Rare disease recognition system based on neural network

Abstract

The invention relates to the technical field of image recognition, in particular to a rare disease recognition system based on a neural network. The system comprises a sample collection module, a sample screening module, a sample scanning module, a sample identification module and a sample evaluation module. Rare disease samples are automatically collected, the efficiency and consistency of data collection are improved, an original data set is preprocessed and screened through the sample screening module, the efficiency of data processing is improved, the data quality is ensured, the sample scanning module accurately segments a focus area in a scanned image through the image recognition technology, and the accuracy of image recognition is improved. The real-time accuracy is calculated through the sample evaluation module, screening parameters are dynamically adjusted according to the real-time accuracy, the recognition accuracy is improved, the model carries out refined learning on a target task, the applicability of the model is improved, meanwhile, dependence on large-scale annotation data is reduced, and the recognition efficiency is improved. And computing power resources required for training the deep learning model are saved.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a rare disease recognition system based on neural network. Background Art

[0002] There are more than 6,000 rare diseases in the world, affecting more than 300 million people. Early and accurate diagnosis has always been an important goal of clinical practice for rare diseases. Among them, about 80% are related to genetic factors. Rare diseases are difficult to diagnose and take a long time to confirm. Accurate identification of rare disease lesions is a great challenge. Nowadays, computer vision is developing rapidly. How to introduce computer vision technology into the work of detecting medical images to improve the efficiency and accuracy of image diagnosis? Among computer vision recognition methods, convolutional neural network (CNN) is a deep learning model that is particularly suitable for processing image data. CNN consists of multiple convolutional layers and pooling layers. It can automatically learn the hierarchical feature representation of images and has strong expression and generalization capabilities. It has achieved great success in image classification, target detection, semantic segmentation and other fields, and is also widely used in image analysis.

[0003] Patent document with publication number CN112472048A discloses a neural network structure for pulse recognition of patients with cardiovascular diseases, including the following steps: S1, pulse data collection; S2, pulse data preprocessing; S3, establishing the neural network structure; S4, training process; It can be seen that the existing disease detection technology lacks an identification system suitable for rare lung diseases, which can quickly and accurately screen the input influencing data set, thereby inputting high-quality images into the CNN model to accurately capture the coarse segmentation area of ​​the lesion area and improve the recognition efficiency. Summary of the invention

[0004] To this end, the present invention provides a rare disease recognition system based on a neural network to overcome the problem in the prior art that the image quality input to the CNN model is low, resulting in low recognition efficiency due to the lack of rapid and accurate screening of a large number of input data sets.

[0005] To achieve the above object, the present invention provides a rare disease identification system based on neural network, comprising:

[0006] A sample collection module, which is connected to an external big data platform to collect rare disease samples at a preset collection cycle to obtain an original data set;

[0007] A sample screening module preprocesses and screens the original data set in sequence to obtain a target data set;

[0008] a sample scanning module, for acquiring each scanned image in the target data set, determining the real-time line area ratio according to the first standard line area ratio and the second standard line area ratio, selecting a processing mode according to the determination result, identifying and segmenting the corresponding lesion area in each scanned image according to the processing mode and the convolutional neural network, and obtaining a coarse segmentation image;

[0009] A sample recognition module recognizes the coarse segmentation image, extracts image features, compares the image features with lesion features, and marks the matching coarse segmentation image as a target image sample;

[0010] The sample evaluation module calculates the percentage of the target image samples correctly predicted by the model to the total image samples, and obtains the real-time accuracy to determine whether to adjust the screening parameters;

[0011] Wherein, the screening parameter is the area proportion of the second standard line.

[0012] Furthermore, the sample screening module includes a pre-processing unit, a primary screening unit and an iterative screening unit, wherein:

[0013] The preprocessing unit is used to obtain the image data in the original data set as the original image, perform contour enhancement processing on the original image, and obtain an intermediate image;

[0014] The primary screening unit performs a primary screening on each of the intermediate images to delete intermediate images with unqualified quality and obtain an intermediate data set;

[0015] The iterative screening unit performs continuous extended scanning on the intermediate images in the intermediate data set to obtain the target data set.

[0016] Furthermore, the sample scanning module includes a window sliding unit, an analysis unit, a selection processing unit and a segmentation unit, wherein:

[0017] The window sliding unit can obtain the total number of pixels of any scanned image, calculate a preset sliding window size according to the sliding scanning coefficient and the total number of pixels, divide the scanned image into a plurality of to-be-scanned areas according to the preset sliding window size, and perform sliding scanning on the to-be-scanned areas;

[0018] The analysis unit is used to obtain the marked line pixels in the area to be scanned, and calculate the unit line pixel ratio and the real-time line area ratio according to the marked line pixels;

[0019] The selection processing unit is used to select a processing mode according to the judgment result, so as to identify and segment the corresponding lesion area in each of the scanned images according to each processing mode and a convolutional neural network to obtain a coarse segmentation image;

[0020] The segmentation unit is used to segment the image to be analyzed and the area to be scanned to obtain a coarse segmentation image.

[0021] Furthermore, the analysis unit includes a determination subunit, a marking subunit and a calculation subunit, wherein:

[0022] The determination subunit is used to compare the pixel ratio of the unit line with the pixel ratio of the standard line;

[0023] The marking subunit is used to mark the image area corresponding to the current sliding window as a line area when the unit line pixel ratio is greater than or equal to the standard line pixel ratio;

[0024] The calculation subunit is used to obtain the percentage of the number of line areas to the total number of areas to be scanned, and obtain the real-time line area ratio.

[0025] Furthermore, the selection processing unit includes a comparison subunit and a processing subunit, wherein:

[0026] The comparison subunit is used to compare the real-time line area ratio according to the first standard line area ratio and the second standard line area ratio;

[0027] The processing subunit is used to enhance the line contour edge of the area to be scanned when the real-time line area ratio is between the first standard line area ratio and the second standard line area ratio to obtain the image to be analyzed.

[0028] Furthermore, the segmentation unit includes a first segmentation subunit, a second segmentation subunit and a third segmentation subunit, wherein:

[0029] When the proportion of the real-time line area is less than the proportion of the first standard line area, the first segmentation subunit performs nodule feature recognition and segmentation on the to-be-scanned area according to a convolutional neural network to obtain a coarse segmentation image;

[0030] The first segmentation subunit identifies nodule features according to a convolutional neural network, segments the area where the nodule features are located, and obtains a coarse segmentation image;

[0031] The second segmentation subunit is used to identify the image to be analyzed to obtain a coarse segmentation image;

[0032] The third segmentation subunit is used to identify and segment the area to be scanned according to a convolutional neural network when the proportion of the real-time line area is greater than the proportion of the second standard line area, so as to identify the non-nodular lung lesion area and obtain a coarse segmentation image.

[0033] Furthermore, the processing subunit performs radial gradient transition processing on each gradient transition layer of the line contour edge to enhance the line contour edge.

[0034] Furthermore, the primary screening unit includes a data deletion subunit and a data enhancement subunit, wherein:

[0035] The data deletion subunit is used to determine whether to delete the intermediate image according to the slice thickness;

[0036] The data enhancement subunit is used to mark the edge contours of the edge pixels of the picture area in the intermediate image area, and adjust each original brightness value to a corresponding modified brightness value in turn, so as to perform data enhancement processing on the brightness data of the intermediate image.

[0037] Furthermore, the data enhancement subunit respectively obtains the brightness value of each pixel point in the inner contour area as a number of original brightness values, calculates the average inner contour brightness according to each original brightness value, obtains the brightness value of each pixel in the outer contour area to calculate the average outer contour brightness, subtracts the average inner contour brightness from the average outer contour brightness to obtain a real-time brightness difference, and adds each original brightness value to the real-time brightness difference in turn to obtain a corresponding corrected brightness value, so as to perform data enhancement processing on the brightness data of the intermediate image.

[0038] Furthermore, the sample evaluation module includes a comparison unit and a correction unit, wherein:

[0039] The comparison unit is set with an accuracy threshold, and the comparison unit compares the real-time accuracy with the accuracy threshold to obtain a comparison result;

[0040] The correction unit obtains the comparison result, and when the real-time accuracy is less than the accuracy threshold, corrects the second standard line area ratio to the second corrected line area ratio.

[0041] Compared with the prior art, the beneficial effects of the present invention are that, by connecting to a big data platform, rare disease samples can be automatically collected to improve the efficiency and consistency of data collection; the original data set can be preprocessed and screened through a sample screening module to improve the efficiency of data processing and ensure data quality; the sample scanning module uses image recognition technology to accurately segment the lesion area in the scanned image, providing an accurate basis for subsequent analysis; the real-time accuracy rate is calculated through a sample evaluation module to dynamically adjust the screening parameters according to the real-time accuracy rate, optimize the recognition process, and improve the recognition accuracy.

[0042] Furthermore, by determining the slice thickness of the image, images with slice thickness that meets the standard are automatically screened out to ensure that sufficient details are captured.

[0043] Furthermore, by setting the scanning coefficient, the corresponding preset sliding window size is calculated according to the actual size of the image to be processed, thereby improving the applicability to image processing.

[0044] Furthermore, when determining that the real-time line area ratio is between the first standard line area ratio and the second standard line area ratio, there may be some blurred edges of the lines. Since blurred or unclear edges are also related to lesions, the line contour edges are first enhanced to ensure the clarity of line recognition and avoid misidentification. After the enhanced processing, the line edge pixels are analyzed again. If blurriness and unclearness are still present, the corresponding area is marked as a lesion area, and associated image processing technology is used to more accurately analyze the features in the image.

[0045] Furthermore, in order to train a high-performance CNN model, thousands to tens of thousands of labeled training images are usually required. These images need to accurately indicate the targets or categories in the images so that the model can learn features and make effective predictions. Therefore, by fine-tuning the proportion of the second standard line area of ​​the screening parameter on a smaller target data set to adapt to the process of adjusting the selection processing mode of the selection processing unit, the model can perform refined learning for the target task to improve its performance on a specific data set, thereby improving the applicability of the model, while also reducing dependence on large-scale labeled data and saving computing power resources required for training deep learning models. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Schematic diagram of the structure of a rare disease identification system based on a neural network according to an embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram of the structure of a sample screening module according to an embodiment of the present invention;

[0048] Figure 3 It is a schematic diagram of the structure of a sample scanning module according to an embodiment of the present invention;

[0049] Figure 4 Schematic diagram of the structure of the analysis unit according to the embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0051] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0052] It should be noted that, in the description of the present invention, terms such as "up", "down", "left", "right", "inside" and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the drawings. This is merely for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation on the present invention.

[0053] In addition, it should be noted that in the description of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0054] See also Figure 1 As shown, it is a schematic diagram of the structure of a rare disease identification system based on a neural network according to an embodiment of the present invention. The present invention provides a rare disease identification system based on a neural network, comprising:

[0055] A sample collection module, which is connected to an external big data platform to collect rare disease samples at a preset collection cycle to obtain an original data set;

[0056] A sample screening module preprocesses and screens the original data set in sequence to obtain a target data set;

[0057] A sample scanning module, which obtains each scanned image in the target data set to identify and segment the corresponding lesion area in each scanned image to obtain a coarse segmented image;

[0058] The sample recognition module recognizes the coarse segmentation image, extracts image features, compares the image features with the lesion features, and marks the matching coarse segmentation image as the target image sample; the sample evaluation module calculates the percentage of the number of target image samples correctly predicted by the model to the total image samples, and obtains the real-time accuracy rate to determine whether to adjust the screening parameters.

[0059] The screening parameter is the area ratio of the second standard line.

[0060] In this embodiment, rare disease recognition results are output through automatic classification and recognition of lung images, and the rare disease recognition results include normal tissue, benign nodules, malignant nodules and non-nodular lesions.

[0061] By connecting to the big data platform, rare disease samples are collected automatically to improve the efficiency and consistency of data collection. The sample screening module preprocesses and screens the original data set to improve data processing efficiency and ensure data quality. The sample scanning module uses image recognition technology to accurately segment the lesion area in the scanned image, providing an accurate basis for subsequent analysis. The sample evaluation module calculates the real-time accuracy rate, dynamically adjusts the screening parameters according to the real-time accuracy rate, optimizes the recognition process, and improves recognition accuracy.

[0062] See also Figure 2 As shown, it is a schematic diagram of the structure of a sample screening module according to an embodiment of the present invention;

[0063] Specifically, the sample screening module includes a preprocessing unit, a primary screening unit and an iterative screening unit, wherein:

[0064] The preprocessing unit is used to obtain the image data in the original data set as the original image, perform contour enhancement processing on the original image, and obtain an intermediate image;

[0065] The primary screening unit performs a primary screening on each of the intermediate images to delete intermediate images with unqualified quality and obtain an intermediate data set;

[0066] The iterative screening unit performs continuous extended scanning on the intermediate images in the intermediate data set to obtain the target data set.

[0067] The primary screening unit includes a data deletion subunit and a data enhancement subunit, wherein:

[0068] The data deletion subunit is used to determine whether to delete the intermediate image according to the slice thickness;

[0069] Among them, when the slice thickness exceeds 5 mm, the intermediate image is deleted;

[0070] The data enhancement subunit is used to mark the edge of the edge pixels of the picture area in the intermediate image area, and adjust each original brightness value to a corresponding modified brightness value in turn, so as to perform data enhancement processing on the brightness data of the intermediate image;

[0071] By judging the slice thickness of the image, the images with slice thickness meeting the standard are automatically screened out to ensure that sufficient details are captured.

[0072] Specifically, the data enhancement processing includes:

[0073] The data enhancement subunit obtains the brightness value of each pixel point in the inner contour area respectively as a number of original brightness values, calculates the average inner contour brightness according to each original brightness value, obtains the brightness value of each pixel in the outer contour area to calculate the average outer contour brightness, subtracts the average inner contour brightness from the average outer contour brightness to obtain a real-time brightness difference, and adds each original brightness value to the real-time brightness difference in turn to obtain a corresponding corrected brightness value.

[0074] By performing data enhancement processing on the brightness values ​​of pixels, the data can be analyzed accurately and reliably, and the brightness values ​​of each pixel in the inner contour area of ​​the image can be balanced and adjusted to the brightness values ​​of each pixel in the outer contour area close to the image area, making the contour edge more obvious.

[0075] The iterative screening unit includes an image cropping subunit and a screening subunit, wherein:

[0076] The image cropping subunit is used to rotate, translate and scale the intermediate image;

[0077] The screening subunit is used to perform continuous extended scanning on the intermediate images in the intermediate data set to delete repeated intermediate images;

[0078] In this embodiment, the hash value of each intermediate image is calculated, and the hash value of each intermediate image is compared with the hash values ​​of other intermediate images through continuous extended scanning. If the Hamming distance between two hash values ​​is less than a threshold, the two images are considered to be duplicates, and one of the images with the same hash value is retained and the others are deleted.

[0079] See also Figure 3 As shown, it is a schematic diagram of the structure of a sample scanning module according to an embodiment of the present invention;

[0080] Specifically, the sample scanning module includes a window sliding unit, an analysis unit, a selection processing unit and a segmentation unit, wherein:

[0081] The window sliding unit can obtain the total number of pixels of any scanned image, calculate a preset sliding window size according to the sliding scanning coefficient and the total number of pixels, divide the scanned image into a plurality of to-be-scanned areas according to the preset sliding window size, and perform sliding scanning on the to-be-scanned areas;

[0082] The analysis unit is used to obtain the marked line pixels in the area to be scanned, and calculate the unit line pixel ratio and the real-time line area ratio according to the marked line pixels;

[0083] The selection processing unit is used to determine the real-time line area ratio according to the first standard line area ratio and the second standard line area ratio, select a processing mode according to the determination result, and perform enhanced processing on the to-be-scanned area according to the processing mode to obtain an image to be analyzed;

[0084] The segmentation unit is used to segment the image to be analyzed and the area to be scanned to obtain a coarse segmentation image.

[0085] In this embodiment, the preset sliding window size is the product of the scanning coefficient and the total number of pixels. The image scanning coefficient represents a set decimal value. It is related to the size and quality requirements of the image to be processed. Generally, the set value is guaranteed to make the preset scanning window size the square of the integer value. Even if the preset sliding window size can be 3×3, it can be set between 0.01-0.1; the sliding window is a rectangle; dividing the scanned image into a number of areas to be scanned means simulating the closure of the current scanning area according to the current sliding window, accurately obtaining the corresponding image area for determination, and ensuring that the determination is carried out smoothly and effectively;

[0086] By setting the scanning coefficient, the corresponding preset sliding window size is calculated according to the actual size of the image to be processed, thereby improving the applicability of image processing.

[0087] See also Figure 4 As shown, it is a schematic diagram of the structure of the analysis unit of an embodiment of the present invention;

[0088] Specifically, the analysis unit includes a determination subunit, a marking subunit and a calculation subunit, wherein:

[0089] The determination subunit is used to compare the pixel ratio of the unit line with the pixel ratio of the standard line;

[0090] The marking subunit is used to mark the image area corresponding to the current sliding window as a line area when the unit line pixel ratio is greater than or equal to the standard line pixel ratio;

[0091] When the pixel ratio of the unit line is less than the pixel ratio of the standard line, the image area corresponding to the current sliding window is not marked;

[0092] The calculation subunit is used to obtain the percentage of the number of line areas to the total number of areas to be scanned, and obtain the real-time line area ratio.

[0093] The standard line pixel ratio set in this embodiment represents the ratio of white line pixels to all pixels in the image area corresponding to the current sliding window, representing the number of white line pixels. Generally, the standard line pixel ratio is set to 5%-15%, preferably, it is set to 10%.

[0094] Specifically, the selection processing unit includes a comparison subunit and a processing subunit, wherein:

[0095] The comparison subunit is used to compare the real-time line area ratio according to the first standard line area ratio and the second standard line area ratio;

[0096] The processing subunit is used to enhance the line contour edge in the picture area when the real-time line area ratio is between the first standard line area ratio and the second standard line area ratio to obtain the image to be analyzed.

[0097] Strengthening the edge of the line contour includes:

[0098] Perform radial gradient transition processing on each gradient transition layer at the edge of the line contour;

[0099] Wherein, determining the number of radial gradient transition layers includes obtaining line contour edge segments as initial radial gradient transition layers, obtaining each pixel constituting the line contour edge segments as each edge center point, and calculating the brightness difference between the adjacent upper and lower layers of pixel points and the edge center point in sequence from each edge center point to obtain a number of real-time brightness difference values. If there is a real-time brightness difference value greater than a preset brightness difference value, then selecting the layer where the pixel point is located as the radial gradient transition layer, and accumulating the initial number of radial gradient transition layers;

[0100] The radial gradient transition process is to adjust the color brightness of each gradient transition layer through a superposition algorithm.

[0101] The overlay algorithm in this embodiment adopts an overlay blending mode algorithm to adjust the color of the previous layer according to the color brightness of one layer. That is, if the color brightness of the current layer is less than 0.5, the color of the previous layer is multiplied by 2 to brighten the color of the upper layer. If the color brightness of the current layer is greater than or equal to 0.5, the color of the upper layer is multiplied by the sum of the brightness of the two and the product of the brightness of the two, so that the color of the upper layer becomes darker. The calculation is performed layer by layer to achieve a gradient effect.

[0102] When determining that the real-time line area ratio is between the first standard line area ratio and the second standard line area ratio, some line edges may be blurred. Since blurred or unclear edges are also related to lesions, the line contour edges are first enhanced to ensure the clarity of line recognition and avoid misidentification. After the enhanced processing, the line edge pixels are analyzed again. If blurriness and unclearness are still present, the corresponding area is marked as a lesion area, and associated image processing technology is used to more accurately analyze the features in the image.

[0103] Specifically, the segmentation unit includes a first segmentation subunit, a second segmentation subunit and a third segmentation subunit, wherein:

[0104] When the proportion of the real-time line area is less than the proportion of the first standard line area, the first segmentation subunit performs nodule feature recognition and segmentation on the to-be-scanned area according to a convolutional neural network to obtain a coarse segmentation image;

[0105] The first segmentation subunit identifies nodule features according to a convolutional neural network, segments the area where the nodule features are located, and obtains a coarse segmentation image;

[0106] The second segmentation subunit identifies contour line segments with unclear edges in the image to be analyzed, and uses the corresponding areas as coarse segmentation images;

[0107] The third segmentation subunit is used to identify and segment the area to be scanned according to a convolutional neural network when the proportion of the real-time line area is greater than the proportion of the second standard line area, so as to identify the non-nodular lung lesion area and obtain a coarse segmentation image.

[0108] In this embodiment, a CNN-based model is used as a classifier, and the transfer learning method is used. The model pre-trained on ImageNet is used as the initialization parameter, and then fine-tuning is performed on the lung image dataset; the nodule features are round or quasi-circular with lobed or burr-shaped edges and cavities inside; the non-nodular lung lesion area is a quasi-circular white lesion with gas and liquid planes inside.

[0109] The standard line area ratio set in this embodiment is used to characterize the area ratio where white lines appear in the image. The first standard line area ratio is set to be close to zero, and the second standard line area ratio is set to be 50%. When the real-time line area ratio is less than the first standard line area ratio, it means that nodules, fat or other features block the lines. In this case, it is necessary to analyze the shape characteristics of the nodules according to the convolutional neural network model to analyze whether the nodules are malignant or benign in order to classify the scanned image. When the real-time line area ratio is greater than the second standard line area ratio, it indicates that most of the white lines are clearly visible and there is no obvious nodule occlusion. However, it is still necessary to further analyze whether there are non-nodular lesions through the convolutional neural network to improve the accuracy of recognition.

[0110] The sample evaluation module includes a comparison unit and a correction unit, wherein:

[0111] The comparison unit is set with an accuracy threshold, and the comparison unit compares the real-time accuracy with the accuracy threshold to obtain a comparison result;

[0112] The correction unit obtains the comparison result,

[0113] If the real-time accuracy is less than the accuracy threshold, the second standard line area ratio is corrected to the second corrected line area ratio, and the next sample evaluation is continued with the second corrected line area ratio;

[0114] If the real-time accuracy is greater than or equal to the accuracy threshold, the screening parameters are not adjusted;

[0115] Among them, Nb2'=Nb2×[1+(Qb-Qs) / Qb], Nb2' is the proportion of the second correction line area, Nb2 is the proportion of the second standard line area, Qs is the real-time accuracy, and Qb is the accuracy threshold.

[0116] In this embodiment, the accuracy threshold is set to 86%. The higher the accuracy, the more accurate the model.

[0117] In order to train a high-performance CNN model, thousands to tens of thousands of labeled training images are usually required. These images need to accurately mark the targets or categories in the images so that the model can learn features and make effective predictions. Therefore, by fine-tuning the proportion of the second standard line area of ​​the screening parameter on a smaller target data set to adapt to the process of adjusting the selection processing mode of the selection processing unit, the model can perform refined learning for the target task to improve its performance on a specific data set, thereby improving the applicability of the model, reducing the dependence on large-scale labeled data, and saving the computing power resources required for training deep learning models.

[0118] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

[0119] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A rare disease identification system based on neural network, characterized in that: include, A sample collection module, which is connected to an external big data platform to collect rare disease samples at a preset collection cycle to obtain an original data set; A sample screening module preprocesses and screens the original data set in sequence to obtain a target data set; a sample scanning module, for acquiring each scanned image in the target data set, determining the real-time line area ratio according to the first standard line area ratio and the second standard line area ratio, selecting a processing mode according to the determination result, identifying and segmenting the corresponding lesion area in each scanned image according to the processing mode and the convolutional neural network, and obtaining a coarse segmentation image; A sample recognition module recognizes the coarse segmentation image, extracts image features, compares the image features with lesion features, and marks the matching coarse segmentation image as a target image sample; The sample evaluation module calculates the percentage of the target image samples correctly predicted by the model to the total image samples, and obtains the real-time accuracy to determine whether to adjust the screening parameters; Wherein, the screening parameter is the area proportion of the second standard line.

2. The rare disease identification system based on neural network according to claim 1, characterized in that: The sample screening module includes a pre-processing unit, a primary screening unit and an iterative screening unit, wherein: The preprocessing unit is used to obtain the image data in the original data set as the original image, perform contour enhancement processing on the original image, and obtain an intermediate image; The primary screening unit performs a primary screening on each of the intermediate images to delete intermediate images with unqualified quality and obtain an intermediate data set; The iterative screening unit performs continuous extended scanning on the intermediate images in the intermediate data set to obtain the target data set.

3. The rare disease identification system based on neural network according to claim 1, characterized in that: The sample scanning module includes a window sliding unit, an analysis unit, a selection processing unit and a segmentation unit, wherein: The window sliding unit can obtain the total number of pixels of any scanned image, calculate a preset sliding window size according to the sliding scanning coefficient and the total number of pixels, divide the scanned image into a plurality of to-be-scanned areas according to the preset sliding window size, and perform sliding scanning on the to-be-scanned areas; The analysis unit is used to obtain the marked line pixels in the area to be scanned, and calculate the unit line pixel ratio and the real-time line area ratio according to the marked line pixels; The selection processing unit is used to select a processing mode according to the judgment result, so as to identify and segment the corresponding lesion area in each of the scanned images according to each processing mode and a convolutional neural network to obtain a coarse segmentation image; The segmentation unit is used to segment the image to be analyzed and the area to be scanned to obtain a coarse segmentation image.

4. The rare disease identification system based on neural network according to claim 3, characterized in that: The analysis unit includes a determination subunit, a marking subunit and a calculation subunit, wherein: The determination subunit is used to compare the pixel ratio of the unit line with the pixel ratio of the standard line; The marking subunit is used to mark the image area corresponding to the current sliding window as a line area when the unit line pixel ratio is greater than or equal to the standard line pixel ratio; The calculation subunit is used to obtain the percentage of the number of line areas to the total number of areas to be scanned, and obtain the real-time line area ratio.

5. The neural network-based rare disease identification system according to claim 3, characterized in that: The selection processing unit includes a comparison subunit and a processing subunit, wherein: The comparison subunit is used to compare the real-time line area ratio according to the first standard line area ratio and the second standard line area ratio; The processing subunit is used to enhance the line contour edge of the area to be scanned when the real-time line area ratio is between the first standard line area ratio and the second standard line area ratio to obtain the image to be analyzed.

6. The neural network-based rare disease identification system according to claim 3, characterized in that: The segmentation unit includes a first segmentation subunit, a second segmentation subunit and a third segmentation subunit, wherein: When the proportion of the real-time line area is less than the proportion of the first standard line area, the first segmentation subunit performs nodule feature recognition and segmentation on the to-be-scanned area according to a convolutional neural network to obtain a coarse segmentation image; The first segmentation subunit identifies nodule features according to a convolutional neural network, segments the area where the nodule features are located, and obtains a coarse segmentation image; The second segmentation subunit is used to identify the image to be analyzed to obtain a coarse segmentation image; The third segmentation subunit is used to identify and segment the area to be scanned according to a convolutional neural network when the proportion of the real-time line area is greater than the proportion of the second standard line area, so as to identify the non-nodular lung lesion area and obtain a coarse segmentation image.

7. The neural network-based rare disease identification system according to claim 5, characterized in that: The processing subunit performs radial gradient transition processing on each gradient transition layer of the line contour edge to strengthen the line contour edge.

8. The neural network-based rare disease identification system according to claim 2, characterized in that: The primary screening unit includes a data deletion subunit and a data enhancement subunit, wherein: The data deletion subunit is used to determine whether to delete the intermediate image according to the slice thickness; The data enhancement subunit is used to mark the edge contours of the edge pixels of the picture area in the intermediate image area, and adjust each original brightness value to a corresponding modified brightness value in turn, so as to perform data enhancement processing on the brightness data of the intermediate image.

9. The neural network-based rare disease identification system according to claim 8, characterized in that: The data enhancement subunit obtains the brightness value of each pixel point in the inner contour area respectively as a number of original brightness values, calculates the average inner contour brightness according to each original brightness value, obtains the brightness value of each pixel in the outer contour area to calculate the average outer contour brightness, subtracts the average inner contour brightness from the average outer contour brightness to obtain a real-time brightness difference, and sequentially adds each original brightness value to the real-time brightness difference to obtain a corresponding corrected brightness value, so as to perform data enhancement processing on the brightness data of the intermediate image.

10. The neural network-based rare disease identification system according to claim 1, characterized in that: The sample evaluation module includes a comparison unit and a correction unit, wherein: The comparison unit is set with an accuracy threshold, and the comparison unit compares the real-time accuracy with the accuracy threshold to obtain a comparison result; The correction unit obtains the comparison result, and when the real-time accuracy is less than the accuracy threshold, corrects the second standard line area ratio to the second corrected line area ratio.

Images