CN113256616A

CN113256616A - Knee joint MRI image classification method and device

Info

Publication number: CN113256616A
Application number: CN202110695411.7A
Authority: CN
Inventors: 余家阔; 王鼎予; 吕晨翀; 丁佳; 袁慧书; 朗宁
Original assignee: Peking University Third Hospital Peking University Third Clinical Medical College; Beijing Yizhun Medical AI Co Ltd
Current assignee: Peking University Third Hospital Peking University Third Clinical Medical College; Beijing Yizhun Medical AI Co Ltd
Priority date: 2021-06-23
Filing date: 2021-06-23
Publication date: 2021-08-13
Anticipated expiration: 2041-06-23
Also published as: CN113256616B

Abstract

The invention relates to a knee joint MRI image classification method and a knee joint MRI image classification device, wherein the method comprises the following steps: acquiring a coronal MRI image and a sagittal MRI image of a knee joint, and preprocessing the coronal MRI image and the sagittal MRI image to obtain a preprocessed coronal MRI image and a preprocessed sagittal MRI image; inputting the preprocessed coronal MRI images and the preprocessed sagittal MRI images into a Siamese ACL network in parallel for primary classification so as to determine whether the coronal MRI images and the sagittal MRI images of the knee joint belong to normal samples or fracture samples; when determining that the coronal MRI image and the sagittal MRI image of the knee joint belong to the fractured samples, classifying the fractured samples again through a ResNextACL network to determine the specific fracture type to which the fractured samples belong. Through the technical scheme, the specific fracture types of the MRI images of the knee joint can be classified.

Description

Knee joint MRI image classification method and device

Technical Field

The present disclosure relates to the field of image classification technologies, and in particular, to a knee joint MRI image classification method and apparatus.

Background

MRI is a standard imaging means for determining whether the anterior cruciate ligament is broken or not, because it is noninvasive and has the advantage of high imaging quality for knee joint parts. Considering the quantity and the details of images in each knee joint MRI examination, accurate judgment of whether the anterior cruciate ligament is broken or not through the knee joint MRI is very time-consuming, and automatic identification of the anterior cruciate ligament breakage through an algorithm has a very valuable clinical application prospect.

The method based on deep learning can automatically learn the image characteristics and is very suitable for modeling the complex relationship between the medical image and the interpretability thereof. In recent years, methods based on deep learning have surpassed traditional image analysis methods in many fields, and particularly have made great progress in the field of medical image research, such as blood vessel segmentation, pulmonary nodule detection and the like. While the application of the prior deep learning in knee joint MRI is mainly limited to cartilage segmentation and cartilage damage detection, a few algorithms for ACL fracture detection in knee joint MRI are generally respectively trained for sagittal MRI and coronal MRI to obtain classification models, and then the classification results are weighted and averaged to obtain the final classification result. The classification algorithm essentially takes coronal MRI and sagittal MRI as two independent inputs, rather than taking coronal MRI and sagittal MRI as parallel inputs, and cannot establish the incidence relation between coronal and sagittal images in the model, so that the classification accuracy is affected. Meanwhile, the current ACL fracture classification model can only judge whether the fracture is generated or not, and cannot output a specific fracture type.

Disclosure of Invention

In order to overcome the problems in the related art, the present disclosure provides a method and an apparatus for classifying MRI images of knee joints, which make full use of information on coronal and sagittal positions of MRI images and output specific fracture types for ACL of fractures.

According to a first aspect of the embodiments of the present disclosure, there is provided a method for classifying MRI images of a knee joint, the method including:

acquiring a coronal MRI image and a sagittal MRI image of a knee joint, and preprocessing the coronal MRI image and the sagittal MRI image to obtain a preprocessed coronal MRI image and a preprocessed sagittal MRI image;

inputting the preprocessed coronal MRI images and the preprocessed sagittal MRI images into a Siamese ACL network in parallel for primary classification so as to determine whether the coronal MRI images and the sagittal MRI images of the knee joint belong to normal samples or fracture samples;

when determining that the coronal MRI image and the sagittal MRI image of the knee joint belong to the fractured samples, classifying the fractured samples again through a ResNextACL network to determine the specific fracture type to which the fractured samples belong.

In one embodiment, preferably, the preprocessing the coronal MRI image and the sagittal MRI image comprises:

performing adaptive histogram equalization processing on the coronal MRI image and the sagittal MRI image, and supplementing a recovery value with frequency greater than a first threshold value to a gray value with frequency less than a second threshold value;

and performing interpolation upsampling processing on axial directions of the coronal MRI image and the sagittal MRI image after equalization processing so as to adjust the resolution of the axial position.

In one embodiment, preferably, the pre-processed coronal MRI image and the pre-processed sagittal MRI image are input to the siameselar network in parallel for preliminary classification to determine whether the coronal MRI image and the sagittal MRI image of the knee joint belong to a normal sample or a fractured sample, including:

inputting the preprocessed coronal MRI image and the preprocessed sagittal MRI image into a Siamese ACL network in parallel, and respectively extracting a feature map of a first coronal position and a feature map of a first sagittal position through a ResAttentNet main network in the Siamese ACL network;

respectively carrying out three-dimensional adaptive average pooling on the characteristic diagram of the first coronal position and the characteristic diagram of the first sagittal position to obtain a 1024-dimensional coronal position characteristic vector and a 1024-dimensional sagittal position characteristic vector;

calculating the distance between the coronal eigenvector and the sagittal eigenvector to obtain a characteristic distance loss;

carrying out element-level averaging on the coronal feature vector and the sagittal feature vector to obtain a new 1024-dimensional feature vector;

inputting the new 1024-dimensional feature vector into a full connection layer to output the probability of ACL fracture;

calculating the focus loss between the output ACL fracture probability and the true value;

and carrying out weighted summation on the characteristic distance loss and the focus loss to obtain a final loss.

In one embodiment, preferably, the resontentionnet backbone network includes a plurality of attention mechanism modules, and the attention mechanism modules are connected with the s-type activation function after sequentially performing downsampling processing, upsampling processing and convolution processing to obtain the attention weight distribution.

In one embodiment, preferably, the reclassifying the fractured samples through the resnextcal network to determine the specific fracture types to which the fractured samples belong includes:

inputting the coronal MRI image and the sagittal MRI image of the fractured sample into a ResNextACL network in parallel, and extracting a feature map of a second coronal position and a feature map of a second sagittal position respectively through a ResNext backbone network of the ResNextACL network;

and merging the characteristic diagram of the second coronal position and the characteristic diagram of the second sagittal position on a channel dimension, obtaining a 2048-dimensional coronal position characteristic vector and a sagittal position characteristic vector through three-dimensional adaptive average pooling, and outputting a specific fracture type through a full-link layer.

According to a second aspect of the embodiments of the present disclosure, there is provided a knee joint MRI image classification apparatus, the apparatus including:

the preprocessing module is used for acquiring coronal MRI images and sagittal MRI images of the knee joint, and preprocessing the coronal MRI images and the sagittal MRI images to obtain preprocessed coronal MRI images and preprocessed sagittal MRI images;

the first classification module is used for inputting the preprocessed coronal MRI images and the preprocessed sagittal MRI images into a Siamese ACL network in parallel to perform primary classification so as to determine whether the coronal MRI images and the sagittal MRI images of the knee joint belong to normal samples or fracture samples;

and the second classification module is used for classifying the fractured samples again through a ResNextACL network when determining that the coronal MRI image and the sagittal MRI image of the knee joint belong to the fractured samples so as to determine the specific fracture types to which the fractured samples belong.

In one embodiment, preferably, the preprocessing module includes:

the first processing unit is used for carrying out adaptive histogram equalization processing on the coronal MRI image and the sagittal MRI image and supplementing a restoration value with frequency greater than a first threshold value to a gray value with frequency less than a second threshold value;

and the second processing unit is used for performing interpolation up-sampling processing on the axial directions of the coronal MRI image and the sagittal MRI image after the equalization processing so as to adjust the resolution of the axial position.

In one embodiment, preferably, the first classification module includes:

the first extraction unit is used for inputting the preprocessed coronal MRI images and the preprocessed sagittal MRI images into a Siamese ACL network in parallel, and extracting a feature map of a first coronal position and a feature map of a first sagittal position respectively through a ResAttentionNet main network in the Siamese ACL network;

the vector extraction unit is used for respectively carrying out three-dimensional self-adaptive average pooling on the characteristic diagram of the first coronal position and the characteristic diagram of the first sagittal position to obtain a 1024-dimensional coronal position characteristic vector and a 1024-dimensional sagittal position characteristic vector;

a first calculating unit, configured to calculate a distance between the coronal eigenvector and the sagittal eigenvector to obtain a characteristic distance loss;

the averaging unit is used for carrying out element-level averaging on the coronal position feature vector and the sagittal position feature vector to obtain a new 1024-dimensional feature vector;

the output unit is used for inputting the new 1024-dimensional feature vector into a full connection layer so as to output the ACL fracture probability;

the second calculation unit is used for calculating the focus loss between the probability of the output ACL fracture and the true value;

and the third calculation unit is used for performing weighted summation calculation on the characteristic distance loss and the focus loss to obtain a final loss.

In one embodiment, preferably, the second classification module includes:

the second extraction unit is used for inputting the coronal MRI image and the sagittal MRI image of the fractured sample into a ResNextACL network in parallel and extracting a feature map of a second coronal position and a feature map of a second sagittal position respectively through a ResNext backbone network of the ResNextACL network;

and the classification unit is used for merging the second coronal characteristic diagram and the second sagittal characteristic diagram in a channel dimension, obtaining 2048-dimensional coronal characteristic vectors and sagittal characteristic vectors through three-dimensional adaptive average pooling, and outputting specific fracture types through a full connection layer.

According to a third aspect of the embodiments of the present disclosure, there is provided a knee joint MRI image classification apparatus, the apparatus including:

a processor;

a memory for storing processor-executable instructions;

wherein the processor is configured to:

According to a fourth aspect of embodiments of the present disclosure, there is provided a computer-readable storage medium having stored thereon computer instructions which, when executed by a processor, implement the steps of the method of any one of the first aspects.

The technical scheme provided by the embodiment of the disclosure can have the following beneficial effects:

in the embodiment of the invention, the image characteristics of knee joint MRI are combined, and targeted preprocessing including self-adaptive equalization and interpolation sampling adjustment of resolution is carried out, so that the preprocessing has the advantages that more effective characteristics can be extracted by a network, and the influence of interference information on the final classification result is reduced. By combining medical common knowledge and the advantages of a twin network, a Siamese ACL classification network for inputting coronal and sagittal MRI images in parallel is designed, and the influence of coronal and sagittal MRI image characteristics on a final classification result is considered by the classification network simultaneously through weighted superposition of two loss functions, so that the problem of insufficient robustness of the classification result caused by single consideration of the coronal or sagittal MRI images is avoided. The invention not only can obtain the prediction result of whether the fracture occurs or not through the neural network classification model, but also designs the ResNextACL classification network aiming at the fractured case and further obtains the specific fracture type.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and together with the description, serve to explain the principles of the disclosure.

Fig. 1 is a flowchart illustrating a knee joint MRI image classification method according to an exemplary embodiment.

Fig. 2 is a flowchart illustrating a step S101 in a classification method of MRI images of a knee joint according to an exemplary embodiment.

Fig. 3 is a flowchart illustrating step S102 in a knee joint MRI image classification method according to an exemplary embodiment.

Fig. 4 is a schematic diagram of a siamesacl network shown in accordance with an exemplary embodiment.

Fig. 5 is a flowchart illustrating step S103 of a knee joint MRI image classification method according to an exemplary embodiment.

Fig. 6 is a diagram illustrating a ResNext structure in accordance with an exemplary embodiment.

Fig. 7 is a block diagram illustrating a knee joint MRI image classification apparatus according to an exemplary embodiment.

Fig. 8 is a block diagram illustrating a preprocessing module in a knee joint MRI image classification device according to an exemplary embodiment.

Fig. 9 is a block diagram illustrating a first classification module of a knee joint MRI image classification device according to an exemplary embodiment.

Fig. 10 is a block diagram illustrating a second classification module of a knee joint MRI image classification device according to an exemplary embodiment.

Detailed Description

Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, like numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the exemplary embodiments below are not intended to represent all implementations consistent with the present disclosure. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present disclosure, as detailed in the appended claims.

Fig. 1 is a flowchart illustrating a knee joint MRI image classification method according to an exemplary embodiment, as shown in fig. 1, the method including:

step S101, acquiring coronal MRI images and sagittal MRI images of a knee joint, and preprocessing the coronal MRI images and the sagittal MRI images to obtain preprocessed coronal MRI images and preprocessed sagittal MRI images;

step S102, inputting the preprocessed coronal MRI images and the preprocessed sagittal MRI images into a Siamese ACL network in parallel for preliminary classification so as to determine whether the coronal MRI images and the sagittal MRI images of the knee joint belong to normal samples or fracture samples;

and step S103, when determining that the coronal MRI image and the sagittal MRI image of the knee joint belong to a fracture sample, reclassifying the fracture sample through a ResNextACL network to determine the specific fracture type to which the fracture sample belongs.

As shown in fig. 2, in one embodiment, preferably, the step S101 includes:

step S201, performing adaptive histogram equalization processing on the coronal MRI image and the sagittal MRI image, and supplementing the gray value with the frequency greater than a first threshold value to the gray value with the frequency less than a second threshold value;

step S202, carrying out interpolation up-sampling processing on axial directions of the coronal MRI image and the sagittal MRI image after the equalization processing so as to adjust the resolution of the axial position.

The direct use of coronal and sagittal MRI images of the knee joint as training sets presents the following problems: the whole knee joint MRI image is dark, the detail information is not clear enough, and the neural network is not favorable for extracting effective characteristics, so that the confidence coefficient of a diagnosis result is reduced. In order to solve the above problems, the adopted preprocessing means comprises: self-adaptive histogram equalization supplements the gray value with high frequency to the gray value with low frequency, so that the brightness of the whole image is improved, and meanwhile, the detailed information can be kept, and the neural network can conveniently extract more effective characteristics. In addition, because the difference between the axial resolution and the non-axial resolution of the MRI image is large, interpolation up-sampling is carried out on the axial direction of the MRI image after the histogram equalization, and the purpose of doing so is mainly to adjust the axial resolution, so that the axial resolution is approximately equal to the resolution of the coronal position and the resolution of the sagittal position, and the occurrence of anisotropic semantic difference during convolution is avoided. After a pre-processing phase, the input part of the network is obtained: coronal and sagittal MRI of the knee joint.

And a data preprocessing stage, which aims to perform targeted processing on the original knee joint MRI image, so that the network can learn the effective characteristics of the ACL part more easily, and thus, accurate classification can be performed. First, the collected data set is updated according to 7: 1: dividing the test data set into a training data set, a verification data set and a test data set, and ensuring that the ratio of positive samples to negative samples is close to 1: 1, avoiding the problem that model prediction has bias caused by sample imbalance as much as possible: model predictions may be more biased towards categories with large amounts of data. Considering the original knee joint MRI size, the adopted equalization method is adaptive histogram equalization for limiting contrast, parameters of the method comprise a limited contrast threshold value and the number of subblocks, and experiments prove that the contrast threshold value parameter range adopted in the invention is 1-3 and the number of subblocks parameter range is 16-64. The contrast threshold is set to overcome the problem of over-amplifying noise in conventional adaptive histogram equalization.

As shown in fig. 3, in one embodiment, preferably, the step S102 includes:

step S301, inputting the preprocessed coronal MRI image and the preprocessed sagittal MRI image into a Siamese ACL network in parallel, and extracting a feature map of a first coronal position and a feature map of a first sagittal position respectively through a ResAttentionNet main network in the Siamese ACL network;

step S302, performing three-dimensional adaptive average pooling on the first coronal feature map and the first sagittal feature map respectively to obtain 1024-dimensional coronal feature vectors and 1024-dimensional sagittal feature vectors;

step S303, calculating the distance between the coronal position feature vector and the sagittal position feature vector to obtain the feature distance loss;

step S304, carrying out element-level averaging on the coronal position feature vector and the sagittal position feature vector to obtain a new 1024-dimensional feature vector;

step S305, inputting the new 1024-dimensional feature vector into a full connection layer to output the ACL fracture probability;

step S306, calculating the focus loss between the output ACL fracture probability and the truth value;

and step S307, carrying out weighted summation on the characteristic distance loss and the focus loss to obtain final loss.

In one embodiment, preferably, the resontentionnet backbone network includes a plurality of attention mechanism modules, and the attention mechanism modules are connected with a sigmoid activation function (sigmoid) to obtain attention weight distribution after sequentially performing downsampling processing, upsampling processing and convolution processing.

In this embodiment, the present invention proposes a classification network siamesacl that can output MRI images of coronal and sagittal knee joints after preprocessing in parallel, and the output of the network at this stage is a 2 × 1 vector (corresponding to whether ACL is broken or not). The main network is Resattention Net, wherein an Attention mechanism (Attention) module is added, the Attention module is proposed and inspired by the Attention mechanism of human brain, the characteristics are calculated according to different influences of different areas on a final task, and the Attention mechanism essentially allocates different weights to the areas with different importance and finally performs weighted summation because the importance weights of the areas to a result of the final task are different, and the distribution of the Attention weights of the different areas is the core of the whole Attention mechanism. The Attention weight distribution is obtained by directly connecting a sigmoid activation function after convolution on the basis of down sampling and up sampling, and the addition of an Attention mechanism improves the capability of effectively extracting features of a network, as shown in FIG. 4. Extracting feature maps of coronal position and sagittal position respectively through backbone network, obtaining 1024 x 1 features as feature vectors respectively through 3D adaptive positive pooling, the design idea of twin network in the invention is to make the distance between the two feature vectors small enough, so we add the L2 distance between coronal and sagittal feature vectors in this section as loss1, the distance between the characteristic vectors is reduced in the training process of the network, an average characteristic vector is obtained at the last part of the network through arithmetic averaging, then a full connection layer is connected to output a two-dimensional vector, the final classification result of whether the network is broken or not is obtained, the focus Loss function (Focol Loss) from the true value is taken as the second partial Loss function Loss2, the total Loss is a weighted sum of Loss1 and Loss2, the optimal weights for loss1 and loss2 were determined experimentally to be 0.3 and 0.7, respectively. In order to test the performance of the model, a comparative test is carried out on test data with the current related method, and the test result is shown in table 1;

TABLE 1 comparative test results with current related methods

Method	Accuracy of	Sensitivity of the probe	Degree of specificity
				Presently related methods	0.867	0.759	0.97
The method of the invention	0.905	0.772	0.965

As shown in fig. 5, in one embodiment, preferably, the step S103 includes:

step S501, inputting the coronal MRI image and the sagittal MRI image of the fractured sample into a ResNextACL network in parallel, and extracting a feature map of a second coronal position and a feature map of a second sagittal position respectively through a ResNext backbone network of the ResNextACL network;

step S502, merging the second coronal feature map and the second sagittal feature map in a channel dimension, obtaining 2048-dimensional coronal feature vectors and sagittal feature vectors through three-dimensional adaptive average pooling, and outputting specific fracture types through a full connection layer.

In this example, the input to the resnextcl network is set to parallel coronal and sagittal MRI fracture images, and the output is a 6 x 1 vector representing the predicted probability values for the six fracture types. The main network of the resextacl classification network is a resext network, the resext structure is shown in fig. 6, and essentially the packet convolution is combined with a residual network, and the invention sets the packet number to 32 by applying the idea of deep separable convolution by controlling the packet number. The structure has the advantages that the network performance is improved on the premise that the number of parameters is not increased, the topological structures of each branch group of ResNext are the same, and the hardware design principle of the GPU is better met. Different from the element-wise averaging of the siemesacl network, the resenextgac respectively obtains characteristic maps of coronal and sagittal MRIs through a main network, then carries out channel-wise containment on the characteristic maps, then carries out three-dimensional 3 dadaptive posing, and finally connects the characteristic vector of the full-connection layer output 6 1 to correspond to six specific fracture types.

As shown in fig. 7, according to a second aspect of the embodiments of the present disclosure, there is provided a knee joint MRI image classification apparatus, including:

the preprocessing module 71 is configured to acquire a coronal MRI image and a sagittal MRI image of the knee joint, and preprocess the coronal MRI image and the sagittal MRI image to obtain a preprocessed coronal MRI image and a preprocessed sagittal MRI image;

a first classification module 72, configured to input the preprocessed coronal MRI image and the preprocessed sagittal MRI image into a siameselar network in parallel for preliminary classification, so as to determine whether the coronal MRI image and the sagittal MRI image of the knee joint belong to a normal sample or a fractured sample;

and the second classification module 73 is used for classifying the fractured samples again through the ResNextACL network when the coronal MRI image and the sagittal MRI image of the knee joint are determined to belong to the fractured samples so as to determine the specific fracture types to which the fractured samples belong.

As shown in fig. 8, in one embodiment, the preprocessing module 71 preferably includes:

a first processing unit 81, configured to perform adaptive histogram equalization on the coronal MRI image and the sagittal MRI image, and supplement the gray value with the frequency greater than the first threshold to the gray value with the frequency less than the second threshold; wherein the first threshold is greater than the second threshold.

And a second processing unit 82, configured to perform interpolation upsampling processing on the axial directions of the coronal MRI image and the sagittal MRI image after the equalization processing, so as to adjust the resolution of the axial position.

As shown in fig. 9, in one embodiment, preferably, the first classification module 72 includes:

the first extraction unit 91 is configured to input the preprocessed coronal MRI image and the preprocessed sagittal MRI image in parallel to the siamese acl network, and extract a feature map of the first coronal position and a feature map of the first sagittal position through a restentionet backbone network in the siamese acl network;

a vector extraction unit 92, configured to perform three-dimensional adaptive average pooling on the feature map of the first coronal position and the feature map of the first sagittal position respectively to obtain a 1024-dimensional coronal position feature vector and a 1024-dimensional sagittal position feature vector;

a first calculating unit 93, configured to calculate a distance between the coronal eigenvector and the sagittal eigenvector to obtain a characteristic distance loss;

an averaging unit 94, configured to perform element-level averaging on the coronal eigenvector and the sagittal eigenvector to obtain a new 1024-dimensional eigenvector;

an output unit 95, configured to input the new 1024-dimensional feature vector to a full connection layer, so as to output a probability of ACL fracture;

a second calculating unit 96 for calculating a focus loss between the probability of the output ACL breaking and the true value;

and a third calculating unit 97, configured to perform weighted summation calculation on the characteristic distance loss and the focus loss to obtain a final loss.

In one embodiment, preferably, the resettitionnet backbone network includes a plurality of attention mechanism modules, and the attention mechanism modules are connected with the S-type activation function after sequentially performing downsampling processing, upsampling processing and convolution processing to obtain the attention weight distribution.

As shown in fig. 10, in one embodiment, preferably, the second classification module 73 includes:

a second extraction unit 1001, configured to input the coronal MRI image and the sagittal MRI image of the fractured sample into a resnextcan network in parallel, and extract a feature map of a second coronal position and a feature map of a second sagittal position through a ResNext backbone network of the resnextcan network, respectively;

and the classification unit 1002 is configured to merge the feature map of the second coronal position and the feature map of the second sagittal position in a channel dimension, obtain a 2048-dimensional coronal position feature vector and a 2048-dimensional sagittal position feature vector through three-dimensional adaptive average pooling, and output a specific fracture type through a full connection layer.

a processor;

a memory for storing processor-executable instructions;

wherein the processor is configured to:

It is further understood that the use of "a plurality" in this disclosure means two or more, as other terms are analogous. "and/or" describes the association relationship of the associated objects, meaning that there may be three relationships, e.g., a and/or B, which may mean: a exists alone, A and B exist simultaneously, and B exists alone. The character "/" generally indicates that the former and latter associated objects are in an "or" relationship. The singular forms "a", "an", and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.

It will be further understood that the terms "first," "second," and the like are used to describe various information and that such information should not be limited by these terms. These terms are only used to distinguish one type of information from another and do not denote a particular order or importance. Indeed, the terms "first," "second," and the like are fully interchangeable. For example, first information may also be referred to as second information, and similarly, second information may also be referred to as first information, without departing from the scope of the present disclosure.

It is further to be understood that while operations are depicted in the drawings in a particular order, this is not to be understood as requiring that such operations be performed in the particular order shown or in serial order, or that all illustrated operations be performed, to achieve desirable results. In certain environments, multitasking and parallel processing may be advantageous.

Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the disclosure disclosed herein. This application is intended to cover any variations, uses, or adaptations of the disclosure following, in general, the principles of the disclosure and including such departures from the present disclosure as come within known or customary practice within the art to which the disclosure pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the disclosure being indicated by the following claims.

It will be understood that the present disclosure is not limited to the precise arrangements described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A method of classifying MRI images of a knee joint, the method comprising:

2. The method of claim 1, wherein pre-processing the coronal and sagittal MRI images comprises:

performing adaptive histogram equalization processing on the coronal MRI image and the sagittal MRI image, and supplementing the gray value with the frequency greater than a first threshold value to the gray value with the frequency less than a second threshold value;

3. The method of claim 1, wherein the pre-processed coronal MRI image and the pre-processed sagittal MRI image are input in parallel to a siamese acl network for preliminary classification to determine whether the coronal MRI image and the sagittal MRI image of the knee joint belong to a normal sample or a fractured sample, comprising:

4. The method of claim 3, wherein the ResAttentNet backbone network comprises a plurality of attention mechanism modules that are connected to an sigmoid activation function to obtain an attention weight distribution after a downsampling process, an upsampling process, and a convolution process in sequence.

5. The method of claim 1, wherein reclassifying the fractured samples through the ResNextACL network to determine the specific fracture types to which the fractured samples belong comprises:

6. A knee MRI image classification apparatus, the apparatus comprising:

7. The apparatus of claim 6, wherein the pre-processing module comprises:

8. The apparatus of claim 6, wherein the first classification module comprises:

9. The apparatus of claim 8, wherein the ResAttentNet backbone network comprises a plurality of attention mechanism modules, and wherein the attention mechanism modules are connected to the sigmoid activation function after the downsampling process, the upsampling process, and the convolution process in sequence to obtain the attention weight distribution.

10. The apparatus of claim 6, wherein the second classification module comprises: