Osteomyelitis lesion analysis method based on medical image registration

Abstract

The invention relates to the technical field of image registration analysis, and particularly discloses an osteomyelitis lesion analysis method based on medical image registration, which ensures the precision of registration information by providing a precise target area and ensures the accuracy of a lesion image evaluation result extracted after registration. The method comprises the following steps: step 1, carrying out a multimode registration simulation experiment on a reference image and a floating image based on a mutual information technology to obtain a registration target area; the reference image is a CT image, and the floating image is a CBCT image; step 2, carrying out test analysis after carrying out spatial transformation on the registration target area, and judging whether the actual mutual information value between the two images meets the requirement or not according to a test analysis result: if so, obtaining the information of the registration target area and extracting the feature of a lesion area; if not, returning to the step 1 to adjust the target area range of multimode registration; 3, performing statistical analysis on lesion area features to obtain lesion area images; and 4, analyzing the lesion area image and outputting an evaluation result.

Description

Technical Field

[0001] The present invention relates to the technical field of image registration analysis, and particularly relates to a method for analyzing osteomyelitis lesions based on medical image registration. Background Art

[0002] There are many types of medical images, and common ones include: X-ray images, CT scan images, MRI magnetic resonance imaging images, ultrasound images, etc. Images can be two-dimensional, three-dimensional or even four-dimensional. The image registration technology has also developed from the initial registration based on external gold markers of positioning devices to the semi-automated era that requires manual feature extraction, and then to today's fully automated registration. With various intelligent algorithms, its accuracy, speed and robustness have been continuously improved.

[0003] In terms of osteomyelitis lesions, through the image registration technology, by aligning and mapping the template image and the detection image spatially, the points (homologous points) corresponding to the same spatial position in the two images are made to correspond one by one, achieving consistency in spatial position, so that doctors can observe the anatomical structure and pathological changes of the osteomyelitis lesion area more comprehensively and intuitively. By registering and analyzing the bone information in the CT image, it can help doctors more accurately evaluate the lesion range and severity of osteomyelitis.

[0004] However, in the existing analysis process of osteomyelitis lesions based on medical image registration technology, when performing spatial transformation each time, due to the offset of pixel points, it is necessary to use interpolation technology to transform the offset positions each time. However, since the position accuracy after transformation cannot be guaranteed, it is easy to cause the target area to shift during the registration process, affecting the registration accuracy. Furthermore, it is easy to cause deviations during the process of extracting the osteomyelitis lesion image after registration, resulting in inaccurate lesion evaluation results. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for analyzing osteomyelitis lesions based on medical image registration, and solve the following technical problems:

[0006] How to provide an accurate target area to ensure the accuracy of registration information, and further ensure the accuracy of the evaluation results of the lesion image extracted after registration.

[0007] The purpose of the present invention can be achieved through the following technical solutions:

[0008] A method for analyzing osteomyelitis lesions based on medical image registration, the method includes:

[0009] Step 1: Perform a multi-modal registration simulation experiment on the reference image and the floating image based on mutual information technology to obtain a registration target area; the reference image is a CT image, and the floating image is a CBCT image;

[0010] Step 2: Perform spatial transformation on the registration target area and then conduct test analysis. Determine whether the actual mutual information value between the two images meets the requirements according to the results of the test analysis:

[0011] If so, obtain the registration target area information and extract the lesion area features;

[0012] If not, return to Step 1 to adjust the target area range of multimodal registration;

[0013] Step 3: Statistically analyze the lesion area features to obtain the lesion area image;

[0014] Step 4: Analyze the lesion area image and output the evaluation results.

[0015] Preferably, the method for obtaining the registration target area by performing a multimodal registration simulation experiment on the reference image and the floating image based on the mutual information technology in Step 1 is as follows:

[0016] Obtain the gray values of all pixels in the reference image and the floating image;

[0017] Based on the gray values of all pixels in the reference image, divide the initial background area and the initial target area of the reference image, and extract the initial target boundary line and the initial target area contour;

[0018] Obtain the mutual information value between the reference image and the floating image through simulated experiment registration. Generate a reference registration target area according to the normalized mutual information value, and extract the reference pixel values of the registration target area.

[0019] Preferably, the method for performing test analysis after spatial transformation on the registration target area in Step 2 is as follows:

[0020] Set the initial position of the reference image, and obtain the test pixel values of the registration target area according to the information of the registered fusion image after spatial transformation;

[0021] Take multiple registration target areas and their test pixel values as a test set and input them into a convolutional neural network model for training, and output the corresponding test parameters of the registration target area after optimization;

[0022] Analyze the test parameters by setting a measure strategy to judge the registration quality.

[0023] Preferably, the method for analyzing the test parameters by the measure strategy includes:

[0024] Obtain the measure coefficient Mea through the formula where n is the total number of registration target areas, and i ∈ [1, n]; δ i is the preset weight coefficient of the i-th registration target area; Pt i is the test parameter of the i-th registration target area; Pti0 is the standard test parameter for the i-th registration target area.

[0025] Preferably, the analysis process for determining whether the actual mutual information value between two images meets the requirements is as follows:

[0026] Compare the measure coefficient Mea with the preset standard measure coefficient threshold interval [Mea A , Mea B :

[0027] If Mea ∈ [Mea A , Mea B , it is determined that the actual mutual information value range is normal;

[0028] If Mea > Mea B , it is determined that the actual mutual information value range is exceeded, and the target area range of multimodal registration is reduced;

[0029] If Mea < Mea A , it is determined that the actual mutual information value range is insufficient, and the target area range of multimodal registration is enlarged.

[0030] Preferably, the method for obtaining the registration target area information and extracting the lesion area features is as follows:

[0031] Use the registration target area image as a sample and input it into a convolutional neural network model based on a graph model for image segmentation training to obtain the bone marrow threshold parameter and extract the bone marrow features;

[0032] Generate a "waiting area" for the pixel nodes of the bone marrow features smaller than the bone marrow threshold parameter;

[0033] Use a clustering algorithm to aggregate the similar pixel nodes of the bone marrow features greater than the bone marrow threshold parameter, supplement the "waiting area" information according to the morphological algorithm, and generate the osteomyelitis lesion area features after clustering.

[0034] Preferably, the method for statistically analyzing the lesion area features in step three to obtain the lesion area image is as follows:

[0035] Input the osteomyelitis lesion area features into the trained convolutional neural network model based on the graph model for optimization, and output the osteomyelitis lesion area graph structure;

[0036] According to the osteomyelitis lesion area graph structure, input it into the optimized convolutional neural network model based on the graph model for prediction, and output the probability that each pixel belongs to the lesion area or the normal area;

[0037] Determine the lesion area image in the bone marrow image according to the set lesion threshold.

[0038] Preferably, the method for analyzing the image of the lesion area in step four to output the evaluation result is as follows:

[0039] Calculate the lesion evaluation coefficient Eva through the formula ;

[0040] where m is the cumulative number of mutual information between the reference image and the floating image, and k ∈ [1, m]; Q k is the conversion rate function of the k-th mutual information; q ijk (i, j) is the joint lesion probability distribution of the reference image pixel point i and the floating image pixel point j of the k-th mutual information; q ik (i) is the lesion probability distribution of the reference image pixel point i of the k-th mutual information; q jk (j) is the lesion probability distribution of the reference image pixel point j of the k-th mutual information;

[0041] Compare the lesion evaluation coefficient Eva with the preset standard lesion evaluation coefficient threshold interval [Eva 1 , Eva 2 :

[0042] If Eva < Eva 1 , it is determined that the degree of osteomyelitis lesion after registration is low;

[0043] If Eva 1 ≤ Eva ≤ Eva 2 , it is determined that the degree of osteomyelitis lesion after registration is normal;

[0044] If Eva > Eva 2 , it is determined that the degree of osteomyelitis lesion after registration is high.

[0045] Advantages of the present invention: After performing a spatial transformation on the obtained registration target area, the present invention adds a test analysis process. Through the test analysis design, the accuracy of the true registration target area after the spatial transformation is ensured. By designing and optimizing machine algorithms such as a convolutional neural network model, further test optimization parameter output is performed, and according to the test results, it is judged whether the actual mutual information value between the two images meets the requirements. If it meets the requirements, the registration target area information is obtained and the characteristics of the lesion area are extracted; otherwise, the error after the current transformation is adjusted, and then the target area range of multimodal registration is adjusted to ensure the accuracy of the gray information calculation area value and reduce the computing power cost in the actual registration analysis process. Also, by using a convolutional neural network model based on a graph model to statistically analyze the extracted lesion area characteristics, the lesion area in the bone marrow image is determined according to the set threshold, and then the lesion area image is obtained; by using the trained GCN model for prediction, the output lesion area image is predicted and the evaluation result of the lesion condition is output, realizing the accurate output of the lesion evaluation result. It is ensured that through the above design combining multimodal registration and machine models, the accurate analysis of osteomyelitis lesion conditions is achieved.

[0046] Of course, it is not necessary for any product implementing the present invention to simultaneously achieve all the advantages described above. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0048] Figure 1 It is a schematic diagram of the steps of a method for analyzing osteomyelitis lesions based on medical image registration according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0050] Please refer to Figure 1 As shown, the present invention is a method for analyzing osteomyelitis lesions based on medical image registration, and the method includes:

[0051] Step 1: Perform a multi-modal registration simulation experiment on the reference image and the floating image based on mutual information technology to obtain the registration target area; the reference image is a CT image, and the floating image is a CBCT image;

[0052] Step 2: After performing a spatial transformation on the registration target area, conduct a test analysis, and determine whether the actual mutual information value between the two images meets the requirements according to the results of the test analysis:

[0053] If so, obtain the registration target area information and extract the lesion area features;

[0054] If not, return to Step 1 to adjust the target area range of the multi-modal registration;

[0055] Step 3: Statistically analyze the lesion area features to obtain the lesion area image;

[0056] Step 4: Analyze the lesion area image and output the evaluation result.

[0057] In the above technical solution, in the existing analysis process of osteomyelitis lesions based on medical image registration technology, when performing a spatial transformation each time, due to the offset of pixel points, it is necessary to use interpolation technology to transform the offset positions each time. However, due to the inability to guarantee the position accuracy after transformation, it is easy to cause the target area to shift during the registration process, affecting the registration accuracy. Furthermore, it is easy to cause deviations during the process of extracting the osteomyelitis lesion image after registration, resulting in inaccurate lesion evaluation results.

[0058] Based on the above technical problems, this design realizes providing an accurate target area to ensure the registration information accuracy and the accuracy of the evaluation result of the lesion image extracted after registration by designing a method for analyzing osteomyelitis lesions based on medical image registration. The specific method implementation process is as follows:

[0059] First, read the osteomyelitis lesion medical image in the standard DICOM format and perform multi-modal registration on the reference image and the floating image based on mutual information technology. Use mutual information as the similarity metric, and realize obtaining the registration target area before image registration through regional feature division; where the reference image is a CT image usually used as the standard image, and the floating image is a CBCT image;

[0060] Then, in the process of calculating the mutual information method for selecting and calculating the target area for traditional registration, usually all gray-scale information in the image is involved in the calculation, which easily leads to deviation in the registration result and even registration failure. In this design, after performing a spatial transformation on the obtained registration target area, a test and analysis process is added. Through the test and analysis design, the accuracy of the actual registration target area after the spatial transformation is ensured. Through design optimization of machine algorithms such as convolutional neural network models, further test optimization parameters are output, and according to the test results, it is judged whether the actual mutual information value between the two images meets the requirements; if it meets the requirements, the registration target area information is obtained and the lesion area features are extracted; otherwise, the error after the current transformation is adjusted, and then the target area range of multimodal registration is adjusted to ensure the accuracy of the gray-scale information calculation area value and reduce the computing power cost in the actual registration analysis process. Then, through the use of a convolutional neural network model based on a graph model, statistical analysis is performed on the extracted lesion area features, and the lesion area in the bone marrow image is determined according to the set threshold, and then the lesion area image is obtained. Finally, through the use of the trained GCN model for prediction, the output lesion area image is predicted and the evaluation result of the lesion condition is output, realizing the accurate output of the lesion evaluation result. It is ensured that through the above design combining multimodal registration and machine models, the accurate analysis of osteomyelitis lesion conditions is achieved.

[0061] As an implementation manner of the present invention, the method for obtaining the registration target area by performing a multimodal registration simulation experiment on the reference image and the floating image based on the mutual information technology in step one is as follows:

[0062] Obtain the gray-scale values of all pixels in the reference image and the floating image;

[0063] Based on the gray-scale values of all pixels in the reference image, divide the initial background area and the initial target area of the reference image, and extract the initial target boundary line and the initial target area contour;

[0064] Through the registration of the simulation experiment, obtain the mutual information value of the reference image and the floating image, generate the reference registration target area according to the mapped mutual information value after normalization processing, and extract the reference pixel values of the registration target area.

[0065] In the above technical solution, the method of obtaining the registration target area through multi-modal registration is to analyze the pixel conditions in the reference image and the floating image in advance to distinguish the background area and the target area. Here, the target area mainly refers to the organizational structure associated with osteomyelitis lesions, mainly including the area of the bone. Since the background area and the target area in the reference image are pre-determined, the initial target boundary line and the initial target area contour are extracted to further select the target area information. And through the simulation experiment, the reference image and the floating image are registered and analyzed to obtain the mutual information value. Here, the mutual information value is determined according to the correlation of all pixels between the images. Therefore, the mutual information value obtained through registration is usually greater than the actual mutual information value. The mutual information value after normalization is used to generate a reference registration target area with a determined reference pixel value of the registration target area, ensuring the process of testing the registration target area obtained by the image fusion after spatial transformation in the next step.

[0066] Among them, the mapping method and the specific process are pre-determined according to various reference calculations and selections in the historical simulation experiment, which will not be elaborated here.

[0067] As an implementation manner of the present invention, the method for testing and analyzing the registration target area after spatial transformation in step two is as follows:

[0068] Set the initial position of the reference image, and obtain the test pixel value of the registration target area according to the information of the registered fusion image after spatial transformation;

[0069] Take multiple registration target areas and their test pixel values as a test set and input them into a convolutional neural network model for training, and output the corresponding test parameters of the registration target area after optimization;

[0070] Analyze the test parameters by setting a measure strategy to judge the registration quality.

[0071] In the above technical solution, the determination of the range of pixel values of the registration target area in a larger range is realized through the acquisition of test parameters. The process of spatial transformation of the registration target area specifically includes: setting the initial position in the reference image as the transformation reference, and then determining the test pixel value of the registration target area according to the information of the registered fusion image after spatial transformation; performing sample tests by determining multiple registration target areas and the corresponding test pixel values, inputting them into a convolutional neural network as a test sample set for training, obtaining the test parameters of each registration target area, and then judging the registration situation according to the size of the test parameters to obtain the best registration quality information, realizing the determination of the mutual information value and improving the registration quality.

[0072] Among them, the test parameters are determined by comparing the reference pixel values extracted in Step 1 with the test pixel values during the current test analysis based on the machine model and optimizing according to the transformation parameters of the pixel points in the registration target area; the test parameters fully reflect the actual landing points of the pixel points in each registration target area and the mutual information values of each test pixel.

[0073] As an implementation manner of the present invention, the method for analyzing test parameters by the measure strategy includes:

[0074] Through the formula Calculate to obtain the measure coefficient Mea, where n is the total number of registration target areas, and i ∈ [1, n]; δ i Is the preset weight coefficient of the i-th registration target area; Pt i Is the test parameter of the i-th registration target area; Pt i0 Is the standard test parameter of the i-th registration target area.

[0075] In the above technical solution, the measure strategy is set to analyze the magnitude of the test parameters through the calculation process. Specifically, through the formula Calculate to obtain the measure coefficient Mea, and compare and analyze by limiting the threshold range of the measure coefficient magnitude, and then judge whether the magnitude of the actual mutual information value meets the normal transformation requirements. Specifically, it is determined by accumulating the change situations of the test parameters of different registration target areas, and the measure coefficient magnitude is determined by averaging the calculations of all registration target areas, and then the range of the actual mutual information value is accurately judged.

[0076] Among them, the preset weight coefficient δ i Is set in advance by fitting according to historical experience data, and the standard test parameter Pt i0 Is obtained by selecting the result of the machine simulation of the actual mutual information value magnitude according to historical experience for the target area after successful registration, and will not be elaborated here.

[0077] As an implementation manner of the present invention, the analysis process for judging whether the magnitude of the actual mutual information value between two images meets the requirements is:

[0078] Compare the measure coefficient Mea with the preset standard measure coefficient threshold interval [Mea A , Mea B :

[0079] If Mea ∈ [Mea A , Mea B , it is judged that the range of the actual mutual information value is normal;

[0080] If Mea > Mea B, it is determined that the actual mutual information value range is exceeded, and the target area range of multimodal registration is reduced;

[0081] If Mea < Mea A , it is determined that the actual mutual information value range is insufficient, and the target area range of multimodal registration is enlarged.

[0082] In the above technical solution, the actual mutual information between the reference image and the floating image is judged by means of comparative analysis, and whether the actual mutual information value is within the normal range is reflected according to the size of its measure coefficient. Specifically, if the measure coefficient Mea belongs to the preset standard measure coefficient threshold interval, that is, falls within the interval [Mea A , Mea B , it is judged that the range of the actual mutual information value is normal, which also reflects that an accurate target area range is selected; if it is greater than the maximum value of the interval, that is, Mea > Mea B , it is considered that the range of the actual mutual information value is small, and the range selection configuration of the current multimodal registration target area needs to be reduced; on the contrary, if it is less than the minimum value of the interval, that is, Mea < Mea A , it is considered that the range of the actual mutual information value is large, and the range selection configuration of the current multimodal registration target area needs to be enlarged.

[0083] As an implementation manner of the present invention, the method for obtaining the registration target area information and extracting the lesion area features is as follows:

[0084] Taking the registration target area image as a sample and inputting it into the convolutional neural network model based on the graph model for image segmentation training to obtain the bone marrow threshold parameter and extract the bone marrow features;

[0085] Generating a "waiting area" for the pixel nodes of the bone marrow features smaller than the bone marrow threshold parameter;

[0086] Using the clustering algorithm to aggregate the similar pixel nodes in the bone marrow features that are greater than the bone marrow threshold parameter, supplementing the information of the "waiting area" according to the morphological algorithm, and generating the osteomyelitis lesion area features after clustering.

[0087] In the above technical solution, by designing the method for extracting the lesion area features by obtaining the registration target area information that meets the requirements, the conventional threshold segmentation method is optimized. By combining the threshold segmentation, the clustering algorithm and the morphological algorithm and performing training and analysis through the convolutional neural network model based on the graph model, the accurate extraction of the regional features of the bone marrow lesion is ensured.

[0088] The Graph Convolutional Network (GCN) - based convolutional neural network model in this design is a deep - learning model specifically designed for processing graph - structured data. It extracts features of the graph by applying convolutional operations on nodes and performs tasks such as classification or regression. The core idea of GCN is to generalize the convolutional operation from the traditional Euclidean data space to non - Euclidean data structures such as graphs. Therefore, the specific idea for extracting bone marrow features is as follows: Each pixel in the bone marrow image is regarded as a node, edges are constructed according to the spatial relationship and similarity between pixels, and weights are assigned to the edges according to the difference in gray values; the boundary line of the registration target area and the contour of the registration target area are used for task division to obtain the loss function; and by regarding the pixels in the bone marrow image as nodes, edges are constructed according to the spatial relationship and similarity between pixels, weights are assigned to the edges, and then a graph structure is constructed. A suitable GCN architecture is selected, and the layer structure of the model is defined, including the input layer, multiple graph convolutional layers, and the output layer to construct a graph - model - based convolutional neural network model.

[0089] As an implementation manner of the present invention, the method for statistically analyzing the lesion area features to obtain the lesion area image in step three is as follows:

[0090] Input the osteomyelitis lesion area features into the trained graph - model - based convolutional neural network model for optimization, and output the osteomyelitis lesion area graph structure;

[0091] According to the osteomyelitis lesion area graph structure, input it into the optimized graph - model - based convolutional neural network model for prediction, and output the probability that each pixel belongs to the lesion area or the normal area;

[0092] Determine the lesion area image in the bone marrow image according to the set lesion threshold.

[0093] In the above - mentioned technical solution, the process of further precisely displaying the lesion area image based on the obtained lesion area features includes the processes of optimization and prediction based on the trained graph - model - based convolutional neural network model. Through optimization processing, the osteomyelitis lesion area graph structure is output, and based on the osteomyelitis lesion area graph structure, further prediction analysis of the model is carried out. Finally, the probability that each pixel belongs to the lesion area or the normal area is output, and the specific lesion area is determined according to the set threshold range of the probability, and the lesion area image is output.

[0094] As an implementation manner of the present invention, the method for analyzing the lesion area image to output the evaluation result in step four is as follows:

[0095] Through the formula Calculate to obtain the lesion evaluation coefficient Eva;

[0096] Among them, m is the cumulative number of times of mutual information between the reference image and the floating image, and k ∈ [1, m]; Q k is the conversion rate function of the k-th mutual information; q ijk (i, j) is the joint lesion probability distribution of the reference image pixel point i and the floating image pixel point j of the k-th mutual information; q ik (i) is the lesion probability distribution of the reference image pixel point i of the k-th mutual information; q jk (j) is the lesion probability distribution of the reference image pixel point j of the k-th mutual information;

[0097] Compare the lesion evaluation coefficient Eva with the preset standard lesion evaluation coefficient threshold interval [Eva 1 , Eva 2 :

[0098] If Eva < Eva 1 , it is determined that the degree of osteomyelitis lesion after registration is low;

[0099] If Eva 1 ≤ Eva ≤ Eva 2 , it is determined that the degree of osteomyelitis lesion after registration is normal;

[0100] If Eva > Eva 2 , it is determined that the degree of osteomyelitis lesion after registration is high.

[0101] In the above technical solution, by evaluating and analyzing the output result of the lesion area image based on the prediction process in step three, the specific evaluation process is to obtain the lesion evaluation coefficient Eva, realize the distribution of the lesion information in the mutual information, and evaluate the degree of osteomyelitis lesion. By comparing and analyzing the lesion evaluation coefficient Eva with the preset standard lesion evaluation coefficient threshold interval, the evaluation result after multimodal registration is obtained, and the accurate estimation of the degree of osteomyelitis lesion is realized.

[0102] Among them, it should be explained that the conversion rate function refers to the probability or proportion of a certain event occurring under certain specific conditions; the conversion rate function Q k in this design refers to the pre-fitting calculation obtained by estimating the influence of the mutual information value in different fusion images on the overall bone marrow lesion according to historical experience, which will not be elaborated here.

[0103] Each embodiment in this specification is described in a progressive manner. The same or similar parts between each embodiment can be referred to each other, and the key points of each embodiment are the differences from other embodiments. In particular, for the device, equipment, and non-volatile computer storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiments.

[0104] The above describes specific embodiments of the present specification. Other embodiments are within the scope of the appended documents. In some cases, the actions or steps recited in this application may be performed in a different order than in the embodiments and still achieve the desired results. Additionally, the processes depicted in the drawings do not necessarily require the particular order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0105] The above content is only an example and illustration of the concept of the present invention. Those skilled in the art of the present technology can make various modifications or supplements to the specific embodiments described or use similar methods for substitution, as long as they do not deviate from the concept of the invention or exceed the scope defined by this application, they should fall within the protection scope of the present invention.

Claims

1. A method for analyzing osteomyelitis lesions based on medical image registration, characterized in that: The method comprises: Step 1: Based on the mutual information technology, a multi-modal registration simulation experiment is performed on the reference image and the floating image to obtain the registration target area; the reference image is a CT image, and the floating image is a CBCT image; Step 2: After spatial transformation of the registration target area, test and analyze it, and judge whether the actual mutual information value between the two images meets the requirements according to the test and analysis results: If so, the registration target area information is obtained and the features of the lesion area are extracted; If not, return to step 1 to adjust the target area range of multi-modal registration; Step 3: Statistically analyzing the characteristics of the lesion area to obtain an image of the lesion area; Step 4: Analyze the image output evaluation results of the lesion area.

2. The method for analyzing osteomyelitis lesions based on medical image registration according to claim 1, characterized in that: In the step 1, the method of performing a multi-modal registration simulation experiment on the reference image and the floating image based on mutual information technology to obtain the registration target area is: Get the grayscale values ​​of all pixels in the reference image and the floating image; Divide the initial background area and the initial target area of ​​the reference image based on the grayscale values ​​of all pixels of the reference image, and extract the initial target boundary line and the initial target area contour; The mutual information value of the reference image and the floating image is obtained through simulation experimental registration, and the reference registration target area is generated according to the normalized mutual information value mapping, and the reference pixel value of the registration target area is extracted.

3. The method for analyzing osteomyelitis lesions based on medical image registration according to claim 2, characterized in that: The method for testing and analyzing the registration target area after spatial transformation in step 2 is: The initial position of the reference image is set, and the test pixel value of the registration target area is obtained according to the registration fusion image information after the spatial transformation; Multiple registration target areas and their test pixel values ​​are used as test sets to input the convolutional neural network model for training, and the corresponding test parameters of the registration target areas are output after optimization; The quality of the registration is determined by setting the measurement strategy and analyzing the test parameters.

4. The method for analyzing osteomyelitis lesions based on medical image registration according to claim 3, characterized in that: The method for analyzing test parameters by measuring strategy includes: By formula Calculate the measurement coefficient Mea, where n is the total number of registration target areas, and i∈[1,n]; δ i Pt is the preset weight coefficient of the i-th registration target area; i is the test parameter of the i-th registration target area; Pt i0 is the standard test parameter for the i-th registration target region.

5. The method for analyzing osteomyelitis lesions based on medical image registration according to claim 4, characterized in that: The analysis process for judging whether the actual mutual information value between two images meets the requirements is as follows: The measurement coefficient Mea is compared with the preset standard measurement coefficient threshold interval [Mea A ,Mea B ] for comparison: If Mea∈[Mea A ,Mea B ], then the actual mutual information value range is judged to be normal; If Mea>Mea B , it is judged that it exceeds the actual mutual information value range, and the target area range of multi-mode registration is reduced; If Mea<Mea A , it is judged that it is less than the actual mutual information value range, so the target area range of multi-modal registration is expanded.

6. The method for analyzing osteomyelitis lesions based on medical image registration according to claim 1, characterized in that: The method of obtaining the registration target area information and extracting the lesion area features is: The registered target area image is input as a sample into the convolutional neural network model based on the graph model for image segmentation training, and the bone marrow threshold parameter is obtained to extract the bone marrow features; Generate a "waiting area" for pixel nodes with bone marrow features that are smaller than the bone marrow threshold parameter; The clustering algorithm is used to obtain similar pixel nodes in the bone marrow features that are greater than the bone marrow threshold parameter for clustering, and the "waiting area" information is supplemented according to the morphological algorithm. After clustering, the characteristics of the osteomyelitis lesion area are generated.

7. The method for analyzing osteomyelitis lesions based on medical image registration according to claim 6, characterized in that: The method for obtaining the image of the lesion area by statistically analyzing the features of the lesion area in step 3 is: Input the features of the osteomyelitis lesion area into the trained graph-based convolutional neural network model for optimization, and output the graph structure of the osteomyelitis lesion area; The optimized graph-based convolutional neural network model is used to predict the osteomyelitis lesion area graph structure and output the probability of each pixel belonging to the lesion area or the normal area. The lesion area image in the bone marrow image is determined according to the set lesion threshold.

8. The method for analyzing osteomyelitis lesions based on medical image registration according to claim 1, characterized in that: The method of analyzing the lesion area image and outputting the evaluation result in step 4 is: By formula The lesion assessment coefficient Eva is calculated; Where m is the cumulative number of mutual information between the reference image and the floating image, and k∈[1,m]; Q k is the conversion rate function of the kth mutual information; q ijk (i, j) is the joint lesion probability distribution of the reference image pixel i and the floating image pixel j of the kth mutual information; q ik (i) is the lesion probability distribution of pixel i in the reference image of the kth mutual information; q jk (j) is the lesion probability distribution of pixel j in the reference image of the kth mutual information; Compare the lesion assessment coefficient Eva with the preset standard lesion assessment coefficient threshold interval [Eva1, Eva2]: If Eva < Eva1, the degree of osteomyelitis lesions after registration is judged to be low; If Eva1≤Eva≤Eva2, the degree of osteomyelitis lesions after registration is judged to be normal; If Eva>Eva2, it is judged that the degree of osteomyelitis lesions after registration is high.

Images