Children nose and throat disease analysis system and analysis method based on endoscopic image

Abstract

The invention relates to an endoscopic image-based child nose and throat disease analysis system and analysis method. The system comprises an image space-time filtering module which is used for obtaining an endoscopic image of the nose and throat part of the child and carrying out space-time filtering to obtain a filtered image; the image artifact removal module is used for carrying out motion artifact removal on the filtered image by adopting a motion artifact removal algorithm to obtain a stable image; the image feature extraction module is used for extracting image features from the stable image; the feature lesion positioning module is used for calculating an image feature mutation index based on the image features and positioning a potential lesion area; the disease probability calculation module is used for inputting the image features into a disease probability recognition convolutional neural network based on the potential lesion region to generate a target disease probability; and the disease report generation module is used for integrating the stable image, the image features and the target disease probability to generate a disease diagnosis report. The system can automatically and accurately analyze nasal and throat diseases of children, and improves the diagnosis efficiency and reliability.

Description

Technical Field

[0001] This invention belongs to the field of medical information technology, and in particular relates to a system and method for analyzing pediatric nasopharyngeal diseases based on endoscopic images. Background Technology

[0002] With the development of medical image processing technology, endoscopic image analysis technology has emerged. This technology can preprocess and analyze the features of nasopharyngeal endoscopic images with computer assistance to improve the objectivity and efficiency of diagnosis.

[0003] In traditional techniques, doctors primarily rely on visual observation of endoscopic images, combined with clinical experience for subjective judgment, or use simple image processing tools for enhancement and measurement. However, the entire analysis process remains largely manual. Current traditional methods suffer from the following problems: low cooperation from pediatric patients; endoscopic images are easily affected by motion artifacts, leading to blurred or distorted images and affecting observation accuracy; manual analysis relies on physician experience, is highly subjective, and easily misses minor lesions or misdiagnoses, resulting in poor diagnostic consistency; traditional methods lack efficient automated lesion localization and quantitative assessment mechanisms, making the diagnostic process time-consuming and failing to meet the demands of efficient clinical practice; and existing technologies often cannot integrate multi-feature information for comprehensive disease probability calculation, resulting in insufficient diagnostic reliability. Summary of the Invention

[0004] Therefore, it is necessary to provide a system and method for analyzing pediatric nasopharyngeal diseases based on endoscopic images that can solve the above problems.

[0005] In a first aspect, this application provides a pediatric nasopharyngeal disease analysis system based on endoscopic images, comprising:

[0006] The image spatiotemporal filtering module is used to acquire endoscopic images of the nasopharynx and larynx of children and perform spatiotemporal filtering on the endoscopic images to obtain filtered images;

[0007] The image artifact removal module is used to remove motion artifacts from the filtered image using a motion artifact removal algorithm to obtain a stable image.

[0008] The image feature extraction module is used to extract image features from stable images;

[0009] The feature lesion localization module is used to calculate the image feature mutation index based on image features, and to locate and identify potential lesion areas based on the image feature mutation index.

[0010] The disease probability calculation module is used to input image features into a preset disease probability recognition convolutional neural network based on potential lesion areas to generate the target disease probability.

[0011] The disease report generation module integrates stable images, image features, and target disease probabilities to generate disease diagnosis reports.

[0012] In one embodiment, the image artifact removal module is further configured to:

[0013] Pixel gradient information is obtained by calculating pixel gradients based on the filtered image.

[0014] Based on pixel gradient information, the filtered image is divided using an adaptive threshold segmentation method to obtain motion artifact regions and the corresponding set of artifact pixels.

[0015] For motion artifact regions, a neighborhood pixel gradient consistency check is performed on the artifact pixel set to obtain the consistency check result;

[0016] Based on the consistency check results, pixels in the artifact pixel set that do not meet the preset consistency conditions are removed to obtain the checked artifact pixel set.

[0017] Based on the verified set of artifact pixels, the artifact offset parameters are fitted by constructing a motion trajectory model.

[0018] By combining the artifact offset parameters, pixel-wise inverse compensation correction is performed on the motion artifact region to obtain a preliminary corrected image;

[0019] A gradient distribution equalization method is used to optimize the pixel transition continuity of the initially corrected image and generate a stable image.

[0020] In one embodiment, the image artifact removal module is further configured to determine the motion artifact region using the following formula:

[0021]

[0022] in, For pixel gradient, In pixels A local neighborhood window centered on the center, For local neighborhood windows The mean of the pixel gradient, For local neighborhood windows The standard deviation of the pixel gradient, For local neighborhood windows The skewness of the pixel gradient, , , and This refers to a preset adaptive adjustment coefficient calibrated based on the imaging characteristics of pediatric nasopharyngeal endoscopy images and the intensity range of common motion artifacts. This represents the global maximum value of the pixel gradient. To preset the high gradient threshold range, The number of pixels.

[0023] In one embodiment, the image features include color features, texture features, and morphological features, and the feature lesion localization module is further used for:

[0024] For image features, the feature standard deviation thresholding method is used to calculate the block abrupt change coefficients of color features, texture features, and morphological features;

[0025] Based on the block mutation coefficient and combined with the preset mutation coefficient threshold, color mutation blocks, texture mutation blocks and shape mutation blocks are located.

[0026] Based on color mutation blocks, texture mutation blocks, and morphological mutation blocks, a neighborhood connectivity analysis method is used to perform block connectivity processing to obtain potential lesion blocks.

[0027] Based on potential lesion blocks, morphological closing operations are used to optimize the block boundaries and form potential lesion regions.

[0028] In one embodiment, the disease report generation module is further configured to:

[0029] Determine the target disease category based on the probability of the target disease;

[0030] Based on the target disease category, the severity level of the disease is calculated by using a preset severity grading index, combined with color features, texture features, and morphological features.

[0031] Based on the target disease category and disease severity level, disease information is annotated in potential lesion areas of stable images to obtain an annotated image, which is then used as a disease diagnosis report.

[0032] In one embodiment, the disease probability recognition convolutional neural network of the disease probability calculation module is constructed in the following manner:

[0033] Obtain normal region templates and lesion region templates, as well as the image feature templates corresponding to the normal region templates and lesion region templates;

[0034] A 3×3 convolution kernel is selected, and the convolutional layer parameters of the 3×3 convolution kernel are set based on the image feature template to obtain the configured convolution kernel;

[0035] Construct comparative learning samples based on normal region templates and lesion region templates;

[0036] Based on the contrastive learning samples, the configured convolutional kernels are trained using the triplet loss function to obtain the feature extraction layer;

[0037] Based on the feature extraction layer, a feature enhancement layer with a channel attention mechanism is embedded to obtain the first neural network;

[0038] According to the preset element weighted summation rules, a feature fusion layer is added to the first neural network to obtain the second neural network;

[0039] Obtain the target disease category of nasopharyngeal diseases in children, and configure a fully connected layer with N neurons for the second neural network according to the target disease category to obtain the third neural network;

[0040] Based on the third neural network, a probability output layer is constructed using the Softmax activation function, and the probability output layer is integrated with the third neural network to form a disease probability recognition convolutional neural network.

[0041] In one embodiment, the disease probability calculation module is further used for:

[0042] Image features of potential lesion areas are input into a disease probability recognition convolutional neural network, and deep lesion features are extracted through a feature extraction layer.

[0043] Based on the characteristics of deep lesions, a channel attention mechanism of the feature enhancement layer is used to enhance the expression of lesion features and form enhanced deep lesion features.

[0044] Based on the enhanced features of deep lesions, combined with image features, the element-weighted summation is performed through the element-weighted summation rules of the feature fusion layer to obtain the weighted fusion features;

[0045] Based on weighted fusion features, a fully connected layer is used to perform dimensional mapping transformation of the target disease category to obtain disease category association features;

[0046] Based on disease category association features, the probability of the target disease is calculated using the Softmax activation function of the probability output layer.

[0047] Secondly, this application also provides a method for analyzing pediatric nasopharyngeal diseases based on endoscopic images, including:

[0048] Acquire endoscopic images of the nasopharynx and larynx of children, and perform spatiotemporal filtering on the endoscopic images to obtain filtered images;

[0049] A motion artifact removal algorithm is used to remove motion artifacts from the filtered image, resulting in a stable image.

[0050] Image features are extracted from stable images;

[0051] Based on image features, image feature mutation indexes are calculated, and potential lesion areas are located and identified according to the image feature mutation indexes.

[0052] Based on potential lesion areas, image features are input into a pre-defined disease probability recognition convolutional neural network to generate the target disease probability;

[0053] Integrate stable images, image features, and target disease probabilities to generate disease diagnosis reports.

[0054] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to realize the functions of the above-mentioned endoscopic image-based pediatric nasopharyngeal disease analysis system.

[0055] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the functions of the above-mentioned endoscopic image analysis system for pediatric nasopharyngeal diseases.

[0056] The aforementioned endoscopic image-based pediatric nasopharyngeal disease analysis system and method preprocesses endoscopic images through an image spatiotemporal filtering module, and combines this with a motion artifact removal algorithm in the image artifact removal module to suppress image blurring and distortion caused by poor cooperation from children, thereby improving image quality. The image feature extraction module and the feature lesion localization module automatically and accurately identify potential lesion areas by calculating image feature mutation indices, reducing reliance on physician experience and avoiding subjective misjudgments. The disease probability calculation module uses a pre-set disease probability recognition convolutional neural network to analyze image features and generate objective target disease probabilities, improving diagnostic reliability. The disease report generation module automatically integrates information to generate diagnostic reports, improving diagnostic efficiency. This system achieves automated and objective analysis of pediatric nasopharyngeal diseases. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a structural diagram of the pediatric nasopharyngeal disease analysis system based on endoscopic images according to the present invention;

[0059] Figure 2 This is a flowchart of the endoscopic image-based method for analyzing pediatric nasopharyngeal diseases according to the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0061] In one embodiment, such as Figure 1 As shown, a pediatric nasopharyngeal disease analysis system based on endoscopic images is provided. This embodiment illustrates the system deployed on a terminal. It is understood that the system can also be deployed on a server, or in an architecture including both a terminal and a server, and is implemented through interaction between the terminal and the server. In a real-time environment (using a terminal as an example), the terminal is typically a medical imaging workstation or a portable computer, directly connected to the pediatric nasopharyngeal endoscopy equipment via USB or a high-speed video interface, meeting the needs of real-time image processing and immediate diagnosis in outpatient clinics or operating rooms. Application scenarios include: After the endoscopic device acquires video streams of the nasopharynx and larynx of a child, the terminal receives the raw image data and performs the following collaborative processing using the deployed system: the image spatiotemporal filtering module performs noise reduction preprocessing on the raw image, and the output filtered image is subjected to motion compensation correction by the image artifact removal module to generate a stable image; the image feature extraction module extracts feature data such as color and texture from the stable image; the feature lesion localization module locates the coordinates of potential lesion areas based on these feature data; the disease probability calculation module calls a locally pre-trained convolutional neural network model to analyze the features of the lesion area and calculate the disease probability; the disease report generation module integrates all intermediate results and probability information, generates a graphic report, and renders it on the terminal display to assist doctors in making quick and objective diagnostic decisions.

[0062] In this embodiment, the system includes:

[0063] The image spatiotemporal filtering module 101 is used to acquire endoscopic images of the nasopharynx and larynx of children and perform spatiotemporal filtering on the endoscopic images to obtain filtered images.

[0064] Among them, the endoscopic images are optical image sequences or single-frame images of the nasopharyngeal region acquired through endoscopic equipment. Spatiotemporal filtering is a filtering technique that combines the time and spatial domains, aiming to suppress random noise and motion blur in the images and improve image quality. In practice, it can be achieved by acquiring a continuous frame image sequence and using spatiotemporal filters (such as recursive filters based on motion compensation or three-dimensional convolutional filters) to smooth the pixel values, reduce spatiotemporal inconsistencies caused by children's movements or equipment shaking, and output filtered images with reduced noise and enhanced details, providing a stable foundation for subsequent processing.

[0065] The image artifact removal module 102 is used to remove motion artifacts from the filtered image using a motion artifact removal algorithm to obtain a stable image.

[0066] Motion artifacts refer to image blurring, ghosting, or distortion caused by the relative motion between the endoscopic equipment and tissues due to poor cooperation from children. Motion artifact removal algorithms are motion compensation and correction techniques based on image analysis. In practice, motion artifacts can be detected and quantified by analyzing the pixel gradient, optical flow, or motion vectors of local feature points in the filtered image. A motion model is constructed to estimate the offset direction and amplitude of the artifacts. Then, pixel compensation, image registration, or inverse transformation are used to correct and repair the artifact regions, outputting a stable image with effectively suppressed motion artifacts and clear anatomical structures, providing a reliable foundation for subsequent feature extraction.

[0067] Image feature extraction module 103 is used to extract image features from a stable image.

[0068] Among them, image features are digital visual attributes that can characterize the physiological state and pathological changes of nasopharyngeal and laryngeal tissues, including at least color features, texture features, and morphological features. In implementation, image analysis techniques such as statistical models, filter banks, morphological operations, or deep learning feature maps can be used to quantitatively analyze pixels or regions in stable images to calculate feature data such as color distribution histograms, local binary patterns, gradient direction histograms, or region contour descriptors, and convert image information into structured feature representations that can be used for subsequent lesion localization and disease identification.

[0069] The feature lesion localization module 104 is used to calculate the image feature mutation index based on image features, and to locate and identify potential lesion areas based on the image feature mutation index.

[0070] Among them, the image feature mutation index is a parameter used to quantify the degree of difference between the feature values ​​of a local area of ​​an image and the feature values ​​of the surrounding normal tissue. In implementation, the image features can be scanned and analyzed by statistical difference analysis, gradient detection or machine learning classifier, the feature mutation coefficient of each local area can be calculated, and the coefficient can be compared with a preset threshold to identify blocks with abnormal feature distribution. Adjacent abnormal blocks can be merged and the boundaries optimized by image segmentation techniques such as region growing, contour extraction or cluster analysis, and potential lesion areas used to identify suspected lesion locations can be output.

[0071] The disease probability calculation module 105 is used to input image features into a preset disease probability recognition convolutional neural network based on potential lesion areas to generate the target disease probability.

[0072] Among them, the disease probability recognition convolutional neural network is a trained deep learning model used to learn the nonlinear mapping relationship between image features and specific diseases. The target disease probability is the quantitative value of the possibility that a potential lesion area belongs to one or more preset disease categories. In implementation, the image features corresponding to the potential lesion area can be used as input, and forward propagation and feature transformation can be performed through multiple convolutional layers, pooling layers and fully connected layers contained in the convolutional neural network. The probability distribution corresponding to each disease category can be calculated through the activation function of the output layer to obtain an objective and quantitative target disease probability for auxiliary diagnosis.

[0073] The disease report generation module 106 is used to integrate stable images, image features, and target disease probabilities to generate disease diagnosis reports.

[0074] Among them, the disease diagnosis report is a structured output document that integrates original image evidence, quantitative feature analysis, and disease identification conclusions. In implementation, information fusion and report generation technologies can be used to take stable images as basic visual evidence, take the image features reflecting tissue characteristics as the basis for quantitative analysis, and integrate them with the target disease probability indicating the disease type. Using methods such as template filling, automatic annotation, or natural language generation, a comprehensive graphic report containing lesion area annotation, feature data summary, disease probability judgment, and diagnostic suggestions can be generated, providing doctors with intuitive and comprehensive diagnostic decision support.

[0075] The aforementioned endoscopic image-based pediatric nasopharyngeal disease analysis system utilizes spatiotemporal filtering and motion artifact removal techniques to effectively suppress image blurring and distortion caused by children's movements, thereby improving image quality. Through automated feature extraction and lesion localization, it reduces reliance on doctors' experience and avoids subjective misjudgments. By using convolutional neural networks to generate objective disease probabilities and automatically integrating them to generate reports, it improves the overall efficiency, consistency, and reliability of diagnosis.

[0076] In one embodiment, the image artifact removal module 102 is further configured to:

[0077] Pixel gradient information is obtained by calculating pixel gradients based on the filtered image.

[0078] Based on pixel gradient information, the filtered image is divided using an adaptive threshold segmentation method to obtain motion artifact regions and the corresponding set of artifact pixels.

[0079] For motion artifact regions, a neighborhood pixel gradient consistency check is performed on the artifact pixel set to obtain the consistency check result;

[0080] Based on the consistency check results, pixels in the artifact pixel set that do not meet the preset consistency conditions are removed to obtain the checked artifact pixel set.

[0081] Based on the verified set of artifact pixels, the artifact offset parameters are fitted by constructing a motion trajectory model.

[0082] By combining the artifact offset parameters, pixel-wise inverse compensation correction is performed on the motion artifact region to obtain a preliminary corrected image;

[0083] A gradient distribution equalization method is used to optimize the pixel transition continuity of the initially corrected image and generate a stable image.

[0084] Specifically, the Sobel, Prewitt, or Scharr operators can be used to calculate the gradient of each pixel in the filtered image. After obtaining the horizontal and vertical gradient components, the gradient magnitude of the pixel is obtained by taking the square root of the sum of the squares. The gradient magnitudes and direction information of all pixels together constitute the pixel gradient information. Based on this pixel gradient information, an adaptive threshold is calculated for each pixel. Pixels that meet this adaptive threshold are considered artifact pixels. All artifact pixels form an artifact pixel set. The continuous or discrete distribution area formed by these artifact pixels in the filtered image is the motion artifact region. For the artifact pixel set in the motion artifact region, a 3×3 or 5×5 local neighborhood window is constructed with each artifact pixel as the center. The absolute value of the difference between the gradient of this pixel and the gradients of other artifact pixels in the neighborhood is calculated. The proportion of pixels with an absolute difference less than a preset consistency threshold is counted to the total number of artifact pixels in the neighborhood. If the proportion reaches the preset threshold, the pixel is considered to have passed the consistency check; otherwise, it has failed. The check results of all pixels are summarized to form the consistency check result. From the artifact pixel set... Pixels that fail the consistency check are removed, and the remaining pixels that pass the check form a set of checked artifact pixels to eliminate false artifact pixels caused by noise interference. Based on the set of checked artifact pixels, if it is an image sequence, a linear motion model or a quadratic polynomial motion model can be used. Using the coordinate changes of artifact pixels in consecutive frames as the basic data, the motion direction, velocity, and offset distance can be obtained by fitting the data using the least squares method or gradient descent method. If it is a single frame image, the offset parameters can be fitted by combining the gradient distribution of the artifact pixel set with the positional relationship of the neighboring normal pixels. The ideal artifact-free coordinates of each artifact pixel are determined according to the artifact offset parameters. The ideal coordinates and the grayscale or color values ​​of the surrounding pixels are extracted from the filtered image by bilinear interpolation or bicubic interpolation. The artifact pixels are then adjusted by reverse compensation to obtain a preliminary corrected image. The gradient histogram of the preliminary corrected image is calculated and normalized. By adjusting the gradient magnitude of different regions, the gradient changes of adjacent pixels tend to be smooth, reducing the pixel transition abruptness caused by correction, and generating a stable image with continuous pixel transitions and no obvious motion artifacts.

[0085] In one embodiment, the image artifact removal module 102 is further configured to determine the motion artifact region using the following formula:

[0086]

[0087] in, For pixel gradient, In pixels A local neighborhood window centered on the center, For local neighborhood windows The mean of the pixel gradient, For local neighborhood windows The standard deviation of the pixel gradient, For local neighborhood windows The skewness of the pixel gradient. , , and This refers to a preset adaptive adjustment coefficient calibrated based on the imaging characteristics of pediatric nasopharyngeal endoscopy images and the intensity range of common motion artifacts. This represents the global maximum value of the pixel gradient. To preset the high gradient threshold range, The number of pixels.

[0088] For example, a local neighborhood window can be determined. The size of each pixel (usually set to 3×3 or 5×5, which can be adaptively adjusted according to the detail density of pediatric nasopharyngeal endoscopy images) is obtained using the Sobel operator isogradient calculation method. pixel gradient Calculate the arithmetic mean of the gradients of all pixels within the local neighborhood window as... The square root of the average of the sum of squared differences between the pixel gradients and the mean within the window is obtained. The result is obtained by dividing the third central moment by the cube of the standard deviation. ; Traverse the entire filtered image and obtain the maximum value among all pixel gradients as Based on the gradient distribution characteristics of motion artifacts in pediatric nasopharyngeal endoscopy images, a preset is made. (For example, setting the gradient range to 50-255 adapts to the common range of 8-bit grayscale levels in endoscopic images), statistical local neighborhood window. The gradient value falls within The number of pixels within the interval is obtained And use the total number of pixels in the local neighborhood window as The four preset adaptive adjustment coefficients α, β, γ, and δ are statistically calibrated using a large number of pediatric nasopharyngeal endoscopy image samples containing known motion artifact regions and normal regions. Optimal values ​​are determined using cross-validation (e.g., α = 1.0, β = 0.4, γ = 0.8, δ = 0.2, fine-tuned according to the imaging characteristics of different endoscopes). Substituting the calculated parameters into the right side of the formula, the adaptive threshold corresponding to each pixel (x, y) is obtained through exponential, multiplicative, addition, and division operations. The gradient of this pixel is then calculated. Compared with the adaptive threshold, if If the value is greater than the adaptive threshold, the pixel is determined to be an artifact pixel. All pixels determined to be artifact pixels together constitute the artifact pixel set. The continuous or discrete distribution area formed by these artifact pixels in the filtered image is the motion artifact region.

[0089] In one embodiment, the image features include color features, texture features, and morphological features, and the feature lesion localization module 104 is further used for:

[0090] For image features, the feature standard deviation thresholding method is used to calculate the block abrupt change coefficients of color features, texture features, and morphological features;

[0091] Based on the block mutation coefficient and combined with the preset mutation coefficient threshold, color mutation blocks, texture mutation blocks and shape mutation blocks are located.

[0092] Based on color mutation blocks, texture mutation blocks, and morphological mutation blocks, a neighborhood connectivity analysis method is used to perform block connectivity processing to obtain potential lesion blocks.

[0093] Based on potential lesion blocks, morphological closing operations are used to optimize the block boundaries and form potential lesion regions.

[0094] Specifically, color features can be obtained by extracting the mean, variance, and proportion of each channel component from the color histogram in the RGB or HSV color space. Texture features can be obtained by calculating the gray-level difference statistics of neighboring pixels using Local Binary Mode (LBP) or by obtaining parameters such as contrast, energy, and entropy through the Gray-Level Co-occurrence Matrix (GLCM). Morphological features can be obtained by extracting the region contour using edge detection algorithms (such as the Canny operator) and then calculating the geometric parameters such as the area, perimeter, roundness, and rectangularity of the contour. Using the feature standard deviation thresholding method, the stable image is uniformly divided into fixed-size sub-blocks of 16×16 pixels or 32×32 pixels. The standard deviation of the feature values ​​for color, texture, and morphological features within each sub-block is calculated. Using the global standard deviation of the feature in the entire stable image as a benchmark, the standard deviation is calculated using the formula:

[0095]

[0096] The block mutation coefficient corresponding to each feature is calculated. Based on a large number of normal and diseased endoscopic images of the nasopharynx and larynx in children, preset mutation coefficient thresholds are statistically calibrated (e.g., color feature threshold is set to 1.5, texture feature threshold to 1.8, and morphological feature threshold to 2.0, which can be fine-tuned according to the actual imaging accuracy). The block mutation coefficient of each sub-block is compared with the corresponding preset threshold. If the color feature mutation coefficient of a sub-block exceeds the color preset threshold, it is determined to be a color mutation block. Similarly, texture mutation blocks and morphological mutation blocks are located. An 8-neighborhood connectivity analysis method can be used to traverse its neighbors, starting from each mutation block. If adjacent sub-blocks in the top, bottom, left, right, and diagonal directions are any type of mutation blocks, they are determined to be connected and merged into a single block. This process is repeated until all connected mutation blocks are merged, resulting in multiple independent potential lesion blocks. Morphological closing operations can be performed on each potential lesion block using 3×3 or 5×5 rectangular structural elements. The dilation operation fills the tiny voids inside the block and connects the small fractures. Then, the erosion operation restores the main outline of the block and smooths the boundaries, eliminating burrs and irregular protrusions at the block edges, forming a potential lesion area with clear boundaries and complete regions.

[0097] In one embodiment, the disease report generation module 106 is further configured to:

[0098] Determine the target disease category based on the probability of the target disease;

[0099] Based on the target disease category, the severity level of the disease is calculated by using a preset severity grading index, combined with color features, texture features, and morphological features.

[0100] Based on the target disease category and disease severity level, disease information is annotated in potential lesion areas of stable images to obtain an annotated image, which is then used as a disease diagnosis report.

[0101] For example, the probabilities of each target disease output by the convolutional neural network for disease probability recognition can be sorted, and the disease category with the highest probability value that exceeds a preset probability threshold (e.g., 0.8, adjusted according to clinical diagnostic accuracy requirements) can be selected as the target disease category. For the determined target disease category, preset corresponding disease severity grading indicators can be called (based on clinical treatment guidelines and a large number of case data; for example, for allergic rhinitis, the grading indicators include the range of HSV color space difference between the diseased mucosa and normal mucosa in color features, the degree of disorder of LBP statistical value of mucosal texture in texture features, and the nasal passage stenosis rate in morphological features; for tonsillitis, the indicators include the proportion of RGB red light components in the tonsil congestion area in color features, the gray-level co-occurrence matrix entropy value of surface follicles in texture features, and the proportion of airway cross-sectional area caused by tonsil enlargement in morphological features). Preset weights are assigned to each feature (e.g., 0.3 for color features, 0.3 for texture features, and 0.4 for morphological features), using the following formula:

[0102]

[0103] Calculate a comprehensive score and determine the severity level of the disease based on the score range (e.g., 0-30 for mild, 31-60 for moderate, and 61-100 for severe). In stable images, potential lesion areas can be selected using a red solid-line rectangle. Disease information can be labeled next to the rectangle, including the target disease category name (e.g., "allergic rhinitis" or "tonsillitis with adenoid hypertrophy"), severity level (e.g., "moderate"), and quantitative data of key features (e.g., "lesion area 2.3 mm², mucosal color difference 18, texture disorder 0.65"). Image acquisition time and basic patient information can be labeled in the corners of the image. An annotated image containing clear lesion location, a clear diagnosis, and quantitative data of key features is generated. This annotated image is a disease diagnosis report that can be directly used for clinical reference.

[0104] In one embodiment, the disease probability calculation module 105's disease probability recognition convolutional neural network is constructed in the following manner:

[0105] Obtain normal region templates and lesion region templates, as well as the image feature templates corresponding to the normal region templates and lesion region templates;

[0106] A 3×3 convolution kernel is selected, and the convolutional layer parameters of the 3×3 convolution kernel are set based on the image feature template to obtain the configured convolution kernel;

[0107] Construct comparative learning samples based on normal region templates and lesion region templates;

[0108] Based on the contrastive learning samples, the configured convolutional kernels are trained using the triplet loss function to obtain the feature extraction layer;

[0109] Based on the feature extraction layer, a feature enhancement layer with a channel attention mechanism is embedded to obtain the first neural network;

[0110] According to the preset element weighted summation rules, a feature fusion layer is added to the first neural network to obtain the second neural network;

[0111] Obtain the target disease category of nasopharyngeal diseases in children, and configure a fully connected layer with N neurons for the second neural network according to the target disease category to obtain the third neural network;

[0112] Based on the third neural network, a probability output layer is constructed using the Softmax activation function, and the probability output layer is integrated with the third neural network to form a disease probability recognition convolutional neural network.

[0113] Specifically, from a database of clinically confirmed pediatric nasopharyngeal endoscopy images, 1000-2000 images containing normal nasopharyngeal tissue regions can be selected as normal region templates, and images of lesion tissues covering common diseases such as allergic rhinitis, tonsillitis, and adenoid hypertrophy can be selected as lesion region templates. Using the color, texture, and morphological feature extraction methods described above, feature data corresponding to the two types of templates are extracted and organized into structured image feature templates. A 3×3 convolution kernel (balancing local feature capture capability and computational efficiency) is selected based on image features... The template's feature dimensions (e.g., 3-channel color + 5-dimensional texture + 4-dimensional morphology forming a 12-dimensional feature space) are used. The convolution kernel's input channels are set to 12 and output channels to 64 (adjustable to 128 depending on feature complexity). The kernel parameters are initialized using a He normal distribution to obtain the configured kernel. Comparative learning samples are constructed based on normal and lesion region templates. Each sample is a triplet structure of "anchor sample - positive sample - negative sample," where the anchor sample is an image feature of any type of template, and the positive sample is of the same type as the anchor sample (same as normal). Other template features (or lesions of the same disease) are used, and negative samples are template features of a different type than the anchor samples. The training set, validation set, and test set are divided in a 7:2:1 ratio. Based on the contrastive learning samples, the configured convolutional kernels are trained using a triplet loss function. During training, the initial learning rate is set to 0.001, the Adam optimizer is used, the batch size is 32, and the training is iterated for 50 rounds until the validation set loss converges. The 3-5 convolutional layers and pooling layers after training are stacked to form a feature extraction layer. Based on this feature extraction layer, a channel attention mechanism with an SENet structure is embedded. The feature enhancement layer performs global average pooling on the feature map output by the feature extraction layer to obtain a global descriptor for each channel. It then learns the importance weights of each channel through two fully connected layers (with the number of hidden neurons being 1 / 4 of the number of channels). After normalization using the sigmoid activation function, these weights are multiplied channel-by-channel with the original feature map to enhance the expression of key features related to the lesion, resulting in the first neural network. A feature fusion layer is then added to the first neural network according to a preset element-weighted summation rule. The preset weights are calibrated through extensive sample experiments (0.7 for deep enhancement features and 0 for original image features).3) The deep features output from the feature enhancement layer are multiplied element-wise with the original image features according to their corresponding weights and then summed to achieve complementary fusion of deep and shallow features, forming a second neural network. Common target disease categories in children's nasopharynx and larynx are obtained (e.g., allergic rhinitis, tonsillitis, adenoid hypertrophy, nasal polyps, and pharyngitis, 5 categories). The N value is set to 5, and a fully connected layer with 5 neurons is configured for the second neural network. The input dimension of this fully connected layer is consistent with the feature dimension output from the feature fusion layer (e.g., 128 dimensions). The fused features are mapped to the disease category dimension through linear transformation to obtain the third neural network. Based on the third neural network, a probability output layer can be constructed using the Softmax activation function. This function converts the linear vector output from the fully connected layer into a probability vector where each element takes values ​​between 0 and 1 and the sum is 1. Each element corresponds to the probability of a target disease category. This probability output layer is connected end-to-end with the third neural network to form a disease probability recognition convolutional neural network.

[0114] In one embodiment, the disease probability calculation module 105 is further configured to:

[0115] Image features of potential lesion areas are input into a disease probability recognition convolutional neural network, and deep lesion features are extracted through a feature extraction layer.

[0116] Based on the characteristics of deep lesions, a channel attention mechanism of the feature enhancement layer is used to enhance the expression of lesion features and form enhanced deep lesion features.

[0117] Based on the enhanced features of deep lesions, combined with image features, the element-weighted summation is performed through the element-weighted summation rules of the feature fusion layer to obtain the weighted fusion features;

[0118] Based on weighted fusion features, a fully connected layer is used to perform dimensional mapping transformation of the target disease category to obtain disease category association features;

[0119] Based on disease category association features, the probability of the target disease is calculated using the Softmax activation function of the probability output layer.

[0120] For example, the color, texture, and morphological features corresponding to potential lesion areas can be quantized and integrated into a 12-dimensional feature vector (3-dimensional for color, 5-dimensional for texture, and 4-dimensional for morphology). This feature vector is then input into a pre-defined disease probability recognition convolutional neural network. Through a feature extraction layer consisting of 3-5 stacked convolutional and pooling layers, multi-scale convolution operations and downsampling are performed on the input feature vector to extract deep lesion features that characterize the essential attributes of the lesion (such as abstract texture patterns, color distribution patterns, and morphological features of the lesion tissue), forming a 128-dimensional deep feature vector. Based on this deep lesion feature, the channel attention mechanism of the SENet structure in the feature enhancement layer is used to perform global average pooling on the deep feature vector to obtain global statistical information for each channel. The importance weights of each channel are learned through two fully connected layers (32 hidden neurons). After the weights are normalized to the 0-1 range using the Sigmoid activation function, they are multiplied channel-by-channel with the original deep lesion feature vector to strengthen the expression of channel features highly correlated with the lesion and suppress interference from irrelevant channels, forming a 128-dimensional deep feature vector. The enhanced deep lesion features remain at 128 dimensions. Based on a preset element-weighted summation rule (where the weight of the enhanced deep lesion features is set to 0.7 and the weight of the original image feature vector is set to 0.3), the enhanced deep lesion features and the original 12-dimensional image feature vector are dimensionally aligned (e.g., the original image feature vector is expanded to 128 dimensions through zero padding), and then the corresponding weights are multiplied element by element and summed to obtain a 128-dimensional weighted fusion feature that combines the advantages of both deep and superficial features. Based on this weighted fusion feature, a target with N neurons (N equal to the target of the pediatric nasopharynx) is used. A fully connected layer (N=5) with 5 disease categories maps 128-dimensional weighted fusion features to an N-dimensional space through linear transformation, resulting in disease category association features where each dimension corresponds to the degree of association with a target disease. Based on these disease category association features, the Softmax activation function of the probability output layer is used to calculate the proportion of each dimension's feature value to the sum of all dimension feature values, converting it into a probability vector with values ​​between 0 and 1 and a sum of 1. The value of each element in this probability vector is the target disease probability for the corresponding target disease category.

[0121] The aforementioned endoscopic image-based pediatric nasopharyngeal disease analysis system preprocesses endoscopic images of children's nasopharyngeal endoscopic lesions through an image spatiotemporal filtering module. Combined with pixel gradient calculation, adaptive threshold segmentation, neighborhood consistency verification, and motion trajectory fitting techniques from the image artifact removal module, motion artifacts are eliminated and image stability is optimized, addressing image blurring and distortion caused by low cooperation from children. The system also extracts multi-dimensional features of color, texture, and morphology through an image feature extraction module. Furthermore, the feature lesion localization module automates lesion localization based on feature mutation indices and morphological closing operations, avoiding the reliance on physician experience, strong subjectivity, and susceptibility to errors inherent in manual analysis. The module addresses the shortcomings of overlooking minute lesions. It employs a convolutional neural network trained through contrastive learning, incorporating channel attention mechanisms and feature fusion layers to integrate multi-feature information for quantitative disease probability calculation, thus improving diagnostic reliability and overcoming the limitations of existing technologies in comprehensively evaluating multiple features. The disease report generation module automatically integrates stable images, feature data, and disease probabilities to generate diagnostic reports with lesion annotations and severity grading, constructing an efficient automated analysis process that significantly reduces diagnostic time, meets the demands of high clinical efficiency, and automates and objectifies the analysis of pediatric nasopharyngeal diseases, improving diagnostic efficiency and consistency.

[0122] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0123] Based on the same inventive concept, this application also provides a method for implementing the endoscopic image-based pediatric nasopharyngeal disease analysis system for pediatric nasopharyngeal diseases described above. The solution provided by this method is similar to the implementation described in the above system. Therefore, the specific limitations of one or more embodiments of the endoscopic image-based pediatric nasopharyngeal disease analysis method provided below can be found in the limitations of the endoscopic image-based pediatric nasopharyngeal disease analysis system described above, and will not be repeated here.

[0124] In one exemplary embodiment, such as Figure 2 As shown, a method for analyzing pediatric nasopharyngeal diseases based on endoscopic images is provided, including:

[0125] S01, acquire endoscopic images of the nasopharynx and larynx of children, and perform spatiotemporal filtering on the endoscopic images to obtain filtered images;

[0126] S02, a motion artifact removal algorithm is used to remove motion artifacts from the filtered image to obtain a stable image;

[0127] S03, extract image features from a stable image;

[0128] S04. Based on image features, the image feature mutation index is calculated, and the potential lesion area is located and identified according to the image feature mutation index.

[0129] S05, based on the potential lesion area, input the image features into the preset disease probability recognition convolutional neural network to generate the target disease probability;

[0130] S06 integrates stable images, image features, and target disease probabilities to generate disease diagnosis reports.

[0131] In one embodiment, the image artifact removal module is further configured to:

[0132] Pixel gradient information is obtained by calculating pixel gradients based on the filtered image.

[0133] Based on pixel gradient information, the filtered image is divided using an adaptive threshold segmentation method to obtain motion artifact regions and the corresponding set of artifact pixels.

[0134] For motion artifact regions, a neighborhood pixel gradient consistency check is performed on the artifact pixel set to obtain the consistency check result;

[0135] Based on the consistency check results, pixels in the artifact pixel set that do not meet the preset consistency conditions are removed to obtain the checked artifact pixel set.

[0136] Based on the verified set of artifact pixels, the artifact offset parameters are fitted by constructing a motion trajectory model.

[0137] By combining the artifact offset parameters, pixel-wise inverse compensation correction is performed on the motion artifact region to obtain a preliminary corrected image;

[0138] A gradient distribution equalization method is used to optimize the pixel transition continuity of the initially corrected image and generate a stable image.

[0139] In one embodiment, the method determines the motion artifact region using the following formula:

[0140]

[0141] in, For pixel gradient, In pixels A local neighborhood window centered on the center, For local neighborhood windows The mean of the pixel gradient, For local neighborhood windows The standard deviation of the pixel gradient, For local neighborhood windows The skewness of the pixel gradient. , , and This refers to a preset adaptive adjustment coefficient calibrated based on the imaging characteristics of pediatric nasopharyngeal endoscopy images and the intensity range of common motion artifacts. This represents the global maximum value of the pixel gradient. To preset the high gradient threshold range, The number of pixels.

[0142] In one embodiment, the method further includes:

[0143] For image features, the feature standard deviation thresholding method is used to calculate the block abrupt change coefficients of color features, texture features, and morphological features;

[0144] Based on the block mutation coefficient and combined with the preset mutation coefficient threshold, color mutation blocks, texture mutation blocks and shape mutation blocks are located.

[0145] Based on color mutation blocks, texture mutation blocks, and morphological mutation blocks, a neighborhood connectivity analysis method is used to perform block connectivity processing to obtain potential lesion blocks.

[0146] Based on potential lesion blocks, morphological closing operations are used to optimize the block boundaries and form potential lesion regions.

[0147] In one embodiment, the method further includes:

[0148] Determine the target disease category based on the probability of the target disease;

[0149] Based on the target disease category, the severity level of the disease is calculated by using a preset severity grading index, combined with color features, texture features, and morphological features.

[0150] Based on the target disease category and disease severity level, disease information is annotated in potential lesion areas of stable images to obtain an annotated image, which is then used as a disease diagnosis report.

[0151] In one embodiment, the disease probability recognition convolutional neural network of the method is constructed in the following manner:

[0152] Obtain normal region templates and lesion region templates, as well as the image feature templates corresponding to the normal region templates and lesion region templates;

[0153] A 3×3 convolution kernel is selected, and the convolutional layer parameters of the 3×3 convolution kernel are set based on the image feature template to obtain the configured convolution kernel;

[0154] Construct comparative learning samples based on normal region templates and lesion region templates;

[0155] Based on the contrastive learning samples, the configured convolutional kernels are trained using the triplet loss function to obtain the feature extraction layer;

[0156] Based on the feature extraction layer, a feature enhancement layer with a channel attention mechanism is embedded to obtain the first neural network;

[0157] According to the preset element weighted summation rules, a feature fusion layer is added to the first neural network to obtain the second neural network;

[0158] Obtain the target disease category of nasopharyngeal diseases in children, and configure a fully connected layer with N neurons for the second neural network according to the target disease category to obtain the third neural network;

[0159] Based on the third neural network, a probability output layer is constructed using the Softmax activation function, and the probability output layer is integrated with the third neural network to form a disease probability recognition convolutional neural network.

[0160] In one embodiment, the method further includes:

[0161] Image features of potential lesion areas are input into a disease probability recognition convolutional neural network, and deep lesion features are extracted through a feature extraction layer.

[0162] Based on the characteristics of deep lesions, a channel attention mechanism of the feature enhancement layer is used to enhance the expression of lesion features and form enhanced deep lesion features.

[0163] Based on the enhanced features of deep lesions, combined with image features, the element-weighted summation is performed through the element-weighted summation rules of the feature fusion layer to obtain the weighted fusion features;

[0164] Based on weighted fusion features, a fully connected layer is used to perform dimensional mapping transformation of the target disease category to obtain disease category association features;

[0165] Based on disease category association features, the probability of the target disease is calculated using the Softmax activation function of the probability output layer.

[0166] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the functions of a pediatric nasopharyngeal disease analysis system based on endoscopic images as described above.

[0167] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above system embodiments.

[0168] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0169] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A pediatric nasopharyngeal disease analysis system based on endoscopic images, characterized in that, The system includes: The image spatiotemporal filtering module is used to acquire endoscopic images of the nasopharynx and larynx of children and perform spatiotemporal filtering on the endoscopic images to obtain filtered images; The image artifact removal module is used to remove motion artifacts from the filtered image using a motion artifact removal algorithm to obtain a stable image. An image feature extraction module is used to extract image features from the stable image; The feature lesion localization module is used to calculate the image feature mutation index based on the image features, and locate and identify potential lesion areas according to the image feature mutation index. The disease probability calculation module is used to input the image features into a preset disease probability recognition convolutional neural network based on the potential lesion area to generate the target disease probability. The disease report generation module is used to integrate the stable image, the image features, and the target disease probability to generate a disease diagnosis report.

2. The system according to claim 1, characterized in that, The image artifact removal module is also used for: Based on the filtered image, pixel gradient calculation is performed to obtain pixel gradient information; Based on the pixel gradient information, the filtered image is divided using an adaptive threshold segmentation method to obtain motion artifact regions and the set of artifact pixels corresponding to the motion artifact regions. For the motion artifact region, a neighborhood pixel gradient consistency check is performed on the artifact pixel set to obtain the consistency check result; Based on the consistency verification result, pixels in the artifact pixel set that do not meet the preset consistency condition are removed to obtain the verified artifact pixel set. Based on the verified set of artifact pixels, the artifact offset parameters are fitted by constructing a motion trajectory model. Based on the artifact offset parameters, pixel-wise inverse compensation correction is performed on the motion artifact region to obtain a preliminary corrected image; A gradient distribution equalization method is used to optimize the pixel transition continuity of the preliminary corrected image, thereby generating the stable image.

3. The system according to claim 2, characterized in that, The image artifact removal module is also used to determine motion artifact regions using the following formula: in, For pixel gradient, In pixels A local neighborhood window centered on the center, For local neighborhood windows The mean of the pixel gradient, For local neighborhood windows The standard deviation of the pixel gradient, For local neighborhood windows The skewness of the pixel gradient. , , and This refers to a preset adaptive adjustment coefficient calibrated based on the imaging characteristics of pediatric nasopharyngeal endoscopy images and the intensity range of common motion artifacts. The global maximum value of the pixel gradient. To preset the high gradient threshold range, The number of pixels.

4. The system according to claim 1, characterized in that, The image features include color features, texture features, and morphological features, and the feature lesion localization module is further used for: For the image features, the feature standard deviation thresholding method is used to calculate the block abrupt change coefficients of the color features, texture features, and morphological features; Based on the block mutation coefficient, and combined with a preset mutation coefficient threshold, color mutation blocks, texture mutation blocks, and morphological mutation blocks are located. Based on the color mutation blocks, texture mutation blocks, and morphological mutation blocks, a neighborhood connectivity analysis method is used to perform block connectivity processing to obtain potential lesion blocks. Based on the potential lesion blocks, morphological closing operations are used to optimize the block boundaries to form the potential lesion regions.

5. The system according to claim 4, characterized in that, The disease report generation module is also used for: Based on the probability of the target disease, the target disease category is determined; Based on the target disease category, the severity level of the disease is calculated by using a preset severity grading index, combined with the color features, texture features, and morphological features. Based on the target disease category and the disease severity level, disease information is annotated in the potential lesion area of ​​the stable image to obtain an annotated image, and the annotated image is used as the disease diagnosis report.

6. The system according to claim 5, characterized in that, The disease probability calculation module's disease probability recognition convolutional neural network is constructed in the following way: Obtain normal region templates and lesion region templates, as well as image feature templates corresponding to the normal region templates and the lesion region templates; A 3×3 convolution kernel is selected, and the convolutional layer parameters of the 3×3 convolution kernel are set based on the image feature template to obtain the configured convolution kernel; Based on the normal region template and the lesion region template, construct a comparative learning sample; Based on the contrastive learning samples, the configured convolutional kernel is trained using the triplet loss function to obtain the feature extraction layer; Based on the feature extraction layer, a feature enhancement layer with a channel attention mechanism is embedded to obtain the first neural network; According to a preset element weighted summation rule, a feature fusion layer is added to the first neural network to obtain the second neural network; Obtain the target disease category of nasopharyngeal diseases in children, and configure a fully connected layer with N neurons for the second neural network according to the target disease category to obtain the third neural network; Based on the third neural network, a probability output layer is constructed using the Softmax activation function, and the probability output layer is integrated with the third neural network to form the disease probability recognition convolutional neural network.

7. The system according to claim 6, characterized in that, The disease probability calculation module is also used for: The image features of the potential lesion area are input into the disease probability recognition convolutional neural network, and the deep lesion features are extracted through the feature extraction layer. Based on the deep lesion characteristics, the channel attention mechanism of the feature enhancement layer is used to enhance the expression of lesion characteristics and form enhanced deep lesion characteristics; Based on the enhanced deep lesion features and combined with the image features, the elements are weighted and summed according to the element weighting rules of the feature fusion layer to obtain the weighted fusion features; Based on the weighted fusion features, the fully connected layer is used to perform dimensional mapping transformation of the target disease category to obtain disease category association features; Based on the disease category association features, the probability of the target disease is calculated using the Softmax activation function of the probability output layer.

8. A method for analyzing pediatric nasopharyngeal diseases based on endoscopic images, characterized in that, The method includes: Spatiotemporal filtering was performed on endoscopic images of the nasopharynx and larynx of children to obtain filtered images; A motion artifact removal method is used to process the filtered image to obtain a stable image; Image features are extracted from the stable image, including color features, texture features, and morphological features; Based on the image features, potential lesion areas are located by calculating a preset image feature mutation index; Based on the potential lesion area, the image features are input into a disease probability recognition convolutional neural network to calculate the probability of the target disease; By integrating the stable image, the image features, and the target disease probability, a disease diagnosis report is generated.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it performs the functions of the system according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it performs the functions of the system according to any one of claims 1 to 7.

Images