Multi-species X-ray picture AI identification method and model training system and method

Abstract

In the species X-ray picture AI recognition model training system and method, a species information collection module collects X-ray pictures of different species, and a part labeling module labels part information on the X-ray pictures of the different species; the species information collection device outputs an X-ray labeling picture A; the X-ray labeling picture A comprises species information and part information; the exception labeling device is used for carrying out exception labeling on the X-ray labeling picture A, wherein the exception labeling is that exception information is labeled in a region corresponding to the part information; the abnormity labeling device outputs an X-ray labeling picture B; the X-ray labeling picture B comprises abnormal information corresponding to the part information; the AI training calculation component is used for AI model training, AI model training is carried out by using the input X-ray labeling picture B, and the AI training calculation component outputs a trained AI model or an AI feature data set; the trained AI model or AI feature data set is combined with AI software and is used for analyzing anomalies in X-ray pictures of multiple species.

Description

Technical Field

[0001] This application belongs to the field of image analysis technology based on X-ray imaging, specifically involving a training system, method, and data storage device for AI recognition models of multi-species X-ray images. Background Technology

[0002] Conventional X-ray imaging devices, such as digital radiography (DR) imaging devices, obtain a two-dimensional planar image by transmitting X-rays through the human body in a single projection.

[0003] CT imaging devices use X-rays to perform continuous tomographic scanning around the human body from multiple angles, and then use computers to reconstruct cross-sectional or three-dimensional images.

[0004] Any image obtained based on the absorption differences that occur when X-rays penetrate human tissue; the interpretation and analysis of such images requires professionally trained doctors for accurate interpretation.

[0005] However, in the context of veterinary hospitals, many veterinarians do not possess the same high level of X-ray image analysis and recognition capabilities as veterinarians in human hospitals. They also cannot utilize specialized X-ray imaging equipment for different body parts, such as dental X-ray imaging devices, as is available in human hospitals.

[0006] Unlike human hospitals, which focus on imaging and analyzing humans and don't require cross-species classification and identification, veterinary hospitals involve imaging various types of pets. Furthermore, the significant differences in physical structure between species make accurate identification and assessment by veterinarians extremely challenging.

[0007] Pet X-ray image analysis is highly dependent on veterinarian experience, and its efficiency in identifying body parts, especially critical organs, is low, with a single analysis taking more than 30 minutes. Furthermore, it requires a high level of veterinary experience; identifying different body parts of different species requires a vast amount of information, making it even more difficult to quickly identify abnormal features.

[0008] Furthermore, veterinary hospitals are typically small and cannot be equipped with multiple types of X-ray imaging devices. They usually desire a single machine that can perform multiple functions, such as being both a conventional digital X-ray imaging device and a CT scanner capable of 3D imaging; ideally, it should also offer automatic identification and diagnostic prompts to improve equipment utilization efficiency, reduce costs, and decrease the workload of veterinarians. The ability to complete one-stop testing and reporting in a cost-effective and efficient manner is one of the core requirements in the veterinary hospital setting.

[0009] The technical solution in this application is applicable to both two-dimensional and three-dimensional imaging scenarios; it can be used for species identification and classification, and can also detect abnormal information in areas such as bones, lungs (e.g., pneumonia), and foreign bodies in the digestive tract. CT-like three-dimensional imaging images can also be used to analyze fine tissue structures, including anatomical details of blood vessels and tumors.

[0010] Artificial intelligence in image recognition technology includes AI training. The AI ​​training process requires labeling the training data. The labeling steps include necessary image segmentation and matting. Masking models are mainly used to accurately segment target regions in images (such as faces and objects) and to label complex target images.

[0011] Masking models are essentially image segmentation techniques in computer vision, and are often used in conjunction with semantic segmentation and instance segmentation.

[0012] ComfyUI enables fine-grained image cutout and background replacement through mask nodes, supporting both images and videos; StableDiffusion utilizes mask redrawing functionality to quickly replace clothing on AI models.

[0013] The masking model in the Rope tool can intelligently handle multi-face replacement and complex occlusion; multimodal models and security protection, large visual models (such as Qwen2.5-VL) integrate masking components to improve target detection accuracy; in the field of cybersecurity, "data masking" technology is used to prevent the leakage of sensitive information during AI training. Summary of the Invention

[0014] The technical problem this application aims to solve is to avoid the low efficiency and high personnel skill requirements in the existing technology when identifying parts of multiple species. It provides a training system and method for AI recognition model of multi-species X-ray images; it can output part information and abnormal information of multiple species, and the output prompt information is more accurate.

[0015] The technical solution proposed in this application to address the aforementioned problems is a multi-species X-ray image AI recognition model training system, comprising a species information collection device, an anomaly labeling device, and an AI training computation component. The species information collection device includes a species information collection module and a part labeling module. The species information collection module collects X-ray images of different species, and the part labeling module labels part information on the X-ray images of different species. The species information collection device outputs an X-ray labeled image A, which includes species information and part information. The anomaly labeling device is used to label anomalies in X-ray labeled image A, whereby the anomaly labeling is performed on the area corresponding to the part information. The anomaly labeling device outputs an X-ray labeled image B, which includes anomaly information corresponding to the part information. The AI ​​training computation component is used for AI model training, employing the input X-ray labeled image B for AI model training, and outputs a trained AI model or AI feature dataset. The trained AI model or AI feature dataset is combined with AI software to analyze anomalies in X-ray images of multiple species.

[0016] Alternatively, the location information may be limited to species information; the anomaly information may be limited to species location information.

[0017] The species may include any one or more of the following: humans, dogs, felines, and fish.

[0018] The location information may include bones, body parts, and organs; the labeled location information uses masking nodes to segment the boundaries of the locations in the image.

[0019] It can be any one or more of the following bones: cranium, frontal bone, parietal bone, occipital bone, sphenoid bone, ethmoid bone, and temporal bone.

[0020] It can be any one or more of the facial bones, including the maxilla, mandible, zygomatic bone, nasal bone, lacrimal bone, inferior nasal concha, palatine bone, vomer, and hyoid bone.

[0021] It can be any one or more of the following: bones, including the trunk bones: spine, cervical vertebrae, thoracic vertebrae, lumbar vertebrae, sacrum, coccyx, sternum, and ribs.

[0022] It can be that the skeleton includes the bones of the limbs: any one or more of the bones of the upper limb, the bones of the upper limb girdle, the clavicle, and the scapula; the bones of the upper limb include any one or more of the humerus, radius, ulna, carpal bones, scaphoid, lunate, triquetrum, pisiform, trapezium, trapezium, capitate, hamate, metacarpals, and phalanges; the bones of the lower limb include any one or more of the bones of the lower limb girdle, hip bone, ischium, pubis, femur, patella, tibia, and fibula.

[0023] It can be any one or more of the following bones: tarsal bones, talus, calcaneus, navicular bone, medial cuneiform, intermediate cuneiform, lateral cuneiform, cuboid, metatarsals, and phalanges.

[0024] It can be any one or more body parts, including the head, torso, and limbs.

[0025] It can be any one or more organs, including the heart, liver, spleen, lungs, and kidneys.

[0026] The abnormal information may include any one or more of the following: traumatic lesion information, joint injury, bone inflammation and infection, metabolic disease, degenerative changes, tumor-related abnormal information, spinal lesion information, foreign body location information, structural abnormalities and effusion information, cardiac lesion information, abnormal liver morphology and density, space-occupying lesion information, and abnormal bone and anatomical structures.

[0027] The technical solution of this application to solve the above problems can also be a multi-species X-ray image AI recognition training method, including step 10: collecting X-ray images; step 20: labeling the X-ray images with species information; step 30: labeling the X-ray images with location information to obtain labeled X-ray images A; step 40: using the labeled X-ray images to train an AI model, outputting a trained AI model A or an AI feature dataset A; the trained AI model A or AI feature dataset A is combined with AI software; used to analyze location information in X-ray images of multiple species.

[0028] The technical solution of this application to solve the above problems can also be that the part information includes bones, body parts, and organs; the labeled part information is segmented by masking nodes to divide the part boundaries in the image.

[0029] The technical solution of this application to solve the above problems can also be a multi-species X-ray image AI recognition method, including step K10: capturing an X-ray image of the object to be detected; step K20: inputting species information into the X-ray image to obtain image data C; step K30: analyzing the image data C with an AI model to obtain an AI detection report; the AI ​​detection report includes the location information of the object to be detected; the AI ​​model is the aforementioned AI model A; or the AI ​​model is combined with the aforementioned feature dataset A to analyze the location information in X-ray images of multiple species.

[0030] The technical solution of this application to solve the above problems can also be a multi-species X-ray image AI recognition training method, including step C10: collecting X-ray images; step C20: labeling the X-ray images with species information; step C30: labeling the X-ray images with location information to obtain labeled X-ray image A; step C31: labeling the labeled X-ray image A with location anomaly information to obtain labeled X-ray image B; step C41: using labeled X-ray image B to train an AI model, outputting trained AI model B or AI feature dataset B; the trained AI model B or AI feature dataset B is combined with AI software; used to analyze location anomaly information in X-ray images of multiple species.

[0031] The technical solution of this application to solve the above problems can also be that the part information includes bones, body parts, and organs; the labeled part information is segmented by masking nodes to divide the part boundaries in the image.

[0032] The technical solution of this application to solve the above problems can also be a multi-species X-ray image AI recognition method, including step K10: capturing an X-ray image of the object to be detected; step K20: inputting species information into the X-ray image to obtain X-ray image data C with species information; step K30: analyzing the image data C with an AI model to obtain an AI detection report; the AI ​​detection report includes the location information and location abnormality information of the detected object; the AI ​​model is the aforementioned AI model B; or the AI ​​model is combined with the aforementioned feature dataset B to analyze the location information and location abnormality information in X-ray images of multiple species.

[0033] The technical solution of this application to solve the above problems can also be a data storage device, used to store the AI ​​model or dataset involved in the above-mentioned multi-species X-ray image AI recognition model training system; used to store the AI ​​model or dataset involved in the above-mentioned multi-species X-ray image AI recognition model training method.

[0034] The beneficial technical effects of this application include that one X-ray detection and analysis device can support the detection of multiple species and multiple detection sites, significantly reducing physical equipment and saving equipment procurement costs.

[0035] The beneficial technical effects of this application include: an X-ray detection and analysis device that can support the analysis of various diseases, significantly reduce the workload of doctors or veterinarians, provide standardized expert analysis results, and significantly reduce the operating costs of pet hospitals or veterinary stations.

[0036] The beneficial technical effects of this application include the ability to output part information and abnormal information of multiple species, and the output prompt information is more accurate.

[0037] The beneficial technical effects of this application include that the part information is limited to the species information; the abnormal information is limited to the species part information, and the output information is more reliable.

[0038] The beneficial technical effects of this application include good species adaptability, applicability to multiple species, and increased compatibility of the equipment in various application scenarios.

[0039] The beneficial technical effects of this application include that the location information includes bones, body parts, and organs, enabling the identification of information from multiple perspectives and reporting of abnormal information, providing rich information at multiple levels.

[0040] The beneficial technical effects of this application include its applicability to the acquisition of abnormal information from various bones, body parts, and organs, resulting in a significant increase in the amount of information and greatly improving system efficiency.

[0041] The beneficial technical effects of this application include that the training and recognition method can output part information, ensuring the accuracy of part information recognition, significantly reducing training costs, and reducing the power consumption of AI computing during the detection process.

[0042] The beneficial technical effects of this application include that the training and recognition method can simultaneously output location information and abnormality information, ensuring the accuracy of the abnormality information.

[0043] The beneficial technical effects of this application include that the setting of the data storage device facilitates data storage and management. Attached Figure Description

[0044] Figure 1 A schematic diagram of a multi-species X-ray image AI recognition model training system. Figure 1 ; Figure 2 A schematic diagram of a multi-species X-ray image AI recognition model training system. Figure 2 ; Figure 3 This is a flowchart illustrating the training method for AI recognition of multi-species X-ray images. Figure 1 ; Figure 4 This is a flowchart illustrating the AI ​​recognition method for multi-species X-ray images. Figure 2 ; Figure 5 This is a flowchart illustrating the training method for AI recognition of multi-species X-ray images. Figure 3 ; Figure 6 This is a flowchart illustrating the AI ​​recognition method for multi-species X-ray images. Figure 4 ; Figure 7 This is the first list of error messages; Figure 8 This is list 2 of the error messages; Figure 9 This is list 3 of the error messages; Figure 10 This is a flowchart illustrating the AI ​​recognition method for multi-species X-ray images. Figure 5 ; Figure 11 This is a flowchart illustrating the AI ​​recognition method for multi-species X-ray images; Figure 12 The kidneys are labeled with masked nodes on X-ray images of selected species M1 in a supine position to form a label map for AI recognition training. Figure 13 In the selected species M2, the kidneys are labeled with masked nodes on lateral X-ray images to form a tag map for AI recognition training; Figure 14 In the selected species M3, the kidneys are labeled with masked nodes on lateral X-ray images to form a tag map for AI recognition training; Figure 15 The liver is labeled with masked nodes on a partial lateral X-ray image of the selected species M4 to form a label map for AI recognition training; Figure 16 The liver is labeled with masked nodes on a partial lateral X-ray image of the selected species M5 to form a label map for AI recognition training. Figure 17 The liver is labeled with masked nodes on a partial lateral X-ray image of the selected species M6 to form a label map for AI recognition training. Figure 18 The lungs are labeled with masked nodes on a partial supine X-ray image of the selected species M7 to form a label map for AI recognition training. Figure 19 The lungs of the selected species M8 are labeled with masked nodes on a partial lateral X-ray image to form a label map for AI recognition training. Figure 20 The process involves using masked nodes to identify bones on a partial lateral X-ray image of the selected species M9, creating a label map for AI recognition training. Figure 21 The process involves using masked nodes to identify the spine on a partial supine X-ray image of the selected species M10, creating a label map for AI recognition training. Figure 22 On a selected species M11, specific locations of bones are marked with masked nodes on a partial supine X-ray image to form a label map, which is used for AI recognition training of the segments of the spine. Figure 23On a selected species M12, specific locations of bones are marked with masked nodes on a partial supine X-ray image to form a label map, which is used for AI recognition training of the segments of the spine. Figure 24 On a selected species M13, in a partial supine X-ray image, the location of bone fractures is marked with mask nodes to form a label map, which is used for AI to identify and determine fractures. Figure 25 On a selected species M14, in a partial supine X-ray image, the location of bone fractures is marked with mask nodes to form a label map, which is used for AI to identify and determine fractures; Figure 26 It is a diagram showing different skeletons of an animal's X-ray image annotated with "3DSlicer"; Figure 27 This is a diagram illustrating how to annotate an X-ray image of an animal using "3DSlicer"; Figure 28 It is an X-ray image of a pet's lungs; Figure 29 It is an X-ray image of a pet's lungs while it is lying face down. Detailed Implementation

[0045] The embodiments of this application will be further described in detail below with reference to the accompanying drawings.

[0046] like Figure 1 and Figure 2 A multi-species X-ray image AI recognition model training system includes a species information collection device, an anomaly labeling device, and an AI training computing component. The species information collection device includes a species information collection module and a body part labeling module. The species information collection module collects X-ray images of different species, and the body part labeling module labels body part information on the X-ray images of different species. The species information collection device outputs an anomaly-labeled X-ray image A, which includes species information and body part information. The anomaly labeling device is used to label anomalies in the anomaly-labeled X-ray image A, whereby the anomaly labeling is performed on the region corresponding to the body part information. The anomaly labeling device outputs an anomaly-labeled X-ray image B, which includes anomaly information corresponding to the body part information. The AI ​​training computing component is used for AI model training, using the input anomaly-labeled X-ray image B for AI model training, and outputs a trained AI model or AI feature dataset. The trained AI model or AI feature dataset is combined with AI software to analyze anomalies in X-ray images of multiple species.

[0047] like Figures 7 to 8 The location information is limited by the species information; the anomaly information is limited by the species location information.

[0048] The species include any one or more of humans, dogs, felines, and fish.

[0049] The location information includes bones, body parts, and organs.

[0050] The skeleton includes the skull: cranium, frontal bone, parietal bone, occipital bone, sphenoid bone, ethmoid bone, temporal bone; the facial bones: maxilla, mandible, zygomatic bone, nasal bone, lacrimal bone, inferior nasal concha, palatine bone, vomer, hyoid bone; the trunk bones: spine, cervical vertebrae, thoracic vertebrae, lumbar vertebrae, sacrum, coccyx, sternum, ribs; the limb bones: upper limb bones, upper limb girdle bones, clavicle, scapula; upper limb bones include humerus, radius, ulna, carpal bones, scapula, lunate, triquetrum, pisiform, trapezium, trapezium, capitate, hamate, metacarpals, phalanges; the lower limb bones include lower limb girdle bones, hip bones, ischium, pubis, femur, patella, tibia, fibula; the tarsal bones: talus, calcaneus, navicular bone, medial cuneiform, intermediate cuneiform, lateral cuneiform, cuboid, metatarsals, phalanges.

[0051] Body parts include the head, torso, and limbs; organs include the heart, liver, spleen, lungs, and kidneys.

[0052] like Figure 3 A multi-species X-ray image AI recognition training method includes step A10: collecting X-ray images; step A20: labeling the X-ray images with species information; step A30: labeling the X-ray images with body part information to obtain labeled X-ray images A; step A40: using the labeled X-ray images to train an AI model, outputting a trained AI model A or an AI feature dataset A; the trained AI model A or AI feature dataset A is combined with AI software to analyze body part information in X-ray images of multiple species.

[0053] like Figure 4 A multi-species X-ray image AI recognition method includes step KA10: capturing an X-ray image of the object to be detected; step KA20: inputting species information into the X-ray image to obtain image data C; step KA30: analyzing the image data C using an AI model to obtain an AI detection report; the AI ​​detection report includes the location information of the detected object; the AI ​​model is the aforementioned AI model A; or the AI ​​model is combined with the aforementioned feature dataset A to analyze the location information in X-ray images of multiple species.

[0054] like Figure 5A multi-species X-ray image AI recognition training method includes the following steps: C10: collecting X-ray images; C20: labeling the X-ray images with species information; C30: labeling the X-ray images with location information to obtain labeled X-ray image A; C31: labeling the labeled X-ray image A with location anomaly information to obtain labeled X-ray image B; C41: using labeled X-ray image B to train an AI model, outputting trained AI model B or AI feature dataset B; the trained AI model B or AI feature dataset B is combined with AI software to analyze location anomaly information in X-ray images of multiple species.

[0055] like Figure 6 A multi-species X-ray image AI recognition method includes step KC10: capturing an X-ray image of the object to be detected; step KC20: inputting species information into the X-ray image to obtain X-ray image data C with species information; step KC30: analyzing the image data C using an AI model to obtain an AI detection report; the AI ​​detection report includes part information and part abnormality information of the detected object; the AI ​​model is the aforementioned AI model B; or the AI ​​model is combined with the aforementioned feature dataset B to analyze part information and part abnormality information in X-ray images of multiple species.

[0056] A data storage device is used to store the AI ​​model or dataset involved in the above-mentioned multi-species X-ray image AI recognition model training system; and to store the AI ​​model or dataset involved in the above-mentioned multi-species X-ray image AI recognition model training method.

[0057] The method of this application is applicable to two-dimensional images obtained by conventional X-ray imaging devices as well as three-dimensional images obtained by CT imaging devices.

[0058] Any image obtained based on the absorption differences caused by X-rays penetrating human tissue is applicable; it can be used for identifying body parts in multiple species, and can also detect abnormalities in bones, lungs (such as pneumonia), and foreign bodies in the digestive tract. CT images can also be used to analyze fine tissue structures, including anatomical details of blood vessels and tumors.

[0059] like Figures 7 to 8 Abnormal information includes traumatic lesions, such as fractures; fractures include complete fractures, bone fissures, and compression fractures (such as the spine); common sites: long bones of the limbs, sternum, pelvis, ribs, and shoulder / knee joints.

[0060] Abnormal information includes joint injuries, such as dislocations and joint space abnormalities; joint space abnormalities include synovitis and arthritis.

[0061] Abnormal information includes bone inflammation and infection, including osteomyelitis and bone destruction, such as worm-eaten or osteolytic changes on imaging.

[0062] Abnormal information includes metabolic diseases, osteoporosis, decreased bone density, trabecular bone thinning, and osteosclerosis.

[0063] Abnormal information includes degenerative changes, bone spur formation, and joint degeneration.

[0064] Abnormal information includes tumor-related abnormalities, such as primary bone tumors, osteosarcoma, osteochondroma, and bone hyperplasia; metastatic bone tumors, cancer bone metastases, and multiple osteolytic defects.

[0065] Abnormal information includes spinal lesions: scoliosis, vertebral compression fracture, skeletal developmental abnormalities, congenital malformations, and epiphyseal line abnormalities.

[0066] Abnormal information includes foreign body location: such as postoperative assessment of the position of internal metal fixation devices.

[0067] Abnormal information includes infectious and inflammatory lesions, including pneumonia, tuberculosis, and lung abscess.

[0068] Abnormal information includes neoplastic lesions; primary lung cancer and large tumors are easily identifiable.

[0069] Abnormal information includes chronic and degenerative diseases, including pulmonary fibrosis and emphysema.

[0070] Abnormal information includes structural abnormalities and effusions; pleural effusion, pneumothorax, rib fractures, cortical bone interruption, dislocation, or angular deformity.

[0071] Abnormal information includes cardiac abnormalities: cardiomegaly, pericardial effusion, mediastinal tumor, or enlarged lymph nodes.

[0072] Abnormal information includes abnormalities in liver morphology and density: changes in liver volume, hepatomegaly, liver atrophy; calcifications, intrahepatic calcifications or calcification spots.

[0073] Abnormal information includes space-occupying lesions, neoplastic lesions, intermediate to advanced liver cancer, metastatic tumors, cysts or abscesses, and ascites.

[0074] Abnormal information in fish includes skeletal and anatomical abnormalities, such as skeletal deformities or injuries, parasitic infections, and observations of organ morphology, such as the swim bladder and fish bones.

[0075] In this application, in the multi-species X-ray image AI recognition model training system, the species information collection device, the anomaly labeling device, and the AI ​​training calculation component can be completed by setting up corresponding computers separately, or they can be integrated into DR or CT imaging equipment.

[0076] Specifically, a data acquisition module can be used to acquire 2D pet X-ray images. The 2D pet X-ray images are preprocessed using a preprocessing module, such as grayscale enhancement, random jittering, and random cropping, and the images are uniformly scaled to 256×256 pixels. A segmentation model is set up and used to output segmentation mask nodes for the lungs, spinal vertebrae, liver, and kidneys, forming a label map. The AI ​​training and computing component includes an analysis module. This module extracts organ size data based on the segmentation mask nodes, compares it with the trained AI model or AI feature dataset and AI software, analyzes anomalies in X-ray images of multiple species, and outputs anomaly information prompts.

[0077] Gray-scale enhancement targets the low-contrast features in areas where bones and organs overlap, thereby improving the accuracy of site information extraction and recognition.

[0078] The segmentation model, such as the cascaded U-Net segmentation model, uses a stochastic gradient descent (SGD) optimizer configuration (learning rate 0.01, momentum 0.9), achieving a Dice coefficient of 58% on the clinical validation set, a 7% improvement over the original U-Net model. After training, the AI ​​model or AI feature dataset can be uploaded and managed by the user; the system does not store default medical data.

[0079] The analysis module extracts organ size data based on the segmentation mask nodes, extracts organ size parameters based on the segmentation results, compares them with the trained AI model or AI feature dataset, and outputs the anomaly probability.

[0080] The anomaly probability is determined by the statistical deviation threshold between organ size and cases in the trained AI model or AI feature dataset.

[0081] A training system and method for multi-species X-ray image AI recognition model based on SGD-optimized U-Net model can be used to assist in the diagnosis of morphological abnormalities in the lungs, spine, liver, and kidneys.

[0082] In this application, a cascaded U-Net segmentation based on SGD optimization is used. First, Hessian grayscale enhancement is used to strengthen the bone edges, and random jitter and cropping are combined to improve the robustness of the model. An SGD optimizer is configured (lr=0.01, momentum=0.9), and it is trained for 100 epochs on a dataset of 5000 labeled images with a batch size of 32. Based on the segmentation results, organ size parameters (such as the aspect ratio of the liver) are extracted, and a bias analysis (±2σ threshold) is performed with the trained AI model or AI feature dataset.

[0083] Through targeted preprocessing, the Dice coefficient for segmenting overlapping bone regions was increased to 58%.

[0084] It achieves 20% faster training convergence compared to the original U-Net model; it supports user-owned private medical record database management, ensuring medical data privacy.

[0085] like Figure 10 The diagram illustrates the process of obtaining a dataset: data preprocessing, segmentation, and database comparison.

[0086] like Figure 11 The flowchart for comparing cases with the database shows that first, size parameters are obtained, and then anomalies are marked. The X-ray data to be analyzed is the digital X-ray image to be analyzed. The AI ​​segmentation module is used to segment the image to obtain a segmented image. Based on the liver size data, the segmented image is searched and identified to confirm the liver size. If the match is normal, the segmentation value result is output; if there is a matching problem, the case database is selected to confirm the specific pathological condition.

[0087] Data preprocessing uses: Hessian enhancement formula:

[0088] I e =I+λ·det(H(I))I e =I+λ·det(H(I)), where λ=0.3 and H is the Hessian matrix; random dithering: brightness adjustment range ±10%, contrast adjustment range ±15%; image uniformly scaled to 256×256 pixels.

[0089] Model training: Dataset: 5000 labeled pet X-ray images; the outlines of the lungs, spine, liver, and kidneys were labeled by veterinarians; Training parameters: SGD optimizer (lr=0.01, momentum=0.9), cross-entropy loss function, 100 training epochs.

[0090] Anomaly Diagnosis: Size Parameter Extraction: Calculate the aspect ratio of the minimum bounding rectangle for each organ; Database Comparison: If the parameters of the current case exceed ±2 times the standard deviation of the historical values ​​in the user's database, an anomaly warning is triggered.

[0091] The path of image AI interpretation is: image acquisition, image preprocessing: feature extraction, model training, recognition and decision-making, feedback and optimization.

[0092] Image Acquisition: Clinical imaging data is acquired using professional imaging equipment such as DR and CT scanners in animal clinics. Images can be in standard medical image DICOM format, acquired via PACS, or in common formats such as JPG / PNG converted from DICOM, or images captured using mobile phones / cameras.

[0093] Image preprocessing: The images acquired in step 1 are processed by denoising, enhancement, scaling, and grayscale conversion to improve image quality and clarity, facilitating subsequent feature extraction.

[0094] Feature extraction: Using specific algorithms developed for animal image recognition, useful feature information is extracted from images, such as low-level features like density, texture, shape, and edges, or high-level features automatically learned by deep learning models.

[0095] Model Training: Building upon feature extraction, machine learning or deep learning algorithms are used to train the model on a large amount of labeled data. The designed model techniques include Support Vector Machines (SVM), Convolutional Neural Networks (CNN), and Recurrent Neural Networks (RNN). By continuously adjusting the model parameters, the model is made capable of accurately identifying targets in images to meet the needs of practical applications.

[0096] Recognition and Decision Making: The trained model is applied to animal images captured by a new veterinary hospital to perform recognition and classification, outputting the recognition results and generating corresponding reports. Based on the recognition results and reports, the system can provide corresponding clinical support decisions.

[0097] Feedback and Optimization: After clinical application in animal diagnosis and treatment, the model is adjusted and optimized based on user feedback and actual results to improve the accuracy and efficiency of identification.

[0098] The above paths together constitute the technical path of AI image interpretation, enabling computers to accurately understand and interpret information in images like professional animal imaging experts, provide accurate interpretation results and reports, and ultimately offer corresponding clinical treatment recommendations.

[0099] like Figures 12 to 25 This method provides X-ray images of different animals and uses masking nodes or circles to identify different organs, bones, the starting positions of bones, and the locations of fractures in X-ray images from different examination positions, forming a label map. After labeling a massive amount of images, the AI ​​is trained to recognize different organs or bones in X-ray images of different species and from different examination positions.

[0100] Figure 12 The labeling method involves using masked nodes to identify the kidneys on an X-ray image of the selected species M1 in a supine position, creating a label image for AI recognition training; the label text is kindey0001.

[0101] Figure 13 The labeling method involves using masked nodes to identify the kidneys on a lateral X-ray image of the selected species M2, creating a label image for AI recognition training; the label text is kindey0002.

[0102] Figure 14 The labeling method involves using masked nodes to identify the kidneys on a lateral X-ray image of the selected species M3, creating a label image for AI recognition training; the label text is kindey0003.

[0103] Figure 15 The liver is labeled using masked nodes on a partial lateral X-ray image of the selected species M4, forming a label image for AI recognition training; the label text is liver0001.

[0104] Figure 16 The liver was labeled using masked nodes on a partial lateral X-ray image of the selected species M5, forming a label image for AI recognition training; the label text is liver0002.

[0105] Figure 17 The liver was labeled using masked nodes on a partial lateral X-ray image of the selected species M6, forming a label image for AI recognition training; the label text is liver0003.

[0106] Figure 18 The lungs were identified by masking nodes on a partial supine X-ray image of the selected species M7, forming a label image for AI recognition training; the label text is lung0001.

[0107] Figure 19 The lungs of the selected species M8 were labeled with masked nodes on a partial lateral X-ray image to form a label image for AI recognition training; the label text is lung0002.

[0108] Figure 20 The process involves using masked nodes to identify bones on a partial lateral X-ray image of the selected species M9, creating a labeled image for AI recognition training; the labeled text is spine0001.

[0109] Figure 21 The labeling method involves using masked nodes to identify the spine on a partial supine X-ray image of the selected species M10, creating a label image for AI recognition training; the label text is spine0003.

[0110] Figure 22 The labeling method involves using masked nodes to mark specific locations of bones on a partial supine X-ray image of the selected species M11, creating a labeled image for AI recognition training of the vertebral segments; the label text is QSW01.

[0111] Figure 23 The labeling method involves using masked nodes to mark specific locations of bones on a partial supine X-ray image of the selected species M12, creating a labeled image for AI recognition training of spinal segments; the label text is QSW02.

[0112] Figure 24 The method involves using masked nodes to mark the fracture locations of bones on a partial supine X-ray image of the selected species M13, creating a labeled image for AI to identify and determine fractures; the labeled text is GZ01.

[0113] Figure 25 The method involves using masked nodes to mark the fracture locations of bones on a partial supine X-ray image of the selected species M14, creating a labeled image for AI to identify and determine fractures; the labeled text is GZ003.

[0114] By recognizing organs or bones in X-ray images from different species and body positions, and further training with pathological images, such as fracture images, AI can acquire the ability to interpret X-ray images across species, organs, and disciplines. For example, traditional doctors—orthopedic surgeons only interpret orthopedic images, and internists only interpret images of the liver, kidneys, or lungs—can, through this specialized training method, interpret X-ray images from various departments, including orthopedics and internal medicine, as well as X-ray images from various species such as dogs, cats, and humans.

[0115] Location information is limited by species information; anomaly information is limited by species location information. This annotation method can classify and constrain various types of information, reducing training and computational load. Furthermore, during application detection, it avoids misidentifying diseases of different organs as other organs or misidentifying specific diseases of different species. For example, it prevents identifying a male testicular disease as a female patient.

[0116] This application involves segmenting, retouching, and annotating a massive amount of X-ray images. Software options for this part include "3DSlicer".

[0117] In the 3DSlicer software, the masking model is deeply integrated in the form of medical image segmentation technology. It mainly achieves pixel-level region identification through label maps and segment editors. Its core logic is consistent with the traditional masking model, but the terminology is different.

[0118] like Figure 26 The example demonstrates how "3DSlicer" annotates different bones in an animal's X-ray image. It shows that different colors are needed to annotate different bones, making the annotation work extremely labor-intensive. "3DSlicer" is one of many software options available; the software is merely the platform. The real challenge lies in scientifically planning the hierarchy and standards of the annotations, which requires a massive investment of human resources.

[0119] like Figure 27The image shows an X-ray image of an animal annotated with "3DSlicer". Different shapes are marked with different colors, and different organs are marked with different colors. The colors correspond to the label names.

[0120] Figure 27 The text provides the colors corresponding to different parts of the skull, teeth, heart, and brain.

[0121] like Figure 28 The image shown is a raw, unlabeled X-ray of a pet's lungs.

[0122] like Figure 29 The image shows an X-ray of a pet's lungs in a prone position. The annotation process requires the annotator to painstakingly draw, using a mouse, the exact location and shape of the bones in the X-ray image, striving for maximum accuracy. Different colors are also used to mark the location and shape of organs such as the heart and lungs. The workload for a single image is enormous; a single project might require annotating hundreds or even thousands of images.

[0123] There are many AI training and recognition software programs for graphics, but none of them are suitable for medical diagnosis.

[0124] The software that can be used includes "DeepFaceLab (DFL)," which is a deepfake face-swapping software.

[0125] The DeepFaceLab (DFL) software package includes the Xseg_TRAIN toolkit, which is specifically designed for training and managing XSeg masking models. Specifically, it allows AI training on annotated masking model images output by 3DSlicer. The trained model can then be used for interpreting X-ray images. Xseg_TRAIN is used to train XSeg models.

[0126] "Xseg_TRAIN" only uses facial feature recognition functions. In the implementation of this application, the software is extended. The software functionality can be extended to train the recognition of a pet's entire body organs and skeleton.

[0127] The core task of the original XSeg model is to accurately identify and segment video frames. The area of ​​the face, especially the area that can It can handle various obstructions (such as hair, glasses, hands, food, hats, bangs, etc.). , So that only the target is replaced during the face-swapping compositing stage Mark the face area to prevent occlusions from being mistakenly replaced or compromising the integrity of the face. .

[0128] The original XSeg model workflow includes the following steps: Data preparation: The user needs to prepare the target video first. The keyframes containing complex occlusions in (data_dst) and the source face video / image (data_src) were manually used to draw precise masks. (That is, marking which obstructions should be retained and which faces should be replaced). This step involves a significant amount of work.

[0129] To start training: After preparing sufficient labeled data, the user runs a program named 5.XSeg)train.bat (or a similar program). The batch script (named XSeg) is the core execution entry point for Xseg_TRAIN. This script initiates the training process of the XSeg model. Procedure.

[0130] Monitoring and Adjustment: During training, users can observe the model's learning performance in the preview window.

[0131] Output Model: Training terminates when the preview results are satisfactory. The trained XSeg model file will be saved. In the model folder.

[0132] Application Model: The trained XSeg model can be applied in the subsequent synthesis stage.

[0133] The workflow of this application is similar to that of the original XSeg model, but it is more complex and requires... It requires more computing resources and time.

[0134] In summary, the software development platforms used in this application, such as "3DSlicer" and "Xseg_TRAIN," were developed by experts in the field of AI. Although this application utilizes these software development platforms, the core aspect lies in the combination of "medical experts" and "AI application experts" to develop a scientific "annotation system," and then iteratively performing the following steps on tens of thousands of X-ray images.

[0135] Step 1: Establishment of the "First Generation Annotation System";

[0136] Step 2: Annotating massive amounts of images;

[0137] Step 3: "AI Training"

[0138] Step 4: "AI Recognition and Verification"

[0139] Step 5: "Annotation System Optimization", proceed to Step 2.

[0140] In this application, the applicant is concerned with the establishment of the "first-generation annotation system" in step 1 above, and step 5: "annotation system optimization".

[0141] The use of the first, second, letters, and combinations of letters and numbers in this application is for convenience of expression only and does not necessarily indicate a relationship of size or temporal order.

[0142] All letters or numbers are merely for convenience of expression, used to refer to the substantive matters related to the preceding and following text, and are not limited by the literal meaning of the words themselves.

[0143] As shown in the accompanying drawings, the above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of the invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A training system for an AI recognition model of multi-species X-ray images, characterized in that: Includes species information collection devices, anomaly labeling devices, and AI training computing components; The species information collection device includes a species information collection module and a part labeling module; The species information collection module collects X-ray images of different species, and the part annotation module annotates the part information of the X-ray images of different species; the species information collection device outputs an annotated X-ray image A. X-ray labeled image A includes species and location information; An anomaly annotation device is used to annotate an X-ray annotated image A, wherein the anomaly annotation is to mark abnormal information in the area corresponding to the location information; the anomaly annotation device outputs X-ray annotated image B; X-ray annotated image B includes abnormal information corresponding to the location information; The AI ​​training computing component is used for AI model training. It uses the input X-ray labeled image B to train the AI ​​model and outputs the trained AI model or AI feature dataset. The trained AI model or AI feature dataset is combined with AI software to analyze anomalies in X-ray images of multiple species.

2. The multi-species X-ray image AI recognition model training system according to claim 1, characterized in that: The location information is limited by the species information; the anomaly information is limited by the species location information.

3. The multi-species X-ray image AI recognition model training system according to claim 1, characterized in that: The species include any one or more of humans, dogs, felines, and fish.

4. The multi-species X-ray image AI recognition model training system according to claim 1, characterized in that: The location information includes bones, body parts, and organs; The labeled area information is obtained by segmenting the area boundaries in the image using masking nodes.

5. The multi-species X-ray image AI recognition model training system according to claim 4, characterized in that: Includes one or more of the following technical features: TA11: Bones include the skull: any one or more of the cranial bones, frontal bone, parietal bone, occipital bone, sphenoid bone, ethmoid bone, and temporal bone; TA12: The skeleton includes facial bones: any one or more of the maxilla, mandible, zygomatic bone, nasal bone, lacrimal bone, inferior nasal concha, palatine bone, vomer, and hyoid bone; TA13: Bones include the trunk bones: any one or more of the spine, cervical vertebrae, thoracic vertebrae, lumbar vertebrae, sacrum, coccyx, sternum, and ribs; TA14: The skeleton includes the bones of the limbs: any one or more of the bones of the upper limb, the bones of the upper limb girdle, the clavicle, and the scapula; The bones of the upper limb include any one or more of the following: humerus, radius, ulna, carpal bones, scaphoid, lunate, triquetrum, pisiform, trapezium, trapezium, capitate, hamate, metacarpals, and phalanges. The bones of the lower limb include any one or more of the bones of the lower limb girdle, hip bone, ischium, pubis, femur, patella, tibia, and fibula; TA15: Bones include any one or more of the following: tarsal bones: talus, calcaneus, navicular bone, medial cuneiform, intermediate cuneiform, lateral cuneiform, cuboid, metatarsals, and phalanges; TA20: Body parts include any one or more of the head, trunk, and limbs; TA30: Organs include any one or more of the heart, liver, spleen, lungs, and kidneys; TA50: The abnormal information includes any one or more of the following: traumatic lesion information, joint injury, bone inflammation and infection, metabolic disease, degenerative changes, tumor-related abnormal information, spinal lesion information, foreign body location information, structural abnormalities and effusion information, cardiac lesion information, abnormal liver morphology and density, space-occupying lesion information, and abnormal bone and anatomical structures.

6. A training method for AI recognition of multi-species X-ray images, characterized in that, include, Step 10: Collect X-ray images; Step 20: Label the species information on the X-ray images; Step 30: Annotate the X-ray image with the annotated area information to obtain an annotated X-ray image A; Step 40: Use the annotated X-ray images to train the AI ​​model and output the trained AI model A or AI feature dataset A; The trained AI model A or AI feature dataset A is combined with AI software to analyze part information in X-ray images of multiple species.

7. The multi-species X-ray image AI recognition training method according to claim 6, characterized in that: The location information includes bones, body parts, and organs; The labeled area information is obtained by segmenting the area boundaries in the image using masking nodes.

8. A multi-species X-ray image AI recognition method, characterized in that, include, Step K10: Obtain an X-ray image of the object to be inspected; Step K20: Input species information into the X-ray image to obtain image data C; Step K30: Analyze the image data C using an AI model to obtain an AI detection report; the AI ​​detection report includes information about the location of the detected object; The AI ​​model is AI model A as described in claim 6 or 7; Alternatively, the AI ​​model can be combined with the feature dataset A described in claim 6 or 7 to analyze part information in X-ray images of multiple species.

9. A multi-species X-ray image AI recognition training method, characterized in that, include, Step C10: Collect X-ray images; Step C20: Label the species information on the X-ray image; Step C30: Mark the regions on the X-ray image to obtain the marked X-ray image A; Step C31: Obtain post-annotated X-ray image B by analyzing the abnormal information of the annotated area in the annotated X-ray image A; Step C41: Use the annotated X-ray image B to train the AI ​​model, and output the trained AI model B or the AI ​​feature dataset B; The trained AI model B or AI feature dataset B is combined with AI software to analyze anomaly information in X-ray images of multiple species.

10. The multi-species X-ray image AI recognition training method according to claim 9, characterized in that: The location information includes bones, body parts, and organs; The labeled area information is obtained by segmenting the area boundaries in the image using masking nodes.

11. A multi-species X-ray image AI recognition method, characterized in that, include, Step K10: Obtain an X-ray image of the object to be inspected; Step K20: Input species information into the X-ray image to obtain X-ray image data C with species information; Step K30: Analyze the image data C using an AI model to obtain an AI detection report; the AI ​​detection report includes the location information and location anomaly information of the detected object; The AI ​​model is AI model B as described in claim 9 or 10; Alternatively, the AI ​​model can be combined with the feature dataset B described in claim 9 or 10 to analyze part information and part abnormality information in X-ray images of multiple species.

12. A data storage device, characterized in that: Used to store the AI ​​model or dataset involved in the multi-species X-ray image AI recognition model training system according to any one of claims 1 to 5; Used to store the AI ​​model or dataset involved in the method of any one of claims 6 to 10.

Images