A method and system for identifying abnormalities in intrinsic fluorescent cervical images

Abstract

The present invention proposes a method and system for identifying abnormal intrinsic fluorescent cervical images, which relate to the technical field of cervical image detection; the method and system comprise the following steps: S1: obtaining a fluorescent grayscale image of a cervical region with lesions collected historically as an original sample; S2: performing boundary fitting and cropping on the original sample; S3: selectively enhancing the contrast of the cropped original sample; S4: using an improved local binary pattern for screening to obtain an image of abnormal cervical surface; S5: using the image of abnormal cervical surface as a sample library, selecting a part of the sample library as a training set and another part as a verification set, using the training set as input for neural network training, and obtaining a trained neural network; S6: obtaining a grayscale image of the current patient's cervix, using the grayscale image of the current patient's cervix as input, sending it into the trained neural network model, and outputting the recognition result of the abnormal area of ​​the cervical image.

Description

Technical Field

[0001] The present invention relates to the technical field of cervical image detection, and in particular to a method and system for identifying abnormalities in intrinsic fluorescent cervical images. Background Art

[0002] Cervical cancer, referred to as cervical cancer, is a malignant tumor that occurs in the female cervix. It ranks second among the most common malignant tumors in women, second only to breast cancer. There are more than 500,000 new confirmed cases worldwide each year, of which more than 140,000 women in my country are diagnosed with cervical cancer each year. With the change of lifestyle, the age of patients is getting younger. Cervical intraepithelial lesions are precancerous lesions, including low-grade squamous intraepithelial lesions LSIL and high-grade squamous intraepithelial lesions HSIL. LSIL covers mild atypical hyperplasia, CINI, flat warts, cervicitis, etc.; HSIL includes moderate to severe atypical hyperplasia, CINII and CINIII, etc. Reliable colposcopy is an important means of early detection of cervical cancer screening, which is conducive to identifying patients with lesions and maintaining family and social stability.

[0003] Currently, commonly used inspection methods include naked eye observation, glacial acetic acid coating observation, conventional colposcopy combined with acetic acid and Lugol's iodine test, Pap smear, liquid-based cytology, etc. However, all of the above methods have a high false negative rate, poor sensitivity, and require a long observation time. They require contact coating of the cervix, which may cause discomfort to the patient, and other pathogens may be introduced during the contact process. None of the above methods can accurately determine the nature of the lesion, and often require a combined cervical biopsy pathological examination for diagnosis. The biopsy often takes more than a week, which is painful for the patient and brings great physical and mental stress to the patient. It also requires a high level of experience for the doctor.

[0004] Therefore, it is very necessary to provide a method and system for identifying abnormalities in intrinsic fluorescent cervical images, which does not require the application of reagents on the cervix and does not require multiple lenses to perform separate imaging. Summary of the invention

[0005] In view of this, the present invention proposes a method and system for identifying abnormalities in intrinsic fluorescent cervical images that does not require the application of reagents and does not require continuous switching of multiple lenses for framing.

[0006] The technical solution of the present invention is achieved in this way:

[0007] In one aspect, the present invention provides a method for identifying abnormalities in intrinsic fluorescent cervical images, comprising the following steps:

[0008] S1: Obtain the fluorescent grayscale image of the lesion in the cervical area collected historically as the original sample;

[0009] S2: perform boundary fitting and cropping on the original sample to obtain the cropped original sample;

[0010] S3: selectively enhance the contrast of the cropped original sample;

[0011] S4: Screening the cropped original sample processed in step S3 using an improved local binary pattern to obtain an image of abnormal cervical surface;

[0012] S5: The images of abnormal cervical surface are used as a sample library, a part of the sample library is selected as a training set, and the other part is selected as a validation set, and the training set is used as input for neural network training to obtain a trained neural network;

[0013] S6: Configure an ultraviolet light source and a camera, use a dilator to expose the cervix, and after the ultraviolet light source irradiates the cervical area for a period of time, obtain a grayscale image of the current patient's cervix through the camera, and use the grayscale image of the current patient's cervix as input to the trained neural network model to output the recognition result of the abnormal area of ​​the cervical image.

[0014] On the basis of the above technical scheme, preferably, the specific content of step S2 is to obtain the outer boundary of the cervix by curve fitting, remove the continuous white area outside the outer boundary of the cervix, or the continuous white area adjacent to the outer boundary of the cervix, retain the image within the outer boundary of the cervix, and obtain the cropped original sample.

[0015] Preferably, the continuous white area adjacent to the outer boundary of the cervix means that the grayscale value of the current pixel is 255, and at least three pixels among eight adjacent pixels have the same grayscale extreme value of 255, and the continuous white area has a smooth boundary.

[0016] Preferably, the selective enhancement of the contrast of the cropped original sample described in step S3 is specifically to obtain a histogram of the grayscale values ​​of all pixels of the cropped original sample: if the histogram is distributed in the grayscale value range [a, b], and the difference between the grayscale peak value and the number of pixels of the grayscale valuation does not exceed 10% of the cropped original sample, then the overall contrast of the cropped original sample is improved; if the histogram is distributed in the grayscale value range [a, b], and the difference between the grayscale peak value and the number of pixels of the grayscale valuation exceeds 10% of the cropped original sample, then No processing is performed on the cropped original samples; if the histogram is normally distributed within the grayscale value range [a, b] and the center value of the normal distribution deviates from the grayscale value 128 by no more than 32 grayscale levels, no contrast enhancement is performed on the cropped original samples; if the histogram is normally distributed within the grayscale value range [a, b] and the center value of the normal distribution deviates from the grayscale value 128 by more than 32 grayscale levels, the contrast of the cropped original samples is enhanced by grayscale logarithmic transformation; the lower limit a of the grayscale value range [a, b] is not less than 40, and the upper limit b is not more than 225.

[0017] Preferably, the overall contrast enhancement of the cropped original sample is performed by increasing the contrast of each pixel of the cropped original sample by 130%-140%; the contrast enhancement of the cropped original sample is performed by using the grayscale logarithmic transformation method, which is performed by using the formula Where x is the grayscale value of each pixel of the cropped original sample, y≤255 is the grayscale value of each pixel of the cropped original sample after grayscale logarithmic transformation, c is a constant, when the center value of the normal distribution of the grayscale values ​​of all pixels of the cropped original sample is in the interval [0, 93], c∈(0, 0.5], when the center value of the normal distribution of the grayscale values ​​of all pixels of the cropped original sample is in the interval (160, 255], c∈(0.5, 1].

[0018] More preferably, the specific content of step S4 is:

[0019] 1) The cropped original sample processed in step S3 is partitioned into 3×3 pixels. If the row or column is not an integer multiple of 3, it is padded to an integer multiple of 3 in the order of row first and then column. The pixel values ​​filled in the padded row or column are all 0;

[0020] 2) Compare the pixel values ​​within the partitions and set the pixel value of the pixel at the center of each partition to x c , the pixel value of the pixel at the non-center of the partition is x i ; Starting from the upper left corner of each partition, the pixel value x of the pixel at the non-center of the partition is incremented clockwise i The pixel value x of the pixel at the center c Perform comparison and generate comparison results The reference value x imax is the maximum pixel value of the pixels in the non-center of the partition, x imin is the minimum pixel value of pixels not located at the center of the partition, For rounding operation, the comparison result is overwritten on the pixel value, and the pixel value in each partition is represented by 0 and 1;

[0021] 3) retaining adjacent regions of pixels with the same comparison result of 1, and removing isolated pixels with the comparison result of 1;

[0022] 4) performing curve fitting on the boundary of the region of pixels whose comparison result is 1, to obtain a closed region to be confirmed; the closed region to be confirmed includes one or a combination of a circle, an ellipse, and a rounded rectangle;

[0023] 5) Manually screen the closed area to be confirmed and retain images of abnormal cervical surface.

[0024] Preferably, the pixel value x of the pixel at the center of each partition is c After selectively enhancing the contrast of the cropped original sample in step S3, the grayscale values ​​of all pixels of the cropped original sample are magnified by a certain multiple according to the median arranged in ascending order; after the pixel values ​​of the pixels at the non-center of the partition are overwritten, the pixel values ​​x of the pixels at the center of the partition are further compared. c The average value of the pixel values ​​of the pixels at the non-center of the partition. If the pixel value x at the center of the partition is c The pixel value x of the pixel at the center of the partition is greater than the average value of the pixel values ​​at the non-center of the partition. c Overwritten to 1; if the pixel value x at the center of the partition c The pixel value x of the pixel at the center of the partition is not greater than the average value of the pixel values ​​at the non-center of the non-partition. c Overwritten to 0.

[0025] Preferably, the neural network training described in step S5 is to perform the following steps: 1) configure the number of nodes in the input layer, hidden layer and output layer, and initialize the weights and biases; 2) use the given initialization weights and biases to execute the forward propagation algorithm, calculate the linear combination value of the input and weights for each layer, and use the activation function to apply to the linear combination value; specify the loss function and cost function of the training neural network model; 3) use the ReLU activation function for non-last layers; for the last layer, use the softmax activation function to calculate the probability values ​​of the results belonging to different categories, combine the probabilities to give the prediction results, and calculate the results of the loss function; 4) use gradient descent to minimize the cost function, continuously update the weights and biases, find the optimal weights and bias values, and obtain the trained neural network.

[0026] Still further preferably, the number of nodes in the output layer does not exceed 10, and the number of training iterations does not exceed 100.

[0027] On the other hand, the present invention also provides an intrinsic fluorescent cervical image abnormality recognition system, which is used to implement the above intrinsic fluorescent cervical image abnormality recognition method, comprising:

[0028] An original sample acquisition unit, used to acquire a fluorescent grayscale image of a lesion in the cervical region collected historically as an original sample;

[0029] A first preprocessing unit is used to perform boundary fitting and cropping on the original sample to obtain a cropped original sample;

[0030] A second preprocessing unit, used for selectively enhancing the contrast of the original sample cropped by the first preprocessing unit;

[0031] A feature extraction unit is used to screen the cropped original sample processed by the second preprocessing unit using an improved local binary pattern to obtain an image of abnormal cervical surface;

[0032] The model training unit uses abnormal cervical surface images as a sample library, selects a part of the sample library as a training set and another part as a validation set, uses the training set as input to train the neural network, and obtains a trained neural network;

[0033] The sampling output unit is used to obtain the grayscale image of the current patient's cervix, take the grayscale image of the current patient's cervix as input, send it into the trained neural network model, and output the recognition result of the abnormal area of ​​the cervical image.

[0034] The method and system for identifying abnormal intrinsic fluorescence cervical images provided by the present invention have the following beneficial effects compared with the prior art:

[0035] The present application configures an ultraviolet light source and a camera, and only obtains grayscale images of lesions in the cervical area. After preprocessing and identifying features of historical images, neural network training is performed to obtain the optimal neural network parameters. It is then convenient to perform feature recognition and output on newly acquired fluorescent images. The application can shorten the detection time, eliminate the discomfort caused by applying reagents, and does not rely on the doctor's experience when using it. It can well overcome the shortcomings of the existing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0037] Figure 1 The present invention is a flowchart of the steps of a method and system for identifying abnormalities in intrinsic fluorescent cervical images. DETAILED DESCRIPTION

[0038] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] like Figure 1 As shown, the present invention provides a method for identifying abnormalities in intrinsic fluorescent cervical images, comprising the following steps:

[0040] S1: Obtain a historically acquired fluorescent grayscale image of a lesion in the cervical region as an original sample.

[0041] This step is mainly to obtain cervical images corresponding to high-grade squamous intraepithelial lesions LSIL and high-grade squamous intraepithelial lesions HSIL confirmed by colposcopy. Each original sample has at least one abnormal area of ​​the cervix, which can be LSIL such as inflammation, or epithelial tissue with raised or depressed areas, with obvious or unclear boundaries.

[0042] S2: Perform boundary fitting and cropping on the original sample to obtain the cropped original sample.

[0043] The specific content of step S2 is to obtain the outer boundary of the cervix by curve fitting, remove the continuous white area outside the outer boundary of the cervix, or the continuous white area adjacent to the outer boundary of the cervix, retain the image within the outer boundary of the cervix, and obtain the cropped original sample. The continuous white area affects the overall pixel distribution of the fluorescent grayscale image, and the white area is highlighted so that the corresponding cervical details cannot be displayed, so it needs to be removed.

[0044] The continuous white area adjacent to the outer boundary of the cervix means that the gray value of the current pixel is 255, and at least three pixels among the eight adjacent pixels have the same gray extreme value of 255, and the continuous white area has a smooth boundary. For the sake of insurance, the boundary of the continuous white area can be offset outward by a certain distance as the actual elimination range, and the offset distance can be an integer multiple of the pixel width of the fluorescent gray image.

[0045] S3: Selectively enhance the contrast of the cropped original sample.

[0046] Specifically, the histogram of the grayscale values ​​of all pixels of the cropped original sample is obtained and analyzed:

[0047] If the histogram is distributed in the grayscale value range [a, b], and the difference between the grayscale peak value and the number of pixels of the grayscale valuation does not exceed 10% of the original sample after cropping, the contrast of the original sample after cropping is improved overall; this situation indicates that the non-extreme pixel value distribution is relatively balanced, the peak is not significant, and the detail features of the image may not be prominent, so the contrast of each pixel is adjusted by overall contrast improvement, so that the details of the image are easy to identify. If the histogram is distributed in the grayscale value range [a, b], and the difference between the grayscale peak value and the number of pixels of the grayscale valuation exceeds 10% of the original sample after cropping, the surface histogram has at least one obvious peak, and the original sample after cropping is not improved overall. Among them, the overall contrast improvement of the original sample after cropping is to improve the contrast of each pixel of the original sample after cropping by 130%-140%, preferably 130%. The lower limit a of the grayscale value range [a, b] is not less than 40, and the upper limit b does not exceed 225.

[0048] If the histogram is normally distributed within the grayscale value range of [a, b] and the center value of the normal distribution deviates from the grayscale value 128 by no more than 32 grayscale values, the contrast enhancement processing is not performed on the cropped original sample.

[0049] If the histogram is normally distributed within the grayscale value range [a, b] and the center value of the normal distribution deviates from the grayscale value 128 by more than 32 grayscale levels, the grayscale logarithmic transformation is used to enhance the contrast of the cropped original sample. The grayscale logarithmic transformation is used to enhance the contrast of the cropped original sample using the formula Where x is the grayscale value of each pixel of the cropped original sample, y≤255 is the grayscale value of each pixel of the cropped original sample using grayscale logarithmic transformation, c is a constant, when the center value of the normal distribution of the grayscale values ​​of all pixels of the cropped original sample is in the interval [0, 96), c∈(0, 0.5], when the center value of the normal distribution of the grayscale values ​​of all pixels of the cropped original sample is in the interval (160, 255], c∈(0.5, 1]. The numerator of this function is a logarithmic function, and the logarithmic function can be expanded The function increases the contrast of pixels in the grayscale value range of [0,96) and appropriately reduces the contrast of brighter areas to make the boundaries of image features clearly displayed. In the entire grayscale range of 0-255, the distribution of pixels is more balanced, rather than concentrated in one side.

[0050] S4: Screening the cropped original sample processed in step S3 using an improved local binary pattern to obtain an image of abnormal cervical surface;

[0051] More preferably, the specific content of step S4 is:

[0052] 1) The cropped original sample processed in step S3 is partitioned into 3×3 pixels. If the row or column is not an integer multiple of 3, it is supplemented to an integer multiple of 3 in the order of row first and column later. The pixel values ​​filled in the supplemented row or column are all 0. For example, if the size of the original sample is 640×480, 640 is not an integer multiple of 3, so it needs to be expanded until it reaches 642 rows, adding 2 rows. The way to add rows is to add a row outside the current first row and the last row respectively. If the number of columns is not a multiple of 3, the columns are expanded and filled with content in the order of left and right first and then back.

[0053] 2) Compare the pixel values ​​within the partitions and set the pixel value of the pixel at the center of each partition to x c , the pixel value of the pixel at the non-center of the partition is x i ; Starting from the upper left corner of each partition, the pixel value x of the pixel at the non-center of the partition is incremented clockwise i The pixel value x of the pixel at the center cPerform comparison and generate comparison results The reference value x imax is the maximum pixel value of the pixels in the non-center of the partition, x imin is the minimum pixel value of pixels not located at the center of the partition, For rounding operation, the comparison result is overwritten on the pixel value, and the pixel value in each partition is represented by 0 and 1. The reference value represents the degree of deviation between the pixel value of the pixel at the center and the pixel value of the pixel at the center in the same partition. For different partitions, the size of the reference value is not exactly the same. This binary overwriting of the image is conducive to quickly screening the brighter pixel area.

[0054] As a preferred implementation, in order to improve the accuracy of identifying abnormal features at each partition, the pixel value x of the pixel at the center of each partition is c In addition to using the pixel value itself multiplied by a fixed multiple, that is, according to step S3, the pixel value after selectively enhancing the contrast of the original sample after cropping is magnified by a certain multiple as the threshold; a uniform pixel value x c , the uniform pixel value x c After selectively enhancing the contrast of the cropped original sample in step S3, the median of the grayscale values ​​of all pixels of the cropped original sample arranged in ascending order is magnified by a certain multiple to be used as the threshold, and the magnification by a certain multiple can be in the range of 1.0-1.2. Regardless of the method used to obtain the pixel value x c After the pixel values ​​of the pixels at the non-center of each partition are overwritten, the pixel value x of the pixel at the center of the partition is further compared. c The average value of the pixel values ​​of the pixels at the non-center of the partition. If the pixel value x at the center of the partition is c The pixel value x of the pixel at the center of the partition is greater than the average value of the pixel values ​​at the non-center of the partition. c Overwritten to 1; if the pixel value x at the center of the partition c The pixel value x of the pixel at the center of the partition is not greater than the average value of the pixel values ​​at the non-center of the non-partition. c Overwritten to 0. In this embodiment, it is preferred to use a uniform pixel value x c Perform pixel value comparison and overwrite.

[0055] 3) The adjacent areas with the same comparison result of 1 are retained, and the isolated pixels with the comparison result of 1 are eliminated. The isolated pixels with the comparison result of 1 may be noise points, which have a small range and no significant reference value, so they can be discarded.

[0056] 4) Curve fitting is performed on the boundary of the region of pixels whose comparison result is 1 to be retained, so as to obtain a closed region to be confirmed; the closed region to be confirmed includes one or a combination of a circle, an ellipse, and a rounded rectangle.

[0057] 5) Manual screening is performed on the closed area to be confirmed, and images with abnormal cervical surface are retained. Manual screening is used to classify and mark images with abnormal cervical surface.

[0058] S5: The images of abnormal cervical surface are used as a sample library, a part of the sample library is selected as a training set, and the other part is selected as a verification set. The training set is used as input for neural network training to obtain a trained neural network.

[0059] The content of the neural network training mentioned in step S5 is to perform the following steps: 1) configure the number of nodes in the input layer, hidden layer and output layer, and initialize the weights and biases; 2) use the given initialization weights and biases to execute the forward propagation algorithm, calculate the linear combination value of the input and weights for each layer, and use the activation function to apply to the linear combination value; specify the loss function and cost function of the training neural network model; 3) use the ReLU activation function for non-last layers; for the last layer, use the softmax activation function to calculate the probability values ​​of the results belonging to different categories, combine the probabilities to give the prediction results, and calculate the results of the loss function; 4) use gradient descent to minimize the cost function, continuously update the weights and biases, find the optimal weights and biases, and obtain the trained neural network.

[0060] As a preferred method of this embodiment, the number of nodes in the output layer does not exceed 10, and the number of training iterations does not exceed 100.

[0061] S6: Configure an ultraviolet light source and a camera, use a dilator to expose the cervix, and after the ultraviolet light source irradiates the cervix area for a period of time, the camera obtains a grayscale image of the current patient's cervix, and uses the grayscale image of the current patient's cervix as input to the trained neural network model to output the recognition result of the abnormal area of ​​the cervix image. The ultraviolet light source outputs long-wave ultraviolet rays with a wavelength of 340nm-400nm.

[0062] On the other hand, the present invention also provides an intrinsic fluorescent cervical image abnormality recognition system, which is used to implement the above intrinsic fluorescent cervical image abnormality recognition method, comprising:

[0063] An original sample acquisition unit, used to acquire a fluorescent grayscale image of a lesion in the cervical region collected historically as an original sample;

[0064] A first preprocessing unit is used to perform boundary fitting and cropping on the original sample to obtain a cropped original sample;

[0065] A second preprocessing unit, used for selectively enhancing the contrast of the original sample cropped by the first preprocessing unit;

[0066] A feature extraction unit is used to screen the cropped original sample processed by the second preprocessing unit using an improved local binary pattern to obtain an image of abnormal cervical surface;

[0067] The model training unit uses abnormal cervical surface images as a sample library, selects a part of the sample library as a training set and another part as a validation set, uses the training set as input to train the neural network, and obtains a trained neural network;

[0068] The sampling output unit is used to obtain the grayscale image of the current patient's cervix, take the grayscale image of the current patient's cervix as input, send it into the trained neural network model, and output the recognition result of the abnormal area of ​​the cervical image.

[0069] The above-mentioned original sample acquisition unit, first preprocessing unit, second preprocessing unit, feature extraction unit and model training unit can all be pre-set and configured on the computer in sequence. The output end of the sampling output unit is connected to the computer for communication, and is used to input the acquired grayscale image of the current patient's cervix to the computer, and the model output by the model training unit on the computer is used for recognition and output.

[0070] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying abnormalities in intrinsic fluorescent cervical images, characterized in that: The steps include: S1: Obtain the fluorescent grayscale image of the lesion in the cervical area collected historically as the original sample; S2: perform boundary fitting and cropping on the original sample to obtain the cropped original sample; S3: selectively enhance the contrast of the cropped original sample; S4: Screening the cropped original sample processed in step S3 using an improved local binary pattern to obtain an image of abnormal cervical surface; S5: The images of abnormal cervical surface are used as a sample library, a part of the sample library is selected as a training set, and the other part is selected as a validation set, and the training set is used as input for neural network training to obtain a trained neural network; S6: configuring an ultraviolet light source and a camera, using a dilator to expose the cervix, and after the ultraviolet light source irradiates the cervix area for a period of time, obtaining a grayscale image of the current patient's cervix through the camera, and sending the grayscale image of the current patient's cervix as input to the trained neural network model, and outputting the recognition result of the abnormal area of ​​the cervix image; The selective enhancement of the contrast of the cropped original sample described in step S3 is specifically to obtain a histogram of the grayscale values ​​of all pixels of the cropped original sample: if the histogram is distributed in the grayscale value range [a, b], and the difference between the number of pixels of the grayscale peak value and the grayscale valuation does not exceed 10% of the cropped original sample, then the overall contrast of the cropped original sample is improved; If the histogram is distributed in the grayscale value range [a, b], and the difference between the grayscale peak value and the grayscale estimated pixel number exceeds 10% of the original sample after clipping, the original sample after clipping will not be processed; If the histogram is normally distributed within the grayscale value range [a, b] and the center value of the normal distribution deviates from the grayscale value 128 by no more than 32 grayscale values, the contrast enhancement processing is not performed on the cropped original sample; If the histogram is normally distributed within the grayscale value range of [a, b] and the center value of the normal distribution deviates from the grayscale value 128 by more than 32 grayscale levels, the contrast of the cropped original sample is enhanced by grayscale logarithmic transformation; The lower limit a of the gray value range [a, b] is not less than 40, and the upper limit b is not more than 225; The overall contrast enhancement of the cropped original sample is to increase the contrast of each pixel of the cropped original sample by 130%-140%; the contrast enhancement of the cropped original sample is performed by grayscale logarithmic transformation, which is to use the formula d>1, where x is the grayscale value of each pixel of the cropped original sample, y≤255 is the grayscale value of each pixel of the cropped original sample after grayscale logarithmic transformation, c is a constant, when the center value of the normal distribution of the grayscale values ​​of all pixels of the cropped original sample is in the interval [0, 96], c∈(0, 0.5], when the center value of the normal distribution of the grayscale values ​​of all pixels of the cropped original sample is in the interval (160, 255], c∈(0.5, 1]; The specific content of step S4 is: 1) The cropped original sample processed in step S3 is partitioned into 3×3 pixels. If the row or column is not an integer multiple of 3, it is padded to an integer multiple of 3 in the order of row first and then column. The pixel values ​​filled in the padded row or column are all 0; 2) Compare the pixel values ​​within the partitions and set the pixel value of the pixel at the center of each partition to x c , the pixel value of the pixel at the non-center of the partition is x i ; Starting from the upper left corner of each partition, the pixel value x of the pixel at the non-center of the partition is incremented clockwise i The pixel value x of the pixel at the center c Perform comparison and generate comparison results The reference value x i max is the maximum pixel value of the pixels in the non-center of the partition, x imin is the minimum pixel value of pixels not located at the center of the partition, For rounding operation, the comparison result is overwritten on the pixel value, and the pixel value in each partition is represented by 0 and 1; 3) retaining adjacent regions of pixels with the same comparison result of 1, and removing isolated pixels with the comparison result of 1; 4) performing curve fitting on the boundary of the region of pixels whose comparison result is 1, to obtain a closed region to be confirmed; the closed region to be confirmed includes one or a combination of a circle, an ellipse, and a rounded rectangle; 5) Manually screen the closed area to be confirmed and retain images of abnormal cervical surface.

2. The method for identifying abnormalities in intrinsic fluorescence cervical images according to claim 1, characterized in that: The specific content of step S2 is to obtain the outer boundary of the cervix by curve fitting, remove the continuous white area outside the outer boundary of the cervix, or the continuous white area adjacent to the outer boundary of the cervix, retain the image within the outer boundary of the cervix, and obtain the cropped original sample.

3. The method for identifying abnormalities in intrinsic fluorescence cervical images according to claim 2, characterized in that: The continuous white area adjacent to the outer boundary of the cervix means that the gray value of the current pixel is 255, and at least three pixels among eight adjacent pixels have the same gray extreme value of 255, and the continuous white area has a smooth boundary.

4. The method for identifying abnormal intrinsic fluorescence cervical images according to claim 1, characterized in that: The pixel value x of the pixel at the center of each partition c After selectively enhancing the contrast of the cropped original sample in step S3, the grayscale values ​​of all pixels of the cropped original sample are magnified by a certain multiple according to the median arranged in ascending order; after the pixel values ​​of the pixels at the non-center of the partition are overwritten, the pixel values ​​x of the pixels at the center of the partition are further compared. c The average value of the pixel values ​​of the pixels at the non-center of the partition. If the pixel value x at the center of the partition is c The pixel value x of the pixel at the center of the partition is greater than the average value of the pixel values ​​at the non-center of the partition. c Overwrite to 1; If the pixel value x at the center of the partition c The pixel value x of the pixel at the center of the partition is not greater than the average value of the pixel values ​​at the non-center of the non-partition. c Overwritten to 0.

5. The method for identifying abnormalities in intrinsic fluorescence cervical images according to claim 1, characterized in that: The neural network training described in step S5 is to perform the following steps: 1) configure the number of nodes in the input layer, hidden layer and output layer, and initialize the weights and biases; 2) use the given initialization weights and biases to execute the forward propagation algorithm, calculate the linear combination value of the input and weight for each layer, and use the activation function to apply to the linear combination value; specify the loss function and cost function of the training neural network model; 3) use the ReLU activation function for non-last layers; for the last layer, use the softmax activation function to calculate the probability value of the result belonging to different categories, combine the probability to give the prediction result, and calculate the result of the loss function; 4) Use gradient descent to minimize the cost function, continuously update the weights and biases, find the optimal values ​​of weights and biases, and obtain the trained neural network.

6. The method for identifying abnormalities in intrinsic fluorescence cervical images according to claim 5, characterized in that: The number of nodes in the output layer does not exceed 10, and the number of training iterations does not exceed 100.

7. An intrinsic fluorescence cervical image abnormality recognition system, used to implement the intrinsic fluorescence cervical image abnormality recognition method according to any one of claims 1 to 6, characterized in that: include: An original sample acquisition unit, used to acquire a fluorescent grayscale image of a lesion in the cervical region collected historically as an original sample; A first preprocessing unit is used to perform boundary fitting and cropping on the original sample to obtain a cropped original sample; A second preprocessing unit, used for selectively enhancing the contrast of the original sample cropped by the first preprocessing unit; Specifically, the grayscale value histogram of all pixels of the cropped original sample is obtained: if the histogram is distributed in the grayscale value range [a, b], and the difference between the grayscale peak value and the number of pixels of the grayscale valuation does not exceed 10% of the cropped original sample, the overall contrast of the cropped original sample is improved; If the histogram is distributed in the grayscale value range [a, b], and the difference between the grayscale peak value and the grayscale estimated pixel number exceeds 10% of the original sample after clipping, the original sample after clipping will not be processed; If the histogram is normally distributed within the grayscale value range [a, b] and the center value of the normal distribution deviates from the grayscale value 128 by no more than 32 grayscale values, the contrast enhancement processing is not performed on the cropped original sample; If the histogram is normally distributed within the grayscale value range of [a, b] and the center value of the normal distribution deviates from the grayscale value 128 by more than 32 grayscale levels, the contrast of the cropped original sample is enhanced by grayscale logarithmic transformation; The lower limit a of the gray value range [a, b] is not less than 40, and the upper limit b is not more than 225; The overall contrast enhancement of the cropped original sample is to increase the contrast of each pixel of the cropped original sample by 130%-140%; the contrast enhancement of the cropped original sample is performed by grayscale logarithmic transformation, which is to use the formula d>1, where x is the grayscale value of each pixel of the cropped original sample, y≤255 is the grayscale value of each pixel of the cropped original sample after grayscale logarithmic transformation, c is a constant, when the center value of the normal distribution of the grayscale values ​​of all pixels of the cropped original sample is in the interval [0, 96], c∈(0, 0.5], when the center value of the normal distribution of the grayscale values ​​of all pixels of the cropped original sample is in the interval (160, 255], c∈(0.5, 1]; The feature extraction unit is used to screen the cropped original samples processed by the second preprocessing unit using an improved local binary pattern to obtain an image of abnormal cervical surface; the specific contents are: 1) The cropped original sample is partitioned into 3×3 pixels. If the row or column is not an integer multiple of 3, it is padded to an integer multiple of 3 in the order of row first and then column. The pixel values ​​filled in the padded row or column are all 0; 2) Compare the pixel values ​​within the partitions and set the pixel value of the pixel at the center of each partition to x c , the pixel value of the pixel at the non-center of the partition is x i ; Starting from the upper left corner of each partition, the pixel value x of the pixel at the non-center of the partition is incremented clockwise i The pixel value x of the pixel at the center c Compare and generate comparison results The reference value x imax is the maximum pixel value of the pixels in the non-center of the partition, x imin is the minimum pixel value of pixels not located at the center of the partition, For rounding operation, the comparison result is overwritten on the pixel value, and the pixel value in each partition is represented by 0 and 1; 3) retaining adjacent regions of pixels with the same comparison result of 1, and removing isolated pixels with the comparison result of 1; 4) performing curve fitting on the boundary of the region of pixels whose comparison result is 1, to obtain a closed region to be confirmed; the closed region to be confirmed includes one or a combination of a circle, an ellipse, and a rounded rectangle; 5) Manually screen the closed area to be confirmed and retain images of abnormal cervical surface; The model training unit uses abnormal cervical surface images as a sample library, selects a part of the sample library as a training set and another part as a validation set, uses the training set as input to train the neural network, and obtains a trained neural network; The sampling output unit is used to obtain the grayscale image of the current patient's cervix, take the grayscale image of the current patient's cervix as input, send it into the trained neural network model, and output the recognition result of the abnormal area of ​​the cervical image.

Images