Hip image femoral neck fracture detection and garden classification model, construction method and system

Abstract

The application discloses a hip image femoral neck fracture detection and Garden classification model, a construction method and a system, and relates to the technical field of medical image analysis. The model fuses an EfficientNetV2 feature extraction network, a feature matching attention module and a YOLOv10 architecture, and can realize hip image target region positioning and complete Garden four-type classification. The construction method collects and screens multi-center image data, carries out labeling auditing and preprocessing enhancement, and ensures performance through 5-fold cross-validation training and external testing. The analysis system comprises data acquisition, preprocessing, model training, image analysis and other modules, supports mainstream image formats, and can be embedded in existing image processing systems. The application solves the problems of traditional classification relying on experience and high error, improves the local feature recognition capability and generalization, provides efficient technical support for hip image analysis, and has strong practicality.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis technology, and in particular to the detection of femoral neck fractures in hip imaging and the Garden classification model, construction method and system. Background Technology

[0002] Hip-related skeletal image analysis is an important research direction in the field of medical image processing. Among them, the image feature identification and classification of the femoral neck region is of great significance for related medical research and clinical support. As a key part connecting the femoral head and femoral shaft, the femoral neck has a complex morphology and structure and is easily affected by various factors in images, making feature extraction and classification of this region a technical challenge.

[0003] Currently, the processing of femoral neck-related images mainly relies on manual analysis or traditional computer-aided techniques. In manual analysis, professionals must manually assess the images based on the Garden classification system, which categorizes femoral neck-related image features into four types, each corresponding to specific morphological manifestations. However, manual analysis has significant limitations: firstly, variations in image quality and blurred feature boundaries can easily lead to judgment errors, especially with low accuracy in distinguishing intermediate types; secondly, manual processing is inefficient, struggling to handle the analysis needs of large-scale image data, and the results are easily influenced by subjective factors such as operator experience and fatigue, making consistency difficult to guarantee.

[0004] While traditional computer-aided techniques have improved processing efficiency to some extent, they still have many shortcomings. Some techniques can only perform binary classification based on the presence or absence of specific features, failing to fully cover the four categories of the Garden classification. Other techniques, although attempting multi-type classification, lack accuracy in feature extraction and are insufficient in recognizing subtle morphological differences, resulting in classification accuracy that is difficult to meet the needs of practical applications. Furthermore, existing technologies have weak generalization capabilities. When faced with differences in imaging equipment, imaging parameters, and uneven data distribution across different medical centers, performance tends to decline significantly, making it difficult to adapt to diverse application scenarios.

[0005] At the image data processing level, traditional technologies are highly dependent on data quality, and their processing effects are poor for low-quality images, images with artifacts, or images captured from unusual angles. Furthermore, existing technologies lack effective feature enhancement mechanisms, making it difficult to focus on key local features in images, thus limiting feature recognition capabilities in complex backgrounds. With the rapid growth of medical image data and the increasing demands for accuracy and efficiency in image analysis, existing technologies can no longer meet the diverse needs of practical applications. There is an urgent need for a technical solution that can automatically, accurately, and efficiently perform femoral neck-related image feature detection and Garden classification. Summary of the Invention

[0006] The purpose of this invention is to provide a model, construction method and system for detecting and classifying femoral neck fractures in hip imaging, thereby achieving automated and standardized analysis of femoral neck-related features in hip imaging, significantly improving image processing efficiency, providing reliable technical support for related medical image analysis work, and meeting the needs of large-scale image data processing and diverse application scenarios.

[0007] To achieve the above objectives, this invention provides a hip imaging femoral neck fracture detection and Garden classification model, comprising a feature extraction network, a feature matching attention module, and a target detection and classification architecture. The feature extraction network uses EfficientNetV2 to extract global features from hip images. The feature matching attention module enhances the attention to local features of the target region in the hip image, performing local feature enhancement processing on the global features. The target detection and classification architecture uses YOLOv10, receives the fused features processed by the feature matching attention module, and outputs the target region localization result and the Garden classification result. The Garden classification includes four categories: Type I, Type II, Type III, and Type IV.

[0008] A method for detecting femoral neck fractures in hip imaging and constructing a Garden classification model, comprising the following steps: S1. Data collection and screening: Collect hip CT and X-ray images of patients with femoral neck fractures from multiple centers, exclude images that do not meet the quality requirements, exceed the preset duration, contain chronic hip diseases, or contain implanted hardware, and obtain a valid raw image dataset. S2. Data Labeling and Review: Draw a bounding box containing the femoral head, greater trochanter, and lesser trochanter in the fracture area of ​​the valid original image data, and label the corresponding Garden classification: For eversion with incomplete fracture, it is judged as non-displaced type I; for complete fracture but no displacement of the fracture ends, it is judged as non-displaced type II; for complete fracture with partial displacement, that is, partial change in the course of the trabeculae, it is judged as displaced type III; for complete fracture with complete displacement and complete change in the parallel course of the trabeculae, it is judged as displaced type IV. S3. Data Preprocessing and Augmentation: The reviewed labeled data is processed by contrast enhancement, Gaussian noise addition, elastic deformation, random cropping, and rotation within a preset angle range to expand the data volume to a preset multiple of the original data volume. The expanded data is consistent with the original training set in terms of gender, age, equipment manufacturer, and collection location distribution. S4. Model Construction: Construct a deep learning model that integrates feature matching and attention modules. Use EfficientNetV2 as the feature extraction network to extract global features of hip images. Enhance the attention of local features in the fracture area through the feature matching and attention module. After fusing the enhanced local features with global features, input them into the YOLOv10 target detection and classification architecture to form an end-to-end fracture detection and classification model. S5. Model Training and Validation: The preprocessed dataset is divided into a training set and an internal test set according to a preset ratio. A 5x cross-validation method is used, and the SGD optimization algorithm is used to train the model for a preset number of rounds. The model performance is evaluated by accuracy, sensitivity, specificity, F1 score and area under the curve. S6. Model generalization test: The trained model is tested using external validation sets from different medical centers to verify its performance under different data distributions and imaging equipment conditions.

[0009] Preferably, the multi-center image data in S1 comes from at least four different data acquisition institutions, covering 800 or more relevant image providers within a preset time period, with a total image data volume of no less than 10,000 images.

[0010] Preferably, in S3, the contrast enhancement adopts CLAHE technology, and the elastic deformation, Gaussian noise addition and rotation processing are all configured with preset parameters; the preset multiplier is 5 times, and the preset angle range is ±15°.

[0011] Preferably, the model described in S4 is built using Python 3.10 and the PyTorch 2.3.1+cu118 open-source library. The model has a sensitivity of not less than 70% and a specificity of not less than 97% for Garden I, and a sensitivity of not less than 94% and an area under the curve of not less than 96% for Garden IV.

[0012] A hip imaging femoral neck fracture detection and Garden classification system, constructed based on a hip imaging femoral neck fracture detection and Garden classification model construction method, includes: Data acquisition module: used to download hip CT and X-ray image data from the image storage system, supports DICOM 3.0 format, and automatically filters valid images that meet the requirements; Data preprocessing module: Used to perform data preprocessing and augmentation operations, outputting standardized training and test data; Model training module: Used to load preprocessed data, execute model building, training and internal validation processes, and generate trained deep learning model files; Image analysis module: Used to receive input hip images, call the trained model, and automatically output the target region bounding box coordinates, Garden classification results, and confidence scores of each classification. Report generation module: This module converts the output of the image analysis module into a standardized analysis report. The report includes fracture localization diagrams, classification results, analysis basis, and model confidence index. System integration module: used to embed the image analysis module and result output module into the existing image processing system, supporting users to manually correct the analysis results and record correction logs.

[0013] Preferably, the image analysis module takes no more than 0.5 seconds to perform a single image analysis, has a consistency rate of no less than 91% with the consensus analysis of professionals, and a classification accuracy of no less than 85% in image data containing a preset proportion of artifacts.

[0014] Preferably, the system further includes a feedback module for collecting user evaluations of the model analysis results, including satisfaction with processing speed, acceptance of analysis accuracy, reliability score of results, and assessment of application value, providing data support for subsequent model optimization.

[0015] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: (1) Breakthrough in classification coverage and accuracy: Breaking through the limitations of existing technologies that rely on multi-focus binary classification detection or partial classification, this model achieves fine-grained classification of all four Garden types for the first time, filling a technological gap. By enhancing local features through the feature matching attention module, the model can accurately capture subtle morphological differences in images, effectively solving the problem of insufficient recognition of small features and low-resolution images by traditional technologies, significantly reducing classification errors. Each classification maintains high sensitivity and specificity, and the overall performance is superior to existing similar models.

[0016] (2) Significantly enhanced generalization ability and environmental adaptability: Based on training and validation using multi-center, multi-source image data, the model can adapt to differences in image equipment, heterogeneity of shooting parameters, and data distribution characteristics of different data acquisition institutions, meeting the needs of diverse application scenarios. Through various preprocessing methods, the negative impacts of image quality differences, shooting angle deviations, and artifacts are effectively reduced. The processing performance of low-quality and special body position images is superior to traditional technologies, and the anti-interference ability is outstanding.

[0017] (3) Significantly improved processing efficiency and practical applicability: The model achieves end-to-end full-process automation, eliminating the need for manual intervention in feature selection or step-by-step processing. The time required for single-case image analysis is significantly reduced, far lower than the efficiency of manual analysis, and can meet the requirements for rapid processing and high real-time performance of large-scale image data. The model supports mainstream image formats and can be seamlessly embedded into existing clinical image storage and processing systems without large-scale modifications to the existing architecture, reducing the cost of technology promotion. At the same time, it provides standardized and repeatable analysis results, reducing consistency issues caused by differences in the subjective experience of operators. The technology is convenient to implement and has outstanding practical value.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the data preprocessing process for the hip imaging femoral neck fracture detection and Garden classification model construction method in Embodiment 1 of the present invention; Figure 2 This is a flowchart of the hip imaging femoral neck fracture detection and Garden classification model construction method of Embodiment 1 of the present invention; Figure 3 This is a graph showing the ROC-AUC curves of the four fractals of the external validation set of the model in Embodiment 2 of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] Example 1: Model Training and Internal Validation 1. Data Preparation (1) Data source: Data from 857 hip imaging providers from Chongqing Ninth People's Hospital, Zizhong County People's Hospital of Neijiang City, Sichuan Province, Bishan District People's Hospital of Chongqing, and Jiangjin District People's Hospital of Chongqing were collected from January 2018 to December 2024. After screening, 10,010 hip CT and X-ray images from 806 imaging providers (356 males and 450 females, with an average age of 71 years) were finally included. Among them, the basic data used for model training and internal validation came from 529 imaging providers from Chongqing Ninth People's Hospital and Bishan District People's Hospital of Chongqing, totaling 7,818 images.

[0024] (2) Data screening criteria: The following image data will be excluded: images with poor quality (inappropriate detail, contrast or film darkness); images with a disease course of more than 4 weeks; images containing chronic hip joint disease; images of subjects with hardware such as screws, plates, wires or pins in their bodies. After screening, 110 images of poor quality and 51 images of subjects that could not be properly labeled were excluded from all images. The final valid labeled data were as follows: Garden I type 1208 images (113 X-ray images, accounting for 9.3%; 1095 CT images, accounting for 90.7%), Garden II type 2312 images (176 X-ray images, accounting for 7.6%; 2136 CT images, accounting for 92.4%), Garden III type 2462 images (769 X-ray images, accounting for 31.2%; 1693 CT images, accounting for 68.8%), and Garden IV type 4028 images (769 X-ray images, accounting for 19.1%; 3259 CT images, accounting for 80.9%).

[0025] (3) Data annotation and review: Two orthopedic surgeons with intermediate or higher professional titles and more than 5 years of work experience from Chongqing Ninth People's Hospital drew bounding boxes containing the femoral head, greater trochanter and lesser trochanter in the target area of ​​the valid original image data, and annotated the corresponding Garden classification (Type I: valgus with incomplete fracture; Type II: complete fracture without displacement; Type III: complete fracture with partial displacement; Type IV: complete fracture with complete displacement). The annotation results were reviewed and confirmed by two other orthopedic surgeons with more than 5 years of experience who were certified by the committee.

[0026] (4) Data preprocessing and augmentation: such as Figure 1 As shown, the CLAHE technique was used to enhance the contrast of the reviewed labeled data (parameters set to clipLimit=2.0, tileGridSize=8×8). At the same time, Gaussian noise was added (σ=0.01), elastic deformation (α=10, σ=5), random cropping and ±15° rotation were performed to expand the data volume to 5 times the original data volume. After expansion, the data remained consistent with the original training set in terms of gender, age, equipment manufacturer and collection location distribution.

[0027] 2. Model Building and Training Model architecture: such as Figure 2 As shown, a deep learning model with a fusion feature matching attention module (FMAM) is constructed. EfficientNetV2 is used as the feature extraction network to extract global features of hip images. The attention of local features in the target region is enhanced by the FMAM module. The enhanced local features are then fused with the global features and input into the YOLOv10 target detection and classification architecture to form an end-to-end detection and classification model.

[0028] Training parameters: The model was built using Python 3.10 and PyTorch 2.3.1 + cu118 open source library, using SGD optimization algorithm, batch size set to 8, training epochs of 200, and 5x cross-validation to ensure evaluation fairness.

[0029] Dataset partitioning: The imaging data of 529 subjects from Chongqing Ninth People's Hospital and Chongqing Bishan District People's Hospital were randomly allocated in an 8:2 ratio to obtain the training and validation set (423 cases, 6255 images) and the internal test dataset (106 cases, 1563 images).

[0030] 3. Verification Results After 5-fold cross-validation, the model achieved an average accuracy of 93.34% (95% CI, 92.75-93.97) for four types of femoral neck fractures of the hip joint. Specifically, the model's sensitivity for Garden I was 73.65% (95% CI, 69.85–79.17), specificity was 97.91% (95% CI, 97.42–98.27), and AUC was 95.51% (95% CI, 94.25–96.68); for Garden II, the sensitivity was 80.07% (95% CI, 77.63–82.20), specificity was 97.10% (95% CI, 96.66–97.52), and AUC was 96.26% (95% CI, 95.64–96.75); and for Garden III, the sensitivity was 78.19% (95% CI, 76.35–79.68), and the specificity was 94.38% (95% CI, ... The specificity for Garden IV was 94.37% (95% CI, 93.93-94.80), with an AUC of 94.37% (95% CI, 93.64-95.32), a specificity of 91.80% (95% CI, 91.21-92.25), and an AUC of 96.89% (95% CI, 96.53-97.13), as shown in Table 1.

[0031] Table 1. Model performance on the test set

[0032] On the internal test dataset (106 cases, 1563 images), the model achieved an average ROC curve (AUC) of 95.78%. The sensitivity, specificity, and other indicators for each subtype were consistent with the cross-validation results, as shown in the table below: Table 2. Model performance in cross-validation

[0033] Example 2: Model Generalization Test 1. Test Data Preparation The test data consisted of 2,192 CT and X-ray images of 277 subjects with femoral neck-related hip imaging provided by Jiangjin District People's Hospital in Chongqing and Zizhong County People's Hospital in Neijiang City, Sichuan Province. These images served as an external validation set to evaluate the model's performance under different data distributions and imaging equipment conditions.

[0034] 2. Testing Methods The model trained in Example 1 is directly applied to an external validation set without additional training or parameter adjustment. The generalization performance of the model is evaluated by metrics such as accuracy, sensitivity, specificity, and AUC.

[0035] 3. Test Results like Figure 3 As shown, the model achieved an average accuracy of 79.27% ​​and an average AUC of 74.77% on the external validation set. Despite the imbalance in data distribution, the model maintained good performance in image classification tasks corresponding to Garden III and IV types with large datasets, indicating that the model has a certain ability to adapt to cross-center data.

[0036] Example 3: Model Performance Comparison Test 1. Comparison Object Twelve orthopedic professionals from four centers were selected and divided into three groups based on their work experience: an expert group (4 people, with more than 10 years of work experience), a senior group (4 people, with 5 years of work experience), and a novice group (4 people, with 1 year of work experience).

[0037] 2. Test Methods The internal test set (1563 images from 106 image providers) in Example 1 was provided to three groups of technical professionals and the model of this invention, respectively. The diagnosis time and AUC value of each group were statistically analyzed and the performance was compared.

[0038] 3. Comparison Results For the same group of clinical patient data, the AUC curve for the expert group's detection and classification was 88, the AUC curve for the advanced group's detection and analysis was 72, and the AUC curve for the novice group's detection and analysis was 43. The model in this application achieved an AUC of 90 on the same dataset, surpassing the expert level. Furthermore, the model's diagnosis time was significantly reduced to 0.5 seconds per case, far lower than the expert group's 5 seconds per case, the advanced group's 8 seconds per case, and the novice group's 15 seconds per case. Specific data are shown in Table 3.

[0039] Table 3. Performance comparison of different groups in terms of diagnostic time

[0040] Example 4: Model-Assisted Application Testing 1. Test Object To test the clinical effectiveness of the model, four novice orthopedic professionals were selected and divided into a model-assisted group and a non-model-assisted group, with two people in each group.

[0041] 2. Testing Methods Both groups used an external validation set (2192 images from 277 subjects) for analysis. The model-assisted group used the analysis results of the model of this invention as a reference during the analysis process, while the group without model assistance relied solely on their own experience for analysis. The analysis accuracy of the two groups was statistically analyzed. At the same time, a simulated clinical workflow test was conducted, selecting 200 prospective images (including images with 15% artifacts) to evaluate the consistency rate between the model and expert consensus diagnosis, as well as the consistency of interpretation results between X-ray and CT. In addition, evaluation opinions of professional and technical personnel on the model were collected through anonymous questionnaires.

[0042] 3. Test Results The analysis accuracy of the model-assisted group was improved to 69.7%, while the analysis accuracy of the unassisted group was 43.6%, indicating that the model of the present invention can effectively assist novice professionals in improving the analysis accuracy of femoral neck-related features in hip imaging.

[0043] In a simulated clinical workflow, the model's classification results for 200 prospective images showed a 91% consistency rate with expert consensus diagnoses. The consistency between X-ray and CT interpretation results was 0.87, and the model output was compatible with DICOM 3.0 format, allowing for seamless embedding into existing PACS systems.

[0044] An anonymous questionnaire was used to collect doctors' support for the diagnostic assistance provided by the model in this application. The questionnaire content and feedback results are shown in the table below.

[0045] Table 4 Questionnaire Survey Results

[0046] In summary, this application proposes a novel deep learning-based system for detecting and classifying femoral neck fractures of the hip joint using the Garden model. This system enables end-to-end fracture detection and classification, ensuring the reliability of the results and significantly reducing the burden on doctors during the diagnostic process.

[0047] The model was trained on 7818 medical images from 529 patients across two centers and internally evaluated using 5-fold cross-validation. Its generalization ability was tested using an external validation set containing 2192 images from 277 patients across two additional centers. Results showed that in classification tasks, the model achieved an overall sensitivity of 82%, accuracy of 75%, and AUC of 78%. However, its sensitivity was relatively low on Garden I and II types, where data volume was limited, due to data imbalance. In detection tasks, the model also demonstrated high performance, but its ability to detect types with smaller sample sizes was slightly insufficient. External validation indicated that although the model's performance decreased under different data distributions, its diagnostic capability remained clinically acceptable. In assisting diagnosis, the model significantly reduced the workload for physicians and met real-time clinical needs.

[0048] The remaining technical features in the above embodiments can be flexibly selected by those skilled in the art to meet different specific practical needs according to actual circumstances. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims. In the above description, numerous specific details have been set forth to provide a thorough understanding of the present invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to implement the present invention. In other instances, to avoid obscuring the present invention, well-known techniques, such as specific construction details, operating conditions, and other technical conditions, have not been specifically described.

[0049] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A hip imaging model for detecting femoral neck fractures and a Garden classification model, characterized in that: The system includes a feature extraction network, a feature matching attention module, and a target detection and classification architecture. The feature extraction network uses EfficientNetV2 to extract global features from hip images. The feature matching attention module enhances the attention to local features of the target region in the hip image, performing local feature enhancement processing on the global features. The target detection and classification architecture uses YOLOv10, receives the fused features processed by the feature matching attention module, and outputs the target region localization result and the Garden classification result. The Garden classification includes four categories: Type I, Type II, Type III, and Type IV.

2. A method for detecting femoral neck fractures in hip imaging and constructing a Garden classification model, characterized in that, The steps are as follows: S1. Data collection and screening: Collect hip CT and X-ray images of patients with femoral neck fractures from multiple centers, exclude images that do not meet the quality requirements, exceed the preset duration, contain chronic hip diseases, or contain implanted hardware, and obtain a valid raw image dataset. S2. Data Labeling and Review: Draw a bounding box containing the femoral head, greater trochanter, and lesser trochanter in the fracture area of ​​the valid original image data, and label the corresponding Garden classification: For eversion with incomplete fracture, it is judged as non-displaced type I; for complete fracture but no displacement of the fracture ends, it is judged as non-displaced type II; for complete fracture with partial displacement, that is, partial change in the course of the trabeculae, it is judged as displaced type III; for complete fracture with complete displacement and complete change in the parallel course of the trabeculae, it is judged as displaced type IV. S3. Data Preprocessing and Augmentation: The reviewed labeled data is processed by contrast enhancement, Gaussian noise addition, elastic deformation, random cropping, and rotation within a preset angle range to expand the data volume to a preset multiple of the original data volume. The expanded data is consistent with the original training set in terms of gender, age, equipment manufacturer, and collection location distribution. S4. Model Construction: Construct a deep learning model that integrates feature matching and attention modules. Use EfficientNetV2 as the feature extraction network to extract global features of hip images. Enhance the attention of local features in the fracture area through the feature matching and attention module. After fusing the enhanced local features with global features, input them into the YOLOv10 target detection and classification architecture to form an end-to-end fracture detection and classification model. S5. Model Training and Validation: The preprocessed dataset is divided into a training set and an internal test set according to a preset ratio. A 5x cross-validation method is used, and the SGD optimization algorithm is used to train the model for a preset number of rounds. The model performance is evaluated by accuracy, sensitivity, specificity, F1 score and area under the curve. S6. Model generalization test: The trained model is tested using external validation sets from different medical centers to verify its performance under different data distributions and imaging equipment conditions.

3. The method for automatic detection of femoral neck fractures in hip imaging and construction of the Garden classification model according to claim 2, characterized in that: The multi-center image data mentioned in S1 comes from at least four different data acquisition institutions, covering 800 or more relevant image providers within a preset time period, with a total image data volume of no less than 10,000 images.

4. The automatic detection and Garden classification model construction of femoral neck fracture in hip imaging according to claim 2, characterized in that: In S3, contrast enhancement uses CLAHE technology, and elastic deformation, Gaussian noise addition, and rotation processing are all configured with preset parameters; the preset multiplier is 5x, and the preset angle range is ±15°.

5. The automatic detection and Garden classification model construction for femoral neck fractures in hip imaging according to claim 2, characterized in that: The model described in S4 was built using Python 3.10 and the PyTorch 2.3.1+cu118 open-source library. The model has a sensitivity of no less than 70% and a specificity of no less than 97% for Garden I, and a sensitivity of no less than 94% and an area under the curve of no less than 96% for Garden IV.

6. A hip imaging femoral neck fracture detection and Garden classification system constructed based on the method described in any one of claims 2-5, characterized in that, include: Data acquisition module: used to download hip CT and X-ray image data from the image storage system, supports DICOM 3.0 format, and automatically filters valid images that meet the requirements; Data preprocessing module: Used to perform data preprocessing and augmentation operations, outputting standardized training and test data; Model training module: Used to load preprocessed data, execute model building, training and internal validation processes, and generate trained deep learning model files; Image analysis module: Used to receive input hip images, call the trained model, and automatically output the target region bounding box coordinates, Garden classification results, and confidence scores of each classification. Report generation module: This module converts the output of the image analysis module into a standardized analysis report. The report includes fracture localization diagrams, classification results, analysis basis, and model confidence index. System integration module: used to embed the image analysis module and result output module into the existing image processing system, supporting users to manually correct the analysis results and record correction logs.

7. The automatic detection and Garden classification system for femoral neck fractures in hip imaging according to claim 6, characterized in that: The image analysis module takes no more than 0.5 seconds to analyze a single image, has a consistency rate of no less than 91% with the consensus analysis of professionals, and a classification accuracy of no less than 85% in image data containing a preset proportion of artifacts.

8. The automatic detection and Garden classification system for hip imaging femoral neck fractures according to claim 6, characterized in that: The system also includes a feedback module for collecting user feedback on the model analysis results, including satisfaction with processing speed, acceptance of analysis accuracy, reliability score of results, and assessment of application value, providing data support for subsequent model optimization.

Images