An automated classification method and system for adenomyosis based on multiplanar MRI images

Abstract

This invention relates to an automatic classification method and system for adenomyosis based on multiplanar MRI images, belonging to the field of adenomyosis classification technology, and solving the problem of inaccurate classification in existing technologies. The method includes the following steps: segmenting multiplanar MRI images of different individuals using multiple trained segmentation models to extract the multiplanar uterine region of interest (ROI) for each individual; constructing a second training set based on the multiplanar uterine ROI and the corresponding adenomyosis classification for each individual; constructing a deep learning model for adenomyosis classification; training the deep learning model based on the second training set to obtain an adenomyosis classification model; inputting the multiplanar MRI images of the individual to be classified into the trained segmentation models to obtain the multiplanar uterine ROI for the individual to be classified; and inputting the multiplanar uterine ROI of the individual to be classified into the adenomyosis classification model to obtain the adenomyosis classification for the individual to be classified. This achieves accurate and efficient adenomyosis classification.

Description

Technical Field

[0001] This invention relates to the field of adenomyosis classification technology, and in particular to an automatic classification method and system for adenomyosis based on multiplanar MRI images. Background Technology

[0002] Adenomyosis is a condition caused by the invasion and growth of endometrial tissue (including glands and stroma) into the myometrium, leading to symptoms such as dysmenorrhea, menorrhagia, and infertility. It commonly affects women of reproductive age, with an incidence rate of 7%–23%, severely impacting patients' quality of life and fertility. Because the etiology of adenomyosis is still unclear, imaging diagnosis plays a crucial role in its classification and clinical management. Magnetic resonance imaging (MRI), due to its advantages such as intuitive imaging, operator-free operation, and multi-parameter, multi-planar imaging, is increasingly used in the diagnosis, classification, and continuous monitoring of adenomyosis after drug treatment.

[0003] Regarding the imaging classification of adenomyosis, the four-subtype classification based on MRI images proposed by Kishi et al. in 2012 is currently the most widely used. Kishi et al. divided adenomyosis into four subtypes based on the location of the lesions: (1) Type I: Adenomyosis lesions infiltrate the inner layer of the uterus without affecting the external structure; (2) Type II: Adenomyosis lesions infiltrate the outer layer of the uterus without affecting the internal structure; (3) Type III: Locally invasive adenomyosis lesions, including adenomyomas and cystic adenomyosis, without affecting the overall structure of the uterus; (4) Type IV: Adenomyosis lesions that do not meet the diagnostic criteria of the above three types. Current research suggests that this classification has important clinical significance for the stratified treatment and efficacy evaluation of adenomyosis patients.

[0004] Traditional imaging classification of adenomyosis mainly relies on manual image interpretation, with radiologists judging the morphology and signal characteristics of lesions based on T2-weighted magnetic resonance imaging (T2WI), which is easily influenced by subjective experience. In addition, the slice thickness of MRI scans is usually thick (4-8mm), which may lead to the loss of some lesion information, resulting in low consistency in classification and further increasing the difficulty of classification.

[0005] With the development of artificial intelligence technologies such as deep learning and reinforcement learning, deep learning has shown superior performance in the field of medical image analysis. However, current studies all use single-plane MRI data (T2WI sagittal plane) to train deep learning models. Adenomyosis lesions may present different characteristics in MRI images of different planes. Single-plane training may lead to insufficient information acquisition by the model, affecting the accuracy of classification. Summary of the Invention

[0006] Based on the above analysis, the present invention aims to provide an automatic classification method and system for adenomyosis based on multiplanar MRI images, in order to solve the problem of inaccurate classification in existing methods.

[0007] On one hand, embodiments of the present invention provide an automatic classification method for adenomyosis based on multiplanar MRI images, comprising the following steps:

[0008] Based on multiple trained segmentation models, multiplanar MRI images of different individuals are segmented to extract the multiplanar uterine ROI of each individual. A second training set is constructed based on the multiplanar uterine ROI of each individual and the corresponding adenomyosis classification.

[0009] A deep learning model for adenomyosis classification is constructed, and the deep learning model is trained based on the second training to obtain an adenomyosis classification model.

[0010] Multiplanar MRI images of the individual to be classified are input into multiple trained segmentation models to obtain the multiplanar uterine ROI of the individual to be classified; the multiplanar uterine ROI of the individual to be classified is input into the adenomyosis classification model to obtain the adenomyosis classification of the individual to be classified.

[0011] Based on further improvements to the above method, multiple trained segmentation models are obtained in the following manner:

[0012] S11. Construct an initial training set, wherein each sample in the initial training set includes an individual's multiplanar MRI image and the corresponding uterine mask; construct multiple segmentation models; each MRI scan plane corresponds to one segmentation model.

[0013] S12. Train the multiple segmentation models based on the initial training set;

[0014] S13. If the number of samples in the initial sample set reaches a preset threshold, multiple trained segmentation models are obtained; otherwise, newly added unlabeled samples are obtained, and the newly added unlabeled samples are segmented based on the multiple segmentation models.

[0015] S15. Obtain the corrected segmentation results for the newly added unlabeled samples, and add the newly added unlabeled samples and their corresponding corrected results to the initial training set; return to step S12.

[0016] Based on the further improvements to the above method, the training loss of each segmentation model is calculated using the following formula:

[0017]

[0018] Where λ1, λ2, and λ3 represent weights. Indicates DICE loss, Indicates focal loss. This indicates boundary loss.

[0019] Based on the further improvement of the above method, the boundary loss is calculated using the following formula:

[0020]

[0021] in, This represents the gradient magnitude indicating the probability that the model predicts the i-th pixel to be in the uterine region. denoted as the gradient magnitude of the uterine mask at the i-th pixel, and n represents the number of pixels in the MRI image.

[0022] Based on the further improvement of the above method, the gradient magnitude of the probability that the i-th pixel is in the uterine region is calculated in the following way:

[0023] The Sobel operator is used to convolve the probability map predicted by the model in the horizontal and vertical directions respectively to obtain the horizontal gradient map and the vertical gradient map.

[0024] Using formula Calculate the gradient magnitude of the probability that the i-th pixel is in the uterine region, where This represents the value of the i-th pixel in the horizontal gradient map. This represents the value of the i-th pixel in the vertical gradient map.

[0025] Based on further improvements to the above method, the deep learning model includes:

[0026] The feature preprocessing module includes multiple branches, each of which is used to extract features from the corresponding uterine ROI in the plane to obtain a feature map;

[0027] The attention module is used to calculate attention features from feature maps based on a cross-plane spatial-channel hybrid mechanism.

[0028] The classification module is used for predicting adenomyosis subtypes based on attention features.

[0029] Based on a further improvement to the above method, the attention module calculates attention features in the following manner:

[0030] Each feature map is converted into a one-dimensional feature sequence. The position code of each element in the sequence is calculated. The position code is concatenated with the feature of the element to obtain the feature sequence of fused spatial location.

[0031] The attention features are obtained by splicing the feature sequences of the fused spatial locations and calculating the spatial-channel hybrid attention of the spliced ​​feature sequences.

[0032] Based on the above method, a further improvement is made, and the position code of the i-th element in the one-dimensional feature sequence is calculated using the following formula:

[0033]

[0034] Where (x,y,z) represents the three-dimensional coordinates of the i-th element in the feature map, and d model Indicates the length of the feature sequence.

[0035] Based on the further improvement of the above method, the formula for calculating attention features is expressed as:

[0036]

[0037] Channel Attenion (F concat )=σ(FC(ReLU(FC(AvgPool(F concat ))))+FC(ReLu(FC(MaxPool(F concat )))))

[0038] Where AvgPool(·) represents global average pooling in the spatial dimension, MaxPool(·) represents global max pooling in the spatial dimension, FC(·) represents a fully connected layer used to model the relationships between channels, and σ(·) represents the Sigmoid activation function. This represents bitwise multiplication, used for channel-weighted spatial attention, and outputs attention features. Attention(F) concat ) represents self-attention.

[0039] On the other hand, embodiments of the present invention provide an automatic classification system for adenomyosis based on multiplanar MRI images, comprising:

[0040] The second training set construction module is used to segment and extract the multiplanar uterine ROI of each individual based on multiple trained segmentation models, and to construct the second training set based on the multiplanar uterine ROI of each individual and the corresponding adenomyosis classification.

[0041] The adenomyosis classification model training module is used to construct a deep learning model for adenomyosis classification, and the deep learning model is trained based on the second training to obtain the adenomyosis classification model.

[0042] The classification prediction module is used to input multiplanar MRI images of an individual to be classified into multiple pre-trained segmentation models to obtain the multiplanar uterine ROI of the individual to be classified; and to input the multiplanar uterine ROI of the individual to be classified into the adenomyosis classification model to obtain the adenomyosis classification of the individual to be classified.

[0043] Compared to existing technologies, this invention first uses multiple trained segmentation models to segment multiplanar MRI images of different individuals, extracting multiplanar uterine ROIs for each individual. This constructs a second training set. A deep learning model for adenomyosis classification is then trained based on this second training set, enhancing the model's understanding of the spatial distribution of adenomyosis lesions. This ensures the integrity of information from different anatomical perspectives, alleviates the limitations of single-planar perspectives on lesion visibility, reduces structural information loss due to the low resolution of thick-slice MRI sequences, and improves classification accuracy. For the individual to be classified, their multiplanar MRI images are input into the segmentation model to obtain the multiplanar uterine ROI. Then, the multiplanar uterine ROI is input into the adenomyosis classification model to quickly and accurately obtain the adenomyosis classification for the individual, improving the efficiency and quality of classification. The results are more consistent with clinical diagnostic procedures, enhancing the model's decision support capabilities for clinicians.

[0044] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0045] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0046] Figure 1 This is a flowchart of an automatic classification method for adenomyosis based on multiplanar MRI images according to an embodiment of the present invention;

[0047] Figure 2 This is a block diagram of the automatic classification posture of adenomyosis based on multiplanar MRI images in an embodiment of the present invention. Detailed Implementation

[0048] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0049] A specific embodiment of the present invention discloses an automatic classification method for adenomyosis based on multiplanar MRI images, such as... Figure 1 As shown, it includes the following steps:

[0050] S1. Based on multiple trained segmentation models, segment the multiplanar MRI images of different individuals to extract the multiplanar uterine ROI of each individual, and construct a second training set based on the multiplanar uterine ROI of each individual and the corresponding adenomyosis classification.

[0051] S2. Construct a deep learning model for adenomyosis classification, and train the deep learning model based on the second training to obtain an adenomyosis classification model.

[0052] S3. Input the multiplanar MRI images of the individual to be classified into multiple trained segmentation models to obtain the multiplanar uterine ROI of the individual to be classified; input the multiplanar uterine ROI of the individual to be classified into the adenomyosis classification model to obtain the adenomyosis classification of the individual to be classified.

[0053] It should be noted that multiplanar MRI images include sagittal MRI images, axial MRI images, and coronal MRI images.

[0054] Compared with existing technologies, the automatic adenomyosis classification method based on multiplanar MRI images provided in this embodiment first uses multiple trained segmentation models to segment multiplanar MRI images of different individuals, extracting multiplanar uterine ROIs for each individual, thus constructing a second training set. A deep learning model for adenomyosis classification is then trained based on this second training set, enhancing the model's understanding of the spatial distribution of adenomyosis lesions. This ensures the integrity of information from different anatomical perspectives, alleviates the limitations of single-planar perspectives on lesion visibility, reduces structural information loss due to the low resolution of thick-slice MRI sequences, and improves classification accuracy. For the individual to be classified, their multiplanar MRI images are input into the segmentation model to obtain the multiplanar uterine ROI; then, the multiplanar uterine ROI is input into the adenomyosis classification model to quickly and accurately obtain the adenomyosis classification for the individual, improving the efficiency and quality of classification. The results are more consistent with clinical diagnostic procedures, enhancing the model's decision support capabilities for clinicians.

[0055] During implementation, pelvic MRI images contain a large amount of background tissue, such as the bladder, rectum, and ovaries, which may interfere with the classification of adenomyosis. To improve the model's classification efficiency, it is necessary to first perform automatic uterine region segmentation on the MRI images, i.e., image semantic segmentation, extracting the uterine ROI (Region of Interest), and removing background noise to improve the accuracy of the subsequent classification model. On the other hand, the classification of adenomyosis is related to the structures invaded by the lesions, and precise localization of the lesions plays an important role in assisting classification.

[0056] Specifically, the following methods are used to obtain multiple trained segmentation models:

[0057] S11. Construct an initial training set, wherein each sample in the initial training set includes an individual's multiplanar MRI image and the corresponding uterine mask; construct multiple segmentation models; each MRI scan plane corresponds to one segmentation model;

[0058] S12. Train the multiple segmentation models based on the initial training set;

[0059] S13. If the number of samples in the initial sample set reaches a preset threshold, multiple trained segmentation models are obtained; otherwise, newly added unlabeled samples are obtained, and the newly added unlabeled samples are segmented based on the multiple segmentation models.

[0060] S15. Obtain the corrected segmentation results for the newly added unlabeled samples, and add the newly added unlabeled samples and their corresponding corrected results to the initial training set; return to step S12.

[0061] In practice, a small sample of data was first labeled. For example, 50 cases of pelvic MRI images were selected, each containing complete images of the sagittal, axial, and coronal planes on T2WI. Senior radiologists manually delineated the overall boundary of the uterus and the boundary of the adenomyosis lesion for each case and layer using open-source medical image annotation software, forming the initial training dataset. This dataset contained sagittal, axial, and coronal MRI images of the 50 individuals, as well as the corresponding uterine masks for the three planes, and was stored in NIfTI format. Correspondingly, three segmentation models were developed.

[0062] In practice, the segmentation model can be the nnU-Net model.

[0063] The three segmentation models constructed based on the initial training set are trained until the preset number of training iterations is reached, at which point training stops. Since the initial training set has a small sample size and does not reach the preset threshold, new unlabeled samples (samples without masking data) are acquired. The MRI images of the newly added unlabeled samples are then segmented into three planes according to the three segmentation models obtained in step S12. Senior radiologists review and correct the prediction results, obtaining the corrected segmentation results for the newly added unlabeled samples. These newly added unlabeled samples and their corresponding corrected results are added to the initial training set, forming a high-quality initial training set of 100 samples. Step S12 is then returned to train the three segmentation models again. If the number of samples in the training set reaches the preset threshold, the three segmentation models are successfully trained. Otherwise, new unlabeled samples are added, and incremental training continues until the three segmentation models are successfully trained. This incremental training method continuously iterates and expands the dataset, optimizes model performance, effectively reduces data annotation costs, improves data construction efficiency, and reduces generalization problems caused by inconsistent doctor annotations.

[0064] Specifically, the training loss for each segmentation model is calculated using the following formula:

[0065]

[0066] Where λ1, λ2, and λ3 represent weights. Indicates DICE loss, Indicates focal loss. This indicates boundary loss.

[0067] Specifically, the DICE loss is calculated using the following formula:

[0068]

[0069] Where n represents the number of pixels in the MRI image, y i p represents the uterine mask value of the i-th pixel. i This represents the probability that the i-th pixel predicted by the model is in the uterine region, and ∈ represents the smoothing coefficient to prevent the denominator from being zero.

[0070] Specifically, the focus loss is calculated using the following formula:

[0071]

[0072] Among them, y i p represents the uterine mask value of the i-th pixel. i α represents the probability that the i-th pixel predicted by the model is in the uterine region, α represents the class weight, and γ represents the focusing parameter, which is used to reduce the loss contribution of easily separable parts.

[0073] To improve the clarity of boundary segmentation, this invention adds a boundary loss to the DICE and Focal losses, taking into account the loss of edge information, thereby improving the model's performance on boundaries.

[0074] Specifically, the boundary loss is calculated using the following formula:

[0075]

[0076] in, This represents the gradient magnitude indicating the probability that the model predicts the i-th pixel to be in the uterine region. denoted as the gradient magnitude of the uterine mask at the i-th pixel, and n represents the number of pixels in the MRI image.

[0077] Specifically, the gradient magnitude of the probability that the i-th pixel is in the uterine region is calculated in the following way:

[0078] The Sobel operator is used to convolve the probability map predicted by the model in the horizontal and vertical directions respectively to obtain the horizontal gradient map and the vertical gradient map.

[0079] Using formula Calculate the gradient magnitude of the probability that the i-th pixel is in the uterine region, where This represents the value of the i-th pixel in the horizontal gradient map. This represents the value of the i-th pixel in the vertical gradient map.

[0080] In practice, the Sobel operator in the horizontal direction is: The Sobel operator in the vertical direction is:

[0081] Since the uterine mask is a two-dimensional value (0 or 1), directly calculating the gradient will result in noise due to the discrete values. In practice, Gaussian blurring can be applied to the uterine mask first to generate smooth values. The gradient can then be calculated using the Sobel operator on the Gaussian blurred uterine mask, in the same way as above.

[0082] After obtaining multiple trained segmentation models, the multiplanar MRI images of different individuals are segmented and the multiplanar uterine ROI of each individual is extracted based on the multiple trained segmentation models. In practice, different individuals may be samples that are not in the initial training set.

[0083] Based on multiple segmentation models, the uterine region of each individual can be segmented into multiple scanning plane MRI images. The minimum bounding cube of each uterine region is determined and cropped as the uterine ROI.

[0084] During implementation, the uterine ROI region can be resampled and adjusted to a standard pixel size, such as 128×128×128 pixels, and then normalized. Each individual can obtain three standard uterine cubes (sagittal, coronal, and axial views). The multiplanar uterine ROI and the corresponding adenomyosis classification are used as samples in the second training set, thus constructing the second training set.

[0085] To improve the accuracy of classification, the deep learning model constructed in this invention includes:

[0086] The feature preprocessing module includes multiple branches, each of which is used to extract features from the corresponding uterine ROI in the plane to obtain a feature map;

[0087] The attention module is used to calculate attention features from feature maps based on a cross-plane spatial-channel hybrid mechanism.

[0088] The classification module is used for predicting adenomyosis subtypes based on attention features.

[0089] In implementation, the feature preprocessing module may include three branches, each corresponding to a scanning plane, and each branch may adopt a CNN structure.

[0090] The three branches have the same structure. Each branch goes through a convolutional layer (feature extraction), batch normalization (feature standardization), ReLU activation function (introducing nonlinearity), and pooling layer (feature dimensionality reduction), and outputs a high-dimensional feature representation (feature map) of its own plane, which serves as the input to the next module.

[0091] Since the Kishi classification of adenomyosis is essentially defined based on the spatial location and extent of infiltration of the lesion within the uterine region, it requires reliance on two core spatial elements: the location and morphology of the lesion, and the extent and spatial relationship of the lesion within the uterine anatomy. This invention improves classification accuracy by integrating key spatial and channel information through an attention mechanism.

[0092] Specifically, the attention module calculates attention features in the following way:

[0093] Each feature map is converted into a one-dimensional feature sequence. The position code of each element in the sequence is calculated. The position code is concatenated with the feature of the element to obtain the feature sequence of fused spatial location.

[0094] The attention features are obtained by splicing the feature sequences of the fused spatial locations and calculating the spatial-channel hybrid attention of the spliced ​​feature sequences.

[0095] In practice, to enhance the model's spatial awareness, the feature maps are positionally encoded.

[0096] During implementation, the feature maps of each plane are transformed into sequence vectors (one-dimensional feature sequences). The three feature sequences are represented as sagittal feature sequence (FSag), coronal feature sequence (FCor), and axial feature sequence (FAx), respectively, as the input sequences of the Transformer.

[0097] To enhance the model's ability to accurately perceive the spatial location of features in 3D images, this invention introduces a 3D position encoding method. Specifically, the position encoding of the i-th element in a one-dimensional feature sequence is calculated using the following formula:

[0098]

[0099] Where (x,y,z) represents the three-dimensional coordinates of the i-th element in the feature map, and d model Indicates the length of the feature sequence.

[0100] By concatenating the positional code with the features of elements in the sequence, that is, by incorporating the positional code as part of the element's features, we obtain a feature sequence of three fused spatial locations. and

[0101] Concatenate the three feature sequences Attention features are calculated based on a spatial-channel hybrid attention mechanism.

[0102] Specifically, the formula for calculating attention features is as follows:

[0103]

[0104] Channel Attenion (F concat )=σ(FC(ReLU(FC(AvgPool(F concat ))))+FC(ReLu(FC(MaxPool(F concat )))))

[0105] Where AvgPool(·) represents global average pooling in the spatial dimension, MaxPool(·) represents global max pooling in the spatial dimension, FC(·) represents a fully connected layer used to model the relationships between channels, and σ(·) represents the Sigmoid activation function. This represents bitwise multiplication, used for channel-weighted spatial attention, and outputs attention features. Attention(F) concat ) represents self-attention.

[0106]

[0107] Where Q represents Query, K represents key, V represents value, and d k This represents the dimension of K.

[0108] While maintaining spatial awareness, the attention module enhances the model's ability to express and fuse key features, strengthens the correlation modeling between features in different planes, and effectively improves the classification accuracy and the robustness of the model.

[0109] The obtained attention features are input into the classification module to complete the Kishi subtyping prediction task for adenomyosis. A fully connected classification layer is used to further map the fused feature vectors to the specific classification task of adenomyosis, and the probability distribution of the predicted category is output through a softmax function. The cross-entropy loss function is used as a supervision signal to optimize the model's classification performance. During training, the Adam classifier is preferentially selected, with an initial learning rate set to 0.001 and 50–100 iterations.

[0110] After obtaining the adenomyosis classification model, the multiplanar MRI images of the individual to be classified are input into multiple trained segmentation models to obtain the multiplanar uterine ROI of the individual to be classified. Inputting the multiplanar uterine ROI of the individual to be classified into the adenomyosis classification model can quickly and accurately obtain the adenomyosis classification of the individual. This is close to the reading logic and workflow of clinicians, has good engineering feasibility and clinical adaptability, and helps to promote the practical application of computer-aided diagnostic technology in the field of gynecological imaging.

[0111] A specific embodiment of the present invention discloses an automatic classification system for adenomyosis based on multiplanar MRI images, such as... Figure 2 As shown, it includes:

[0112] The second training set construction module is used to segment and extract the multiplanar uterine ROI of each individual based on multiple trained segmentation models, and to construct the second training set based on the multiplanar uterine ROI of each individual and the corresponding adenomyosis classification.

[0113] The adenomyosis classification model training module is used to construct a deep learning model for adenomyosis classification, and the deep learning model is trained based on the second training to obtain the adenomyosis classification model.

[0114] The classification prediction module is used to input multiplanar MRI images of an individual to be classified into multiple pre-trained segmentation models to obtain the multiplanar uterine ROI of the individual to be classified; and to input the multiplanar uterine ROI of the individual to be classified into the adenomyosis classification model to obtain the adenomyosis classification of the individual to be classified.

[0115] The above-described method and system embodiments are based on the same principles, and their related aspects can be referenced from each other to achieve the same technical effects. For specific implementation processes, please refer to the foregoing embodiments, which will not be repeated here.

[0116] Electronic device example:

[0117] One specific embodiment of this application discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the automatic classification method for adenomyosis based on multiplanar MRI images in the method embodiment.

[0118] Examples of readable storage media:

[0119] One specific embodiment of this application discloses a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the automatic classification method for adenomyosis based on multiplanar MRI images in the method embodiment.

[0120] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0121] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An automatic classification method of adenomyosis based on multi-planar MRI images, characterized in that, The method comprises the following steps: Based on the trained multiple segmentation models, the multi-planar MRI images of different individuals are segmented to extract the multi-planar uterus ROIs of each individual, and a second training set is constructed based on the multi-planar uterus ROIs of each individual and the corresponding adenomyosis classification; A deep learning model for adenomyosis classification is constructed, and the deep learning model is trained based on the second training set to obtain an adenomyosis classification model; The multi-planar MRI images of the individual to be classified are input into the trained multiple segmentation models to obtain the multi-planar uterus ROIs of the individual to be classified; and the multi-planar uterus ROIs of the individual to be classified are input into the adenomyosis classification model to obtain the adenomyosis classification of the individual to be classified. The trained multiple segmentation models are obtained in the following manner: S11, an initial training set is constructed, each sample in the initial training set comprising multi-planar MRI images of an individual and a corresponding uterus mask; a plurality of segmentation models are constructed; each MRI scanning plane corresponds to a segmentation model; S12, the multiple segmentation models are trained based on the initial training set; S13, if the number of samples in the initial sample set reaches a preset threshold, the trained multiple segmentation models are obtained; otherwise, a newly added unlabeled sample is obtained, and the newly added unlabeled sample is segmented based on the multiple segmentation models; S14, a correction result of the segmentation result of the newly added unlabeled sample is obtained, and the newly added unlabeled sample and the corresponding correction result are added to the initial training set; Return to step S12.

2. The method of claim 1, wherein the method further comprises: The training loss of each segmentation model is calculated using the following formula: ； wherein, and denotes a weight, denotes a DICE loss, denotes a focal loss, denotes a boundary loss.

3. The method of claim 2, wherein the step of automatically classifying the adenomyosis is performed by using a deep learning model trained to classify the adenomyosis based on the multi-planar MRI images. The boundary loss is calculated using the following formula: ； wherein, denotes a gradient magnitude of the probability that the i-th pixel point is a uterus region predicted by the model, denotes a gradient magnitude of the uterus mask of the i-th pixel point, and n denotes a number of pixels of the MRI image.

4. The method of claim 3, wherein the step of automatically classifying the adenomyosis is performed by using a machine learning algorithm. The gradient amplitude of the probability that the i-th pixel point is in the uterus region is calculated in the following manner: The probability map predicted by the model is respectively convolved in the horizontal direction and the vertical direction using Sobel operators to obtain a horizontal gradient map and a vertical gradient map; The gradient magnitude of the probability that the ith pixel point is in the uterus region is calculated using the formula wherein denotes the value of the ith pixel point in the horizontal gradient map, denotes the value of the ith pixel point in the vertical gradient map.

5. The method of claim 1, wherein the method further comprises: The deep learning model comprises: A feature preprocessing module comprising a plurality of branches, each branch being configured to extract features from a corresponding planar uterus ROI to obtain a feature map; An attention module configured to calculate attention features from the feature map based on a cross-plane spatial-channel hybrid mechanism; A classification module configured to predict adenomyosis classification based on the attention features.

6. The method of claim 5, wherein the step of automatically classifying the adenomyosis is performed by using a deep learning model trained to classify the adenomyosis based on the multi-planar MRI images. The attention module calculates the attention features in the following manner: Each feature map is converted into a one-dimensional feature sequence, the position encoding of each element in the sequence is calculated, the position encoding is spliced with the features of the element to obtain a feature sequence fused with spatial positions; The feature sequence fused with spatial positions is spliced, and spatial-channel hybrid attention of the spliced feature sequence is calculated to obtain the attention features.

7. The method of claim 6, wherein the step of automatically classifying the adenomyosis is performed by using a deep learning model trained to classify the adenomyosis based on the multi-planar MRI images. The position encoding of the i-th element in the one-dimensional feature sequence is calculated using the following formula: ； where (x, y, z) represents the three-dimensional coordinates of the i-th element in the feature map, denotes the length of the feature sequence.

8. The method of claim 6, wherein the step of automatically classifying the adenomyosis is performed by using a deep learning model trained to classify the adenomyosis based on the multi-planar MRI images. The attention feature calculation formula is represented as: ； wherein, denotes spatial dimension global average pooling, denotes spatial dimension global max pooling, FC(·) denotes a fully connected layer for modeling the interrelation between channels, denotes a Sigmoid activation function, denotes a bit-wise multiplication for channel attention weighting spatial attention, output attention features, denotes self-attention.

9. An automatic classification system for adenomyosis based on multi-planar MRI images, characterized by, It comprises: A second training set construction module configured to segment the multi-planar MRI images of different individuals based on the trained multiple segmentation models to extract the multi-planar uterus ROIs of each individual, and to construct a second training set based on the multi-planar uterus ROIs of each individual and the corresponding adenomyosis classification; The adenomyosis typing model training module is configured to construct a deep learning model for adenomyosis typing, and train the deep learning model based on the second training set to obtain an adenomyosis typing model. The typing prediction module is configured to input the multi-plane MRI image of the individual to be typed into the trained segmentation models to obtain the multi-plane uterus ROI of the individual to be typed, and input the multi-plane uterus ROI of the individual to be typed into the adenomyosis typing model to obtain the adenomyosis typing of the individual to be typed. The trained segmentation models are obtained in the following manner: S11, constructing an initial training set, each sample in the initial training set including the multi-plane MRI image of an individual and a corresponding uterus mask, and constructing a plurality of segmentation models, each MRI scanning plane corresponding to a segmentation model; S12, training the plurality of segmentation models based on the initial training set; S13, if the number of samples in the initial sample set reaches a preset threshold, the trained plurality of segmentation models are obtained; otherwise, obtaining new unlabeled samples, and segmenting the new unlabeled samples based on the plurality of segmentation models; S14, obtaining the correction result of the segmentation result of the new unlabeled samples, and adding the new unlabeled samples and the corresponding correction result to the initial training set; returning to step S12.

Images