An AI-based automatic segmentation and annotation system for medical images

Abstract

The present invention relates to the field of medical technology, and in particular to an AI-based automatic medical image segmentation and annotation system, comprising an image data acquisition module, a data preprocessing module, a feature extraction module, an automatic segmentation module and an intelligent annotation module, wherein the image data acquisition module is used to acquire medical image data, the data preprocessing module improves image quality by noise suppression, artifact removal and modality alignment, the feature extraction module extracts feature maps of potential lesion areas, the automatic segmentation module combines a full convolutional network, a bidirectional long short-term memory network and an attention mechanism to achieve accurate segmentation of organ areas and lesion areas, and the intelligent annotation module uses medical knowledge graphs and semantic embedding technology to annotate the segmentation results with anatomical names, lesion types and pathological characteristics; the present invention has the advantages of high segmentation accuracy, comprehensive annotation and strong scalability, and is helpful to improve the automation and intelligence level of medical image analysis.

Description

Technical Field

[0001] The present invention relates to the field of medical technology, and in particular to an AI-based automatic medical image segmentation and annotation system. Background Art

[0002] Medical imaging is an important basis for modern medical diagnosis and treatment, and is widely used in disease detection, surgical planning, and efficacy evaluation. With the rapid development of medical imaging technologies such as CT, MRI, and X-ray, a large amount of high-resolution, multimodal imaging data has been generated, bringing new opportunities for medical image analysis.

[0003] Although the existing medical image segmentation and annotation technologies have improved in terms of automation, there are still many technical bottlenecks. For complex lesion areas (such as lesions with irregular shapes, low contrast or similar adjacent tissues), the segmentation model is not robust enough and it is difficult to generate high-precision segmentation results. Annotation technology is usually limited to simple identification of anatomical structures and lacks fine-grained annotation of lesion types and pathological characteristics, resulting in insufficient clinical practicality of the annotation results. Summary of the invention

[0004] The present invention provides an AI-based automatic medical image segmentation and annotation system.

[0005] An AI-based medical image automatic segmentation and annotation system includes an image data acquisition module, a data preprocessing module, a feature extraction module, an automatic segmentation module and an intelligent annotation module, wherein;

[0006] The image data acquisition module is used to acquire medical image data, which includes CT images, MRI images and X-ray images, and to perform format conversion and standardization processing on the acquired medical image data, so as to convert the medical image data into a unified DICOM format;

[0007] The data preprocessing module is used to preprocess the collected medical image data;

[0008] The feature extraction module is used to extract features from the preprocessed medical image data and extract feature maps of potential lesion areas;

[0009] The automatic segmentation module uses a segmentation model constructed by a fully convolutional network combined with a bidirectional long short-term memory network to automatically segment the feature map output by the feature extraction module to generate a segmentation result including an organ area and a lesion area;

[0010] The intelligent labeling module combines the medical knowledge graph to label the segmentation results generated by the automatic segmentation module, and the labeling content includes the anatomical name, lesion type and pathological characteristics of the target area.

[0011] Optionally, the image data acquisition module includes:

[0012] Equipment connection: The image data acquisition module establishes a connection with the medical imaging equipment through the communication protocol interface. The medical imaging equipment includes CT imaging equipment, MRI imaging equipment and X-ray imaging equipment;

[0013] Real-time image acquisition: using the acquisition control program to obtain real-time generated medical image data from medical imaging equipment, including CT images, MRI images and X-ray images;

[0014] Format conversion: convert the collected medical image data into a unified DICOM format;

[0015] Image data standardization: unify the size, resolution and grayscale range of medical image data;

[0016] Verification of acquisition results: consistency check of medical imaging data after format conversion and standardization;

[0017] Output and transfer: The medical imaging data that has completed format conversion and standardization processing will be transferred to the data preprocessing module to ensure smooth data connection.

[0018] Optionally, the preprocessing includes noise suppression, which is achieved by detecting the types of noise in the medical imaging data, including Gaussian noise and speckle noise, and using an adaptive noise suppression algorithm to reduce different types of noise. The noise reduction effect is evaluated by peak signal-to-noise ratio and structural similarity indicators. If the preset quality standards are not met, the algorithm parameters are adjusted and reprocessed.

[0019] Optionally, the feature extraction module includes:

[0020] Region of interest detection: Receive preprocessed medical image data, detect significant areas in the image by gradient amplitude calculation, mark areas with larger gradient amplitude as potential lesion candidate areas, screen potential lesion candidate areas based on grayscale range and connectivity analysis, merge the screened significant areas, and generate a binary mask matrix of the region of interest;

[0021] Feature extraction: Calculate the geometric features and grayscale statistical features of the region of interest, and use the grayscale co-occurrence matrix to calculate the texture features of the region of interest;

[0022] Feature map generation: Integrate geometric features, grayscale statistical features and texture features to generate a feature map of the potential lesion area.

[0023] Optionally, the automatic segmentation module implements preliminary region segmentation through a fully convolutional network, specifically including:

[0024] Input feature map loading: receiving the feature map output by the feature extraction module;

[0025] Convolution operation extracts segmentation features: Use a fully convolutional network to perform pixel-level convolution operations on the input feature map to generate a segmentation feature map;

[0026] Downsampling and upsampling: extract multi-scale features through downsampling operations in the fully convolutional network, and restore the segmentation features to the same spatial resolution as the original feature map through upsampling operations;

[0027] Generation of preliminary segmentation results: The restored segmentation feature map is processed by activation function and threshold to generate preliminary segmentation results.

[0028] Optionally, the automatic segmentation module combines a bidirectional long short-term memory network to refine the segmentation boundary, specifically including:

[0029] Boundary feature extraction: Extract boundary features from the preliminary segmentation results;

[0030] Boundary serialization: convert boundary features into boundary sequences;

[0031] Boundary optimization: Use a bidirectional long short-term memory network to optimize the boundary sequence and generate an optimized boundary sequence;

[0032] Optimize boundary mapping: Map the optimized boundary sequence back to the original space to generate refined segmentation boundaries;

[0033] Refined segmentation result output: The optimized boundary is merged into the preliminary segmentation result to generate a refined segmentation mask.

[0034] Optionally, the automatic segmentation module enhances the segmentation capability of complex lesion areas by introducing an attention mechanism, specifically including:

[0035] Attention weight calculation: Calculate the attention weight of each pixel based on the input feature map and refined segmentation results;

[0036] Attention-weighted feature generation: Use attention weights to weight the input feature map to generate an enhanced feature map;

[0037] Complex area optimization: The enhanced feature map is integrated into the refined segmentation result to further improve the segmentation effect of complex lesion areas;

[0038] Final segmentation result output: Output the final segmentation result, including accurate segmentation of organ area and lesion area.

[0039] Optionally, the intelligent labeling module combines the medical knowledge graph to label the anatomical name of the target area on the segmentation result generated by the automatic segmentation module, specifically including:

[0040] Segmentation result loading: receiving the segmentation result output by the automatic segmentation module;

[0041] Anatomical knowledge matching: Based on the medical knowledge graph, the standard names and spatial feature descriptions of anatomical regions are loaded, including the location, shape, and size range of organs;

[0042] Spatial position comparison: compare the regional features (such as center point coordinates and area) in the segmentation results with the standard regional features in the medical knowledge graph to match the most appropriate anatomical name;

[0043] Anatomical annotation generation: assign the anatomical names in the comparison results to the corresponding segmented regions, and attach the corresponding regional information (such as area, boundary coordinates);

[0044] Labeling result output: Output the labeling results including the anatomical name for subsequent labeling of lesion type and pathological features.

[0045] Optionally, the intelligent annotation module combines semantic embedding technology with a medical terminology library to annotate the lesion type and pathological features of the target area, specifically including:

[0046] Annotation model loading: Load the annotation model built based on semantic embedding technology. The model pre-training data includes standard medical terminology libraries (such as SNOMED CT, UMLS) and clinical case data of segmented areas.

[0047] Semantic feature extraction of segmented regions: Extract the semantic features of the region from the segmentation results, including geometric features, grayscale statistical features, and texture features, and use encoding functions to map the regional features to the semantic space;

[0048] Medical terminology matching: compare the regional semantic embedding with the semantic embedding of the medical terminology library and calculate the matching score;

[0049] Lesion type and pathological feature annotation: The best matching medical term is selected as the annotation result according to the matching score. The annotation content includes lesion type (such as mass, inflammation, ischemia) and pathological features (such as size, boundary characteristics, signal intensity);

[0050] Annotation verification and output: Perform consistency checks on the generated annotation results to ensure that the annotation content conforms to the logical relationships in the medical knowledge graph (such as the association between anatomical location and lesion type), and output the final annotation results containing lesion type and pathological characteristics for clinical use or further analysis.

[0051] Beneficial effects of the present invention:

[0052] In this invention, the automatic segmentation module combines a fully convolutional network, a bidirectional long short-term memory network and an attention mechanism, and realizes accurate segmentation of complex lesion areas through multi-stage optimization. The fully convolutional network can extract multi-scale features of the image, generate preliminary segmentation results, and ensure the overall accuracy of the segmentation; the bidirectional long short-term memory network uses boundary sequence information to optimize the segmentation boundary, effectively solving the boundary blur problem of weak contrast areas; the attention mechanism further focuses on complex lesion areas, strengthens the model's ability to capture features of difficult-to-segment areas, and greatly improves the robustness and accuracy of the segmentation effect.

[0053] In this invention, the intelligent annotation module uses medical knowledge graph and semantic embedding technology to realize all-round intelligent annotation of segmentation results. Through the anatomical knowledge comparison function of the medical knowledge graph, the anatomical name can be added to the target area quickly and accurately; through the semantic embedding technology combined with the medical terminology library, the lesion type and pathological characteristics of the segmented area are intelligently matched to ensure the clinical practicality and comprehensiveness of the annotation results. The automated processing of the entire process not only significantly improves the annotation efficiency, but also avoids the subjective bias that may exist in manual annotation. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0055] Figure 1 A schematic diagram of a system flow of an embodiment of the present invention;

[0056] Figure 2 Schematic diagram of the feature extraction module flow in an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. At the same time, it is explained here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments, and those skilled in the art may also adopt other alternatives to implement some known technologies; and the accompanying drawings are only for more specific description of the embodiments, and are not intended to specifically limit the present invention.

[0058] It should be noted that the references to "one embodiment", "embodiment", "exemplary embodiments", "some embodiments" and the like in the specification indicate that the embodiments described may include specific features, structures or characteristics, but not every embodiment may include the specific features, structures or characteristics. In addition, when a specific feature, structure or characteristic is described in conjunction with an embodiment, it should be within the knowledge of a person skilled in the art to implement such feature, structure or characteristic in conjunction with other embodiments (whether or not explicitly described).

[0059] In general, a term can be understood, at least in part, from its use in context. For example, depending, at least in part, on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending, at least in part, on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0060] like Figure 1-Figure 2 As shown, an AI-based medical image automatic segmentation and annotation system includes an image data acquisition module, a data preprocessing module, a feature extraction module, an automatic segmentation module and an intelligent annotation module, wherein;

[0061] The image data acquisition module is used to acquire medical image data, including CT images, MRI images and X-ray images. The module is connected to medical imaging equipment, supports real-time image input, and performs format conversion and standardization processing on the acquired medical image data, converting the medical image data into a unified DICOM format;

[0062] The data preprocessing module is used to preprocess the collected medical imaging data;

[0063] The feature extraction module is used to extract features from the preprocessed medical image data and extract feature maps of potential lesion areas;

[0064] The automatic segmentation module uses a segmentation model constructed by a fully convolutional network combined with a bidirectional long short-term memory network to automatically segment the feature map output by the feature extraction module and generate segmentation results including organ areas and lesion areas;

[0065] The intelligent labeling module combines the medical knowledge graph to label the segmentation results generated by the automatic segmentation module. The labeling content includes the anatomical name, lesion type and pathological characteristics of the target area.

[0066] The image data acquisition module includes:

[0067] Equipment connection: The image data acquisition module establishes a connection with the medical imaging equipment through the communication protocol interface. The medical imaging equipment includes CT imaging equipment, MRI imaging equipment and X-ray imaging equipment;

[0068] Real-time image acquisition: using the acquisition control program to obtain real-time generated medical image data from medical imaging equipment, including CT images, MRI images and X-ray images;

[0069] Format conversion: convert the collected medical image data into a unified DICOM format. During the format conversion process, add image metadata, including image acquisition time, device type, parameter settings (such as resolution, layer thickness), etc., for use by subsequent processing modules;

[0070] Image data standardization: Unify the size, resolution, and grayscale range of medical image data to ensure that medical image data from different devices and modalities have consistent input standards, including:

[0071] For CT images, grayscale standardization is performed to map the HU value to a predefined range;

[0072] For MRI images, correct the inhomogeneity of signal intensity;

[0073] Acquisition result verification: consistency check of medical imaging data after format conversion and standardization, including format integrity verification, image quality assessment (such as clarity and contrast), and metadata matching check;

[0074] If data anomalies are found, an acquisition error log is generated and corrected by re-acquisition or parameter adjustment to ensure that the output image data meets the quality requirements of subsequent processing;

[0075] Output and transfer: The medical imaging data that has completed format conversion and standardization processing will be transferred to the data preprocessing module to ensure smooth data connection.

[0076] Preprocessing includes noise suppression. By detecting the noise types in medical image data, including Gaussian noise and speckle noise, the adaptive noise suppression algorithm is used to reduce the noise of different types to ensure that the image edge and texture features are retained while denoising. The noise reduction effect is evaluated by the peak signal-to-noise ratio and structural similarity indicators. If it does not meet the preset quality standards, the algorithm parameters are adjusted and reprocessed. The specific steps are as follows:

[0077] (1) Noise analysis and classification include:

[0078] (1.1) Input image data detection: receiving medical image data ,in represents the two-dimensional pixel value distribution, and is the spatial coordinate of medical imaging data;

[0079] Use the noise detection model to perform preliminary analysis on the image data and extract pixel intensity distribution and spatial frequency characteristics;

[0080] (1.2) Noise type identification includes:

[0081] (1.21) Gaussian noise detection: Calculate the mean of the image and standard deviation , detect whether there is a Gaussian distribution feature:

[0082] , ;

[0083] in is the total number of pixels, It is The intensity value of each pixel. If the distribution conforms to the Gaussian characteristic, it is judged that Gaussian noise exists;

[0084] (1.22) Speckle noise detection: by image contrast Detecting granular noise:

[0085] ;

[0086] in, and are the maximum and minimum pixel values ​​of the image, respectively. When the contrast is too large, speckle noise may exist;

[0087] (2) Adaptive noise suppression algorithm includes:

[0088] (2.1) Noise processing strategy selection: Select an appropriate noise reduction algorithm based on the noise classification results, including

[0089] (2.11) Gaussian noise: For Gaussian noise, a Gaussian smoothing filter is used for noise reduction, which is expressed as:

[0090] ;

[0091] in, is the standard deviation of the filter kernel, and is the spatial coordinate offset of the filter kernel, is the coordinate in the denoised image The pixel value of the position;

[0092] (2.12) Speckle noise: For speckle noise, wavelet transform is used to denoise it, which is expressed as:

[0093] ;

[0094] in, is the image after wavelet denoising, is the wavelet basis function, is the wavelet coefficient, retain the main coefficient Achieve noise reduction;

[0095] (2.2) Edge and texture protection: A gradient preservation mechanism is embedded in the denoising process, expressed as:

[0096] ;

[0097] in is the edge gradient value after denoising, It is a preset gradient threshold used to protect edge information and retain high-frequency texture features through frequency domain analysis;

[0098] (3) Noise reduction result evaluation includes:

[0099] (3.1) Calculation of performance indicators: The performance indicators calculated include peak signal-to-noise ratio and structural similarity. The peak signal-to-noise ratio (PSNR) is calculated as:

[0100] ;

[0101] Among them, MSE is the mean square error;

[0102] The structural similarity (SSIM) is calculated as:

[0103] ;

[0104] in, and are the mean and variance of the original image and the denoised image respectively, is their covariance, is a constant;

[0105] (3.2) Verification of evaluation results: Perform quality evaluation on the medical image data after noise reduction. If the PSNR is higher than 30dB and the SSIM is higher than 0.95, the noise reduction effect is considered qualified. If it does not meet the standards, adjust the filter parameters. or Threshold , rerun the noise reduction process.

[0106] The feature extraction module includes:

[0107] Region of interest detection: Receive preprocessed medical image data, detect significant areas in the image by gradient amplitude calculation, mark areas with larger gradient amplitude as potential lesion candidate areas, screen potential lesion candidate areas based on grayscale range and connectivity analysis, merge the screened significant areas, and generate a binary mask matrix of the region of interest, including:

[0108] (1) Salient region recognition: The salient regions in the detected image are calculated using the gradient magnitude, expressed as:

[0109] ;

[0110] in, is the image at coordinates The gradient amplitude at , indicates the magnitude of the intensity change at that position. The image is -The gradient in the direction indicates the rate of change of the image in the horizontal direction. The image is -Directional gradient, which indicates the rate of change of the image in the vertical direction;

[0111] The area with larger gradient amplitude is marked as a potential lesion candidate area. Points >60 are marked as areas with large gradient amplitudes;

[0112] (2) Region of interest screening: Screen candidate regions based on grayscale range and connectivity analysis:

[0113] (2.1) Exclude areas with abnormal grayscale values ​​(too high or too low);

[0114] (2.2) Retain salient regions with connectivity characteristics;

[0115] (3) Region of interest mask generation: Merge the filtered salient regions to generate a binary mask matrix of the region of interest ,in Indicates the area of ​​interest. Indicates background;

[0116] Feature extraction: Based on , calculate the geometric features and grayscale statistical features of the region of interest, and use the gray level co-occurrence matrix to calculate the texture features of the region of interest. The specific steps are as follows:

[0117] based on , the geometric features of the region of interest are calculated as:

[0118] ;

[0119] in is the area of ​​the region of interest, is the circumference, is the shape compactness, which is used to describe the shape characteristics of the region of interest;

[0120] Calculate the grayscale statistical characteristics of the region of interest, including the mean , Standard Deviation and texture contrast , expressed as:

[0121] ;

[0122] ;

[0123] The gray-level co-occurrence matrix (GLCM) is used to calculate the texture features of the region of interest, including contrast, entropy, and correlation, which is expressed as:

[0124] ;

[0125] Contrast: Contrast ;

[0126] Entropy ;

[0127] Indicates grayscale and In the distance and angle The co-occurrence probability under ;

[0128] Feature map generation: Generate a feature map of the potential lesion area by integrating geometric features, grayscale statistical features, and texture features. , expressed as:

[0129] ;

[0130] in, It is the feature fusion weight, which is dynamically adjusted according to the importance of the feature.

[0131] The automatic segmentation module implements preliminary region segmentation through a fully convolutional network, including:

[0132] Input feature map loading: receiving the feature map output by the feature extraction module ;

[0133] Convolution operation extracts segmentation features: Use a fully convolutional network to perform pixel-level convolution operations on the input feature map to generate a segmentation feature map, which is expressed as:

[0134] ;

[0135] Among them, Conv represents the convolution operation, is the segmentation feature map;

[0136] Downsampling and upsampling: Multi-scale features are extracted through downsampling operations in the fully convolutional network, and the segmentation features are restored to the same spatial resolution as the original feature map through upsampling operations, expressed as:

[0137] ;

[0138] in, represents the feature map after upsampling operation, is the upsampling operator;

[0139] Generating preliminary segmentation results: The restored segmentation feature map is processed by activation function and threshold to generate preliminary segmentation results, which are expressed as:

[0140] ;

[0141] in, represents the activation function (such as softmax), is the segmentation threshold, is the preliminary segmentation mask;

[0142] Preliminary segmentation result output: Output As input for subsequent segmentation refinement.

[0143] The automatic segmentation module combines the bidirectional long short-term memory network to refine the segmentation boundaries, including:

[0144] Boundary feature extraction: from the preliminary segmentation results Extract boundary features from ,in Represents the boundary pixels of the segmented area. The boundary features are calculated as follows:

[0145] ;

[0146] in, is the boundary of the initial segmentation result;

[0147] Boundary serialization: Boundary features Converted to a boundary sequence, represented as , where each Indicates the position of boundary pixels in the sequence and their related feature values;

[0148] Boundary optimization: Use a bidirectional long short-term memory network to optimize the boundary sequence and generate an optimized boundary sequence , expressed as:

[0149] ;

[0150] in, is a bidirectional long short-term memory network, is the boundary sequence, is the optimized boundary sequence;

[0151] Optimize boundary mapping: The optimized boundary sequence Map back to the original space to generate refined segmentation boundaries ;

[0152] Refined segmentation result output: The refined boundary Fuse it into the preliminary segmentation result to generate a refined segmentation mask, expressed as:

[0153] ;

[0154] Output As an intermediate state of the final segmentation result, is the preliminary segmentation result. is the refined boundary.

[0155] The automatic segmentation module introduces an attention mechanism to enhance the segmentation capability of complex lesion areas, including:

[0156] Attention weight calculation: Based on the input feature map And refined segmentation results , calculate the attention weight of each pixel , expressed as:

[0157] ;

[0158] in, To query features, is the key feature, is the dimension of the feature;

[0159] Attention-weighted feature generation: Using attention weights Weight the input feature map to generate an enhanced feature map , expressed as:

[0160] ;

[0161] in, is the value feature;

[0162] Complex area optimization: Enhanced feature map It is integrated into the refined segmentation result to further improve the segmentation effect of complex lesion areas, which is expressed as:

[0163] ;

[0164] Final segmentation result output: Output the final segmentation result , including accurate segmentation of organ areas and lesion areas.

[0165] The intelligent labeling module combines the medical knowledge graph to label the anatomical name of the target area of ​​the segmentation results generated by the automatic segmentation module, including:

[0166] Segmentation result loading: receiving the segmentation results output by the automatic segmentation module ,in, including masks of organ regions and lesion regions;

[0167] Anatomical knowledge matching: Based on the medical knowledge graph, the standard names and spatial feature descriptions of anatomical regions are loaded, including the location, shape, and size range of organs;

[0168] Spatial position comparison: compare the regional features (such as center point coordinates and area) in the segmentation results with the standard regional features in the medical knowledge graph to match the most appropriate anatomical name, expressed as: ;

[0169] in, Indicates the anatomical name corresponding to the most consistent standard regional feature obtained through matching, Refers to the medical knowledge graph that includes all standard areas. It means to solve the input that makes a function reach its maximum value. represents the region similarity function, is the target area in the segmentation result, It is a standard area in the knowledge graph;

[0170] Anatomical annotation generation: Assign the anatomical names in the comparison results to the corresponding segmented regions, and attach the corresponding regional information (such as area, boundary coordinates), expressed as:

[0171] ;

[0172] in, For the final generated annotation, Indicates the anatomical name, which comes from the comparison result. represents the area of ​​the target region, Indicates the boundary information of the target area;

[0173] Labeling result output: Output the labeling results including the anatomical name for subsequent labeling of lesion type and pathological features.

[0174] The intelligent annotation module combines semantic embedding technology with the medical terminology library to annotate the lesion type and pathological characteristics of the target area, including:

[0175] Annotation model loading: Load the annotation model built based on semantic embedding technology. The model pre-training data includes standard medical terminology libraries (such as SNOMED CT, UMLS) and clinical case data of segmented areas.

[0176] Semantic feature extraction of segmentation region: from segmentation results The semantic features of the region are extracted, including geometric features, grayscale statistical features and texture features, and the encoding function is used to map the regional features to the semantic space, which is expressed as:

[0177] ;

[0178] in, is the encoding function, is the segmentation result, A semantic embedding representation representing the segmented region;

[0179] Medical terminology comparison: embedding semantics of segmented regions Semantic embedding with medical terminology Compare and calculate the matching score , expressed as:

[0180] ;

[0181] in, is the matching score, is the semantic embedding of the segmented region, For semantic embedding of medical terminology, represents cosine similarity;

[0182] Lesion type and pathological feature annotation: The best matching medical term is selected as the annotation result according to the matching score. The annotation content includes lesion type (such as mass, inflammation, ischemia) and pathological features (such as size, boundary characteristics, signal intensity);

[0183] Annotation verification and output: Perform consistency checks on the generated annotation results to ensure that the annotation content conforms to the logical relationships in the medical knowledge graph (such as the association between anatomical location and lesion type), and output the final annotation results containing lesion type and pathological characteristics for clinical use or further analysis.

[0184] The present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention. In order to make the public have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, but those skilled in the art can fully understand the present invention without the description of these details. In addition, in order to avoid unnecessary confusion about the essence of the present invention, well-known methods, processes, procedures, components and circuits are not described in detail.

[0185] The above are only preferred embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An AI-based medical image automatic segmentation and annotation system, characterized in that: It includes image data acquisition module, data preprocessing module, feature extraction module, automatic segmentation module and intelligent labeling module, among which; The image data acquisition module is used to acquire medical image data, which includes CT images, MRI images and X-ray images, and to perform format conversion and standardization processing on the acquired medical image data, so as to convert the medical image data into a unified DICOM format; The data preprocessing module is used to preprocess the collected medical image data; The feature extraction module is used to extract features from the preprocessed medical image data and extract feature maps of potential lesion areas; The automatic segmentation module uses a segmentation model constructed by a fully convolutional network combined with a bidirectional long short-term memory network to automatically segment the feature map output by the feature extraction module, and generates segmentation results including organ areas and lesion areas. The automatic segmentation module enhances the segmentation capability of complex lesion areas by introducing an attention mechanism, specifically including: Attention weight calculation: Calculate the attention weight of each pixel based on the input feature map and segmentation results; Attention-weighted feature generation: Use attention weights to weight the input feature map to generate an enhanced feature map; Complex area optimization: The enhanced feature map is integrated into the segmentation result to further improve the segmentation effect of complex lesion areas; Final segmentation result output: Output the final segmentation result, including the segmentation of organ area and lesion area; The intelligent annotation module combines the medical knowledge graph to annotate the segmentation results generated by the automatic segmentation module, and the annotation content includes the anatomical name, lesion type and pathological characteristics of the target area; The automatic segmentation module implements preliminary region segmentation through a fully convolutional network, specifically including: Input feature map loading: receiving the feature map output by the feature extraction module; Convolution operation extracts segmentation features: Use a fully convolutional network to perform pixel-level convolution operations on the input feature map to generate a segmentation feature map; Downsampling and upsampling: extract multi-scale features through downsampling operations in the fully convolutional network, and restore the segmentation features to the same spatial resolution as the original feature map through upsampling operations; Generating preliminary segmentation results: The restored segmentation feature map is processed by activation function and threshold to generate preliminary segmentation results; The automatic segmentation module combines a bidirectional long short-term memory network to refine the segmentation boundary, specifically including: Boundary feature extraction: Extract boundary features from the preliminary segmentation results; Boundary serialization: convert boundary features into boundary sequences; Boundary optimization: Use a bidirectional long short-term memory network to optimize the boundary sequence and generate an optimized boundary sequence; Optimize boundary mapping: Map the optimized boundary sequence back to the original space to generate refined segmentation boundaries; Refined segmentation result output: The optimized boundary is merged into the preliminary segmentation result to generate a refined segmentation mask.

2. The AI-based medical image automatic segmentation and annotation system according to claim 1, characterized in that: The image data acquisition module comprises: Equipment connection: The image data acquisition module establishes a connection with the medical imaging equipment through the communication protocol interface. The medical imaging equipment includes CT imaging equipment, MRI imaging equipment and X-ray imaging equipment; Real-time image acquisition: using the acquisition control program to obtain real-time generated medical image data from medical imaging equipment, including CT images, MRI images and X-ray images; Format conversion: convert the collected medical image data into a unified DICOM format; Image data standardization: unify the size, resolution and grayscale range of medical image data; Verification of acquisition results: consistency check of medical imaging data after format conversion and standardization; Output and transfer: The medical imaging data that has completed format conversion and standardization processing is transferred to the data preprocessing module.

3. The AI-based medical image automatic segmentation and annotation system according to claim 2, characterized in that: The preprocessing includes noise suppression, which detects the noise types in the medical image data, including Gaussian noise and speckle noise, uses an adaptive noise suppression algorithm to reduce noise of different types, and evaluates the noise reduction effect through peak signal-to-noise ratio and structural similarity indicators. If it does not meet the preset quality standards, adjust the algorithm parameters and reprocess.

4. The AI-based medical image automatic segmentation and annotation system according to claim 3, characterized in that: The feature extraction module comprises: Region of interest detection: Receive preprocessed medical image data, detect significant areas in the image by gradient amplitude calculation, mark areas with larger gradient amplitude as potential lesion candidate areas, screen potential lesion candidate areas based on grayscale range and connectivity analysis, merge the screened significant areas, and generate a binary mask matrix of the region of interest; Feature extraction: Calculate the geometric features and grayscale statistical features of the region of interest, and use the grayscale co-occurrence matrix to calculate the texture features of the region of interest; Feature map generation: Integrate geometric features, grayscale statistical features and texture features to generate a feature map of the potential lesion area.

5. The AI-based medical image automatic segmentation and annotation system according to claim 1, characterized in that: The intelligent labeling module combines the medical knowledge graph to label the anatomical name of the target area on the segmentation result generated by the automatic segmentation module, specifically including: Segmentation result loading: receiving the segmentation result output by the automatic segmentation module; Anatomical knowledge matching: Based on the medical knowledge graph, the standard names and spatial feature descriptions of anatomical regions are loaded, including the location, shape, and size range of organs; Spatial position comparison: compare the regional features in the segmentation results with the standard regional features in the medical knowledge graph to match the most appropriate anatomical name; Anatomical annotation generation: assign the anatomical names in the comparison results to the corresponding segmented regions and attach the corresponding regional information; Labeling result output: Output the labeling results including the anatomical name for subsequent labeling of lesion type and pathological features.

6. The AI-based medical image automatic segmentation and annotation system according to claim 5, characterized in that: The intelligent annotation module combines semantic embedding technology with the medical terminology library to annotate the lesion type and pathological characteristics of the target area, specifically including: Annotation model loading: Load the annotation model built based on semantic embedding technology. The model pre-training data includes the standard medical terminology library and clinical case data of the segmented area. Semantic feature extraction of segmented regions: Extract the semantic features of the region from the segmentation results, including geometric features, grayscale statistical features, and texture features, and use encoding functions to map the regional features to the semantic space; Medical terminology matching: compare the regional semantic embedding with the semantic embedding of the medical terminology library and calculate the matching score; Lesion type and pathological feature annotation: The best matching medical term is selected as the annotation result according to the matching score. The annotation content includes lesion type and pathological features. Annotation verification and output: Perform consistency check on the generated annotation results to ensure that the annotation content conforms to the logical relationship in the medical knowledge graph, and output the final annotation results containing lesion types and pathological characteristics.

Images