Nuclear magnetic image-based endometrial cancer classification diagnosis system

Abstract

The invention provides a nuclear magnetic image-based endometrial cancer classification and diagnosis system, and the working principle is that the advantages of artificial intelligence in image recognition and visual interpretation are utilized, a deep learning theory is adopted, a picture sample generated by clinical nuclear magnetic resonance examination is taken as a data basis, and image data are sliced, labeled and transformed to obtain an endometrial cancer classification and diagnosis result. A region of interest is selected from the data, and stage and classification diagnosis of the endometrial cancer is realized through a deep learning neural network model; therefore, the method does not need to perform an operation on the patient, can perform stage diagnosis on the endometrial cancer of the patient in a non-invasive manner, has the advantages of being non-invasive, accurate and the like, and can assist a doctor to perform professional and accurate diagnosis on the endometrial cancer in a classified manner; meanwhile, compared with common doctors, the method has higher diagnosis accuracy and higher diagnosis speed, can assist primary doctors in providing diagnosis opinions, and has important significance on treatment of endometrial cancer, improvement of the survival rate of patients and reduction of public health burden.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision detection technology such as intelligent medical auxiliary diagnosis, image processing, and deep learning, and in particular relates to an endometrial cancer classification diagnosis system based on nuclear magnetic resonance images. Background Art

[0002] For the diagnosis of endometrial cancer, the commonly used diagnostic methods in clinical practice are: magnetic resonance imaging, transvaginal ultrasound, computed tomography, hysteroscopy, histopathological sections, etc. As a non-invasive examination method, nuclear magnetic resonance has good soft tissue contrast resolution, so it is considered to be the best imaging examination method for preoperative endometrial cancer staging assessment. However, due to the low resolution limit of magnetic resonance images, the diagnostic results are greatly affected by the subjective influence of physicians, and there are few physicians with rich experience, so accurate diagnosis in clinical practice is still difficult.

[0003] In view of the above problems, the present invention proposes a method for staging and classification diagnosis of endometrial cancer based on nuclear magnetic resonance images, which provides a solution with practical medical diagnostic significance and assists doctors in staging and classification diagnosis of endometrial cancer. Summary of the invention

[0004] To solve the above problems, the present invention provides an endometrial cancer classification diagnosis system based on nuclear magnetic resonance images, which can assist doctors in staging and classification diagnosis of endometrial cancer, thereby effectively alleviating the pressure of primary clinical diagnosis, and is of great significance for the treatment of endometrial cancer, improving the survival rate of patients, and reducing the burden on public health.

[0005] A classification and diagnosis system for endometrial cancer based on nuclear magnetic resonance images, comprising a preprocessing module, a uterine region segmentation module, and a multimodal classification module, wherein the multimodal classification module comprises a two-dimensional model unit, a three-dimensional model unit, a fusion unit, and a classification unit;

[0006] The preprocessing module is used to convert the abdominal magnetic resonance image into three-dimensional data of a set standard; wherein the three-dimensional data includes multiple layers of two-dimensional plane data, and the layers of the three-dimensional data are divided along the height direction of the human body;

[0007] The uterus region segmentation module is used to select the two-dimensional plane data of the layer where the endometrial cancer lesion area is located from the three-dimensional data as the key layer, and to extract the minimum bounding box containing the endometrial cancer lesion area from all the key layers;

[0008] The two-dimensional model unit takes the two-dimensional plane data of each layer in the minimum bounding box as input, thereby extracting the intra-layer features of the two-dimensional plane data of each layer in the minimum bounding box;

[0009] The three-dimensional model unit uses the three-dimensional data corresponding to the minimum bounding box as input to extract the inter-layer features of the two-dimensional plane data of each layer in the minimum bounding box;

[0010] The fusion unit is used to linearly weight the intra-layer features and the inter-layer features according to the dimension of the number of layers to obtain multimodal features;

[0011] The classification unit is used to stage endometrial cancer lesions based on multimodal features.

[0012] Furthermore, the key layer positioning method is:

[0013] The three-dimensional data is input into the trained YOLO positioning model, and the YOLO positioning model selects the layer number of the layer where the endometrial cancer lesion area is located. The loss function L used in the training of the YOLO positioning model is loc as follows:

[0014]

[0015] Among them, λ coord is the set loss weight, M is the number of key layers, i is the key layer number, (x i ,y i ) is the actual coordinate of the center point of the i-th key layer, (w i ,h i ) is the actual width and height of the i-th key layer, is the predicted coordinate of the center point of the i-th key layer output by the YOLO positioning model, is the predicted width and height of the i-th key layer output by the YOLO positioning model.

[0016] Furthermore, the method for extracting the minimum bounding box is:

[0017] Each key layer is input into the trained YOLO segmentation model in turn, and the YOLO segmentation model extracts the minimum envelope of each key layer containing the endometrial cancer lesion area. The minimum envelope of all key layers is combined to obtain the minimum bounding box containing the endometrial cancer lesion area. The loss function L used in the training of the YOLO segmentation model is seg as follows:

[0018]

[0019] Among them, H is the height of the key layer, W is the width of the key layer, and y i,j is the true label of the pixel in the i-th row and j-th column in the key layer, where y i,j =1 means that the pixel point in the i-th row and j-th column belongs to the endometrial cancer lesion area, i,j= 0 means that the pixel point in the i-th row and j-th column does not belong to the endometrial cancer lesion area, P i,j is the probability that the pixel in the i-th row and j-th column in the key layer output by the YOLO segmentation model belongs to the endometrial cancer lesion area.

[0020] Furthermore, the two-dimensional model unit adopts a 4-layer improved Unet network to extract intra-layer features of each layer of two-dimensional plane data in the minimum bounding box; wherein, based on the original Unet network, the improved Unet network adds a Tansformer self-attention subunit between the encoder and the decoder, and adds a fine-grained feature extraction subunit in each convolution layer, wherein the fine-grained feature extraction subunit is composed of sequentially cascaded small receptive field convolution layers, multi-scale convolution layers and residual connections.

[0021] Furthermore, the three-dimensional model unit adopts a 4-layer improved Vnet network to extract inter-layer features of each layer of two-dimensional plane data in the minimum bounding box; wherein, the improved Vnet network adds 8 stacked self-attention layers between the encoder and the decoder on the basis of the original Vnet network, and each self-attention layer is composed of sequentially cascaded Transformer self-attention sub-units, multi-head attention mechanism, feedforward neural network and residual connection.

[0022] Furthermore, the loss function L used when training the three-dimensional model unit is correlation for:

[0023]

[0024] Wherein, n=2, ..., N-1, n is the layer number of each layer of two-dimensional plane data in the three-dimensional data corresponding to the minimum bounding box, and N is the total number of layers of three-dimensional data corresponding to the minimum bounding box. is the eigenvalue at the position (x, y) of the two-dimensional plane data of the n-1th layer, is the average eigenvalue of the n-1th layer of two-dimensional plane data, is the eigenvalue at the position (x, y) of the n-th layer of two-dimensional plane data, is the average eigenvalue of the n-th layer of two-dimensional plane data, is the eigenvalue at the position (x, y) of the two-dimensional plane data of the n+1th layer, is the average eigenvalue of the two-dimensional plane data of the n+1th layer.

[0025] Furthermore, the eigenvalue at any position (x, y) of the nth layer of two-dimensional plane data in the three-dimensional data corresponding to the minimum bounding box is is the conditional probability P(D) that the pixel at the position (x, y) of the n-th layer of two-dimensional plane data belongs to the endometrial cancer lesion area n(x,y)|(D n-1 (x,y),D n+1 (x,y)));

[0026] Conditional probability P(D n (x,y)|(D n-1 (x,y),D n+1 The calculation method of (x,y))) is as follows:

[0027]

[0028] Among them, D n (x, y) is the probability that the pixel at the position (x, y) of the n-th layer of two-dimensional plane data belongs to the endometrial cancer lesion area, D n-1 (x, y) is the probability that the pixel at the position (x, y) of the n-1th layer of two-dimensional plane data belongs to the endometrial cancer lesion area, D n+1 (x, y) is the probability that the pixel point at the position (x, y) of the n+1th layer of two-dimensional plane data belongs to the endometrial cancer lesion area, α is the set proportional coefficient, and Z is the set normalization constant.

[0029] Beneficial effects:

[0030] The present invention provides an endometrial cancer classification diagnosis system based on nuclear magnetic resonance images. The working principle is to utilize the advantages of artificial intelligence in image recognition and visual interpretation, adopt deep learning theory, use picture samples generated by clinical nuclear magnetic resonance examination as data basis, slice, annotate and transform the image data, select regions of interest from the data, and use a deep learning neural network model to diagnose the staging classification of endometrial cancer. It can be seen that the present invention does not require surgery on the patient, and can non-invasively diagnose the patient's endometrial cancer staging. It has the advantages of being non-invasive and accurate, and can assist doctors in making professional and accurate diagnosis of endometrial cancer classification. At the same time, compared with ordinary physicians, the present invention has higher diagnostic accuracy and faster diagnostic speed, can assist grassroots doctors in providing diagnostic opinions, can effectively alleviate the pressure of grassroots clinical diagnosis, and is of great significance to the treatment of endometrial cancer, improving the survival rate of patients, and reducing the burden of public health. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the endometrial cancer classification diagnosis system based on nuclear magnetic resonance images of the present invention.

[0032] Figure 2 It is a schematic diagram of the uterus region segmentation model of the present invention.

[0033] Figure 3 It is a schematic diagram of the multimodal diagnostic model for endometrial cancer of the present invention. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0035] like Figure 1 As shown, a classification and diagnosis system for endometrial cancer based on nuclear magnetic resonance images includes a preprocessing module, a uterine region segmentation module, and a multimodal classification module, wherein the multimodal classification module includes a two-dimensional model unit, a three-dimensional model unit, a fusion unit, and a classification unit;

[0036] The preprocessing module is used to convert the abdominal magnetic resonance images into three-dimensional data of set standards; wherein the three-dimensional data includes multiple layers of two-dimensional plane data, and the layers of the three-dimensional data are divided along the height direction of the human body; for example, the first layer is close to the chest cavity, and the last layer is close to the legs; the preprocessing methods of the abdominal magnetic resonance images include but are not limited to data slicing, data annotation, data screening and data transformation methods.

[0037] The uterine region segmentation module is used to filter out the two-dimensional plane data of the layer where the endometrial cancer lesion area is located from the three-dimensional stereo data as the key layer, and to extract the minimum bounding box containing the endometrial cancer lesion area from all key layers; wherein the segmentation method adopted by the uterine region segmentation module includes but is not limited to a deep learning neural network algorithm based on the YOLO backbone network.

[0038] The two-dimensional model unit takes the two-dimensional plane data of each layer in the minimum bounding box as input, thereby extracting the intra-layer features of the two-dimensional plane data of each layer in the minimum bounding box;

[0039] The three-dimensional model unit uses the three-dimensional data corresponding to the minimum bounding box as input to extract the inter-layer features of the two-dimensional plane data of each layer in the minimum bounding box;

[0040] The fusion unit is used to linearly weight the intra-layer features and the inter-layer features according to the dimension of the number of layers to obtain multimodal features;

[0041] The classification unit is used to stage endometrial cancer lesions according to multimodal features, such as stage I (IA, IB), stage II (IIA, IIB, IIC), stage III (IIIA, IIIC) and stage IV; wherein the multimodal diagnosis model of endometrial cancer adopted by the classification unit includes but is not limited to a deep learning neural network algorithm based on a Transformer backbone network, and the diagnosis model uses a large amount of endometrial cancer magnetic resonance imaging data and annotated labels as training samples, and compares them with the actual diagnosis results of doctors to verify the diagnostic ability of the model.

[0042] It should be noted that if Figure 2 As shown in the figure, since the collected MRI data is the data of the patient's entire abdominal cavity, there is interference from other organs, and the upper and lower layers close to the end do not contain the uterus, and all layers containing the uterus do not all contain lesions. Therefore, it is necessary to screen the layers containing the lesion area as the key layer, and exclude the interference of the good uterus and other organs in the abdominal cavity in subsequent identification through preliminary screening. As long as the lesion area is identified, the layer is considered to be the key layer.

[0043] Furthermore, the key layer positioning method is:

[0044] The three-dimensional data is input into the trained YOLO positioning model, and the YOLO positioning model selects the layer number of the layer where the endometrial cancer lesion area is located. The loss function L used in the training of the YOLO positioning model is loc as follows:

[0045]

[0046] Among them, λ coord is the set loss weight, M is the number of key layers, i is the key layer number, (x i ,y i ) is the actual coordinate of the center point of the i-th key layer, (w i ,h i ) is the actual width and height of the i-th key layer, is the predicted coordinate of the center point of the i-th key layer output by the YOLO positioning model, is the predicted width and predicted height of the i-th key layer output by the YOLO positioning model.

[0047] Furthermore, the method for extracting the minimum bounding box is:

[0048] Each key layer is input into the trained YOLO segmentation model in turn, and the YOLO segmentation model extracts the minimum envelope of each key layer containing the endometrial cancer lesion area. The minimum envelope of all key layers is combined to obtain the minimum bounding box containing the endometrial cancer lesion area. The loss function L used in the training of the YOLO segmentation model is seg as follows:

[0049]

[0050] Among them, H is the height of the key layer, W is the width of the key layer, and y i,j is the true label of the pixel in the i-th row and j-th column in the key layer, where y i,j =1 means that the pixel point in the i-th row and j-th column belongs to the endometrial cancer lesion area, i,j= 0 means that the pixel point in the i-th row and j-th column does not belong to the endometrial cancer lesion area, P i,j is the probability that the pixel in the i-th row and j-th column in the key layer output by the YOLO segmentation model belongs to the endometrial cancer lesion area.

[0051] Furthermore, if Figure 3 As shown, the two-dimensional model unit uses a 4-layer improved Unet network to extract the intra-layer features of each layer of two-dimensional plane data in the minimum bounding box; wherein, based on the original Unet network, the improved Unet network adds a Tansformer self-attention subunit between the encoder and the decoder, and adds a fine-grained feature extraction subunit in each convolution layer, wherein the fine-grained feature extraction subunit is composed of a sequentially cascaded small receptive field convolution layer, a multi-scale convolution layer and a residual connection, and the specific functions of the fine-grained feature extraction subunit are as follows:

[0052] Small receptive field convolution: Use a 3*3 convolution kernel to capture details.

[0053] Multi-scale convolution layer: 1*1, 3*3, 5*5 and other convolution kernels are added in parallel to capture details at different scales.

[0054] Residual connection: Perform residual connection on the input features to retain the original information and improve stability.

[0055] Furthermore, the three-dimensional model unit uses a 4-layer improved Vnet network to extract inter-layer features of each layer of two-dimensional plane data in the minimum bounding box; wherein, the improved Vnet network adds 8 stacked self-attention layers between the encoder and the decoder on the basis of the original Vnet network, and each self-attention layer is composed of sequentially cascaded Transformer self-attention subunits, multi-head attention mechanisms, feedforward neural networks, and residual connections. Specifically, the encoded sequence is input to the first Transformer self-attention subunit, and the global context information is extracted through the multi-head self-attention mechanism, and then the features are enhanced through the feedforward neural network; the sequence data processed by the 8-layer Transformer is reshaped back to a 3D shape to continue the subsequent decoding process.

[0056] The loss function L used when training the three-dimensional model unit correlation for:

[0057]

[0058] Wherein, n=2, ..., N-1, n is the layer number of each layer of two-dimensional plane data in the three-dimensional data corresponding to the minimum bounding box, and N is the total number of layers of three-dimensional data corresponding to the minimum bounding box. is the eigenvalue at the position (x, y) of the two-dimensional plane data of the n-1th layer, is the average eigenvalue of the n-1th layer of two-dimensional plane data, is the eigenvalue at the position (x, y) of the n-th layer of two-dimensional plane data, is the average eigenvalue of the n-th layer of two-dimensional plane data, is the eigenvalue at the position (x, y) of the two-dimensional plane data of the n+1th layer, is the average eigenvalue of the two-dimensional plane data of the n+1th layer.

[0059] It should be noted that the numerator The sum of the products of the two layers of two-dimensional plane data at each position after removing the mean is calculated to quantify their relationship; the denominator is used to normalize the numerator and constrain the correlation value to be in the range of [-1,1][-1,1][-1,1].

[0060] At the same time, the eigenvalue at any position (x, y) of the nth layer of two-dimensional plane data in the three-dimensional data corresponding to the minimum bounding box is the conditional probability P(D) that the pixel at the position (x, y) of the n-th layer of two-dimensional plane data belongs to the endometrial cancer lesion area n (x,y)|(D n-1 (x,y),D n+1 (x,y)));

[0061] Conditional probability P(D n (x,y)|(D n-1 (x,y),D n+1 The calculation method of (x,y))) is as follows:

[0062]

[0063] Among them, D n (x, y) is the probability that the pixel at the position (x, y) of the n-th layer of two-dimensional plane data belongs to the endometrial cancer lesion area, D n-1 (x, y) is the probability that the pixel at the position (x, y) of the n-1th layer of two-dimensional plane data belongs to the endometrial cancer lesion area, D n+1 (x, y) is the probability that the pixel at the position (x, y) of the n+1th layer of two-dimensional plane data belongs to the endometrial cancer lesion area, α is the set proportional coefficient, and Z is the set normalization constant to ensure that the total probability is 1.

[0064] That is, P(D n (x,y)|(D n-1 (x,y),D n+1(x,y))) represents the conditional probability that the position (x,y) of the n-th layer of two-dimensional plane data belongs to the lesion area under the condition of the lesion status of the n-1th layer and the n+1th layer; the exponential function in the calculation formula of the conditional probability is used to map the function value to a positive number to ensure that the probability is non-negative and can effectively express the impact of position differences on the probability; α is a hyperparameter that controls the impact of the difference in lesion status of adjacent layers on the lesion probability of the n-th layer of two-dimensional plane data, which is determined by network training. A larger α will increase the sensitivity of the lesion area change to the probability of the current layer; |D n-1 (x,y)-D n+1 (x,y)| represents the difference in the lesion status of adjacent layers at the same position. Assuming that the lesion area is continuous, if D n-1 (x,y) and D n+1 If the difference between (x, y) is large, the probability that the position (x, y) is the lesion area in the nth layer will decrease.

[0065] For example, suppose that the collected three-dimensional stereo data is processed into three-dimensional standard size data of 512*512*32, where 512*512 represents the resolution of each sampling scanning layer, and 32 means that for the abdominal cavity of a patient, a total of 32 slice samples were collected; since the lesion areas are distributed in different layers, and the lesion areas between adjacent layers are positionally associated, the inter-layer feature refers to the lesion position association information between slice samples of different layers, and the intra-layer feature refers to the position information of the lesion area in a single slice.

[0066] That is to say, since the tumor tissue is continuous, the tumor area detected in each layer will not change suddenly, but will continue to grow or shrink. Therefore, position association means that if the (x, y) pixel point of the nth layer is identified as a lesion area, then the position and its adjacent positions (such as (x±5, y±5)) of the n+1th layer are more likely to be tumors. Of course, since the shape of the tumor will change continuously between layers, the judgment of the n+1th layer should be related to both the nth layer and the n+2th layer.

[0067] Furthermore, the fusion unit linearly weights the corresponding positions of the output results of the two-dimensional model unit and the three-dimensional model unit in the three-dimensional scale, and the weighting ratio is obtained according to the training results. Stereoscopic image data refers to the results obtained by the training of the three-dimensional model unit, because the output result of the model is a three-dimensional lesion position segmentation array, so it is called stereoscopic data. Single-layer image data refers to the results obtained by the training of the two-dimensional model unit, because the output result of the model for each layer is a two-dimensional array. For a case patient, multiple two-dimensional arrays will be output (the number is equal to the number of layers, that is, there are also 32 two-dimensional arrays). Multimodal features refer to the dual features of two-dimensional and three-dimensional. To be precise, they are not converted from associated information and fine-grained features, but are composed of associated information and fine-grained feature recognition results. Since the data dimensions of the two features are different, they are called multimodal features.

[0068] It should be noted that due to the different equipment in different hospitals, the single-layer resolution and number of data layers of the collected MRI data are different. The standard value is 512*512*32. Based on this, the method used by the preprocessing module to convert the patient's abdominal MRI image into a set standard three-dimensional stereo data is:

[0069] Determine whether the collected MRI data is larger than the standard value of 512*512*32. If the data size is larger than the standard value, use its geometric center as the origin and perform center cropping. If the data size is smaller than the standard value, use its geometric center as the origin and fill zeros around the data to unify the data format of the input network model for subsequent training.

[0070] In summary, the present invention proposes a system for assisting diagnosis of the staging classification of endometrial cancer of patients based on nuclear magnetic resonance data; the present invention first scans the abdominal cavity of the patient through the nuclear magnetic resonance instrument of the hospital to obtain the nuclear magnetic resonance data of the abdominal cavity of the patient; after pre-processing the data, the data is input into the uterine region segmentation model, and the image layer where the uterus containing the lesion area is located is identified by the model, and the uterine region in the layer is segmented; finally, the segmented uterine key layer is input into the multimodal classification model of endometrial cancer, and the endometrial cancer is accurately diagnosed by combining two-dimensional and three-dimensional multimodal data, so as to assist doctors in staging endometrial cancer tumors. This invention can realize non-invasive and accurate diagnosis of endometrial cancer, can effectively alleviate medical pressure, and assist doctors in diagnosis.

[0071] Of course, the present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may certainly make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A classification and diagnosis system for endometrial cancer based on nuclear magnetic resonance images, characterized in that: It includes a preprocessing module, a uterine region segmentation module, and a multimodal classification module, wherein the multimodal classification module includes a two-dimensional model unit, a three-dimensional model unit, a fusion unit, and a classification unit; The preprocessing module is used to convert the abdominal magnetic resonance image into three-dimensional data of a set standard; wherein the three-dimensional data includes multiple layers of two-dimensional plane data, and the layers of the three-dimensional data are divided along the height direction of the human body; The uterus region segmentation module is used to select the two-dimensional plane data of the layer where the endometrial cancer lesion area is located from the three-dimensional data as the key layer, and to extract the minimum bounding box containing the endometrial cancer lesion area from all the key layers; The two-dimensional model unit takes the two-dimensional plane data of each layer in the minimum bounding box as input, thereby extracting the intra-layer features of the two-dimensional plane data of each layer in the minimum bounding box; The three-dimensional model unit uses the three-dimensional data corresponding to the minimum bounding box as input to extract the inter-layer features of the two-dimensional plane data of each layer in the minimum bounding box; The fusion unit is used to linearly weight the intra-layer features and the inter-layer features according to the dimension of the number of layers to obtain multimodal features; The classification unit is used to stage endometrial cancer lesions based on multimodal features.

2. The endometrial cancer classification diagnosis system based on nuclear magnetic resonance images as claimed in claim 1, characterized in that: The positioning method of the key layer is: The three-dimensional data is input into the trained YOLO positioning model, and the YOLO positioning model selects the layer number of the layer where the endometrial cancer lesion area is located. The loss function L used in the training of the YOLO positioning model is loc as follows: Among them, λ coord is the set loss weight, M is the number of key layers, i is the key layer number, (x i ,y i ) is the actual coordinate of the center point of the i-th key layer, (w i ,h i ) is the actual width and actual height of the i-th key layer, is the predicted coordinate of the center point of the i-th key layer output by the YOLO positioning model, is the predicted width and height of the i-th key layer output by the YOLO positioning model.

3. The endometrial cancer classification diagnosis system based on nuclear magnetic resonance images as claimed in claim 1, characterized in that: The method for extracting the minimum bounding box is as follows: Each key layer is input into the trained YOLO segmentation model in turn, and the YOLO segmentation model extracts the minimum envelope of each key layer containing the endometrial cancer lesion area. The minimum envelope of all key layers is combined to obtain the minimum bounding box containing the endometrial cancer lesion area. The loss function L used in the training of the YOLO segmentation model is seg as follows: Among them, H is the height of the key layer, W is the width of the key layer, and y i,j is the true label of the pixel in the i-th row and j-th column in the key layer, where y i,j =1 means that the pixel point in the i-th row and j-th column belongs to the endometrial cancer lesion area, i,j = 0 means that the pixel point in the i-th row and j-th column does not belong to the endometrial cancer lesion area, P i,j is the probability that the pixel in the i-th row and j-th column in the key layer output by the YOLO segmentation model belongs to the endometrial cancer lesion area.

4. The endometrial cancer classification diagnosis system based on nuclear magnetic resonance images as claimed in claim 1, characterized in that: The two-dimensional model unit uses a 4-layer improved Unet network to extract the intra-layer features of each layer of two-dimensional plane data in the minimum bounding box; wherein, based on the original Unet network, the improved Unet network adds a Tansformer self-attention subunit between the encoder and the decoder, and adds a fine-grained feature extraction subunit in each convolution layer, wherein the fine-grained feature extraction subunit is composed of a sequentially cascaded small receptive field convolution layer, a multi-scale convolution layer and a residual connection.

5. The endometrial cancer classification diagnosis system based on nuclear magnetic resonance images as claimed in claim 1, characterized in that: The three-dimensional model unit uses a 4-layer improved Vnet network to extract the inter-layer features of each layer of two-dimensional plane data in the minimum bounding box; among them, the improved Vnet network adds 8 stacked self-attention layers between the encoder and the decoder on the basis of the original Vnet network, and each self-attention layer is composed of sequentially cascaded Transformer self-attention sub-units, a multi-head attention mechanism, a feedforward neural network and a residual connection.

6. The endometrial cancer classification diagnosis system based on nuclear magnetic resonance images as claimed in claim 1, characterized in that: The loss function L used when training the three-dimensional model unit correlation for: Wherein, n=2, ..., N-1, n is the layer number of each layer of two-dimensional plane data in the three-dimensional data corresponding to the minimum bounding box, and N is the total number of layers of three-dimensional data corresponding to the minimum bounding box. is the eigenvalue at the position (x, y) of the two-dimensional plane data of the n-1th layer, is the average eigenvalue of the n-1th layer of two-dimensional plane data, is the eigenvalue at the position (x, y) of the n-th layer of two-dimensional plane data, is the average eigenvalue of the n-th layer of two-dimensional plane data, is the eigenvalue at the position (x, y) of the two-dimensional plane data of the n+1th layer, is the average eigenvalue of the two-dimensional plane data of the n+1th layer.

7. The endometrial cancer classification diagnosis system based on nuclear magnetic resonance images as claimed in claim 6, characterized in that: The eigenvalue at any position (x, y) of the nth layer of two-dimensional plane data in the three-dimensional data corresponding to the minimum bounding box is the conditional probability P(D) that the pixel at the position (x, y) of the n-th layer of two-dimensional plane data belongs to the endometrial cancer lesion area n (x,y)|(D n-1 (x,y),D n+1 (x,y))); Conditional probability P(D n (x,y)|(D n-1 (x,y),D n+1 The calculation method of (x,y))) is as follows: Among them, D n (x, y) is the probability that the pixel at the position (x, y) of the n-th layer of two-dimensional plane data belongs to the endometrial cancer lesion area, D n-1 (x, y) is the probability that the pixel at the position (x, y) of the n-1th layer of two-dimensional plane data belongs to the endometrial cancer lesion area, D n+1 (x, y) is the probability that the pixel point at the position (x, y) of the n+1th layer of two-dimensional plane data belongs to the endometrial cancer lesion area, α is the set proportional coefficient, and Z is the set normalization constant.

Images