Joint cavity positioning method and device based on deep learning model

Abstract

The invention relates to the technical field of articular cavity positioning, and particularly provides an articular cavity positioning method and device based on a deep learning model, and the method comprises the steps: S1, obtaining an arthroscopic image, and carrying out the preprocessing of the arthroscopic image, so as to obtain an enhanced image; s2, acquiring articular cavity position information according to a pre-trained articular cavity positioning model and the enhanced image; the pre-training of the joint cavity positioning model comprises the following steps: A1, training a pre-constructed deep learning model by using a pre-collected general medical image data set to obtain a preliminary positioning model; a2, expanding a pre-collected arthroscopic image data set of the rare bone joint disease to obtain an expanded image data set, and then dynamically adjusting the regularization intensity of the preliminary positioning model based on a self-adaptive regularization coefficient function according to an articular cavity form diversity index of the expanded image data set to obtain an articular cavity positioning model; according to the method, the articular cavity of the rare osteoarticular disease can be accurately positioned.

Description

Technical Field

[0001] This application relates to the technical field of joint cavity positioning. Specifically, it relates to a joint cavity positioning method and device based on a deep learning model. Background Art

[0002] The joint structure of the human body is complex, and there are a large number of blood vessels, nerves, tendons, ligaments and other structures around the joint. When positioning the joint cavity, doctors usually need to find the joint cavity through repeated puncture attempts and gradual corrections. This traditional surgical method not only takes a long time, but also easily causes unnecessary damage to human tissues. In order to shorten the operation time and reduce the damage to human tissues, the existing technology has begun to study using the joint cavity positioning model obtained based on the deep learning model to position the joint cavity.

[0003] However, when applying the above joint cavity positioning technology to the joint cavity positioning of rare bone and joint diseases in a rare bone and joint disease specialty medical center, unprecedented challenges are faced. This is because rare bone and joint diseases have the characteristic of low incidence, that is, the amount of image data corresponding to rare bone and joint diseases available for model training is seriously insufficient. And the deep learning model is a data-driven model, that is, the deep learning model needs to rely on a large amount of image data for training to ensure that the model can accurately identify and position the joint cavity. Therefore, the existing technology has the problem that the joint cavity positioning model obtained based on the deep learning model cannot accurately position the joint cavity of rare bone and joint diseases due to the insufficient amount of training data corresponding to rare bone and joint diseases.

[0004] In addition, the pathological features of rare bone and joint diseases often show a high degree of heterogeneity, and the joint cavity morphology may be significantly different in different patients or at different disease stages of the same patient, that is, the joint cavity morphology shows the characteristic of diversity. The diversity of the joint cavity morphology will make it difficult for the joint cavity positioning model obtained based on the deep learning model to learn general features, resulting in a decrease in the generalization ability and positioning accuracy of the joint cavity positioning model. That is, the existing technology also has the problem that the generalization ability and positioning accuracy of the joint cavity positioning model decrease due to the diversity of the joint cavity morphology.

[0005] In response to the above problems, there is currently no effective technical solution. It should be noted that the above information disclosed in this part is only used to understand the background of the inventive concept of the present invention, and therefore may include information that does not constitute the prior art. Summary of the Invention

[0006] The purpose of this application is to provide a method and device for arthroscopic cavity localization based on a deep learning model, which can effectively solve the problems that the arthroscopic cavity localization model obtained based on the deep learning model cannot accurately localize the arthroscopic cavity of rare osteoarticular diseases due to insufficient training data corresponding to rare osteoarticular diseases, and the generalization ability and localization accuracy of the arthroscopic cavity localization model decrease due to the diversity of arthroscopic cavity morphology.

[0007] In the first aspect, this application provides a method for arthroscopic cavity localization based on a deep learning model, which is used to localize the arthroscopic cavity of rare osteoarticular diseases. The method for arthroscopic cavity localization based on a deep learning model includes the following steps: S1. Obtain arthroscopic images and preprocess the arthroscopic images to obtain enhanced images. The preprocessing includes contrast enhancement; S2. Obtain arthroscopic cavity position information according to a pre-trained arthroscopic cavity localization model and the enhanced images. The arthroscopic cavity localization model is used to localize the arthroscopic cavity in the enhanced images and output the arthroscopic cavity position; The pre-training of the arthroscopic cavity localization model includes: A1. Use a pre-collected general medical image dataset to train a pre-constructed deep learning model to obtain a preliminary localization model; A2. Expand the pre-collected arthroscopic image dataset of rare osteoarticular diseases to obtain an expanded image dataset, and then dynamically adjust the regularization strength of the preliminary localization model based on an adaptive regularization coefficient function according to the arthroscopic cavity morphology diversity index of the expanded image dataset to obtain an arthroscopic cavity localization model.

[0008] The method for arthroscopic cavity localization based on a deep learning model provided by this application is equivalent to first enabling the deep learning model to learn the general features of medical images and the basic knowledge of arthroscopic cavity localization by training the pre-constructed deep learning model with a pre-collected general medical image dataset, and then enabling the deep learning model to learn the unique features of the arthroscopic cavity of rare osteoarticular diseases on the premise of avoiding overfitting by dynamically adjusting the regularization strength of the preliminary localization model based on an adaptive regularization coefficient function according to the arthroscopic cavity morphology diversity index of the expanded image dataset. Therefore, the arthroscopic cavity localization model of this application has good generalization ability and localization accuracy, and this arthroscopic cavity localization model can more accurately localize the arthroscopic cavity of rare osteoarticular diseases, that is, this application can effectively solve the problems that the arthroscopic cavity localization model obtained based on the deep learning model cannot accurately localize the arthroscopic cavity of rare osteoarticular diseases due to insufficient training data corresponding to rare osteoarticular diseases, and the generalization ability and localization accuracy of the arthroscopic cavity localization model decrease due to the diversity of arthroscopic cavity morphology.

[0009] Optionally, step A2 includes: A21. The arthroscopic image dataset of rare osteoarticular diseases collected in advance is augmented by randomly rotating, scaling, and flipping each image in the dataset to obtain an augmented image dataset, which includes multiple augmented images. A22. Use the pre - constructed image quality assessment model to obtain the image quality scores of each augmented image, so as to obtain multiple image quality scores. A23. Normalize all the image quality scores to the interval [0, 1] to obtain the quality weights corresponding to each augmented image, and obtain the joint cavity morphological feature vectors corresponding to each augmented image based on the local binary pattern histogram corresponding to the augmented image. A24. Perform weighted average calculation according to all the quality weights and all the joint cavity morphological feature vectors to obtain the joint cavity morphological diversity index. A25. Dynamically adjust the regularization strength of the preliminary localization model based on the adaptive regularization coefficient function according to the joint cavity morphological diversity index of the augmented image dataset to obtain the joint cavity localization model.

[0010] Since this technical solution is equivalent to introducing image quality assessment and joint cavity morphological diversity index during data augmentation, and using the joint cavity morphological diversity index to dynamically adjust the regularization strength of model training, this technical solution can effectively solve the problem of low accuracy of joint cavity morphological diversity caused by uneven quality of augmented images, thereby effectively improving the generalization ability and localization accuracy of the joint cavity localization model.

[0011] Optionally, the image quality assessment model is used to extract the clarity feature, contrast feature, and noise level feature of each augmented image, and output the corresponding clarity score, contrast score, and noise score. Step A22 includes: A221. For each augmented image, calculate its corresponding image quality score based on the weighted fusion algorithm according to its corresponding clarity score, contrast score, and noise score.

[0012] Since this technical solution can first use the image quality assessment model to independently evaluate its clarity, contrast, and noise level, and output the corresponding scores, and then calculate the image quality score corresponding to each augmented image based on the weighted fusion algorithm according to the clarity score, contrast score, and noise score, it can avoid the situation that the image quality score cannot accurately reflect the overall quality of the image due to evaluating the quality of the augmented image based on only one dimension. Thus, it can effectively improve the training efficiency and localization accuracy of the preliminary localization model, and further effectively improve the generalization ability and localization accuracy of the joint cavity localization model.

[0013] Optionally, step A221 includes: A2211. For each augmented image, divide the augmented image into a plurality of non-overlapping image patches; A2212. Obtain the gradient distribution feature, color histogram feature, and texture feature of each image patch. The gradient distribution feature is obtained by calculating the gradient magnitude and direction of each image patch. The color histogram feature is obtained by counting the color distribution of each image patch in the HSV color space. The texture feature is obtained by calculating the gray-level co-occurrence matrix of each image patch; A2213. Calculate the sharpness confidence, contrast confidence, and noise confidence of each image patch according to the gradient distribution feature, color histogram feature, and texture feature, and perform normalization processing on the sharpness confidence, contrast confidence, and noise confidence based on the softmax function to obtain the sharpness weight, contrast weight, and noise weight of the weighted fusion algorithm; A2214. Calculate the corresponding image quality score based on the weighted fusion algorithm according to its corresponding sharpness score, contrast score, and noise score.

[0014] This technical solution is equivalent to a method of adaptively determining weights according to the local features of different regions in the augmented image. Therefore, this technical solution can make the image quality score obtained based on the weighted fusion algorithm more refined, thereby improving the accuracy of the image quality score, and further improving the training effect and positioning accuracy of the joint cavity positioning model.

[0015] Optionally, step A2211 includes: A22111. For each augmented image, calculate the gradient magnitude image of the augmented image; A22112. Select the pixel point with the largest gradient magnitude in the gradient magnitude image as the seed point; A22113. Expand from the seed point to the neighborhood, and merge the pixel points in the neighborhood whose difference in gradient magnitude from the seed point is less than the preset threshold into the current image patch until there are no pixel points in the gradient magnitude image whose difference in gradient magnitude from the seed point is less than the preset threshold; A22114. Analyze whether there are unpartitioned pixel points in the gradient magnitude image. If so, select the pixel point with the largest gradient magnitude among the unpartitioned pixel points as the new seed point, and return to step A22113. If not, execute step 2212.

[0016] Since this technical solution can merge pixel points with similar gradient magnitudes into the same image patch, that is, the gradient features of the pixel points within the same image patch are similar, this technical solution can effectively improve the uniformity of the image features within the image patch, thereby effectively further improving the accuracy of the image quality score.

[0017] Optionally, step A22111 includes: A221111. For each augmented image, convert the augmented image to the HSV color space and perform blurring on the V channel of the HSV color space to obtain an illumination intensity map; A221112. Calculate the illumination compensation coefficient corresponding to each pixel point in the augmented image according to the illumination intensity map; A221113. Adjust the RGB value corresponding to each pixel point in the augmented image according to the illumination compensation coefficient; A221114. Calculate the gradient magnitude image of the augmented image.

[0018] Since before calculating the gradient magnitude image of the augmented image, this technical solution first performs illumination equalization processing on the augmented image by obtaining the illumination intensity map corresponding to the augmented image and adjusting the RGB values of the pixel points in the augmented image based on the illumination intensity map, this technical solution can effectively reduce the influence of uneven illumination on the calculation of the gradient magnitude image, thereby effectively improving the accuracy of the gradient magnitude image.

[0019] Optionally, step A25 includes: A251. Obtain the high-frequency subband energy intensity feature corresponding to each augmented image based on wavelet transform, and obtain the loss weight corresponding to each augmented image according to the high-frequency subband energy intensity feature and the first preset conversion relationship; A252. Dynamically adjust the regularization strength of the preliminary localization model based on the adaptive regularization coefficient function according to the joint cavity shape diversity index of the augmented image dataset, and minimize the cross-entropy loss function of the preliminary localization model by using the Adam optimizer to multiply the cross-entropy loss value of each augmented image by the corresponding loss weight during the adjustment process to obtain the joint cavity localization model.

[0020] Optionally, step S2 includes: S21. Match the feature points of the current frame image with the feature points of the previous frame image to obtain a pair of matching feature points; S22. Estimate the homography matrix of the current frame image relative to the previous frame image based on the RANSAC algorithm. The homography matrix describes the transformation relationship of the arthroscopic view: S23. Decompose the homography matrix into a rotation matrix and a translation vector, and perform rotation and translation operations on the convolution kernel of the pre-trained joint cavity localization model according to the rotation matrix and the translation vector to make the joint cavity localization model adapt to the current view; S24. Input the enhanced image into the joint cavity localization model to obtain the joint cavity position information.

[0021] Optionally, the preprocessing further includes edge sharpening.

[0022] In a second aspect, the present application also provides an arthroscopic cavity positioning device based on a deep learning model for positioning the arthroscopic cavity of rare osteoarticular diseases. The arthroscopic cavity positioning device based on the deep learning model includes the following steps: An enhanced image acquisition module, configured to acquire arthroscopic images and preprocess the arthroscopic images to obtain enhanced images. The preprocessing includes contrast enhancement; An arthroscopic cavity position information acquisition module, configured to acquire arthroscopic cavity position information according to a pre-trained arthroscopic cavity positioning model and the enhanced images. The arthroscopic cavity positioning model is used to position the arthroscopic cavity in the enhanced images and output the arthroscopic cavity position; The pre-training of the arthroscopic cavity positioning model includes: A1. Training a pre-constructed deep learning model using a pre-collected general medical image dataset to obtain a preliminary positioning model; A2. Augmenting a pre-collected arthroscopic image dataset of rare osteoarticular diseases to obtain an augmented image dataset, and then dynamically adjusting the regularization strength of the preliminary positioning model based on an adaptive regularization coefficient function according to the arthroscopic cavity morphological diversity index of the augmented image dataset to obtain the arthroscopic cavity positioning model.

[0023] The arthroscopic cavity positioning device provided by the present application is equivalent to first enabling the deep learning model to learn the general features of medical images and the basic knowledge of arthroscopic cavity positioning by training the pre-constructed deep learning model using the pre-collected general medical image dataset, and then enabling the deep learning model to learn the specific features of the arthroscopic cavity of rare osteoarticular diseases on the premise of avoiding overfitting by dynamically adjusting the regularization strength of the preliminary positioning model based on the adaptive regularization coefficient function according to the arthroscopic cavity morphological diversity index of the augmented image dataset. Therefore, the arthroscopic cavity positioning model of the present application has good generalization ability and positioning accuracy, and the arthroscopic cavity positioning model can more accurately position the arthroscopic cavity of rare osteoarticular diseases, that is, the present application can effectively solve the problems that the arthroscopic cavity positioning model obtained based on the deep learning model cannot accurately position the arthroscopic cavity of rare osteoarticular diseases due to insufficient training data volume corresponding to rare osteoarticular diseases and the generalization ability and positioning accuracy of the arthroscopic cavity positioning model decrease due to the diversity of arthroscopic cavity morphology.

[0024] As can be seen from the above, a joint cavity localization method and device based on a deep learning model provided by this application is equivalent to first training a pre-constructed deep learning model through a pre-collected general medical image dataset so that the deep learning model learns the general features of medical images and the basic knowledge of joint cavity localization, and then dynamically adjusting the regularization strength of the preliminary localization model according to the joint cavity morphology diversity index of the augmented image dataset through an adaptive regularization coefficient function so that the deep learning model learns the unique features of the joint cavity of rare osteoarticular diseases on the premise of avoiding overfitting. Therefore, the joint cavity localization model of this application has good generalization ability and localization accuracy, and this joint cavity localization model can more accurately locate the joint cavity of rare osteoarticular diseases, that is, this application can effectively solve the problems that the joint cavity localization model obtained based on the deep learning model cannot accurately locate the joint cavity of rare osteoarticular diseases due to insufficient training data volume corresponding to rare osteoarticular diseases and the generalization ability and localization accuracy of the joint cavity localization model decrease due to the diversity of joint cavity morphology. Description of the Drawings

[0025] Figure 1 It is a flowchart of a joint cavity localization method based on a deep learning model provided by an embodiment of this application.

[0026] Figure 2 It is a schematic structural diagram of a joint cavity localization device based on a deep learning model provided by an embodiment of this application.

[0027] Reference Signs: 1. Enhanced Image Acquisition Module; 2. Joint Cavity Position Information Acquisition Module. Detailed Embodiments

[0028] Next, the technical solutions in the embodiments of this application will be clearly and completely described in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. The components of the embodiments of this application usually described and illustrated in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the drawings is not intended to limit the scope of the claimed application, but only represents the selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative efforts fall within the scope of protection of this application.

[0029] It should be noted that similar reference numerals and letters indicate similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of this application, terms such as "first" and "second" are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0030] In the first aspect, as Figure 1 shown, the present application provides a method for arthroscopic cavity localization based on a deep learning model, which is used to localize the arthroscopic cavity of rare osteoarticular diseases. The method for arthroscopic cavity localization based on a deep learning model includes the following steps: S1. Obtain arthroscopic images and preprocess the arthroscopic images to obtain enhanced images. The preprocessing includes contrast enhancement; S2. Obtain arthroscopic cavity position information according to a pre-trained arthroscopic cavity localization model and the enhanced images. The arthroscopic cavity localization model is used to localize the arthroscopic cavity in the enhanced images and output the position of the arthroscopic cavity; The pre-training of the arthroscopic cavity localization model includes: A1. Train a pre-constructed deep learning model using a pre-collected general medical image dataset to obtain a preliminary localization model; A2. Expand the pre-collected arthroscopic image dataset of rare osteoarticular diseases to obtain an expanded image dataset, and then dynamically adjust the regularization strength of the preliminary localization model based on the arthroscopic cavity morphological diversity index of the expanded image dataset according to an adaptive regularization coefficient function to obtain the arthroscopic cavity localization model.

[0031] The preprocessing operation in step S1 aims to improve the quality of arthroscopic images. The preprocessing includes contrast enhancement (such as histogram equalization or contrast-limited adaptive histogram equalization algorithm). The contrast enhancement is used to enhance the contrast of arthroscopic images, so that the boundary between the arthroscopic cavity and surrounding tissues is clearer and to provide higher-quality input data for the subsequent arthroscopic cavity localization model, thereby effectively improving the accuracy of arthroscopic cavity localization.

[0032] In step S2, the pre-trained arthroscopic cavity localization model is the core to realize the automatic localization of the arthroscopic cavity. The model can receive the preprocessed enhanced images as input and output the position information of the arthroscopic cavity (such as the bounding box or segmentation mask of the arthroscopic cavity in the enhanced images).

[0033] The pre-training process of the joint cavity localization model is divided into step A1 and step A2. In step A1, first, the pre-constructed deep learning model is trained using a pre-collected general medical image dataset to obtain a preliminary localization model. This general medical image dataset can include various medical images corresponding to common bone and joint diseases (such as CT images and MRI images, etc.). This embodiment can enable the deep learning model to learn the general features of medical images and the basic knowledge of joint cavity localization by training the pre-constructed deep learning model using the pre-collected general medical image dataset (equivalent to pre-training the deep learning model using general data). Even when the data volume of rare bone and joint diseases is insufficient, the model also has a certain joint cavity localization ability. Since rare bone and joint diseases have the characteristic of low incidence, that is, the arthroscopic image dataset of rare bone and joint diseases pre-collected contains a small amount of data, so in step A2, the arthroscopic image dataset of rare bone and joint diseases can be augmented using existing data augmentation methods to increase the amount of image data corresponding to rare bone and joint diseases that can be used for model training and increase the diversity of the image data corresponding to rare bone and joint diseases used for model training, so that the joint cavity localization model has higher generalization ability. Then, calculate the joint cavity morphology diversity index of the augmented image dataset. This index is used to evaluate the degree of change in the joint cavity morphology in the augmented dataset. The joint cavity morphology diversity index can be calculated based on the morphological features of the image. After that, based on the adaptive regularization coefficient function, the regularization strength of the preliminary localization model is dynamically adjusted according to the joint cavity morphology diversity index. Regularization is a commonly used technique to prevent model overfitting. This embodiment can limit the complexity of the model by adding a regularization term to the loss function of the deep learning model. The adaptive regularization coefficient function of this embodiment can dynamically adjust the regularization strength according to the joint cavity morphology diversity. Specifically, when the joint cavity morphology diversity is relatively high, this embodiment increases the regularization strength to prevent the model from overfitting to a specific morphology of the joint cavity, thereby effectively improving the generalization ability of the model; when the joint cavity morphology diversity is relatively low, this embodiment decreases the regularization strength to enable the model to better learn the unique features of the joint cavity of rare bone and joint diseases. Therefore, the joint cavity localization model obtained after training in step A2 can more accurately localize the joint cavity of rare bone and joint diseases. It should be understood that this embodiment is equivalent to first training the model using the general medical image dataset to obtain a general joint cavity localization model (equivalent to a general model applicable to common bone and joint diseases), and then training the general joint cavity localization model using the augmented arthroscopic image dataset of rare bone and joint diseases to obtain a joint cavity localization model applicable to rare bone and joint diseases.

[0034] A method for arthroscopic cavity localization based on a deep learning model provided by this application is equivalent to first training a pre-constructed deep learning model through a pre-collected general medical image dataset so that the deep learning model learns the general features of medical images and the basic knowledge of arthroscopic cavity localization, and then making the deep learning model learn the unique features of the arthroscopic cavity of rare osteoarticular diseases on the premise of avoiding overfitting by dynamically adjusting the regularization strength of the preliminary localization model based on an adaptive regularization coefficient function according to the arthroscopic cavity morphology diversity index of the augmented image dataset. Therefore, the arthroscopic cavity localization model of this application has good generalization ability and localization accuracy, and this arthroscopic cavity localization model can more accurately localize the arthroscopic cavity of rare osteoarticular diseases, that is, this application can effectively solve the problems that the arthroscopic cavity localization model obtained based on the deep learning model cannot accurately localize the arthroscopic cavity of rare osteoarticular diseases due to insufficient training data volume corresponding to rare osteoarticular diseases and the generalization ability and localization accuracy of the arthroscopic cavity localization model decrease due to the diversity of arthroscopic cavity morphology.

[0035] In some preferred embodiments, step A2 includes: A21. Augment the pre-collected arthroscopic image dataset of rare osteoarticular diseases by randomly rotating, scaling, and flipping each image in the pre-collected arthroscopic image dataset of rare osteoarticular diseases to obtain an augmented image dataset, where the augmented image dataset includes multiple augmented images; A22. Use the pre-constructed image quality assessment model to obtain the image quality scores of each augmented image to obtain multiple image quality scores; A23. Normalize all image quality scores to the interval [0, 1] to obtain the quality weights corresponding to each augmented image, and obtain the arthroscopic cavity morphology feature vectors corresponding to each augmented image based on the local binary pattern histogram corresponding to the augmented image; A24. Perform weighted average calculation according to all quality weights and all arthroscopic cavity morphology feature vectors to obtain the arthroscopic cavity morphology diversity index; A25. Dynamically adjust the regularization strength of the preliminary localization model based on the adaptive regularization coefficient function according to the arthroscopic cavity morphology diversity index of the augmented image dataset to obtain the arthroscopic cavity localization model.

[0036] Step A21 can implement data augmentation for the pre - collected arthroscopic image dataset of rare osteoarticular diseases by randomly rotating, scaling, and flipping each image in the dataset, thereby increasing the diversity of training data. In step A22, the image quality assessment model can be a pre - trained convolutional neural network configured to receive image inputs and output quality scores. Step A23 can linearly map all quality scores to the [0, 1] interval by using the existing min - max normalization method. The quality weights in step A23 are directly determined by the normalized quality scores. The arthroscopic cavity shape feature vector is obtained by calculating the local binary pattern histogram of each augmented image. The local binary pattern histogram can capture the local texture information of the image and serve as a representation of the arthroscopic cavity shape. The arthroscopic cavity shape diversity index in step A24 is calculated by weighted average. This index can comprehensively reflect the image quality and the arthroscopic cavity shape distribution of the augmented dataset. In the weighted average, each arthroscopic cavity shape feature vector is weighted by its corresponding quality weight. The higher the quality weight, the greater the contribution of the corresponding image to the calculation of the diversity index. Specifically, when the arthroscopic cavity shape diversity index is relatively high, it indicates that the arthroscopic cavity shapes in the dataset vary greatly. At this time, increasing the regularization strength can improve the generalization ability of the model and prevent the model from overfitting to arthroscopic cavities of specific shapes. On the contrary, when the arthroscopic cavity shape diversity index is relatively low, the regularization strength can be appropriately reduced to enable the model to better learn the subtle features in the data. Since this embodiment is equivalent to introducing image quality assessment and arthroscopic cavity shape diversity index during the data augmentation process and using the arthroscopic cavity shape diversity index to dynamically adjust the regularization strength of model training, this embodiment can effectively solve the problem of low accuracy of arthroscopic cavity shape diversity caused by uneven quality of augmented images, thereby effectively improving the generalization ability and localization accuracy of the arthroscopic cavity localization model.

[0037] In some preferred embodiments, the image quality assessment model is used to extract the sharpness feature, contrast feature, and noise level feature of each augmented image and output the corresponding sharpness score, contrast score, and noise score. Step A22 includes: A221. For each augmented image, calculate its corresponding image quality score based on the weighted fusion algorithm according to its corresponding sharpness score, contrast score, and noise score.

[0038] Among them, the image quality assessment model of this embodiment is configured to analyze three key quality dimensions of the augmented image: sharpness, contrast, and noise level. This embodiment can extract sharpness features based on image gradient information. For example, calculate the Laplacian operator response or gradient magnitude of the image. The higher the response value or magnitude, the sharper the image. This embodiment can extract contrast features based on the gray distribution of the image. For example, calculate the standard deviation or information entropy of the image gray values. The larger the standard deviation or the higher the information entropy, the higher the contrast of the image. This embodiment can extract noise level features by analyzing the frequency domain characteristics of the image or the fluctuations of local pixel values. For example, calculate the energy of the high-frequency components of the image or the variance of pixel values in a local area. The lower the energy or the smaller the variance, the lower the noise level of the image. After obtaining the sharpness features, contrast features, and noise level features, the image quality assessment model converts these features into corresponding scores. The conversion of scores can be achieved by means of linear mapping, non-linear mapping, or look-up table, etc. In step A221, a weighted fusion algorithm is used to integrate the sharpness score, contrast score, and noise score to obtain the final image quality score. The specific implementation method of the weighted fusion algorithm can be various. For example, a weight can be assigned to each score, and then the weighted average is calculated as the image quality score. This weight can be set according to experience or obtained through learning. Since this embodiment can first use the image quality assessment model to independently evaluate its sharpness, contrast, and noise level, and output corresponding scores, and then calculate the image quality score corresponding to each augmented image based on the sharpness score, contrast score, and noise score by means of the weighted fusion algorithm, it can avoid the situation that the image quality score cannot accurately reflect the overall quality of the image due to evaluating the quality of the augmented image based on only one dimension, thereby effectively improving the training efficiency and positioning accuracy of the preliminary positioning model, and further effectively improving the generalization ability and positioning accuracy of the joint cavity positioning model. Specifically, the formula of the weighted fusion algorithm of this embodiment is: ; where Q represents the image quality score, w c represents the sharpness weight, C represents the sharpness score, w o represents the contrast weight, O represents the contrast score, w n represents the noise weight, and N represents the noise score.

[0039] In some preferred embodiments, step A221 includes: A2211. For each augmented image, divide the augmented image into multiple non-overlapping image blocks; A2212. Obtain the gradient distribution features, color histogram features, and texture features of each image patch. The gradient distribution features are obtained by calculating the gradient magnitude and direction of each image patch. The color histogram features are obtained by statistically analyzing the color distribution of each image patch in the HSV color space. The texture features are obtained by calculating the gray-level co-occurrence matrix of each image patch. A2213. Calculate the sharpness confidence, contrast confidence, and noise confidence of each image patch based on the gradient distribution features, color histogram features, and texture features, and perform normalization processing on the sharpness confidence, contrast confidence, and noise confidence based on the Aoftmax function to obtain the sharpness weight, contrast weight, and noise weight of the weighted fusion algorithm. A2214. Calculate the corresponding image quality score based on the weighted fusion algorithm according to its corresponding sharpness score, contrast score, and noise score.

[0040] Step A2211 can divide the augmented image into multiple non-overlapping image patches by using a grid of a fixed size to divide the augmented image. In this embodiment, the gradient distribution features can be obtained by calculating the gradient magnitude and direction of the image patches using the Sobel operator or Canny operator. The color histogram features can be obtained by statistically analyzing the pixel distribution of each color channel of the image patches in the HSV color space. The texture features can be obtained by calculating statistics such as the energy, contrast, correlation, and entropy of the gray-level co-occurrence matrix, which can characterize the texture information of the image patches. The sharpness confidence, contrast confidence, and noise confidence in step A2213 can be obtained through a pre-trained classifier or regression model, which uses the gradient distribution features, color histogram features, and texture features extracted in step A2212 as inputs and outputs the corresponding confidence scores. The Aoftmax function is used to normalize the confidence scores to allocate more reasonable weights for the sharpness, contrast, and noise levels, so that high-quality features (features with higher confidence) obtain higher weights. In step A2214, the weighted fusion algorithm can adopt a linear weighting or non-linear weighting method to perform weighted fusion on the sharpness score, contrast score, and noise score with the corresponding weights obtained in step A2213 to obtain the final image quality score. This embodiment is equivalent to a method of adaptively determining weights according to the local features of different regions in the augmented image. Therefore, this embodiment can make the image quality score obtained based on the weighted fusion algorithm more refined, thereby improving the accuracy of the image quality score and further enhancing the training effect and positioning accuracy of the joint cavity positioning model.

[0041] In some preferred embodiments, step A2211 includes: A22111. For each augmented image, calculate the gradient magnitude image of the augmented image. A22112. Select the pixel point with the largest gradient magnitude in the gradient magnitude image as the seed point; A22113. Expand from the seed point to the neighborhood, and merge the pixel points in the neighborhood whose difference in gradient magnitude from the seed point is less than a preset threshold into the current image block until there are no pixel points in the gradient magnitude image whose difference in gradient magnitude from the seed point is less than the preset threshold; A22114. Analyze whether there are unpartitioned pixel points in the gradient magnitude image. If so, select the pixel point with the largest gradient magnitude among the unpartitioned pixel points as the new seed point, and return to step A22113. If not, execute step 2212.

[0042] In step A22111, edge detection operators such as Sobel operator, Prewitt operator, or Canny operator can be used to calculate the gradient magnitude image. These operators can effectively detect the sudden change of pixel gray values in the image, so as to obtain the gradient magnitude image. That is, step A22111 simply highlights the edge information of the image by calculating and expanding the gradient magnitude image of the image. Since the pixel points with the largest gradient magnitude are usually located in the regions with rich edge information in the image, selecting such points as the seed points can ensure that the image blocks grow from the edge regions of the image. Therefore, step A22112 selects the pixel point with the largest gradient magnitude in the gradient magnitude image as the seed point. The expansion from the seed point to the neighborhood in step A22113 can be realized by eight-neighborhood or four-neighborhood growth. The setting of the preset threshold in this embodiment needs to be adjusted according to the actual application scenario and image characteristics to control the size and shape of the image block. It should be understood that this embodiment is equivalent to iteratively selecting the pixel point with the largest gradient magnitude as the seed point and merging the pixel points with similar gradient magnitudes into the same image block by means of region growth centered on the seed point. Since this embodiment can merge the pixel points with similar gradient magnitudes into the same image block, that is, the gradient features of the pixel points within the same image block are similar, this embodiment can effectively improve the uniformity of the image features within the image block, thereby effectively further improving the accuracy of the image quality score.

[0043] In some preferred embodiments, step A22111 includes: A221111. For each expanded image, convert the expanded image to the HSV color space, and perform blurring processing on the V channel of the HSV color space to obtain the illumination intensity map; A221112. Calculate the illumination compensation coefficient corresponding to each pixel point in the expanded image according to the illumination intensity map; A221113. Adjust the RGB value corresponding to each pixel point in the expanded image according to the illumination compensation coefficient; A221114. Calculate the gradient magnitude image of the enlarged image.

[0044] Step A22111 aims to preliminarily process the image illumination. By converting the enlarged image to the HSV color space, step A22111 separates the color information and brightness information of the image, facilitating the separate processing of the brightness information. The V channel represents the brightness information, and this channel is blurred, for example, using Gaussian blur, with the aim of smoothing the illumination intensity map and reducing noise interference, thereby obtaining the overall illumination intensity distribution of the image. Step A221112 is used to calculate the illumination compensation coefficient. Specifically, the illumination compensation coefficient for each pixel point is calculated by the ratio of the average brightness value of the illumination intensity map to the brightness value of each pixel point. That is, the calculation formula for the illumination compensation coefficient is: ; where k(x,y) represents the illumination compensation coefficient of the pixel point with coordinates (x,y) in the illumination intensity map, V mean represents the average brightness value of the illumination intensity map, and V(x,y) represents the brightness value of the pixel point with coordinates (x,y) in the illumination intensity map. The illumination compensation coefficient of a pixel point is negatively correlated with its brightness value. A high brightness value results in a small compensation coefficient, while a low brightness value leads to a large illumination compensation coefficient. It should be understood that the number of pixel points in the illumination intensity map is equal to the number of pixel points in the enlarged image, that is, each pixel point in the illumination intensity map corresponds to a pixel point in the enlarged image. Step A221113 is used to adjust the RGB values of the image according to the illumination compensation coefficient to correct the image brightness difference caused by uneven illumination. The adjusted RGB values are obtained by multiplying the original RGB values by the illumination compensation coefficient. The RGB value adjustment formula is: ; where, , and represent the RGB values of the pixel point with coordinates (x,y) in the illumination intensity map after adjustment, and R(x,y), G(x,y), and B(x,y) represent the RGB values of the pixel point with coordinates (x,y) in the illumination intensity map before adjustment. Step A221114 is used to calculate the gradient magnitude image of the enlarged image after illumination compensation, providing a basis for subsequent image block division. Since before calculating the gradient magnitude image of the enlarged image, this embodiment first performs illumination equalization processing on the enlarged image by obtaining the illumination intensity map corresponding to the enlarged image and adjusting the RGB values of the pixel points in the enlarged image based on the illumination intensity map, this embodiment can effectively reduce the influence of uneven illumination on the calculation of the gradient magnitude image, thereby effectively improving the accuracy of the gradient magnitude image.

[0045] In some preferred embodiments, step A25 includes: A251. Obtain the high-frequency subband energy intensity feature corresponding to each augmented image based on wavelet transform, and obtain the loss weight corresponding to each augmented image according to the high-frequency subband energy intensity feature and the first preset conversion relationship; A252. Dynamically adjust the regularization strength of the preliminary localization model based on the joint cavity morphology diversity index of the augmented image dataset according to the adaptive regularization coefficient function, and use the Adam optimizer during the adjustment process to minimize the cross-entropy loss function of the preliminary localization model by multiplying the cross-entropy loss value of each augmented image by the corresponding loss weight, so as to obtain the joint cavity localization model.

[0046] The wavelet transform in step A251 can be Haar wavelet transform, Daubechies wavelet transform or Symlets wavelet transform. The wavelet transform in step A251 can decompose the augmented image into multiple subbands with different frequencies, and the high-frequency subband energy intensity feature is extracted from these subbands. Since noise is mainly concentrated in the high-frequency part of the image, the high-frequency subband energy intensity feature obtained in step A251 can reflect the noise intensity and noise level of the augmented image. The first preset conversion relationship in this embodiment is preferably a pre-constructed mapping relationship between the high-frequency subband energy intensity feature and the loss weight. Specifically, the high-frequency subband energy intensity feature is negatively correlated with the loss weight, that is, this embodiment is equivalent to assigning a lower loss weight to an image with a large noise intensity and a higher loss weight to an image with a small noise intensity. Since in the process of adjusting the regularization strength, this embodiment uses the Adam optimizer to minimize the cross-entropy loss function of the preliminary localization model by multiplying the cross-entropy loss value of each augmented image by the corresponding loss weight, so that the model pays more attention to low-noise images and less attention to the influence of high-noise images during model training. Therefore, this embodiment can effectively improve the robustness and localization accuracy of the joint cavity localization model.

[0047] In some specific embodiments, step A251 can further include: First, perform wavelet decomposition on each augmented image to obtain subband coefficients of multiple frequency bands; then, calculate the energy of each high-frequency subband, and the energy calculation method can be to calculate the sum of the squares of the subband coefficients; then, perform weighted summation on the energies of all high-frequency subbands to obtain the high-frequency subband energy intensity feature; finally, according to the high-frequency subband energy intensity feature, through a preset conversion relationship, such as a linear function, a non-linear function or a look-up table, obtain the loss weight corresponding to each augmented image. As a preferred embodiment, the first preset conversion relationship can be set as a Sigmoid function, the input of which is the normalized high-frequency subband energy intensity feature, and the output is the loss weight in the interval [0,1]. Thus, the effect that the higher the high-frequency subband energy intensity feature, the lower the loss weight can be achieved.

[0048] In some preferred embodiments, step S2 includes: S21. Matching the feature points of the current frame image with the feature points of the previous frame image to obtain pairs of matching feature points; S22. Estimating the homography matrix of the current frame image relative to the previous frame image based on the RANSAC algorithm. The homography matrix describes the transformation relationship of the arthroscopic view: S23. Decomposing the homography matrix into a rotation matrix and a translation vector, and performing rotation and translation operations on the convolution kernels of the pre-trained arthroscopic cavity localization model according to the rotation matrix and the translation vector, so that the arthroscopic cavity localization model adapts to the current view; S24. Inputting the enhanced image into the arthroscopic cavity localization model to obtain arthroscopic cavity position information.

[0049] Step S21 can implement feature point matching by using SIFT, SURF or ORB feature matching algorithms. These algorithms can extract distinctive feature points in the image and establish the correspondence of feature points between the current frame image and the previous frame image through descriptor matching to obtain pairs of matching feature points. The RANSAC algorithm in step S2 is used to robustly estimate the homography matrix from the pairs of matching feature points. The RANSAC algorithm effectively excludes the interference of incorrect matching point pairs through random sampling and iterative optimization, so as to obtain an accurate homography matrix. The homography matrix can describe the mapping relationship of pixel points between two frame images and reflects the transformation of the arthroscopic view. Step S23 can decompose the homography matrix by using the singular value decomposition method to obtain a rotation matrix and a translation vector. The rotation matrix and the translation vector respectively represent the rotation and displacement changes of the arthroscopic view. The convolution kernels of the pre-trained arthroscopic cavity localization model are the key components in the model for extracting image features. This embodiment can make the convolution kernels adapt to the image features under the current view by applying the rotation matrix and the translation vector to the convolution kernels, so as to ensure that the arthroscopic cavity localization model can still effectively extract features when the view changes. This embodiment is equivalent to introducing a view adaptability adjustment mechanism in the arthroscopic cavity localization process. Therefore, this embodiment can effectively solve the problem of decreased localization accuracy caused by the change of the arthroscopic view, and thus effectively improve the robustness and accuracy of the arthroscopic cavity localization model.

[0050] In some specific embodiments, in step S21, the ORB feature matching algorithm is used to perform feature point matching to quickly and robustly find the corresponding relationship of feature points between the current frame image and the previous frame image. In step S22, the specific number of iterations of the RANSAC algorithm is set to 1000 times to ensure the accuracy and stability of the homography matrix estimation. In step S23, the homography matrix is decomposed into a rotation matrix and a translation vector by the singular value decomposition method, and the obtained rotation matrix and translation vector are directly applied to the convolution kernels of all convolutional layers of the joint cavity localization model. This embodiment can effectively realize the adaptive adjustment of the arthroscopic view to ensure that the joint cavity localization model can still accurately locate the joint cavity when the arthroscopic view changes, thereby effectively improving the practical value of the joint cavity localization model.

[0051] In some preferred embodiments, the preprocessing further includes edge sharpening. Edge sharpening is a technique used to enhance the edge contrast in an image to make the boundaries between different regions clearer. This embodiment can make the boundary of the joint cavity clearer and reduce the degree of blurriness by edge sharpening the arthroscopic image, so that the joint cavity localization model can more accurately identify and locate the joint cavity.

[0052] As can be seen from the above, a joint cavity localization method based on a deep learning model provided by the present application is equivalent to first training a pre-constructed deep learning model through a pre-collected general medical image dataset to enable the deep learning model to learn the general features of medical images and the basic knowledge of joint cavity localization, and then dynamically adjusting the regularization strength of the preliminary localization model according to the joint cavity morphology diversity index of the augmented image dataset based on an adaptive regularization coefficient function to enable the deep learning model to learn the unique features of the joint cavity of rare osteoarticular diseases on the premise of avoiding overfitting. Therefore, the joint cavity localization model of the present application has good generalization ability and localization accuracy, and the joint cavity localization model can more accurately locate the joint cavity of rare osteoarticular diseases, that is, the present application can effectively solve the problems that the joint cavity localization model obtained based on the deep learning model cannot accurately locate the joint cavity of rare osteoarticular diseases due to insufficient training data corresponding to rare osteoarticular diseases and the generalization ability and localization accuracy of the joint cavity localization model decrease due to the diversity of joint cavity morphology.

[0053] Second aspect, as Figure 2 shown, the present application also provides a joint cavity localization device based on a deep learning model for locating the joint cavity of rare osteoarticular diseases. The joint cavity localization device based on the deep learning model includes the following steps: An enhanced image acquisition module 1, configured to acquire an arthroscopic image and perform preprocessing on the arthroscopic image to obtain an enhanced image, and the preprocessing includes contrast enhancement; The joint cavity position information acquisition module 2 obtains joint cavity position information according to a pre-trained joint cavity positioning model and an enhanced image. The joint cavity positioning model is used to locate the joint cavity in the enhanced image and output the joint cavity position. The pre-training of the joint cavity positioning model includes: A1. Training a pre-constructed deep learning model using a pre-collected general medical image dataset to obtain a preliminary positioning model. A2. Augmenting a pre-collected arthroscopic image dataset of rare osteoarticular diseases to obtain an augmented image dataset, and then dynamically adjusting the regularization strength of the preliminary positioning model based on an adaptive regularization coefficient function according to the joint cavity morphological diversity index of the augmented image dataset to obtain the joint cavity positioning model.

[0054] A joint cavity positioning device based on a deep learning model provided by the present application includes an enhanced image acquisition module 1 and a joint cavity position information acquisition module 2. The joint cavity positioning based on a deep learning model provided by this embodiment is used to execute a joint cavity positioning method based on a deep learning model provided in the above first aspect. The principle of the joint cavity positioning device based on a deep learning model provided by this embodiment is the same as the principle of the joint cavity positioning method based on a deep learning model provided in the above first aspect, and will not be elaborated here in detail.

[0055] As can be seen from the above, a joint cavity positioning method and device based on a deep learning model provided by the present application is equivalent to first enabling the deep learning model to learn the general features of medical images and the basic knowledge of joint cavity positioning by training a pre-constructed deep learning model using a pre-collected general medical image dataset, and then enabling the deep learning model to learn the unique features of the joint cavity of rare osteoarticular diseases on the premise of avoiding overfitting by dynamically adjusting the regularization strength of the preliminary positioning model based on an adaptive regularization coefficient function according to the joint cavity morphological diversity index of the augmented image dataset. Therefore, the joint cavity positioning model of the present application has good generalization ability and positioning accuracy, and the joint cavity positioning model can more accurately locate the joint cavity of rare osteoarticular diseases, that is, the present application can effectively solve the problems that the joint cavity positioning model obtained based on the deep learning model cannot accurately locate the joint cavity of rare osteoarticular diseases due to insufficient training data volume corresponding to rare osteoarticular diseases and the generalization ability and positioning accuracy of the joint cavity positioning model decrease due to the diversity of joint cavity morphology.

[0056] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the above division of units is only a logical function division. In actual implementation, there may be other division methods. For another example, multiple units or components can be combined or integrated into another robot, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection between each other can be through some communication interfaces. The indirect coupling or communication connection of devices or units can be in electrical, mechanical or other forms.

[0057] In addition, in each embodiment of the present application, the functional modules can be integrated together to form an independent part, or each module can exist separately, or two or more modules can be integrated to form an independent part.

[0058] In this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0059] The above are only the embodiments of the present application and are not used to limit the protection scope of the present application. For those skilled in the art, the present application can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for joint cavity localization based on a deep learning model, characterized in that, For positioning the joint cavity of rare osteoarticular diseases, the joint cavity positioning method based on a deep learning model includes the following steps: S1. Obtain arthroscopic images and preprocess the arthroscopic images to obtain enhanced images. The preprocessing includes contrast enhancement; S2. Obtain joint cavity position information according to a pre-trained joint cavity positioning model and the enhanced images. The joint cavity positioning model is used to position the joint cavity in the enhanced images and output the joint cavity position; The pre-training of the joint cavity positioning model includes: A1. Use a pre-collected general medical image dataset to train a pre-constructed deep learning model to obtain a preliminary positioning model; A2. Expand the pre-collected arthroscopic image dataset of rare osteoarticular diseases to obtain an expanded image dataset, and then dynamically adjust the regularization strength of the preliminary positioning model based on an adaptive regularization coefficient function according to the joint cavity morphological diversity index of the expanded image dataset to obtain a joint cavity positioning model.

2. The arthroscopic cavity positioning method based on a deep learning model according to claim 1, wherein Step A2 includes: A21. Expand the pre-collected arthroscopic image dataset of rare osteoarticular diseases by randomly rotating, scaling, and flipping each image in the pre-collected arthroscopic image dataset of rare osteoarticular diseases to obtain an expanded image dataset. The expanded image dataset includes multiple expanded images; A22. Use a pre-constructed image quality assessment model to obtain the image quality scores of each of the expanded images to obtain multiple image quality scores; A23. Normalize all the image quality scores to the interval [0, 1] to obtain the quality weights corresponding to each of the expanded images, and obtain the joint cavity morphological feature vectors corresponding to each of the expanded images based on the local binary pattern histogram corresponding to the expanded images; A24. Perform weighted average calculation according to all the quality weights and all the joint cavity morphological feature vectors to obtain the joint cavity morphological diversity index; A25. Dynamically adjust the regularization strength of the preliminary positioning model based on an adaptive regularization coefficient function according to the joint cavity morphological diversity index of the expanded image dataset to obtain a joint cavity positioning model.

3. The joint cavity localization method based on a deep learning model according to claim 2, wherein The image quality assessment model is used to extract the clarity feature, contrast feature, and noise level feature of each of the expanded images and output the corresponding clarity score, contrast score, and noise score. Step A22 includes: A221. For each of the expanded images, calculate its corresponding image quality score based on a weighted fusion algorithm according to its corresponding clarity score, contrast score, and noise score.

4. The joint cavity localization method based on a deep learning model according to claim 3, characterized in that Step A221 includes: A2211. For each of the expanded images, divide the expanded image into multiple non-overlapping image blocks; A2212. Obtain the gradient distribution features, color histogram features, and texture features of each of the image patches. The gradient distribution features are obtained by calculating the gradient magnitude and direction of each image patch. The color histogram features are obtained by statistically analyzing the color distribution of each image patch in the HSV color space. The texture features are obtained by calculating the gray-level co-occurrence matrix of each image patch. A2213. Calculate the clarity confidence, contrast confidence, and noise confidence of each image patch based on the gradient distribution features, color histogram features, and texture features, and perform normalization processing on the clarity confidence, contrast confidence, and noise confidence based on the Aoftmax function to obtain the clarity weight, contrast weight, and noise weight of the weighted fusion algorithm. A2214. Calculate the corresponding image quality score based on the weighted fusion algorithm according to its corresponding clarity score, contrast score, and noise score.

5. The joint cavity localization method based on a deep learning model according to claim 4, wherein Step A2211 includes: A22111. For each of the augmented images, calculate the gradient magnitude image of the augmented image. A22112. Select the pixel point with the largest gradient magnitude in the gradient magnitude image as the seed point. A22113. Expand from the seed point to the neighborhood, and merge the pixel points in the neighborhood whose difference in gradient magnitude from the seed point is less than a preset threshold into the current image patch until there are no pixel points in the gradient magnitude image whose difference in gradient magnitude from the seed point is less than the preset threshold. A22114. Analyze whether there are unpartitioned pixel points in the gradient magnitude image. If so, select the pixel point with the largest gradient magnitude among the unpartitioned pixel points as the new seed point and return to step A22113. If not, execute step 2212.

6. The joint cavity localization method based on a deep learning model according to claim 5, characterized in that, Step A22111 includes: A221111. For each of the augmented images, convert the augmented image to the HSV color space and perform blurring processing on the V channel of the HSV color space to obtain the illumination intensity map. A221112. Calculate the illumination compensation coefficient corresponding to each pixel point in the augmented image according to the illumination intensity map. A221113. Adjust the RGB values corresponding to each pixel point in the augmented image according to the illumination compensation coefficient. A221114. Calculate the gradient magnitude image of the augmented image.

7. The arthroscopic cavity positioning method based on a deep learning model according to claim 2, characterized in that Step A25 includes: A251. Obtain the high-frequency subband energy intensity features corresponding to each of the augmented images based on wavelet transform, and obtain the loss weights corresponding to each of the augmented images according to the high-frequency subband energy intensity features and the first preset conversion relationship. A252. Dynamically adjust the regularization strength of the preliminary localization model based on the adaptive regularization coefficient function according to the joint cavity shape diversity index of the augmented image dataset, and minimize the cross-entropy loss function of the preliminary localization model by multiplying the cross-entropy loss value of each augmented image by the corresponding loss weight using the Adam optimizer during the adjustment process to obtain the joint cavity localization model.

8. A method for arthroscopic cavity localization based on a deep learning model according to claim 1, characterized in that, Step S2 includes: S21. Match the feature points of the current frame image with those of the previous frame image to obtain the matched feature point pairs; S22. Estimate the homography matrix of the current frame image relative to the previous frame image based on the RANSAC algorithm, where the homography matrix describes the transformation relationship of the arthroscopic view: S23. Decompose the homography matrix into a rotation matrix and a translation vector, and perform rotation and translation operations on the convolution kernels of the pre-trained joint cavity localization model according to the rotation matrix and the translation vector, so that the joint cavity localization model adapts to the current view; S24. Input the enhanced image into the joint cavity localization model to obtain the joint cavity position information.

9. The method for joint cavity positioning based on a deep learning model according to claim 1, wherein, The preprocessing further includes edge sharpening.

10. An arthroscopic cavity positioning device based on a deep learning model, characterized in that, For localizing the joint cavity of rare osteoarticular diseases, the joint cavity localization device based on the deep learning model includes the following steps: An enhanced image acquisition module, configured to acquire an arthroscopic image and preprocess the arthroscopic image to obtain an enhanced image, where the preprocessing includes contrast enhancement; A joint cavity position information acquisition module, configured to obtain joint cavity position information according to a pre-trained joint cavity localization model and the enhanced image, where the joint cavity localization model is used to localize the joint cavity in the enhanced image and output the joint cavity position; The pre-training of the joint cavity localization model includes: A1. Train a pre-constructed deep learning model using a pre-collected general medical image dataset to obtain a preliminary localization model; A2. Expand the pre-collected arthroscopic image dataset of rare osteoarticular diseases to obtain an expanded image dataset, and then dynamically adjust the regularization strength of the preliminary localization model based on the joint cavity morphology diversity index of the expanded image dataset according to the adaptive regularization coefficient function to obtain the joint cavity localization model.

Images