Damage image diagnosis method and system based on artificial intelligence

Abstract

The invention relates to the technical field of data processing, and discloses a damage image diagnosis method and system based on artificial intelligence. The injury image diagnosis method based on artificial intelligence comprises the following steps: acquiring sequence image data of an injured part of a patient; performing quality optimization and contrast enhancement on the sequence image data to obtain an image sequence; feature learning and target area detection are carried out on the image sequence, an anatomical structure indicating damage is automatically identified and marked, and a corresponding target area positioning frame is generated; according to the invention, the damaged area can be identified and positioned more accurately, so that the diagnosis accuracy is improved. Through space calibration and sequence image alignment, the method can accurately evaluate the change of damage along with time, and has important significance for chronic disease or curative effect monitoring. The composite feature representation fuses image features from different time points, and a comprehensive view angle is provided for understanding the overall condition and progress of damage.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to an artificial intelligence-based injury imaging diagnosis method and system. Background Art

[0002] Anterior cruciate ligament (ACL) injury of the knee is a common injury in sports activities, and timely and accurate diagnosis is of great significance for patient treatment planning and surgical intervention. Traditional diagnostic methods mainly rely on physician experience and two-dimensional imaging data, such as X-rays or MRI scans, to assess the injury. However, the limitation of these traditional imaging data is that it is difficult to fully display the precise location of the ligament injury and the complex surrounding anatomical structures, which poses a challenge to the accuracy of surgical positioning, increases the difficulty of surgery and the risk of postoperative complications. In addition, the lack of real-time navigation during surgery limits the surgeon's perception of the surgical area and accurate operation.

[0003] To address these challenges, some existing technologies attempt to improve the accuracy and intuitiveness of surgical positioning using 3D reconstruction or virtual reality. While these technologies can enhance the visualization of surgical planning to a certain extent, they still rely on the physician's interpretation of imaging data and the alignment of preoperative planning with actual intraoperative conditions. These methods are not only time-consuming but can also lead to inaccurate positioning due to differences in interpretation and fluctuating intraoperative conditions.

[0004] Existing technologies are limited in efficiency and accuracy when it comes to integrating preoperative planning with intraoperative realities. Due to a lack of advanced tools capable of processing and analyzing sequential imaging data, surgeons may not have a deep and comprehensive understanding of the surgical area, limiting the precision of surgical positioning. These technical approaches also typically lack automated recognition and classification of injury features, nor do they include mechanisms for real-time navigation, thus failing to fully meet the demands for fast and accurate surgical navigation. Summary of the Invention

[0005] The present invention provides an artificial intelligence-based injury imaging diagnostic method and system, which aims to automatically identify the injury location through deep learning analysis of serial imaging data, and provide accurate injury classification and importance assessment, thereby providing more precise surgical planning and real-time navigation for the treatment of ACL injuries, and solving the technical problems mentioned above.

[0006] A first aspect of the present invention provides an artificial intelligence-based injury imaging diagnosis method, the artificial intelligence-based injury imaging diagnosis method comprising: Acquiring sequential image data of the patient's damaged part; wherein the sequential image data includes at least two-dimensional medical images taken at two or more time periods; performing quality optimization and contrast enhancement on the sequence image data to obtain an image sequence; Performing feature learning and target region detection on the image sequence, automatically identifying and marking anatomical structures indicating damage, and generating a corresponding target region positioning framework; Performing spatial calibration and alignment on the target area positioning frame to obtain damage image features; fusing the damage image features to obtain a composite feature representation; The composite feature representation is input into the trained diagnostic model for damage classification and importance assessment to obtain damage classification results and damage assessment results; and a diagnostic result is generated based on the damage classification results and damage assessment results.

[0007] Optionally, in a first implementation of the first aspect of the present invention, the performing quality optimization and contrast enhancement on the sequence image data to obtain an image sequence includes: Denoising the sequence image data to obtain denoised sequence image data; Performing artifact removal on the denoised image sequence to obtain artifact-removed sequence image data; Performing image contrast enhancement on the image sequence after artifact removal to obtain image contrast enhanced sequence image data; Performing pseudo-color processing on the sequence image data after image contrast enhancement to obtain pseudo-color processed sequence image data; performing sequence calibration on the pseudo-color processed sequence image data to obtain sequence-calibrated sequence image data; The sequence image data after sequence calibration is subjected to sequence enhancement to obtain an image sequence.

[0008] Optionally, in a second implementation of the first aspect of the present invention, performing sequence enhancement on the sequence image data after sequence calibration to obtain an image sequence includes: Based on the preset Laplace operator, the sequence image data after sequence calibration is calculated to generate an edge image; Perform weighted fusion of sequence image data and edge image to generate weighted fusion image; Performing histogram equalization processing on the weighted fusion image to obtain a weighted fusion image after histogram equalization processing; Based on the preset rigid registration algorithm, the weighted fusion images after histogram equalization processing are rigidly registered to obtain an image sequence.

[0009] Optionally, in a third implementation of the first aspect of the present invention, performing feature learning and target region detection on the image sequence, automatically identifying and marking anatomical structures indicating damage, and generating a corresponding target region positioning framework includes: Perform feature extraction on the input image sequence to obtain the feature vector representation of each frame of image; Model the feature vector representation of each frame of image to generate the global feature representation of the image sequence; Based on the preset Faster R-CNN model, target detection is performed on the global feature representation of the input image sequence to generate the target area positioning frame; Segment the target area positioning frame, identify the bone structure of the fracture site, and mark the specific bone name to obtain anatomical structure annotation information; The target region positioning frame and the anatomical structure annotation information are integrated to generate a corresponding target region positioning frame; wherein the positioning frame includes the coordinates and size of the target region positioning frame and the corresponding anatomical structure annotation information.

[0010] Optionally, in a fourth implementation of the first aspect of the present invention, performing target detection on the global feature representation of the input image sequence based on a preset Faster R-CNN model to generate a target region positioning frame includes: Obtain a preset Faster R-CNN model; wherein the preset Faster R-CNN model includes at least a region proposal network layer, an anchor box alignment layer, a pooling layer, and a fully connected layer; Generate potential damage candidate regions based on the global feature representation of the input image sequence through the region proposal network layer, and output damage candidate region suggestions; The anchor frame alignment layer aligns the damage candidate region proposal output by the region proposal network layer with the predefined anchor frame, determines the boundary of the candidate region, and outputs the damage candidate region; Performing pooling processing on the candidate damage area through the pooling layer to generate a fixed-size regional feature map; The fully connected layer performs category determination and bounding box regression on the regional feature map to generate a target region positioning frame.

[0011] Optionally, in a fifth implementation of the first aspect of the present invention, fusing the damage image features to obtain a composite feature representation includes: Performing feature extraction at different levels on the damaged image features to obtain local features and global features of the damaged image; wherein the local features of the damaged image represent the detailed structure of the damaged image, and the global features represent the context information of the entire damaged image; Performing feature complementation on the local features and the global features to obtain complementary information; Performing feature fusion on the complementary information, the local features, and the global features, adjusting the scales of multi-source features, and obtaining fused features; The fusion features are abstracted and optimized through a preset deep learning network to generate a composite feature representation.

[0012] A second aspect of the present invention provides an artificial intelligence-based injury imaging diagnosis system, the artificial intelligence-based injury imaging diagnosis system comprising: An acquisition module, configured to acquire sequential image data of the patient's damaged part; wherein the sequential image data comprises at least two-dimensional medical images captured at two or more time periods; A contrast enhancement module, configured to optimize the quality and enhance the contrast of the sequence image data to obtain an image sequence; a detection module, configured to perform feature learning and target region detection on the image sequence, automatically identify and mark anatomical structures indicating damage, and generate a corresponding target region positioning framework; an alignment module, configured to perform spatial calibration and alignment on the target region positioning frame to obtain damage image features; A fusion module, configured to fuse the damage image features to obtain a composite feature representation; The diagnosis module is used to input the composite feature representation into the trained diagnosis model to perform damage classification and importance assessment to obtain damage classification results and damage assessment results; and generate a diagnosis result based on the damage classification results and damage assessment results.

[0013] The third aspect of the present invention provides an artificial intelligence-based injury imaging diagnostic device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the artificial intelligence-based injury imaging diagnostic device executes the above-mentioned artificial intelligence-based injury imaging diagnostic method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium, which stores instructions that, when executed on a computer, enable the computer to execute the above-mentioned artificial intelligence-based injury imaging diagnosis method.

[0015] The technical solution provided by the present invention has the following beneficial effects: The present invention provides an artificial intelligence-based injury imaging diagnosis method and system, which obtains sequential image data of a patient's damaged area; optimizes the quality and contrast of the sequential image data to obtain an image sequence; performs feature learning and target region detection on the image sequence, automatically identifies and annotates anatomical structures indicating the injury, and generates a corresponding target region positioning framework; spatially calibrates and aligns the target region positioning framework to obtain injury image features; fuses the injury image features to obtain a composite feature representation; inputs the composite feature representation into a trained diagnostic model for injury classification and importance assessment, obtaining injury classification and assessment results; and generates a diagnostic result based on the injury classification and assessment results. By optimizing the quality and contrast of the sequential image data and applying feature learning and target region detection techniques, the present invention can more accurately identify and locate the injury area, thereby improving diagnostic accuracy. By performing spatial calibration and alignment of the sequential images, the present invention can accurately assess the changes in the injury over time, which is of great significance for chronic disease or therapeutic efficacy monitoring. The composite feature representation integrates image features from different time points, providing a comprehensive perspective for understanding the overall condition and progression of the injury. By using trained diagnostic models to classify and assess the severity of injuries, doctors can make further diagnostic and treatment decisions based on the information provided, improving the level of support for medical decision-making. Accurate injury assessment and classification results, combined with comprehensive diagnostic results, can help medical personnel provide more precise treatment plans for patients and improve the overall quality of medical services. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Schematic diagram of an embodiment of an artificial intelligence-based injury imaging diagnosis method according to an embodiment of the present invention; Figure 2 Schematic diagram of an embodiment of an artificial intelligence-based injury imaging diagnostic system in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The embodiments of the present invention provide an artificial intelligence-based injury imaging diagnosis method and system. The terms "first", "second", "third", "fourth", etc. (if any) in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way are interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or are inherent to these processes, methods, products or devices.

[0018] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 An embodiment of the damage imaging diagnosis method based on artificial intelligence in the embodiment of the present invention includes: Step 101: Acquire sequential image data of a patient's damaged part; wherein the sequential image data includes at least two-dimensional medical images captured at two or more time periods; It is understandable that the execution subject of the present invention can be an artificial intelligence-based injury imaging diagnosis system, or a terminal or a server, and the specific implementation is not limited here. The embodiment of the present invention is described by taking the server as the execution subject as an example.

[0019] Specifically, in an embodiment of the present invention, the artificial intelligence-based injury imaging diagnosis method includes the steps of obtaining serial imaging data of the patient's damaged part. This process involves the following detailed steps: Step 101: Obtaining sequential image data of the patient's damaged area Function: Obtain serial image data of the patient's damaged parts from the medical image storage system, ensure the integrity and quality of the image data, and provide basic data for subsequent image processing and diagnosis.

[0020] Implementation process: Data request and retrieval: Send a data request from the hospital's Picture Archiving and Communication System (PACS) to obtain the patient's image sequence data. The PACS system is a professional system used by hospitals to store and manage medical images. It supports the DICOM (Digital Imaging and Communications in Medicine) standard to ensure data standardization and compatibility.

[0021] For example, the PACS system is used to retrieve the patient's anterior cruciate ligament injury scan sequence at different time points to ensure that the image data at multiple time points are included.

[0022] Anterior cruciate ligament (ACL) injuries of the knee can be assessed using a variety of MRI sequences. Commonly used sequence imaging data include: T1-weighted imaging: This is used to assess changes in the bone marrow and anatomy. T1 sequences are less effective than other sequences for visualizing the ACL itself but are helpful for assessing associated anatomy.

[0023] T2-weighted imaging: This is used to assess edema and fluid accumulation. T2 sequences are particularly helpful in demonstrating high signal intensity at the site of ACL injury and can pinpoint the location of fluid accumulation.

[0024] Proton Density Weighted Imaging (PD): This sequence is very useful for showing details of the ACL ligament, especially when used with fat suppression technology, it can well show subtle changes in soft tissue.

[0025] Fat-suppressed T2-weighted imaging: This technique suppresses fat signals to enhance the contrast of soft tissue lesions. Fluid signals in ACL injuries are more clearly displayed in this sequence.

[0026] STIR sequence (Short TI Inversion Recovery): is a special fat-suppressed sequence that is very sensitive to bone marrow edema and soft tissue damage.

[0027] Gradient echo imaging: Sometimes used to evaluate ligament and cartilage injuries but probably less frequently than other sequences.

[0028] Data format conversion and verification: Ensure that acquired image data is in a unified format, usually DICOM. Perform necessary format conversions to ensure data availability and integrity.

[0029] For example: Perform format verification on DICOM files extracted from the PACS system to ensure the integrity of each file and convert them into a standard format that can be used in processing software, such as the NIfTI format.

[0030] Time series data collation: The acquired imaging data should be organized in chronological order to ensure temporal consistency. Time series data should include images of the patient taken at different time points to reflect the dynamic changes of the lesion.

[0031] For example: Images of anterior cruciate ligament injuries of the knee joint at different time points are sorted according to the shooting time to ensure that the image sequence reflects the changing process of the patient's condition.

[0032] Image data preprocessing: The acquired image data is pre-processed, including image normalization and denoising, to improve the quality of the image data and provide high-quality input for subsequent processing.

[0033] For example, perform Gaussian filtering and denoising on images of anterior cruciate ligament injuries of the knee joint to standardize the grayscale value range of the image and ensure that the image data at different time points are within the same grayscale value range.

[0034] Data storage and management: The pre-processed image data is stored in the database of the image processing system to ensure data security and traceability.

[0035] For example: Store the processed image sequence of anterior cruciate ligament injury of the knee joint in the database of the image processing system, and record the relevant information of each image file, such as shooting time, image number, patient ID, etc.

[0036] Example Taking the image sequence processing of anterior cruciate ligament injury of the knee as an example, the specific implementation process is as follows: Data request and retrieval: The radiologist sends a data request to the PACS system to retrieve the patient's anterior cruciate ligament injury image series, which includes image data at multiple time points.

[0037] Data format conversion and verification: Verify the DICOM files extracted from the PACS system to ensure the integrity of the files and convert the DICOM files into NIfTI format for subsequent processing.

[0038] Time series data collation: The injured images of the anterior cruciate ligament of the knee joint at different time points were sorted according to the shooting time and organized into complete time series data.

[0039] Image data preprocessing: Gaussian filtering was used to remove noise from the injured anterior cruciate ligament of the knee joint images, and the grayscale value range was standardized to ensure the consistency of the image data.

[0040] Data storage and management: The pre-processed anterior cruciate ligament injury image sequence of the knee joint is stored in the database of the image processing system, and the relevant information is recorded to ensure the security and traceability of the data.

[0041] Step 102: Optimizing the quality and enhancing the contrast of the sequence image data to obtain an image sequence; Specifically, in an embodiment of the present invention, step 102 involves optimizing the quality and enhancing the contrast of the sequence image data to obtain a clearer image sequence. The specific steps are as follows: Step 102: Optimize the quality and enhance the contrast of the sequence image data Function: Optimize the acquired image sequence to remove noise, correct artifacts, enhance contrast, and improve the overall image quality and detail visibility.

[0042] Implementation process: Image denoising processing: Function: Remove noise from images and improve image clarity.

[0043] Implementation process: Image denoising algorithms (such as Gaussian filtering, mean filtering, and non-local mean algorithm) are used to denoise the image sequence.

[0044] Gaussian filtering: Use convolution operations to smooth images and reduce high-frequency noise.

[0045] Mean filtering: Take the average value of neighboring pixels for smoothing to reduce noise.

[0046] Non-local means algorithm: Calculates the weighted average of similar pixel blocks to remove noise while retaining image details.

[0047] Image artifact removal: Function: Eliminate artifacts in images and improve the realism of images.

[0048] Implementation process: Use artifact removal algorithms (such as Fourier transform and beamforming artifact removal algorithms) for processing.

[0049] Fourier transform: Converts the image to the frequency domain and filters out specific frequency components to remove artifacts.

[0050] Beam artifact removal algorithm: corrects the beam propagation path and eliminates image artifacts.

[0051] For example: Apply Fourier transform to MRI image sequences to filter out high-frequency artifacts and enhance image clarity.

[0052] Image contrast enhancement: Function: Enhance the contrast of the image and make the details stand out more.

[0053] Implementation process: Use contrast enhancement algorithms (such as histogram equalization and adaptive histogram equalization) for processing.

[0054] Histogram equalization: Adjust the grayscale distribution of the image to make the contrast evenly distributed.

[0055] Adaptive Histogram Equalization (CLAHE): Performs histogram equalization in local areas to avoid over-enhancement.

[0056] For example, apply CLAHE to an X-ray image sequence, set the local window size to 8×8, and enhance the image contrast.

[0057] Image pseudo-color processing: Function: Improve the visualization effect of images through pseudo-color processing technology.

[0058] Implementation process: Apply pseudo-color processing technology (such as grayscale-color mapping and heat map mapping) for processing.

[0059] Grayscale-color mapping: Map grayscale values to color space to enhance the visual effect of the image.

[0060] Heat map mapping: Use colors to represent different grayscale values to make image features more obvious.

[0061] For example, heatmap mapping is applied to an ultrasound image sequence to map the grayscale values 0-255 to a red-blue gradient to obtain a pseudo-color image.

[0062] Image sequence calibration: Function: Perform spatial and temporal calibration on image sequences to ensure image continuity and consistency.

[0063] Implementation process: Use image registration algorithms (such as rigid registration and elastic registration) for calibration.

[0064] Rigid registration: Align image sequences through operations such as translation, rotation, and scaling.

[0065] Elastic registration: Correct the image through a nonlinear deformation model to ensure image consistency.

[0066] For example, rigid registration is applied to cardiac MRI image sequences, using a mutual information-based optimization method to ensure precise spatial alignment of each image frame.

[0067] Image sequence enhancement: Function: Perform global and local enhancement on image sequences to improve the overall quality of the image.

[0068] Implementation process: Apply image enhancement algorithms (such as Laplace sharpening and contrast-limited adaptive enhancement) for processing.

[0069] Laplace sharpening: Calculates the second-order derivative of an image to enhance edge details.

[0070] Contrast-limited adaptive enhancement: Limits the contrast enhancement range to avoid over-enhancement.

[0071] For example, Laplace sharpening can be applied to PET image sequences to enhance edge details and improve image sharpness.

[0072] Example Taking the image sequence processing of anterior cruciate ligament injury of the knee as an example, the specific implementation process is as follows: Image denoising processing: Gaussian filtering was used to denoise the damaged image sequence, and the filter standard deviation σ was set to 1.5 to remove noise and obtain a smooth image.

[0073] Image artifact removal: Fourier transform method was applied to the injury image sequence to filter out high-frequency artifacts and enhance image clarity.

[0074] Image contrast enhancement: Adaptive histogram equalization (CLAHE) was applied to the lesion image sequence, with the local window size set to 8×8 to enhance the image contrast.

[0075] Image pseudo-color processing: Grayscale-color mapping is applied to the damaged image sequence, mapping the grayscale value range of 0-255 to a red-blue gradient to obtain a pseudo-color processed image.

[0076] Image sequence calibration: The lesion image series were rigidly registered using a mutual information-based optimization method to ensure accurate spatial alignment of each image frame.

[0077] Image sequence enhancement: Laplace sharpening is applied to damaged image sequences to enhance edge details and improve image sharpness.

[0078] Step 103: performing feature learning and target region detection on the image sequence, automatically identifying and marking anatomical structures indicating damage, and generating a corresponding target region positioning framework; Specifically, in an embodiment of the present invention, step 103 involves performing feature learning and target region detection on the image sequence, automatically identifying and marking the anatomical structure indicating the injury, and generating a corresponding target region positioning framework. The specific steps are as follows: Step 103: Perform feature learning and target area detection on the image sequence Function: Through deep learning algorithms, the system extracts and analyzes features of image sequences, automatically identifies and marks damaged anatomical structures, and generates a corresponding target area positioning framework.

[0079] Implementation process: Feature Learning Function: Extract depth features from image sequences to provide a basis for subsequent target area detection.

[0080] Implementation process: Use pre-trained convolutional neural networks (such as ResNet, VGG, and DenseNet) to extract features from image sequences. Convolutional neural networks gradually extract high-level features of images through multiple layers of convolution and pooling operations.

[0081] For example, the ResNet50 model is used to extract features from an input image sequence of anterior cruciate ligament injury of the knee joint, and a feature vector representation of each frame of the image is obtained.

[0082] Feature fusion and sequence modeling Function: Fuse the frame-level features of image sequences to construct a global feature representation.

[0083] Implementation process: The extracted frame-level features are modeled using a long short-term memory network (LSTM) or a temporal convolutional network (TCN) to capture the temporal information of the image sequence. This step ensures the continuity and consistency of information in the temporal dimension of the image sequence.

[0084] For example, a bidirectional LSTM network is used to model the feature vector of each frame of an image and generate a global feature representation of the image sequence.

[0085] Target area detection Function: Automatically detect and locate the target area of damage in the image sequence.

[0086] Implementation process: Use an object detection network (such as Faster R-CNN, YOLO, and SSD) to detect global features. The detection network generates candidate regions through a region proposal network (RPN), and further classifies and accurately locates the target region.

[0087] For example, the Faster R-CNN model is used to perform target detection on the global features of the input and generate a localization frame for the target area containing the damage in each frame.

[0088] Anatomical structure recognition and annotation Function: Identify and mark the anatomical structures in the target area and indicate the location of the injury.

[0089] Implementation process: Use pre-trained anatomical structure recognition models (such as U-Net and Mask R-CNN) to finely segment the target area and identify specific anatomical structures.

[0090] The identified anatomical structures are matched with predefined anatomical atlases, and the specific anatomical structure names and locations are marked.

[0091] For example: Use the U-Net model to segment the detected target area, identify the bone structure at the fracture site, and label the specific bone names.

[0092] Generate target area positioning framework Function: Generate the corresponding target area positioning framework based on the detection and recognition results.

[0093] Implementation process: The positioning frame of the target area and the anatomical structure annotation information are integrated to generate a positioning frame containing the target area position and annotation. The positioning frame includes the coordinates and size of the target area and the corresponding anatomical structure annotation information.

[0094] For example, generate a positioning frame containing the fracture site, mark the coordinates of the fracture location (x, y, w, h) and the bone name (such as "femoral neck").

[0095] Example Taking the image sequence processing of anterior cruciate ligament injury of the knee as an example, the specific implementation process is as follows: Feature Learning The ResNet50 model is used to extract features from the input anterior cruciate ligament injury image sequence to obtain the feature vector representation of each frame of the image.

[0096] Feature fusion and sequence modeling A bidirectional LSTM network is used to model the feature vector of each frame of image and generate a global feature representation of the image sequence.

[0097] Target area detection The Faster R-CNN model is used to perform target detection on the input global features and generate a localization frame of the target area containing the damage in each frame.

[0098] Anatomical structure recognition and annotation Use the U-Net model to segment the detected target area, identify the specific anatomical structure in the head image (such as a specific area of the brain), and label it.

[0099] Generate target area positioning framework Generate a localization frame containing the head injury site, annotate the coordinates (x, y, w, h) of the injury location and the name of the anatomical structure (such as "frontal lobe").

[0100] Step 104: performing spatial calibration and alignment on the target area positioning frame to obtain damage image features; Specifically, in this embodiment of the present invention, step 104 involves spatially calibrating and aligning the target region positioning frame to obtain the damaged image features. The specific steps are as follows: Step 104: Perform spatial calibration and alignment on the target area positioning frame to obtain damage image features Function: Perform precise spatial calibration and alignment on the generated target area positioning frame to ensure accurate matching between the positioning frame and the actual anatomical structure, ultimately obtaining an image that reflects the location and characteristics of the injury.

[0101] Implementation process: Spatial transformation model establishment Function: Establish a spatial transformation model in an image sequence for calibration and alignment of target areas.

[0102] Implementation process: Key points (such as anatomical landmarks or feature points) are selected and annotated in the image sequence to establish a spatial transformation model. Common methods include affine transformation, rigid transformation, and elastic transformation.

[0103] For example, in an image of anterior cruciate ligament injury, key anatomical points (such as brain boundary points and bony landmarks) are selected for annotation and a rigid transformation model is established.

[0104] Spatial calibration Function: Perform preliminary spatial calibration on the target area positioning frame to ensure the accurate position of the positioning frame in the image sequence.

[0105] Implementation process: A keypoint-based spatial calibration algorithm is used to perform preliminary calibration of the target area through translation, rotation, and scaling operations.

[0106] For example, in an image sequence of anterior cruciate ligament injury of the knee, the positioning frame is preliminarily calibrated through affine transformation based on the selected anatomical key points.

[0107] Image registration Function: Further align the target area in the image sequence to ensure the consistency of images at different time points.

[0108] Implementation process: Use image registration algorithms (such as mutual information-based and gradient descent-based registration algorithms) to align image sequences to ensure the consistency of the target area in images at different time points.

[0109] For example, in an image sequence of anterior cruciate ligament injury of the knee, an image registration algorithm based on mutual information is used to align CT images at different time points to ensure spatial consistency of the injury area.

[0110] Detail correction and optimization Function: Perform detail correction and optimization on the calibrated and aligned target areas to improve calibration accuracy.

[0111] Implementation process: Detail correction algorithms (such as thin plate spline transformation and elastic registration) are used to perform further detail correction on the target area to ensure accurate matching of the actual anatomical structure.

[0112] For example, in an image of anterior cruciate ligament injury in the knee joint, thin plate spline transformation is applied to the aligned injury area for detail correction to ensure that the injury area accurately matches the actual anatomical structure.

[0113] Feature extraction Function: Extract damage image features from calibrated and aligned images.

[0114] Implementation process: Use feature extraction algorithms (such as HOG, SIFT, and deep learning feature extraction) to extract damage features from the calibrated and aligned images, including morphological features, texture features, and spatial features.

[0115] For example, in an image of anterior cruciate ligament injury of the knee joint, a deep convolutional neural network is used to extract the morphological and texture features of the injury area and generate an injury image feature vector.

[0116] Feature integration and representation Function: Integrate the extracted damage features to form a comprehensive damage image feature representation.

[0117] Implementation process: The extracted morphological features, texture features and spatial features are integrated, and feature fusion algorithms (such as feature splicing and feature weighted fusion) are used to generate comprehensive feature representation.

[0118] For example, the morphological features and texture features extracted from the injury image of the anterior cruciate ligament of the knee are weightedly fused to generate a comprehensive injury image feature representation for subsequent diagnosis and analysis.

[0119] Example Taking the image sequence processing of anterior cruciate ligament injury of the knee as an example, the specific implementation process is as follows: Spatial transformation model establishment Brain boundary points and bony landmarks were selected in the anterior cruciate ligament injury image of the knee joint to establish a rigid transformation model.

[0120] Spatial calibration Based on the selected anatomical key points, the positioning frame is preliminarily calibrated through affine transformation.

[0121] Image registration A mutual information-based image registration algorithm is used to align CT images at different time points to ensure spatial consistency of the damaged area.

[0122] Detail correction and optimization Thin plate spline transformation is applied to the aligned lesion areas for detail correction, ensuring an accurate match with the actual anatomy.

[0123] Feature extraction A deep convolutional neural network is used to extract the morphological and texture features of the damaged area from the calibrated and aligned images.

[0124] Feature integration and representation The extracted morphological features and texture features are weightedly fused to generate a comprehensive damage image feature representation for subsequent diagnosis and analysis.

[0125] Step 105: Fusing the damage image features to obtain a composite feature representation; Specifically, in the embodiment of the present invention, step 105 involves fusing the damage image features to obtain a composite feature representation. The specific steps are as follows: Step 105: Fusing the damage image features to obtain a composite feature representation Function: Fuse various damage features extracted from images to generate a composite feature representation containing multi-dimensional information, providing a rich feature basis for further diagnosis and analysis.

[0126] Implementation process: Feature classification and standardization Function: Classify and standardize different types of features.

[0127] Implementation process: The extracted damage features are divided into categories such as morphological features, texture features, and spatial features.

[0128] All kinds of features are standardized so that they are in the same numerical range, ensuring that different features have the same dimension during the fusion process.

[0129] For example, the grayscale value range of the morphological features is normalized to [0, 1], and the gray-level co-occurrence matrix (GLCM) value of the texture features is normalized.

[0130] Feature dimensionality reduction Function: Perform dimensionality reduction on high-dimensional features, reduce redundant information, and retain the main features.

[0131] Implementation process: Use principal component analysis (PCA) or linear discriminant analysis (LDA) to reduce the dimensionality of high-dimensional features and retain the main feature components.

[0132] For example, PCA is used to reduce the dimensionality of high-dimensional texture features, retaining the first few principal components with 95% variance.

[0133] Feature fusion Function: Fuse different types of features to generate comprehensive feature representation.

[0134] Implementation process: Use feature fusion methods (such as feature splicing, feature weighted fusion, and feature voting fusion) to fuse morphological features, texture features, and spatial features.

[0135] For example: The morphological feature vector, texture feature vector and spatial feature vector are spliced into a comprehensive feature vector through feature splicing.

[0136] Feature weighting Function: Assign weights to different features according to their importance and generate weighted feature representations.

[0137] Implementation process: Different weights are assigned to different features based on their importance in the classifier (e.g., feature importance evaluation by random forest or XGBoost models).

[0138] For example, based on the feature importance scores of the random forest model, different weights are assigned to morphological features, texture features, and spatial features to generate weighted feature representations.

[0139] Feature fusion representation Function: Generate the final composite feature representation for subsequent classification and analysis.

[0140] Implementation process: The weighted features are integrated to generate a composite feature representation containing multi-dimensional information.

[0141] For example: Integrate the weighted morphological features, texture features, and spatial feature vectors into a composite feature vector as the final composite feature representation.

[0142] Example Taking the image sequence processing of anterior cruciate ligament injury of the knee as an example, the specific implementation process is as follows: Feature classification and standardization The morphological features, texture features and spatial features extracted from the injured images of the anterior cruciate ligament of the knee are classified and standardized so that the numerical range of all features is unified in [0,1].

[0143] Feature dimensionality reduction PCA is used to reduce the dimensionality of high-dimensional texture features, retaining the first few principal components with 95% variance, reducing feature dimensions and retaining the main information.

[0144] Feature fusion Through the feature splicing method, the standardized morphological feature vector, the texture feature vector after dimensionality reduction and the spatial feature vector are spliced into a comprehensive feature vector.

[0145] Feature weighting According to the feature importance evaluation of the random forest model, different weights are assigned to the morphological features, texture features and spatial features to generate weighted feature representation.

[0146] Feature fusion representation The weighted morphological features, texture features and spatial feature vectors are integrated into a composite feature vector as the final composite feature representation.

[0147] Step 106: Input the composite feature representation into the trained diagnostic model to perform damage classification and importance assessment to obtain damage classification results and damage assessment results; and generate a diagnostic result based on the damage classification results and damage assessment results.

[0148] Specifically, in an embodiment of the present invention, step 106 involves inputting the composite feature representation into the trained diagnostic model to perform damage classification and importance assessment, obtaining damage classification results and damage assessment results, and generating a diagnostic result based on the results. The specific steps are as follows: Step 106: Input the composite feature representation into the trained diagnostic model to perform damage classification and importance assessment, and obtain damage classification results and damage assessment results; generate a diagnostic result based on the damage classification results and damage assessment results. Function: Classify and assess the importance of injuries through pre-trained diagnostic models, generate comprehensive diagnostic results, and provide a basis for clinical decision-making.

[0149] Implementation process: Composite feature input Function: Input composite feature representation into the trained diagnostic model for processing.

[0150] Implementation process: The composite feature representation generated in step 105 (a comprehensive vector containing morphological features, texture features, and spatial features) is input into a trained deep learning model or machine learning classifier.

[0151] For example, the composite feature vector extracted and fused from the anterior cruciate ligament injury image of the knee joint is input into the trained ResNet classification model.

[0152] Injury classification Function: Classify the damage type based on the input composite feature representation.

[0153] Implementation process: Use a trained classification model (such as convolutional neural network, support vector machine, random forest, etc.) to classify the input features and output the probability distribution of damage types.

[0154] For example, use a trained ResNet model to classify the input composite features to obtain the probability distribution of head injury types, such as concussion and intracranial hemorrhage.

[0155] Materiality Assessment Function: Assess the relative importance and severity of each injury type.

[0156] Implementation process: The severity and importance of each injury type are evaluated based on the output probability and feature importance index of the classification model.

[0157] For example, based on the output probability of the classification model and the feature importance score, the severity of head injury can be evaluated, such as the severity classification of concussion.

[0158] Generate classification and evaluation results Function: Generate damage classification results and importance assessment results, providing basic data for generating diagnosis results.

[0159] Implementation process: The final damage classification and assessment results are determined based on the classification probability and importance evaluation score output by the model.

[0160] For example, based on the probability distribution output by the model, the injury classification result of the anterior cruciate ligament injury image of the knee joint is determined (such as a confirmed concussion) and a severity assessment score is provided.

[0161] Diagnosis result generation Function: Generate comprehensive diagnostic results based on damage classification and importance assessment results.

[0162] Implementation process: The injury classification and assessment results are combined to generate a detailed diagnosis report. The report should include information such as injury type, injury location, severity, and clinical recommendations.

[0163] For example, a diagnosis result of an injury image of the anterior cruciate ligament of the knee joint is generated, and the report content includes: a diagnosis of concussion, the injury site is located in the right temporal lobe, the severity is graded as moderate, and further observation and reexamination are recommended.

[0164] Example Taking the image sequence processing of anterior cruciate ligament injury of the knee as an example, the specific implementation process is as follows: Composite feature input The composite feature vector extracted and fused from the anterior cruciate ligament injury image of the knee joint is input into the trained ResNet classification model.

[0165] Injury classification Use the trained ResNet model to classify the input composite features and output the probability distribution of head injury types, such as concussion and intracranial hemorrhage.

[0166] Materiality Assessment Based on the output probability and feature importance index of the classification model, the severity and importance of each injury type are evaluated, such as the severity classification of concussion.

[0167] Generate classification and evaluation results According to the probability distribution output by the model, the injury classification result of the anterior cruciate ligament of the knee injury image is determined, and a severity assessment score is provided. If a concussion is confirmed, the severity grade is moderate.

[0168] Diagnosis result generation The injury classification and assessment results were combined to generate a detailed diagnosis. The report included: a confirmed concussion, injury located in the right temporal lobe, moderate severity, and a recommendation for further observation and reexamination.

[0169] Another embodiment of the artificial intelligence-based injury imaging diagnosis method according to the embodiment of the present invention includes: optimizing the quality and enhancing the contrast of the sequence image data to obtain an image sequence, including: Denoising the sequence image data to obtain denoised sequence image data; Performing artifact removal on the denoised image sequence to obtain artifact-removed sequence image data; Performing image contrast enhancement on the image sequence after artifact removal to obtain image contrast enhanced sequence image data; Performing pseudo-color processing on the sequence image data after image contrast enhancement to obtain pseudo-color processed sequence image data; performing sequence calibration on the pseudo-color processed sequence image data to obtain sequence-calibrated sequence image data; The sequence image data after sequence calibration is subjected to sequence enhancement to obtain an image sequence.

[0170] Specifically, the implementation process: Image denoising Function: Remove noise from images and improve image clarity.

[0171] Implementation process: Use Gaussian filtering to denoise the image sequence. Gaussian filtering smoothes the image through convolution operations and reduces high-frequency noise.

[0172] For example, apply Gaussian filtering to the image sequence of anterior cruciate ligament injury of the knee joint, set the filter standard deviation σ to 1.5, remove noise, and obtain the denoised sequence image data.

[0173] Artifact Removal Function: Eliminate artifacts in images and improve the realism of images.

[0174] Implementation process: Artifact removal is performed using Fourier transform technology. Fourier transform converts the image into the frequency domain and filters out specific frequency components to remove artifacts.

[0175] For example: Apply the Fourier transform method to the head MRI image sequence to filter out high-frequency artifacts and obtain the artifact-removed sequence image data.

[0176] Image contrast enhancement Function: Enhance the contrast of the image and make the details stand out more.

[0177] Implementation process: The image sequence is contrast enhanced using adaptive histogram equalization (CLAHE). CLAHE performs histogram equalization in local regions to avoid over-enhancement and improve image contrast.

[0178] For example, CLAHE is applied to an X-ray image sequence, the local window size is set to 8×8, and the image contrast is enhanced to obtain the sequence image data after image contrast enhancement.

[0179] Image pseudo-color processing Function: Improve the visualization effect of images through pseudo-color processing technology.

[0180] Implementation process: Apply pseudo-color processing techniques (such as grayscale-color mapping and heat map mapping) to process the image. Grayscale-color mapping maps grayscale values to color space to enhance the visual effect of the image.

[0181] For example, heat map mapping is applied to an ultrasound image sequence, and the grayscale value 0-255 is mapped to a red-blue gradient to obtain the pseudo-color processed sequence image data.

[0182] Image sequence calibration Function: Perform spatial and temporal calibration on image sequences to ensure image continuity and consistency.

[0183] Implementation process: Calibration is performed using image registration algorithms (such as rigid registration and elastic registration). Rigid registration aligns image sequences through translation, rotation, scaling, and other operations. Elastic registration corrects images using a nonlinear deformation model to ensure image consistency.

[0184] For example, rigid registration is applied to cardiac MRI image sequences, and a mutual information-based optimization method is used to ensure that each frame of the image is accurately aligned in space, thereby obtaining sequence-calibrated image data.

[0185] Image sequence enhancement Function: Perform global and local enhancement on image sequences to improve the overall quality of the image.

[0186] Implementation process: Apply image enhancement algorithms (such as Laplace sharpening and contrast-limited adaptive enhancement) to the image. Laplace sharpening enhances edge details by calculating the second-order derivative of the image. Contrast-limited adaptive enhancement limits the contrast enhancement range to avoid over-enhancement.

[0187] For example, Laplace sharpening is applied to a PET image sequence to enhance edge details and improve image sharpness, thereby obtaining an image sequence.

[0188] Example Taking the image sequence processing of anterior cruciate ligament injury of the knee as an example, the specific implementation process is as follows: Image denoising The damaged image sequence was denoised using Gaussian filtering, and the filter standard deviation σ was set to 1.5 to remove noise and obtain the denoised sequence image data.

[0189] Artifact Removal The Fourier transform method is applied to the damage image sequence to filter out high-frequency artifacts and obtain the artifact-removed sequence image data.

[0190] Image contrast enhancement Adaptive histogram equalization (CLAHE) was applied to the damage image sequence, and the local window size was set to 8×8 to enhance the image contrast. The sequence image data after image contrast enhancement was obtained.

[0191] Image pseudo-color processing Grayscale-color mapping is applied to the damaged image sequence, and the grayscale value range of 0-255 is mapped to a red-blue gradient to obtain the pseudo-color processed sequence image data.

[0192] Image sequence calibration The injury image sequence is rigidly registered, and an optimization method based on mutual information is used to ensure the precise spatial alignment of each frame of the image, thereby obtaining sequence image data after sequence calibration.

[0193] Image sequence enhancement Laplace sharpening is applied to the damaged image sequence to enhance edge details and improve image sharpness to obtain an image sequence.

[0194] Through the above detailed steps, the embodiment of the present invention achieves quality optimization and contrast enhancement of the image sequence, ensures high quality and high availability of the image, and provides a reliable imaging basis for subsequent artificial intelligence diagnosis.

[0195] Another embodiment of the artificial intelligence-based injury imaging diagnosis method according to the embodiment of the present invention includes: performing sequence enhancement on the sequence image data after sequence calibration to obtain an image sequence, including: Based on the preset Laplace operator, the sequence image data after sequence calibration is calculated to generate an edge image; Perform weighted fusion of sequence image data and edge image to generate weighted fusion image; Performing histogram equalization processing on the weighted fusion image to obtain a weighted fusion image after histogram equalization processing; Based on the preset rigid registration algorithm, the weighted fusion images after histogram equalization processing are rigidly registered to obtain an image sequence.

[0196] Specifically, the implementation process: Edge image generation Function: Based on the preset Laplace operator, perform edge detection on the calibrated image data and generate an edge image.

[0197] Implementation process: The Laplacian operator is applied to perform convolution operation on the calibrated image data and the second-order derivative is calculated to detect the edges in the image.

[0198] For example, the Laplacian operator is applied to an image sequence of anterior cruciate ligament injury of the knee joint to generate an edge image containing image edge information.

[0199] Weighted Fusion Function: Perform weighted fusion of sequence image data and edge image to generate a fused image with greater clarity and edge details.

[0200] Implementation process: Weights are assigned to the sequence image data and edge images respectively, and the two are linearly weighted fused to enhance image details.

[0201] For example, set the weight of the original image data to 0.7 and the weight of the edge image to 0.3, and perform weighted fusion of the two to generate a weighted fused image.

[0202] Histogram equalization Function: Perform histogram equalization on the weighted fusion image to improve the overall contrast of the image.

[0203] Implementation process: The histogram equalization technology is applied to redistribute the grayscale values of the weighted fusion image to balance the grayscale distribution of the image and improve the contrast.

[0204] For example, histogram equalization is performed on a weighted fusion image sequence of anterior cruciate ligament injury of the knee joint to make the grayscale value distribution of the image more uniform and the details clearer.

[0205] Rigid registration Function: Based on the preset rigid registration algorithm, the weighted fused images after histogram equalization are spatially aligned to ensure the continuity and consistency of the images.

[0206] Implementation process: A rigid registration algorithm is applied to accurately align the image sequence through translation, rotation and scaling operations to ensure the spatial consistency of each frame in the sequence.

[0207] For example, a mutual information-based rigid registration algorithm is applied to the processed image sequence of anterior cruciate ligament injury of the knee joint to ensure the precise spatial alignment of each frame of the image and generate a consistent image sequence.

[0208] Example Taking the image sequence processing of anterior cruciate ligament injury of the knee as an example, the specific implementation process is as follows: Edge image generation The Laplace operator is used to perform edge detection on the injured image data of the anterior cruciate ligament of the knee joint after sequence calibration to generate an edge image containing image edge information.

[0209] Weighted Fusion The original image data and the edge image are assigned weights of 0.7 and 0.3 respectively, and linear weighted fusion is performed to generate a weighted fused image with clarity and edge details.

[0210] Histogram equalization The weighted fused image is processed with histogram equalization to enhance the overall contrast of the image and make the details more prominent.

[0211] Rigid registration A rigid registration algorithm based on mutual information is used to spatially align the processed anterior cruciate ligament injury image sequence to ensure the spatial consistency of each frame of the image and generate the final image sequence.

[0212] Through the above detailed steps, the embodiment of the present invention realizes the enhancement processing of the image data after sequence calibration, ensures the high quality and high availability of the image, and provides a reliable imaging basis for subsequent artificial intelligence diagnosis.

[0213] Another embodiment of the artificial intelligence-based injury imaging diagnosis method according to the embodiment of the present invention includes: performing feature learning and target region detection on the image sequence, automatically identifying and marking anatomical structures indicating injury, and generating a corresponding target region positioning framework, including: Perform feature extraction on the input image sequence to obtain the feature vector representation of each frame of image; Model the feature vector representation of each frame of image to generate the global feature representation of the image sequence; Based on the preset Faster R-CNN model, target detection is performed on the global feature representation of the input image sequence to generate the target area positioning frame; Segment the target area positioning frame, identify the bone structure of the fracture site, and mark the specific bone name to obtain anatomical structure annotation information; The target region positioning frame and the anatomical structure annotation information are integrated to generate a corresponding target region positioning frame; wherein the positioning frame includes the coordinates and size of the target region positioning frame and the corresponding anatomical structure annotation information.

[0214] Specifically, in an embodiment of the present invention, another embodiment of the artificial intelligence-based injury imaging diagnosis method includes performing feature learning and target region detection on an image sequence, automatically identifying and marking anatomical structures indicating injury, and generating a corresponding target region positioning framework. The specific steps are as follows: Steps: Perform feature learning and target area detection on the image sequence, automatically identify and mark the anatomical structure indicating the injury, and generate the corresponding target area positioning framework Function: Through deep learning algorithms, the system extracts and analyzes features of image sequences, automatically identifies and marks damaged anatomical structures, and generates a corresponding target area positioning framework.

[0215] Implementation process: Feature extraction Function: Extract depth features from image sequences to provide a basis for subsequent target area detection.

[0216] Implementation process: Use pre-trained convolutional neural networks (such as ResNet, VGG, and DenseNet) to extract features from image sequences. Convolutional neural networks gradually extract high-level features of images through multiple layers of convolution and pooling operations.

[0217] For example, the ResNet50 model is used to extract features from an input image sequence of anterior cruciate ligament injury of the knee joint, and a feature vector representation of each frame of the image is obtained.

[0218] Global feature representation modeling Function: Model the feature vector of each frame of image and generate a global feature representation of the image sequence.

[0219] Implementation process: The extracted frame-level features are modeled using a long short-term memory network (LSTM) or a temporal convolutional network (TCN) to capture the temporal information of the image sequence. This step ensures the continuity and consistency of information in the temporal dimension of the image sequence.

[0220] For example, a bidirectional LSTM network is used to model the feature vector of each frame of an image and generate a global feature representation of the image sequence.

[0221] Target area detection Function: Automatically detect and locate the target area of damage in the image sequence.

[0222] Implementation process: Use an object detection network (such as Faster R-CNN, YOLO, and SSD) to detect global features. The detection network generates candidate regions through a region proposal network (RPN), and further classifies and accurately locates the target region.

[0223] For example, the Faster R-CNN model is used to perform target detection on the global features of the input and generate a localization frame for the target area containing the damage in each frame.

[0224] Anatomical structure recognition and annotation Function: Identify and mark the anatomical structures in the target area and indicate the location of the injury.

[0225] Implementation process: Use pre-trained anatomical structure recognition models (such as U-Net and Mask R-CNN) to finely segment the target area and identify specific anatomical structures.

[0226] The identified anatomical structures are matched with predefined anatomical atlases, and the specific anatomical structure names and locations are marked.

[0227] For example: Use the U-Net model to segment the detected target area, identify the bone structure at the fracture site, and label the specific bone names.

[0228] Generate target area positioning framework Function: Generate the corresponding target area positioning framework based on the detection and recognition results.

[0229] Implementation process: The positioning frame of the target area and the anatomical structure annotation information are integrated to generate a positioning frame containing the target area position and annotation. The positioning frame includes the coordinates and size of the target area and the corresponding anatomical structure annotation information.

[0230] For example, generate a positioning frame containing the fracture site, mark the coordinates of the fracture location (x, y, w, h) and the bone name (such as "femoral neck").

[0231] Example Taking the image sequence processing of anterior cruciate ligament injury of the knee as an example, the specific implementation process is as follows: Feature extraction The ResNet50 model is used to extract features from the input anterior cruciate ligament injury image sequence to obtain the feature vector representation of each frame of the image.

[0232] Global feature representation modeling A bidirectional LSTM network is used to model the feature vector of each frame of image and generate a global feature representation of the image sequence.

[0233] Target area detection The Faster R-CNN model is used to perform target detection on the input global features and generate a localization frame of the target area containing the damage in each frame.

[0234] Anatomical structure recognition and annotation Use the U-Net model to segment the detected target area, identify the specific anatomical structure in the head image (such as a specific area of the brain), and label it.

[0235] Generate target area positioning framework Generate a localization frame containing the head injury site, annotate the coordinates (x, y, w, h) of the injury location and the name of the anatomical structure (such as "frontal lobe").

[0236] Through the above detailed steps, the embodiment of the present invention realizes feature learning and target area detection of image sequences, automatically identifies and labels anatomical structures indicating injuries, generates a corresponding target area positioning framework, and provides accurate positioning and labeling information for subsequent injury diagnosis and treatment.

[0237] Another embodiment of the artificial intelligence-based injury imaging diagnosis method according to the embodiment of the present invention includes: performing target detection on the global feature representation of the input image sequence based on the preset Faster R-CNN model to generate a target area positioning frame, including: Obtain a preset Faster R-CNN model; wherein the preset Faster R-CNN model includes at least a region proposal network layer, an anchor box alignment layer, a pooling layer, and a fully connected layer; Generate potential damage candidate regions based on the global feature representation of the input image sequence through the region proposal network layer, and output damage candidate region suggestions; The anchor frame alignment layer aligns the damage candidate region proposal output by the region proposal network layer with the predefined anchor frame, determines the boundary of the candidate region, and outputs the damage candidate region; Performing pooling processing on the candidate damage area through the pooling layer to generate a fixed-size regional feature map; The fully connected layer performs category determination and bounding box regression on the regional feature map to generate a target region positioning frame.

[0238] Specifically, in an embodiment of the present invention, another embodiment of the artificial intelligence-based injury imaging diagnosis method includes performing target detection on the global feature representation of the input image sequence based on a preset Faster R-CNN model to generate a target region positioning frame. The specific steps are as follows: Steps: Based on the preset Faster R-CNN model, target detection is performed on the global feature representation of the input image sequence to generate the target area positioning frame Function: Use the Faster R-CNN model to detect global features in image sequences and automatically generate target region positioning boxes for further analysis and annotation of anatomical structures.

[0239] Implementation process: Get the preset Faster R-CNN model Function: Load and initialize the preset Faster R-CNN model, including its main components.

[0240] Implementation process: Obtain the Faster R-CNN model from the pre-trained model library and load its weight parameters. This model includes at least the following main components: Region Proposal Network (RPN) layer, anchor box alignment layer, pooling layer, and fully connected layer.

[0241] Example: Load the pre-trained Faster R-CNN model, whose base network is ResNet50, and load the model's weight file.

[0242] The region proposal network layer generates candidate regions Function: Generate potential damage candidate regions based on feature representation through the region proposal network layer and output damage candidate region proposals.

[0243] Implementation process: The global feature representation of the input image sequence is input to the region proposal network layer. RPN generates a series of anchor boxes on the feature map through a sliding window and evaluates whether each anchor box contains the target area.

[0244] Each anchor box is classified and regressed through the Softmax classifier and bounding box regressor to output potential damage candidate areas.

[0245] For example, perform RPN processing on the global feature map of the anterior cruciate ligament injury image of the knee joint to generate several potential candidate injury region suggestions.

[0246] Anchor box alignment layer performs anchor box alignment Function: Align the damage candidate region proposal output by the region proposal network layer with the pre-defined anchor box to determine the boundary of the candidate region.

[0247] Implementation process: Align the anchor boxes of the damage candidate regions output by the RPN. Calculate the overlap between the proposed region and the predefined anchor boxes, and select the optimal candidate region by the maximum overlap (IoU).

[0248] Output the aligned damage candidate regions, including accurate bounding box information.

[0249] For example, in the candidate region proposal of anterior cruciate ligament injury images of the knee joint, the region with the highest overlap with the predefined anchor point box is selected as the final candidate region.

[0250] The pooling layer generates a fixed-size regional feature map Function: Pool the candidate damage areas to generate a fixed-size regional feature map.

[0251] Implementation process: ROI Pooling (Region of Interest Pooling) is used to map the aligned candidate regions to a fixed-size feature map. ROI Pooling scales the candidate regions to the same size through segmentation and pooling operations, making it easier to process them later.

[0252] For example, the candidate region detected in the anterior cruciate ligament injury image of the knee joint is mapped into a fixed-size feature map of 7×7 through ROI Pooling.

[0253] Fully connected layers perform category determination and bounding box regression Function: Perform category determination and bounding box regression on the regional feature map through the fully connected layer to generate the target area positioning box.

[0254] Implementation process: The pooled region feature map is fed into the fully connected layer for category determination and bounding box regression. Category determination outputs the classification probability of each region, and bounding box regression outputs the precise bounding box coordinates of each region.

[0255] For example, the regional feature map of the anterior cruciate ligament injury image of the knee joint is classified to determine whether the area is an injury area, and its bounding box coordinates are regressed to generate the final target area positioning frame.

[0256] Example Taking the image sequence processing of anterior cruciate ligament injury of the knee as an example, the specific implementation process is as follows: Get the preset Faster R-CNN model Load the Faster R-CNN model from the pre-trained model library, using the ResNet50 network as the base network, and load the weight file.

[0257] The region proposal network layer generates candidate regions The global feature representation of the anterior cruciate ligament injury image of the knee joint is input into the RPN layer to generate several candidate injury region proposals.

[0258] Anchor box alignment layer performs anchor box alignment The candidate regions output by the RPN layer are aligned with the anchor boxes, the region with the highest overlap with the predefined anchor box is selected, and the aligned candidate regions are output.

[0259] The pooling layer generates a fixed-size regional feature map The aligned candidate regions are mapped into a fixed-size feature map of 7×7 through ROI Pooling.

[0260] Fully connected layers perform category determination and bounding box regression The regional feature map is input into the fully connected layer for category determination and bounding box regression to generate the target area positioning box of the head injury.

[0261] Through the above detailed steps, the embodiment of the present invention realizes target detection in image sequences based on the Faster R-CNN model and automatically generates target area positioning frames, providing a basis for subsequent anatomical structure recognition and annotation.

[0262] Another embodiment of the artificial intelligence-based injury image diagnosis method according to the embodiment of the present invention includes: fusing the injury image features to obtain a composite feature representation, including: Performing feature extraction at different levels on the damaged image features to obtain local features and global features of the damaged image; wherein the local features of the damaged image represent the detailed structure of the damaged image, and the global features represent the context information of the entire damaged image; Performing feature complementation on the local features and the global features to obtain complementary information; Performing feature fusion on the complementary information, the local features, and the global features, adjusting the scales of multi-source features, and obtaining fused features; The fusion features are abstracted and optimized through a preset deep learning network to generate a composite feature representation.

[0263] Specifically, in an embodiment of the present invention, another embodiment of the damage image diagnosis method based on artificial intelligence includes fusing damage image features to obtain a composite feature representation. The specific steps are as follows: Step: Fuse the damage image features to obtain composite feature representation Function: Through feature extraction and feature fusion technologies at different levels, the local and global features of the injury image are integrated to generate a composite feature representation containing rich information to improve the accuracy and robustness of diagnosis.

[0264] Implementation process: Local feature and global feature extraction Function: Extract features of damaged images at different levels to obtain local and global features of damaged images.

[0265] Implementation process: Local feature extraction: Use shallow convolutional layers in a convolutional neural network (CNN) to extract detailed structural features of the image. These features typically include details such as edges and textures.

[0266] Global feature extraction: Use deep convolutional layers in CNN to extract the overall contextual information of the image. These features usually contain the macroscopic structure and global layout of the image.

[0267] For example, when processing images of anterior cruciate ligament injuries in the knee, shallow convolutional layers extract local features, such as the detailed structure of the fracture; deep convolutional layers extract global features, such as the overall shape of the skull and the overall layout of the brain.

[0268] Complementary features Function: Complement local features and global features to obtain complementary information.

[0269] Implementation process: Through feature complementation technology, the detail information in local features is fused with the context information in global features to enhance the richness and completeness of feature representation.

[0270] For example, the local fracture details of anterior cruciate ligament injury images of the knee are complemented with the global skull shape information, so that the feature representation contains both details and overall structure.

[0271] Feature fusion Function: Fuse complementary information, local features and global features, adjust the scale of multi-source features, and obtain fused features.

[0272] Implementation process: Use feature fusion algorithms (such as feature concatenation and weighted feature fusion) to fuse complementary information, local features, and global features. Feature concatenation concatenates different feature vectors to form a new feature vector; weighted feature fusion performs a weighted average of features based on their importance.

[0273] For example, for an image of anterior cruciate ligament injury of the knee joint, local features, global features, and complementary information are spliced together to form a fused feature vector containing multi-source information.

[0274] Feature scale adjustment Function: Adjust the scale of multi-source features so that they can be fused at the same scale.

[0275] Implementation process: Through normalization technology or scale transformation method, the scales of different features are adjusted so that the fused feature vectors are within the same scale range, which is convenient for subsequent processing.

[0276] For example, the fusion features of the anterior cruciate ligament injury image of the knee joint are normalized and the feature values are adjusted to the same range to ensure that different features are fused at the same scale.

[0277] Feature abstraction and optimization Function: Abstract and optimize fusion features through a preset deep learning network to generate composite feature representations.

[0278] Implementation process: Use pre-trained deep learning networks (such as multi-layer perceptrons and deep convolutional neural networks) to further abstract and optimize the fused features. Through multi-layer nonlinear transformations, high-level abstract features are extracted, enhancing the discriminative power of feature representation.

[0279] For example, the damaged image features of the anterior cruciate ligament of the knee joint after fusion are input into a deep convolutional neural network. After processing through multiple layers of convolution and fully connected layers, a composite feature representation containing rich information is generated.

[0280] Example Taking the image processing of anterior cruciate ligament injury of the knee as an example, the specific implementation process is as follows: Local feature and global feature extraction Using the pre-trained ResNet50 model, the shallow convolutional layers extract local features (such as fracture details) and the deep convolutional layers extract global features (such as the overall shape of the skull).

[0281] Complementary features The fracture details in the local features are complemented with the skull shape in the global features to obtain complementary features containing both details and overall information.

[0282] Feature fusion Through the feature splicing method, local features, global features and complementary features are spliced to form a fused feature vector.

[0283] Feature scale adjustment The fused feature vector is normalized to adjust the feature values to the same scale range.

[0284] Feature abstraction and optimization The normalized fused feature vector is input into the deep convolutional neural network, and after processing by multiple layers of convolution and fully connected layers, a composite feature representation containing rich information is generated.

[0285] The above describes the damage imaging diagnosis method based on artificial intelligence in the embodiment of the present invention. The following describes the damage imaging diagnosis system based on artificial intelligence in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, an artificial intelligence-based injury imaging diagnosis system includes: An acquisition module, configured to acquire sequential image data of the patient's damaged part; wherein the sequential image data comprises at least two-dimensional medical images captured at two or more time periods; A contrast enhancement module, configured to optimize the quality and enhance the contrast of the sequence image data to obtain an image sequence; a detection module, configured to perform feature learning and target region detection on the image sequence, automatically identify and mark anatomical structures indicating damage, and generate a corresponding target region positioning framework; an alignment module, configured to perform spatial calibration and alignment on the target region positioning frame to obtain damage image features; A fusion module, configured to fuse the damage image features to obtain a composite feature representation; The diagnosis module is used to input the composite feature representation into the trained diagnosis model to perform damage classification and importance assessment to obtain damage classification results and damage assessment results; and generate a diagnosis result based on the damage classification results and damage assessment results.

[0286] The present invention also provides an artificial intelligence-based injury imaging diagnostic device, which includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor executes the steps of the artificial intelligence-based injury imaging diagnostic method in the above-mentioned embodiments.

[0287] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of the artificial intelligence-based injury imaging diagnosis method.

[0288] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0289] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0290] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An artificial intelligence-based injury imaging diagnosis method, characterized in that: The artificial intelligence-based injury imaging diagnosis method comprises the following steps: Acquiring sequential image data of the patient's damaged part; wherein the sequential image data includes at least two-dimensional medical images taken at two or more time periods; performing quality optimization and contrast enhancement on the sequence image data to obtain an image sequence; Performing feature learning and target region detection on the image sequence, automatically identifying and marking anatomical structures indicating damage, and generating a corresponding target region positioning framework; Performing spatial calibration and alignment on the target area positioning frame to obtain damage image features; fusing the damage image features to obtain a composite feature representation; The composite feature representation is input into the trained diagnostic model for damage classification and importance assessment to obtain damage classification results and damage assessment results; a diagnostic result is generated based on the damage classification results and damage assessment results; wherein the diagnostic result is a probability value for predicting the injury disease.

2. The artificial intelligence-based injury imaging diagnosis method according to claim 1, characterized in that: The step of optimizing the quality and enhancing the contrast of the sequence image data to obtain an image sequence includes: Denoising the sequence image data to obtain denoised sequence image data; Performing artifact removal on the denoised image sequence to obtain artifact-removed sequence image data; Performing image contrast enhancement on the image sequence after artifact removal to obtain image contrast enhanced sequence image data; Performing pseudo-color processing on the sequence image data after image contrast enhancement to obtain pseudo-color processed sequence image data; performing sequence calibration on the pseudo-color processed sequence image data to obtain sequence-calibrated sequence image data; The sequence image data after sequence calibration is subjected to sequence enhancement to obtain an image sequence.

3. The artificial intelligence-based injury imaging diagnosis method according to claim 2, characterized in that: The step of performing sequence enhancement on the sequence image data after sequence calibration to obtain an image sequence includes: Based on the preset Laplace operator, the sequence image data after sequence calibration is calculated to generate an edge image; Perform weighted fusion of sequence image data and edge image to generate weighted fusion image; Performing histogram equalization processing on the weighted fusion image to obtain a weighted fusion image after histogram equalization processing; Based on the preset rigid registration algorithm, the weighted fusion images after histogram equalization processing are rigidly registered to obtain an image sequence.

4. The artificial intelligence-based injury imaging diagnosis method according to claim 1, characterized in that: The performing of feature learning and target region detection on the image sequence, automatically identifying and marking anatomical structures indicating damage, and generating a corresponding target region positioning framework includes: Perform feature extraction on the input image sequence to obtain the feature vector representation of each frame of image; Model the feature vector representation of each frame of image to generate the global feature representation of the image sequence; Based on the preset Faster R-CNN model, target detection is performed on the global feature representation of the input image sequence to generate the target area positioning frame; Segment the target area positioning frame, identify the bone structure of the fracture site, and mark the specific bone name to obtain anatomical structure annotation information; The target region positioning frame and the anatomical structure annotation information are integrated to generate a corresponding target region positioning frame; wherein the positioning frame includes the coordinates and size of the target region positioning frame and the corresponding anatomical structure annotation information.

5. The artificial intelligence-based injury imaging diagnosis method according to claim 4, characterized in that: The method of performing target detection on the global feature representation of the input image sequence based on the preset Faster R-CNN model and generating a target region positioning frame includes: Obtain a preset Faster R-CNN model; wherein the preset Faster R-CNN model includes at least a region proposal network layer, an anchor box alignment layer, a pooling layer, and a fully connected layer; Generate potential damage candidate regions based on the global feature representation of the input image sequence through the region proposal network layer, and output damage candidate region suggestions; The anchor frame alignment layer aligns the damage candidate region proposal output by the region proposal network layer with the predefined anchor frame, determines the boundary of the candidate region, and outputs the damage candidate region; Performing pooling processing on the candidate damage area through the pooling layer to generate a fixed-size regional feature map; The fully connected layer performs category determination and bounding box regression on the regional feature map to generate a target region positioning frame.

6. The artificial intelligence-based injury imaging diagnosis method according to claim 1, characterized in that: The fusion of the damage image features to obtain a composite feature representation includes: Performing feature extraction at different levels on the damaged image features to obtain local features and global features of the damaged image; wherein the local features of the damaged image represent the detailed structure of the damaged image, and the global features represent the context information of the entire damaged image; Performing feature complementation on the local features and the global features to obtain complementary information; Performing feature fusion on the complementary information, the local features, and the global features, adjusting the scales of multi-source features, and obtaining fused features; The fusion features are abstracted and optimized through a preset deep learning network to generate a composite feature representation.

7. An artificial intelligence-based injury imaging diagnostic system, characterized in that: The artificial intelligence-based injury imaging diagnostic system includes: An acquisition module, configured to acquire sequential image data of the patient's damaged part; wherein the sequential image data comprises at least two-dimensional medical images captured at two or more time periods; A contrast enhancement module, configured to optimize the quality and enhance the contrast of the sequence image data to obtain an image sequence; a detection module, configured to perform feature learning and target region detection on the image sequence, automatically identify and mark anatomical structures indicating damage, and generate a corresponding target region positioning framework; an alignment module, configured to perform spatial calibration and alignment on the target region positioning frame to obtain damage image features; A fusion module, configured to fuse the damage image features to obtain a composite feature representation; The diagnosis module is used to input the composite feature representation into the trained diagnosis model to perform damage classification and importance assessment to obtain damage classification results and damage assessment results; and generate a diagnosis result based on the damage classification results and damage assessment results.

8. An artificial intelligence-based injury imaging diagnostic device, characterized in that: The artificial intelligence-based injury imaging diagnostic device includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor calls the instructions in the memory to enable the artificial intelligence-based injury imaging diagnosis device to execute the artificial intelligence-based injury imaging diagnosis method as described in any one of claims 1-6.

9. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the artificial intelligence-based injury imaging diagnosis method according to any one of claims 1 to 6 is implemented.

Images