Cervical picture image classification method and system based on machine learning

Abstract

The invention relates to a machine learning-based cervical picture image classification method and system. The method comprises the following steps of: S1, collecting a cervical picture and extracting a region of interest; s2, determining a first feature representing a local color feature based on the region of interest, and determining a second feature representing a structural relationship feature; and S3, inputting the first feature and the second feature into an artificial intelligence model to obtain a classification result. According to the method, priori knowledge extraction is carried out by means of local color information and global texture organization structure information, medical logic is embedded in the calculation process, a machine learning model is assisted to carry out rapid classification, the classification complexity is reduced, and the classification efficiency of cervical pictures is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image classification, and particularly relates to a cervical picture image classification method and system based on machine learning.

Background Art

[0002] Cervical cancer is one of the malignant tumors that seriously threaten the physical and mental health of women. At present, the main methods for cervical cancer screening include human papillomavirus nucleic acid testing and cytological examination. During the implementation of the organized cervical cancer screening work for women of appropriate age in various regions of China, it has been found that the lack of the number and ability of cytologists, especially in grass-roots areas, is an important factor affecting the detection rate of cervical cytological abnormalities. Cervical cytological examination is an important means for cervical cancer screening. It mainly involves pathologists observing the characteristics of cervical cells under a microscope with the naked eye, evaluating cell types and morphological characteristics to provide a basis for clinical diagnosis and treatment. However, manual diagnosis is time-consuming and laborious, with strong subjectivity in the process, requiring high professional knowledge and experience, and the training cycle of pathologists is long.

[0003] Artificial intelligence, as an outstanding representative of modern technology, has demonstrated its powerful application potential and value in multiple fields. In the field of healthcare, the application of artificial intelligence has become increasingly common and in-depth. With the help of advanced algorithms and big data processing capabilities, artificial intelligence can diagnose diseases quickly and accurately. The following is the classification and typical cases in the main directions: One is in the field of healthcare, which can be used for medical image diagnosis: Artificial intelligence analyzes images such as X-rays and CTs through deep learning. For example, the lung nodule screening system of Alibaba Health can identify the risk of lung cancer in seconds. It can also be used for decision-making assistance and drug research and development: IBM Watson Health can analyze medical data to recommend treatment plans; AlphaFold predicts protein structures to accelerate the development of new drugs. In addition, health management has also been widely applied: Smart wearable devices (such as Huawei's electrocardiogram bracelet) can monitor users' health data in real time and give warnings.

[0004] Currently, the use of artificial intelligence in cervical cytology screening has also become a hot topic. Extensive research has been conducted on developing automatic diagnostic analysis methods for cervical cell images using computer intelligence and machine learning to automatically screen for cervical cancer and improve detection accuracy. Artificial intelligence-assisted diagnosis in cervical cytology: It refers to after cervical exfoliated cells are made into slides through a base solution, digital slides are generated through imaging by a digital slide scanner, and then analyzed by computer technology based on artificial intelligence. The artificial intelligence gives a negative or positive judgment result, and screens out suspicious lesion cells to assist cytopathologists in making cytopathological diagnoses more quickly and accurately. The scanned images should highly restore the characteristics of the original glass slides, and be clear, colorless, seamless, and support single-layer and multi-layer scanning. Deep learning and other artificial intelligence technologies are used to detect and intelligently analyze digital pathology images, and the positive and suspected positive cell fields are screened and presented to pathologists to assist cytopathologists in finally making cytopathological diagnoses. In recent years, some detection models based on artificial intelligence technology have been applied to the intelligent detection of cervical cytology images, greatly reducing the dependence on manual features and saving the cell segmentation step, thereby improving the efficiency of automatic detection of cervical lesions. Using artificial intelligence technology, it can identify abnormal patterns in cell images like a highly trained detective, and gradually improve the accuracy through continuous learning. When a pathologist gets a slide, they can observe whether the cells are cancerous just by looking through a microscope. However, if the diseased tissues are unevenly distributed on the slide, then the doctor needs to repeatedly move the slide to observe. This is not only time-consuming and laborious but may also lead to missed diagnoses. In fact, the same problem exists during artificial intelligence processing. When classifying pictures, the large image is first divided into blocks, then the diseased areas are automatically tracked and identified, magnified and analyzed, and finally the case-level interpretation is completed and the results are output; but because the acquisition method of cervical pictures is different from that of other images, there is a great correlation between the final image results and the slicing method and process. Therefore, in order to obtain a reference classification result, the same acquisition result needs to be repeatedly sliced and processed multiple times, and there may be large differences between each sliced image, and repeated image classification and a large amount of processing are required; in addition, the interior of the same image block is often chaotic, which causes great trouble to the intelligent model. And directly training the whole slide with the artificial intelligence model is also difficult to obtain a high accuracy due to the large differences in the image results, and the training efficiency is very low and the training complexity is very high. In addition, in the existing technology, the method of image classification is often based on the grayscale of the pictures, but the colors in cervical pictures contain a lot of information, and discarding this information will obviously cause information loss; based on the above problems, the present invention extracts prior knowledge by means of comprehensive local color information and global texture tissue structure information, embeds medical logic in the calculation process, and assists the machine learning model in rapid classification, reducing the classification complexity and improving the classification efficiency of cervical pictures.

Summary of the Invention

[0005] To solve the above problems in the prior art, the present invention proposes a method and system for classifying cervical picture images based on machine learning. The method includes:

[0006] Step S1: Extract the region of interest from the collected cervical pictures;

[0007] Step S2: Determine the first feature representing the local color feature and the second feature representing the structural relationship feature based on the region of interest;

[0008] The specific method for determining the first feature representing the local color feature is as follows: Set a sliding window to slide within the region of interest, and obtain the local color feature within the sliding window during the sliding process; Represent the local color feature in a histogram to construct the first feature; where: the sliding step of the sliding window is less than the size of the sliding window;

[0009] The specific method for determining the second feature representing the structural relationship feature is as follows: Set a sliding window to slide within the region of interest, and obtain the structural relationship feature between the sliding window and the adjacent windows during the sliding process; After splicing the structural relationship features, form the second feature; The specific method for extracting the structural relationship feature between the sliding window and the adjacent windows is as follows:

[0010] Step S2B21: Calculate the gray-scale mean value of the sliding window w3 and the gray-scale mean value of each of the adjacent windows k of this sliding window 8 k = 1 to 8;

[0011] Step S2B22: If then this adjacent window is called a difference window k1; where: trg is the gray-scale difference threshold;

[0012] Step S2B23: Determine the number N of difference windows k1 to determine the occurrence probability as N / 8;

[0013] Step S2B24: Calculate the contrast, energy, entropy, and homogeneity of the sliding window w3 based on the occurrence probability of N / 8;

[0014] Step S3: Input the first feature and the second feature into the artificial intelligence model to obtain the classification result.

[0015] Furthermore, a segmentation model based on U-Net is used to locate the cervical region as the region of interest.

[0016] Furthermore, the cervical pictures are H&E stained cervical pathological sections collected by a 40-fold microscope.

[0017] Further, image preprocessing is performed before extracting the region of interest.

[0018] Further, the artificial intelligence model is a machine learning model.

[0019] Further, the determination of the first feature representing local color features is specifically as follows:

[0020] Step S2A1: Set the sliding window to include a first window and a second window, and the window size of each window is W×H; there is a preset interval size INT between the first window and the second window; set the initial position of the sliding window at the upper left corner of the region of interest; set the sliding step size to WX, and INT > W;

[0021] Step S2A2: Make the sliding window slide within the region of interest in the preset sliding direction according to the sliding step size, and extract the local color features within the sliding window after each slide; the local color features include the calculated value of the average gray scale gw of the first window and the second window, the calculated value of the average brightness b w of the calculated value and / or the calculated value of the saturation s w of the calculated value;

[0022] Step S2A3: Represent the local color features in a histogram to construct the first feature; specifically: Represent each of the local color features in a histogram H it ; where: t is the type of the local color feature; t = g, b, w indicates that its type is average gray scale, average brightness and saturation; it = 0~IT-1 is the histogram value range of the histogram element number it of type t; t(w) is the local color feature value of type t of the sliding window w in the region of interest, δ is the Dirac function, and its value is 1 when its parameter (t(w)-it) falls into [0, 1), otherwise it is 0;

[0023] H it＝0～IT-1 = ∑ w δ(t(w)-it) (1);

[0024] Step S2A4: Put the histogram representation H it into the t-th row and the i-th column of the matrix to form the first feature in matrix form.

[0025] A cervical picture image classification system based on machine learning, and the cervical picture image classification system based on machine learning is used to implement the above-mentioned cervical picture image classification method based on machine learning.

[0026] A cervical picture image classification device based on machine learning, and the cervical picture image classification device based on machine learning is used to implement the above-mentioned cervical picture image classification method based on machine learning.

[0027] A cervical picture image classification control server based on machine learning, and the cervical picture image classification server based on machine learning is used to implement the above-mentioned cervical picture image classification method based on machine learning.

[0028] A cervical picture image classification chip based on machine learning, and the cervical picture image classification chip based on machine learning is used to implement the above-mentioned cervical picture image classification method based on machine learning.

[0029] The beneficial effects of the present invention include:

[0030] (1) By using local color information and global texture organizational structure information as prior knowledge extraction, medical logic is embedded in the calculation process; the information acquisition amount and the sensitivity of local pixel points are reduced through a sliding window, and the semi-sliding extraction method eliminates the differential impact of the acquisition and slicing methods on the color of the image to a certain extent, further comprehensively acquires all color features, and all are represented by a continuous numerical histogram, making the local color features global, discovering the randomness of the spatial distribution of color changes in the picture, to adapt to the characteristics of frequent spatial changes in color feature situations; the gradual change features of adjacent regions can be captured by the window sliding method that does not completely slide out each time.

[0031] (2) Combining the structural relationship characteristics brought by the cervical picture and its acquisition method, through an 8-neighbor comparison method during the sliding process, and through a probability calculation method on the basis of regional comparison, the unevenness of the detected cell nuclei and the regular changes in cell arrangement are discovered, so as to fully obtain the structural information as prior information; the complexity of subsequent classification calculation is reduced.

[0032] (3) Since the information acquisition granularity is reduced to the sliding window granularity, it is hardware-friendly and has low complexity, improving the classification efficiency and feasibility.

BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The drawings described herein are used to provide a further understanding of the present invention, form a part of this application, but do not constitute an improper limitation to the present invention. In the drawings:

[0034] Figure 1 It is a schematic diagram of the cervical picture image classification method based on machine learning provided by the present invention.

DETAILED DESCRIPTION

[0035] The present invention will be described in detail below in conjunction with the drawings and specific embodiments, and the schematic embodiments and descriptions are only used to explain the present invention, but do not limit the present invention.

[0036] The present invention proposes a cervical picture image classification method and system based on machine learning, as shown in the attachedFigure 1 As shown, the method includes the following steps:

[0037] Step S1: When collecting cervical pictures, the extraction of the region of interest should be carried out; specifically: the cervical region is located as the region of interest based on the U-Net segmentation model;

[0038] Preferably: The cervical pictures are H&E stained cervical pathological sections, collected through a 40x microscope; when the original pixel density is less than 0.25μm / pixel, it is recommended to use a 20x objective lens; the field of view diameter is 25mm; the scanning speed of 15mm×15mm does not exceed 150s; the scanning resolution is less than or equal to 0.25vm / pixel; the scanned images are required to be accurately focused, and the cell morphology, cell membrane, and internal details of the cell nucleus are relatively intact; the performance of out-of-focus scanned images is manifested as blurred cell structures, double images, etc.;

[0039] Preferably: Image preprocessing is carried out before the extraction of the region of interest; specifically: first, the effectiveness of cervical features is enhanced; then, image data preprocessing is carried out; among them: the enhancement of the effectiveness of cervical features is specifically: first, nuclear segmentation is carried out, and HoVer-Net (for example: the SOTA model for medical image segmentation) is used to accurately separate overlapping cells; then, irrelevant regions are removed, and interference regions such as mucus and blood are removed based on threshold segmentation; the image data preprocessing is specifically: first, image normalization is carried out to unify the resolution, and then color space conversion, histogram equalization, and noise processing are carried out in sequence;

[0040] Preferably: The unified image resolution is such as 512×512 or 1024×1024; RGB to HSV or Lab space is used to enhance the contrast for color space conversion; the CLAHE algorithm is used to optimize the local contrast for histogram equalization; the noise processing is to use median filtering to remove salt-and-pepper noise and non-local means denoising (NLM) to retain details;

[0041] Preferably: After preprocessing, image data augmentation is carried out, including: rotation (±30°), flipping, translation, scaling, elastic deformation (simulating cell deformation), GAN generation, and Mixup;

[0042] Step S2: Based on the region of interest, a first feature representing local color features and a second feature representing structural relationship features are determined;

[0043] The determination of the first feature representing the local color feature is specifically as follows: Set a sliding window to slide within the region of interest, and obtain the local color feature within the sliding window during the sliding process; Represent the local color feature as a histogram of continuous values to construct the first feature; where: The sliding step of the sliding window is smaller than the size of the sliding window; that is, after each slide, it does not completely slide out; Specifically, it includes the following steps:

[0044] Step S2A1: Initialize the sliding window size to W×H; Set the initial position of the sliding window at the upper left corner of the region of interest; Set the sliding step to WX, and WX < W, H < W;

[0045] Preferably: The sliding window is rectangular, and 2H <= W;

[0046] Preferably: WX = (1 / 3 to 1 / 2)×W;

[0047] In alternative method 1: Step S2A1 is specifically: Set the sliding window to include a first window and a second window, and the window size of each window is W×H; There is a preset interval size INT between the first window and the second window; Set the initial position of the sliding window at the upper left corner of the region of interest; Set the sliding step to WX, and INT > W;

[0048] Preferably: W = H;

[0049] Alternatively: 2H = W, and WX = (1 / 2 to 1)×W;

[0050] Step S2A2: Make the sliding window w slide within the region of interest in the preset sliding direction according to the sliding step, and extract the local color feature within the sliding window after each slide; The local color feature includes the average gray scale g w 、average brightness b w and / or saturation s w ; w is the sliding window number; Since the sliding step is smaller than the size of the sliding window, therefore, during the sliding process, some pixel values are not slid out of the window range, resulting in repeated feature extraction;

[0051] Preferably: The preset sliding direction is to slide in the order from left to right and from top to bottom starting from the initial position;

[0052] Since the core feature of cervical images is their color characteristics. For example: normal tissue: the cell nuclei are lightly stained (light blue), the cytoplasm is uniform, and the background is clean; lesion area: inflammation: the cell nuclei are slightly enlarged and the staining is deepened; low-grade lesion (LSIL): the cell nuclei are significantly enlarged, the nuclear-cytoplasmic ratio is increased, and the chromatin is rough; high-grade lesion (HSIL): the nuclei are darkly stained, irregular in shape, and the nuclear membrane is wrinkled; canceration: the cell nuclei are significantly atypic, extremely darkly stained, and necrotic areas may appear. Therefore, discovering and extracting the color characteristics of cervical images can obtain effective prior indications. However, different pictures and different image acquisition methods will have a great differential impact on the color of the images, causing great "troubles" to the artificial intelligence model during the training process and making it impossible to achieve the training effect. Describing the color (not just grayscale) through a histogram can discover the disorderliness of the spatial distribution of color changes in the picture. By the way of sliding the window that does not completely slide out each time, the gradual change characteristics of adjacent areas can be captured, such as the changes and continuity of the arrangement of cervical cells. Further, considering that the more prominent feature of cervical pictures in color characteristics is disorderliness, obtaining prior knowledge based on regional comparison can more significantly discover this color characteristic. Therefore, it can be through replaceable method 1;

[0053] In replaceable method 1: The specific step S2A2 is: making the sliding window slide in the region of interest along the preset sliding direction according to the sliding step length, and extracting the local color characteristics within the sliding window after each slide; the local color characteristics include the calculated value gw of the average grayscale, the calculated value b of the average lightness w and / or the calculated value s of the saturation w ;

[0054] Preferably: The calculated value is the mean, maximum value and / or minimum value of the first window and the second window;

[0055] Step S2A3: Represent the local color characteristics by a histogram of continuous values to construct the first feature; specifically: Based on the following formula (1), each of the local color characteristics is respectively represented by a histogram of continuous values H it ; where: t is the type of local color characteristic; t = g, b, w indicates that its type is average grayscale, average lightness and saturation; it = 0 to IT - 1 is the histogram value range of the histogram element it of type t; t(w) is the local color characteristic value of type t of the sliding window w in the region of interest, δ is the Dirac function, which is 1 when its parameter (t(w) - it) falls within [0, 1), otherwise 0; for the numerical value t(w) of each local color characteristic of the sliding window, this numerical value is a continuous numerical value, and the number of sliding windows whose local color characteristic values in the region of interest belong to the range [it, it + 1) is obtained through this function;

[0056] H it＝0～IT-1 = ∑ w δ(t(w) - it) (1);

[0057] Preferably: for average gray scale, IT = 256; from 0 to IT - 1, the average gray scale is divided into 256 value ranges; for average lightness, IT = 100; from 0 to IT - 1, the lightness is divided into 100 value ranges on average; for saturation, IT = 100; from 0 to IT - 1, the lightness is divided into 100 value ranges;

[0058] Step S2A4: Place the histogram representation H of continuous numerical values it into the t-th row and i-th column of the matrix to form the first feature in matrix form;

[0059] Alternatively, place H ig , H ib , H is sequentially into a vector to form the first feature in vector form;

[0060] The second feature for determining the feature representing the structural relationship is specifically: Set a sliding window to slide within the region of interest, and obtain the structural relationship features within the sliding window during the sliding process; Concatenate the structural relationship features to form the second feature; where: The sliding step size of the sliding window is less than the size of the sliding window;

[0061] The above steps specifically include the following steps:

[0062] Step S2B1: Set the sliding window size to W3 * W3, set the initial position of the sliding window at the upper left corner of the region of interest; Set the sliding step size to WX, and 0.5W3 <= WX <= W3;

[0063] Step S2B2: Make the sliding window slide within the region of interest in the preset sliding direction according to the sliding step size, and extract the structural relationship features between the sliding window and the adjacent window after each slide; The structural relationship features include the contrast, energy, entropy, and / or homogeneity between the sliding window and the adjacent window; For cervical pictures, its slicing method and acquisition method have a great impact on the structural relationship features, which may bring completely wrong information to the classification result; During the sliding process, the 8 - adjacent comparison method can reduce this influence degree, find the comparison situation of the structural features in 8 directions, and discover the unevenness of detecting cell nuclei and capture the regular changes of cell arrangement based on the relatively large granularity of the window; Instead of discovering information for each cell nucleus one by one;

[0064] The extraction of the structural relationship features between the sliding window and the adjacent window specifically includes the following steps:

[0065] Step S2B21: Calculate the grayscale mean of the sliding window w3 and the grayscale mean of each of the adjacent windows k adjacent to the sliding window 8 k = 1 to 8;

[0066] Step S2B22: If then this adjacent window is called a difference window k1; where: trg is the grayscale difference threshold; this threshold is a preset value, for example: set to 50 to 120;

[0067] Step S2B23: Determine the number N of difference windows k1;

[0068] Step S2B24: Calculate the contrast, energy, entropy, and / or homogeneity of the sliding window w3; specifically: calculate the sliding window contrast C based on the following formulas (2)-(5) w3 , energy e w3 , entropy s w3 and / or homogeneity h w3 ; where: is the grayscale mean of the difference window k1; the larger the contrast value, the clearer the structure; the larger the energy, the smaller and more uniform the structural difference; the larger the entropy value, the more chaotic the structure; the larger the homogeneity, the less chaotic the structure;

[0069]

[0070] Preferably: The preset sliding direction starts from the initial position and slides in the order from left to right and then from top to bottom;

[0071] Alternatively: The preset sliding direction starts from the initial position and slides in the order from top to bottom and then from left to right;

[0072] Alternatively: The preset sliding direction is a random slide;

[0073] Step S2B3: Arrange the structural relationship features of the sliding windows in the order of the sliding window numbers to form a second feature; at this time, each element in the second feature is a 4-tuple;

[0074] Alternatively: Arrange the contrast, energy, entropy, and homogeneity of each window in the order of the window numbers to form each row in the second feature; at this time, the second feature is a 4*|w3| matrix; |w3| is the number of sliding windows;

[0075] Alternatively: For each sliding window, calculate the comprehensive value of its structural relationship features, and arrange these comprehensive values in the order of the sliding window numbers to form the second feature; The comprehensive method can be selected as the weighted sum method;

[0076] Replaceable: Represent the structural relationship features as a histogram of continuous values to construct the second feature; the representation method using the histogram is the same as before; here, the contrast, energy, entropy, and homogeneity can all be normalized to the 0-100 space in terms of data value range, and then the histogram elements with 100 value ranges are constructed for them;

[0077] Step S3: Input the first feature and the second feature into the artificial intelligence model to obtain a classification result; specifically: after splicing the first feature and the second feature for input, input them into the artificial intelligence model to obtain a classification result; preferably: the classification result includes normal, inflammation, low-grade lesion, high-grade lesion, canceration, etc.;

[0078] Preferably: the artificial intelligence model is a machine learning model; preprocess the first feature and the second feature respectively, so that the first feature is Z-score standardized, and after standardization, its mean is 0 and the variance is 1; perform Min-Max normalization on the second feature and scale its elements to the [0, 1] interval; splice the preprocessed first feature and the second feature along the feature dimension to generate a fused feature vector, and input the fused feature vector into the machine learning model; the machine learning model is SVM, random forest or XGBoost / LightGBM; since the prior knowledge has been introduced, the artificial intelligence model for subsequent image classification can choose a machine learning model with relatively small computational complexity, without the need for a large amount of training, and the complexity will be significantly reduced;

[0079] Replaceable: Use the first feature and the second feature as the prior knowledge fused feature of the artificial intelligence model; specifically: after fusion, use it as the input weight or output weight of the artificial intelligence model; set the first feature encoder and the second feature encoder, and respectively use the first feature and the second feature as the weight vectors wt1 and wt2; for the newly input cervical image X, perform weighted adjustment, and set X = X ⊙ (αwt1 + βwt2), where: α, β are learnable parameters, and the initial values are set to 0.5 and are adaptively adjusted during training; ⊙ represents element-wise multiplication, which is used to enhance important channels and suppress noise; convert the first feature containing local color information and the second feature containing global texture organizational structure information into the dynamic weight parameters of the model, rather than simply input features, so that the prior knowledge directly affects the calculation process of the model. In this way, the model can dynamically adjust the importance of different features or neurons during the inference process, enhance the sensitivity to key pathological patterns, so that the prior features can be converted into dynamic weights, and the model can not only fuse traditional handcrafted features and deep features, but also embed medical logic in the calculation process;

[0080] Preferably, the artificial intelligence model is a two-branch model. The first branch is used to input the first feature; the second branch is used to input the second feature. Among them: The first branch adopts the traditional feature processing method, enabling the input second feature to pass through 2-3 fully connected layers to gradually compress the feature dimension (such as 128→64→32) to avoid overfitting; using Batch Normalization (BatchNorm) and Dropout (dropout rate about 30%) to enhance the generalization ability; as much as possible retaining the interpretability of local color features through the processing of the first branch, while learning the complementary expression formed with the second feature; The second branch adopts deep feature processing, and the input dimension of the second feature matches the output of the pre-trained model (such as the 1024-dimensional feature of ResNet-50); using an N*N convolutional layer for normalization to adapt to the high-dimensional feature distribution; using the GELU activation function to replace ReLU to enhance the non-linear expression ability; a higher proportion of Dropout (about 50%) can be added to suppress overfitting; for the first feature and the second feature, feature fusion is performed in an adaptive weighted fusion manner, and an attention mechanism is used to dynamically calculate the weights of the outputs of the two branches; when one branch is discriminative in a certain sample (such as a significant abnormality in the nuclear-cytoplasmic ratio), a higher weight (up to 70%) is assigned;

[0081] Alternatively, the artificial intelligence model includes a first model and a second model; the first model is used to input the first feature, and the second model is used to input the second feature; the outputs of the first model and the second model are fused to obtain a classification result;

[0082] Based on the same inventive concept, the present invention also provides a cervical picture image classification system based on machine learning, and the system is used to complete the above-mentioned cervical picture image classification method based on machine learning;

[0083] A computer program (also known as a program, software, software application, script, or code) can be written in any form of programming language, including assembly or interpreted languages, declarative or procedural languages, and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, object, or other unit suitable for use in a computing environment. A computer program may or may not correspond to a file in a file system. The program can be stored as part of a file that holds other programs or data (such as one or more scripts in a markup language document), in a single file dedicated to the program, or in multiple cooperating files (such as files that store one or more modules, subroutines, or code portions). A computer program can be deployed to execute on one computer or on multiple computers located at one site or distributed across multiple sites and interconnected by a communication network.

[0084] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) that contain computer-usable program code.

[0085] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0086] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing devices to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0087] These computer program instructions can also be loaded onto a computer or other programmable data processing devices, such that a series of operation steps are executed on the computer or other programmable devices to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable devices provide steps for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications or equivalent replacements can still be made to the specific embodiments of the present invention. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention shall be covered by the protection scope of the claims of the present invention.

Claims

1. A cervical picture image classification method based on machine learning, characterized in that, The method includes: Step S1: When collecting cervical pictures, the extraction of the region of interest should be performed. Step S2: Based on the region of interest, determine the first feature representing the local color feature and the second feature representing the structural relationship feature. The specific method for determining the first feature representing the local color feature is as follows: Set a sliding window to slide within the region of interest. During the sliding process, obtain the local color feature within the sliding window; Represent the local color feature in the form of a histogram to construct the first feature. Among them, the sliding step size of the sliding window is smaller than the size of the sliding window. The specific method for determining the second feature representing the structural relationship feature is as follows: Set a sliding window to slide within the region of interest. During the sliding process, obtain the structural relationship feature between the sliding window and the adjacent window; After splicing the structural relationship features, form the second feature. The specific method for extracting the structural relationship feature between the sliding window and the adjacent window is as follows: Step S2B21: Calculate the grayscale mean of the sliding window w3 and the grayscale mean of each of the adjacent windows k to the sliding window 8 k = 1 to 8; Step S2B22: If then this adjacent window is called a difference window k1; where: trg is the gray difference threshold; Step S2B23: Determine the number N of the difference windows k1 to determine the occurrence probability of N / 8. Step S2B24: Calculate the contrast, energy, entropy, and homogeneity of the sliding window w3 based on the occurrence probability of N / 8. Step S3: Input the first feature and the second feature into the artificial intelligence model to obtain the classification result.

2. The cervical picture image classification method based on machine learning according to claim 1, characterized in that Use the U-Net-based segmentation model to locate the cervical region as the region of interest.

3. The cervical picture image classification method based on machine learning according to claim 2, characterized in that The cervical pictures are H&E-stained cervical pathological sections, collected through a 40x microscope.

4. The cervical picture image classification method based on machine learning according to claim 3, characterized in that, Perform image preprocessing before extracting the region of interest.

5. The cervical picture image classification method based on machine learning according to claim 4, wherein The artificial intelligence model is a machine learning model.

6. The cervical picture image classification method based on machine learning according to claim 5, wherein The specific method for determining the first feature representing the local color feature is as follows: Step S2A1: Set the sliding window to include a first window and a second window, and the window size of each window is W×H; There is a preset interval size INT between the first window and the second window. Set the initial position of the sliding window at the upper left corner of the region of interest; Set the sliding step size to WX, and INT>W. Step S2A2: Make the sliding window slide within the region of interest in the preset sliding direction according to the sliding step size. After each slide, extract the local color feature within the sliding window. The local color feature includes the calculated value of the average gray scale gw of the first window and the second window, the calculated value of the average brightness b w and / or the calculated value of the saturation s w ; Step S2A3: Represent the local color feature in the form of a histogram to construct the first feature. Specifically: Each of the local color features is respectively represented by a histogram H based on the following formula (1) it ; where: t is the type of the local color feature; t = g, b, w indicates that its type is average gray scale, average lightness, and saturation; it = 0 to IT - 1 is the histogram value range of the histogram element number it of type t; t(w) is the local color feature value of type t of the sliding window w in the region of interest, δ is the Dirac function, which has a value of 1 when its argument (t(w) - it) falls within [0, 1), and 0 otherwise; H it＝0～IT-1 = ∑ w δ(t(w) - it) (1); Step S2A4: Place the histogram representation H it in the t-th row and i-th column of the matrix to form the first feature in matrix form.

7. A cervical picture image classification system based on machine learning, characterized in that, The machine learning-based cervical picture image classification system is used to implement the machine learning-based cervical picture image classification method described in any one of claims 1-6 above.

8. A cervical picture image classification device based on machine learning, characterized in that, The machine learning-based cervical picture image classification device is used to implement the machine learning-based cervical picture image classification method described in any one of claims 1-6 above.

9. A cervical picture image classification control server based on machine learning, characterized in that, The machine learning-based cervical picture image classification server is used to implement the machine learning-based cervical picture image classification method described in any one of claims 1-6 above.

10. A cervical picture image classification chip based on machine learning, characterized in that, The machine learning-based cervical picture image classification chip is used to implement the machine learning-based cervical picture image classification method described in any one of claims 1-6 above.

Images