CN117788454A

CN117788454A - Method for improving efficiency of cervical cancer liquid-based cell screening analysis system

Info

Publication number: CN117788454A
Application number: CN202410018661.0A
Authority: CN
Inventors: 方润; 白友龙; 曾新华; 廖乘胜; 曾立波; 曾立平
Original assignee: Hangzhou Xinuo Intelligent Medicine Co ltd
Current assignee: Hangzhou Xinuo Intelligent Medicine Co ltd
Priority date: 2024-01-04
Filing date: 2024-01-04
Publication date: 2024-03-29

Abstract

The invention discloses a method for improving efficiency of a cervical cancer liquid-based cell screening analysis system, which relates to the technical field of cervical cancer liquid-based cell screening and solves the problems of low screening efficiency and high misjudgment, wherein an image denoising module, a contrast enhancement module, a sharpness enhancement module and a morphological transformation module are used for preprocessing a cervical cancer liquid-based cell map, improving image quality, a saliency target detection algorithm is used for eliminating areas which are not required to be detected through an encoder stage and a decoder stage of an image segmentation neural network unet, analyzing time of the analysis system is shortened, the saliency target detection algorithm is used for generating a saliency probability map through a saliency map fusion module of the image segmentation neural network unet, and a self-adaptive threshold classifier is used for carrying out binarization processing on the saliency probability map, so that accurate screening of cervical cancer liquid-based cells is realized, analyzing efficiency is improved, labor and screening cost are reduced.

Description

Method for improving efficiency of cervical cancer liquid-based cell screening analysis system

Technical Field

The invention relates to the technical field of cervical cancer liquid-based cell screening, and in particular relates to a method for improving the efficiency of a cervical cancer liquid-based cell screening analysis system.

Background

The liquid-based cell screening is a method for detecting cervical cells and performing cytological classification diagnosis by adopting a liquid-based thin-layer cell detection system, and is a currently internationally advanced cytological examination technology for cervical cancer. The scanning system acquires a batch of images via the camera. The analysis system then analyzes the batch of images. The analysis system improves the analysis efficiency of the analysis system on the single slide, shortens the analysis time, improves the practicability of medical equipment and improves the economy. The prior method comprises the following steps: acquiring a spliced large image of a scanning image of 500 slides by scanning 500 liquid-based cell slides; labeling the cell staining area by a new generation automatic labeling tool X-AnyLabeling; carrying out data enhancement on the image through an image enhancement tool, namely an image enhancement tool; 500 pictures were taken according to 8:1: the scale of 1 is divided into a training set, a test set and a validation set. The neural network is trained through nested unet, and a model is obtained. For the new pictures, the liquid basal cell zone was inferred by a trained model.

In the prior art, the method for improving the efficiency of the cervical cancer liquid-based cell screening analysis system has a plurality of defects, on one hand, the pretreatment of spliced cervical cancer liquid-based cell images is lacked, the image quality is affected, and on the other hand, the existing method increases the analysis time of the analysis system by analyzing each scanned image, influences the efficiency of cervical cancer liquid-based cell screening, causes overlarge labor cost and overlarge screening cost, cannot accurately identify a cell area with significance, and greatly improves the risk of misjudgment.

Disclosure of Invention

Aiming at the defects of the technology, the invention discloses a method for improving the efficiency of a cervical cancer liquid-based cell screening analysis system, an image pretreatment mechanism is used for preprocessing a cervical cancer liquid-based cell map through an image denoising module, a contrast enhancement module, a sharpness enhancement module and a morphological transformation module, so that the problem of influencing image quality is solved, a salient target detection algorithm is used for eliminating areas which do not need to be detected through an encoder stage and a decoder stage of an image segmentation neural network unet, the analysis time of the analysis system is shortened, the analysis efficiency is improved, a salient map fusion module of the image segmentation neural network unet is used for generating a salient probability map through a salient probability mapping module, an entropy sampling module, a continuous splicing module and a fusion generation module, and an adaptive threshold classifier is used for carrying out binarization processing on the salient probability map, so that the cervical cancer liquid-based cell screening is realized, the analysis efficiency is improved, the labor cost is reduced, and the screening cost is reduced.

The invention adopts the following technical scheme:

the method for improving the efficiency of the cervical cancer liquid-based cell screening analysis system comprises the following steps:

step one, splicing a liquid-based cell slide scanned by a scanning system into a cervical cancer liquid-based cell map by adopting a computer vision library, wherein the size of the cervical cancer liquid-based cell map is 536 x 468;

Preprocessing a cervical cancer liquid-based cytogram through an image preprocessing mechanism, wherein the image preprocessing mechanism comprises an image denoising module, a contrast enhancement module, a sharpness enhancement module and a morphological transformation module, the output end of the image denoising module is connected with the input end of the contrast enhancement module, the output end of the contrast enhancement module is connected with the input end of the sharpness enhancement module, and the output end of the sharpness enhancement module is connected with the input end of the morphological transformation module;

performing salient object detection on the preprocessed cervical cancer liquid-based cell map by adopting a salient object detection algorithm, and identifying a cell region with saliency, wherein the salient object detection algorithm performs feature extraction by using residual U blocks with different heights in an encoder stage of an image segmentation neural network unet to remove a region without detection, and a decoder stage restores the high-wide dimension of an output feature map to the high-wide dimension of an input feature map to obtain six feature extraction maps, wherein the six feature extraction maps comprise a sixth feature extraction map Sup6, a fifth feature extraction map Sup5, a fourth feature extraction map Sup4, a third feature extraction map Sup3, a second feature extraction map Sup2 and a first feature extraction map Sup1;

In the third step, the memory stores program instructions of a saliency target detection algorithm through a computer readable storage medium, the program instructions are executed by the processor to realize the saliency target detection algorithm to perform saliency target detection on the preprocessed cervical cancer liquid-based cell map, and a cell area with saliency is identified;

generating a saliency probability map by the saliency target detection algorithm through a saliency map fusion module of the image segmentation neural network unet, wherein the saliency map fusion module of the image segmentation neural network unet comprises a saliency probability mapping module, an entropy sampling module, a continuous splicing module and a fusion generating module;

in the fourth step, the output end of the saliency probability mapping module is connected with the input end of the entropy sampling module, the output end of the entropy sampling module is connected with the input end of the continuous splicing module, and the output end of the continuous splicing module is connected with the input end of the fusion generating module;

and fifthly, performing binarization processing on the saliency probability map by adopting an adaptive threshold classifier, dividing the saliency probability map into a target area and a non-target area, wherein the target area is used as a liquid-based cell area to realize cervical cancer liquid-based cell screening.

As a further technical scheme of the invention, the image denoising module performs smoothing and denoising operations on the cervical cancer liquid-based cell map by adopting mean filtering, the mean filtering processes perform averaging operations on pixels in the cervical cancer liquid-based cell map through a 5x5 mean filter template to obtain a pixel neighborhood average value, the mean filtering processes replace pixel values in the cervical cancer liquid-based cell map according to the pixel neighborhood average value to realize smoothing and denoising, the contrast enhancement module adjusts brightness and color of the image by adopting stretching gray scale transformation, the stretching gray scale transformation performs linear scaling of a minimum value and a maximum value on the cervical cancer liquid-based cell map by linear stretching to realize gray scale expansion, the stretching gray scale transformation performs nonlinear transformation on the cervical cancer liquid-based cell map by an exponential function to effectively enhance detail information of a low gray scale region, and finally, the stretching gray scale transformation redistributes occurrence frequencies of gray scales in the cervical cancer liquid-based cell map through histogram equalization to realize contrast enhancement to be 1500:1.

According to the technical scheme, the sharpness enhancement module adopts a guide average sharpening filter to sharpen the edges and details of the cervical cancer liquid-based cell map, the guide average sharpening filter is used for sharpening the cervical cancer liquid-based cell map while filtering the cervical cancer liquid-based cell map through a visible light image fusion algorithm, the visible light image fusion algorithm is used for carrying out weighted average on the cervical cancer liquid-based cell map through scale decomposition and nonlinear filtering, the visible light image fusion algorithm is used for carrying out reconstruction sharpening on the weighted average cervical cancer liquid-based cell map through Laplacian pyramid inverse transformation, the visual sharpening effect of the cervical cancer liquid-based cell map is achieved, the morphological transformation module is used for removing noise points and holes in the cervical cancer liquid-based cell map through rectangular structural elements and corrosion operation, the morphological transformation module is used for carrying out separation degree processing on the object contour in the cervical cancer liquid-based cell map through expansion nuclear structural elements and expansion operation, and the cervical cancer liquid-based cell map is clear in object contour through iterative expansion operation.

As a further embodiment of the present invention, the working method of the encoder stage and the decoder stage of the image segmentation neural network unet is as follows:

s1, inputting a preprocessed cervical cancer liquid-based cell map as a feature map into 7 convolution layer residual error U blocks RSU-7 of a first-stage encoder En_1, wherein in the encoding process of the 7 convolution layer residual error U blocks RSU-7, the 7 convolution layer residual error U blocks RSU-7 perform feature extraction and compression on the input feature map through 2 convolution operations and 5 downsampling operations, and in the decoding process of the 7 convolution layer residual error U blocks, the 7 convolution layer residual error U blocks RSU-7 restore the height and width of the input feature map through 5 upsampling operations;

s2, inputting the result output by the S1 into 6 convolution layer residual U blocks RSU-6 of a second-stage encoder En_2, wherein in the encoding process of the 6 convolution layer residual U blocks RSU-6, the 6 convolution layer residual U blocks RSU-6 further process and compress an input feature map through 2 convolution operations and 4 downsampling operations, and in the decoding process of the 6 convolution layer residual U blocks RSU-6, the 6 convolution layer residual U blocks RSU-6 restore the height and width of the input feature map through 4 upsampling operations;

s3, inputting a result output by the second-stage encoder En_2 into 5 convolution layer residual U blocks RSU-5 of the third-stage encoder En_3, wherein in the encoding process of the 5 convolution layer residual U blocks RSU-5, the 5 convolution layer residual U blocks RSU-5 further extract and compress the input feature map through 2 convolution operations and 3 downsampling operations, and in the decoding process of the 5 convolution layer residual U blocks RSU-5, the 5 convolution layer residual U blocks RSU-5 further restore the height and width of the input feature map through 3 upsampling operations;

S4, inputting a result output by the third-stage encoder En_3 into 4 convolution layer residual U blocks RSU-4 of the fourth-stage encoder En_4, wherein in the encoding process of the 4 convolution layer residual U blocks RSU-4, the 4 convolution layer residual U blocks RSU-4 further extract and compress the input feature map through 2 convolution operations and 2 downsampling operations, and in the decoding process of the 4 convolution layer residual U blocks RSU-4, the 4 convolution layer residual U blocks RSU-4 further restore the width of the input feature map through 2 upsampling operations;

s5, inputting a result output by the fourth-stage encoder En_4 into 4 convolution layer expansion version residual error U blocks RSU-4F of the fifth-stage encoder En_5, wherein in the encoding process of the 4 convolution layer expansion version residual error U blocks RSU-4F, the 4 convolution layer expansion version residual error U blocks RSU-4F further extract and compress the input characteristic diagram through 2 convolution operations and 2 expansion convolution operations, and in the decoding process of the 4 convolution layer residual error U blocks RSU-4, the 4 convolution layer residual error U blocks RSU-4 further restore the width of the input characteristic diagram through 2 expansion convolution operations;

s6, inputting the result output by the fifth-stage encoder En_5 into 4 convolution layer expanded residual error U blocks RSU-4F of a sixth-stage encoder En_6, wherein the sixth-stage encoder En_6 obtains a decoder network layer 5 output result De_5 through up-sampling operation, and the up-sampling operation obtains a decoder network layer 4 output result De_4, a decoder network layer 3 output result De_3, a decoder network layer 2 output result De_2 and a decoder network layer 1 output result De_1;

S7, the output of the sixth-level encoder En_6, the decoder network 5 th layer output result De_5, the decoder network 4 th layer output result De_4, the decoder network 3 rd layer output result De_3, the decoder network 2 nd layer output result De_2 and the decoder network 1 st layer output result De_1 is passed through a convolution layer of 3*3, and the obtained feature map is restored to the size of the input image by a bilinear interpolation method, so that a sixth feature extraction map Sup6, a fifth feature extraction map Sup5, a fourth feature extraction map Sup4, a third feature extraction map Sup3, a second feature extraction map Sup2 and a first feature extraction map Sup1 are respectively obtained.

As a further embodiment of the present invention, the convolution operation performs convolution calculation by using a space-time sampling graph convolution kernel, the space-time sampling graph convolution kernel corrects and adjusts each pixel point of the feature graph by using a point multiplication operation function and convolution properties, and extracts feature information of different scales, the convolution operation realizes extraction of more abstract feature information by stacking multiple layers of convolutions, and the feature information extraction calculation formula is as follows:

in the formula (1), H is the characteristic information extracted by the convolution operation, c is the scale value corresponding to the characteristic information extracted by the convolution operation, b is the number of convolution kernels of the space-time sampling graph, and x is the scale value of the input characteristic graph; g is the weighted value of the point multiplication operation function to each pixel point;

The downsampling operation reduces the space size of the feature map by half under the condition that the depth of the feature map is not changed through a maximum pooling layer, the maximum pooling layer calculates the maximum element of a local area in the input feature map through a plurality of rounds of maximum pooling functions to create an output feature map, the plurality of rounds of maximum pooling functions realize the space size of the feature map which is regulated and output through a pooling window with the stride of 2, and the calculation formula of the maximum element of the local area in the input feature map is as follows:

in the formula (2), Q is the largest element of a local area in the input feature map, y is the feature value of the input feature map, i is a scale index corresponding to the feature value of the input feature map, epsilon is the size of a pooling window of a maximum pooling layer, and s is the spatial size of the input feature map;

the up-sampling operation repairs and complements the down-sampled output feature map by adopting a transposed convolution network, the transposed convolution network recovers the resolution of the output feature map to the same resolution of the input feature map by carrying out deconvolution conversion on the output feature map, the transposed convolution network splices the deconvolution converted feature map by a bicubic interpolation function to recover the output feature map to be the high and wide dimension of the input feature map, and the deconvolution conversion calculation formula is as follows:

In the formula (3), D is an output characteristic diagram after deconvolution conversion, f is a resolution difference value between the output characteristic diagram and an input characteristic diagram, j is a moving step length of deconvolution conversion of a convolution kernel in a transposed convolution network, and N is the number of channels contained in the output characteristic diagram in the transposed convolution network;

the calculation formula of the bicubic interpolation function is as follows:

in the formula (4), K is the spliced output characteristic diagram, a is the difference of height dimension between the output characteristic diagram and the input characteristic diagram, p is the interpolation coefficient matrix of the bicubic interpolation function, and r is the difference of width dimension between the output characteristic diagram and the input characteristic diagram;

the expansion convolution operation adopts a cavity convolution kernel to enlarge 97% convolution sampling rate, so that a multiscale receptive field for an input feature map is realized, the cavity convolution kernel adjusts the cavity aperture in the convolution kernel through a compressed residual error network to obtain a 7x7 effective convolution kernel, the cavity convolution kernel extracts feature information of different scales according to the 7x7 effective convolution kernel, and the calculation formula of the cavity aperture is as follows:

in the formula (5), T is the hole aperture, w is the convolution sampling rate, k is the number of holes in the convolution kernel, n is the average increment of the compressed residual network when the hole aperture is adjusted each time, and l is the number of times the compressed residual network adjusts the hole aperture.

As a further embodiment of the present invention, the saliency probability mapping module outputs saliency probability maps corresponding to six feature maps through a 3×3 convolution layer and an S-shaped sigmoid function, the 3×3 convolution layer performs standardization processing on the feature maps through a batch normalization function, the 3×3 convolution layer performs weight attention on feature pixels in the feature maps in the standardization processing process through an attention adding mechanism, the batch normalization function applies a leavable scaling and translation operation to the feature maps according to feature pixel variances and means, and outputs the feature maps of the standardization mapping, the S-shaped sigmoid function converts the output of the 3×3 convolution layer into the saliency probability maps through a spatial pooling operation, and the spatial pooling operation performs nonlinear mapping on the output of the 3×3 convolution layer through a self-adaptive nonlinear transformation to obtain the saliency probability maps; the entropy sampling module adopts information entropy dimension sampling to carry out entropy sampling on the convolution output of a 3 multiplied by 3 convolution layer before an S-shaped sigmoid function, the information entropy dimension sampling generates a three-dimensional countermeasure network through information entropy sampling shear wave transformation, each pixel point of a feature map is regarded as a random variable to obtain information entropy, the entropy sampling module adopts a full-connection layer to map the information entropy to output six feature maps, the full-connection layer trains weights and offsets through a reverse propagation loss function, the reverse propagation loss function updates the values of the weights and the offsets according to information entropy gradient information, and the updating process is iterated until convergence is achieved.

As a further embodiment of the present invention, the continuous stitching module uses a nonlinear residual image stitching concat operation to stitch six feature images to obtain a salient target feature image Sup0, the nonlinear residual image stitching concat operation performs channel number adjustment on each feature image through nonlinear residual convolution check to achieve that the six feature images have the same channel number, and the nonlinear residual image stitching concat operation performs element-by-element addition to stitch the six feature images in the channel direction to obtain the salient target feature image Sup0; the fusion generation module converts the saliency target feature map Sup0 through a 1X 1 convolution layer and an S-type sigmoid hyperbolic activation function to obtain a prediction probability map, the 1X 1 convolution layer segments different branch features in the saliency target feature map Sup0 of the convolution layer through target detection and semantic segmentation, the S-type sigmoid hyperbolic activation function analyzes resolution probabilities of different branch features through a spatial resolution analysis engine, and the spatial resolution analysis engine fuses the resolution probabilities of different branch features into the prediction probability map through a spatial interaction network and structural time-varying probability transformation.

As a further embodiment of the present invention, the adaptive threshold classifier determines a binarization threshold of the saliency probability map by using a binarization threshold algorithm, the binarization threshold algorithm analyzes correlations between resolution probabilities of different branch features and liquid-based cell regions through a global threshold network, the binarization threshold algorithm selects a binarization threshold through a cumulative histogram of the resolution probabilities and the liquid-based cell regions, and the adaptive threshold classifier selects a target region and a non-target region in the prediction probability map according to the binarization threshold, and eliminates the non-target region which does not need to be detected through a morphological segmentation processing engine, thereby accurately identifying the cervical cancer liquid-based cell region with saliency.

The invention has positive and beneficial effects different from the prior art:

the invention discloses a method for improving the efficiency of a cervical cancer liquid-based cell screening analysis system, which is characterized in that an image pretreatment mechanism is used for preprocessing a cervical cancer liquid-based cell map through an image denoising module, a contrast enhancement module, a sharpness enhancement module and a morphological transformation module, so that the image quality is improved, a salient target detection algorithm is used for removing areas which are not needed to be detected through an encoder stage and a decoder stage of an image segmentation neural network unet, the analysis time of the analysis system is reduced, the analysis efficiency is improved, a salient map fusion module of the image segmentation neural network unet is used for generating a salient probability map through a salient probability mapping module, an entropy sampling module, a continuous splicing module and a fusion generation module, and a self-adaptive threshold classifier is used for carrying out binarization processing on the salient probability map, so that the cervical cancer liquid-based cell screening is realized, the analysis efficiency is improved, the labor cost is reduced, and the screening cost is reduced.

Drawings

For a clearer description of an embodiment of the invention or of a technical solution in the prior art, the drawings that are necessary for the description of the embodiment or of the prior art will be briefly described, it being obvious that the drawings in the description below are only some embodiments of the invention, from which, without inventive faculty, other drawings are obtained for a person skilled in the art, in which:

FIG. 1 is a flow chart of a method for improving the efficiency of a cervical cancer liquid-based cell screening assay system according to the present invention;

FIG. 2 is a diagram of a network structure of a codec model of an image segmentation neural network employed in the present invention;

FIG. 3 is a diagram of a 7 convolutional layer residual U block RSU-7 network structure employed in the present invention;

FIG. 4 is a flowchart of a method for operating a codec model of an image-splitting neural network employed in the present invention;

fig. 5 is a schematic diagram of an image pretreatment mechanism used in the present invention.

Detailed Description

The preferred embodiments of the present invention will be described below with reference to the accompanying drawings, it being understood that the embodiments described herein are for illustration and explanation of the present invention only, and are not intended to limit the present invention.

As shown in fig. 1 to 5, a method for improving the efficiency of a cervical cancer liquid-based cell screening assay system comprises the steps of:

In a specific embodiment, the computer vision library may splice the liquid-based cell slide scanned by the scanning system into a cervical cancer liquid-based cell map, where the size of the cervical cancer liquid-based cell map is 536×468: 1. all cell slide pictures were read and pre-treated, for example: noise removal, brightness and contrast adjustment, normalized image size, etc. 2. For each cell slide picture, a feature extraction algorithm (such as SIFT, SURF or ORB) is used to extract its key points and descriptors, and calculate its corresponding matching relationship. 3. According to the matching relation, an image stitching algorithm (such as an algorithm based on panoramic stitching or local area stitching) is used for stitching all the cell slide images into a large image. 4. And performing post-processing and analysis, such as cropping, scaling or rotating the large image to meet the actual requirements. At the same time, operations such as abnormality detection and morphological feature analysis are also required. In this process, the computer vision library may use a variety of algorithms and tools to implement the functionality of the different steps. For example, SIFT, SURF, or ORB functions provided in OpenCV may be used in feature extraction; the stitcher function provided in OpenCV can be used in image stitching or a stitching algorithm can be realized by the user; in the aspect of post-processing, proper functions can be selected or programs can be developed according to requirements. It should be noted that the parameter setting and adjustment between the different steps in the process affect the final result, so repeated testing and adjustment are required to achieve the optimal effect.

In particular embodiments, image pre-processing mechanisms are an important step in computer vision for the purpose of improving image quality, reducing noise, and enhancing image features. The image pretreatment mechanism comprises an image denoising module, a contrast enhancement module, a sharpness enhancement module and a morphological transformation module, wherein the image denoising module replaces the value of a central pixel by the average value in the pixel neighborhood, so that the effect of smooth noise reduction is achieved. Specifically, when the cervical cancer liquid-based cytogram is processed using a 5×5 mean filter template, for each pixel, gray values of 5×5 pixels (including itself) around it are summed, and the average value is taken as a new gray value for that pixel. In this way, noise in the image can be effectively reduced, and the mean filtering retains better image edge information than other filtering methods. The contrast enhancement module expands the gray level of the image by stretching the gray level transformation to perform linear or non-linear transformation, thereby achieving the effect of enhancing the details and contrast of the image. Specifically, when the cervical cancer liquid-based cytogram is processed by using the tensile gray scale transformation, the following steps may be taken: linear stretching: and linearly scaling the minimum value and the maximum value in the original image to be between 0 and 255, so as to realize the expansion of the gray level. This step can effectively increase the dynamic range of the image without affecting the relative differences between the various regions in the image. Nonlinear transformation: for low gray areas (such as background, etc.), nonlinear transformation is adopted to further enhance the detail information. Common methods include exponential functions, power functions, and the like. Histogram equalization: the contrast of the whole image is improved by increasing the difference between the different gray levels in the image by reassigning the frequency of occurrence of the individual gray levels. It should be noted that too aggressive histogram equalization may result in loss of local detail information. Through the steps, the contrast ratio of the cervical cancer liquid-based cytogram can be enhanced to 1500:1, so that detailed information in the image can be better displayed.

In a specific embodiment, the convolutional layer residual U block is an encoder-decoder architecture design based on a deep convolutional neural network (Convolutional Neural Networks, CNN), and is mainly used in an image semantic segmentation task. The method consists of a plurality of layers of convolution operation and residual connection, and can effectively extract image features and reduce difficulty of feature matching. The convolution layer residual U block adopts an encoder-decoder architecture, and combines an encoder and a decoder through residual connection, so that consistency and continuity of feature expression are maintained, and the situation that a segmentation task fails due to feature loss is avoided. Specifically, the convolutional layer residual U block consists of one downsampling operation, two convolutional layer residual modules, and one upsampling operation. Each convolution layer residual module consists of two convolution operations and a skip connection, which implements residual learning (residual learning) and can directly connect the original features to the back of the convolution layer, thereby reducing the feature loss due to the convolution operations. The multi-stage convolution layer residual U block combination can realize layer-by-layer extraction and compression of image features, and finally a semantic segmentation result is obtained, so that better performance is realized in an image segmentation task.

The convolutional layer expanded residual U block is an encoder-decoder structure in a convolutional neural network and is specially used for semantic segmentation tasks. The method is mainly characterized in that the expansion convolution operation is adopted, and semantic segmentation can be carried out by utilizing the information with a larger receptive field, so that the segmentation accuracy is improved. The dilation convolution operation is a variation of the convolution operation that can expand the receptive field by increasing the number of holes (Kernel Dilated Factor) of the convolution kernel. More specifically, the dilation convolution operation often employs a convolution kernel with holes that spatially have a larger receptive field range when convolving with the input feature map, so that the contextual information of the image can be better captured. In the convolution layer expanded residual U block, each residual module not only comprises a conventional convolution transverse connection and a residual jump connection, but also adopts an expanded convolution operation, so that the scope of the receptive field is increased, and more context information can be considered. Meanwhile, the closer the residual module of the decoder is, the more the previous characteristic information needs to be considered, so as to ensure the accuracy of the result. In a word, the convolution layer expanded residual U block adopts a plurality of technologies such as residual connection, encoder-decoder structure, expanded convolution operation and the like in a convolution neural network, and can better capture image characteristics and context information, thereby realizing higher performance in an image segmentation task.

the calculation formula of the bicubic interpolation function is as follows:

In a specific embodiment, an application platform of a coding and decoding model of an image segmentation neural network adopted by a saliency target detection algorithm can be constructed, and in the working process of the platform, a hardware platform can be constructed, for example, the following components are constructed for processing: central processing units, graphics processors, storage devices, monitors and keyboard mice, network interfaces and other auxiliary devices, the following is a more detailed description of embodiments:

1. central Processing Unit (CPU): the method is used for controlling the whole platform, training and testing the coding and decoding model of the image segmentation neural network, and carrying out subsequent processing and management.

2. Graphics Processor (GPU): and the operation and the image processing of the pixel level are specially carried out, so that the operation efficiency of the coding and decoding model of the image segmentation neural network is improved.

3. A storage device: for storing data such as raw data sets, training and testing data sets, feature maps, saliency maps, and the like.

4. Monitor and keyboard mouse: the system is used for interacting with the platform, and displaying and adjusting training and testing results of the coding and decoding model of the image segmentation neural network.

5. Network interface: for transmitting data into the platform including raw images, training and testing data sets, etc.

6. Other auxiliary devices: including, for example, servers, routing devices, etc., for connecting multiple processors and storage devices, support the expansion and management of the platform.

Through the combination of the hardware components, an application platform of a coding and decoding model of the high-performance and high-efficiency image segmentation neural network can be built, and training, testing, optimizing and applying of a saliency target detection algorithm are supported.

The convolution operation is a common image processing method in deep learning, and feature information of different scales is extracted by performing convolution calculation on an input feature map. In this process, the spatio-temporal sample map convolution kernel plays a key role. The convolution kernel of the space-time sampling graph corrects and adjusts each pixel point of the feature graph through a point multiplication operation function and convolution property, so that feature information of different scales is extracted. Specifically, the convolution kernel may be thought of as a filter that performs a weighted summation of each pixel in the input feature map with pixels in its neighboring region and outputs a new value. Through the combination of convolution kernels with different sizes and numbers, the input data can be analyzed and understood in multiple layers, multiple angles and multiple directions. In deep learning models, multiple layers of convolution operations are typically stacked to achieve higher-level, more abstract feature extraction. With the increase of the layer number, the model can gradually understand more complex and abstract information in the data, and gradually realize the tasks of fine and accurate classification and identification. Therefore, in the deep learning model, convolution operation performs convolution calculation on an input feature map through a space-time sampling map convolution check, extracts feature information of different scales, and extracts and analyzes more abstract feature information by stacking multiple layers of convolutions.

The dilation convolution operation is an image processing method that can improve the performance of the model by performing multiscale receptive field processing on the input feature map. Under the condition that the 97% convolution sampling rate is enlarged by adopting the cavity convolution kernel, the receptive field in a larger range can be realized, and the problem that the number of layers and the parameter quantity are required to be increased in the traditional convolution operation is avoided. Specifically, in the methods described herein, the hole aperture in the hole convolution kernel is adjusted by compressing the residual network to yield a 7x7 effective convolution kernel. This process can achieve more efficient, accurate feature extraction with a relatively small number of model parameters maintained. Meanwhile, when the 7x7 effective convolution kernel is adopted, the characteristic information can be extracted from different scales, so that the understanding and analysis capability of the model on input data is further enhanced. In summary, the expansion convolution operation adopts the cavity convolution kernel to expand the 97% convolution sampling rate, and adjusts the cavity aperture through the compressed residual network to obtain the 7x7 effective convolution kernel, so as to realize the multi-scale receptive field of the input feature map and the extraction of the feature information with different scales. The statistical table of the calculation results of the hole diameters is shown in table 1:

As shown in table 2, four test groups are set, two methods are adopted to calculate the hole aperture respectively, the method 1 introduces a corresponding compensation network, the characteristic of the compensation network is utilized to offset the parameter change caused by the interrupt signal change, the hole aperture is obtained, the method 2 is that the hole convolution kernel obtains a 7x7 effective convolution kernel by adjusting the hole aperture in the convolution kernel through a compressed residual network, the error of the method 1 is larger than that of the method 2, and the hole convolution kernel obtains the 7x7 effective convolution kernel through the compression residual network.

In a specific embodiment, a saliency map fusion model of the image segmentation neural network is one of the important components to implement the saliency target detection algorithm. The saliency map fusion model of the image segmentation neural network comprises the following modules: the saliency probability mapping module is an important component part in a saliency map fusion model of the image segmentation neural network, and outputs saliency probability mapping corresponding to six feature maps through a 3X 3 convolution layer and an S-shaped sigmoid function. Specifically, this module comprises the following steps: 3 x 3 convolutional layer: the layer receives an input original image and extracts a plurality of feature maps through a series of convolution operations. These feature maps may reflect the degree of similarity and the degree of difference between different regions in the input image. In the convolution process, a batch normalization function is adopted to carry out standardization processing on the feature map so as to reduce the problem of internal covariate displacement during training. Attention mechanism: the mechanism can weight each pixel point in the feature map, so that important areas and information are focused better. Specifically, during the normalization process, attention to the feature pixel weight is achieved by adding an attention mechanism. Batch normalization function: the function applies a leachable scaling and translation operation to the feature map according to the feature pixel variance and the mean value, and outputs a normalized mapped feature map. Therefore, the distribution of the characteristic diagrams is more stable, and the convergence speed and generalization capability of the network are improved. Sigmoid function of S shape: the function converts the output of the 3 x 3 convolutional layer into a saliency probability map through a spatial pooling operation. Specifically, the output of the 3×3 convolution layer is subjected to nonlinear mapping by adopting adaptive nonlinear transformation to obtain saliency probability mapping, so that important target information in an image is captured better. The entropy sampling module is an important component in the image segmentation neural network, and performs entropy sampling on the convolution output of the 3X 3 convolution layer before the S-shaped sigmoid function through information entropy dimension sampling, so that the robustness and generalization capability of the network are improved. Specifically, the entropy sampling module processes and analyzes the input image through the steps of information entropy dimension sampling, a full connection layer, a reverse propagation loss function and the like, and finally six feature graphs are generated. The feature maps can be used as input of a subsequent target detection algorithm and guide the algorithm to classify, identify, divide and the like the input image. Meanwhile, due to the adoption of an entropy sampling technology, the module has certain robustness and generalization capability, and is more effective in coping with input data in a complex scene.

The continuous stitching module is a key component in the image segmentation neural network, and is used for stitching the six feature images through nonlinear residual image stitching concat operation to obtain a salient target feature image Sup0. Specifically, the continuous stitching module performs stitching on six feature images in the channel direction through nonlinear residual image stitching concat operation to obtain a saliency target feature image Sup0. The feature maps can be used as input of a subsequent target detection algorithm and guide the algorithm to classify, identify, divide and the like the input image. Meanwhile, as a nonlinear residual convolution kernel and an element-by-element addition method are adopted, the module has certain nonlinear transformation and information fusion capacity, and is more effective in the aspects of extracting important features in input data, enhancing network perception capacity and the like. The fusion generation module is an important component in the image segmentation neural network, converts the saliency target feature map Sup0 through a 1X 1 convolution layer and an S-type sigmoid hyperbolic activation function to obtain a prediction probability map, and particularly converts the saliency target feature map Sup0 through the 1X 1 convolution layer and the S-type sigmoid hyperbolic activation function to obtain the prediction probability map. The feature maps can be used as input of a subsequent target detection algorithm and guide the algorithm to classify, identify, divide and the like the input image. Meanwhile, due to the adoption of methods such as space interaction network and structure time transformation, the module has certain information fusion and space resolution capability, and is more effective in the aspects of improving network performance, enhancing target detection effect and the like.

The adaptive threshold classifier is an important component in the image segmentation neural network, determines a binarization threshold value of the saliency probability map through a binarization threshold algorithm, and selects a target region and a non-target region in the prediction probability map according to the threshold value. Specifically, this classifier includes the following steps: binarization thresholding algorithm: the step analyzes the correlation of the resolution probability of different branch features and the liquid-based cell area through a global threshold network, and selects a binarization threshold through the cumulative histogram of the resolution probability and the liquid-based cell area. The global threshold network herein refers to a neural network structure that can perform global statistics and analysis on input data, and can effectively extract global features in the input data and convert them into a form suitable for subsequent processing and analysis. Target area selection: the step selects a target region and a non-target region in the predictive probability map based on the binarization threshold. Specifically, a pixel point in the saliency probability map equal to or higher than the binarization threshold is regarded as a target region, and a pixel point smaller than the threshold is regarded as a non-target region. Morphological segmentation: according to the method, a morphological segmentation processing engine is used for eliminating non-target areas which are not required to be detected, and cervical cancer liquid-based cell areas with significance are accurately identified. Morphological segmentation here refers to an image processing technique that allows shape transformation and structural optimization of the target region, which can efficiently extract key information in the target region and convert it into a form suitable for subsequent analysis and processing. Therefore, the self-adaptive threshold classifier determines the binarization threshold of the saliency probability map through a binarization threshold algorithm, selects a target area and a non-target area in the prediction probability map according to the threshold, and realizes accurate identification of the liquid-based cell area through a morphological segmentation processing engine. The technical methods can effectively improve network performance, enhance target detection effect and the like.

While specific embodiments of the present invention have been described above, it will be understood by those skilled in the art that these specific embodiments are by way of example only, and that various omissions, substitutions, and changes in the form and details of the methods and systems described above may be made by those skilled in the art without departing from the spirit and scope of the invention. For example, it is within the scope of the present invention to combine the above-described method steps to perform substantially the same function in substantially the same way to achieve substantially the same result. Accordingly, the scope of the invention is limited only by the following claims.

Claims

1. A method for improving the efficiency of a cervical cancer liquid-based cell screening analysis system is characterized by comprising the following steps: the method comprises the following steps:

2. The method for improving the efficiency of a cervical cancer liquid-based cell screening assay system according to claim 1, wherein: the image denoising module performs smoothing and denoising operations on the cervical cancer liquid-based cell map by adopting mean filtering, the mean filtering processes perform average operation on pixels in the cervical cancer liquid-based cell map through a 5x5 mean filter template to obtain a pixel neighborhood average value, the mean filtering processes replace pixel values in the cervical cancer liquid-based cell map according to the pixel neighborhood average value to realize smooth denoising, the contrast enhancement module adjusts brightness and color of the image by adopting stretching gray level transformation, the stretching gray level transformation performs linear scaling of minimum and maximum values on the cervical cancer liquid-based cell map by linear stretching to realize gray level expansion, the stretching gray level transformation performs nonlinear transformation on the cervical cancer liquid-based cell map by an exponential function to effectively enhance detail information of a low gray level region, and finally, the stretching gray level transformation redistributes occurrence frequencies of various gray levels in the cervical cancer liquid-based cell map by histogram equalization to realize contrast enhancement to 1500:1.

3. The method for improving the efficiency of a cervical cancer liquid-based cell screening assay system according to claim 1, wherein: the sharpness enhancement module adopts a guiding average sharpening filter to sharpen the edge and detail of the cervical cancer liquid-based cell map, the guiding average sharpening filter is used for sharpening the cervical cancer liquid-based cell map while filtering the cervical cancer liquid-based cell map through a visible light image fusion algorithm, the visible light image fusion algorithm is used for carrying out weighted average on the cervical cancer liquid-based cell map through scale decomposition and nonlinear filtering, the visible light image fusion algorithm is used for carrying out reconstruction sharpening on the weighted average cervical cancer liquid-based cell map through Laplacian pyramid inverse transformation, the visual sharpening effect of the cervical cancer liquid-based cell map is achieved, the morphological transformation module is used for removing noise points and holes in the cervical cancer liquid-based cell map through rectangular structural elements and corrosion operation, the morphological transformation module is used for carrying out separation degree processing on the object outline in the cervical cancer liquid-based cell map through expansion nuclear structural elements and expansion operation, and the cervical cancer liquid-based cell map is used for obtaining clear object outline through iterative expansion operation.

4. The method for improving the efficiency of a cervical cancer liquid-based cell screening assay system according to claim 1, wherein: the working method of the encoder stage and the decoder stage of the image segmentation neural network unet comprises the following steps:

5. The method for improving the efficiency of a cervical cancer liquid-based cell screening assay system according to claim 4, wherein: the convolution operation carries out convolution calculation on an input feature map through a space-time sampling map convolution kernel, the space-time sampling map convolution kernel corrects and adjusts each pixel point of the feature map through a point multiplication operation function and convolution properties, feature information of different scales is extracted, the convolution operation realizes extraction of more abstract feature information through stacking multiple layers of convolutions, and a feature information extraction calculation formula is as follows:

the calculation formula of the bicubic interpolation function is as follows:

6. The method for improving the efficiency of a cervical cancer liquid-based cell screening assay system according to claim 1, wherein: the saliency probability mapping module outputs saliency probability mapping corresponding to six feature images through a 3X 3 convolution layer and an S-shaped sigmoid function, the 3X 3 convolution layer performs standardization processing on the feature images through a batch normalization function, the 3X 3 convolution layer realizes the weight attention of feature pixels in the feature images in the standardization processing process through an attention adding mechanism, the batch normalization function applies learning scaling and translation operation to the feature images according to feature pixel variances and means to output the feature images of the standardization mapping, the S-shaped sigmoid function converts the output of the 3X 3 convolution layer into the saliency probability mapping through a spatial pooling operation, and the spatial pooling operation performs nonlinear mapping on the output of the 3X 3 convolution layer through self-adaptive nonlinear transformation to obtain the saliency probability mapping; the entropy sampling module adopts information entropy dimension sampling to carry out entropy sampling on the convolution output of a 3 multiplied by 3 convolution layer before an S-shaped sigmoid function, the information entropy dimension sampling generates a three-dimensional countermeasure network through information entropy sampling shear wave transformation, each pixel point of a feature map is regarded as a random variable to obtain information entropy, the entropy sampling module adopts a full-connection layer to map the information entropy to output six feature maps, the full-connection layer trains weights and offsets through a reverse propagation loss function, the reverse propagation loss function updates the values of the weights and the offsets according to information entropy gradient information, and the updating process is iterated until convergence is achieved.

7. The method for improving the efficiency of a cervical cancer liquid-based cell screening assay system according to claim 1, wherein: the continuous splicing module splices six feature images by adopting nonlinear residual image splicing concat operation to obtain a salient target feature image Sup0, the nonlinear residual image splicing concat operation carries out channel number adjustment on each feature image through nonlinear residual convolution check to realize that the six feature images have the same channel number, and the nonlinear residual image splicing concat operation carries out element-by-element addition to splice the six feature images in the channel direction to obtain the salient target feature image Sup0; the fusion generation module converts the saliency target feature map Sup0 through a 1X 1 convolution layer and an S-type sigmoid hyperbolic activation function to obtain a prediction probability map, the 1X 1 convolution layer segments different branch features in the saliency target feature map Sup0 of the convolution layer through target detection and semantic segmentation, the S-type sigmoid hyperbolic activation function analyzes resolution probabilities of different branch features through a spatial resolution analysis engine, and the spatial resolution analysis engine fuses the resolution probabilities of different branch features into the prediction probability map through a spatial interaction network and structural time-varying probability transformation.

8. The method for improving the efficiency of a cervical cancer liquid-based cell screening assay system according to claim 1, wherein: the self-adaptive threshold classifier adopts a binarization threshold algorithm to determine a binarization threshold of the saliency probability map, the binarization threshold algorithm analyzes the correlation between the resolution probabilities of different branch features and the liquid-based cell region through a global threshold network, the binarization threshold algorithm selects the binarization threshold through the cumulative histogram of the resolution probabilities and the liquid-based cell region, the self-adaptive threshold classifier selects a target region and a non-target region in the prediction probability map according to the binarization threshold, and the self-adaptive threshold classifier eliminates the non-target region which does not need to be detected through a morphological segmentation processing engine to accurately identify the cervical cancer liquid-based cell region with saliency.