Target detection method and system for cervical cell image

Abstract

The invention provides a target detection method and system for a cervical cell image, and relates to the technical field of target detection, and the specific scheme comprises the steps: obtaining a to-be-detected cervical cell image; inputting the cervical cell image into a trained target detection model to obtain a candidate target area of the cervical cells; wherein the target detection model is used for fusing low-level feature information and high-level information based on enhanced semantic information and position information by adopting a multi-level connected feature pyramid network on the basis of YOLOv5, and generating candidate target areas through fused features; according to the method, the extraction capability of cell characteristics and the perception capability of the model are improved, so that better detection performance is obtained.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and in particular, relates to a target detection method and system for cervical cell images. Background Art

[0002] Cervical cancer is a common female malignancy, with an increasing incidence and a younger age trend. Medical research has shown that the earlier cervical cancer is detected, the higher the cure rate. The traditional cervical cell screening method is for professional physicians to manually analyze and screen the Pap smear of cervical cells. The Pap staining method is a cell staining method used to examine cervical precancerous lesions and cervical cancer. The collected cervical cells are fixed and stained on a glass slide with Pap staining fluid, which can make the nucleus and cytoplasm of cervical cells appear in different colors. The traditional manual screening method has high requirements on the professional ability of the testers, and it also takes a lot of time and effort. When conducting large-scale cervical cancer screening, the efficiency may not meet the requirements, and it is easy to cause misdiagnosis and missed diagnosis.

[0003] With the continuous development of artificial intelligence-assisted diagnosis technology, the use of intelligent technology to assist in the diagnosis process has brought great benefits to the screening and prevention of cervical cancer; and the target detection task is an important branch in the field of medical image research. It mainly analyzes the nuclear-cytoplasmic ratio and the morphological changes of cells, extracts cell features, and serves as the basis for judging cell categories, which lays an important theoretical foundation for subsequent detection, analysis and diagnosis; At present, the target detection method for cervical cells has been applied, but there are still some limitations. There may be cell clusters in the cervical cell image after Papanicolaou staining, which increases the difficulty of the cell detection task; At the same time, abnormal cells have different states at different times, which also increases the complexity of the cell detection task.

[0004] Therefore, the existing cervical cell image target detection method has insufficient cell feature extraction capability and cannot obtain better detection performance. Summary of the invention

[0005] In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides a target detection method and system for cervical cell images, which improve the ability to extract cell features and the perception ability of the model, thereby obtaining better detection performance.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0007] A first aspect of the present invention provides a method for object detection in a cervical cell image.

[0008] A method for detecting a target in a cervical cell image, comprising:

[0009] Acquire the cervical cell image to be detected;

[0010] Input the cervical cell image into the trained target detection model to obtain the candidate target area of ​​cervical cells;

[0011] Among them, the target detection model is based on YOLOv5, adopts a multi-level connected feature pyramid network, fuses low-level feature information and high-level information based on enhanced semantic information and position information, and generates candidate target areas through the fused features.

[0012] Furthermore, the target detection model includes a feature extraction network, a feature fusion network and a detection network;

[0013] The feature extraction network uses an improved CSPDarknet53 to extract feature maps of different sizes and levels from cervical cell images based on enhanced semantic information and position information;

[0014] The feature fusion network adopts a multi-level connected feature pyramid network to fuse low-level feature information with high-level information based on enhanced semantic information and position information to obtain fused features;

[0015] The detection network uses an improved detection head to generate bounding box coordinates, category probabilities and target confidences of candidate target regions.

[0016] Furthermore, the improved CSPDarknet53 adds compression, dimensionality reduction and fusion of channel features and spatial features to CSPDarknet53 to obtain enhanced semantic information and location information.

[0017] Furthermore, the multi-level connected feature pyramid network fuses features of multiple levels on feature maps of different sizes and levels based on one convolution and one information filtering operation.

[0018] Furthermore, the improved detection heads are increased from the original three target detection heads to four target detection heads, thereby alleviating the neglect of small cell information.

[0019] Furthermore, the target detection model constructs an objective function using classification loss, confidence loss, and positioning loss, and is trained using the constructed training data set.

[0020] Furthermore, the construction of the training data set is specifically as follows:

[0021] Get the cervical cell images and the corresponding JSON format coordinate files to form the original data set;

[0022] The cervical cell images in the dataset and the corresponding JSON format coordinate files are preprocessed to obtain the final training dataset.

[0023] A second aspect of the present invention provides a target detection system for cervical cell images.

[0024] A target detection system for cervical cell images, including an acquisition module and a detection module:

[0025] The acquisition module is configured to: acquire the cervical cell image to be detected;

[0026] The detection module is configured to: input the cervical cell image into the trained target detection model to obtain a candidate target area of ​​the cervical cell;

[0027] Among them, the target detection model is based on YOLOv5, adopts a multi-level connected feature pyramid network, fuses low-level feature information and high-level information based on enhanced semantic information and position information, and generates candidate target areas through the fused features.

[0028] The third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the target detection method for cervical cell images as described in the first aspect of the present invention.

[0029] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the target detection method for cervical cell images as described in the first aspect of the present invention are implemented.

[0030] One or more of the above technical solutions have the following beneficial effects:

[0031] The present invention adopts a multi-level connection method in the feature fusion network of YOLOv5 to fuse low-level feature information and high-level information, complement each other, improve the utilization rate of information, enhance the learning ability of the network, and reduce the information loss in the process of cervical cell image feature extraction to obtain better detection performance.

[0032] The present invention adds compression, dimensionality reduction and fusion of channel features and spatial features to the feature extraction network of cervical cell images, obtains enhanced semantic information and position information, highlights important feature information, improves the perception ability of the model, and further improves the model performance.

[0033] The present invention alleviates the neglect of small cell information due to the increase of network depth by adding a prediction head to the high-resolution feature layer part on the basis of the three prediction heads of the original YOLOv5 model.

[0034] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0036] Figure 1 This is a flow chart of the method of the first embodiment.

[0037] Figure 2 This is a simplified structural diagram of the target detection model of the first embodiment.

[0038] Figure 3 This is a detailed structural diagram of the target detection model of the first embodiment.

[0039] FIG. 4( a ) is a structural diagram of a CBAM module of the first embodiment.

[0040] FIG4( b ) is a structural diagram of the BRA module of the first embodiment.

[0041] FIG. 4( c ) is a structural diagram of a CSP module of the first embodiment.

[0042] FIG. 4( d ) is a structural diagram of the C3 module of the first embodiment.

[0043] FIG4( e ) is a structural diagram of the SPPF module of the first embodiment.

[0044] Figure 5 This is an example diagram of the detection results of the first embodiment. DETAILED DESCRIPTION

[0045] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0046] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0047] Embodiment 1

[0048] In one embodiment of the present disclosure, a method for detecting a target in a cervical cell image is provided, comprising the following steps:

[0049] Step S1: obtaining a cervical cell image to be detected;

[0050] Step S2: input the cervical cell image into the trained target detection model to obtain the candidate target area of ​​cervical cells;

[0051] Among them, the target detection model is based on YOLOv5, adopts a multi-level connected feature pyramid network, fuses low-level feature information and high-level information based on enhanced semantic information and position information, and generates candidate target areas through the fused features.

[0052] Furthermore, the target detection model includes a feature extraction network, a feature fusion network and a detection network;

[0053] The feature extraction network uses an improved CSPDarknet53 to extract feature maps of different sizes and levels from cervical cell images based on enhanced semantic information and position information;

[0054] The feature fusion network adopts a multi-level connected feature pyramid network to fuse low-level feature information with high-level information based on enhanced semantic information and position information to obtain fused features;

[0055] The detection network uses an improved detection head to generate bounding box coordinates, category probabilities and target confidences of candidate target regions.

[0056] Furthermore, the improved CSPDarknet53 adds compression, dimensionality reduction and fusion of channel features and spatial features to CSPDarknet53 to obtain enhanced semantic information and location information.

[0057] Furthermore, the multi-level connected feature pyramid network fuses features of multiple levels on feature maps of different sizes and levels based on one convolution and one information filtering operation.

[0058] Furthermore, the improved detection heads are increased from the original three target detection heads to four target detection heads, thereby alleviating the neglect of small cell information.

[0059] Furthermore, the target detection model constructs an objective function using classification loss, confidence loss, and positioning loss, and is trained using the constructed training data set.

[0060] Furthermore, the construction of the training data set is specifically as follows:

[0061] Get the cervical cell images and the corresponding JSON format coordinate files to form the original data set;

[0062] The cervical cell images in the dataset and the corresponding JSON format coordinate files are preprocessed to obtain the final training dataset.

[0063] As a classic algorithm for single-stage target detection tasks, YOLO has been continuously developed in recent years and has been proven to be useful for medical target detection. It can be seen that the YOLO series shows great potential in improving medical diagnosis. Therefore, this embodiment improves the existing YOLOv5 to obtain a target detection model. The improvement is summarized as follows: based on the YOLOv5 algorithm, in the feature extraction network part, compression, dimensionality reduction and fusion of channel features and spatial features are added to obtain enhanced semantic information and location information, and the ability to extract cell features is improved; in the feature fusion network part, a multi-level connected feature pyramid network is reconstructed to fuse the extracted multi-scale features to obtain better detection performance; in the detection network part, a prediction head for high-resolution feature maps is added to predict small target cells.

[0064] Based on the above improvements, the process of the target detection method of this embodiment is summarized as follows: first, the feature extraction network (CSPDarknet53) is used to extract feature maps of different sizes and levels in the cervical cell image through a series of convolution and pooling operations; secondly, the feature information of different levels is fully utilized to fuse the feature maps of different sizes and levels with multi-level connections; the feature maps after feature fusion are predicted and output to generate candidate target areas of cervical cells. The detection results are shown in Figure 4.

[0065] The following describes in detail the implementation process of the target detection method for cervical cell images in this embodiment from the perspective of the construction and training process of the target detection model. Figure 1 As shown, the specific steps are:

[0066] Step 1: Collect cervical cell images and preprocess them. The preprocessed cervical cell images form a training data set and are divided into a test set, a validation set, and a training set.

[0067] In this embodiment, the SIPaKMeD dataset is selected as the original dataset, the JSON format coordinate file of the original dataset is processed to generate a TXT format coordinate file, and the dataset is divided into a training set, a validation set, and a test set.

[0068] The cervical cell images are preprocessed by mosaic, flipping, rotation, cropping, contrast and other methods to compensate for the imbalance of sample data.

[0069] Step 2: Build an object detection model

[0070] Figure 2 It is a simple structural diagram of the target detection model. Figure 3 is a detailed structural diagram of the target detection model, the target detection model, such as Figure 2 , Figure 3 As shown, it includes feature extraction network, feature fusion network and detection network:

[0071] The feature extraction network is a feature extraction network based on enhanced semantic information and position information, which extracts feature maps of different sizes and levels;

[0072] The feature fusion network is a feature pyramid network based on multi-level connections;

[0073] The detection network is a detection network of four detection heads, which is used to generate bounding box coordinates, category probabilities and target confidences of candidate target areas.

[0074] In this embodiment, the feature extraction network adopts the improved CSPDarknet53, the feature fusion network adopts the improved feature pyramid network, and the detection network adopts the improved YOLOv5 prediction output network, which is described in detail below:

[0075] 1. Improved CSPDarknet53

[0076] The existing feature extraction network has five groups of convolutional layers. The input size of the first group of convolutions is 640×640, and the output size is reduced by 32 times, with a reduction of 2 times each time. The final output feature map size is 20×20. After five reductions, five feature layers from C1 to C5 are formed.

[0077] As the network depth increases, although deep optimization helps improve feature extraction capabilities, it easily leads to excessive attention to high-level semantic information, thereby neglecting the effective use of low-level spatial location information. To compensate for this deficiency, the integration of channel dimension and spatial dimension information is added after the C5 layer of the feature extraction network to better utilize global and local features. Through this improvement, the network's attention to key areas in target detection is enhanced, and the accuracy of target content and detection location in cervical cell images is improved.

[0078] CSPDarknet53 is an efficient feature extraction network based on the Darknet53 structure and introduces the CSP module to optimize performance. As shown in Figure 4(c), the CSP module shows that the input feature map is divided into two paths: one that passes through two CBL (convolution, BN, Leaky ReLU) and one convolution, and the other that passes through convolution directly. The outputs of the two paths are combined through the Concat operation, and then processed by BN, Leaky ReLU and CBL to finally output the result.

[0079] The CSP module reduces the amount of computation and improves network performance by branching and fusing features. As shown in Figure 4(d), the schematic diagram of the C3 module in the feature extraction network is shown. The C3 module adopts a dual-branch design. The input feature map passes through two convolution blocks, one of which extracts deeper features through multiple Bottleneck modules, and the other branch maintains a relatively simple convolution process; finally, the two branches are merged through the `Concat` operation and further fuse features through convolution; through this structural design, the C3 module enhances the network's feature extraction capability while maintaining good computational efficiency.

[0080] As shown in Figure 4(e), the SPPF module in the feature extraction network extracts feature information under different receptive fields through multiple 2D maximum pooling operations. The purpose is to reduce the computational complexity while maintaining a good feature extraction effect. After the SPPF module, the input feature map is first subjected to a convolution process (CBL), and then three 2D maximum pooling operations are performed in sequence. The feature map after each pooling is concatenated with the original input feature map (Concat); finally, the concatenated result is further processed through another CBL layer. This module realizes multi-scale feature extraction through hierarchical pooling operations, which enhances the model's perception of objects of different scales.

[0081] The improved CSPDarknet53 adds the CBAM attention mechanism in the output stage of feature extraction, increases the compression, dimensionality reduction and fusion of channel features and spatial features, obtains enhanced semantic information and position information, and improves the ability to extract cell features; Figure 4(a) is a schematic diagram of channel spatial attention (CBAM), which is a dual attention mechanism that combines channel and space. Its working principle is to first pass the input feature map through the channel attention module to focus on the importance of different feature channels; then, pass it through the spatial attention module to capture the spatial position information in the feature map; finally, after two weighted operations, a feature map containing richer contextual information is output.

[0082] 2. Improved Feature Pyramid Network

[0083] The feature pyramid network consists of FPN (Feature Pyramid Network) and PAN (Path Aggregation Network). FPN transmits strong semantic features from top to bottom, and PAN transmits strong positioning features from bottom to top. The combination of the two performs multi-scale information fusion on features at different levels.

[0084] In order to improve the utilization rate of the underlying feature information, the improved FPN performs multi-level feature connections to achieve deeper feature fusion:

[0085] (1) C5 layer: First, the C5 feature layer is convolved to reduce the number of channels to 512, generating C5' while keeping the spatial resolution unchanged. Then, the C4 feature layer is convolved to generate C4' with the same spatial resolution as C5, but C4' introduces the BRA attention mechanism to filter information and effectively remove redundant information.

[0086] (2) Fusion of C4 and C5: After convolution and filtering, C4' and C5' are concat-operated in the channel dimension to form the P4 feature map. P4 is upsampled again to generate a higher resolution (40×40) P4', at which point high-level features are passed downward.

[0087] (3) The C3 layer participates in multi-level connections: The C3 layer is convolved at the same time to generate C3' with the same resolution as P4'. Here, the multi-level connection of the feature pyramid network is specifically manifested in the concat operation of P4', C3', and C4 to generate a new P3 feature map, realizing the effective transfer of high-level features to low-level features.

[0088] (4) Generation of P2: Upsample P3 to generate higher resolution P3', and perform convolution and information filtering on the C2 layer to generate C2'. Then, P3', C2', and C3 are concat-operated to form the P2 layer, which further enhances the multi-level connection and fusion of feature information.

[0089] (5) Final formation of P1: Upsample P2 to generate high-resolution P2', and concat it with C2 to generate the P1 feature map. Here, multi-level feature fusion is used to ensure that more abundant feature information is obtained at different scales, ensuring that the final detection output is more accurate.

[0090] In the second half of feature fusion, the improved PAN network enhances the further connection of multi-level features, as described below:

[0091] (1) D1 feature layer: First, P1 is processed to generate a 160×160 D1 feature map, and D1 is used for the first prediction. D1 is then convolved to reduce its resolution to 80×80 and concat-operated with the upper layer P2 feature map to generate D2, which further enhances the fusion of low-level and high-level features and increases the capture of details.

[0092] (2) D2 feature layer: Similarly, after convolution, the resolution of D2 is reduced to 40×40 and concat-operated with the P3 feature map to generate D3, which adds more high-level semantic information and improves the detection capability of small objects.

[0093] (3) Generation of D3 and D4: Finally, D3 is reduced to 20×20 by convolution and concatenated with P4 to generate the final D4 feature map, completing the fourth prediction. At this time, the high-level features are fully combined with the lower-level features, enhancing the fusion and detection effect of multi-scale features.

[0094] In the multi-level connection, the BRA attention mechanism is introduced to solve the problem of information redundancy during feature fusion. This mechanism enhances the fusion of global and local information and effectively improves the model's ability to extract features from complex cervical cell images.

[0095] 3. Improved YOLOv5 prediction output network

[0096] In order to further enhance the detection of small target cells, the original YOLOv5 prediction output network was modified from the original three target detection heads to four target detection heads (Head), and the feature fusion of different resolutions was further refined to improve the detection effect of the target. The main function of the detection part is to generate a detection anchor frame and output the target's category probability, confidence, and position information of the bounding box; by clustering the input training coordinates, an anchor frame that can better cover the real cervical cells is generated, thereby improving the detection accuracy and positioning effect of small targets. .

[0097] Calculate the classification loss and use the Softmax function to determine the category of cervical cells; calculate the confidence loss and use the binary cross entropy loss function to calculate the confidence of the target of each bounding box and whether the prediction contains the target; calculate the positioning loss and use SIoU to determine the degree of overlap between the bounding box and the true bounding box, and make further adjustments to the subsequent suggested boxes; output the classification and positioning information of the predicted target after weighted calculation and non-maximum suppression of the three types of loss functions.

[0098] Step 3: The target detection model is trained using the cervical cell images in the training set, and the cervical cell images in the validation set are used for validation during the training process to correct the training process, thereby obtaining a trained target detection model;

[0099] Step 4: Input the cervical cell images in the test set into the trained target detection model and output the detection results.

[0100] The experimental environment and parameter settings of this embodiment are as follows: the network uses a 3090 GPU, a batch size of 8, a training model with 70 training rounds, a stochastic gradient descent algorithm with a momentum factor of 0.937, and an initial learning rate of 0.001. Figure 5 The test results are shown.

[0101] Embodiment 2

[0102] In one embodiment of the present disclosure, a target detection system for cervical cell images is provided, including an acquisition module and a detection module:

[0103] The acquisition module is configured to: acquire the cervical cell image to be detected;

[0104] The detection module is configured to: input the cervical cell image into the trained target detection model to obtain a candidate target area of ​​the cervical cell;

[0105] Among them, the target detection model is based on YOLOv5, adopts a multi-level connected feature pyramid network, fuses low-level feature information and high-level information based on enhanced semantic information and position information, and generates candidate target areas through the fused features.

[0106] Embodiment 3

[0107] The purpose of this embodiment is to provide a computer-readable storage medium.

[0108] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a target detection method for cervical cell images as described in the first embodiment of the present disclosure.

[0109] Embodiment 4

[0110] The purpose of this embodiment is to provide an electronic device.

[0111] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the target detection method for cervical cell images as described in the first embodiment of the present disclosure are implemented.

[0112] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for detecting a target in a cervical cell image, characterized in that: include: Acquire the cervical cell image to be detected; Input the cervical cell image into the trained target detection model to obtain the candidate target area of ​​cervical cells; Among them, the target detection model is based on YOLOv5, adopts a multi-level connected feature pyramid network, fuses low-level feature information and high-level information based on enhanced semantic information and position information, and generates candidate target areas through the fused features.

2. The target detection method of cervical cell image according to claim 1, characterized in that: The target detection model includes a feature extraction network, a feature fusion network and a detection network; The feature extraction network uses an improved CSPDarknet53 to extract feature maps of different sizes and levels from cervical cell images based on enhanced semantic information and position information; The feature fusion network adopts a multi-level connected feature pyramid network to fuse low-level feature information with high-level information based on enhanced semantic information and position information to obtain fused features; The detection network uses an improved detection head to generate bounding box coordinates, category probabilities and target confidences of candidate target regions.

3. The target detection method of cervical cell image according to claim 2, characterized in that: The improved CSPDarknet53 adds compression, dimensionality reduction and fusion of channel features and spatial features to CSPDarknet53 to obtain enhanced semantic information and location information.

4. The target detection method of cervical cell image according to claim 2, characterized in that: The multi-level connected feature pyramid network fuses features of multiple levels based on one convolution and one information filtering operation on feature maps of different sizes and levels.

5. The target detection method of cervical cell image according to claim 2, characterized in that: The improved detection heads are increased from the original three target detection heads to four target detection heads, thereby alleviating the neglect of small cell information.

6. The target detection method of cervical cell image according to claim 1, characterized in that: The target detection model constructs an objective function using classification loss, confidence loss, and positioning loss, and is trained using the constructed training data set.

7. The target detection method of cervical cell image according to claim 6, characterized in that: The construction of the training data set is specifically as follows: Get the cervical cell images and the corresponding JSON format coordinate files to form the original data set; The cervical cell images in the dataset and the corresponding JSON format coordinate files are preprocessed to obtain the final training dataset.

8. A target detection system for cervical cell images, characterized in that: Including acquisition module and detection module: The acquisition module is configured to: acquire the cervical cell image to be detected; The detection module is configured to: input the cervical cell image into the trained target detection model to obtain a candidate target area of ​​the cervical cell; Among them, the target detection model is based on YOLOv5, adopts a multi-level connected feature pyramid network, fuses low-level feature information and high-level information based on enhanced semantic information and position information, and generates candidate target areas through the fused features.

9. An electronic device, comprising: a memory for non-transitory storage of computer readable instructions; a processor for executing the computer-readable instructions; Wherein, when the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 7 is executed.

10. A storage medium, characterized in that: The computer-readable instructions are non-transitorily stored, wherein when the computer-readable instructions are executed by a computer, the method of any one of claims 1 to 7 is performed.

Images