A multi-level classification method for knee cartilage injury based on deep learning multi-modal MRI
By employing a deep learning-based multimodal MRI method, combined with dragonfly optimization and region similarity transformation function for image enhancement, and constructing a multi-level classification model, the accuracy and efficiency issues of knee cartilage injury detection were resolved, achieving high-precision classification of cartilage injury levels.
Patent Information
- Application Number
- CN202211349434.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-31
- Publication Date
- 2025-12-16
- Estimated Expiration
- 2042-10-31
AI Technical Summary
Existing technologies for detecting knee cartilage damage are time-consuming, labor-intensive, and lack accuracy. Traditional machine learning methods are not accurate enough in classifying the degree of knee cartilage damage, and simple transfer learning cannot achieve a high level of accuracy.
A deep learning-based multimodal MRI method was adopted, which combined dragonfly optimization and region similarity transformation function for image enhancement. A classification model with convolutional layer, fully connected layer, max pooling layer and softmax classifier was constructed. Multi-level classification of knee joint cartilage injury was achieved through multi-feature fusion.
It improves the accuracy and robustness of knee cartilage injury classification, enabling rapid and accurate grading based on image features, conforming to the standards of the International Association for Cartilage Repair, and enhancing the model's generalization ability.
Smart Images
Figure CN115601346B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of knee cartilage injury grade classification, and in particular to a multi-level classification method for knee cartilage injury based on multi-modal MRI and deep learning. BACKGROUND
[0002] Knee cartilage injury is an important cause of knee pain, mainly because the knee cartilage is thin, and the knee cartilage exists in the complex structure of the knee joint cavity, and there has been great difficulty in detecting the knee cartilage. Since the treatment measures for different grades of knee cartilage injury are different in clinical practice, and the prognosis of patients is also significantly different, accurate classification of knee cartilage injury is of great significance for the development of clinical treatment.
[0003] In clinical practice, the MRI examination results of patients with knee cartilage injury are mostly reviewed by traditional manual reading, which is time-consuming and laborious, and the accuracy and reliability are not ideal. With the rise of artificial intelligence related technology, machine learning models are widely used in the classification of various diseases, such as COVID-19, acute leukemia, diabetic retinopathy, and brain tumors. Traditional machine learning classification methods require a large number of samples, and the image feature values and probability values need to be set manually, which has the disadvantages of unstable results and high false positive rate, and the feasibility is low in practical application. For the classification of knee cartilage injury level, simple transfer learning cannot achieve high classification accuracy. SUMMARY
[0004] The present application provides a multi-level classification method for knee cartilage injury based on multi-modal MRI and deep learning to overcome the above technical problems.
[0005] A multi-level classification method for knee cartilage injury based on multi-modal MRI and deep learning, comprising,
[0006] S1, obtaining an MRI data set of knee cartilage, the MRI data set comprising a set of sagittal images of PdWI-fs, a set of sagittal images of T1WI, and a set of sagittal images of T2 mapping, dividing the MRI data set into five levels, the five levels being normal, first grade, second grade, third grade, and fourth grade;
[0007] S2, dividing the MRI data set into a training set and a test set, performing image enhancement on the images in the training set based on dragonfly optimization and regional similarity transformation function, the image enhancement being to enhance the contrast of the images, and obtaining the boundary features of the knee cartilage, cartilage edema, and subchondral bone in the images;
[0008] S3, a knee cartilage injury classification model based on deep learning is constructed, the model includes a convolution layer, a full connection layer, a maximum pooling layer and a Softmax classifier, the convolution layer is used to extract features in the image, the features refer to cartilage features, cartilage edema features and subchondral bone features, the pooling layer is used to select the features extracted in the convolution layer, the full connection layer is used to map the feature space obtained by the convolution layer and the pooling layer to the sample label space, and the Softmax classifier is used to classify the image, wherein the cartilage features, the cartilage edema features and the subchondral bone features are fused in each full connection layer;
[0009] S4, the initial value of the training parameter in the knee cartilage injury classification model is set and the iteration termination condition is set, the knee cartilage injury classification model is trained according to the training set, when the iteration termination condition is met, the optimal knee cartilage injury classification model is obtained, the optimal knee cartilage injury classification model is tested according to the test set and the prediction value is obtained, and the optimal knee cartilage injury classification model is evaluated through the confusion matrix.
[0010] Preferably, the image enhancement of the image in the training set based on the dragonfly optimization and the regional similarity transformation function comprises,
[0011] S21, for each image in the training set, each pixel point in the image is regarded as a dragonfly, an adaptability function is constructed according to the entropy value of the image, the total number of image edge intensities and the number of image edges, the adaptability function is formula (1), the adaptability value of each dragonfly is calculated in sequence according to the adaptability function, the maximum adaptability value and the corresponding pixel point set are obtained,
[0012]
[0013] Wherein, H(I(x, y)) is the entropy value of the image, C edge (x, y)) is the number of edge pixels of the image, G(I(x, y)) is the sobel value of the image, M*N is the image size, I(x, y) is the input image, (x, y) is the pixel position of the image,
[0014] S22, the pixels in the pixel point set obtained in S21 are transformed according to formula (2), and a transformed image is obtained,
[0015]
[0016] Wherein, I(x, y) and O(x, y) are the input image and the output image respectively, is the standard deviation value in the smooth window, Max is the maximum gray value in the image, a, b, c, d are constants, and the ranges are a∈[0, 3], b∈[0, 5], c∈[0, 1], d∈[0, 1], respectively, S(x, y) is the similarity value of the pixel at the pixel point (x, y) in the pixel point set and the surrounding z pixel points, z represents the number of pixel points, and the surrounding is the pixel point adjacent to the center with the pixel point (x, y) as the center, and the similarity value S(x, y) of the pixel point (x, y) in the pixel point set is calculated according to formula (3),
[0017]
[0018] wherein I pq is the distance between pixels p and q, and ω is a normalization coefficient.
[0019] Preferably, the fusing of the cartilage feature, the cartilage edema feature and the subchondral bone feature in each full connection layer comprises feature fusion according to formula (4),
[0020]
[0021] wherein, is the fused feature, σ is a convolution kernel, and * is convolution; X1, X2,..., X c , Y1, Y2,..., Y c and Z1, Z2,..., Z c respectively represent the knee cartilage feature, the cartilage edema feature and the subchondral bone feature, and C represents the number of features.
[0022] Preferably, the knee cartilage injury classification model further comprises a batch normalization layer, and the batch normalization layer is used for batch normalization processing of data obtained by the full connection layer in the knee cartilage injury classification model according to formula (5),
[0023]
[0024] wherein N represents the number of channels shared by each feature mapping internal vector, i represents the number of convolution kernels, j=max(0, i-n / 2), and ε, n, α, β are all hyperparameters,
[0025] and the batch normalization processed data is transmitted to the first maximum pooling layer.
[0026] Preferably, the setting of the iteration termination condition comprises but is not limited to setting a fixed number of iterations.
[0027] The application discloses a multi-level classification system for knee cartilage injury based on deep learning of multi-modal MRI, which comprises a data acquisition module, an image enhancement module, a model construction module, a model training and test evaluation module and a model application module.
[0028] The data acquisition module is used for acquiring an MRI data set of a knee cartilage, wherein the MRI data set comprises a sagittal image set of PdWI-fs, a sagittal image set of T1WI and a sagittal image set of T2 mapping, the MRI data set is divided into five levels, namely normal, first grade, second grade, third grade and fourth grade, and the MRI data set is divided into a training set and a test set.
[0029] The image enhancement module is used for performing image enhancement on the images in the training set based on dragonfly optimization and a regional similarity transformation function, wherein the image enhancement is to enhance the contrast of the images, and the boundary features of the knee cartilage, cartilage edema and subchondral bone in the images are acquired.
[0030] The model construction module is used for constructing a knee cartilage injury classification model based on deep learning, wherein the model comprises a convolution layer, a full connection layer, a maximum pooling layer and a Softmax classifier, the convolution layer is used for extracting features in the images, the features refer to cartilage features, cartilage edema features and subchondral bone features, the pooling layer is used for selecting the features extracted in the convolution layer, the full connection layer is used for mapping the feature space of the convolution layer and the pooling layer to a sample label space, and the Softmax classifier is used for classifying the images, wherein the cartilage features, the cartilage edema features and the subchondral bone features are fused in each full connection layer.
[0031] The model training and test evaluation module is used for training and testing the knee cartilage injury classification model, wherein initial values of training parameters in the knee cartilage injury classification model and an iteration termination condition are set, the knee cartilage injury classification model is trained according to the training set, the optimal knee cartilage injury classification model is acquired when the iteration termination condition is met, the optimal knee cartilage injury classification model is tested according to the test set and a prediction value is acquired, and the optimal knee cartilage injury classification model is evaluated through a confusion matrix.
[0032] The model application module is used for classifying knee cartilage injury images according to the optimal knee cartilage injury classification model, and the categories corresponding to the images are acquired.
[0033] The application provides a multi-level classification method for knee cartilage injury based on multi-modal MRI and deep learning, which classifies knee cartilage injury into five levels according to the international cartilage repair association standard, so that the classification standard is more convincing; the dragonfly optimization and regional similarity transformation function are used for feature extraction of knee cartilage, so that the deep learning model is more robust; the deep learning network model is improved, so that the generalization ability is stronger, and the classification accuracy is improved; the knee cartilage injury can be classified according to the image features, and the doctor can quickly and accurately obtain the knee cartilage injury image classification result. BRIEF DESCRIPTION OF DRAWINGS
[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.
[0035] Figure 1 is a method flowchart of the present application;
[0036] Figure 2 is a hospital real data of sagittal plane under different modalities of the present application;
[0037] Figure 3 is a flowchart of knee cartilage feature extraction based on dragonfly optimization of the present application;
[0038] Figure 4 is the feature extraction result based on dragonfly optimization and regional similarity transformation function algorithm of the present application;
[0039] Figure 5 is the confusion matrix and detailed results of knee cartilage classification of the present application;
[0040] Figure 6 is a flowchart of the embodiment of the present application;
[0041] Figure 7 is an experimental result graph of extracting different features of the present application;
[0042] Figure 8 is the model performance under different learning rates of the present application. DETAILED DESCRIPTION
[0043] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.
[0044] Figure 1 For the flowchart of the method of the present application, as shown in Figure 1 , the method of the present embodiment can include:
[0045] S1, acquiring an MRI data set of knee cartilage, since there is no open magnetic resonance data set for detection of knee cartilage injury at present, the present embodiment creates a data set from real hospital data provided by the Second Affiliated Hospital of Dalian Medical University and the Third People's Hospital of Dalian City. The MRI data set includes a set of sagittal images of proton fat suppression sequence images (PdWI-fs), a set of sagittal images of T1 weighted imaging (T1WI), a set of sagittal images of T2 mapping, some images are as shown in Figure 2 , the sagittal plane is usually used for classification of knee cartilage injury, the gender and age ratio in the data set is balanced, and the resolution is 640*640.
[0046] The MRI data set is divided into five levels according to the International Cartilage Repair Society standard, and the five levels are normal (Level 0), first level (Level I), second level (Level II), third level (Level III) and fourth level (Level IV);
[0047] S2, dividing the MRI data set into a training set and a test set, specifically, 70% of the MRI data set is the training set, and 30% of the MRI data set is the test set, image enhancement is performed on the images in the training set based on the dragonfly optimization and the regional similarity transformation function, the main calculation process of the dragonfly algorithm optimization is as shown in Figure 3 . The dragonfly algorithm optimization process is mainly realized by simulating the hunting behavior of dragonflies. This optimization mechanism can be represented as five parts of dragonfly population separation, aggregation, hunting, formation and escape. Dragonflies tend to see more dragonflies to adjust the flight path, so that the population becomes a group in the last stage of optimization, that is, converges to the global optimum, and the fitness reaches the maximum.
[0048] Image enhancement is performed on the images in the training set based on the dragonfly optimization and the regional similarity transformation function, which includes,
[0049] S21, for each image in the training set, each pixel in the image is regarded as a dragonfly, and a fitness function is constructed according to the entropy value of the image, the total number of image edge strength and the number of edges of the image, the fitness function is formula (1), the fitness value of each dragonfly is calculated in turn according to the fitness function, the maximum fitness value and the corresponding pixel point set are obtained,
[0050]
[0051] wherein H(I(x,y)) is the entropy value of the image, C edge (I(x,y)) is the number of edge pixels of the image, G(I(x,y)) is the sobel value of the image, M*N is the image size, I(x,y) is the input image, (x,y) is the pixel position of the image,
[0052] S22, by considering the density distribution similarity between adjacent pixels, the pixels in the pixel point set obtained in S21 are transformed according to formula (2), and a transformed image is obtained,
[0053]
[0054] wherein I(x,y) and O(x,y) are the input image and the output image respectively, is the standard deviation value in the smoothing window, Max is the maximum gray value in the image, a, b, c, d are constants, the ranges of a, b, c, d are a∈[0,3], b∈[0,5], c∈[0,1], d∈[0,1] respectively, S(x,y) is the similarity value of the pixel at pixel point (x,y) in the pixel point set and the surrounding z pixel points, z represents the number of pixel points, the surrounding is the pixel point adjacent to the center, the similarity value S(x,y) of the pixel point (x,y) in the pixel point set is calculated according to formula (3),
[0055]
[0056] wherein I pq =|p-q|, which is the distance between pixels p and q, ω is a normalization coefficient.
[0057] The contrast of the knee cartilage image is enhanced by the dragonfly optimization and regional similarity transformation function algorithm, and the boundary features of the cartilage, cartilage edema and subchondral bone are highlighted, as shown in Figure 4
[0058] S3, a deep learning-based knee cartilage injury classification model is constructed, the model includes a convolutional layer, a fully connected layer, a max pooling layer and a Softmax classifier, the convolutional layer is used to extract features in the image, the features refer to cartilage features, cartilage edema features and subchondral bone features, the pooling layer is used to select the features extracted in the convolutional layer, the fully connected layer is used to map the feature space obtained by the convolutional layer and the pooling layer to the sample label space, the Softmax classifier is used to classify the image, the knee cartilage injury classification model includes 7 convolutional layers, 3 fully connected layers, 3 max pooling layers and a Softmax classifier,
[0059] Each feature map of the convolutional layer is almost always obtained by combining multiple feature maps calculated by the upper layer. The convolutional layer mainly performs feature extraction of the image, and its calculation process is as shown in formula (4),
[0060]
[0061] Wherein, and respectively refer to the input and output layers, indicates an excitation function, U j indicates a set of input feature maps, indicates the i-th element in the convolution kernel of the v-th layer, indicates a bias value, and * indicates a convolution process;
[0062] The pooling layer is also called a down-sampling layer, which generates a sampling result of the input feature map. The pooling layer does not change the number of feature maps, but changes the size of the features. The calculation method of the pooling layer is as shown in formula (5),
[0063]
[0064] Wherein, down(·) indicates a pooling function, which usually indicates a maximum pool, an average pool or a random pool, this layer reduces the feature dimension while retaining useful structural information from the image;
[0065] In order to obtain more complex features of knee cartilage injury, the cartilage features, cartilage edema features and subchondral bone features are fused according to formula (6) in each fully connected layer,
[0066]
[0067] Wherein, ζ is the fused feature, σ is the convolution kernel, and * is the convolution; X1, X2,..., X c , Y1, Y2,..., Y c and Z1, Z2,..., Z crespectively represent the knee cartilage feature, the cartilage edema feature, and the subchondral bone feature, and C represents the number of features;
[0068] In order to overcome the problem of overfitting, a batch normalization layer is used in front of the pooling layer in the embodiment. The batch normalization layer can help the training network to be faster and more stable, less sensitive to initial random weights, and solve the problem of internal covariate shift.
[0069] The batch normalization layer refers to the function of lateral inhibition of the neural network in biology, normalizes the adjacent data, and improves the generalization ability of the model. The layer can also make the local maximum pixel value as the excitation of the next layer to accelerate the convergence. In mathematics, the batch normalized result can be listed as formula (7) shows:
[0070]
[0071] wherein N represents the number of channels shared by each feature mapping internal vector, i represents the number of convolution kernels, j = max(0, i-n / 2), and ε, n, α, and β are hyperparameters;
[0072] The knee cartilage injury level is classified by the Softmax function, and the Softmax function is shown in formula (8),
[0073]
[0074] wherein, is the output value of the i-th node, K is the number of output nodes, i.e., the number of categories of classification, and the Softmax function can convert the output value of multi-classification into a probability distribution with a range of [0, 1] and 1.
[0075] S4, setting the initial value of the training parameter in the knee cartilage injury classification model and the iteration termination condition, the initial value of the training parameter is shown in Table 1,
[0076]
[0077] The iteration termination condition includes but is not limited to setting a fixed number of iterations, training the knee cartilage injury classification model according to the training set, inputting the training set into the constructed model for training, performing multi-feature fusion by extracting the features of the multi-modal knee cartilage injury image, then calculating the loss rate and other indicators of each batch of pictures in the training set, performing back propagation, and updating the parameter value, wherein each batch of pictures refers to training once using all samples in the training set,
[0078] When the iteration termination condition is met, the optimal knee cartilage injury classification model is obtained, the optimal knee cartilage injury classification model is tested according to the test set, and the prediction value is obtained. All samples in the test set are input into the trained network model for prediction.
[0079] The optimal knee cartilage injury classification model is evaluated by a confusion matrix, which is also called a likelihood table or an error matrix. It is a result analysis table for evaluating models or classifiers in the field of machine learning. Each column represents the predicted value, and each row represents the actual value. Building a confusion matrix includes first obtaining a test set, then inputting the test set into the knee cartilage injury classification model to obtain the predicted labels of the test set, and constructing a confusion matrix of the classification results according to the predicted labels and the true labels of the test set. The test set is tested by the classifier to obtain the predicted labels of the test samples, and then the true labels of the test samples are combined to construct a confusion matrix. The developed model is comprehensively evaluated using the confusion matrix, including obtaining statistical prediction results and classification recall (Recall), precision (Precision), F-measure and accuracy (Accuracy). In order to clearly prove the success of the proposed model of knee cartilage injury grade classification method, the confusion matrix and detailed results are shown in Figure 5 .
[0080] The implementation process of the embodiment is shown in Figure 6 First, the cartilage features, cartilage edema features and subchondral bone features are extracted based on the dragonfly optimization and regional similarity transformation function, and then input into the reconstructed AlexNet model for multi-feature fusion and training. Finally, the model is tested for classification of five grades of knee cartilage injury.
[0081] A multi-level classification system for knee cartilage injury based on deep learning of multi-modal MRI includes a data acquisition module, an image enhancement module, a model construction module, a model training and test evaluation module, and a model application module,
[0082] The data acquisition module is used to acquire the MRI data set of the knee cartilage, which includes a set of sagittal images of PdWI-fs, a set of sagittal images of T1WI, and a set of sagittal images of T2 mapping. The MRI data set is divided into five levels, namely normal, first grade, second grade, third grade and fourth grade. The MRI data set is divided into a training set and a test set.
[0083] The image enhancement module is used to enhance the contrast of the images in the training set based on dragonfly optimization and regional similarity transformation function, and to obtain the boundary features of the knee cartilage, cartilage edema and subchondral bone in the images.
[0084] The model construction module is configured to construct a deep learning-based knee cartilage injury classification model, the model comprising a convolutional layer, a fully connected layer, a max pooling layer and a Softmax classifier, the convolutional layer being configured to extract features in an image, the features being cartilage features, cartilage edema features and subchondral bone features, the pooling layer being configured to select the features extracted in the convolutional layer, the fully connected layer being configured to map the feature space obtained by the convolutional layer and the pooling layer to a sample label space, and the Softmax classifier being configured to classify the image, wherein the cartilage features, the cartilage edema features and the subchondral bone features are fused in each fully connected layer;
[0085] The model training and test evaluation module is configured to train the knee cartilage injury classification model and perform testing, set initial values of training parameters and iteration termination conditions in the knee cartilage injury classification model, train the knee cartilage injury classification model according to a training set, obtain an optimal knee cartilage injury classification model when the iteration termination conditions are met, test the optimal knee cartilage injury classification model according to a test set and obtain a prediction value, and evaluate the optimal knee cartilage injury classification model through a confusion matrix
[0086] The model application module is configured to classify knee cartilage injury images according to the optimal knee cartilage injury classification model and obtain categories corresponding to the images.
[0087] In order to make a comparison, the trained VGG16Net, VGG19Net, Squeezenet and AlexNet are applied to the real hospital data set established in this embodiment. The accuracy, test time and F-measure after the experiment are shown in Table 2. As can be seen from Table 2, the proposed model is 0.0692s slower than the nearest competitor in test time, but it achieves better classification accuracy and F-measure compared with all the comparative algorithms.
[0088]
[0089] In this embodiment, multiple features of the knee cartilage are extracted, i.e., cartilage features, cartilage edema features and subchondral bone features. In order to prove the importance of the multiple feature extraction, comparative experiments are performed on the method using real hospital data, including no feature extraction, only one feature extraction and multiple feature extraction, and the experimental results are shown in Table 3. Figure 7
[0090] The numerical setting of the learning rate has a great influence on the proposed model, as shown in Table 4. Figure 8 The experimental results show that when the learning rate is 0.0001, the recall rate and the precision are closest to 1, the curve of F-measure is most gentle at different levels of results, and the accuracy is the highest. In this case, the classification effect is the best.
[0091] Beneficial effects of the whole: the application provides a multi-level classification method for knee cartilage injury based on deep learning of multi-modal MRI, which classifies knee cartilage injury into five levels according to the standard of International Cartilage Repair Society, so that the classification standard is more persuasive; the feature extraction of knee cartilage is based on dragonfly optimization and regional similarity transformation function, so that the deep learning model is more robust; the deep learning network model is improved, so that the generalization ability is stronger, and the classification accuracy is improved; the knee cartilage injury can be classified according to the image features, which can help doctors quickly and accurately obtain the classification results of knee cartilage injury images.
[0092] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.
Claims
1. A deep learning-based multi-modal MRI multi-level classification method for knee cartilage injury, characterized in that, The knee cartilage injury classification model comprises the following steps: S1, acquiring an MRI data set of knee cartilage, the MRI data set comprising a set of sagittal images of PdWI-fs, a set of sagittal images of T1WI, and a set of sagittal images of T2 mapping, and dividing the MRI data set into five levels, namely normal, first grade, second grade, third grade, and fourth grade; S2, dividing the MRI data set into a training set and a test set, performing image enhancement on the images in the training set based on dragonfly optimization and a regional similarity transformation function, the image enhancement being enhancement of the contrast of the images, and acquiring boundary features of knee cartilage, cartilage edema, and subchondral bone in the images; S3, constructing a knee cartilage injury classification model based on deep learning, the model comprising a convolution layer, a fully connected layer, a max-pooling layer, and a Softmax classifier, the convolution layer being used to extract features in the images, the features being cartilage features, cartilage edema features, and subchondral bone features, the pooling layer being used to select the features extracted in the convolution layer, the fully connected layer being used to map the features obtained by the convolution layer and the pooling layer from a feature space to a sample label space, and the Softmax classifier being used to classify the images, wherein the cartilage features, the cartilage edema features, and the subchondral bone features are fused in each fully connected layer; S4, setting initial values of training parameters and an iteration termination condition in the knee cartilage injury classification model, training the knee cartilage injury classification model according to the training set, acquiring an optimal knee cartilage injury classification model when the iteration termination condition is met, testing the optimal knee cartilage injury classification model according to the test set and acquiring a predicted value, and evaluating the optimal knee cartilage injury classification model through a confusion matrix.
2. The method of claim 1, wherein the method is characterized by: The image enhancement of the images in the training set based on dragonfly optimization and a regional similarity transformation function comprises the following steps: S21, for each image in the training set, regarding each pixel point in the image as a dragonfly, constructing an fitness function according to the entropy value of the image, the total number of image edge strengths, and the number of edges of the image, the fitness function being formula (1), sequentially calculating the fitness value of each dragonfly according to the fitness function, acquiring a maximum fitness value and a pixel point set corresponding to the maximum fitness value, where H(I(x, y)) is the entropy value of the image, C edge (I(x, y)) is the number of edge pixels of the image, G(I(x, y)) is the sobel value of the image, M*N is the image size, I(x, y) is the input image, and (x, y) is the pixel position of the image. S22, transforming the pixels in the pixel point set acquired in S21 according to formula (2) to acquire a transformed image, where I(x, y) and O(x, y) are the input image and the output image respectively, is the standard deviation value in the smoothing window, Max is the maximum gray value in the image, a, b, c, d are constants, and their ranges are a∈[0, 3], b∈[0, 5], c∈[0, 1], d∈[0, 1] respectively, S(x, y) is the similarity value of the pixel at the pixel point (x, y) in the pixel point set and the surrounding z pixel points, z represents the number of pixel points, the surrounding is the pixel point adjacent to the center with the pixel point (x, y) as the center, and the similarity value S(x, y) of the pixel point (x, y) in the pixel point set is calculated according to formula (3). where I pq = |p - q|, i.e. the distance between pixels p and q, and ω is a normalization coefficient. 3.The deep learning-based multi-modal MRI multi-level classification method for knee cartilage injury according to claim 1, characterized in that, The fusion of the cartilage features, the cartilage edema features, and the subchondral bone features in each fully connected layer comprises feature fusion according to formula (4), Wherein, ζ is the fused feature, σ is the convolution kernel, * is convolution; X1, X2,..., X c , Y1, Y2,..., Y c and Z1, Z2,..., Z c respectively represent the knee cartilage feature, the cartilage edema feature, the subchondral bone feature, and C represents the number of features.
4. The method of claim 1, wherein the method is characterized by: The knee cartilage injury classification model further comprises a batch normalization layer, the batch normalization layer being used to perform batch normalization processing on the data obtained by the fully connected layer in the knee cartilage injury classification model according to formula (5), wherein N represents the number of channels shared by internal vectors of each feature mapping, i represents the number of convolution kernels, j = max(0, i-n / 2), and ε, n, α, and β are all hyperparameters, and the data subjected to the batch normalization processing is transmitted to the first max-pooling layer. 5.The method of claim 1, wherein the method is characterized by, The iteration termination condition comprises, but is not limited to, a fixed number of iterations.
6. A deep learning-based multi-level classification system for knee cartilage injury based on multi-modal MRI, characterized in that, The method comprises a data acquisition module, an image enhancement module, a model construction module, a model training and test evaluation module, and a model application module. The data acquisition module is used for acquiring an MRI data set of a knee joint cartilage, wherein the MRI data set comprises a sagittal image set of PdWI-fs, a sagittal image set of T1WI, and a sagittal image set of T2 mapping. The MRI data set is divided into five levels, namely normal, first grade, second grade, third grade, and fourth grade. The MRI data set is divided into a training set and a test set. The image enhancement module is used for image enhancement of images in the training set based on dragonfly optimization and regional similarity transformation functions, wherein the image enhancement is contrast enhancement of the images, and boundary features of knee joint cartilage, cartilage edema, and subchondral bone in the images are obtained. The model construction module is used for constructing a knee joint cartilage injury classification model based on deep learning, wherein the model comprises a convolution layer, a full connection layer, a maximum pooling layer, and a Softmax classifier. The convolution layer is used for extracting features in the images, wherein the features refer to cartilage features, cartilage edema features, and subchondral bone features. The pooling layer is used for selecting the features extracted in the convolution layer. The full connection layer is used for spatial mapping of the features obtained by the convolution layer and the pooling layer to sample label space. The Softmax classifier is used for classifying the images, wherein the cartilage features, the cartilage edema features, and the subchondral bone features are fused in each full connection layer. The model training and test evaluation module is used for training and testing the knee joint cartilage injury classification model. Initial values of training parameters in the knee joint cartilage injury classification model and iteration termination conditions are set. The knee joint cartilage injury classification model is trained according to the training set. When the iteration termination conditions are met, the optimal knee joint cartilage injury classification model is obtained. The optimal knee joint cartilage injury classification model is tested according to the test set, and a prediction value is obtained. The optimal knee joint cartilage injury classification model is evaluated by a confusion matrix. The model application module is used for classifying knee joint cartilage injury images according to the optimal knee joint cartilage injury classification model, and obtaining a class corresponding to the images.