Cervical cancer segmentation method with enhanced forward and reverse lesion areas at multiple angles

Abstract

The invention relates to a cervical cancer segmentation method for enhancing forward and reverse lesion areas at multiple angles. The method comprises the following steps: acquiring an original cervical cancer related image and preprocessing the original cervical cancer related image; constructing a cervical cancer segmentation model; inputting the preprocessed image into a cervical cancer segmentation model; feature maps Out1, Out2 and Out3 which are output after fusion are obtained; obtaining a fused lesion area feature map Outpu1, and obtaining a fused lesion area feature map Outpu1; and a finally fused lesion area feature map Outpu2 is obtained. According to the invention, a cervical cancer lesion area can be identified more accurately, and misdiagnosis and missed diagnosis are reduced; the downsampling stage is reduced, the calculation efficiency is improved, the model performance is improved, and a large amount of cervical cancer image data can be quickly processed; the method is high in operation speed and low in calculation cost, can quickly give out a cervical cancer segmentation result in practical application, is suitable for large-scale cervical cancer screening work, particularly provides reliable auxiliary support for early screening of cervical cancer in areas with limited health resources, and has wide application prospects and popularization value.

Description

Technical Field

[0001] The invention relates to the technical field of medical image processing, and in particular to a cervical cancer segmentation method with forward and reverse lesion area enhancement at multiple angles. Background Art

[0002] Cervical cancer is a common gynecological malignancy that threatens women's life and health. It has a high incidence and mortality rate, and the distribution of cases is uneven. The mortality rate is on the rise in underdeveloped areas. Early detection and treatment can significantly improve the cure rate, so regular screening for cervical cancer is essential.

[0003] The Pap smear method is often used to screen for cervical cancer in economically developed areas, but in areas with limited health resources, the cytological diagnosis system is imperfect, and the iodine staining test based on colposcopy has become the preferred screening method. However, this method is highly dependent on the clinical experience of doctors, and there is a lack of experienced doctors in underdeveloped areas, which can easily lead to misdiagnosis and missed diagnosis.

[0004] With the development of computer technology, deep learning methods have provided new breakthroughs in medical diagnosis and treatment. In the field of medical image segmentation, convolutional neural networks (CNNs) and attention mechanisms are widely used, but CNNs are insufficient in extracting global feature information. The existing channel attention mechanism often ignores position information, and the network upsampling module fails to fully improve the representation details of the predicted feature map. In addition, Transformer lacks spatial perception bias when processing global context information, and the local detail modeling of medical image segmentation needs to be improved. Summary of the invention

[0005] In order to solve the limitations of convolutional neural networks in extracting global features, the traditional attention mechanism fails to fully utilize spatial position information, and the existing upsampling module cannot effectively enhance the details of the lesion area, the purpose of the present invention is to provide a cervical cancer segmentation method that can more accurately identify cervical cancer lesions, reduce the occurrence of misdiagnosis and missed diagnosis, and can quickly process a large amount of cervical cancer image data and quickly give cervical cancer segmentation results under multiple angles with positive and negative lesion area enhancement.

[0006] To achieve the above object, the present invention adopts the following technical solution: a cervical cancer segmentation method with forward and reverse lesion area enhancement under multiple angles, the method comprising the following steps in order:

[0007] (1) collecting original cervical cancer-related images and preprocessing them to obtain preprocessed images;

[0008] (2) Constructing a cervical cancer segmentation model, which includes an improved Swin-Transformer model, a multi-angle attention mechanism, and an upsampling module for forward and reverse lesion area enhancement;

[0009] (3) Input the preprocessed image into the cervical cancer segmentation model;

[0010] (4) The three-stage feature maps Feature1, Feature2, and Feature3 extracted by the improved Swin-Transformer model are respectively input into the multi-angle attention mechanism module, and features are extracted and fused from four angles: key pixels, surrounding pixels, horizontal direction, and vertical direction, to obtain the fused output feature maps Out1, Out2, and Out3 respectively;

[0011] (5) Input Out2 and Out3 into the upsampling module for positive and negative lesion area enhancement, enhance the lesion area and suppress the non-lesion area, and obtain the fused lesion area feature map Output1;

[0012] (6) Out1 and Output1 are input into the upsampling module for forward and reverse lesion area enhancement to obtain the final fused lesion area feature map Output2.

[0013] In step (2), the improved Swin-Transformer model refers to improving the encoder of the Swin-Transformer model on the basis of the Swin-Transformer model, and modifying the four stages for achieving 32-fold downsampling to three stages for achieving 16-fold downsampling.

[0014] The step (4) specifically refers to: extracting feature information of different scales from the input feature map, assuming that the input feature map is X∈R c×h×w , c is the number of channels of the input feature map, h is the height of the input feature map, and w is the width of the input feature map; for key pixel information, a convolution kernel with a size of k = 3 × 3 and a dilation rate of d = 1 is used for feature extraction; for information around key pixels, a convolution kernel with a size of k = 3 × 3 and a dilation rate of d = 2 is used for feature extraction; for information at two angles, horizontal and vertical, the input feature map is converted into a one-dimensional feature by adaptive average pooling.

[0015] The pixel point and its surrounding information are concatenated as follows to obtain the feature map Z':

[0016] Z'=Cat[Conv k＝3,d＝1 (X),Conv k＝3,d＝2 (X)]

[0017] In the formula, Conv means convolution; Cat means fusion;

[0018] It is processed by convolution and batch normalization operations, followed by nonlinear activation, as shown in the following formula:

[0019] Z = ReLU(BN(Conv(Z')))

[0020] In the formula, Z represents the feature map after nonlinear activation, BN represents the batch normalization operation, and ReLU represents the nonlinear activation operation;

[0021] After the nonlinear activation, the feature map Z is divided into Z1' and Z2' by the Split function, and the convolution operation and Sigmoid function are used to generate the important feature map, that is, the feature Z1 of the key pixel and the feature Z2 of the surrounding pixels:

[0022] Z1',Z2'=Split(Z)

[0023] Z1=Sigmoid(Conv(Z1'))

[0024] Z2=Sigmoid(Conv(Z2'))

[0025] Similarly, for the feature information in the horizontal and vertical directions, the input feature map is first subjected to global pooling operations in the horizontal and vertical directions, and further convolution operations are performed to extract deeper feature information Y', followed by batch normalization and nonlinear activation operations on Y'; finally, the feature is decomposed into horizontal features Y h ' and vertical feature Y w ', and calculate the weight distribution through the Sigmoid activation function to generate the horizontal feature weight map Y h and vertical feature weight map Y w :

[0026] Y'=Cat[AvgPooL h (X)AvgPooL w (X)]

[0027] Y = ReLU(BN(Conv(Y')))

[0028] Y h ',Y w '=Split(Y)

[0029] Y h =Sigmoid(Conv(Y h '))

[0030] Y w =Sigmoid(Conv(Y w '))

[0031] Among them, AvgPooL h Represents the average pooling in the horizontal direction, AvgPooL wrepresents the average pooling in the vertical direction; Y represents the intermediate feature map of the activation process;

[0032] Then, the pixels from the key pixel Z1, the surrounding pixel Z2, and the horizontal direction Y are integrated. h and vertical direction Y w The feature importance map of is used to redistribute the weights of the input feature map as follows:

[0033] Out=X×Y h ×Y w ×Z1×Z2

[0034] In the formula, Out is the feature map output after fusion; Feature1, Feature2, and Feature are respectively output through the multi-angle attention mechanism module to obtain the feature maps Out1, Out2, and Out3 after fusion.

[0035] The step (5) specifically refers to: the upsampling module for forward and reverse lesion area enhancement simultaneously enhances the detailed features of the lesion area and the edge of the lesion area from both forward and reverse angles;

[0036] Input feature maps Out2 and Out3 of different resolution sizes. The size of Out2 is twice that of feature map Out3. Therefore, Out3 is upsampled before feature fusion to make it the same size as Out2.

[0037] The positive lesion area enhancement method of the upsampling module for positive and negative lesion area enhancement is as follows:

[0038] First, the features of the two feature maps are fused through convolution to obtain the initial target area importance map Feature:

[0039] Feature=ReLU(BN(Conv(Cat(Out2,Out3))))

[0040] In the formula, BN represents batch normalization operation, ReLU represents nonlinear activation operation; Conv represents convolution, and Cat represents fusion;

[0041] Then, the feature map is normalized using the Sigmoid function to generate the weight feature map of the lesion area. Map :

[0042] Forward Map =Sigmoid(Conv(Feature))

[0043] Then, the adaptive weight is used to set the image area greater than or equal to the weight μ to 1, and the image area less than the weight β to 0. Through binarization processing, the focus area map of the lesion part in the image is obtained. AdpMap :

[0044] Forward AdpMap [Forward Map ≥μ]=1

[0045] Forward AdpMap [Forward Map <β]=0

[0046] The reverse lesion enhancement method of the upsampling module for forward and reverse lesion area enhancement is as follows:

[0047] First, the two feature maps Out2 and Out3 are subjected to their own convolution feature extraction operations to refine their respective detail features and obtain feature maps Out2' and Out3':

[0048] Out2' = ReLU(BN(Conv(Out2)))

[0049] Out3' = ReLU(BN(Conv(Out3)))

[0050] Then, the distribution of the area of ​​interest is obtained through the Sigmoid function, and the feature maps Out2” and Out3” are obtained:

[0051] Out2"=Sigmoid(Conv(Out2'))

[0052] Out3"=Sigmoid(Conv(Out3'))

[0053] Then, the target area is segmented from the reverse perspective, and the feature distribution of each of the two feature maps is subtracted from 1, and the feature distribution of each of the two feature maps is summed to generate the reverse feature map Reverse Map , further refine the target area characteristics:

[0054] Reverse Map =(1-Sigmoid(Conv(Out2”)))+(1-Sigmoid(Conv(Out3”)))

[0055] Next, the adaptive weight is used to set the image area greater than or equal to the weight γ to 1, and the image area less than the weight γ to 0. The reverse feature weight distribution map of the lesion area is further screened out through importance weighting. AdpMap :

[0056] Reverse AdpMap [Reverse Map ≥γ]=1

[0057] Reverse AdpMap [Reverse Map <γ]=0

[0058] Combine the initial target area importance map Feature and Forward AdpMap 、Reverse AdpMap The three are multiplied and weighted, and then reintegrated using residual connections to obtain the fused lesion area feature map Output1:

[0059] Output1=Feature+Feature×Forward AdpMap ×Reverse AdpMap .

[0060] It can be seen from the above technical scheme that the beneficial effects of the present invention are: First, high-precision segmentation: through the improved Swin-Transformer model, combined with the multi-angle attention mechanism module and the upsampling module for positive and negative lesion area enhancement, the cervical cancer lesion area can be more accurately identified. The experimental results show that the mIoU value reaches 82.5%, and it performs well in indicators such as accuracy, specificity and precision. Compared with traditional methods, it can more accurately assist doctors in the early diagnosis of cervical cancer and reduce the occurrence of misdiagnosis and missed diagnosis; Second, efficient feature extraction and calculation: The present invention optimizes the encoder of the Swin-Transformer model and reduces the downsampling stage. The computational efficiency has been improved. The multi-angle attention mechanism module can more comprehensively extract feature information of different scales and directions. The upsampling module enhances the features of the lesion area and edge details from both positive and negative angles, making the feature extraction more complete, thereby improving the model performance and being able to quickly process a large amount of cervical cancer image data; Third, low cost and high availability: The present invention has fast computing speed and low computing cost. In practical applications, it can quickly give cervical cancer segmentation results and is suitable for large-scale cervical cancer screening. Especially for areas with limited health resources, it can make up for the lack of experienced doctors to a certain extent, and provide reliable auxiliary support for early screening of cervical cancer. It has broad application prospects and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic diagram of the structure of the cervical cancer segmentation model in the present invention;

[0062] Figure 2 It is a schematic diagram of the structure of the up-sampling module used for positive and negative lesion area enhancement in the present invention. DETAILED DESCRIPTION

[0063] A cervical cancer segmentation method with forward and reverse lesion area enhancement under multiple angles, the method comprising the following steps in order:

[0064] (1) collecting original cervical cancer-related images and preprocessing them to obtain preprocessed images;

[0065] (2) Constructing a cervical cancer segmentation model, which includes an improved Swin-Transformer model, a multi-angle attention mechanism, and an upsampling module for forward and reverse lesion area enhancement;

[0066] (3) Input the preprocessed image into the cervical cancer segmentation model, such as Figure 1 As shown, Figure 1 MAA is a multi-angle attention mechanism module, and BLEU is an upsampling module for positive and negative lesion area enhancement;

[0067] (4) The three-stage feature maps Feature1, Feature2, and Feature3 extracted by the improved Swin-Transformer model are respectively input into the multi-angle attention mechanism module, and features are extracted and fused from four angles: key pixels, surrounding pixels, horizontal direction, and vertical direction, to obtain the fused output feature maps Out1, Out2, and Out3 respectively;

[0068] (5) Input Out2 and Out3 into the upsampling module for positive and negative lesion area enhancement, enhance the lesion area and suppress the non-lesion area, and obtain the fused lesion area feature map Output1;

[0069] (6) Out1 and Output1 are input into the upsampling module for forward and reverse lesion area enhancement to obtain the final fused lesion area feature map Output2.

[0070] In step (2), the improved Swin-Transformer model refers to improving the encoder of the Swin-Transformer model on the basis of the Swin-Transformer model, modifying the four stages of 32-fold downsampling to three stages of 16-fold downsampling, and proving through ablation experiments that the segmentation result of 16-fold downsampling is better than the result of 32-fold downsampling, and is more suitable for the network model of the present invention. The three stages of feature extraction can gradually capture a wider range of contextual information by increasing the receptive field of the window layer by layer, thereby enhancing the coherence and globality of the model's feature representation of cervical image regions of different sizes and colors.

[0071] The present invention uses a hierarchical Swin-Transformer encoder as the backbone network. The input image is gradually divided into patches of different resolutions. The network reduces the resolution of the feature map layer by layer by learning the weight correlation and interaction between the patches of the feature map in the current layer, thereby better capturing the global information of the colposcopic image and refining the local details of the lesion area.

[0072] The step (4) specifically refers to: extracting feature information of different scales from the input feature map, assuming that the input feature map is X∈R c×h×w , c is the number of channels of the feature map, h is the height of the feature map, and w is the width of the feature map; for the key pixel information, a convolution kernel with a size of k = 3 × 3 and a dilation rate of d = 1 is used for feature extraction; for the information around the key pixel, a convolution kernel with a size of k = 3 × 3 and a dilation rate of d = 2 is used for feature extraction; for the information of the horizontal and vertical directions, the input feature map is converted into a one-dimensional feature by adaptive average pooling respectively;

[0073] The pixel point and its surrounding information are concatenated as follows to obtain the feature map Z':

[0074] Z'=Cat[Conv k＝3,d＝1 (X),Conv k＝3,d＝2 (X)]

[0075] In the formula, Conv means convolution; Cat means fusion;

[0076] It is processed by convolution and batch normalization operations, followed by nonlinear activation, as shown in the following formula:

[0077] Z = ReLU(BN(Conv(Z')))

[0078] In the formula, Z represents the feature map after nonlinear activation, BN represents the batch normalization operation, and ReLU represents the nonlinear activation operation;

[0079] After the nonlinear activation, the feature map Z is divided into Z1' and Z2' by the Split function, and the convolution operation and Sigmoid function are used to generate the important feature map, that is, the feature Z1 of the key pixel and the feature Z2 of the surrounding pixels:

[0080] Z1',Z2'=Split(Z)

[0081] Z1=Sigmoid(Conv(Z1'))

[0082] Z2=Sigmoid(Conv(Z2'))

[0083] Similarly, for the feature information in the horizontal and vertical directions, the input feature map is first subjected to global pooling operations in the horizontal and vertical directions, and further convolution operations are performed to extract deeper feature information Y', followed by batch normalization and nonlinear activation operations on Y'; finally, the feature is decomposed into horizontal features Y h ' and vertical feature Y w ', and calculate the weight distribution through the Sigmoid activation function to generate the horizontal feature weight map Y h and vertical feature weight map Y w :

[0084] Y'=Cat[AvgPooL h (X)AvgPooL w (X)]

[0085] Y = ReLU(BN(Conv(Y')))

[0086] Y h ',Y w '=Split(Y)

[0087] Y h =Sigmoid(Conv(Y h '))

[0088] Y w =Sigmoid(Conv(Y w '))

[0089] Among them, AvgPooL h Represents the average pooling in the horizontal direction, AvgPooL w represents the average pooling in the vertical direction; Y represents the intermediate feature map of the activation process;

[0090] Then, the pixels from the key pixel Z1, the surrounding pixel Z2, and the horizontal direction Y are integrated. h and vertical direction Y w The feature importance map of is used to redistribute the weights of the input feature map as follows:

[0091] Out=X×Y h ×Y w ×Z1×Z2

[0092] In the formula, Out is the feature map output after fusion; Feature1, Feature2, and Feature are respectively output through the multi-angle attention mechanism module to obtain the feature maps Out1, Out2, and Out3 after fusion.

[0093] like Figure 1As shown in the figure, the multi-angle attention mechanism understands the semantic relationship between the horizontal and vertical directions of image features from the horizontal and vertical perspectives, namely the two branches on the right in the MAA module diagram. At the same time, taking each pixel in the image as a reference, convolution with different expansion rates is used to obtain the semantic information of each pixel and its surrounding pixels, namely the two branches on the left in the MAA module diagram. Finally, all feature information obtained from four different angles is fused to obtain multi-angle attention weights that emphasize both space and channels, providing sufficient feature expression for feature extraction of each part of the network.

[0094] The step (5) specifically refers to: the upsampling module for forward and reverse lesion area enhancement simultaneously enhances the detailed features of the lesion area and the edge of the lesion area from both forward and reverse angles, such as Figure 2 As shown;

[0095] Input feature maps Out2 and Out3 of different resolution sizes. The size of Out2 is twice that of feature map Out3. Therefore, Out3 is upsampled before feature fusion to make it the same size as Out2.

[0096] The positive lesion area enhancement method of the upsampling module for positive and negative lesion area enhancement is as follows:

[0097] First, the features of the two feature maps are fused through convolution to obtain the initial target area importance map Feature:

[0098] Feature=ReLU(BN(Conv(Cat(Out2,Out3))))

[0099] In the formula, BN represents batch normalization operation, ReLU represents nonlinear activation operation; Conv represents convolution, and Cat represents fusion;

[0100] Then, the feature map is normalized using the Sigmoid function to generate the weight feature map of the lesion area. Map :

[0101] Forward Map =Sigmoid(Conv(Feature))

[0102] Then, the adaptive weight is used to set the image area greater than or equal to the weight μ to 1, and the image area less than the weight β to 0. Through binarization processing, the focus area map of the lesion part in the image is obtained. AdpMap :

[0103] Forward AdpMap [Forward Map ≥μ]=1

[0104] Forward AdpMap [Forward Map <β]=0

[0105] The reverse lesion enhancement method of the upsampling module for forward and reverse lesion area enhancement is as follows:

[0106] First, the two feature maps Out2 and Out3 are subjected to their own convolution feature extraction operations to refine their respective detail features and obtain feature maps Out2' and Out3':

[0107] Out2' = ReLU(BN(Conv(Out2)))

[0108] Out3' = ReLU(BN(Conv(Out3)))

[0109] Then, the distribution of the area of ​​interest is obtained through the Sigmoid function, and the feature maps Out2” and Out3” are obtained:

[0110] Out2"=Sigmoid(Conv(Out2'))

[0111] Out3"=Sigmoid(Conv(Out3'))

[0112] Then, the target area is segmented from the reverse perspective, and the feature distribution of each of the two feature maps is subtracted from 1, and the feature distribution of each of the two feature maps is summed to generate the reverse feature map Reverse Map , further refine the target area characteristics:

[0113] Reverse Map =(1-Sigmoid(Conv(Out2”)))+(1-Sigmoid(Conv(Out3”)))

[0114] Next, the adaptive weight is used to set the image area greater than or equal to the weight γ to 1, and the image area less than the weight γ to 0. The reverse feature weight distribution map of the lesion area is further screened out through importance weighting. AdpMap :

[0115] Reverse AdpMap [Reverse Map ≥γ]=1

[0116] Reverse AdpMap [Reverse Map <γ]=0

[0117] Combine the initial target area importance map Feature and Forward AdpMap 、Reverse AdpMap The three are multiplied and weighted, and then reintegrated using residual connections to obtain the fused lesion area feature map Output1:

[0118] Output1=Feature+Feature×Forward AdpMap ×Reverse AdpMap .

[0119] Similarly, step (6) is obtained according to step (5), except that the input parameters are different.

[0120] In summary, the present invention can more accurately identify cervical cancer lesion areas through the improved Swin-Transformer model, combined with a multi-angle attention mechanism module and an upsampling module for positive and negative lesion area enhancement. The experimental results show that the mIoU value reaches 82.5%, and it performs well in terms of accuracy, specificity and precision. Compared with traditional methods, it can more accurately assist doctors in early diagnosis of cervical cancer and reduce the occurrence of misdiagnosis and missed diagnosis. The encoder of the Swin-Transformer model is optimized, the downsampling stage is reduced, and the calculation efficiency is improved. The multi-angle attention mechanism module can more comprehensively extract feature information of different scales and directions. The upsampling module enhances the lesion area and edge detail features from both positive and negative angles, making the feature extraction more perfect, thereby improving the model performance and being able to quickly process a large amount of cervical cancer image data. The present invention has fast computing speed and low computing cost. In practical applications, it can quickly give cervical cancer segmentation results, which is suitable for large-scale cervical cancer screening. Especially for areas with limited health resources, it can make up for the lack of experienced doctors to a certain extent, and provide reliable auxiliary support for early screening of cervical cancer. It has broad application prospects and promotion value.

Claims

1. A cervical cancer segmentation method with forward and reverse lesion area enhancement under multiple angles, characterized by: The method comprises the following steps in order: (1) collecting original cervical cancer-related images and preprocessing them to obtain preprocessed images; (2) Constructing a cervical cancer segmentation model, which includes an improved Swin-Transformer model, a multi-angle attention mechanism, and an upsampling module for forward and reverse lesion area enhancement; (3) Input the preprocessed image into the cervical cancer segmentation model; (4) The three-stage feature maps Feature1, Feature2, and Feature3 extracted by the improved Swin-Transformer model are respectively input into the multi-angle attention mechanism module, and features are extracted and fused from four angles: key pixels, surrounding pixels, horizontal direction, and vertical direction, to obtain the fused output feature maps Out1, Out2, and Out3 respectively; (5) Inputting Out2 and Out3 into the upsampling module for positive and negative lesion area enhancement, enhancing the lesion area and suppressing the non-lesion area, and obtaining the fused lesion area feature map Output1; (6) Out1 and Output1 are input into the upsampling module for forward and reverse lesion area enhancement to obtain the final fused lesion area feature map Output2.

2. The cervical cancer segmentation method with forward and reverse lesion area enhancement under multiple angles according to claim 1, characterized in that: In step (2), the improved Swin-Transformer model refers to improving the encoder of the Swin-Transformer model on the basis of the Swin-Transformer model, and modifying the four stages for achieving 32-fold downsampling to three stages for achieving 16-fold downsampling.

3. The cervical cancer segmentation method with forward and reverse lesion area enhancement under multiple angles according to claim 1, characterized in that: The step (4) specifically refers to: extracting feature information of different scales from the input feature map, assuming that the input feature map is X∈R c×h×w , c is the number of channels of the input feature map, h is the height of the input feature map, and w is the width of the input feature map; for key pixel information, a convolution kernel with a size of k = 3 × 3 and a dilation rate of d = 1 is used for feature extraction; for information around key pixels, a convolution kernel with a size of k = 3 × 3 and a dilation rate of d = 2 is used for feature extraction; for information at two angles, horizontal and vertical, the input feature map is converted into a one-dimensional feature by adaptive average pooling. The pixel point and its surrounding information are concatenated as follows to obtain the feature map Z': Z'=Cat[Conv k＝3,d＝1 (X),Conv k＝3,d＝2 (X)] In the formula, Conv means convolution; Cat means fusion; It is processed by convolution and batch normalization operations, followed by nonlinear activation, as shown in the following formula: Z = ReLU(BN(Conv(Z'))) In the formula, Z represents the feature map after nonlinear activation, BN represents the batch normalization operation, and ReLU represents the nonlinear activation operation; After the nonlinear activation, the feature map Z is divided into Z1' and Z2' by the Split function, and the convolution operation and Sigmoid function are used to generate the important feature map, that is, the feature Z1 of the key pixel and the feature Z2 of the surrounding pixels: Z1',Z2'=Split(Z) Z1=Sigmoid(Conv(Z1')) Z2=Sigmoid(Conv(Z2')) Similarly, for the feature information in the horizontal and vertical directions, the input feature map is first subjected to global pooling operations in the horizontal and vertical directions, and further convolution operations are performed to extract deeper feature information Y', followed by batch normalization and nonlinear activation operations on Y'; finally, the feature is decomposed into horizontal features Y h ' and vertical feature Y w ', and calculate the weight distribution through the Sigmoid activation function to generate the horizontal feature weight map Y h and vertical feature weight map Y w : Y'=Cat[AvgPooL h (X)AvgPooL w (X)] Y = ReLU(BN(Conv(Y'))) AND h ',AND w '=Split(Y) Y h =Sigmoid(Conv(Y h ')) Y w =Sigmoid(Conv(Y w ')) Among them, AvgPooL h Represents the average pooling in the horizontal direction, AvgPooL w represents the average pooling in the vertical direction; Y represents the intermediate feature map of the activation process; Then, the pixels from the key pixel Z1, the surrounding pixel Z2, and the horizontal direction Y are integrated. h and vertical direction Y w The feature importance map of is used to redistribute the weights of the input feature map as follows: Out=X×Y h ×Y w ×Z1×Z2 In the formula, Out is the feature map output after fusion; Feature1, Feature2, and Feature are respectively output through the multi-angle attention mechanism module to obtain the feature maps Out1, Out2, and Out3 after fusion.

4. The cervical cancer segmentation method with forward and reverse lesion area enhancement under multiple angles according to claim 1, characterized in that: The step (5) specifically refers to: the upsampling module for forward and reverse lesion area enhancement simultaneously enhances the detailed features of the lesion area and the edge of the lesion area from both forward and reverse angles; Input feature maps Out2 and Out3 of different resolution sizes. The size of Out2 is twice that of feature map Out3. Therefore, Out3 is upsampled before feature fusion to make it the same size as Out2. The positive lesion area enhancement method of the upsampling module for positive and negative lesion area enhancement is as follows: First, the features of the two feature maps are fused through convolution to obtain the initial target area importance map Feature: Feature=ReLU(BN(Conv(Cat(Out2,Out3)))) In the formula, BN represents batch normalization operation, ReLU represents nonlinear activation operation; Conv represents convolution, and Cat represents fusion; Then, the feature map is normalized using the Sigmoid function to generate the weight feature map of the lesion area. Map : Forward Map =Sigmoid(Conv(Feature)) Then, the adaptive weight is used to set the image area greater than or equal to the weight μ to 1, and the image area less than the weight β to 0. Through binarization processing, the focus area map of the lesion part in the image is screened. AdpMap : Forward AdpMap [Forward Map ≥μ]=1 Forward AdpMap [Forward Map <β]=0 The reverse lesion enhancement method of the upsampling module for forward and reverse lesion area enhancement is as follows: First, the two feature maps Out2 and Out3 are subjected to their own convolution feature extraction operations to refine their respective detail features and obtain feature maps Out2' and Out3': Out2' = ReLU(BN(Conv(Out2))) Out3' = ReLU(BN(Conv(Out3))) Then, the distribution of the area of ​​interest is obtained through the Sigmoid function, and the feature maps Out2” and Out3” are obtained: Out2"=Sigmoid(Conv(Out2')) Out3"=Sigmoid(Conv(Out3')) Then, the target area is segmented from the reverse perspective, and the feature distribution of each of the two feature maps is subtracted from 1, and the feature distribution of each of the two feature maps is summed to generate the reverse feature map Reverse Map , further refine the target area characteristics: Reverse Map =(1-Sigmoid(Conv(Out2”)))+(1-Sigmoid(Conv(Out3”))) Next, the adaptive weight is used to set the image area greater than or equal to the weight γ to 1, and the image area less than the weight γ to 0. The reverse feature weight distribution map of the lesion area is further screened out through importance weighting. AdpMap : Reverse AdpMap [Reverse Map ≥γ]=1 Reverse AdpMap [Reverse Map <γ]=0 Combine the initial target area importance map Feature and Forward AdpMap 、Reverse AdpMap The three are multiplied and weighted, and then reintegrated using residual connections to obtain the fused lesion area feature map Output1: Output1=Feature+Feature×Forward AdpMap ×Reverse AdpMap 。

Images