Ovarian cancer image deep analysis method based on adaptive fusion network

Abstract

The invention belongs to the technical field of ovarian cancer image research and analysis, and particularly relates to an ovarian cancer image deep analysis method based on a self-adaptive fusion network, which comprises the steps of random horizontal flipping, color jitter and random cutting and erasing, and realizes the diversification of training data, thereby improving the generalization ability of a model; through a hierarchical feature extraction mechanism, introducing two models of a convolutional neural network (CNN) and a visual converter (ViT), and respectively capturing low-level detail features and high-level semantic features of the image; through a feature adaptive fusion method, the weights of CNN and ViT can be dynamically adjusted according to task requirements, and low-level and high-level features are dynamically fused, so that the model can comprehensively learn local details and global structure information of an image; a loss function Balanced Dice Loss combining Focal Loss and Dice Loss is introduced, the weight of a difficult sample is adjusted, the problems that data categories are unbalanced and prediction of a model in a small area is difficult are solved, and therefore more accurate ovarian cancer image classification is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of ovarian cancer image research and analysis, and in particular relates to an ovarian cancer image depth analysis method based on an adaptive fusion network. Background Art

[0002] Ovarian cancer is a highly heterogeneous and lethal gynecological malignancy, and its development is accompanied by complex molecular biological characteristics and multi-stage pathological changes. Image analysis plays a vital role in the early diagnosis of ovarian cancer. Medical imaging can effectively detect and monitor the tumor morphology, location and spread of ovarian cancer. Ovarian cancer image data is highly heterogeneous. The lesion areas of different patients have different manifestations. The size, shape, density of the tumor and its relationship with the surrounding tissue vary greatly. Existing methods do not consider its high heterogeneity when classifying ovarian cancer images. Summary of the invention

[0003] In view of the technical problem that the above-mentioned existing methods do not consider the high heterogeneity of ovarian cancer images when classifying them, the present invention provides a deep analysis method of ovarian cancer images based on an adaptive fusion network.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0005] A method for deep analysis of ovarian cancer images based on an adaptive fusion network comprises the following steps:

[0006] S1. First, the input medical image data is preprocessed and the image is uniformly resized to 224×224 using standardized size adjustment technology to ensure the uniformity of the input data;

[0007] S2. Perform data enhancement on the adjusted images by using techniques such as random horizontal flipping, color jittering, random cropping, and random erasing to increase the diversity of training data and avoid model overfitting.

[0008] S3, introduce convolutional neural network CNN and visual transformer ViT for feature extraction, CNN is used to extract low-level features in the image, and ViT is used to extract high-level features;

[0009] S4, based on the adaptive fusion method, the features extracted by CNN and ViT are dynamically fused to adaptively adjust the weights of different features and optimize the feature fusion effect;

[0010] S5. Use the loss function Balanced Dice Loss, which is a combination of Focal Loss and Dice Loss, for model training to solve the problems of data category imbalance and model prediction difficulties in small areas;

[0011] S6, introduces the Adam optimizer and OneCycleLR learning rate scheduling strategy to make the model converge quickly during training;

[0012] S7. The performance of the model was evaluated through various indicators such as accuracy, F1 score, AUC-ROC, etc. during the training and validation process. The final classification effect of the model was evaluated using statistical analysis methods to provide support for the early diagnosis of ovarian cancer.

[0013] The method for preprocessing the input medical image data in S2 is:

[0014] S2.1, randomly flip the medical image data horizontally to generate mirror data;

[0015] S2.2, randomly rotate the image, and the angle range is set to [-30°, 30°] to simulate samples at different shooting angles;

[0016] S2.3, introduce color jitter to adjust the brightness, contrast, saturation and hue of the image respectively, and the perturbation range of each parameter is between [0.8, 1.2];

[0017] S2.4, randomly crop the image, set the cropping ratio range to [80%, 100%], and then adjust the cropped image to the standard size (224×224);

[0018] S2.5, randomly erase some areas in the image, the size range of the erased area is [2%, 10%], simulating the scene of partial information missing;

[0019] S2.6. Complete the normalization operation of the enhanced image and normalize the pixel values ​​to the range of [-1,1] to ensure data consistency.

[0020] The method of introducing convolutional neural network CNN and visual transformer ViT for feature extraction in S3 is:

[0021] The convolutional neural network is used to extract low-level features of the input image and generate feature vectors containing local information such as edges and textures.

[0022] y=W*X+b

[0023] Among them, X is the input image, W is the convolution kernel, b is the bias term, y is the output feature vector, and * represents the convolution operation;

[0024] The high-level features of the input image are extracted through the visual transformer, and the semantic feature vector is generated using the global self-attention mechanism;

[0025] ViT constructs a global representation of the image by dividing the image into multiple small blocks and linearly embedding each image block into a vector. Then, the self-attention mechanism is used to weight these image blocks to capture the relationship between different regions in the image. The calculation formula of self-attention is:

[0026]

[0027] Where Q is the query matrix, K is the key matrix, V is the value matrix, and softmax is used to calculate the attention weights. is the normalization factor; by averaging all features output by ViT, the global semantic information of the image is extracted, and the calculation formula is:

[0028]

[0029] where F i is the output of the i-th feature block, and N is the number of feature blocks.

[0030] The method for dynamically fusing the features extracted by CNN and ViT in S4 to adaptively adjust the weights of different features is:

[0031] A dynamic weighting mechanism is used to fuse low-level features and high-level features. The specific method is as follows: according to the feature importance of CNN and ViT, the weight generation module g(·) is used to automatically calculate the weight according to the input features:

[0032] F hybrid =g CNN (F CNN , F ViT )·F CNN +g ViT (F CNN , F ViT )·F ViT

[0033] where g CNN (·), and g ViT (·) are two dynamic weight generation functions, which are dynamically adjusted according to the importance of input features through model training to satisfy g CNN(·) +g ViT(·) =1;

[0034] The adaptively fused CNN features and ViT features are concatenated to form a comprehensive feature vector for subsequent classification tasks;

[0035] The fused features are input into the classification module, which consists of a fully connected layer and a Sigmoid activation function to output the classification probability value.

[0036] The method for model training using the loss function Balanced Dice Loss combined with Focal Loss and Dice Loss in S5 is:

[0037] Focal Loss is introduced to solve the problem of class imbalance. It gives higher weights to difficult-to-classify samples and reduces the impact of easy-to-classify samples on the loss, so that the model can focus more on distinguishing difficult samples. The formula is:

[0038] FL(p t )=―α t (1―p t ) γ log(p t )

[0039] Among them, p t is the probability of the correct category predicted by the model, α t is the category balance factor, which is used to adjust the weights of each category, and γ is the focus factor, which controls the attention paid to difficult-to-classify samples;

[0040] Dice Loss is introduced to optimize the model's prediction ability for tiny areas (such as boundary areas) by measuring the overlap between the prediction and the true label. It is particularly suitable for medical image classification. Its formula is:

[0041]

[0042] Among them, P represents the predicted area, G represents the real area, and Dice Loss optimizes the performance of the model in category segmentation by calculating the ratio of the intersection and the total area.

[0043] Balanced Dice Loss combines the advantages of Focal Loss and Dice Loss to solve the problem of category imbalance and optimize the performance of the model in fine-grained areas. Its formula is:

[0044] Balanced Dice Loss=λ·Focal Loss+(1―λ)·Dice Loss

[0045] Where λ∈[0,1] is the weight coefficient, and the weight ratio of Focal Loss and Dice Loss can be dynamically adjusted according to the characteristics of the data set to achieve the best effect.

[0046] The method of introducing the Adam optimizer and the OneCycleLR learning rate scheduling strategy in S6 is:

[0047] The Adam optimizer is introduced for model training, combined with momentum and adaptive learning rate adjustment mechanism to improve the efficiency and stability of model parameter updating; the formula of the Adam optimizer is:

[0048]

[0049] Where: θ t is the model at step t, η is the learning rate, m t is the first-order momentum estimate of the gradient, v t is the second-order momentum estimate of the gradient, and λ is the weight decay coefficient, which is used to control model overfitting;

[0050] The OneCycleLR learning rate scheduling strategy is introduced to dynamically adjust the learning rate according to the number of training rounds to prevent the model from falling into the local optimum and accelerate the training convergence. The learning rate is divided into two stages during the training process: first slowly increase, and then gradually decrease after reaching the maximum value to ensure a more stable training process. Assuming the total number of training steps is T, the formula for the change of the learning rate is:

[0051]

[0052] Among them, η min is the initial learning rate, η max is the maximum learning rate, T up is the number of steps in the learning rate ramp-up phase.

[0053] The method for evaluating the performance of the model through multiple indicators in the training and verification process in S7 is:

[0054] A variety of evaluation indicators are used to measure the classification effect of the model. Accuracy is used to measure the proportion of correctly classified samples to the total number of samples. The formula is:

[0055]

[0056] Among them, TP is true positive, TN is true negative, FP is false positive, and FN is false negative;

[0057] F1-Score is the comprehensive evaluation model's precision and recall. Its formula is:

[0058]

[0059] in,

[0060] AUC (Area Under the Curve) is the area under the ROC (Receiver Operating Characteristic) curve, which is used to evaluate the performance of the model under different classification thresholds. The ROC curve is a curve drawn with the false positive rate FPR as the horizontal axis and the true positive rate TPR as the vertical axis.

[0061]

[0062] in The closer the AUC value is to 1, the stronger the model's ability to distinguish is, and the closer it is to 0.5, the better the model's performance is.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] The present invention utilizes convolutional neural network (CNN) and visual transformer (ViT) to extract low-level features and high-level features of images, combines cross-level feature adaptive fusion method, fully explores the complementarity of low-level features and high-level features, optimizes feature fusion effect, and effectively improves the accuracy and robustness of ovarian cancer image classification. In order to solve the problem of image category imbalance, the present invention introduces Balanced Dice Loss loss function, combined with Focal Loss and Dice Loss, which is particularly suitable for small area prediction in ovarian cancer images. The model performance is evaluated by multiple indicators such as accuracy, F1 score, AUC-ROC, etc. during training and verification, and the model is further hyperparameterized. Finally, the present invention uses statistical analysis method to evaluate the classification effect of the model to ensure that the early diagnosis of ovarian cancer is more accurate. Through these technologies, the present invention can effectively improve the performance of ovarian cancer image classification and provide important support for the formulation of clinical diagnosis and treatment plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

[0066] The structures, proportions, sizes, etc. illustrated in this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with the technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantial technical significance. Any structural modification, change in proportion or adjustment of size shall still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.

[0067] Figure 1 It is a schematic diagram of the process of the present invention;

[0068] Figure 2 It is a model diagram of the present invention;

[0069] Figure 3 The loss curve, accuracy curve, F1 score curve and ROC curve of the present invention in the training set and the validation set are shown;

[0070] Figure 4 It is a comparison chart of the present invention and the existing classification method;

[0071] Figure 5 is a graph of the prediction performance of the present invention and different prediction methods on the TCIA dataset;

[0072] Figure 6 It is a comparison diagram of the confusion matrix of the present invention under different training parameters. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. These descriptions are only to further illustrate the features and advantages of the present invention, rather than to limit the claims of the present invention. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0074] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0075] This embodiment provides a flowchart of a method for deep analysis of ovarian cancer images based on an intelligent adaptive fusion network under the guidance of precision medicine. Figure 1 , Figure 2 As shown, the steps of the method include:

[0076] Step 1: Data preprocessing of ovarian cancer images, including random horizontal flipping, color jittering, random cropping and erasing, includes the following steps:

[0077] Step 1.1: Perform random horizontal flipping operations on the medical image data to generate mirror data and increase data diversity;

[0078] Step 1.2: Perform random rotation on the image with the angle range set to [-30°, 30°] to simulate samples at different shooting angles and help the model adapt to the deformation and changes of the image at different angles;

[0079] Step 1.3: Perform color jittering on the image, and set the perturbation range of each parameter between [0.8, 1.2] to adjust the brightness, contrast, saturation, and hue of the image. Simulate differences in shooting conditions and enhance the diversity and robustness of the data;

[0080] Step 1.4: Perform random cropping on the image with a cropping ratio range of [80%, 100%], adjust the cropped image to a standard size (224×224), increase the spatial transformation of the image, and simulate detection in different areas;

[0081] Step 1.5: Randomly erase some areas in the image. The size range of the erased area is set to [2%, 10%] to simulate the situation where part of the image information is lost, helping the model to better handle incomplete or partially occluded images;

[0082] Step 1.6: Normalize the image and normalize the pixel values ​​of the image to the interval [-1, 1] to ensure data consistency.

[0083] Through the above data preprocessing steps, the diversity of image data can be effectively increased, the generalization ability of the model can be improved, and overfitting can be prevented, providing richer training data for the ovarian cancer image classification task.

[0084] Step 2: Introduce two models, Convolutional Neural Network (CNN) and Visual Transformer (ViT), to capture low-level detail features and high-level semantic features of the image respectively.

[0085] The convolutional neural network is used to extract low-level features of the input image and generate feature vectors containing local information such as edges and textures.

[0086] y=W*X+b

[0087] Among them, X is the input image, W is the convolution kernel, b is the bias term, y is the output feature vector, and * represents the convolution operation.

[0088] The high-level features of the input image are extracted through the visual transformer, and the semantic feature vector is generated using the global self-attention mechanism;

[0089] ViT constructs a global representation of the image by dividing the image into multiple patches and linearly embedding each patch into a vector. Then, the self-attention mechanism is used to weight these patches to capture the relationship between different regions in the image. The calculation formula for self-attention is:

[0090]

[0091] Where Q is the query matrix, K is the key matrix, V is the value matrix, and softmax is used to calculate the attention weights. is the normalization factor. By performing average pooling on all features output by ViT, the global semantic information of the image is extracted. The calculation formula is:

[0092]

[0093] where F i is the output of the i-th feature block, and N is the number of feature blocks.

[0094] Step 3: Adopt the feature adaptive fusion method to dynamically fuse low-level and high-level features to comprehensively learn the local details and global structural information of the image.

[0095] A dynamic weighting mechanism is used to fuse low-level features and high-level features. The specific method is as follows: according to the feature importance of CNN and ViT, a weight generation module g is used to generate (·) , automatically calculate weights based on input features:

[0096] F hybrid =g CNN (F CNN , F ViT )·F CNN +g ViT (F CNN , F ViT )·F ViT

[0097] where g CNN(·) , and g ViT(·) are two dynamic weight generation functions, which are dynamically adjusted according to the importance of input features through model training to satisfy g CNN(·) +g ViT(·) =1.

[0098] The adaptively fused CNN features and ViT features are concatenated to form a comprehensive feature vector for subsequent classification tasks.

[0099] The fused features are input into the classification module, which consists of a fully connected layer and a Sigmoid activation function to output the classification probability value.

[0100] Step 4: Introduce the loss function Balanced Dice Loss that combines Focal Loss and Dice Loss to solve the problems of data category imbalance and difficulty in model prediction in small areas.

[0101] Focal Loss is introduced to solve the problem of class imbalance. It gives higher weights to difficult-to-classify samples and reduces the impact of easy-to-classify samples on the loss, so that the model can focus more on distinguishing difficult samples. The formula is:

[0102] FL(p t )=-α i (1-p i ) γ log(p i )

[0103] Among them, p t is the probability of the correct category predicted by the model, α t is the category balance factor, which is used to adjust the weights of each category, and γ is the focusing factor, which controls the attention paid to difficult-to-classify samples.

[0104] Dice Loss is introduced to optimize the model's prediction ability for tiny areas (such as boundary areas) by measuring the overlap between the prediction and the true label. It is particularly suitable for medical image classification. Its formula is:

[0105]

[0106] Among them, P represents the predicted area, G represents the real area, and Dice Loss optimizes the performance of the model in category segmentation by calculating the ratio of the intersection and the total area.

[0107] Balanced Dice Loss combines the advantages of Focal Loss and Dice Loss to solve the problem of category imbalance and optimize the performance of the model in fine-grained areas. Its formula is:

[0108] Balanced Dice Loss=λ·Focal Loss+(1―λ)·DiceLoss

[0109] Where λ∈[0,1] is the weight coefficient, and the weight ratio of Focal Loss and Dice Loss can be dynamically adjusted according to the characteristics of the data set to achieve the best effect.

[0110] Step 5: Use the Adam optimizer and OneCycleLR learning rate scheduling strategy to train the model to improve the convergence speed and stability of the model during training.

[0111] The Adam optimizer (Adam with Weight Decay) is introduced for model training, combined with momentum and adaptive learning rate adjustment mechanism to improve the efficiency and stability of model parameter update. The formula of Adam optimizer is:

[0112]

[0113] Where: θ t is the model at step t, η is the learning rate, m t is the first-order momentum estimate of the gradient, v t is the second-order momentum estimate of the gradient, and λ is the weight decay coefficient, which is used to control model overfitting.

[0114] The OneCycleLR learning rate scheduling strategy is introduced to dynamically adjust the learning rate according to the number of training rounds to prevent the model from falling into the local optimum and accelerate the training convergence. The learning rate is divided into two stages during the training process: first slowly increase, and then gradually decrease after reaching the maximum value to ensure a more stable training process. Assuming the total number of training steps is T, the formula for the change of the learning rate is:

[0115]

[0116] Among them, η min is the initial learning rate, η max is the maximum learning rate, T up is the number of steps in the learning rate ramp-up phase.

[0117] Step 6: Perform a comprehensive quantitative analysis of the model’s performance by using a variety of performance evaluation metrics during the training and validation phases.

[0118] During the training and validation phases, a variety of evaluation indicators are used to measure the classification effect of the model. Accuracy is used to measure the proportion of correctly classified samples to the total number of samples. The formula is:

[0119]

[0120] Among them, TP is true positive, TN is true negative, FP is false positive, and FN is false negative.

[0121] F1-Score, which comprehensively evaluates the precision and recall of the model, is calculated as follows:

[0122]

[0123] in,

[0124] AUC (Area Under the Curve) is the area under the ROC (Receiver Operating Characteristic) curve, which is used to evaluate the performance of the model at different classification thresholds. The ROC curve is a curve drawn with the false positive rate (FPR) as the horizontal axis and the true positive rate (TPR) as the vertical axis.

[0125]

[0126] in The closer the AUC value is to 1, the stronger the model's ability to distinguish is, and the closer it is to 0.5, the better the model's performance is.

[0127] Different from the existing technology, such as Figure 3-6 As shown, the ovarian cancer image deep analysis method based on intelligent adaptive fusion network under the guidance of precision medicine described in the present invention uses convolutional neural network (CNN) and visual transformer (ViT) to extract low-level features and high-level features of the image, and combines the cross-level feature adaptive fusion method to fully explore the complementarity of low-level features and high-level features, optimize the feature fusion effect, and effectively improve the accuracy and robustness of ovarian cancer image classification. In order to solve the problem of image category imbalance, the present invention introduces the Balanced Dice Loss loss function, which is combined with Focal Loss and DiceLoss, and is particularly suitable for the prediction of small areas in ovarian cancer images. The model performance is evaluated by multiple indicators such as accuracy, F1 score, AUC-ROC, etc. during training and verification, and the model is further hyperparameterized. Finally, the present invention uses a statistical analysis method to evaluate the classification effect of the model to ensure that the early diagnosis of ovarian cancer is more accurate. Through these technologies, the present invention can effectively improve the performance of ovarian cancer image classification and provide important support for the formulation of clinical diagnosis and treatment plans.

[0128] Only the preferred embodiments of the present invention are described in detail above, but the present invention is not limited to the above embodiments. Various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention, and various changes should be included in the protection scope of the present invention.

Claims

1. A method for deep analysis of ovarian cancer images based on an adaptive fusion network, characterized in that: The following steps are involved: S1. First, the input medical image data is preprocessed and the image is uniformly resized to 224×224 using standardized size adjustment technology to ensure the uniformity of the input data; S2. Perform data enhancement on the adjusted images by using techniques such as random horizontal flipping, color jittering, random cropping, and random erasing to increase the diversity of training data and avoid model overfitting. S3, introduce convolutional neural network CNN and visual transformer ViT for feature extraction, CNN is used to extract low-level features in the image, and ViT is used to extract high-level features; S4, based on the adaptive fusion method, the features extracted by CNN and ViT are dynamically fused to adaptively adjust the weights of different features and optimize the feature fusion effect; S5. Use the loss function Balanced Dice Loss, which is a combination of Focal Loss and Dice Loss, for model training to solve the problems of data category imbalance and model prediction difficulties in small areas; S6, introduces the Adam optimizer and OneCycleLR learning rate scheduling strategy to make the model converge quickly during training; S7. The performance of the model was evaluated through various indicators such as accuracy, F1 score, AUC-ROC, etc. during the training and validation process. The final classification effect of the model was evaluated using statistical analysis methods to provide support for the early diagnosis of ovarian cancer.

2. The method for deep analysis of ovarian cancer images based on an adaptive fusion network according to claim 1, characterized in that: The method for preprocessing the input medical image data in S2 is: S2.1, randomly flip the medical image data horizontally to generate mirror data; S2.2, randomly rotate the image, and the angle range is set to [-30°, 30°] to simulate samples at different shooting angles; S2.3, introduce color jitter to adjust the brightness, contrast, saturation and hue of the image respectively, and the perturbation range of each parameter is between [0.8, 1.2]; S2.4, randomly crop the image, set the cropping ratio range to [80%, 100%], and then adjust the cropped image to the standard size (224×224); S2.5, randomly erase some areas in the image, the size range of the erased area is [2%, 10%], simulating the scene of partial information missing; S2.

6. Complete the normalization operation of the enhanced image and normalize the pixel values ​​to the range of [-1,1] to ensure data consistency.

3. The method for deep analysis of ovarian cancer images based on an adaptive fusion network according to claim 1, characterized in that: The method of introducing convolutional neural network CNN and visual transformer ViT for feature extraction in S3 is: The convolutional neural network is used to extract low-level features of the input image and generate feature vectors containing local information such as edges and textures. y=W*X+b Among them, X is the input image, W is the convolution kernel, b is the bias term, y is the output feature vector, and * represents the convolution operation; The high-level features of the input image are extracted through the visual transformer, and the semantic feature vector is generated using the global self-attention mechanism; ViT constructs a global representation of the image by dividing the image into multiple small blocks and linearly embedding each image block into a vector. Then, the self-attention mechanism is used to weight these image blocks to capture the relationship between different regions in the image. The calculation formula of self-attention is: Where Q is the query matrix, K is the key matrix, V is the value matrix, and softmax is used to calculate the attention weights. is the normalization factor; by averaging all features output by ViT, the global semantic information of the image is extracted, and the calculation formula is: Among them, F i is the output of the i-th feature block, and N is the number of feature blocks.

4. The method for deep analysis of ovarian cancer images based on an adaptive fusion network according to claim 1, characterized in that: The method for dynamically fusing the features extracted by CNN and ViT in S4 to adaptively adjust the weights of different features is: A dynamic weighting mechanism is used to fuse low-level features and high-level features. The specific method is as follows: according to the feature importance of CNN and ViT, a weight generation module g is used to generate (·) , automatically calculate weights based on input features: F hybrid =g CNN (F CNN ,F ViT )·F CNN +g ViT (F CNN ,F ViT )·F ViT where g CNN(·) , and g ViT(·) are two dynamic weight generation functions, which are dynamically adjusted according to the importance of input features through model training to satisfy g CNN(·) +g ViT(·) =1; The adaptively fused CNN features and ViT features are concatenated to form a comprehensive feature vector for subsequent classification tasks; The fused features are input into the classification module, which consists of a fully connected layer and a Sigmoid activation function to output the classification probability value.

5. The method for deep analysis of ovarian cancer images based on an adaptive fusion network according to claim 1, characterized in that: The method for model training using the loss function Balanced Dice Loss combined with Focal Loss and Dice Loss in S5 is: Focal Loss is introduced to solve the problem of class imbalance. It gives higher weights to difficult-to-classify samples and reduces the impact of easy-to-classify samples on the loss, so that the model can focus more on distinguishing difficult samples. The formula is: FL(p t )=―a t (1-p t ) γ log(p t ) Among them, p t is the probability of the correct category predicted by the model, α t is the category balance factor, which is used to adjust the weights of each category, and γ is the focus factor, which controls the attention paid to difficult-to-classify samples; Dice Loss is introduced to optimize the model's prediction ability for tiny areas (such as boundary areas) by measuring the overlap between the prediction and the true label. It is particularly suitable for medical image classification. Its formula is: Among them, P represents the predicted area, G represents the real area, and Dice Loss optimizes the performance of the model in category segmentation by calculating the ratio of the intersection and the total area. Balanced Dice Loss combines the advantages of Focal Loss and Dice Loss to solve the problem of category imbalance and optimize the performance of the model in fine-grained areas. Its formula is: Balanced Dice Loss=λ·Focal Loss+(1―λ)·Dice Loss Among them, λ∈[0,1] is the weight coefficient, and the weight ratio of Focal Loss and Dice Loss can be dynamically adjusted according to the characteristics of the data set to achieve the best effect.

6. The method for deep analysis of ovarian cancer images based on an adaptive fusion network according to claim 1, characterized in that: The method of introducing the Adam optimizer and the OneCycleLR learning rate scheduling strategy in S6 is: The Adam optimizer is introduced for model training, combined with momentum and adaptive learning rate adjustment mechanism to improve the efficiency and stability of model parameter updating; the formula of the Adam optimizer is: Where: θ t is the model at step t, η is the learning rate, m t is the first-order momentum estimate of the gradient, v t is the second-order momentum estimate of the gradient, and λ is the weight decay coefficient, which is used to control model overfitting; The OneCycleLR learning rate scheduling strategy is introduced to dynamically adjust the learning rate according to the number of training rounds to prevent the model from falling into the local optimum and accelerate the training convergence. The learning rate is divided into two stages during the training process: first slowly increase, and then gradually decrease after reaching the maximum value to ensure a more stable training process. Assuming the total number of training steps is T, the formula for the change of the learning rate is: Among them, η min is the initial learning rate, η max is the maximum learning rate, T up is the number of steps in the learning rate ramp-up phase.

7. The method for deep analysis of ovarian cancer images based on an adaptive fusion network according to claim 1, characterized in that: The method for evaluating the performance of the model through multiple indicators in the training and verification process in S7 is: A variety of evaluation indicators are used to measure the classification effect of the model. Accuracy is used to measure the proportion of correctly classified samples to the total number of samples. The formula is: Among them, TP is true positive, TN is true negative, FP is false positive, and FN is false negative; F1-Score is the comprehensive evaluation model's precision and recall. Its formula is: in, AUC (Area Under the Curve) is the area under the ROC (Receiver Operating Characteristic) curve, which is used to evaluate the performance of the model under different classification thresholds. The ROC curve is a curve drawn with the false positive rate FPR as the horizontal axis and the true positive rate TPR as the vertical axis. AUC=∫0 1 TPR(t)·d(FPR(t)) in The closer the AUC value is to 1, the stronger the model's ability to distinguish is, and the closer it is to 0.5, the better the model's performance is.

Images