the grayscale value of the r-th pixel in the first region, and tanh() represents the hyperbolic tangent function.
[0102] q' r -q r represents the first difference, and the hyperbolic tangent function is used to normalize the first difference.
[0103] By adopting the above method, the grayscale characteristic value of each pixel in the first area can be obtained.
[0104] Map the grayscale eigenvalues of all pixels in the first region to the interval [0, 255]. Use the resulting grayscale eigenvalue mapping to replace the original grayscale values, and use the replaced first region as the target region image. Mapping the grayscale eigenvalues of all pixels in the first region to the interval [0, 255] is a prior art method and will not be further described here.
[0105] At this point, this embodiment has obtained the target area image of the person to be tested. The target area image is the image of the area of interest of the spine of the person to be tested after grayscale correction. It is marked in the current MRI image on the computer to display the position and range of the area of interest of the spine of the person to be tested in an intuitive manner. When the doctor subsequently analyzes whether the person to be tested has a spinal infection and the degree of infection, he can directly analyze the target area image, saving judgment time and improving detection efficiency.
[0106] This embodiment first obtains historical MRI images of the spines of multiple reference persons, historical MRI images of the spines of the persons to be tested, and current MRI images, and uses the historical MRI images of the spines of the reference persons and the historical MRI images of the spines of the persons to be tested as reference images, and screens ROI regions based on the reference images of each person's spine. The ROI regions are regions that require special attention, and then all reference images are divided into multiple categories based on the difference characteristics of the position distribution and the difference characteristics of the shape distribution of the ROI regions in the reference images of each pair, and an evolution model corresponding to each category of reference images is constructed in combination with the position distribution of the ROI regions in each category of reference images. The evolution model can reflect the changes in the ROI regions in the reference images of the corresponding category. Since the time interval between the current MRI image of the spine of the person to be tested and the previous adjacent historical MRI image is short in acquisition time, the similarity of the features in the two MRI images is high. Therefore, the current MRI image is put into the evolution model where the previous historical MRI image is located, and combined with the grayscale value of the pixel points in the predicted scan image, the target area image of the person to be tested is determined. The target area image is the image of the area of interest of the spine of the person to be tested after the grayscale is corrected. When the doctor subsequently analyzes whether the person to be tested has a spinal infection and the degree of infection, he can directly analyze the target area image, which saves detection time and can effectively improve the accuracy of the doctor's diagnosis of the spinal infection of the person to be tested.
[0107] In other embodiments, a computer-assisted analysis device for spinal infection MRI images is provided, comprising a memory and a processor. The memory is used to store executable program code, and the processor is used to call and run the executable program code from the memory, so that the device executes the method performed by the computer-assisted analysis system for spinal infection MRI images described above. The device can specifically be a chip, component, or module, and the chip may include a connected processor and memory; wherein the memory is used to store instructions, and when the processor calls and executes the instructions, the chip can execute the method performed by the computer-assisted analysis system for spinal infection MRI images provided in the above embodiments.
[0108] In other embodiments, a computer program product is also provided. When the computer program product is run on a computer, it enables the computer to execute the above-mentioned related steps to implement the method performed by the computer-aided analysis system for spinal infection MRI images provided in the above-mentioned embodiment.
[0109] In other embodiments, a computer-readable storage medium is also provided, in which a computer program code is stored. When the computer program code is run on a computer, the computer executes the above-mentioned related method steps to implement the method performed by the computer-assisted analysis system for spinal infection MRI images provided in the above-mentioned embodiment.
[0110] Among them, the provided devices, computer program products, and computer-readable storage media are all used to execute the corresponding methods provided above. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding methods provided above, and will not be repeated here.
[0111] It should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A computer-aided analysis system for spinal infection MRI images, comprising a memory and a processor, characterized in that: The processor executes the computer program stored in the memory to implement the following steps: Obtaining historical MRI images of the spine of a reference person, historical MRI images of the spine of the person to be tested, and current MRI images; Obtaining a predicted scan image of each person and cell state values corresponding to pixels in the reference image based on a reference image of the spine of each person, the reference image comprising a historical MRI image of the spine of the reference person and a historical MRI image of the spine of the person to be tested; and screening a ROI region in the reference image based on changes in the cell state values corresponding to the same person; All reference images are divided into multiple categories based on the difference characteristics of the position distribution and shape distribution of the ROI regions in the reference images. The corresponding evolution model is constructed based on the position distribution of the ROI regions in each category of reference images. The target area image of the person to be tested is obtained by integrating the position of the previous historical MRI image of the current MRI image of the spine of the person to be tested in the evolution model and the grayscale value of the pixel point in the predicted scan image.
2. The computer-aided analysis system for spinal infection MRI images according to claim 1, characterized in that: The predicted scan image includes: For any person: all reference images of the said person are input into the LSTM model to obtain the predicted scan image of the said person.
3. The computer-aided analysis system for spinal infection MRI images according to claim 2, characterized in that: The cell state value corresponding to the pixel point in the reference image includes: obtaining the cell state value of each pixel point in each reference image through the LSTM model.
4. The computer-aided analysis system for spinal infection MRI images according to claim 1, characterized in that: The screening of the ROI region in the reference image according to the change of the cell state value corresponding to the same person includes: For any person: For any pixel point in the candidate image: obtaining a forward difference sequence of the cell state value of the any pixel point; taking the normalized average value of the absolute values of the elements in the forward difference sequence as the first probability value that the any pixel point belongs to the ROI area; Based on the first likelihood values of all pixels in the candidate image belonging to the ROI region, clustering all pixels in the candidate image to obtain clusters; Screening the ROI region in the candidate image according to the distribution of the first likelihood values of the pixels in each cluster; The candidate image is any other reference image of the person except the first reference image.
5. The computer-aided analysis system for spinal infection MRI images according to claim 4, characterized in that: The screening of the ROI region in the candidate image according to the distribution of the first likelihood values of the pixels in each cluster includes: A cluster in which the average value of the first likelihood values of all pixels in the cluster is greater than the average value of the first likelihood values of all pixels in the candidate image is determined as the ROI region in the candidate image.
6. The computer-aided analysis system for spinal infection MRI images according to claim 1, characterized in that: According to the difference characteristics of the position distribution and the shape distribution of the ROI regions in the reference images, all reference images are divided into multiple categories, including: For any reference image, each ROI region in the reference image and the ROI region in each reference image other than the reference image having the smallest Euclidean distance with each ROI region in the reference image form a matching region pair; For any ROI region, determining a variation factor of the ROI region according to the area of the ROI region; determining a morphological coefficient of the ROI region according to the shape distribution of the ROI region; the variation factor and the morphological coefficient constitute a feature group of the ROI region; For any two reference images: using the Euclidean distance between the feature groups of two ROI regions in the matching region pair of the any two reference images as the feature distance of the corresponding matching region pair; obtaining a second likelihood value that the any two reference images belong to the same class based on the position difference and feature distance between the two ROI regions in all matching region pairs of the any two reference images, wherein the position difference and the feature distance are both negatively correlated with the second likelihood value; All reference images are clustered based on the second likelihood values of the pairwise reference images, and all reference images in the same cluster are regarded as one class.
7. The computer-aided analysis system for spinal infection MRI images according to claim 6, characterized in that: The determining of the morphological coefficient of any ROI region according to the shape distribution of any ROI region includes: Extract the binary mask of any ROI area and obtain the lesion contour point set through the findContours function of OpenCV; Uniformly sample the contour points in the contour point set to construct a shape context descriptor; based on the shape context descriptor, obtain the first morphological vector, the second morphological vector, and the third morphological vector of any ROI region: The morphological coefficient of any ROI region is obtained by integrating the first morphological vector, the second morphological vector and the third morphological vector.
8. The computer-aided analysis system for spinal infection MRI images according to claim 1, characterized in that: The step of constructing a corresponding evolution model based on the position distribution of the ROI region in each type of reference image includes: For any type of reference image: Calculating the total area of all ROI regions in each reference image of any one type of reference images respectively; sorting all reference images in any one type of reference images in ascending order of total area to obtain an image sequence corresponding to any one type of reference images; Mapping all reference images in the image sequence into a three-dimensional rectangular coordinate system, wherein the X axis of the three-dimensional rectangular coordinate system represents the order of the corresponding reference images in the image sequence, and all reference images in the image sequence are mapped in the YOZ plane of the three-dimensional rectangular coordinate system; In a three-dimensional rectangular coordinate system, the SIFT algorithm is used to match the pixel points of any ROI area in each reference image in the three-dimensional rectangular coordinate system with the pixel points at the corresponding position in the next adjacent image to obtain multiple matching point sequences; The least squares method is used to fit all the pixels in each matching point sequence into straight lines, and the fitting lines corresponding to all the edge points of the same region of interest are spatially fitted into quadratic surfaces; the closed area enclosed by all quadratic surfaces is the evolution model of this type of reference image.
9. The computer-aided analysis system for spinal infection MRI images according to claim 1, characterized in that: The method of synthesizing the position of the previous historical MRI image of the current MRI image of the spine of the person to be tested in the evolution model and the grayscale value of the pixel point in the predicted scan image to obtain the target area image of the person to be tested includes: Mapping the current MRI image to the position of the previous historical MRI image in the evolution model where the previous historical MRI image is located, and taking the region in the current MRI image mapping result that is at the same position as the region of the previous historical MRI image in the evolution model as the first region in the current MRI image; The grayscale values of the pixels in the first area of the current MRI image are corrected using the grayscale values of the pixels at positions corresponding to the first area in the predicted scan image of the person to be tested, to obtain the target area image.
10. The computer-aided analysis system for spinal infection MRI images according to claim 9, characterized in that: The method of correcting the grayscale values of the pixels in the first area of the current MRI image using the grayscale values of the pixels at positions corresponding to the first area in the predicted scan image of the person to be tested to obtain the target area image includes: For any pixel point in the first area: calculating a first difference between the grayscale value of the pixel point and the grayscale value of the pixel point in the predicted scan image of the person to be tested, which is located at the same position as the pixel point; calculating the sum of a constant 1 and a normalized value of the first difference; and multiplying the grayscale value of the pixel point in the predicted scan image of the person to be tested, which is located at the same position as the pixel point, by the sum, to record the product as the grayscale feature value of the pixel point; The grayscale eigenvalues of all pixels in the first region are mapped to the interval [0, 255], the original grayscale values are replaced by the mapping results of the grayscale eigenvalues, and the replaced first region is used as the target region image.