Neural network model for vulva disease classification and training method thereof

Abstract

The invention discloses a vulva disease classification neural network model and a training method thereof, and the model comprises an image obtaining module which is used for obtaining a vulva image and inputting the vulva image into a Swin Transformer module; the Swin Transformer module is used for segmenting the vulva image into a plurality of non-overlapped image blocks with fixed sizes, converting each image block into an image block sequence, carrying out hierarchical multi-scale feature extraction on the image block sequence through an attention mechanism to obtain a multi-scale hierarchical feature image, carrying out vulva disease category prediction on the feature image, and carrying out vulva disease category prediction on the multi-scale hierarchical feature image to obtain a vulva disease category prediction result; outputting the predicted vulva disease category and the corresponding probability; through the neural network model for vulva disease classification, full-process automation from image input to disease prediction is realized, subjective deviation of manual diagnosis is avoided, and the accuracy of vulva disease diagnosis and classification is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of neural network models, and in particular to a neural network model for vulvar disease classification and a training method thereof. Background Art

[0002] Because vulvar diseases are numerous, have complex clinical manifestations and special locations, and are located at the intersection of gynecology and dermatology, they are blind spots and difficulties for gynecologists and dermatologists.

[0003] Traditional diagnosis of vulvar diseases relies primarily on the physician's subjective experience and is subject to visual limitations. However, due to the differences between gynecology and dermatology, physicians in different departments may have limited knowledge and understanding of clinical diagnosis and treatment. This can lead to inadequate early screening, resulting in numerous misdiagnoses and missed diagnoses, and delayed treatment. Consequently, patients miss the optimal treatment window and suffer immensely. Furthermore, vulvar disease image datasets are much smaller than those for common skin diseases, and annotation requires the involvement of specialized gynecologists or dermatologists, which is costly.

[0004] Therefore, the existing technology has obvious deficiencies in the diagnosis and classification of vulvar diseases, and more advanced methods are needed to improve the accuracy and timeliness of the diagnosis and classification of vulvar diseases. It is necessary to provide a neural network model for vulvar disease classification and its training method to solve the above problems. Summary of the Invention

[0005] The present invention provides a neural network model for vulvar disease classification and a training method thereof, so as to solve the problem of inaccurate diagnosis and classification of vulvar diseases in the prior art.

[0006] In a first aspect, an embodiment of the present invention provides a neural network model for vulvar disease classification, comprising:

[0007] An image acquisition module, configured to acquire a vulva image and input the vulva image into a Swin Transformer module;

[0008] The Swin Transformer module is used to segment the vulva image into multiple non-overlapping fixed-size tiles, convert each tile into a tile sequence, perform hierarchical multi-scale feature extraction on the tile sequence through an attention mechanism, obtain a multi-scale, hierarchical feature map, predict the vulvar disease category on the feature map, and output the predicted vulvar disease category and corresponding probability.

[0009] Optionally, also include:

[0010] Diagnostic suggestion generation is used to generate corresponding diagnostic suggestions based on the expert database and the predicted vulvar disease category and corresponding probability.

[0011] Optionally, the Swin Transformer module includes:

[0012] An image preprocessing layer, configured to segment the vulva image into non-overlapping fixed-size tiles, and convert each of the tiles into a tile sequence by flattening and linear projection;

[0013] A hierarchical multi-scale feature extraction layer, which feeds the tile sequence into a hierarchical network structure consisting of four stages, each of which contains a pair of sliding window Transformer blocks. Each sliding window Transformer block extracts a multi-scale, hierarchical feature map through local window self-attention and shifting mechanisms.

[0014] A pooling layer, configured to perform global average pooling on the multi-scale, hierarchical feature map to integrate the multi-scale features into a feature vector of fixed length;

[0015] The classification layer is used to output the predicted vulvar disease category and the corresponding probability based on the feature vector and the Softmax function.

[0016] Optionally, the vulvar disease categories include: at least one of vulvar lichen sclerosus, vulvar lichen simplex chronicus, vulvar lichen planus, vulvar vitiligo, vulvar HPV-related precancerous lesions, vulvar HPV-unrelated precancerous lesions, vulvar Paget's disease and vulvar squamous cell carcinoma.

[0017] In a second aspect, an embodiment of the present invention provides a neural network model training method for vulvar disease classification, comprising:

[0018] Obtaining an original sample data set, and preprocessing the original sample data set to obtain a target sample data set;

[0019] Divide the target sample data set into a training set, a validation set, and a test set according to a set ratio;

[0020] Based on the training set, validation set and test set, the pre-trained neural network model for vulvar disease classification is trained by cross-validation method to obtain a trained neural network model for vulvar disease classification.

[0021] Optionally, the preprocessing of the original sample data set to obtain a target sample data set includes:

[0022] The original sample data set is expanded using the OpenCV library to obtain a target sample data set.

[0023] Optionally, the pre-trained neural network model for vulvar disease classification is a neural network model for vulvar disease classification with a frozen hierarchical multi-scale feature extraction layer;

[0024] Accordingly, the neural network model for vulvar disease classification is pre-trained, including:

[0025] Based on the ImageNet dataset, the hierarchical multi-scale feature extraction layer in the neural network model for vulvar disease classification is pre-trained, and the parameters of the pre-trained hierarchical multi-scale feature extraction layer are frozen to obtain a pre-trained neural network model for vulvar disease classification.

[0026] Optionally, the pre-trained neural network model for vulvar disease classification is trained by a cross-validation method based on the training set, the validation set, and the test set to obtain a trained neural network model for vulvar disease classification, including:

[0027] Iteratively train the pre-trained neural network model for vulvar disease classification based on the training set;

[0028] Each time it is iterated, the neural network model for vulvar disease classification after iterative training is verified based on the validation set, and the model parameters and validation indicators corresponding to the current iteration are saved;

[0029] After all iterations of the training set are completed, the model with the best verification index is selected as the candidate neural network model for vulvar disease classification;

[0030] The candidate neural network model for vulvar disease classification is tested based on the test set. If the test passes, the candidate neural network model for vulvar disease classification is used as the trained neural network model for vulvar disease classification.

[0031] Optionally, the step of testing the candidate neural network model for vulvar disease classification based on a test set, and using the candidate neural network model for vulvar disease classification as a trained neural network model for vulvar disease classification if the test passes, includes:

[0032] Testing the candidate neural network model for vulvar disease classification based on the test set, and calculating performance indicators corresponding to the candidate neural network model for vulvar disease classification;

[0033] If the performance indicator is within a preset performance indicator threshold range, the candidate neural network model for vulvar disease classification is used as a trained neural network model for vulvar disease classification.

[0034] (1) The technical solution of the embodiment of the present invention is to construct a neural network model for vulvar disease classification. In terms of model architecture design, by introducing the Swin Transformer module and constructing a hierarchical multi-scale feature extraction mechanism, the module divides the vulvar image into a fixed-size block sequence, and uses the local window attention and shift window mechanism to collaboratively extract multi-scale features, directly achieving efficient modeling of the local details and global structure of the lesion, significantly improving the feature representation capability. Compared with the traditional CNN model that is limited by the fixed receptive field or the computational redundancy of the full-attention Transformer, this design reduces the computational complexity while maintaining the ability to capture long-range dependencies, effectively improving the model's recognition accuracy for complex lesions such as the depigmentation of vulvar lichen sclerosus and the ulcer morphology of vulvar squamous cell carcinoma.

[0035] (2) At the data preprocessing level, to address the core pain points of vulvar disease data scarcity and imbalanced categories, the OpenCV library was used to perform targeted enhancement on the original samples, and a balanced dataset was constructed using pathologically confirmed samples, directly addressing the model bias problem caused by sample skew. Compared with existing technologies that rely on manual experience due to insufficient data and have a high rate of missed diagnosis, this method effectively improves the recall rate of rare diseases (such as vulvar Paget's disease and HPV-unrelated precancerous lesions) and significantly enhances the robustness of the model in complex clinical scenarios.

[0036] (3) In terms of model training, a two-stage strategy of "ImageNet pre-training freeze + cross-validation fine-tuning" is adopted. The parameters of the feature extraction layer are pre-trained based on ImageNet transfer learning, and the classification layer is dynamically optimized using cross-validation. This directly reduces the model's dependence on small-scale labeled data, greatly accelerates the model convergence speed, and effectively improves training efficiency. Compared with the bottleneck of traditional end-to-end training requiring tens of thousands of labeled samples, this method only requires thousands of samples to achieve a high verification accuracy rate, and the generalization ability and stability of the model are ensured by the hard threshold of the test set (such as sensitivity ≥ 98%), ensuring that it can still maintain high accuracy when facing vulvar images of diverse sources and varying quality in real clinical environments.

[0037] (4) In terms of clinical application value, by adding a diagnostic suggestion generation module and linking with the expert database, individualized diagnosis and treatment plans are automatically generated based on the predicted probability, directly converting the binary classification results into actionable clinical decision support. Compared with the existing pure classification model that lacks treatment guidance and relies on the doctor's secondary judgment, this design shortens the time spent on the entire diagnosis process and improves the standardization of primary care doctors in handling complex vulvar diseases. At the same time, it assists doctors in assessing confidence through probability visualization (such as prompting re-examination at low probability), effectively reducing the risk of missed diagnosis / misdiagnosis.

[0038] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present invention, nor is it intended to limit the scope of the present invention. Other features of the present invention will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0040] Figure 1 A network architecture diagram of a neural network model for vulvar disease classification provided in Example 1 of the present invention;

[0041] Figure 2 A flowchart of a neural network model training method for vulvar disease classification provided in Example 2 of the present invention;

[0042] Figure 3 This is a flowchart of a specific neural network model training method for vulvar disease classification provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0043] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0044] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0045] Example 1:

[0046] Figure 1 This is a network architecture diagram of a neural network model for vulvar disease classification provided in Example 1 of the present invention. This embodiment is applicable to the diagnosis and classification of vulvar diseases, and the model can be applied to a terminal or a server. The terminal can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, and portable wearable devices, and the server can be implemented as an independent server or a server cluster composed of multiple servers. The server can use the neural network model for vulvar disease classification and its corresponding training method to complete the diagnosis and classification task of vulvar diseases, and send the classification prediction results of the model to the terminal for the terminal user to view and analyze. Figure 1 As shown, the model includes:

[0047] An image acquisition module, configured to acquire a vulva image and input the vulva image into a Swin Transformer module;

[0048] Among them, the vulva image can refer to a color image containing the structure and morphology of the female external genitalia (i.e., the vulva); the SwinTransformer module can refer to a module that extracts features from the vulva image and predicts the corresponding vulva disease category and corresponding probability, for example, the Swin Transformer Tiny architecture.

[0049] Specifically, the vulva image can be obtained through a terminal device equipped with a camera, such as an electronic colposcope; and the obtained vulva image is uploaded to the Swin Transformer module.

[0050] The Swin Transformer module is used to segment the vulva image into multiple non-overlapping fixed-size tiles, convert each tile into a tile sequence, perform hierarchical multi-scale feature extraction on the tile sequence through an attention mechanism, obtain a multi-scale hierarchical feature map, predict the vulvar disease category on the feature map, and output the predicted vulvar disease category and corresponding probability.

[0051] Among them, fixed-size tiles can refer to dividing the input vulva image into non-overlapping local areas of fixed size. Tile sequence can refer to flattening each tile into a vector and linearly projecting it to a hidden dimension (such as C = 96). The attention mechanism can refer to reducing computational complexity by calculating local attention. Hierarchical multi-scale features can refer to the hierarchical features output by each stage obtained by downsampling at different stages, and the features output by the previous stage can be used as the features input by the next stage; then, multi-scale information is integrated through jump connections to achieve cross-stage feature fusion and obtain multi-scale features. Multi-scale hierarchical feature maps can refer to high-resolution shallow features and low-resolution deep features; vulvar disease category can refer to the type of vulvar lesion determined by vulvar image; and the corresponding probability of vulvar disease category can refer to the probability of vulvar disease belonging to a certain category.

[0052] Specifically, such as Figure 1 As shown, the normalized vulva image is divided into non-overlapping fixed-size tiles (e.g., 4×4 pixels). Each tile is converted into a low-dimensional feature vector by flattening (4×4×3 → 48 dimensions) and linear projection (48 → C, e.g., C = 96). All tile features are arranged in order to form a tile sequence (e.g., length = H / 4×W / 4). SwinTransformer gradually extracts multi-scale features through four stages. Each stage contains multiple SwinTransformer Block pairs (window attention + shifted window attention). For example, the data of the four stages in Swin Transformer are shown in Table 1:

[0053] Table 1

[0054]

[0055]

[0056] Among them, self-attention is calculated within a local window (such as 7×7) to reduce computational complexity while retaining local features. By shifting the window (for example, shifting half a window to the upper left), cross-window information interaction is achieved, solving the problem of information isolation in the local window. By merging the blocks of each stage (such as 2×2 merging), not only the resolution is reduced, but the number of channels is doubled, achieving multi-scale feature fusion, and obtaining multi-scale hierarchical feature maps of multiple channels. Gradually transition from high-resolution local features (Stage 1) to low-resolution global features (Stage 4), taking into account both details and semantic information. The feature map of each channel is average pooled in the spatial dimension (i.e., height and width), and a fixed-length one-dimensional feature vector is output, which is equal to the number of channels. The predicted vulvar disease category and corresponding probability are then output through a linear classifier (i.e., fully connected layer).

[0057] In this embodiment, a neural network model for vulvar disease classification is constructed. The model includes an image acquisition module that acquires a vulvar image and inputs the vulvar image into a Swin Transformer module. The Swin Transformer module segments the vulvar image into multiple non-overlapping fixed-size tiles, converts each tile into a tile sequence, and performs hierarchical multi-scale feature extraction on the tile sequence using an attention mechanism to obtain a multi-scale, hierarchical feature map. The vulvar disease category is predicted from the feature map, and the predicted vulvar disease category and corresponding probability are output. This solves the problem of inaccurate diagnosis and classification of vulvar diseases in the prior art, automates the entire process from image input to disease prediction, avoids the subjective bias of manual diagnosis, and outputs a probability of the classification result to assist doctors in assessing confidence. Furthermore, the Swin Transformer is applied to the field of vulvar diseases. Its shifting window mechanism reduces computational complexity while maintaining the ability to model long-range dependencies of vulvar images, thereby improving the accuracy of vulvar disease diagnosis and classification.

[0058] Optionally, before the Swin Transformer module segments the received vulva image, it is necessary to ensure that the image size is uniform and to align it with a pre-trained model through normalization; wherein the pre-trained model may refer to a neural network model for vulvar disease classification that freezes the four stages of feature extraction.

[0059] Specifically, the ImageNet dataset can be used to pre-train the four feature extraction stages of an initial neural network model for vulvar disease classification. The parameters of these four pre-trained stages are then frozen to create a pre-trained neural network model for vulvar disease classification. This effectively improves training efficiency, enabling faster model convergence in the early stages of training, reducing reliance on large-scale labeled vulvar disease data, and lowering computational costs, significantly saving computing resources and time.

[0060] In this embodiment, before the Swin Transformer module segments the received vulva image, the image size is ensured to be uniform and aligned with the pre-trained model through normalization processing; ensuring perfect matching with the preset tile segmentation mechanism of Swin Transformer to avoid feature extraction errors or boundary information loss due to size differences; avoiding the additional overhead caused by dynamic adjustment, which is conducive to the acceleration of image processing.

[0061] Optionally, the model also includes:

[0062] The diagnosis suggestion generating module is used to generate corresponding diagnosis suggestions based on the expert database and the predicted vulvar disease categories and corresponding probabilities.

[0063] The expert database includes various expert consensuses and the latest clinical treatment guidelines. Diagnostic recommendations can refer to treatment recommendations corresponding to the vulvar disease category and its predicted probability. For example, if the vulvar disease category is VLSC with a predicted probability of 98.79%, the corresponding treatment recommendations generated by the expert database may be: 1. No biopsy required. 2. Mometasone once daily or 0.1% tacrolimus twice daily. Recheck the vulvar condition every 3-4 weeks. If the vulva has fully recovered, discontinue the medication. If the vulvar lesions have diminished, reduce the dosage until the vulvar skin returns to normal and discontinue the medication.

[0064] Specifically, a correspondence between vulvar disease categories and their probabilities and the expert database can be established in advance. After the predicted vulvar disease category and the corresponding probability are determined, the corresponding treatment recommendations can be directly determined from the expert database based on the correspondence, assisting doctors in making quick and accurate diagnosis and treatment decisions.

[0065] In this embodiment, the diagnosis suggestion generation module can directly generate standardized treatment suggestions based on the predicted vulvar disease category and corresponding probability through the expert database, providing doctors with accurate auxiliary decision support, improving diagnostic efficiency, and reducing the workload of doctors.

[0066] Optionally, the Swin Transformer module includes:

[0067] An image preprocessing layer, configured to segment the vulva image into non-overlapping fixed-size tiles, and convert each of the tiles into a tile sequence by flattening and linear projection;

[0068] Flattening refers to the process of converting a two-dimensional image tile from a matrix into a one-dimensional vector, preserving the original pixel values ​​but removing the spatial structure. Linear projection, on the other hand, refers to the process of mapping a high-dimensional pixel vector into a low-dimensional feature space using a learnable weight matrix, which is used to extract abstract features and adapt to the input dimensions of the Transformer.

[0069] Specifically, the uniformly sized and normalized vulva image is segmented into non-overlapping fixed-size tiles using non-overlapping windows. Each tile is flattened and linearly projected into a low-dimensional feature vector. A learnable positional encoding is added to each tile feature to preserve spatial location information. All tile features are arranged in sequence, and the final tile sequence (length = H / 4 × W / 4) is output.

[0070] A feature extraction layer, configured to input the tile sequence into a hierarchical network structure consisting of four stages, each of which contains a pair of sliding window Transformer blocks. Each sliding window Transformer block extracts a multi-scale hierarchical feature map through window self-attention and shifting mechanisms.

[0071] The hierarchical network structure can refer to a multi-scale feature extraction architecture consisting of four progressively downsampling stages; each stage processes input features via a sliding window Transformer block; the sliding window Transformer block consists of window self-attention and its shifted version (i.e., shifted window self-attention). Window self-attention can refer to computing self-attention within a local window (e.g., 7×7) to reduce computational complexity. Shifted window self-attention can refer to enabling cross-window information interaction by shifting the window; for example, shifting the window to the lower right by (window size / 2) pixels.

[0072] Specifically, such as Figure 1 As shown in the figure, the feature extraction layer is a hierarchical network structure consisting of four progressive downsampling stages (Stage1-Stage4), each of which contains a pair of sliding window Transformer blocks. Each sliding window Transformer block can improve computational efficiency through the local window self-attention mechanism; in particular, the Swin Transformer has significant advantages when processing high-resolution images. The shifted window self-attention is used to achieve cross-window information interaction, solving the problem of information isolation at the window boundary. Through the patch merging (such as 2×2 merging) downsampling method, the resolution is reduced (i.e., the resolution is halved) while the channel dimension is increased (i.e., the number of channels is doubled at each stage), gradually extracting feature maps of different levels and scales.

[0073] A pooling layer, configured to perform global average pooling on the multi-scale hierarchical feature map to integrate the multi-scale features into a feature vector of fixed length;

[0074] Global Average Pooling (GAP) refers to taking the average of the spatial dimensions (height and width) of each channel to generate a 1×C vector. For example, a feature map with 768 channels will be converted into a 768-dimensional vector. Specifically, by performing a global average pooling (GAP) operation on the feature map output from the final stage (Stage 4), multi-level and multi-scale features are integrated into a fixed-length feature vector. This feature vector is a global semantic vector that combines local details and global context information.

[0075] The classification layer is used to output the predicted vulvar disease category and the corresponding probability based on the feature vector and the Softmax function.

[0076] The Softmax function may refer to converting unnormalized scores (logits) into a probability distribution such that the sum of the probabilities of all categories is 1.

[0077] Specifically, the feature vector output by global average pooling (GAP) is mapped to the number of target categories (e.g., the total number of vulvar disease categories) to generate unnormalized scores (logits). The unnormalized scores (logits) are converted into a probability distribution. Based on the probability distribution, the category with the highest probability is selected as the predicted result, i.e., the predicted vulvar disease category. This probability is the corresponding probability of the predicted vulvar disease category.

[0078] In this embodiment, the vulvar image is segmented into multiple non-overlapping fixed-size tiles using the Swin Transformer module, and each tile is converted into a tile sequence. The tile sequence is subjected to hierarchical multi-scale feature extraction through the attention mechanism to obtain a multi-scale, hierarchical feature map. The vulvar disease category is predicted on the feature map, and the predicted vulvar disease category and corresponding probability are output. This achieves collaborative modeling of local lesions and global features, significantly improving the recognition accuracy of complex vulvar lesions (such as white spots, ulcers, etc.); the hierarchical processing of the tile sequence can simultaneously capture micro-textures (such as abnormal skin keratinization) and macro-structural features (such as lesion area distribution), overcoming the defects of the limited receptive field of traditional CNN; the entire process from image input to disease prediction is automated, avoiding the subjective bias of manual diagnosis, and the classification results are accompanied by probability output to assist doctors in assessing confidence. Furthermore, the Swin Transformer is applied to the field of vulvar diseases. Its shifting window mechanism reduces the amount of computation while maintaining the ability to model long-range dependencies of vulvar images, thereby improving the accuracy of vulvar disease diagnosis and classification.

[0079] Optionally, the vulvar disease categories include: at least one of vulvar lichen sclerosus, vulvar lichen simplex chronicus, vulvar lichen planus, vulvar vitiligo, vulvar HPV-related precancerous lesions, vulvar HPV-unrelated precancerous lesions, vulvar Paget's disease and vulvar squamous cell carcinoma.

[0080] Specifically, vulvar lichen sclerosus (VLS) is a rare chronic inflammatory skin disease characterized by thinning, pallor, loss of elasticity and itching of the skin, which can lead to anatomical changes and functional disorders in severe cases. Vulvar lichen simplex chronicus (VLSC) is a chronic vulvar skin disease characterized by thickened and leathery skin with obvious hypopigmentation or loss of pigmentation. Vulvar lichen planus (VLP) is a chronic inflammatory skin and mucosal disease characterized by purple-red or white flat papules and plaques with reticular white lines (Wickham lines) visible on the surface, often accompanied by erosions or ulcers. Vulvar vitiligo is an acquired depigmentation disease characterized by the appearance of clearly demarcated milky white or porcelain white patches on the vulva, usually without subjective symptoms (such as itching or pain). HPV-related vulvar precancerous lesions, also known as vulvar high-grade squamous intraepithelial lesions (VHSIL), are a type of vulvar precancerous lesion caused by HPV infection and characterized by single or multiple erythematous, leukoplastic, hyperpigmented, verrucous, or plaque-like lesions. HPV-unrelated vulvar precancerous lesions include differentiated intraepithelial neoplasia (dVIN), vulvar acanthosis with altered differentiation (VAAD), and differentiated exophytic vulvar intraepithelial lesions (DEVIL). DEVIL is associated with chronic inflammation and is characterized by single, poorly defined white or erythematous lesions with a rough surface, hyperkeratosis, or erosions. Variant differentiated vulvar acanthosis is a condition resembling chronic vulvar dermatitis characterized by thickened, hyperkeratotic plaques. Differentiated exophytic vulvar intraepithelial lesions are exophytic (verrucous or papillary) growths characterized by well-differentiated squamous epithelium with abnormal keratinization and the absence of HPV-related changes. Vulvar Paget's disease is a rare intraepidermal adenocarcinoma characterized by the presence of Paget cells within the epidermis and clinically presenting as chronic eczematous erythema. Vulvar squamous cell carcinoma (VSCC) is a vulvar malignancy characterized by erythematous, leukoplakia, nodular, or verrucous lesions.

[0081] In this embodiment, through the neural network model of vulvar disease classification, based on the characteristics of each vulvar disease category, accurate diagnosis of vulvar disease categories and corresponding probabilities is achieved, avoiding missed diagnosis and misdiagnosis, significantly improving the early identification rate of vulvar diseases, and facilitating early intervention and treatment of vulvar squamous cell carcinoma and precancerous lesions; and based on the expert database, it assists in formulating personalized treatment recommendations, provides doctors with accurate auxiliary decision support, improves diagnostic efficiency, and reduces the workload of doctors.

[0082] Example 2:

[0083] Figure 2 This is a flowchart of a neural network model training method for vulvar disease classification provided by the second embodiment of the present invention. This embodiment is applicable to the case of training any of the above-mentioned neural network models for vulvar disease classification. Figure 2 As shown, the method includes:

[0084] S110 : Acquire an original sample data set, and preprocess the original sample data set to obtain a target sample data set.

[0085] The original sample dataset may refer to an original sample dataset obtained by collecting standard samples of vulva images through an electronic colposcope. The preprocessing may be an extended process. The target sample dataset may refer to the original sample dataset after the extension process.

[0086] Specifically, the image data in the original sample dataset includes vulvar lichen sclerosus (VLS), vulvar lichen simplex chronicus (VLSC), vulvar lichen planus, vulvar vitiligo, HPV-related vulvar precancerous lesions (high-grade vulvar lesions (VHSIL)), HPV-unrelated vulvar precancerous lesions (differentiated intraepithelial neoplasia (dVIN), variant differentiated vulvar acanthosis (VAAD), and differentiated exophytic vulvar intraepithelial lesion (DEVIL), vulvar Paget's disease, vulvar cancer, and vulvar images from healthy individuals. With the exception of normal vulvar images, all vulvar disease images have been pathologically confirmed. All disease and healthy individuals have signed informed consent for enrollment; the enrolled patients range in age from 5 to 95 years. All image data will be anonymized after collection and stored in the original sample database. The category distribution of each image data in the original sample dataset is shown in Table 2.

[0087] Table 2

[0088] Vulvar disease categories Number of images Composition ratio (%) Vulvar lichen sclerosus (VLS) 965 0.349 Vulvar lichen simplex chronicus (VLSC) 648 0.234 vulvar lichen planus 44 0.016 Vulvar vitiligo 44 0.016 HPV-related precancerous lesions of the vulva 586 0.212 Paget's disease of the vulva 126 0.045 vulvar squamous cell carcinoma 77 0.028 HPV-unrelated precancerous lesions of the vulva 45 0.016 normal 229 0.083

[0089] As shown in Table 1, the distribution of disease images in the original sample dataset is severely unbalanced, which can easily lead to overfitting during model training. Therefore, this paper uses the OpenCV library to perform image expansion techniques such as random image rotation, brightness perturbation, chromaticity perturbation, local contrast enhancement, and random horizontal flipping on the original sample dataset. This effectively improves the diversity and effectiveness of the sample data, significantly increasing image variability while maintaining the integrity of lesion structural features. The total number of images after preprocessing has been expanded to 16,868, and the category distribution has become more balanced, as shown in Table 3.

[0090] Table 3

[0091]

[0092]

[0093] In this embodiment, by expanding the original sample data set, a target sample data set with balanced category distribution is obtained, which significantly improves the stability of model training and classification accuracy, while avoiding model bias caused by data skewness and ensuring the reliability of vulvar disease classification prediction results.

[0094] S120: Divide the target sample data set into a training set, a validation set, and a test set according to a set ratio.

[0095] Among them, the set ratio can refer to a ratio of 7:2:1; the training set can refer to a sample data set for training a neural network model for vulvar disease classification; the validation set can refer to a sample data set for verifying the trained neural network model for vulvar disease classification; and the test set can refer to a sample data set for testing the trained neural network model for vulvar disease classification.

[0096] Specifically, the target sample dataset can be randomly divided into a training set, a validation set, and a test set in a ratio of 7:2:1. The number of sample images in the training set, validation set, and test set can be 11,807, 3,373, and 1,688, respectively. During the division process, it is ensured that the training set, validation set, and test set do not overlap.

[0097] In this embodiment, by setting a ratio to divide the target sample data set into a non-overlapping training set, a validation set, and a test set, the generalization ability of the neural network model training for vulvar disease classification can be effectively improved.

[0098] S130. Based on the training set, the validation set, and the test set, the pre-trained neural network model for vulvar disease classification is trained by a cross-validation method to obtain a trained neural network model for vulvar disease classification.

[0099] Among them, cross-validation is a commonly used model evaluation and optimization technique in statistics and machine learning. Its core idea is to divide the data set multiple times and perform training and verification in a cycle to more accurately evaluate the generalization ability of the model and avoid evaluation bias caused by the randomness of a single data division.

[0100] Specifically, the pre-trained neural network model for vulvar disease classification is trained using the training set, and indicators, such as accuracy, are calculated on the current validation set. After the training of the training set is completed, the calculated indicators of each training validation are summarized to determine the model with the optimal calculation indicators; the optimal model is tested using the test set. If it meets the preset performance indicator threshold (such as the accuracy threshold), the optimal model is determined to be the trained neural network model for vulvar disease classification; if not, the neural network model for vulvar disease classification is retrained until a trained neural network model for vulvar disease classification is obtained.

[0101] In this embodiment, by expanding the original sample data set, a target sample data set with balanced category distribution is obtained, which significantly improves the stability of model training and classification accuracy, while avoiding model bias caused by data skew, and ensuring the reliability of vulvar disease classification prediction results; by setting a ratio to divide the target sample data set into non-overlapping training set, validation set and test set, the generalization ability of the neural network model training for vulvar disease classification can be effectively improved; through the training set, validation set and test set, the pre-trained neural network model for vulvar disease classification is trained by cross-validation method to obtain a trained neural network model for vulvar disease classification, which can effectively balance the model performance and provide a more reliable model foundation for subsequent clinical deployment.

[0102] Optionally, the preprocessing of the original sample data set to obtain a target sample data set includes:

[0103] The original sample data set is expanded by using the Open CV library to obtain a target sample data set.

[0104] Among them, Open CV (Open Source Computer Vision Library) is an open source computer vision and machine learning software library that provides a wealth of tools and algorithms to support development from basic image processing to complex real-time vision applications.

[0105] Specifically, the following enhancement operations were performed on the images in the original sample dataset using the Open CV library: random rotation within the range of 40° to 280°, brightness perturbation, chromaticity perturbation, local contrast enhancement, and random horizontal flipping. Furthermore, to improve the diversity and effectiveness of the enhancement effect, the above operations were applied to each image in a random combination of 1 to 3, thereby significantly increasing the image variability while maintaining the integrity of the lesion structural features. After data enhancement, the total number of images in the target sample dataset expanded to 16,868, and the category distribution tended to be balanced.

[0106] In this embodiment, the original sample dataset is expanded through the Open CV library to obtain the target sample dataset, which significantly improves the stability of model training and classification accuracy, while avoiding model bias caused by data skewness and ensuring the reliability of vulvar disease classification prediction results.

[0107] Optionally, the pre-trained neural network model for vulvar disease classification is a neural network model for vulvar disease classification with a frozen feature extraction layer;

[0108] Accordingly, the neural network model for vulvar disease classification is pre-trained, including:

[0109] Based on the ImageNet dataset, the feature extraction layer in the neural network model for vulvar disease classification is pre-trained, and the parameters of the pre-trained feature extraction layer are frozen to obtain a pre-trained neural network model for vulvar disease classification.

[0110] The ImageNet dataset may refer to one of the most widely used benchmark datasets in the field of computer vision, containing approximately 1.2 million training images, 50,000 validation images, and 100,000 test images. The feature extraction layer may refer to a hierarchical network structure comprising four progressive downsampling stages (Stage 1-Stage 4).

[0111] Specifically, the feature extraction layer in the ImageNet pre-trained neural network model for vulvar disease classification can be used as initialization, and the parameters of the pre-trained feature extraction layer can be frozen to obtain the pre-trained neural network model for vulvar disease classification. Then, the subsequent layers (such as the classification layer) can be fine-tuned through transfer learning to adapt to the vulvar disease classification task.

[0112] In this embodiment, the feature extraction layer in the neural network model of vulvar disease classification pre-trained by ImageNet is used as initialization, and then the subsequent layers are fine-tuned through transfer learning to adapt to the vulvar disease classification task, which effectively improves the training efficiency, enables the model to converge faster in the early training stage, and reduces the dependence on large-scale labeled data of vulvar diseases; enhances the generalization ability and reduces the risk of overfitting; reduces the computational cost and significantly saves computing resources and time.

[0113] Optionally, the pre-trained neural network model for vulvar disease classification is trained by a cross-validation method based on the training set, the validation set, and the test set to obtain a trained neural network model for vulvar disease classification, including:

[0114] Iteratively train the pre-trained neural network model for vulvar disease classification based on the training set;

[0115] Each time it is iterated, the neural network model for vulvar disease classification after iterative training is verified based on the validation set, and the model parameters and validation indicators corresponding to the current iteration are saved;

[0116] After all iterations of the training set are completed, the model with the best verification index is selected as the candidate neural network model for vulvar disease classification;

[0117] The candidate neural network model for vulvar disease classification is tested based on the test set. If the test passes, the candidate neural network model for vulvar disease classification is used as the trained neural network model for vulvar disease classification.

[0118] Among them, the validation indicators can refer to accuracy, precision, recall, F1 score, AUC-ROC, confusion matrix, etc., and at least one of sensitivity (reducing missed diagnosis), specificity (reducing misdiagnosis), positive predictive value (PPV), etc.

[0119] Specifically, Figure 3 This is a flowchart of a specific neural network model training method for vulvar disease classification provided by Example 2 of the present invention. Figure 3As shown, the pre-trained neural network model for vulvar disease classification is iteratively trained using the training set; after each iteration, the neural network model for vulvar disease classification after the current iteration training is verified using the validation set, and the model parameters and validation indicators corresponding to the current iteration are saved to evaluate the performance of the current iteration model. When all iterations of the training set are completed, the model with the best validation indicator is selected as the candidate neural network model for vulvar disease classification. The candidate neural network model for vulvar disease classification is then tested using the test set. The test set is only used once, and the model or model parameters cannot be adjusted based on the test results. If the test passes, the candidate neural network model for vulvar disease classification is used as the trained neural network model for vulvar disease classification. If the test fails, the model is retrained and / or the model deployment is updated.

[0120] In this embodiment, based on the training set, validation set and test set, the pre-trained neural network model for vulvar disease classification is trained by cross-validation to obtain a trained neural network model for vulvar disease classification; cross-validation reduces evaluation errors through multiple data partitioning (such as K-fold cross-validation), ensuring that the model performance evaluation is more robust; maximizes the use of training samples, avoids the key information that may be lost in a single partition, and improves the generalization ability of the model. The model parameters can be adjusted more accurately through the validation set and cross-validation to prevent overfitting or underfitting; the average performance of multiple cross-validation results can reduce the impact of randomness, make the model training process more statistically significant, support reliability verification of clinical applications, and combine quantitative evaluation with hard standards through the test set to ensure the reliability of the final deployed model.

[0121] Optionally, the step of testing the candidate neural network model for vulvar disease classification based on a test set, and using the candidate neural network model for vulvar disease classification as a trained neural network model for vulvar disease classification if the test passes, includes:

[0122] Testing the candidate neural network model for vulvar disease classification based on the test set, and calculating performance indicators corresponding to the candidate neural network model for vulvar disease classification;

[0123] If the performance indicator is within a preset performance indicator threshold range, the candidate neural network model for vulvar disease classification is used as a trained neural network model for vulvar disease classification.

[0124] Among them, the performance indicator may refer to at least one of accuracy, recall, confusion matrix, AUC-ROC, sensitivity, specificity and clinical indicators; among them, accuracy is used to determine the proportion of correct overall predictions; precision, recall and F1 score are used to measure the fine performance of the classifier (especially suitable for unbalanced data); confusion matrix is ​​used to display the prediction status of each category and identify easily confused categories; AUC-ROC is used to evaluate the model's ability to sort positive and negative samples (binary classification or multi-classification); sensitivity is used to reduce false negatives (such as avoiding missed diagnosis of diseases); specificity is used to reduce false positives (such as avoiding misdiagnosis of healthy samples; clinical indicators include positive predictive value (PPV), negative predictive value (NPV), etc. The preset performance indicator threshold may refer to a pre-set threshold corresponding to each performance indicator, which can be flexibly set according to actual application.

[0125] Specifically, an independent test set (not involved in training and verification) is retained from the target sample data set to simulate the real application scenario and evaluate the generalization ability of the model. The model with the best performance in the verification set (such as the model with the highest accuracy or the largest AUC) is selected as the candidate neural network model for vulvar disease classification. The candidate neural network model for vulvar disease classification is loaded, and the candidate neural network model for vulvar disease classification is tested through the test set, and the performance index corresponding to the candidate neural network model for vulvar disease classification is calculated, for example, the accuracy is 98.5%; if the performance index is within the preset performance index threshold range, for example, the accuracy threshold is 98%, then the performance index of the candidate neural network model for vulvar disease classification is within the preset performance index threshold range, and the candidate neural network model for vulvar disease classification is used as the trained neural network model for vulvar disease classification.

[0126] Furthermore, a judgment can be made by combining multiple performance indicators. For example, the accuracy and sensitivity of the candidate neural network model for vulvar disease classification can be tested on a test set. If the accuracy is within the accuracy threshold range and the sensitivity is within the sensitivity threshold range, then the candidate model performance indicator is determined to be within the preset performance indicator threshold range, and the candidate neural network model for vulvar disease classification is used as the trained neural network model for vulvar disease classification. If the accuracy is not within the accuracy threshold range and the sensitivity is within the sensitivity threshold range, or if the accuracy is within the accuracy threshold range and the sensitivity is not within the sensitivity threshold range, or if the accuracy is not within the accuracy threshold range and the sensitivity is not within the sensitivity threshold range, then the neural network model for vulvar disease classification needs to be retrained.

[0127] In this embodiment, the candidate neural network model for vulvar disease classification is tested using a test set, and the performance index corresponding to the candidate neural network model for vulvar disease classification is calculated; if the performance index is within a preset performance index threshold range, the candidate neural network model for vulvar disease classification is used as the trained neural network model for vulvar disease classification; objective performance testing is performed using a test set, and the true generalization ability of the model can be evaluated in an unbiased manner to avoid false high precision caused by overfitting; preset performance index thresholds (for example, accuracy ≥ 95%, AUC ≥ 0.9) can ensure that the model meets clinical usability standards, exclude unqualified candidate models, reduce manual intervention, and accelerate the transformation process from research and development to clinical practice.

[0128] Example 3:

[0129] In actual application scenarios, a corresponding application system can be built based on the neural network model for vulvar disease classification using the Flask framework. This system can implement key functions such as image uploading, calling the neural network model for vulvar disease classification, diagnosis output, and result display. The system's front-end interface is simple and intuitive. Doctors can input images and view the diagnostic report automatically generated by the system. The report content includes:

[0130] (1) Identification results of vulvar disease categories;

[0131] (2) probability of vulvar disease diagnosis;

[0132] (3) Personalized treatment recommendations generated based on expert database.

[0133] For example, if the input vulva image features thinning of the skin, paleness (ivory or porcelain white patches), epidermal atrophy (such as labial atrophy, clitoral hood adhesions), purpura or ecchymosis, the diagnostic results output by the neural network model based on vulvar disease classification may be: diagnostic category: VLS; predicted probability: 98.78%; treatment recommendation:

[0134] 1. Multiple deep biopsies.

[0135] 2. If VLS is pathologically confirmed, lifelong treatment is required.

[0136] Option 1: Glucocorticoid (mometasone):

[0137] Initial: once a day for 4 weeks. Reduction: change to once every 2 days for 4 weeks; change to twice a week for 4 weeks; change to 1-2 times a week for maintenance phase, lifelong use (adjust dosage according to vulvar skin condition).

[0138] Option 2: Tacrolimus ointment, adults (0.1% concentration):

[0139] Initial: once every 2 days for 1-3 months, change to once a day for 1 month, change to once every 2 days in the maintenance phase (adjust according to skin condition).

[0140] For children, tacrolimus ointment (0.03%):

[0141] Initial: 2 times a day for 1-3 months, then gradually reduce to once every 2 days (adjust according to skin condition).

[0142] 3. Follow-up: Review 1-3 months after the initial treatment, and then every 6 months to 1 year.

[0143] In this embodiment, the neural network model based on the classification of vulvar diseases and the corresponding application system built through the Flask framework can be applied to clinical scenarios such as primary medical institutions and gynecological clinics, and can significantly improve the efficiency and standardization of primary screening and auxiliary diagnosis of vulvar diseases.

[0144] Example 4:

[0145] In some embodiments, the neural network model training method for vulvar disease classification can be implemented as a computer program, which is tangibly contained in a computer-readable storage medium, such as a storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed on an electronic device via a ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the processor, one or more steps of the neural network model training method for vulvar disease classification described above can be performed. Alternatively, in other embodiments, the processor can be configured to execute the neural network model training method for vulvar disease classification by any other appropriate means (for example, by means of firmware).

[0146] Various embodiments of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard parts (ASSPs), system on chips (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0147] Computer programs for implementing the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when the computer program is executed by the processor, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The computer program may be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0148] In the context of the present invention, a computer-readable storage medium stores computer instructions that are used to implement the neural network model training method for vulvar disease classification provided by the present invention when the processor executes the instructions. The computer-readable storage medium can be a tangible medium that can contain or store a computer program for use by an instruction execution system, device or apparatus or used in conjunction with an instruction execution system, device or apparatus. The computer-readable storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the above. Alternatively, the computer-readable storage medium can be a machine-readable signal medium. More specific examples of machine-readable storage media would include electrical connections based on one or more wires, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, or any suitable combination of the above.

[0149] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a cathode ray tube (CRT) or a liquid crystal display (LCD monitor)) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0150] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with embodiments of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.

[0151] A computing system may include a client and a server. The client and server are generally remote from each other and typically interact via a communication network. The client-server relationship arises through computer programs running on the respective computers and establishing a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, a host product within a cloud computing service ecosystem that addresses the management difficulties and limited scalability of traditional physical hosts and virtual private server (VPS) services.

[0152] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in the present invention can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solution of the present invention can be achieved. This is not limited herein.

[0153] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A neural network model for vulvar disease classification, characterized in that: include: An image acquisition module, configured to acquire a vulva image and input the vulva image into a Swin Transformer module; The Swin Transformer module is used to segment the vulva image into multiple non-overlapping fixed-size tiles, convert each tile into a tile sequence, perform hierarchical multi-scale feature extraction on the tile sequence through an attention mechanism, obtain a multi-scale hierarchical feature map, predict the vulvar disease category on the feature map, and output the predicted vulvar disease category and corresponding probability.

2. The neural network model for vulvar disease classification according to claim 1, characterized in that: Also includes: Diagnostic suggestion generation is used to generate corresponding diagnostic suggestions based on the expert database and the predicted vulvar disease category and corresponding probability.

3. The neural network model for vulvar disease classification according to claim 1, characterized in that: The SwinTransformer module includes: An image preprocessing layer, configured to segment the vulva image into non-overlapping fixed-size tiles, and convert each of the tiles into a tile sequence by flattening and linear projection; A feature extraction layer, configured to input the tile sequence into a hierarchical network structure consisting of four stages, each of which contains a pair of sliding window Transformer blocks. Each sliding window Transformer block extracts a multi-scale hierarchical feature map through a local window self-attention and shift mechanism; A pooling layer, configured to perform global average pooling on the multi-scale hierarchical feature map to integrate the multi-scale features into a feature vector of fixed length; The classification layer is used to output the predicted vulvar disease category and the corresponding probability based on the feature vector and the Softmax function.

4. The neural network model for vulvar disease classification according to claim 1, characterized in that: The vulvar disease categories include: at least one of vulvar lichen sclerosus, vulvar lichen simplex chronicus, vulvar lichen planus, vulvar vitiligo, vulvar HPV-related precancerous lesions, vulvar HPV-unrelated precancerous lesions, vulvar Paget's disease and vulvar squamous cell carcinoma.

5. A neural network model training method for vulvar disease classification according to any one of claims 1 to 4, characterized in that: include: Obtaining an original sample data set, and preprocessing the original sample data set to obtain a target sample data set; Divide the target sample data set into a training set, a validation set, and a test set according to a set ratio; Based on the training set, validation set and test set, the pre-trained neural network model for vulvar disease classification is trained by cross-validation method to obtain a trained neural network model for vulvar disease classification.

6. The neural network model training method for vulvar disease classification according to claim 5, characterized in that: The preprocessing of the original sample data set to obtain a target sample data set includes: The original sample data set is expanded using the OpenCV library to obtain a target sample data set.

7. The neural network model training method for vulvar disease classification according to claim 5, characterized in that: The pre-trained neural network model for vulvar disease classification is a neural network model for vulvar disease classification with a frozen feature extraction layer; Accordingly, the neural network model for vulvar disease classification is pre-trained, including: Based on the ImageNet dataset, the feature extraction layer in the neural network model for vulvar disease classification is pre-trained, and the parameters of the pre-trained feature extraction layer are frozen to obtain a pre-trained neural network model for vulvar disease classification.

8. The neural network model training method for vulvar disease classification according to claim 5, characterized in that: The method of training the pre-trained neural network model for vulvar disease classification based on the training set, the validation set, and the test set by cross-validation to obtain a trained neural network model for vulvar disease classification includes: Iteratively training the pre-trained neural network model for vulvar disease classification based on the training set; Each time it is iterated, the neural network model for vulvar disease classification after iterative training is verified based on the verification set, and the verification index corresponding to the current iterative model is saved; After all iterations of the training set are completed, the model with the best verification index is selected as the candidate neural network model for vulvar disease classification; The candidate neural network model for vulvar disease classification is tested based on the test set. If the test passes, the candidate neural network model for vulvar disease classification is used as the trained neural network model for vulvar disease classification.

9. The neural network model training method for vulvar disease classification according to claim 8, characterized in that: The step of testing the candidate neural network model for vulvar disease classification based on the test set and using the candidate neural network model for vulvar disease classification as a trained neural network model for vulvar disease classification if the test passes includes: Testing the candidate neural network model for vulvar disease classification based on the test set, and calculating performance indicators corresponding to the candidate neural network model for vulvar disease classification; If the performance indicator is within a preset performance indicator threshold range, the candidate neural network model for vulvar disease classification is used as a trained neural network model for vulvar disease classification.

Images