Cervical tct image detection device based on lesion region relationship modeling

Abstract

The application discloses a kind of based on pathological region relationship modeling cervical TCT image detection device, belong to medical image processing technical field, including: in extracting the multiscale feature of entire cell smear image and extracting the feature of segmented small image block, through the pathological region relationship modeling of these two features, and using attention mechanism to extract key features, using the key features to detect pathological region, improve the detection accuracy, help doctors more accurately and easily find pathological cells and suspected pathological cells, can effectively improve the work efficiency of doctors, provide greater possibility for early treatment of patients.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image processing, and particularly relates to a cervical TCT image detection device based on lesion area relationship modeling. BACKGROUND

[0002] The cervical cancer screening process consists of a three-step diagnostic procedure of "cytology-colposcopy-histology". In clinical practice, after the cell sample is made into a cell smear, the doctor needs to observe the whole slide under a microscope and find the diseased cells and cells that may have potential lesions in a large number of cells, and finally the number of diseased cells and potential diseased cells in the whole slide and the severity of the disease are comprehensively considered to give the final diagnosis. However, under the microscope, the whole cell smear is very large, and the doctor needs to move the field of view constantly to observe the situation of each part of the whole slide, and there are usually hundreds of millions of cells on the whole cell smear. Without the help of computers, the whole manual reading process is very time-consuming and laborious, and many doctors need to work hard to complete a large number of examinations.

[0003] With the rapid development of deep learning technology in the field of computer vision, deep learning technology has also been applied to the field of cervical cancer screening technology as an auxiliary means in the TCT screening reading process to reduce the workload of doctors and further improve the accuracy of diagnosis.

[0004] For the task of assisting doctors in cervical TCT lesion area detection, a rectangular box needs to be used to mark the lesion area and lesion category on the microscopic image corresponding to the cervical TCT cell smear. However, due to the problems of inconsistent magnification when imaging by different microscopes, obvious different image features of single cells and cell clusters on the cell smear, and the need to predict the same lesion category, the existing target detection method and cervical TCT lesion area detection method have problems such as being unable to accurately judge the lesion category of the lesion area and producing redundant nested prediction boxes for cell clusters.

[0005] The reasons for the great difference in image features between single cells and cell clusters include: there can be a large number of cells gathered together on the cell smear, and the doctor cannot label a single rectangular frame for each cell when labeling the image, but can only label the entire cell cluster using a rectangular frame. This labeling method leads to great differences in image features between single lesion cells and a cell cluster under the same lesion category, but the model needs to predict the same lesion category, which will cause the model to encounter contradictions during the learning process and affect the accuracy of the final prediction result. In addition, the detection model uses an image pyramid to improve the detection accuracy of small targets, and the local features of the cell cluster and the features of the single cell are very similar under high magnification, which leads the model to predict many redundant and mutually nested prediction boxes, which greatly reduces the detection accuracy of the model. SUMMARY

[0006] In view of the above, the purpose of the present application is to provide a cervical TCT image lesion area detection method based on lesion area relationship modeling to improve the detection accuracy of the lesion area.

[0007] To achieve the above-mentioned purpose of the application, the present application provides a cervical TCT image detection device based on lesion area relationship modeling, comprising a memory and a processor, the memory is used to store a computer program, and the processor executes the computer program to realize the following steps:

[0008] Obtain the cell smear image and the corresponding lesion area positioning and classification label to form the cell smear training sample data;

[0009] Construct a lesion area detection neural network, including a first feature extraction module, a second feature extraction module, a key feature extraction module, and a detection module, the first feature extraction module is used to extract the multi-scale features of the cell smear image and flatten them to form the first part of the subsequent token sequence, the second feature extraction module is used to extract the features of the corresponding multiple image blocks of the cell smear image and flatten and splice them to form the second part of the subsequent token sequence, the key feature extraction module is used to model the lesion area relationship of the first part of the subsequent token sequence and the second part of the subsequent token sequence, and extract the key features using the attention mechanism, and the detection module is used to detect the lesion area according to the key features;

[0010] Put all the cell smear training sample data into the lesion area detection neural network for training, and continuously optimize it by updating the parameters;

[0011] Use the parameter-optimized lesion area detection neural network to detect the lesion area of the TCT image.

[0012] Preferably, the first feature extraction module comprises a first residual network, a feature pyramid, and a plurality of convolution layer networks; after the cell smear image is extracted by the first residual network, the extracted features are input into the feature pyramid to extract multi-scale features, each scale feature is input into a corresponding convolution layer network to obtain a single token sequence through convolution operation and flattening operation, and all single token sequences output by the convolution layer networks are spliced to form a first part of a subsequent token sequence.

[0013] Preferably, the second feature extraction module comprises a plurality of second branch networks and a small transformer encoder, each second branch network comprising a second residual network, a pooling operation and a flattening operation, each small image block is input into a corresponding second branch network to form a single token through the pooling operation and the flattening operation, and all single tokens corresponding to the second branch networks are spliced and input into the small transformer encoder to obtain a second part of a subsequent token sequence through encoding operation.

[0014] Preferably, the key feature extraction module comprises a transformer encoder, an MLP, and a transformer decoder, the first part of the subsequent token sequence and the second part of the token sequence are spliced and input into the transformer encoder to obtain an encoded feature, the encoded feature is input into the MLP to obtain a mapping feature, the mapping feature is input into the transformer decoder as a Q channel of a multi-head attention mechanism, and the encoded feature is input into the transformer decoder as a K channel and a V channel of the multi-head attention mechanism, the transformer decoder adopts the multi-head attention mechanism to perform decoding operation to output a key feature.

[0015] Preferably, the transformer encoder comprises n encoding blocks connected in sequence, the first encoding block encodes the initial splicing result to obtain an encoded feature, the n th encoding block from the second encoding block encodes an input feature, the n th encoding block comprises the initial splicing feature and a historical encoded feature, the historical encoded feature comprises encoded features output by all encoding blocks before the n th encoding block, and n is a natural number greater than or equal to 2.

[0016] Preferably, the transformer decoder comprises n decoding blocks connected in sequence, each decoding block adopts a multi-head attention mechanism for decoding operation, the K and V channels of each decoding block adopt the encoding features output by the transformer encoder, the Q channel of the first decoding block adopts the mapping features output by the MLP, the Q channel of the n decoding block from the second decoding block adopts the decoding features output by the previous decoding block, the decoding features output by the last decoding block are taken as the key features, and n is a natural number greater than or equal to 2.

[0017] Preferably, the detection module adopts an MLP to perform logical regression on the input key features to locate and classify the lesion area.

[0018] Preferably, during training, the prediction results output by the lesion area detection neural network and the lesion area positioning and classification labels are matched by using the Hungarian algorithm to calculate a positioning loss and a classification loss, wherein the positioning loss is used to measure the error between the lesion area annotation position and the positioning result of the lesion area after detection, and the classification loss is used to measure the error between the lesion area classification label and the classification result of the lesion area after detection.

[0019] To achieve the above-mentioned purposes, the application further provides a cervical TCT image detection device based on lesion area relationship modeling, comprising a data acquisition unit, a model construction unit, a training unit, an application unit,

[0020] The data acquisition unit is used to acquire cell smear images and corresponding lesion area positioning and classification labels to form cell smear training sample data;

[0021] The model construction unit is used to construct a lesion area detection neural network, comprising a first feature extraction module, a second feature extraction module, a key feature extraction module, and a detection module, the first feature extraction module is used to extract multi-scale features of the cell smear image and flatten them to form a first part of the subsequent token sequence, the second feature extraction module is used to extract features of multiple image blocks corresponding to the cell smear image and flatten and splice them to form a second part of the subsequent token sequence, the key feature extraction module is used to model the lesion area relationship of the first part of the subsequent token sequence and the second part of the subsequent token sequence, and extract key features by using an attention mechanism, and the detection module is used to detect the lesion area according to the key features;

[0022] The training unit is used to input all the cell smear training sample data into the lesion area detection neural network for training, and continuously optimize by updating parameters;

[0023] The application unit is used to detect the lesion area of the TCT image by using the parameter-optimized lesion area detection neural network.

[0024] Compared with the prior art, the technical effects of the present application at least include:

[0025] On the basis of extracting multi-scale features of the entire cell smear image and extracting features of the segmented small image blocks, by modeling the lesion region relationship of the two parts of features, and using the attention mechanism to extract key features, the key features are used for lesion region detection, which improves the detection accuracy, helps doctors to more accurately and easily find lesion cells and suspected lesion cells, and can effectively improve the work efficiency of doctors, and provides greater possibility for early treatment of patients. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0027] Figure 1 is a flowchart of the cervical TCT image detection method based on lesion region relationship modeling provided by the embodiments of the present application;

[0028] Figure 2 is a structural diagram of the lesion region detection neural network provided by the embodiments of the present application;

[0029] Figure 3 is a structural diagram of the transformer encoder provided by the embodiments of the present application;

[0030] Figure 4 is a structural diagram of the transformer decoder provided by the embodiments of the present application;

[0031] Figure 5 is a structural diagram of the cervical TCT image detection device based on lesion region relationship modeling provided by the embodiments of the present application. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the present application more clear and explicit, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and do not limit the protection scope of the present application.

[0033] In order to solve the problem that the magnification of cell smear detection is inconsistent in the prior art, and the image features of single cells and cell clusters are greatly different but need to be predicted into the same lesion category, which further leads to inaccurate detection of the lesion region of cervical TCT image, the embodiments provide a cervical TCT image detection method and device based on lesion region relationship modeling.

[0034] As Figure 1 shown, the embodiment provides a cervical TCT image lesion area detection method based on lesion area relationship modeling, comprising the following steps:

[0035] S110, acquiring a cell smear image and corresponding lesion area positioning and classification label to form a cell smear training sample data.

[0036] In the embodiment, cervical TCT cell smears are prepared through steps such as collection, rinsing, mixing, filtering, adsorption and settlement, staining, and fixing, and the images of the cell smears are photographed by a microscope, and the obtained cell smear images include images obtained by different magnifications and different staining agents.

[0037] The entire cell smear image is cut into images of a fixed window size by using a sliding window algorithm; a plurality of windows with typical lesion cells are selected from a plurality of cell smear images for labeling; the lesion cells and suspected lesion cells on the image are labeled by using a rectangular frame for each independent window image, and the lesion area labeling position and the lesion area classification label are used, wherein the label is divided into five categories, which are ASCUS, LSIL, ASCH, HSIL, and SQCA, corresponding to suspected low-grade lesions, low-grade lesions, suspected high-grade lesions, high-grade lesions, and squamous cell carcinoma.

[0038] In the embodiment, all labeled window images and corresponding lesion area labeling positions and lesion area classification labels are integrated to obtain a data set, and the mean and variance information of the images are counted, and the five categories are divided into the same training set and validation set in a ratio of 4:1. Before division, the image data needs to be randomly arranged to prevent similar image data from being aggregated in a certain set. The image data in the training set is first normalized and then deformed to a size of 1792*896. For cervical TCT cell smear image data, only the most basic data augmentation (normalization and flipping) is needed to achieve good training results. In addition, it should be noted that when the image is flipped, the labeling frame also needs to be flipped.

[0039] S120, constructing a lesion area detection neural network, including a first feature extraction module, a second feature extraction module, a key feature extraction module, and a detection module.

[0040] Figure 2 The structure diagram of the lesion area detection neural network provided by the embodiment is shown in Figure 2 .

[0041] Firstly, the first feature extraction module is used to extract the multi-scale features of the cell smear image and flatten to form the first part of the subsequent token sequence, preferably including a residual network Resnet-152, a feature pyramid FPN, and a plurality of convolutional layer networks. The three-channel RGB image data of the 1792*896 size cell smear sample is input into the first feature extraction module, the multi-scale image features are extracted by Resnet-152, and the multi-scale features are further extracted by inputting the feature pyramid FPN. The FPN features of different scales are respectively aligned in channel number by using a 1*1 convolutional layer, the two-dimensional structure of the image is flattened to a one-dimensional sequence, and all the features are spliced as the first part of the subsequent transformer token sequence input.

[0042] Secondly, the second feature extraction module is used to extract the features of the corresponding multiple image blocks of the cell smear image and flatten and splice to form the second part of the subsequent token sequence, preferably including: a plurality of second branch networks for extracting features of small image blocks after cell smear segmentation, a small transformer encoder, each second branch network containing a second residual network, a pooling operation and a flattening operation. Specifically, the 1792*896 size three-channel RGB image is divided into 7*7 size patches, each patch is sent into a small convolutional network ResNet-50 to extract features, and the features of each patch are converted into a token by using a pooling layer and a flattening operation. All tokens are spliced into a sequence and input into a small transformer encoder. The output result of the small transformer encoder is the second part of the subsequent token sequence.

[0043] Thirdly, the key feature extraction module is used to model the relationship between the first part and the second part of the subsequent token sequence, and extract key features by using an attention mechanism, preferably including a transformer encoder, an MLP, and a transformer decoder. Specifically, the first and second parts of the above token sequence are spliced and input into the key feature extraction module. The obtained encoded features are input into the MLP for mapping, and the obtained mapping features are input into the transformer decoder as the Q channel of the multi-head attention mechanism. The encoded features are simultaneously input into the transformer decoder as the K and V channels of the multi-head attention mechanism. The transformer decoder adopts the multi-head attention mechanism for decoding operation, and outputs the key features.

[0044] Among them, as Figure 3As shown, the transformer encoder includes n sequentially connected encoding blocks, the first encoding block encodes the initial splicing result to obtain the encoded feature, the n th encoding block from the second encoding block encodes the input features of all previous encoding blocks, the input features of all previous encoding blocks include the initial splicing feature and the historical encoded feature, the historical encoded feature includes the encoded features output by all previous encoding blocks of the n th encoding block, and n is a natural number greater than or equal to 2.

[0045] As shown in Figure 4 The transformer decoder includes n sequentially connected decoding blocks, each decoding block adopts a multi-head attention mechanism for decoding operation, the K and V channels of each decoding block adopt the encoded features output by the transformer encoder, the Q channel of the first decoding block adopts the mapping features output by the MLP, the Q channel of the n th decoding block from the second decoding block adopts the decoding features output by the previous decoding block, and the decoding features output by the last decoding block are used as the key features, and n is a natural number greater than or equal to 2.

[0046] Finally, the detection module is used to detect the lesion area according to the key features, and preferably, an MLP is adopted. Specifically, all the key features output by the last layer of the transformer decoder are sent into the MLP for lesion area detection.

[0047] S130, all cell smear training sample data are sent into the lesion area detection neural network for training, and the parameters are updated to continuously optimize to the best state.

[0048] During the training process, after the image of the training sample data is extracted into the token sequence by the ResNet-152 and the feature pyramid and the ResNet-50 and the small transformer encoder two branch networks, the token sequence output by the transformer decoder is sent into the transformer encoder and the decoder to model the relationship between the lesion area, the token sequence output by the transformer decoder is sent into the MLP to predict the final prediction result. The prediction result and the lesion area positioning and classification label use the Hungarian algorithm to calculate the bilateral matching relationship, so that the lesion area positioning and classification label at most matches one detection result. The matched detection result and the lesion area positioning and classification label calculate the L1 positioning loss and the BCE classification loss. Specifically, after the loss calculation is completed, the network parameters are updated, and the validation set is also used for verification in the iteration process. After the iteration is completed, the network with the best verification effect is saved.

[0049] S140, using the parameter-optimized lesion area detection neural network to detect the lesion area of the TCT image.

[0050] In the embodiment, the lesion area detection neural network optimized by parameters is used for TCT image lesion area detection, including: inputting the cervical TCT lesion cell image data to be predicted into the lesion area neural network for detection, outputting the lesion area positioning position and the corresponding classification probability value, and screening the lesion area positioning position with the classification probability value greater than 0.5 from the classification probability value, and the category with the maximum classification probability value is the category of the lesion area positioning position.

[0051] For the problem of inconsistent magnification, in the embodiment, a multi-scale and multi-resolution image feature pyramid is constructed by ResNet-152 and FPN and is flattened into a token sequence and sent into the subsequent transformer, the semantic relationship between the image feature maps at multiple scales is established in the transformer, which can promote the network to better capture the key features of the image at different scales, and can also use the multi-head attention mechanism in the transformer which is not sensitive to the token order to make the network have stronger robustness to the magnification of the input image.

[0052] For the problem that single cells and cell clusters have great differences in image features but need to be predicted into the same lesion category, in the embodiment, a method of extracting features by ResNet-50 and a small transformer encoder after image division into patches is designed. This method does not have the traditional image feature pyramid structure, and will not produce redundant and nested prediction boxes because the features of single cells and local features of cell clusters from different pyramid levels are very similar. In addition, the features extracted by the two methods are spliced into the same token sequence and sent into the transformer for prediction. Through this result, the transformer can exert its powerful ability to model the relationship between features, model the relationship between the captured lesion areas, extract key features, and find accurate lesion areas, thereby improving the detection performance of the network.

[0053] Based on the same inventive concept, the embodiment also provides a cervical TCT image detection device based on lesion area relationship modeling, including a memory and a processor, the memory is used to store a computer program, and the processor implements the steps of the cervical TCT image detection method based on lesion area relationship modeling provided by the above-mentioned embodiment when executing the computer program, including the following steps:

[0054] S110, acquiring a cell smear image and corresponding lesion area positioning and classification label to form a cell smear training sample data;

[0055] S120, constructing a lesion area detection neural network, including a first feature extraction module, a second feature extraction module, a key feature extraction module, and a detection module;

[0056] S130, all cell smear training sample data is sent into the lesion area detection neural network for training, and is constantly optimized to the best state by updating parameters;

[0057] S140, the lesion area detection neural network with optimized parameters is used for TCT image lesion area detection.

[0058] In the embodiment, the memory can be a volatile memory such as RAM at a near end, and can also be a non-volatile memory such as ROM, FLASH, a floppy disk, a mechanical hard disk, and the like, and can also be a storage cloud at a far end. The processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), that is, the steps of the cervical TCT image detection method based on lesion area relationship modeling can be implemented through these processors.

[0059] Based on the same inventive concept, the embodiment also provides a cervical TCT image detection device 500 based on lesion area relationship modeling, comprising a data acquisition unit 510, a model construction unit 520, a training unit 530, an application unit 540,

[0060] The data acquisition unit 510 is configured to acquire cell smear images and corresponding lesion area positioning and classification labels, and form cell smear training sample data.

[0061] The model construction unit 520 is configured to construct a lesion area detection neural network, comprising a first feature extraction module, a second feature extraction module, a key feature extraction module, and a detection module. The first feature extraction module is configured to extract multi-scale features of the cell smear images and flatten to form a first part of a subsequent token sequence. The second feature extraction module is configured to extract features of a plurality of image blocks corresponding to the cell smear images and flatten and splice to form a second part of the subsequent token sequence. The key feature extraction module is configured to model the lesion area relationship of the first part of the subsequent token sequence and the second part of the subsequent token sequence, and extract key features by using an attention mechanism. The detection module is configured to detect the lesion area according to the key features.

[0062] The training unit 530 is configured to send all cell smear training sample data into the lesion area detection neural network for training, and constantly optimize by updating parameters.

[0063] The application unit 540 is configured to use the lesion area detection neural network with optimized parameters to detect the lesion area of the TCT image.

[0064] It should be noted that the cervical TCT image detection device provided by the above embodiment should be illustrated by the division of the above functional units when performing cervical TCT image detection, and the above functions can be completed by different functional units according to needs, that is, the internal structure of the terminal or server is divided into different functional units to complete all or part of the above described functions. In addition, the cervical TCT image detection device and the cervical TCT image detection generation method provided by the above embodiment belong to the same concept, and the specific implementation process is described in the cervical TCT image detection method embodiment, which will not be repeated here.

[0065] The specific embodiments described above describe the technical solutions and advantages of the present application in detail. It should be understood that the above description is only the most preferred embodiment of the present application and is not intended to limit the present application. Any modification, supplement and equivalent replacement within the principle range of the present application should be included in the protection scope of the present application.

Claims

1. A cervical TCT image detection device based on lesion area relationship modeling, comprising a memory and a processor, the memory is used to store a computer program, characterized in that, The processor implements the following steps when executing the computer program: Obtaining a cell smear image and a corresponding lesion area positioning and classification label to form a cell smear training sample data; A lesion area detection neural network is constructed, including a first feature extraction module, a second feature extraction module, a key feature extraction module, and a detection module. The first feature extraction module is used to extract multi-scale features of the cell smear image and flatten them to form a first part of the subsequent token sequence. The second feature extraction module is used to extract features of multiple image blocks corresponding to the cell smear image and flatten and splice them to form a second part of the subsequent token sequence. The key feature extraction module is used to model the relationship between the first part of the subsequent token sequence and the second part of the subsequent token sequence, and extract key features using an attention mechanism. The detection module is used to detect the lesion area according to the key features. The key feature extraction module includes a transformer encoder, an MLP, and a transformer decoder. The first part of the subsequent token sequence and the second part of the token sequence are spliced and input into the transformer encoder to obtain encoded features. The encoded features are input into the MLP for mapping, and the mapped features are input into the transformer decoder as the Q channel of the multi-head attention mechanism. The encoded features are also input into the transformer decoder as the K and V channels of the multi-head attention mechanism. The transformer decoder performs decoding operation using the multi-head attention mechanism to output the key features. All cell smear training sample data is input into the lesion area detection neural network for training, and the parameters are updated to continuously optimize. The lesion area detection neural network with optimized parameters is used for TCT image lesion area detection.

2. The cervical TCT image detection device based on lesion area relationship modeling according to claim 1, wherein, The first feature extraction module includes a first residual network, a feature pyramid, and multiple convolution layer networks. After the cell smear image is processed by the first residual network to extract features, the extracted features are input into the feature pyramid to extract multi-scale features. Each scale feature is input into the corresponding convolution layer network to obtain a single token sequence through convolution operation and flattening operation. All single token sequences output by the convolution layer networks are spliced to form the first part of the subsequent token sequence.

3. The cervical TCT image detection device based on lesion area relationship modeling of claim 1, wherein, The second feature extraction module includes multiple second branch networks for extracting features of small image blocks after cell smear segmentation and a small transformer encoder. Each second branch network includes a second residual network, a pooling operation, and a flattening operation. Each small image block is input into the corresponding second branch network to form a single token through the pooling operation and the flattening operation. All single tokens corresponding to the second branch networks are spliced and input into the small transformer encoder to obtain the second part of the subsequent token sequence through encoding operation.

4. The cervical TCT image detection device based on lesion area relationship modeling of claim 1, wherein, The transformer encoder comprises n sequentially connected encoding blocks, the first encoding block encodes the initial spliced result to obtain an encoded feature, the n th encoding block from the second encoding block encodes the input feature, the input feature of the n th encoding block comprises the initial spliced feature and a historical encoded feature, the historical encoded feature comprises the encoded features output by all the encoding blocks before the n th encoding block, and n is a natural number greater than or equal to 2.

5. The cervical TCT image detection device based on lesion area relationship modeling of claim 1, wherein, The transformer decoder comprises n sequentially connected decoding blocks, each decoding block adopts a multi-head attention mechanism for decoding operation, the K and V channels of each decoding block adopt the encoded features output by the transformer encoder, the Q channel of the first decoding block adopts the mapping feature output by the MLP, the Q channel of the n th decoding block from the second decoding block adopts the decoding feature output by the previous decoding block, and the decoding feature output by the last decoding block is taken as the key feature, and n is a natural number greater than or equal to 2.

6. The cervical TCT image detection device based on lesion area relationship modeling of claim 1, wherein, The detection module adopts an MLP to perform logical regression on the input key feature to locate and classify the lesion area.

7. The cervical TCT image detection device based on lesion area relationship modeling of claim 1, wherein, During training, the prediction results output by the lesion area detection neural network and the lesion area positioning and classification labels are bidirectionally matched using the Hungarian algorithm, so as to calculate a positioning loss and a classification loss, the positioning loss is used to measure the error between the lesion area annotation position and the lesion area positioning result after detection, and the classification loss is used to measure the error between the lesion area classification label and the lesion area classification result after detection.

8. A device for detecting cervical TCT images based on lesion area relationship modeling, characterized in that, The method comprises the steps of acquiring a cell smear image and corresponding lesion area positioning and classification labels to form a cell smear training sample data, constructing a lesion area detection neural network, training the lesion area detection neural network, and applying the trained lesion area detection neural network. The model construction unit is configured to construct a lesion area detection neural network, which comprises a first feature extraction module, a second feature extraction module, a key feature extraction module, and a detection module. The first feature extraction module is configured to extract multi-scale features of the cell smear image and flatten them to form a first part of a subsequent token sequence. The second feature extraction module is configured to extract features of a plurality of image blocks corresponding to the cell smear image and flatten and splice them to form a second part of the subsequent token sequence. The key feature extraction module is configured to model the lesion area relationship of the first part of the subsequent token sequence and the second part of the subsequent token sequence, and extract a key feature using an attention mechanism. The detection module is configured to detect a lesion area according to the key feature. The key feature extraction module comprises a transformer encoder, an MLP, a transformer decoder, and the first part and the second part of the token sequence are spliced and input into the transformer encoder, the obtained coding features are input into the MLP for mapping, the obtained mapping features are input into the transformer decoder as the Q channel of the multi-head attention mechanism, the coding features are input into the transformer decoder as the K and V channels of the multi-head attention mechanism, the transformer decoder adopts the multi-head attention mechanism for decoding operation, and the key features are output. The training unit is configured to input all cell smear training sample data into the lesion area detection neural network for training, and continuously optimize by updating parameters. The application unit is configured to use the parameter-optimized lesion area detection neural network to detect the lesion area of the TCT image.

Images