A Method and System for Lesion Image Recognition Based on Deep Learning Models

By combining deep learning models with biochemical indicators to identify lesion images, this method addresses the diversity and complexity of lesion identification, achieving intelligent and automated lesion recognition and improving identification accuracy and diagnostic efficiency.

CN121459136BActive Publication Date: 2026-11-17JIANG SU AI YING YI LIAO KE JI YOU XIAN GONG SI
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202511920312.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-12-18
Publication Date
2026-11-17
Estimated Expiration
2045-12-18

AI Technical Summary

Technical Problem

Existing medical image recognition technologies struggle to effectively identify diverse and complex lesions, and the lack of sufficient integration between biochemical indicators and imaging data results in incomplete and inaccurate lesion identification.

Method used

A lesion image recognition method based on a deep learning model is adopted. High-quality training data is generated through image preprocessing and manual labeling. An LSTM model is established for lesion segmentation and feature extraction. Combined with biochemical index data, a comprehensive lesion recognition coefficient is generated to achieve intelligent and automated lesion recognition.

Benefits of technology

It improves the accuracy and reliability of lesion identification, simplifies doctors' workflow, provides richer information support, and enhances diagnostic efficiency and timeliness.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121459136B_ABST
    Figure CN121459136B_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for lesion image recognition based on a deep learning model, relating to the field of medical device technology. The method includes the following steps: acquiring medical images of various lesion types and preprocessing them to generate sample images; labeling lesion regions, extracting image feature parameters, and constructing a sample training dataset. Based on this dataset, a segmentation and recognition model is trained. The medical images to be recognized are preprocessed, and the trained segmentation and recognition model is used to identify and segment lesion regions. The predicted values ​​of image feature parameters are obtained by inputting them into a feature extraction model, and a biochemical discrimination influence index is calculated by combining patient biochemical index data. A comprehensive lesion recognition coefficient is generated based on the lesion texture complexity index, edge influence index, and biochemical discrimination influence index, and compared with a set threshold to complete lesion image recognition. This improves the accuracy and reliability of lesion recognition.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of medical device technology, specifically to a method and system for lesion image recognition based on a deep learning model. Background Technology

[0002] With the development of medical imaging technology, medical imaging plays an increasingly important role in the diagnosis, monitoring, and treatment of diseases. In recent years, computer vision and deep learning technologies have made significant progress in medical image analysis, especially in the automatic identification and segmentation of lesions. However, existing medical image recognition technologies still face many challenges.

[0003] First, the diversity and complexity of medical images make lesion identification a challenging task. Different types of lesions (such as tumors, inflammation, and other lesions) vary in shape, size, color, and texture, and are often accompanied by changes in surrounding tissue. This diversity makes it difficult for traditional image processing methods to effectively identify and segment lesion regions, thus affecting clinical treatment decisions.

[0004] Secondly, the correlation between biochemical markers and tumor lesions is also an important research area. Biochemical markers (such as erythrocyte sedimentation rate, white blood cell count, carcinoembryonic antigen (CEA) levels, and alpha-fetoprotein (AFP) levels have a significant impact on lesion characteristics and patient survival prognosis. However, current technologies mostly process imaging and biochemical data separately, failing to fully utilize their complementary nature. This lack of comprehensive analysis may result in incomplete and inaccurate lesion identification and assessment.

[0005] Therefore, to address the aforementioned issues, a deep learning-based medical image recognition method has been developed. Through image preprocessing, feature extraction, model training, and comprehensive analysis, this method can effectively improve the accuracy and reliability of lesion identification, which has significant clinical implications.

[0006] In the prior art, patent announcement number CN110232383B discloses a lesion image recognition method and system based on a deep learning model. This method, by sequentially segmenting the smallest organ / tissue image, recognizing the corresponding organ / tissue image, performing deep learning model prediction, and marking the prediction results on the acquired medical image, can replace doctors in identifying potential lesions in medical images and automatically mark lesions on the image to prompt further diagnosis and timely confirmation of the condition. This reduces the workload of doctors, facilitates timely diagnosis of lesions, and avoids delays in treatment, especially in the early detection of lesions. However, while this system can assist doctors in identifying potential lesions, the final clinical judgment still relies on the doctor's professional knowledge and experience. Furthermore, the performance of the deep learning model is highly dependent on the quality and quantity of training data. If the training data is insufficient or lacks diversity, the model may not be able to effectively identify all types of lesions. Therefore, relying solely on the model for judgment reduces the accuracy and effectiveness of the recognition system.

[0007] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for lesion image recognition based on a deep learning model, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A method for lesion image recognition based on a deep learning model, comprising the following steps:

[0011] Several medical images of different lesion types were acquired and preprocessed to obtain sample images. The lesion regions in the sample images were marked by manual labeling, and the image feature parameters of the lesion marked regions were obtained to generate a sample training dataset.

[0012] A segmentation and recognition model is established based on a sample training dataset. The sample images are used as input, and the labels of the lesion regions in the images and the image feature parameters are used as labels to train the segmentation and recognition model.

[0013] The medical image of the lesion to be identified is obtained and input into the trained segmentation and recognition model to obtain the predicted value of the image feature parameters of the medical image of the lesion to be identified, and at the same time, the biochemical index data of the patient corresponding to the medical image of the lesion to be identified.

[0014] Based on the obtained image feature parameter prediction values, the complexity of lesion texture and the impact of lesion edge are characterized, and the biochemical impact is evaluated based on biochemical index data;

[0015] Based on the complexity of lesion texture and the influence of lesion edges, combined with biochemical effects, a comprehensive lesion recognition coefficient is generated. The comprehensive lesion recognition coefficient is compared with the set lesion judgment threshold to generate different comparison results, and the corresponding lesion recognition judgment result is issued to complete the lesion image recognition.

[0016] Furthermore, the image feature parameters include color contrast, texture entropy, lesion edge intensity, lesion contour perimeter, and lesion contour area; the lesion regions in the sample images are marked manually, specifically by marking the images using a labeling tool, and the different lesion types specifically include hemorrhage, inflammation, and tumor;

[0017] The image preprocessing specifically includes: image enhancement and denoising preprocessing, wherein wavelet transform is used to denoise each lesion medical image, and bilateral filtering is used to enhance each lesion medical image.

[0018] The method for generating the training sample dataset is as follows: the labeled sample images are mapped one-to-one with the corresponding image feature parameters to form a corresponding grid, and the resulting grid is recorded as the training sample dataset.

[0019] Furthermore, a segmentation recognition model is established based on the sample training dataset. This model is constructed using an LSTM model, and activation functions and optimization algorithms are selected. The Tanh function is chosen as the activation function, and Adam is chosen as the optimization algorithm for the LSTM model. The formula for the Tanh function is:

[0020]

[0021] In the formula, Represents the Tanh function, with the independent variable... This represents the weighted sum of the neuron's inputs, that is, the result of the weighted sum of the inputs received by the neuron from the previous layer;

[0022] Simultaneously, the hyperparameters of the LSTM model are set, including: number of network layers, number of iterations, learning rate, batch size, number of training iterations, number of batches, and number of hidden layer neurons.

[0023] The network is set to a 5-layer network structure, the number of iterations is set to 200, the learning rate is set to 0.001, the batch size is set to 32, the number of training iterations is set to 100, the batch size is set to 128, and the number of hidden layer neurons is set to 32.

[0024] The input to the trained segmentation and recognition model is a medical image of the lesion, and the output is an image with the lesion region marked, as well as the corresponding predicted values ​​of image feature parameters, including the predicted values ​​of color contrast, texture entropy, lesion edge intensity, lesion contour perimeter, and lesion contour area.

[0025] Furthermore, based on the obtained predicted values ​​of image feature parameters, the lesion texture complexity index and the lesion edge influence index are calculated. The specific formula for calculating the lesion texture complexity index is as follows:

[0026]

[0027] In the formula, This represents the complexity index of lesion texture. The normalized value of the predicted color contrast of the labeled region image. The predicted texture entropy value of the labeled region image. The predicted maximum texture entropy value for all labeled region images;

[0028] The formula used to calculate the lesion margin impact index is as follows:

[0029]

[0030] In the formula, The lesion margin impact index, This is the predicted perimeter of the lesion outline in the labeled region image. The predicted area of ​​the lesion contour in the labeled region image. The intensity of the lesion edge in the labeled region image. The circularity of the lesion outline in the labeled region image.

[0031] Furthermore, the circularity of the lesion outline in the labeled region image. The specific formula used for the calculation is as follows:

[0032]

[0033] The biochemical data includes erythrocyte sedimentation rate, white blood cell count, carcinoembryonic antigen (CEA) level, and alpha-fetoprotein (AFP) level. A biochemical discrimination influence index is calculated based on this data, and the specific formula for the biochemical discrimination influence index is as follows:

[0034]

[0035] In the formula, To influence the biochemical discrimination index, The normalized value of erythrocyte sedimentation rate for the corresponding patient. This represents the normalized value of the white blood cell count for the corresponding patient. and These are the normalized values ​​of the carcinoembryonic antigen (CEA) and alpha-fetoprotein (AFP) levels for the corresponding patients.

[0036] Furthermore, based on the obtained lesion texture complexity index and lesion edge influence index, combined with the biochemical discrimination influence index, a comprehensive lesion identification coefficient is generated. The formula used to calculate the comprehensive lesion identification coefficient is as follows:

[0037]

[0038] In the formula, For comprehensive lesion identification coefficient, , and These are the weighting coefficients for the biochemical discrimination influence index, the lesion edge influence index, and the lesion texture complexity index, respectively. and , and All are greater than 0.

[0039] Furthermore, the comprehensive lesion identification coefficient is compared with the set lesion judgment threshold to generate different comparison results and issue corresponding lesion identification judgment results. The specific judgment logic is as follows:

[0040] when If the lesion in the medical image is determined to be a tumor, other methods should be used for further diagnosis and treatment.

[0041] when When the lesion in the medical image is determined to be inflammation, anti-inflammatory treatment should be administered.

[0042] when At that time, the lesion in the medical image to be identified was determined to be hemorrhage;

[0043] in Threshold for judging lesions.

[0044] This invention also provides a lesion image recognition system based on a deep learning model. This system is used to execute the aforementioned lesion image recognition method based on a deep learning model, and includes:

[0045] The sample data acquisition module is used to acquire several medical images of different lesion types, perform image preprocessing to obtain sample images, manually label the lesion regions in the sample images, obtain the image feature parameters of the lesion labeling regions, and generate a sample training dataset.

[0046] The prediction model training module is used to build a segmentation and recognition model based on the sample training dataset. It takes the sample image as input and uses the label of the lesion area in the image and the image feature parameters as labels to train the segmentation and recognition model.

[0047] The biochemical index acquisition module is used to acquire medical images of lesions to be identified, input them into the trained segmentation and recognition model, and obtain the predicted values ​​of image feature parameters of the medical images of lesions to be identified, as well as the biochemical index data of the patients corresponding to the medical images of lesions to be identified.

[0048] The influence parameter analysis module is used to characterize the complexity of lesion texture and the influence of lesion edge based on the obtained image feature parameter prediction values, and to evaluate the biochemical influence based on biochemical index data.

[0049] The lesion image recognition module is used to generate a comprehensive lesion recognition coefficient based on the complexity of lesion texture and the influence of lesion edge, combined with biochemical influence. The comprehensive lesion recognition coefficient is compared with the set lesion judgment threshold to generate different comparison results and issue the corresponding lesion recognition judgment result, thus completing the lesion image recognition.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] First, image preprocessing and manual labeling are used to comprehensively analyze different types of medical images, making the labeling of lesion areas more accurate, thereby providing high-quality training data for deep learning models and improving the training efficiency and accuracy of the models.

[0052] Secondly, the established segmentation and recognition model, through precise segmentation and comprehensive feature extraction of the lesion region, can fully capture the shape, texture, and structural information of the lesion. By extracting multi-dimensional features such as color contrast, texture entropy, edge intensity, contour perimeter, and contour area, it provides a comprehensive description of the lesion, which helps to improve the lesion recognition rate. In addition, by combining the biochemical index data of the medical image to be identified, multi-dimensional analysis of the lesion can be achieved.

[0053] Furthermore, combining biochemical indicators with lesion characteristics yields comprehensive indicators such as the lesion texture complexity index and the biochemical discrimination influence index. This multi-source data-based analysis method not only improves the accuracy of lesion identification but also provides clinicians with richer information support, enabling better decision-making. Finally, by generating a comprehensive lesion identification coefficient and comparing it with a set judgment threshold, intelligent and automated lesion identification can be achieved. This process simplifies the physician's workflow, improves diagnostic efficiency, and provides patients with more timely medical intervention. Attached Figure Description

[0054] Figure 1This is a schematic diagram of the overall method flow of the present invention;

[0055] Figure 2 The fitted curve is the normalized value of erythrocyte sedimentation rate minus the normalized value of white blood cell count;

[0056] Figure 3 The curve showing the fitting of normalized erythrocyte sedimentation rate (ESR) versus normalized carcinoembryonic antigen (CEA) level.

[0057] Figure 4 The fitted curve of normalized white blood cell count - normalized alpha-fetoprotein level;

[0058] Figure 5 Color mapping of the influence index of normalized white blood cell count on biochemical discrimination;

[0059] Figure 6 This is a schematic diagram of the overall system structure of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0061] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0062] Example:

[0063] Please see Figures 1-5 The present invention provides a technical solution:

[0064] A method for lesion image recognition based on a deep learning model, comprising the following steps:

[0065] Step 1: Acquire several medical images of different lesion types and perform image preprocessing to obtain sample images. Mark the lesion regions in the sample images using manual labeling and obtain the image feature parameters of the lesion-marked regions to generate a sample training dataset.

[0066] The lesion regions in the sample images were manually labeled. Specifically, the LabelImg tool was used to label the images, selecting irregular curves to annotate the lesion contours. After all annotations were completed, the results were saved in an appropriate format (such as XML, JSON, CSV, etc.) for subsequent model training and analysis. The different lesion types specifically included hemorrhage, inflammation, and tumors.

[0067] The image preprocessing specifically includes: image enhancement and denoising preprocessing, wherein wavelet transform is used to denoise each lesion medical image, and bilateral filtering is used to enhance each lesion medical image.

[0068] The method for denoising and enhancing medical images of lesions is as follows: The denoising method of wavelet transform is used to denoise the distortion-corrected image. The specific steps of the wavelet transform denoising method include: decomposing the distortion-corrected image through wavelet transform to obtain wavelet coefficients of the image at different scales and directions; thresholding the wavelet coefficients, setting low-amplitude wavelet coefficients to zero while retaining high-amplitude wavelet coefficients; and performing an inverse transform on the thresholded wavelet coefficients to reconstruct the image from the processed coefficients, thus completing the image denoising process.

[0069] Bilateral filtering was used to enhance the details of the lesion in the medical image. The specific formula used for the filtering transformation is as follows:

[0070]

[0071] In the formula, The coordinate vector within the image coordinate system. Coordinate vector grayscale value at that location grayscale value Gray values ​​after bilateral filtering transformation Both are Gaussian functions, among which The formula used is:

[0072]

[0073]

[0074] In the formula, The coordinate vector within the image coordinate system. Coordinate vector grayscale value at that location and They are respectively The standard deviation.

[0075] Color contrast typically refers to the degree of difference in color distribution within an image. It can be obtained through the following steps: converting the image from the RGB color space to the HSV or Lab color space for better color analysis; calculating a histogram using functions in OpenCV; and then using this histogram to calculate the color contrast.

[0076] Texture entropy is a parameter used to describe the complexity of image texture. It is calculated by calculating the gray-level co-occurrence matrix of the image, which describes the spatial relationship between pixels of different gray levels. The probability distribution of each gray level is calculated through the gray-level co-occurrence matrix, and then the entropy formula is used to calculate the texture entropy.

[0077] Edge strength describes the salience of edges in an image and is usually obtained using edge detection algorithms, such as Canny edge detection or the Sobel operator. Edge strength can be calculated by measuring the gradient magnitude of the edge image or by using the output of the edge detection algorithm, which is usually the average intensity value of the edge pixels.

[0078] The perimeter of an object's outline refers to the total length of its boundary in an image. Edge detection algorithms (such as Canny) are used to obtain the edge map, and then a contour detection function is applied. Similarly, edge detection and contour detection methods are used to obtain the contour area.

[0079] The method for generating the training sample dataset is as follows: the labeled sample images are mapped one-to-one with the corresponding image feature parameters to form a corresponding grid, and the resulting grid is recorded as the training sample dataset.

[0080] Step 2: Establish a segmentation and recognition model based on the sample training dataset. Use the sample images as input and the labels of the lesion areas in the images and the image feature parameters as labels to train the segmentation and recognition model.

[0081] A segmentation recognition model is established based on a sample training dataset. This model is constructed using an LSTM model, and activation functions and optimization algorithms are selected. The Tanh function is chosen as the activation function, and Adam is chosen as the optimization algorithm for the LSTM model. The formula for the Tanh function is:

[0082]

[0083] In the formula, Represents the Tanh function, with the independent variable... This represents the weighted sum of the neuron's inputs, that is, the result of the weighted sum of the inputs received by the neuron from the previous layer;

[0084] Simultaneously, the hyperparameters of the LSTM model are set, including: number of network layers, number of iterations, learning rate, batch size, number of training iterations, number of batches, and number of hidden layer neurons.

[0085] The network is set to a 5-layer network structure, the number of iterations is set to 200, the learning rate is set to 0.001, the batch size is set to 32, the number of training iterations is set to 100, the batch size is set to 128, and the number of hidden layer neurons is set to 32.

[0086] The input to the trained segmentation and recognition model is a medical image of the lesion, and the output is an image with the lesion region marked, as well as the corresponding predicted values ​​of image feature parameters, including the predicted values ​​of color contrast, texture entropy, lesion edge intensity, lesion contour perimeter, and lesion contour area.

[0087] LSTM (Laser-Based Memory) can capture nonlinear relationships in data through its complex network structure. This is particularly important for extracting features such as color contrast and texture entropy, as the relationships between these features are often nonlinear. Feature parameters (such as color contrast and edge intensity) typically change over time in image sequences. LSTM can adapt to input sequences of varying lengths, making it particularly effective when handling irregular or variable-length time series. LSTM can automatically learn features without requiring the manual design of complex feature extraction algorithms. Through training, LSTM can learn effective ways to extract features such as color contrast and edge intensity, thereby improving the accuracy and reliability of feature extraction.

[0088] Step 3: Obtain the medical image of the lesion to be identified, input it into the trained segmentation and recognition model, and obtain the predicted value of the image feature parameters of the medical image of the lesion to be identified, as well as the biochemical index data of the patient corresponding to the medical image of the lesion to be identified.

[0089] Obtain medical images of the lesions to be identified, and perform image preprocessing. The preprocessing steps are the same as those described above and will not be repeated here.

[0090] The biochemical indicators include erythrocyte sedimentation rate, white blood cell count, carcinoembryonic antigen level, and alpha-fetoprotein level.

[0091] Simultaneously, the biochemical indicators of the patient corresponding to the medical images of the lesions to be identified are obtained as follows: Biochemical indicators are usually measured through blood samples. Medical staff will use aseptic techniques to draw a certain amount of blood from the patient. ESR is usually performed using the Western or Capillary method. Anticoagulated blood is placed in a special test tube, and the sedimentation rate of red blood cells is measured over a certain period of time. The white blood cell count can be performed using an automated blood analyzer (such as a fully automated blood cell counter). The method for detecting carcinoembryonic antigen (CEA) levels is: detection using immunological methods (such as enzyme-linked immunosorbent assay (ELISA) or chemiluminescence immunoassay). The method for detecting alpha-fetoprotein (AFP) levels is: similarly, detection can be performed using enzyme-linked immunosorbent assay (ELISA), chemiluminescence immunoassay, or radioimmunoassay.

[0092] Step 4: Based on the obtained image feature parameter prediction values, characterize the lesion texture complexity and lesion edge influence, and evaluate the biochemical influence based on biochemical index data.

[0093] Based on the obtained predicted values ​​of image feature parameters, the lesion texture complexity index and the lesion edge influence index are calculated. The specific formula for calculating the lesion texture complexity index is as follows:

[0094]

[0095] In the formula, This represents the complexity index of lesion texture. The normalized value of the predicted color contrast of the labeled region image. The predicted texture entropy value of the labeled region image. The predicted maximum texture entropy value for all labeled region images;

[0096] It should be noted that the lesion texture complexity index This value is used to represent the texture complexity of the lesion area; the higher the value, the more complex the texture.

[0097] Color contrast is used to reflect the diversity of colors in an image. Higher color contrast means that the lesion area contains more color variations. High color contrast is usually associated with complex structures and morphologies, increasing the probability of lesions. Therefore, the normalized value of the color contrast prediction is... With lesion texture complexity index Proportional, through the logarithmic function It can smoothly handle the effects of color contrast variations. The specific normalization method involves extracting the raw color contrast values ​​from all images to be analyzed and calculating the maximum and minimum values ​​in the dataset. Let the maximum value be... The minimum value is Using a linear normalization formula, the predicted color contrast values ​​of the labeled region image are converted into normalized predicted color contrast values ​​of the labeled region image. The linear normalization formula is a conventional existing technique and will not be elaborated here.

[0098] Texture entropy is an important parameter describing texture complexity. A higher entropy value indicates a more complex and irregular texture, and the more complex and irregular the texture, the greater the probability of the presence of lesions. Therefore, the predicted texture entropy value of the labeled region image... With lesion texture complexity index Proportional, through an exponential function In the denominator, the higher the entropy value, the greater the texture complexity of the lesion. Used to normalize the predicted texture entropy values ​​of the labeled region image.

[0099] The formula used to calculate the lesion margin impact index is as follows:

[0100]

[0101] In the formula, The lesion margin impact index, This is the predicted perimeter of the lesion outline in the labeled region image. The predicted area of ​​the lesion contour in the labeled region image. The intensity of the lesion edge in the labeled region image. The circularity of the lesion outline in the labeled region image.

[0102] It should be noted that the lesion margin impact index It indicates the complexity of the lesion's edge and the degree of its impact on surrounding tissues. The higher the value, the more complex the structure of the lesion area and the more severe the lesion.

[0103] Among these factors, the perimeter of the lesion is a crucial influence on its marginal characteristics. A longer perimeter often indicates a more complex and tortuous edge, potentially leading to a greater impact on surrounding tissues. The ratio of perimeter to area (i.e., the relationship between perimeter and area) provides a quantitative assessment of the impact of the lesion's margins, while edge intensity reflects the clarity of the lesion's edges. Clearer edges may indicate a more significant impact on surrounding tissues. Edge intensity and perimeter together constitute the molecular component; taking the square root of the sum of the squares of perimeter and edge intensity enhances the reflection of their combined characteristics, ultimately leading to a more comprehensive assessment. It reflects the overall characteristics of the edge.

[0104] The concept of roundness is introduced to adjust for the edge influence index. A value closer to 1 indicates a shape closer to a circle, meaning relatively regular edges. Lesions, however, are often irregular in shape, potentially exhibiting serrated or protruding features, or lobulated, spiculated, or infiltrative edges. Therefore, roundness... Influence index of lesion margin Inversely proportional, through It emphasizes the non-linear relationship between shape and edge complexity, while also indicating that when... When reduced, the lesion margin influence index Significantly increased.

[0105] The circularity of the lesion outline in the labeled region image The specific formula used for the calculation is as follows:

[0106]

[0107] The biochemical discriminant influence index is calculated based on biochemical index data. The specific formula for the biochemical discriminant influence index is as follows:

[0108]

[0109] In the formula, To influence the biochemical discrimination index, The normalized value of erythrocyte sedimentation rate for the corresponding patient. This represents the normalized value of the white blood cell count for the corresponding patient. and These are the normalized values ​​of the carcinoembryonic antigen (CEA) and alpha-fetoprotein (AFP) levels for the corresponding patients.

[0110] The specific method used for normalization preprocessing is as follows: data can be normalized by subtracting the maximum and minimum values. This method compresses all data values ​​into a fixed range (usually 0 to 1), eliminating the influence of different units or dimensions between indicators, and making each indicator have the same dimension, which is convenient for subsequent calculations.

[0111] It should be noted that the biochemical discrimination index These are used to assess the importance of certain biochemical indicators in determining the impact of lesions; higher values ​​may indicate more severe or complex lesions.

[0112] Among them, the normalized value of erythrocyte sedimentation rate corresponding to the patient It is an indicator reflecting the state of inflammation or infection, and is usually used clinically to determine the degree of inflammatory response. The higher the value, the greater the probability that the patient has inflammation. Therefore, it corresponds to the patient's normalized erythrocyte sedimentation rate (ESR). Impact of biochemical discrimination index Proportional, through the square An increase in the degree of inflammation indicates the impact of The effect is non-linearly enhanced, reflecting high The value may be associated with a more severe pathological condition.

[0113] The normalized white blood cell count, representing a patient's white blood cell count, is commonly used to assess the immune response or infection status. White blood cells are an important component of the immune system, responsible for fighting infections and other diseases. When the body experiences infection or inflammation, the white blood cell count usually increases; a high white blood cell count often indicates the presence of infection, inflammation, or other pathological conditions. Therefore, the normalized white blood cell count corresponds to the patient's... Impact of biochemical discrimination index Proportional, through the logarithmic function This is to prevent extremely high volatility from having an excessive impact on the results and to ensure stability.

[0114] Carcinoembryonic antigen (CEA) levels are commonly used to monitor the presence and activity of tumors. Increased CEA levels are generally associated with the presence or development of tumors; therefore, a normalized value for the CEA level in a patient is needed. The higher the level, the more severe the lesion. Alpha-fetoprotein (AFP) levels are similar to carcinoembryonic antigen (CEA), and an increase in the normalized value of AFP levels is usually associated with the presence of a tumor. Therefore, an exponential function can be used to determine the relative severity of the lesion. The denominator indicates a direct proportional relationship, while the exponential relationship indicates the influence of the biochemical discrimination index when alpha-fetoprotein (AFP) and carcinoembryonic antigen (CEA) levels increase. Significantly increased, used to indicate a significant impact on the assessment of lesion severity.

[0115] Step 5: Based on the complexity of lesion texture and the influence of lesion edge, combined with biochemical effects, generate a comprehensive lesion recognition coefficient. Compare the comprehensive lesion recognition coefficient with the set lesion judgment threshold to generate different comparison results, issue the corresponding lesion recognition judgment result, and complete the lesion image recognition.

[0116] Based on the obtained lesion texture complexity index and lesion edge influence index, combined with the biochemical discrimination influence index, a comprehensive lesion identification coefficient is generated. The formula used to calculate the comprehensive lesion identification coefficient is as follows:

[0117]

[0118] In the formula, For comprehensive lesion identification coefficient, , and These are the weighting coefficients for the biochemical discrimination influence index, the lesion edge influence index, and the lesion texture complexity index, respectively. and , and All are greater than 0.

[0119] It should be noted that the comprehensive lesion identification coefficient This index is used to indicate the severity of lesions; the higher the value, the more severe the lesion. Since the relative relationships between the biochemical discrimination influence index, lesion edge influence index, and lesion texture complexity index and lesion identification have already been explained above, they will not be repeated here.

[0120] Among them, through the exponential function The significant impact of the biochemical discrimination index on lesion identification is indicated by a logarithmic function. To prevent extremely high volatility from having an excessive impact on the results and to ensure the stability of the model.

[0121] In clinical practice, biochemical indicators (such as tumor markers) are crucial for assessing disease condition and risk. They directly reflect a patient's physiological state and pathological changes. The lesion margin impact index reflects the lesion structure and its influence on surrounding tissues, which is also very important in lesion assessment. Although margin impact is a key factor, its importance is generally slightly lower than that of biochemical indicators. While the lesion texture complexity index helps assess lesion characteristics based on morphology and texture features, its clinical value is relatively lower than that of biochemical indicators and margin impact. Therefore, it is set as follows: and , and All are greater than 0.

[0122] The comprehensive lesion identification coefficient is compared with the set lesion judgment threshold to generate different comparison results and issue the corresponding lesion identification judgment result. The specific judgment logic is as follows:

[0123] when If the lesion in the medical image is determined to be a tumor, other methods should be used for further diagnosis and treatment.

[0124] when When the lesion in the medical image is determined to be inflammation, anti-inflammatory treatment should be administered.

[0125] when At that time, the lesion in the medical image to be identified was determined to be hemorrhage;

[0126] in Threshold for judging lesions.

[0127] Please see Figure 6The present invention also provides a lesion image recognition system based on a deep learning model, wherein the lesion image recognition system based on a deep learning model is used to perform the above-described lesion image recognition method based on a deep learning model, comprising:

[0128] The sample data acquisition module is used to acquire several medical images of different lesion types, perform image preprocessing to obtain sample images, manually label the lesion regions in the sample images, obtain the image feature parameters of the lesion labeling regions, and generate a sample training dataset.

[0129] The prediction model training module is used to build a segmentation and recognition model based on the sample training dataset. It takes the sample image as input and uses the label of the lesion area in the image and the image feature parameters as labels to train the segmentation and recognition model.

[0130] The biochemical index acquisition module is used to acquire medical images of lesions to be identified, input them into the trained segmentation and recognition model, and obtain the predicted values ​​of image feature parameters of the medical images of lesions to be identified, as well as the biochemical index data of the patients corresponding to the medical images of lesions to be identified.

[0131] The influence parameter analysis module is used to characterize the complexity of lesion texture and the influence of lesion edge based on the obtained image feature parameter prediction values, and to evaluate the biochemical influence based on biochemical index data.

[0132] The lesion image recognition module is used to generate a comprehensive lesion recognition coefficient based on the complexity of lesion texture and the influence of lesion edge, combined with biochemical influence. The comprehensive lesion recognition coefficient is compared with the set lesion judgment threshold to generate different comparison results and issue the corresponding lesion recognition judgment result, thus completing the lesion image recognition.

[0133] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0134] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0135] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0136] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for lesion image recognition based on a deep learning model, characterized in that, The specific steps include: Several medical images of different lesion types were acquired and preprocessed to obtain sample images. The lesion regions in the sample images were marked by manual labeling, and the image feature parameters of the lesion marked regions were obtained to generate a sample training dataset. A segmentation and recognition model is established based on a sample training dataset. The sample images are used as input, and the labels of the lesion regions in the images and the image feature parameters are used as labels to train the segmentation and recognition model. The medical image of the lesion to be identified is obtained and input into the trained segmentation and recognition model to obtain the predicted value of the image feature parameters of the medical image of the lesion to be identified, and at the same time, the biochemical index data of the patient corresponding to the medical image of the lesion to be identified. Based on the obtained image feature parameter prediction values, the complexity of lesion texture and the impact of lesion edge are characterized, and the biochemical impact is evaluated based on biochemical index data; Based on the complexity of lesion texture and the influence of lesion edge, combined with biochemical effects, a comprehensive lesion recognition coefficient is generated. The comprehensive lesion recognition coefficient is compared with the set lesion judgment threshold to generate different comparison results and issue the corresponding lesion recognition judgment result, thus completing the lesion image recognition. Based on the obtained predicted values ​​of image feature parameters, the lesion texture complexity index and the lesion edge influence index are calculated. The specific formula for calculating the lesion texture complexity index is as follows: ; In the formula, This represents the complexity index of lesion texture. The normalized value of the predicted color contrast of the labeled region image. The predicted texture entropy value of the labeled region image. The predicted maximum texture entropy value for all labeled region images; The formula used to calculate the lesion margin impact index is as follows: ; In the formula, The lesion margin impact index, This is the predicted perimeter of the lesion outline in the labeled region image. The predicted area of ​​the lesion contour in the labeled region image. The intensity of the lesion edge in the labeled region image. The circularity of the lesion outline in the labeled region image; The circularity of the lesion outline in the labeled region image The specific formula used for the calculation is as follows: ; The biochemical data includes erythrocyte sedimentation rate, white blood cell count, carcinoembryonic antigen (CEA) level, and alpha-fetoprotein (AFP) level. A biochemical discrimination influence index is calculated based on this data, and the specific formula for the biochemical discrimination influence index is as follows: ; In the formula, To influence the biochemical discrimination index, The normalized value of erythrocyte sedimentation rate for the corresponding patient. This represents the normalized value of the white blood cell count for the corresponding patient. and These are the normalized values ​​of the carcinoembryonic antigen (CEA) and alpha-fetoprotein (AFP) levels for the corresponding patients. Based on the obtained lesion texture complexity index and lesion edge influence index, combined with the biochemical discrimination influence index, a comprehensive lesion identification coefficient is generated. The formula used to calculate the comprehensive lesion identification coefficient is as follows: ; In the formula, For comprehensive lesion identification coefficient, , and These are the weighting coefficients for the biochemical discrimination influence index, the lesion edge influence index, and the lesion texture complexity index, respectively. and , and All are greater than 0.

2. The lesion image recognition method based on a deep learning model according to claim 1, characterized in that: The image feature parameters include color contrast, texture entropy, lesion edge intensity, lesion contour perimeter, and lesion contour area; the lesion regions in the sample images are marked manually, specifically by marking the images using a labeling tool, and the different lesion types specifically include hemorrhage, inflammation, and tumor; The image preprocessing specifically includes: image enhancement and denoising preprocessing, wherein wavelet transform is used to denoise each lesion medical image, and bilateral filtering is used to enhance each lesion medical image. The method for generating the sample training dataset is as follows: the labeled sample images are mapped one-to-one with the corresponding image feature parameters to form a corresponding grid, and the resulting grid is recorded as the sample training dataset.

3. The lesion image recognition method based on a deep learning model according to claim 2, characterized in that: A segmentation recognition model is established based on a sample training dataset. This model is constructed using an LSTM model, and activation functions and optimization algorithms are selected. The Tanh function is chosen as the activation function, and Adam is chosen as the optimization algorithm for the LSTM model. The formula for the Tanh function is: ; In the formula, Represents the Tanh function, with the independent variable... This represents the weighted sum of the neuron's inputs, that is, the result of the weighted sum of the inputs received by the neuron from the previous layer; Simultaneously, the hyperparameters of the LSTM model are set, including: number of network layers, number of iterations, learning rate, batch size, number of training iterations, number of batches, and number of hidden layer neurons. The network is set to a 5-layer network structure, the number of iterations is set to 200, the learning rate is set to 0.001, the batch size is set to 32, the number of training iterations is set to 100, the batch size is set to 128, and the number of hidden layer neurons is set to 32. The input to the trained segmentation and recognition model is a medical image of the lesion, and the output is an image with the lesion region marked, as well as the corresponding predicted values ​​of image feature parameters, including the predicted values ​​of color contrast, texture entropy, lesion edge intensity, lesion contour perimeter, and lesion contour area.

4. The lesion image recognition method based on a deep learning model according to claim 1, characterized in that: The comprehensive lesion identification coefficient is compared with the set lesion judgment threshold to generate different comparison results and issue the corresponding lesion identification judgment result. The specific judgment logic is as follows: when If the lesion in the medical image is determined to be a tumor, other methods should be used for further diagnosis and treatment. when When the lesion in the medical image is determined to be inflammation, anti-inflammatory treatment should be administered. when At that time, the lesion in the medical image to be identified was determined to be hemorrhage; in Threshold for judging lesions.

5. A lesion image recognition system based on a deep learning model, characterized in that: The lesion image recognition system based on a deep learning model is used to execute the lesion image recognition method based on a deep learning model according to any one of claims 1-4, comprising: The sample data acquisition module is used to acquire several medical images of different lesion types, perform image preprocessing to obtain sample images, manually label the lesion regions in the sample images, obtain the image feature parameters of the lesion labeling regions, and generate a sample training dataset. The prediction model training module is used to build a segmentation and recognition model based on the sample training dataset. It takes the sample image as input and uses the label of the lesion area in the image and the image feature parameters as labels to train the segmentation and recognition model. The biochemical index acquisition module is used to acquire medical images of lesions to be identified, input them into the trained segmentation and recognition model, and obtain the predicted values ​​of image feature parameters of the medical images of lesions to be identified, as well as the biochemical index data of the patients corresponding to the medical images of lesions to be identified. The influence parameter analysis module is used to characterize the complexity of lesion texture and the influence of lesion edge based on the obtained image feature parameter prediction values, and to evaluate the biochemical influence based on biochemical index data. The lesion image recognition module is used to generate a comprehensive lesion recognition coefficient based on the complexity of lesion texture and the influence of lesion edge, combined with biochemical influence. The comprehensive lesion recognition coefficient is compared with the set lesion judgment threshold to generate different comparison results and issue the corresponding lesion recognition judgment result, thus completing the lesion image recognition.

Citation Information

Patent Citations

  • A method and system for lesion image recognition based on a deep learning model

    CN110232383B

  • Identification and classification method for lesions in medical images

    CN120219799A

  • Machine learning-based lesion identification using focused recursive synthetic training data

    WO2025049550A2