A training method for early screening model of non-alcoholic fatty liver disease

Abstract

The present invention relates to the technical field of liver image processing, and in particular to a method for training an early screening model for non-alcoholic fatty liver disease. The method effectively screens out fat infiltration areas and obtains the significance of the fat layer through the grayscale value of the pixel area and the regional shape characteristics. Processing is performed in the grayscale histogram, and the probability of each peak as a fibrosis peak is analyzed except for the highest peak. Finally, the fibrosis peak is judged and the fibrosis significance is obtained. Automatic labeling of liver CT images is achieved through these two features. The early screening model for non-alcoholic fatty liver disease can be trained with the automatically labeled liver CT images. The embodiment of the present invention identifies features in liver CT images, and facilitates the training of the early screening model for non-alcoholic fatty liver disease through automatic labeling, thereby ensuring the efficiency of training and the accuracy of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of liver image processing, and in particular to a method for training an early screening model for non-alcoholic fatty liver disease. Background Art

[0002] Non-alcoholic fatty liver disease (NAFLD) is a liver disease associated with metabolic disorders (such as obesity and insulin resistance), characterized by fat accumulation in hepatocytes (steatosis) in the absence of a history of excessive alcohol consumption. NAFLD progresses through steatosis, inflammation, and hepatocellular damage, potentially progressing to liver fibrosis, cirrhosis, or liver cancer. Therefore, automated screening for NAFLD at an early stage allows for effective intervention.

[0003] Existing technologies enable automated screening and analysis using liver CT image analysis models. These models are typically machine learning models, trained by inputting a large number of labeled images to learn the image features. However, the actual training process often involves a large number of training images, and manual labeling of these images can result in errors and omissions, which in turn affects the training efficiency and accuracy of the model. Summary of the Invention

[0004] In order to solve the technical problem of poor training efficiency of the early screening model for non-alcoholic fatty liver disease in the prior art, the present invention aims to provide a method for training an early screening model for non-alcoholic fatty liver disease. The technical solution adopted is as follows:

[0005] The present invention proposes a method for training a model for early screening of non-alcoholic fatty liver disease, the method comprising:

[0006] Obtain liver CT images of early non-alcoholic fatty liver disease;

[0007] Obtaining each pixel region in the liver CT image; obtaining a fat infiltration probability for each pixel region based on the grayscale value and regional shape characteristics of each pixel region; screening out the fat infiltration region based on the fat infiltration probability and obtaining the fat layer prominence of the liver CT image;

[0008] In liver CT images, a grayscale histogram of the liver region excluding the fatty infiltration region is obtained. For each peak in the grayscale histogram except the highest peak, the fibrosis peak is determined based on the peak similarity and grayscale value distance of each peak relative to the highest peak. The fibrosis significance is determined based on the size of the pixel area corresponding to the fibrosis peak.

[0009] Liver CT images are labeled according to the prominence of the fat layer and fibrosis; the labeled liver CT images are used to train an early screening model for non-alcoholic fatty liver disease.

[0010] Furthermore, the method for obtaining the pixel area includes:

[0011] All pixels are clustered according to the grayscale value of each pixel in the liver CT image, and the clustering area corresponding to each cluster is the pixel area.

[0012] Furthermore, the method for acquiring the regional shape feature includes:

[0013] Obtain the minimum circumscribed circle area of ​​the pixel area; obtain the maximum distance and the minimum distance between the boundary pixel points of the pixel area, and use the ratio of the minimum distance to the maximum distance as the initial near-circularity of the pixel area; multiply the ratio of the area of ​​the pixel area to the minimum circumscribed circle area by the initial near-circularity to obtain the near-circularity; and use the ratio of the near-circularity to the normalized pixel area area as the area shape feature.

[0014] Furthermore, the method for obtaining the fat infiltration probability includes:

[0015] A grayscale value difference between an average grayscale value of a pixel region and an average grayscale value of the entire liver CT image is obtained, and a product of the grayscale value difference and the regional shape feature is normalized to obtain the fat infiltration probability.

[0016] Furthermore, the method for screening the fat infiltration area includes:

[0017] If the fat infiltration probability is greater than a preset first threshold, the corresponding pixel area is regarded as a fat infiltration area.

[0018] Furthermore, the method for obtaining the prominence of the fat layer includes:

[0019] The area ratio of the fatty infiltration area to the entire liver CT image was used as the fat layer prominence.

[0020] Furthermore, the method for determining the fibrosis peak includes:

[0021] Obtaining the abscissa difference and ordinate difference of the peak points between each peak and the highest peak; and using the ratio of the abscissa difference to the ordinate difference as a judgment indicator;

[0022] If the judgment index of other peaks except the highest peak in the grayscale histogram is greater than the preset second threshold, the peak with the largest judgment index is selected as the fibrosis peak;

[0023] If there is no other peak in the grayscale histogram whose judgment index is greater than the preset second threshold, it is determined that there is no fibrosis in the liver.

[0024] Furthermore, the method for obtaining the fibrosis significance includes:

[0025] If a fibrosis peak exists, for a pixel region, a negative correlation mapping is performed between the difference between the average grayscale value of the pixel region and the abscissa of the peak point of the fibrosis peak to obtain a fibrosis peak correlation; a negative correlation mapping is performed between the difference between the average grayscale value of the pixel region and the abscissa of the peak point of the highest peak to obtain a normal peak correlation; if the fibrosis peak correlation is greater than the normal peak correlation, the pixel region is determined to be a fibrosis region; the product of the area ratio of all fibrosis regions to the total area of ​​the liver CT image and the fat layer prominence is used as the fibrosis prominence;

[0026] If there is no fibrosis peak, the fibrosis significance is judged to be 0.

[0027] Furthermore, the marking of the liver CT image according to the fat layer significance and the fibrosis significance includes:

[0028] If the fat layer significance is less than a preset third threshold, it is marked as low risk;

[0029] If the fat layer significance is greater than or equal to the third threshold and less than or equal to the fourth threshold, and the fibrosis significance is less than or equal to the fifth threshold, then it is marked as medium risk;

[0030] If the fat layer significance is greater than a fourth threshold and the fibrosis significance is greater than a fifth threshold, it is marked as high risk.

[0031] Furthermore, the non-alcoholic fatty liver disease early screening model is a convolutional neural network comprising a fully connected layer.

[0032] The present invention has the following beneficial effects:

[0033] The present invention realizes automatic labeling of liver CT images, and analyzes liver CT images in combination with the characteristics of localized fat infiltration and localized fibrosis in the early stage of non-alcoholic fatty liver disease. Considering that fat infiltration will show the characteristics of highlight, small area and circle in liver CT images, the grayscale value of the pixel area and the regional shape characteristics can be used to effectively screen out the fat infiltration area and obtain the fat layer significance. Further considering that fibrosis will produce obvious texture information in the image, a clear bimodal distribution will be produced in the grayscale histogram, so it is processed in the grayscale histogram, and the probability of each peak as a fibrosis peak is analyzed except for the highest peak, and finally the fibrosis peak is judged and the fibrosis significance is obtained. Automatic labeling of liver CT images is realized by these two features. The early screening model for non-alcoholic fatty liver disease can be trained by automatically labeled liver CT images. The embodiment of the present invention identifies the features in the liver CT image and facilitates the training of the early screening model for non-alcoholic fatty liver disease through automatic labeling, thereby ensuring the efficiency of training and the accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. 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 any creative work.

[0035] Figure 1 A flow chart of a method for training a model for early screening of non-alcoholic fatty liver disease provided by one embodiment of the present invention;

[0036] Figure 2 A schematic diagram of a liver CT image of early non-alcoholic fatty liver disease including an infiltrated area provided by one embodiment of the present invention;

[0037] Figure 3 An image of severe fatty liver provided by one embodiment of the present invention;

[0038] Figure 4 A schematic diagram of a CT image comparison between fibrotic liver tissue and normal liver tissue provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0039] To further illustrate the technical means and effectiveness of the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effectiveness of a model training method for early screening of non-alcoholic fatty liver disease proposed by the present invention. In the following description, references to different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics of one or more embodiments may be combined in any suitable manner.

[0040] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0041] The specific scheme of the early screening model training method for non-alcoholic fatty liver disease provided by the present invention is described in detail below with reference to the accompanying drawings.

[0042] See also Figure 1 , which shows a flow chart of a method for training a model for early screening of non-alcoholic fatty liver disease provided by one embodiment of the present invention, the method comprising:

[0043] Step S1: Obtaining a liver CT image of early non-alcoholic fatty liver disease.

[0044] Because the present invention aims to process a training set for a non-alcoholic fatty liver disease early screening model, it first requires obtaining a large number of training images. These training data should consist of liver CT images that already demonstrate early-stage non-alcoholic fatty liver disease. Processing these liver CT images allows the non-alcoholic fatty liver disease early screening model to learn the CT image characteristics of early-stage non-alcoholic fatty liver disease. Specific liver CT images can be obtained from public databases or internal hospital databases and will not be detailed here.

[0045] It should be noted that the grayscale value of each pixel in the CT image is the density value during the CT scanning process, and the embodiments of the present invention are described in terms of grayscale values.

[0046] It should be noted that the present invention is mainly aimed at the analysis of the liver segment area in the liver CT image. Other tissues such as the blood vessels of the liver are not within the analysis scope of the embodiment of the present invention. Therefore, irrelevant tissues can be eliminated based on the existing image segmentation technology to ensure the efficiency of subsequent analysis. The details are not repeated here. Those skilled in the art can choose semantic segmentation networks or simple preset mask segmentation methods for segmentation.

[0047] Step S2: Obtain each pixel region in the liver CT image; obtain the fat infiltration probability of each pixel region based on the grayscale value and regional shape characteristics of each pixel region; filter out the fat infiltration region based on the fat infiltration probability and obtain the fat layer prominence of the liver CT image.

[0048] See also Figure 2 , Figure 2 FIG1 shows a schematic diagram of a liver CT image of an early stage non-alcoholic fatty liver disease including an infiltrated area provided by an embodiment of the present invention. Figure 2 As shown in the figure, the area pointed by the arrow is the fatty infiltration area. For early non-alcoholic fatty liver disease, the fat on the liver is localized or diffuse, with obvious grayscale differences from normal liver tissue, and forms a circular or spotted shape. Figure 3 In contrast to the image of severe fatty liver provided by an embodiment of the present invention, Figure 2 This early symptom has obvious characteristics, and Figure 3 The fat in the image has completely infiltrated normal liver tissue, the boundaries are no longer clear, and significant fibrosis is present. Therefore, in this embodiment of the present invention, because the liver CT images used are of early-stage non-alcoholic fatty liver disease, the fat-infiltrated areas can be screened and the prominence of the fat layer quantified by analyzing the grayscale differences and shapes between the various regions in the image.

[0049] In the embodiment of the present invention, each pixel region in the liver CT image is first obtained, that is, each pixel region is composed of pixels of the same gray value. Figure 2 As shown, the fatty infiltration areas of early-stage non-alcoholic fatty liver disease are characterized by high brightness, a nearly circular shape, and a small area. Therefore, the fatty infiltration probability of each pixel region can be determined based on its grayscale value and regional shape characteristics. This probability can be used to identify fatty infiltration areas. Specifically, the greater the fatty infiltration probability of a pixel region, the more likely it is to be fatty infiltrated. Further, by calculating the area or number of fatty infiltration areas, the prominence of the fat layer in the current liver CT image can be determined. In other words, a greater fat layer prominence indicates a more pronounced fatty infiltration area in the image.

[0050] Preferably, in an embodiment of the present invention, the method for acquiring a pixel area includes:

[0051] All pixels in the liver CT image are clustered according to the grayscale value of each pixel, and the clustering area corresponding to each cluster is the pixel area. In the embodiment of the present invention, the clustering algorithm adopts a density clustering algorithm. The specific clustering algorithm is a technical means well known to those skilled in the art and will not be described in detail here.

[0052] In other embodiments of the present invention, each pixel region may also be obtained by a region growing algorithm. The specific method is well known to those skilled in the art and will not be described in detail here.

[0053] Preferably, in an embodiment of the present invention, considering that the fat infiltration area is nearly circular and small in area, the method for acquiring the area shape feature includes:

[0054] Get the minimum circumscribed circle area of ​​a pixel region.

[0055] The maximum and minimum distances between boundary pixels of a pixel region are obtained, and the ratio of the minimum and maximum distances is used as the initial approximation to circularity of the pixel region. Specifically, the closer the minimum and maximum distances are, the smaller the aspect ratio of the pixel region, the closer it is to a circle, and the greater the initial approximation to circularity.

[0056] The ratio of the pixel area to the minimum circumscribed circle area is multiplied by the initial near-circularity to obtain the near-circularity. The minimum circumscribed circle area is always greater than or equal to the pixel area area. Therefore, the closer this ratio is to 1, the more circular the pixel area is, and the greater the near-circularity obtained when combined with the initial near-circularity.

[0057] The ratio of the near-circularity to the normalized pixel area is used as the regional shape feature. Because the pixel area should be negatively correlated with the probability of fat infiltration, meaning the smaller the area, the more likely it is a fat-infiltrated area, the regional shape feature is obtained by normalizing the pixel area and dividing the near-circularity by the normalized area. A larger regional shape feature indicates that the pixel area more closely matches the shape features of a fat-infiltrated area.

[0058] In the embodiments of the present invention, the normalization method adopts range normalization, and other basic mathematical algorithms may also be adopted, which will not be described in detail here.

[0059] Preferably, in the embodiment of the present invention, considering that the regional shape feature is quantified such that the larger the value, the more likely the pixel region belongs to the fat infiltration region, the method for obtaining the fat infiltration probability includes:

[0060] Because fatty infiltration areas are prominently highlighted in liver CT images, their average grayscale values ​​should be significantly higher. Therefore, the grayscale difference between the average grayscale value of the pixel region and the average grayscale value of the entire liver CT image is calculated. The product of this grayscale difference and the regional shape characteristics is then normalized to obtain the probability of fatty infiltration.

[0061] In the embodiment of the present invention, if the fat infiltration probability is greater than a preset first threshold, the corresponding pixel area is regarded as a fat infiltration area. In the embodiment of the present invention, because the fat infiltration probability has been normalized, the first threshold is set to 0.88.

[0062] Preferably, in the embodiment of the present invention, the area ratio of the fatty infiltration area to the entire liver CT image is used as the fat layer prominence. That is, the larger the area ratio of the fatty infiltration area to the entire area, the more prominent the fat layer and the more severe the liver cell damage.

[0063] Step S3: In the liver CT image, a grayscale histogram of the liver area excluding the fatty infiltration area is obtained; for each peak in the grayscale histogram except the highest peak, the fibrosis peak is determined based on the peak similarity of each peak relative to the highest peak and the grayscale value distance; the fibrosis significance is obtained based on the size of the pixel area corresponding to the fibrosis peak.

[0064] After the liver produces a fat layer, it will try to repair the damage and produce excessive fibrous tissue. This fibrous tissue will gradually replace the normal liver tissue and form fibrosis. This fibrosis is also an important sign of liver damage. The more severe the fibrosis, the more severe the liver function decline. Therefore, for the liver CT images of early non-alcoholic fatty liver disease, attention should also be paid to the surface fibrosis. Figure 4 As shown, Figure 4 FIG1 shows a schematic diagram of a CT image comparison between fibrotic liver tissue and normal liver tissue provided by an embodiment of the present invention. Figure 4 The left side of the middle image is a fibrotic liver, and the right side is a normal liver area. It can be clearly seen that the area circled on the left is a clear fibrotic area. The fibrotic area has different pixel values ​​compared to normal liver tissue, and the distribution is more significant.

[0065] If fibrosis is present in the liver, the image will primarily consist of normal liver pixels and fibrotic pixels. If fibrosis is absent, the image will primarily consist of normal liver pixels. This embodiment of the present invention takes into account that liver fibrosis may also be accompanied by areas of fatty infiltration, which also constitute a type of pixel. To ensure analysis accuracy, this fatty infiltration area is removed, and a grayscale histogram of the liver area excluding the fatty infiltration is obtained. This grayscale histogram primarily contains one or two significant peaks, facilitating subsequent analysis.

[0066] For a grayscale histogram, the highest peak is the peak containing the most pixels. Because the embodiments of the present invention analyze liver CT images of early-stage non-alcoholic fatty liver disease, the fibrotic region is smaller than the normal region. Therefore, normal liver pixels should be the most abundant in the image, meaning the highest peak can be considered the peak formed by normal liver pixels. Therefore, the highest peak can be used as a basis for determining whether other peaks are fibrosis peaks. For each peak, the closer its peak is to the highest peak and the greater the grayscale value distance, indicating that the grayscale histogram exhibits a distinct bimodal distribution, the more likely that the peak is a fibrosis peak. Based on this, the fibrosis peak can be identified, and the fibrosis significance can be determined based on the size of the pixel area corresponding to the fibrosis peak.

[0067] Preferably, in an embodiment of the present invention, the method for determining the fibrosis peak includes:

[0068] The abscissa and ordinate differences between each peak and the highest peak are obtained; the ratio of these differences is used as a judgment indicator. Specifically, the abscissa difference represents the distance between the grayscale values ​​of the two peaks, while the ordinate difference represents the difference in pixel frequencies corresponding to the two peaks. For a peak, the smaller the abscissa difference and the larger the ordinate difference, the farther the peak is from the highest peak. If both peaks are significant, the peak is more likely to be a fibrosis peak.

[0069] If the judgment index for any peak other than the highest peak in the grayscale histogram exceeds the preset second threshold, the peak with the largest judgment index is selected as the fibrosis peak. This means that if multiple peaks have judgment indexes exceeding the preset second threshold, the peak with the largest judgment index is selected as the fibrosis peak. In this embodiment of the present invention, after normalization of the judgment index, the second threshold is set to 0.9.

[0070] If there is no other peak in the grayscale histogram whose judgment index is greater than the preset second threshold, it is determined that there is no fibrosis in the liver.

[0071] Preferably, in the embodiment of the present invention, considering that the fiber peak is mainly determined based on the peak point, it is impossible to directly determine the grayscale interval of the fibrosis area accurately. Therefore, the embodiment of the present invention adopts a method of analyzing the correlation of grayscale values ​​to determine the correspondence between the pixel area and the double peak, thereby obtaining the fibrosis significance, which specifically includes:

[0072] If there is a fibrosis peak, for a pixel area, the difference between the average grayscale value of the pixel area and the horizontal coordinate of the peak point of the fibrosis peak is negatively correlated to obtain the fibrosis peak correlation. The difference between the average grayscale value of the pixel area and the horizontal coordinate of the peak point of the highest peak is negatively correlated to obtain the normal peak correlation. If the fibrosis peak correlation is greater than the normal peak correlation, the pixel area is judged to be a fibrosis area. The product of the area ratio of all fibrosis areas to the total area of ​​the liver CT image and the fat layer significance is used as the fibrosis significance. That is, the larger the area ratio of all fibrosis areas to the total area of ​​the liver CT image, the more severe the fibrosis, and because fibrosis is accompanied by fat infiltration, the fibrosis significance is further obtained by combining the fat layer significance.

[0073] In this embodiment of the present invention, the difference between the average grayscale value of the pixel region and the abscissa of the peak point of the fibrosis peak is the absolute value of the difference between the two values. The inverse of this difference is raised to the power of an exponential function with a natural constant as the base. The output of the exponential function is the negative correlation mapping result of the difference, namely the fibrosis peak correlation. The same applies to the normal peak correlation.

[0074] Step S4: Label the liver CT image according to the fat layer significance and the fibrosis significance; and use the labeled liver CT image to train a non-alcoholic fatty liver disease early screening model.

[0075] The greater the fat layer significance obtained in the above steps, the more areas of fatty infiltration are present; the greater the fibrosis significance, the more severe the liver fibrosis. Therefore, these two indicators can be used to label liver CT images, and then the labeled CT images can be used as training data to train a model for early screening of non-alcoholic fatty liver disease.

[0076] Preferably, in an embodiment of the present invention, if the prominence of the fat layer is less than a preset third threshold, it is marked as low risk;

[0077] If the fat layer significance is greater than or equal to the third threshold and less than or equal to the fourth threshold, and the fibrosis significance is less than or equal to the fifth threshold, then it is marked as medium risk;

[0078] If the fat layer significance is greater than a fourth threshold and the fibrosis significance is greater than a fifth threshold, it is marked as high risk.

[0079] In the embodiment of the present invention, since the fat layer significance is a percentage data and the fibrosis significance is a combination of two percentage data, both data can be converted into percentage form, and the third threshold is set to 15%, the fourth threshold is set to 30%, and the fifth threshold is set to 25%.

[0080] In this embodiment of the present invention, the early screening model for non-alcoholic fatty liver disease is a convolutional neural network with fully connected layers. During use, the image to be assessed is input into the model, which then outputs its risk level. It should be noted that the training methods for convolutional neural networks with fully connected layers are well known to those skilled in the art and will not be further described in this embodiment of the present invention.

[0081] In summary, the present invention effectively screens out fat infiltration areas and obtains the significance of the fat layer through the grayscale value of the pixel area and the regional shape characteristics. Processing is performed in the grayscale histogram, and the probability of each peak as a fibrosis peak is analyzed except for the highest peak. Finally, the fibrosis peak is determined and the fibrosis significance is obtained. Automatic labeling of liver CT images is achieved through these two features. The early screening model for non-alcoholic fatty liver disease can be trained with the automatically labeled liver CT images. The embodiment of the present invention identifies features in liver CT images and facilitates the training of the early screening model for non-alcoholic fatty liver disease through automatic labeling, thereby ensuring the efficiency of training and the accuracy of the model.

[0082] It should be noted that the order in which the embodiments of the present invention are described above is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0083] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

Claims

1. A method for training a model for early screening of non-alcoholic fatty liver disease, characterized in that: The method comprises: Obtain liver CT images of early non-alcoholic fatty liver disease; Obtaining each pixel region in the liver CT image; obtaining a fat infiltration probability for each pixel region based on the grayscale value and regional shape characteristics of each pixel region; screening out the fat infiltration region based on the fat infiltration probability and obtaining the fat layer prominence of the liver CT image; In liver CT images, a grayscale histogram of the liver region excluding the fatty infiltration region is obtained. For each peak in the grayscale histogram except the highest peak, the fibrosis peak is determined based on the peak similarity and grayscale value distance of each peak relative to the highest peak. The fibrosis significance is determined based on the size of the pixel area corresponding to the fibrosis peak. Liver CT images are labeled according to the prominence of the fat layer and fibrosis; the labeled liver CT images are used to train an early screening model for non-alcoholic fatty liver disease.

2. The early screening model training method for non-alcoholic fatty liver disease according to claim 1, characterized in that: The method for obtaining the pixel area includes: All pixels are clustered according to the grayscale value of each pixel in the liver CT image, and the clustering area corresponding to each cluster is the pixel area.

3. The early screening model training method for non-alcoholic fatty liver disease according to claim 1, characterized in that: The method for obtaining the regional shape feature includes: Obtain the minimum circumscribed circle area of ​​the pixel area; obtain the maximum distance and the minimum distance between the boundary pixel points of the pixel area, and use the ratio of the minimum distance to the maximum distance as the initial near-circularity of the pixel area; multiply the ratio of the area of ​​the pixel area to the minimum circumscribed circle area by the initial near-circularity to obtain the near-circularity; and use the ratio of the near-circularity to the normalized pixel area area as the area shape feature.

4. The early screening model training method for non-alcoholic fatty liver disease according to claim 3, characterized in that: The method for obtaining the fat infiltration probability includes: A grayscale value difference between an average grayscale value of a pixel region and an average grayscale value of the entire liver CT image is obtained, and a product of the grayscale value difference and the regional shape feature is normalized to obtain the fat infiltration probability.

5. The early screening model training method for non-alcoholic fatty liver disease according to claim 1, characterized in that: The method for screening the fat infiltration area comprises: If the fat infiltration probability is greater than a preset first threshold, the corresponding pixel area is regarded as a fat infiltration area.

6. The method for early screening model training for non-alcoholic fatty liver disease according to claim 1, characterized in that: The method for obtaining the fat layer prominence includes: The area ratio of the fatty infiltration area to the entire liver CT image was used as the fat layer prominence.

7. The early screening model training method for non-alcoholic fatty liver disease according to claim 1, characterized in that: The method for determining the fibrosis peak includes: Obtaining the abscissa difference and ordinate difference of the peak points between each peak and the highest peak; and using the ratio of the abscissa difference to the ordinate difference as a judgment indicator; If the judgment index of other peaks except the highest peak in the grayscale histogram is greater than the preset second threshold, the peak with the largest judgment index is selected as the fibrosis peak; If there is no other peak in the grayscale histogram whose judgment index is greater than the preset second threshold, it is determined that there is no fibrosis in the liver.

8. The method for early screening model training for non-alcoholic fatty liver disease according to claim 1, characterized in that: The method for obtaining the fibrosis significance includes: If a fibrosis peak exists, for a pixel region, a negative correlation mapping is performed between the difference between the average grayscale value of the pixel region and the abscissa of the peak point of the fibrosis peak to obtain a fibrosis peak correlation; a negative correlation mapping is performed between the difference between the average grayscale value of the pixel region and the abscissa of the peak point of the highest peak to obtain a normal peak correlation; if the fibrosis peak correlation is greater than the normal peak correlation, the pixel region is determined to be a fibrosis region; the product of the area ratio of all fibrosis regions to the total area of ​​the liver CT image and the fat layer prominence is used as the fibrosis prominence; If there is no fibrosis peak, the fibrosis significance is judged to be 0.

9. The method for early screening model training for non-alcoholic fatty liver disease according to claim 1, characterized in that: The marking of the liver CT image according to the fat layer significance and the fibrosis significance includes: If the fat layer significance is less than a preset third threshold, it is marked as low risk; If the fat layer significance is greater than or equal to the third threshold and less than or equal to the fourth threshold, and the fibrosis significance is less than or equal to the fifth threshold, then it is marked as medium risk; If the fat layer significance is greater than a fourth threshold and the fibrosis significance is greater than a fifth threshold, it is marked as high risk.

10. The method for early screening model training for non-alcoholic fatty liver disease according to claim 1, characterized in that: The non-alcoholic fatty liver disease early screening model is a convolutional neural network comprising a fully connected layer.

Images