An AI-based method for medical image analysis and processing

Abstract

This invention discloses an AI-based medical image analysis and processing method, belonging to the field of image analysis technology. It involves collecting multimodal medical image data from patients using imaging equipment and preprocessing it; fusing the preprocessed multimodal medical image data and extracting biological tissue features to construct a biological tissue feature library, marking abnormal biological tissue features; training an AI visual detection model based on the marked biological tissue feature library according to the medical image analysis task to identify and classify different abnormal biological tissues; and automatically detecting abnormal biological tissue regions in the images using the AI ​​visual detection model. This invention enables high-precision analysis of medical images through AI technology, detecting minute abnormal changes. Furthermore, through quantitative analysis of image features, it can assess the severity of lesions and predict disease progression trends, allowing for timely detection before lesions cause obvious symptoms, thus reducing the risk of disease progression.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, and specifically to an AI-based medical image analysis and processing method. Background Technology

[0002] Medical imaging technology is an indispensable part of modern medicine, widely used in the early diagnosis of diseases, the formulation of treatment plans, and the monitoring of diseases. Common types of medical imaging include X-rays, CT (computed tomography), MRI (magnetic resonance imaging), ultrasound, and PET (positron emission tomography). With the continuous advancement of medical technology and the increasing aging of the population, the demand for medical imaging examinations is growing. The rapid development of AI technology has provided new solutions for medical image analysis, promoting the development of medical image analysis towards automation and intelligence.

[0003] For example, the AI-based medical image analysis and processing method disclosed in Chinese Patent Publication No. CN116823701A can perform repair and segmentation processing on medical images, effectively reducing the noise components of medical images and facilitating the synchronous processing of different sub-image parts of medical images. It can also perform biological tissue feature identification and comparison processing on medical images to accurately locate lesion sources and improve the accuracy of medical image processing.

[0004] In existing technologies, the identification and comparison of biological tissue features in medical images can quickly and accurately locate lesions, solving the problems of long processing time and significant noise interference in the processing results of digital medical images, which reduce the reliability and accuracy of digital medical image processing. However, due to the large differences in biological tissue features among different patients, some cases may fail to accurately identify abnormal tissues, leading to misdiagnosis or missed diagnosis of abnormal biological tissues. Therefore, an AI-based medical image analysis and processing method is proposed to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-based medical image analysis and processing method to solve the problems mentioned in the background art.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] An AI-based medical image analysis and processing method includes the following steps:

[0008] Step 1: Collect multimodal medical image data of the patient from the imaging equipment and preprocess it;

[0009] Step 2: Fuse the preprocessed multimodal medical image data, extract biological tissue features, construct a biological tissue feature library, and label abnormal biological tissue features.

[0010] Step 3: Based on the medical image analysis task, train the AI ​​visual detection model in conjunction with the labeled biological tissue feature library to identify and classify different abnormal biological tissues;

[0011] Step 4: Use an AI visual detection model to automatically detect abnormal biological tissue regions in the image, and segment the detected abnormal biological tissues to obtain their boundaries and ranges.

[0012] Step 5: Based on the segmentation results and the detection results of the AI ​​visual detection model, analyze the characteristics of the abnormal biological tissue area to determine whether there is a lesion;

[0013] Step 6: Based on the analysis results of the AI ​​visual detection model, determine the severity of the existing lesions, and then provide targeted recommendations.

[0014] A further improvement to the technical solution of this invention lies in the following: In step one, the collection and preprocessing of multimodal medical image data includes:

[0015] It connects to the hospital's imaging equipment (CT, MRI, PET, ultrasound, etc.) and obtains patients' image data through the hospital's information system (PACS, i.e., image storage and transmission system) to ensure the integrity and accuracy of the data. At the same time, it records the patient's personal information (name, age, gender, medical record number, etc.) as well as the time and type of imaging examination, assigns a unique identifier to the patient, and matches the patient's image data with personal information.

[0016] For multiple imaging modalities (CT+MRI, PET+CT, etc.) of the same patient, the images of different modalities are spatially registered to ensure that the images have the same geometric relationship and are integrated into a unified folder to facilitate subsequent processing and ensure the spatial resolution and temporal series consistency of the data of different modalities.

[0017] The integrated multimodal medical image data is preprocessed, including format conversion, noise reduction, and correction.

[0018] Multimodal medical image data, after quality checks, is stored on a local server to ensure data security and accessibility. A database management system (MySQL, MongoDB, etc.) is used to record metadata, including patient information, examination information, image data format, storage location, etc., and a data index is established to facilitate quick querying and retrieval of image data later.

[0019] A further improvement to the technical solution of this invention lies in that: in step two, the construction process of the biological tissue feature library includes:

[0020] The deep learning-based fusion method uses convolutional neural networks to extract image features, fuses preprocessed multimodal medical image data of the same patient, and verifies the fused image data to verify whether the fused image retains the key information of the multimodal images, including texture information, morphological information and functional metabolic information.

[0021] Using computer vision and image processing techniques, biological tissue features, including morphological features, texture features, functional features and density features, are extracted from fused multimodal image data. The extracted biological tissue features are selected and optimized, redundant and irrelevant features are removed, and features relevant to abnormal tissue detection are retained.

[0022] A biological tissue feature library is constructed, and the extracted and optimized biological tissue features are stored in the biological tissue feature library to form a structured feature dataset. The biological tissue feature library includes feature information of normal and abnormal biological tissues for subsequent analysis, comparison and identification.

[0023] The biological tissue features in the biological tissue feature database are analyzed, and abnormal biological tissue features are labeled. The labeling content includes the abnormal tissue type (tumor, inflammation, etc.), the location and extent of the abnormal tissue, and the clinical diagnostic information of the abnormal tissue.

[0024] A further improvement to the technical solution of this invention lies in: the specific process of retaining the features related to abnormal tissue detection is as follows:

[0025] For morphological features, including shape, size, and edges; for texture features, including gray-level co-occurrence matrix (GLCM) features, local binary pattern (LBP) features, and wavelet transform features; for functional features, including metabolic information and blood perfusion information; and for density features, including pixel value distribution and histogram features.

[0026] Selecting and optimizing the extracted biological tissue features includes features related to abnormal tissue morphology, such as irregular shape and blurred edges; features related to abnormal tissue texture, such as texture heterogeneity and texture roughness; features related to abnormal tissue function, such as metabolic abnormalities and abnormal blood perfusion; and features related to abnormal tissue density, such as density inhomogeneity and density differences.

[0027] A further improvement to the technical solution of this invention lies in the fact that, in step three, the training process of the AI ​​visual detection model includes:

[0028] Labeled feature datasets are obtained from biological tissue feature databases, and a model architecture based on convolutional neural networks is selected in conjunction with medical image analysis tasks. The feature datasets are divided into training and testing sets. The feature datasets contain feature information of multimodal medical images and corresponding labels (normal or abnormal biological tissues), and the feature datasets cover different patients, different disease types and different image modalities.

[0029] This paper describes how to build a CNN model using a deep learning framework, defining the network structure, convolutional layers, pooling layers, and fully connected layers. Regularization techniques are incorporated to prevent overfitting. A multi-task learning module (combining classification and segmentation) is designed to fully utilize feature information. The CNN model is trained using training data to construct an AI visual detection model. During training, model parameters, including learning rate, batch size, and number of iterations, are adjusted to optimize performance. The training process is monitored to evaluate the model's generalization ability and prevent overfitting. The trained model is evaluated using test data, and metrics such as accuracy, recall, and F1 score are calculated to measure model performance. Based on the evaluation results, the model is optimized.

[0030] The trained AI visual detection model is deployed into the medical image analysis system. The deployed model is used to extract features from new image data and map them to corresponding category labels in order to identify and classify different abnormal biological tissues.

[0031] A further improvement to the technical solution of this invention lies in that the process of segmenting the detected abnormal biological tissue includes:

[0032] Load the pre-trained AI visual detection model, input the medical image data to be detected into the model, and the model automatically extracts image features through convolutional layers, pooling layers and fully connected layers, and classifies each pixel in the image to determine whether it belongs to abnormal biological tissue.

[0033] The AI ​​visual detection model outputs a segmentation mask for abnormal biological tissue based on the classification results and sets a distinction threshold. It converts the probability map into a binary mask to clearly distinguish between abnormal and normal tissue. The region with a value of 1 in the binary mask is represented as abnormal biological tissue, and the region with a value of 0 is represented as normal tissue. The segmentation mask is a two-dimensional or three-dimensional matrix with the same size as the input image, where the value of each pixel or voxel represents the probability that it belongs to abnormal tissue.

[0034] A series of morphological operations are applied to process the segmentation mask. Through multiple iterations of erosion and dilation operations, the boundary of the segmentation mask is optimized to better fit the actual shape of the abnormal tissue, thereby optimizing the segmentation result. The morphological operations include erosion, dilation, opening, and closing operations.

[0035] The edge detection algorithm (Canny edge detection) is used to extract the boundary of the abnormal tissue from the optimized segmentation mask to obtain the boundary contour of the abnormal biological tissue. The boundary is represented by a series of coordinate points to describe the contour of the abnormal tissue. Based on the extracted boundary contour, the boundary coordinates and range of the abnormal biological tissue are calculated, including area or volume, centroid position and maximum diameter. The area or volume represents the two-dimensional area or three-dimensional volume of the abnormal tissue, the centroid position represents the centroid position of the abnormal tissue to describe its spatial distribution, and the maximum diameter represents the longest span of the abnormal tissue.

[0036] A further improvement to the technical solution of the present invention is that: in step five, the process of determining whether a lesion exists includes:

[0037] From the segmentation mask output by the AI ​​visual detection model, the classification probability of each pixel or region is parsed out, and the region of abnormal biological tissue is extracted based on the segmentation mask. The boundary coordinates and range of the obtained abnormal biological tissue are analyzed, including area or volume, centroid position and maximum diameter.

[0038] By integrating morphological features, texture features, intensity features and spatial features, a comprehensive feature description is formed, and based on medical knowledge and clinical experience, lesion judgment criteria such as area or volume, centroid location and maximum diameter are set.

[0039] The fused features are compared with the lesion judgment criteria to determine whether there is a lesion in the abnormal area. If a lesion is found, it is further classified as benign, malignant, or inflammatory.

[0040] A further improvement to the technical solution of the present invention is that the lesion judgment criteria are as follows:

[0041] Set area threshold A T and volume threshold V T , if A>A T Or volume V>V T This indicates the presence of a lesion;

[0042] If the centroid position shifts significantly over time, it suggests that the lesion is undergoing dynamic changes. If the GLCM contrast of the texture feature or the standard deviation of the pixel intensity of the intensity feature deviates significantly from the normal range, it suggests that a lesion is present.

[0043] Set the maximum diameter threshold D T If D max >D T This indicates the presence of a lesion.

[0044] A further improvement to the technical solution of the present invention is that: in step six, the process of judging the severity of the lesion includes:

[0045] Based on the established lesion judgment criteria, the abnormal biological tissue characteristics exhibiting lesion behavior are analyzed. According to medical knowledge and clinical experience, a grading standard for the severity of lesions is set, dividing the severity of lesions into three levels: mild, moderate, and severe. Among them, the lesion characteristics of mild lesion are slight, such as small area or volume, regular shape, and little impact on surrounding tissues; the lesion characteristics of moderate lesion are obvious, such as moderate area or volume, irregular shape, and some impact on surrounding tissues; and the lesion characteristics of severe lesion are severe, such as large area or volume, extremely irregular shape, and significant impact or invasion on surrounding tissues.

[0046] The abnormal biological tissue characteristics identified as exhibiting pathological behavior are comprehensively analyzed to calculate the severity evaluation coefficient of the lesion and quantify the severity of the lesion.

[0047] By combining the lesion severity evaluation coefficient and the lesion grade classification, a corresponding severity threshold is matched for each lesion grade;

[0048] For each severity level of the lesion, corresponding recommended measures are provided.

[0049] Due to the adoption of the above technical solution, the technical progress achieved by this invention compared to the prior art is as follows:

[0050] 1. This invention provides an AI-based medical image analysis and processing method. Through AI technology, medical images can be analyzed with high precision, detecting minute abnormal changes. Furthermore, through quantitative analysis of image features, the severity of lesions can be assessed and the development trend of diseases can be predicted. This method can detect lesions in time before they cause obvious symptoms, reducing the risk of disease progression.

[0051] 2. This invention provides an AI-based medical image analysis and processing method. By extracting and analyzing key features in the images, it reduces misdiagnosis and missed diagnosis caused by human factors. Through deep learning algorithms, AI can quickly and accurately identify lesion areas, improving the accuracy of diagnosis. At the same time, the automated processing of AI also greatly shortens the diagnosis time, improves the efficiency of doctors' work, and allows doctors to focus more on the treatment and care of patients. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0053] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0054] Figure 2 This is a schematic diagram of the process for determining the presence of lesions according to the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Example 1, as Figure 1 As shown, this invention provides an AI-based medical image analysis and processing method, comprising the following steps:

[0057] Step 1: Collect multimodal medical image data of patients from imaging equipment and preprocess it. Connect to the hospital's imaging equipment (CT, MRI, PET, ultrasound, etc.) and obtain patient image data through the hospital's information system (PACS, i.e., image storage and transmission system) to ensure data integrity and accuracy. Simultaneously, record the patient's personal information (name, age, gender, medical record number, etc.) and the time and type of imaging examination. Assign a unique identifier to each patient and match the patient's image data with their personal information. For multiple imaging modalities of the same patient (CT+MRI, PET+CT, etc.), spatially register the images of different modalities to ensure that the images have the same geometric relationships and integrate them into a unified folder for subsequent processing, ensuring spatial resolution and temporal series consistency of different modal data. Preprocess the integrated multimodal medical image data, including format conversion, noise reduction, and correction. Specifically, convert the image data output from different devices to HDF5 format to ensure data format compatibility. Use filtering algorithms (Gaussian filtering, median filtering, etc.) to remove noise from the images, reduce artifacts and interference, and then perform geometric correction (rotation). The images are rotated, panned, and scaled to ensure proper alignment. Furthermore, for CT and MRI images, intensity correction is performed to normalize pixel values ​​to a uniform range (0-1 or 0-255). Then, the preprocessed multimodal medical image data undergoes data quality checks, including integrity checks, consistency checks, and outlier checks. Integrity checks verify the completeness of the image data, identifying missing slices or data corruption. Consistency checks ensure temporal and spatial consistency of the multimodal data, avoiding errors caused by equipment differences or varying examination times. Outlier checks detect outliers (such as pixel values ​​exceeding the normal range) for correction or removal. The multimodal medical image data after quality checks is stored on a local server to ensure data security and accessibility. A database management system (MySQL, MongoDB, etc.) is used to record metadata, including patient information, examination information, image data format, and storage location. A data index is also established for easy and rapid subsequent querying and retrieval of image data.

[0058] Step 2: Fuse preprocessed multimodal medical image data and extract biological tissue features, construct a biological tissue feature library, and label abnormal biological tissue features. A deep learning-based fusion method is used, employing a convolutional neural network to extract image features, fusing preprocessed multimodal medical image data from the same patient, and validating the fused image data to verify whether it retains key information from the multimodal images, including texture, morphology, and functional metabolic information. By comparing with the original images, the effectiveness of the fused images in preserving anatomical details and functional information is evaluated. Specifically, the CNN automatically learns features from different modalities, using a two-branch CNN architecture to process anatomical images (CT, MRI) and functional images (PET) separately. After feature extraction through convolution operations, an attention mechanism is used to fuse the features. For multi-scale information fusion, local Laplacian filtering is combined to decompose the image into... Multi-scale approximate and residual images are used, and then super-resolution anatomical images are extracted using CNN. These images are then fused with functional images. Using computer vision and image processing techniques, biological tissue features, including morphological, texture, functional, and density features, are extracted from the fused multimodal image data. The extracted biological tissue features are selected and optimized, removing redundant and irrelevant features and retaining features relevant to abnormal tissue detection. A biological tissue feature library is constructed, and the extracted and optimized biological tissue features are stored in the biological tissue feature library to form a structured feature dataset. The biological tissue feature library includes feature information of normal and abnormal biological tissues for subsequent analysis, comparison, and identification. The biological tissue features in the biological tissue feature library are analyzed, and abnormal biological tissue features are labeled. The labeling content includes the abnormal tissue type (tumor, inflammation, etc.), the location and extent of the abnormal tissue, and the clinical diagnostic information of the abnormal tissue.

[0059] In addition, features related to abnormal tissue detection are retained, specifically through the following process:

[0060] For morphological features, these include shape, size, and edges; for texture features, they include gray-level co-occurrence matrix (GLCM) features, local binary pattern (LBP) features, and wavelet transform features; for functional features, they include metabolic information and blood perfusion information; and for density features, they include pixel value distribution and histogram features. Shape features describe the geometry of biological tissues, such as circular, elliptical, or irregular shapes; size features are characterized by measuring the area, volume, or diameter of the tissue; edge features reflect changes in tissue boundaries, such as smoothness, jaggedness, or blurring; and gray-level co-occurrence matrix features are obtained by calculating the relationships between pixel gray values. Spatial correlation is used to describe the uniformity, contrast, and correlation of textures. Local binary mode features reflect the texture patterns of local areas and have a certain robustness to changes in illumination. Wavelet transform features extract detailed information of textures through multi-scale analysis. Metabolic information features represent features reflecting tissue metabolic activities extracted from PET images. Blood perfusion information features represent features reflecting tissue blood perfusion extracted from MRI or CT perfusion imaging. Pixel value distribution features characterize tissue density by statistically analyzing the distribution of pixel values ​​in the image. Histogram features reflect the density distribution of tissues by calculating the gray-level histogram of the image.

[0061] The extracted biological tissue features are selected and optimized. Features related to abnormal tissue morphology include irregular shape and blurred edges; features related to abnormal tissue texture include texture heterogeneity and texture roughness; features related to abnormal tissue function include metabolic abnormalities and abnormal blood perfusion; and features related to abnormal tissue density include density inhomogeneity and density differences. Among these, abnormal tissues usually have irregular shapes, and retaining irregularity indicators in shape features helps in the detection of abnormal tissues; the edges of abnormal tissues are relatively blurred or jagged, and the blurring or irregularity indicators in edge features are retained; the texture of abnormal tissues is relatively complex and heterogeneous, and features reflecting texture heterogeneity (GLCM contrast, LBP complexity) are retained; the texture of abnormal tissues is relatively rough, and features reflecting texture roughness are retained; abnormal tissues have high metabolic activity, and functional features reflecting metabolic abnormalities are retained; the blood perfusion of abnormal tissues differs from that of normal tissues, and functional features reflecting abnormal blood perfusion are retained; the density distribution of abnormal tissues is relatively uneven, and features reflecting density inhomogeneity are retained; the density difference between abnormal tissues and surrounding normal tissues is large, and features reflecting density differences are retained.

[0062] Step 3: Based on the medical image analysis task, train an AI visual detection model using a labeled biological tissue feature library to identify and classify different abnormal biological tissues. Obtain the labeled feature dataset from the biological tissue feature library, and select a convolutional neural network-based model architecture based on the medical image analysis task. Divide the feature dataset into training and testing sets. The feature dataset contains feature information of multimodal medical images and corresponding labels (normal or abnormal biological tissues), and covers different patients, different disease types, and different image modalities. Use a deep learning framework to build a CNN model, defining the network structure, convolutional layers, pooling layers, fully connected layers, etc. Regularization techniques are added to the model to prevent overfitting. Design a multi-task learning model. The system combines classification and segmentation to fully utilize feature information. A convolutional neural network model is trained using training data to build an AI visual detection model. During training, model parameters are adjusted to optimize performance, including learning rate, batch size, and number of iterations. The training process is monitored to evaluate the model's generalization ability and prevent overfitting. The trained model is evaluated using test data, and metrics such as accuracy, recall, and F1 score are calculated to measure performance. Based on the evaluation results, the model is optimized. The trained AI visual detection model is then deployed to a medical image analysis system. The deployed model is used to extract features from new image data and map them to corresponding category labels to identify and classify different abnormal biological tissues.

[0063] Step 4: Utilize an AI visual detection model to automatically detect abnormal biological tissue regions in the image, segment the detected abnormal biological tissue, and obtain its boundaries and extent. Load the pre-trained AI visual detection model, input the medical image data to be detected into the model, and the model automatically extracts image features through convolutional layers, pooling layers, and fully connected layers, classifying each pixel in the image to determine whether it belongs to abnormal biological tissue. The AI ​​visual detection model outputs a segmentation mask for abnormal biological tissue based on the classification results, sets a discrimination threshold, and converts the probability map into a binary mask to clearly distinguish between abnormal and normal tissue. Regions with a value of 1 in the binary mask represent abnormal biological tissue, and regions with a value of 0 represent normal tissue. The segmentation mask is a two-dimensional or three-dimensional matrix with the same size as the input image, where the value of each pixel or voxel represents its probability of belonging to abnormal tissue. A series of morphological operations are applied to process the segmentation mask, and through multiple iterations of erosion and dilation operations, the boundaries of the segmentation mask are optimized to better fit the abnormal tissue. The actual shape of the abnormal tissue is used to optimize the segmentation results. Morphological operations include erosion, dilation, opening, and closing operations. Erosion is used to remove small holes or noise points in the segmentation results, making the boundaries of the abnormal tissue smoother. Dilation is used to fill small holes in the segmentation results, making the boundaries of the abnormal tissue more complete. The opening operation is eroded first and then dilated to remove small objects or connecting parts. The closing operation is dilated first and then eroded to fill small holes or broken parts. The boundary of the abnormal tissue is extracted from the optimized segmentation mask using an edge detection algorithm (Canny edge detection) to obtain the boundary contour of the abnormal biological tissue. The boundary is represented by a series of coordinate points to describe the contour of the abnormal tissue. Based on the extracted boundary contour, the boundary coordinates and range of the abnormal biological tissue are calculated, including area or volume, centroid position, and maximum diameter. Area or volume represents the calculated two-dimensional area or three-dimensional volume of the abnormal tissue. Centroid position represents the calculated centroid position of the abnormal tissue to describe its spatial distribution. Maximum diameter represents the calculated longest span of the abnormal tissue.

[0064] The expression for the position of the centroid is:

[0065]

[0066] In the formula, The location of the center of mass. and Let x and y represent the x and y coordinates of the centroid, respectively. i and y i Let x and y represent the x and y coordinates of the i-th boundary point, respectively, and n be the total number of boundary points. The centroid position reflects the center position of the abnormal tissue in the image, and its value will change with the position of the abnormal tissue.

[0067] The expression for calculating two-dimensional area is:

[0068]

[0069] In the formula, A represents the two-dimensional area of ​​the anomalous tissue. The two-dimensional area formula is based on the shoelace formula and is used to calculate the area of ​​polygons. n+1 =x1 and y n+1 =y1, which represents the closure of the polygon. The value of A is a non-negative real number, which represents the area of ​​the abnormal tissue. The area A changes with the size of the abnormal tissue and is used to assess the severity of the lesion area.

[0070] The expression for calculating the three-dimensional area is:

[0071]

[0072] In the formula, V represents the three-dimensional volume of the abnormal tissue. Let V be the three-dimensional coordinates of the i-th boundary point. V is a non-negative real number representing the volume of the abnormal tissue. The volume V varies with the size of the abnormal tissue and is used to assess the severity of the three-dimensional lesion region. · represents the dot product, and × represents the cross product. This represents the three-dimensional coordinate vector of the j-th boundary point. Let j and k be the three-dimensional coordinate vectors of the k-th boundary point. j and k are index variables used to traverse all possible combinations of points. j is traversed from i+1 to n, ensuring that j is always greater than i. k is traversed from j+1 to n, ensuring that k is always greater than j.

[0073] The expression for calculating the maximum diameter is:

[0074]

[0075] In the formula, D max The maximum diameter of the anomalous tissue is given by the square root, which is used to calculate the Euclidean distance between two points, D. max The value range is non-negative real number. The maximum diameter reflects the longest span of the abnormal tissue, and its value will change with the size of the abnormal tissue.

[0076] Step 5: Based on the segmentation results and the detection results of the AI ​​visual detection model, analyze the characteristics of the abnormal biological tissue area to determine whether there is a lesion;

[0077] Step 6: Based on the analysis results of the AI ​​visual detection model, determine the severity of the existing lesions, and then provide targeted recommendations.

[0078] Example 2, as Figure 2As shown, based on Embodiment 1, the present invention provides a technical solution: Preferably, in step five, the process of determining whether a lesion exists includes:

[0079] From the segmentation mask output by the AI ​​visual detection model, the classification probability of each pixel or region is parsed out. Based on the segmentation mask, regions of abnormal biological tissue are extracted. The boundary coordinates and extent of the obtained abnormal biological tissue are analyzed, including area or volume, centroid location, and maximum diameter. Morphological features, texture features, intensity features, and spatial features are fused to form a comprehensive feature description. Based on medical knowledge and clinical experience, lesion judgment criteria for area or volume, centroid location, and maximum diameter are set. The fused features are compared with the lesion judgment criteria to determine whether there is a lesion in the abnormal region. If a lesion is found, further investigation is conducted on the lesion. The lesions are classified as benign, malignant, or inflammatory. Benign lesions are close to the normal range but have slight abnormalities, such as a slightly larger area or volume than the normal range. They usually grow slowly, have clear boundaries, and are distinctly separated from surrounding tissues. They are non-invasive. Malignant lesions deviate significantly from the normal range, such as a rapid increase in area or volume, unstable centroid position, complex texture, rapid growth, blurred boundaries, and may invade surrounding tissues. They have the ability to metastasize. Inflammatory lesions are characterized by irregular boundaries, uneven intensity, and may be accompanied by inflammatory reactions in surrounding tissues. Symptoms include local redness, swelling, heat, and pain. On imaging, they may appear as an abnormal area with unclear boundaries.

[0080] The criteria for diagnosing lesions are as follows:

[0081] Set area threshold A T and volume threshold V T , if A>A T Or volume V>V T If the centroid position shifts significantly over time, it indicates a dynamic change in the lesion. If the GLCM contrast of the texture feature or the standard deviation of the pixel intensity of the intensity feature significantly deviates from the normal range, it indicates a lesion. Set a maximum diameter threshold D. T If D max >D T This indicates the presence of a lesion;

[0082] Step six, the process of determining the severity of the lesion, includes:

[0083] Based on established lesion assessment criteria, the abnormal biological tissue characteristics exhibiting pathological behavior are analyzed. According to medical knowledge and clinical experience, a grading standard for lesion severity is established, classifying lesion severity into three levels: mild, moderate, and severe. Mild lesions are characterized by subtle features, such as small area or volume, regular shape, and minimal impact on surrounding tissues. Moderate lesions are characterized by obvious features, such as moderate area or volume, irregular shape, and some impact on surrounding tissues. Severe lesions are characterized by severe features, such as large area or volume, extremely irregular shape, and significant impact or invasion on surrounding tissues. The identified abnormal biological tissue characteristics exhibiting pathological behavior are comprehensively analyzed to calculate a lesion severity evaluation coefficient, quantifying the severity of the lesion. Combining the lesion severity evaluation coefficient and the lesion level classification, corresponding severity thresholds are assigned to each lesion level. For each severity level, corresponding recommendations are provided. For mild lesions, it is recommended to have imaging follow-up every 3-6 months to monitor lesion development, adjust lifestyle habits such as diet and sleep to promote health, and, if necessary, administer drug treatment. Controlling disease progression is crucial. Mild lesions typically do not pose a serious threat to a patient's life and health, but close monitoring is necessary to identify and address potential risks promptly. For moderate lesions, immediate and detailed medical examinations are required to determine the nature and cause of the lesion. Based on the examination results, a personalized treatment plan should be developed, such as surgery, radiotherapy, or chemotherapy. Strengthening the patient's daily care and nutritional support is essential to improve their immunity. Moderate lesions may have some impact on a patient's life and health, therefore prompt and detailed medical examinations are necessary to determine the nature and cause of the lesion and develop an appropriate treatment plan. For severe lesions, an emergency treatment procedure should be initiated immediately, involving multidisciplinary consultation to develop a comprehensive treatment plan. Close monitoring of the patient's vital signs is crucial, with timely adjustments to the treatment plan. Comprehensive daily care and psychological support should be provided to improve the patient's quality of life. Severe lesions may pose a serious threat to a patient's life and health, therefore immediate and detailed emergency treatment procedures are necessary, involving multidisciplinary consultation to develop a comprehensive treatment plan. Simultaneously, close monitoring of the patient's vital signs is essential, with timely adjustments to the treatment plan and comprehensive daily care and psychological support provided.

[0084] The expression for the severity assessment coefficient of the lesion is:

[0085]

[0086] In the formula, SS is the severity evaluation coefficient of the lesion; the higher the value, the more severe the lesion. A is the area or volume of the abnormal tissue, A0 is the baseline value of the area or volume used for normalization, and D represents the maximum area or volume of the normal tissue. maxD0 is the baseline value for the maximum diameter of the abnormal tissue, used for normalization, and represents the maximum diameter of the normal tissue. ΔC is the change in centroid position, i.e., the offset compared to the previous examination, used to assess the dynamic changes of the lesion. α, β, and γ are weighting coefficients used to adjust the contribution of each feature to the severity assessment. SS ranges from 0 to 1, where 0 represents no lesion and 1 represents the most severe lesion. When A and D... max Or, when ΔC increases, the SS value increases, indicating an increase in the severity of the lesion;

[0087] Multiple lesion grades correspond to multiple severity thresholds, where the severity thresholds include an upper limit threshold and a lower limit threshold;

[0088] Multiple lesion grades and multiple severity thresholds satisfy the following relationship:

[0089] Mild lesion grade 0 <SS≤SS L ;

[0090] Moderate lesion grade SS L <SS≤SS M ;

[0091] Severe lesion grade SS M <SS<1;

[0092] Wherein, SS is the severity evaluation coefficient for the lesion. L SS represents the lower threshold corresponding to moderate lesion level and the upper threshold corresponding to mild lesion level. M SS represents the lower threshold corresponding to severe lesion level and the upper threshold corresponding to moderate lesion level. L =0.33, SS M =0.67.

[0093] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations 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. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI-based medical image analysis and processing method, characterized in that, Includes the following steps: Step 1: Collect multimodal medical image data of the patient from the imaging equipment and preprocess it; Step 2: Fuse the preprocessed multimodal medical image data, extract biological tissue features, construct a biological tissue feature library, and label abnormal biological tissue features. Step 3: Based on the medical image analysis task, train the AI ​​visual detection model in conjunction with the labeled biological tissue feature library to identify and classify different abnormal biological tissues; Step 4: Use the AI ​​visual detection model to automatically detect abnormal biological tissue regions in the image, segment the detected abnormal biological tissue to obtain its boundary and range, load the pre-trained AI visual detection model, input the medical image data to be detected into the model, the model automatically extracts image features through convolutional layer, pooling layer and fully connected layer structure, classifies each pixel in the image, and determines whether it belongs to abnormal biological tissue. The AI ​​visual detection model outputs a segmentation mask for abnormal biological tissues based on the classification results, sets a distinction threshold, and converts the probability map into a binary mask to clearly distinguish between abnormal and normal tissues. The region with a value of 1 in the binary mask is represented as abnormal biological tissue, and the region with a value of 0 is represented as normal tissue. Step 5: Based on the segmentation results and the detection results of the AI ​​visual detection model, analyze the characteristics of the abnormal biological tissue region to determine whether a lesion exists. The process of determining whether a lesion exists includes: From the segmentation mask output by the AI ​​visual detection model, the classification probability of each pixel or region is parsed out, and the region of abnormal biological tissue is extracted based on the segmentation mask. The boundary coordinates and range of the obtained abnormal biological tissue are analyzed, including area or volume, centroid position and maximum diameter. By integrating morphological features, texture features, intensity features and spatial features, a comprehensive feature description is formed, and based on medical knowledge and clinical experience, lesion judgment criteria such as area or volume, centroid location and maximum diameter are set. The fused features are compared with the lesion judgment criteria to determine whether there is a lesion in the abnormal area. If there is a lesion, it is further classified as benign, malignant or inflammatory. The criteria for diagnosing the lesions are as follows: Set area threshold A T and volume threshold V T If area A > A T Or volume V>V T This indicates the presence of a lesion; If the centroid position shifts significantly over time, it suggests that the lesion is undergoing dynamic changes. If the GLCM contrast of the texture feature or the standard deviation of the pixel intensity of the intensity feature deviates significantly from the normal range, it suggests that a lesion is present. Set the maximum diameter threshold D T If D max >D T This indicates the presence of a lesion; Step Six: Based on the analysis results of the AI ​​visual detection model, determine the severity of the existing lesions, and then output targeted recommendations. The process of determining the severity of the lesions includes: Based on the established criteria for lesion diagnosis, the characteristics of abnormal biological tissues exhibiting lesion behavior are analyzed. Furthermore, based on medical knowledge and clinical experience, a grading standard for the severity of lesions is established, classifying the severity of lesions into three levels: mild, moderate, and severe. The abnormal biological tissue characteristics identified as exhibiting pathological behavior are comprehensively analyzed to calculate the severity evaluation coefficient of the lesion and quantify the severity of the lesion. By combining the lesion severity evaluation coefficient and the lesion grade classification, a corresponding severity threshold is matched for each lesion grade; For each severity level of the lesion, corresponding recommended measures are provided; The expression for the severity assessment coefficient of the lesion is: In the formula, SS is the severity evaluation coefficient of the lesion, A is the area or volume of the abnormal tissue, A0 is the baseline value of the area or volume, representing the maximum area or volume of the normal tissue, and D... max D0 represents the maximum diameter of the abnormal tissue, D0 is the baseline value of the maximum diameter, ΔC represents the maximum diameter of the normal tissue, ΔC is the change in the centroid position, which is the offset compared to the previous examination, and α, β and γ are weighting coefficients.

2. The AI-based medical image analysis and processing method according to claim 1, characterized in that: In step one, the collection and preprocessing of multimodal medical image data includes: It connects to the hospital's imaging equipment and obtains the patient's imaging data through the hospital's information system. At the same time, it records the patient's personal information, as well as the time and type of imaging examination, assigns a unique identifier to the patient, and matches the patient's imaging data with personal information. For multiple imaging modalities of the same patient, images of different modalities are spatially registered and integrated into a unified folder; The integrated multimodal medical image data is preprocessed, including format conversion, noise reduction, and correction. After data quality checks, the multimodal medical image data is stored on a local server. A database management system is used to record the metadata of the data and to create a data index.

3. The AI-based medical image analysis and processing method according to claim 2, characterized in that: In step two, the construction process of the biological tissue feature library includes: The deep learning-based fusion method uses convolutional neural networks to extract image features, fuses preprocessed multimodal medical image data of the same patient, and verifies the fused image data to verify whether the fused image retains the key information of the multimodal images, including texture information, morphological information and functional metabolic information. Using computer vision and image processing techniques, biological tissue features, including morphological features, texture features, functional features and density features, are extracted from fused multimodal image data. The extracted biological tissue features are selected and optimized, redundant and irrelevant features are removed, and features relevant to abnormal tissue detection are retained. A biological tissue feature library is constructed, and the extracted and optimized biological tissue features are stored in the biological tissue feature library to form a structured feature dataset. The biological tissue feature library includes feature information of normal and abnormal biological tissues. The biological tissue features in the biological tissue feature database are analyzed, and abnormal biological tissue features are labeled. The labeling content includes the abnormal tissue type, abnormal tissue location and extent, and abnormal tissue clinical diagnostic information.

4. The AI-based medical image analysis and processing method according to claim 3, characterized in that: The specific process for retaining the features related to abnormal tissue detection is as follows: For morphological features, including shape, size, and edges; for texture features, including gray-level co-occurrence matrix features, local binary pattern features, and wavelet transform features; for functional features, including metabolic information and blood perfusion information; and for density features, including pixel value distribution and histogram features. Selecting and optimizing the extracted biological tissue features includes features related to abnormal tissue morphology, such as irregular shape and blurred edges; features related to abnormal tissue texture, such as texture heterogeneity and texture roughness; features related to abnormal tissue function, such as metabolic abnormalities and abnormal blood perfusion; and features related to abnormal tissue density, such as density inhomogeneity and density differences.

5. The AI-based medical image analysis and processing method according to claim 4, characterized in that: In step three, the training process of the AI ​​visual detection model includes: Labeled feature datasets are obtained from biological tissue feature databases, and a model architecture based on convolutional neural networks is selected in conjunction with medical image analysis tasks. The feature datasets are divided into training and testing sets. The feature datasets contain feature information of multimodal medical images and corresponding labels, and cover different patients, different disease types and different image modalities. We use a deep learning framework to build a CNN model, use training set data to train a convolutional neural network model, build an AI visual detection model, adjust model parameters to optimize performance during training, use test set data to evaluate the trained model, and optimize the model based on the evaluation results. The trained AI visual detection model is deployed into the medical image analysis system. The deployed model is used to extract features from new image data and map them to corresponding category labels in order to identify and classify different abnormal biological tissues.

6. The AI-based medical image analysis and processing method according to claim 5, characterized in that: Step four involves segmenting the detected abnormal biological tissue, including: A series of morphological operations are applied to process the segmentation mask. Through multiple iterations of erosion and dilation operations, the boundary of the segmentation mask is optimized to better fit the actual shape of the abnormal tissue, thereby optimizing the segmentation result. The morphological operations include erosion, dilation, opening, and closing operations. The boundary of the abnormal tissue is extracted from the optimized segmentation mask using an edge detection algorithm to obtain the boundary contour of the abnormal biological tissue. Based on the extracted boundary contour, the boundary coordinates and range of the abnormal biological tissue are calculated, including area or volume, centroid location, and maximum diameter.

Images