Fracture X-ray film AI auxiliary diagnosis system

Abstract

The invention discloses a fracture X-ray film AI auxiliary diagnosis system. The system adopts a deep learning architecture based on a convolutional neural network, and comprises an image preprocessing module, a deep learning recognition module, a diagnosis result output module and a user interaction interface module. The image preprocessing module carries out de-noising processing by combining adaptive median filtering with a bilateral filtering technology, and an improved Canny edge detector is combined with a deep learning edge enhancement network to identify skeleton features. The deep learning recognition module constructs a feature extraction network based on a ResNet-50 pre-training model, and adopts a weighted cross entropy loss function and an Adam optimizer for training. And the diagnosis result output module generates a visual fracture line label by applying a gradient weighted class activation mapping technology, and automatically generates a structured report containing fracture types, severity evaluation and diagnosis and treatment suggestions. The system supports DICOM and JPG format image input, the diagnosis accuracy reaches 78%-80%, and the fracture diagnosis efficiency and accuracy are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging artificial intelligence technology, and in particular to an AI-assisted diagnosis system for fracture X-rays. Background Art

[0002] Fracture diagnosis, a fundamental component of emergency medicine and orthopedics, has traditionally relied primarily on manual interpretation of X-ray images. In the current diagnostic process, physicians determine the presence of a fracture based on the continuity and alignment of bones in X-rays, as well as the morphological characteristics of the fracture line. They then assess the fracture type, location, and severity. However, this diagnostic approach, which relies solely on manual experience, has significant limitations. First, X-rays, as two-dimensional grayscale images, have relatively low contrast, and bone boundaries are easily blended with surrounding soft tissue or shadows, making it difficult to identify subtle fracture lines. This is particularly true in incomplete fractures in patients with osteoporosis or children. Second, diagnostic accuracy is highly dependent on the physician's clinical experience and expertise, and interpretations can vary subjectively between physicians, a problem particularly prominent in primary care settings. Furthermore, in emergency settings, physicians are required to process large volumes of imaging data in a short period of time. This excessive workload can lead to fatigue-induced misdiagnosis or missed diagnoses, hindering timely patient care.

[0003] In recent years, artificial intelligence technology has made significant progress in the field of medical imaging diagnosis, with deep learning algorithms demonstrating excellent performance in applications such as lung nodule detection, diabetic retinopathy screening, and skin cancer identification. Convolutional neural networks, with their powerful feature extraction and pattern recognition capabilities, provide a new technical path for the automated analysis of medical images. However, existing AI medical imaging diagnostic systems still face numerous technical challenges in the field of fracture identification. Most current research focuses on the identification of fractures at a single anatomical site or specific type, lacking a universal diagnostic system applicable to multiple sites and types of fractures. Furthermore, the accuracy of existing algorithms still needs to be improved when dealing with complex fracture morphologies, overlapping bone structures, and pathological fractures. More importantly, many existing systems only provide simple binary classification results and are unable to provide clinicians with detailed, practical information such as fracture location, type analysis, and treatment recommendations.

[0004] AI-assisted diagnosis of fracture X-rays also faces unique challenges in data quality and model training. Medical imaging data often suffers from an imbalanced sample distribution, with normal images often far outnumbering pathological images. This imbalance can easily lead models to favor predicting normal findings, thereby reducing sensitivity to fracture cases. Furthermore, the annotation of fracture images requires specialized medical knowledge and extensive clinical experience, and acquiring high-quality annotated data is costly and time-consuming. At the algorithmic level, the effective integration of multi-scale feature information, the handling of image noise, and the improvement of the ability to identify subtle fracture lines remain key bottlenecks in current technological development. Furthermore, clinical applications require systems with good interpretability, providing physicians with intuitive diagnostic evidence and confidence assessments. However, existing deep learning models are often considered "black box" systems, lacking transparency in their decision-making processes, limiting their widespread application in clinical settings. Summary of the Invention

[0005] Based on the above purpose, the present invention provides an AI-assisted diagnosis system for fracture X-rays

[0006] It includes an image preprocessing module, a deep learning recognition module, a diagnosis result output module and a user interaction interface module; the image preprocessing module includes an image format conversion unit, a size standardization unit and a data enhancement unit. The image format conversion unit supports X-ray image input in DICOM and JPG formats. The size standardization unit adjusts the input image to the standard size of 224×224 pixels and converts it into a tensor format. The data enhancement unit performs geometric transformation and photometric transformation operations on the fracture sample; the deep learning recognition module adopts a deep learning model based on a convolutional neural network, including multi-layer convolution layers, activation function layers, pooling layers, flattening layers, fully connected layers and output layers. The convolution layer extracts fracture features through convolution kernels, the pooling layer uses maximum pooling or average pooling to reduce spatial resolution, the fully connected layer learns the discriminant features between fractures and normal states, and the output layer outputs the fracture probability distribution through the softmax function; the diagnosis result output module uses gradient weighted class activation mapping technology to generate fracture line annotations and automatically generates a structured diagnosis report containing fracture type, severity assessment and treatment recommendations.

[0007] Furthermore, the image preprocessing module also includes a denoising and smoothing processing unit and an edge detection unit. The denoising and smoothing processing unit uses an adaptive median filtering algorithm combined with bilateral filtering technology to remove image noise, and the edge detection unit uses an improved Canny edge detector combined with a deep learning edge enhancement network to identify bone edge features.

[0008] Furthermore, the geometric transformation of the data enhancement unit includes random rotation in the range of -15 degrees to +15 degrees, random scaling in the range of 0.9 to 1.1 times, random translation in the range of 5% of the image width and height, and horizontal flipping with a probability of 50%. The photometric transformation includes brightness adjustment in the range of 0.8 to 1.2 times the original brightness, histogram equalization contrast enhancement, and gamma correction with a gamma value between 0.8 and 1.2.

[0009] Furthermore, the deep learning recognition module adopts the pre-trained ResNet-50 model as the feature extraction backbone network, uses the pre-trained weights of the ImageNet dataset for transfer learning, and redesigns the output layer into 2 neurons corresponding to the normal and fracture categories respectively.

[0010] Furthermore, the training process of the deep learning recognition module adopts a weighted cross-entropy loss function to assign higher weights to fracture samples to solve the sample imbalance problem. The optimizer adopts the Adam algorithm, and the learning rate adopts a progressive adjustment strategy. The initial learning rate is set to 0.001 and gradually reduced to 0.0001 during the training process.

[0011] Furthermore, the user interaction interface module includes a homepage function entrance, a customer management module and an image upload module. The homepage integrates the full-body posture assessment, custom measurement, local ROM assessment and fracture X-ray AI-assisted diagnosis function entrances. The customer management module supports the creation of new customer files and customer list viewing. The image upload module sets the image quality detection program and provides upload prompt information.

[0012] Furthermore, the diagnosis result output module includes a confidence assessment unit and a report format conversion unit. The confidence assessment unit calculates the probability value of fracture identification and provides accuracy level annotation. The report format conversion unit supports the output of PDF documents, Word documents and structured data formats.

[0013] Furthermore, the system also includes a statistical analysis module to record system usage and diagnostic results data, divide the training set, validation set and test set in an 8:1:1 ratio, set the training cycle to 10 times, the batch size to 32, and adopt the early stopping mechanism and model checkpoint saving mechanism to prevent overfitting.

[0014] Furthermore, the deep learning edge enhancement network adopts the U-Net architecture, including an encoder and a decoder. The encoder extracts multi-scale features, and the decoder restores edge details and outputs an enhanced edge image, which is combined with the traditional Canny algorithm to achieve accurate identification of fracture lines.

[0015] Beneficial effects of the present invention:

[0016] The fracture X-ray AI-assisted diagnosis system of the present invention has significant technical advantages and practical value compared with the existing technology, which is mainly reflected in the improvement of diagnostic accuracy, improvement of clinical application efficiency and optimization of medical resource allocation.

[0017] In terms of diagnostic accuracy, this system effectively addresses the subjectivity and experience-based interpretation inherent in traditional manual interpretation by utilizing a deep learning architecture based on the ResNet-50 pre-trained model, combined with a pre-processing algorithm specifically designed for the characteristics of X-ray images. The system currently achieves a diagnostic accuracy rate of 78% to 80%, and this accuracy is expected to increase further with the continued accumulation of sample data and optimization of the algorithm. In particular, the AI ​​algorithm is able to detect subtle fracture line identification and complex fracture morphology analysis, significantly reducing missed and misdiagnoses.

[0018] In terms of clinical efficiency, this system fully automates the entire process from image upload to diagnostic report generation, significantly reducing diagnostic time. While traditional manual interpretation of X-rays typically takes 10-30 minutes, this system can complete preliminary diagnostic analysis within minutes, saving valuable time for rapid treatment of emergency patients. Furthermore, the standardized diagnostic reports and visual fracture line annotations provided by the system provide physicians with intuitive diagnostic evidence, reducing their workload and improving overall diagnostic and treatment efficiency.

[0019] In terms of technological innovation, this system uniquely combines an improved Canny edge detection algorithm with a deep learning edge enhancement network, effectively addressing the technical challenges of low contrast and blurred bone boundaries in X-ray images. By employing gradient-weighted class activation mapping (GLM) technology to visualize fracture localization, the interpretability and credibility of AI diagnostic results are enhanced. The rational application of data augmentation strategies addresses the imbalance of fracture samples and avoids the training instability and medical authenticity risks associated with using generative adversarial networks.

[0020] To optimize the user experience, the system features a simple and intuitive interface that supports multiple image formats, including DICOM and JPG, to meet the compatibility requirements of diverse medical devices. One-click operation and intelligent quality control capabilities lower the barrier to entry for system users, enabling healthcare professionals at primary care facilities to easily utilize AI-assisted diagnostic services. Multi-format report output supports seamless integration with existing hospital information systems, eliminating the need for additional system upgrade costs.

[0021] In terms of medical resource allocation, this system can effectively alleviate the uneven distribution of high-quality medical resources. By bringing expert-level diagnostic capabilities to primary healthcare institutions in the form of AI, it improves fracture diagnosis at grassroots hospitals and reduces the need for patient referrals to larger hospitals. This not only reduces patients' medical expenses and time costs, but also helps achieve the goal of tiered diagnosis and treatment and optimizes the overall efficiency of medical resource allocation.

[0022] In terms of quality control and standardization, the system establishes unified diagnostic standards and evaluation systems, reducing diagnostic variability between doctors and improving the consistency and repeatability of diagnostic results. The system's built-in statistical analysis capabilities continuously monitor diagnostic quality, providing hospital management with a scientific basis for decision-making and promoting continuous improvement in medical quality.

[0023] From a long-term development perspective, this system offers excellent scalability and upgrade potential. Its modular system architecture facilitates future functional expansion, enabling the gradual integration of more types of medical imaging diagnostic capabilities. Continuous data accumulation and algorithm optimization will further enhance system performance, laying a solid foundation for building a more comprehensive intelligent medical diagnostic platform. The realization of these comprehensive benefits not only promotes the industrialization of medical imaging AI technology but also significantly contributes to improving the overall quality and efficiency of medical services. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0025] Figure 1 This is a schematic diagram of the overall architecture of the AI-assisted fracture X-ray diagnosis system of the present invention;

[0026] Figure 2 This is a schematic diagram of the deep learning convolutional neural network model structure of the present invention. DETAILED DESCRIPTION

[0027] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0028] It should be noted that references in the specification to "one embodiment," "an embodiment," "an exemplary embodiment," "some embodiments," etc. indicate that the described embodiments may include specific features, structures, or characteristics, but not every embodiment necessarily includes such specific features, structures, or characteristics. In addition, when specific features, structures, or characteristics are described in conjunction with an embodiment, it is within the knowledge of persons skilled in the relevant art to implement such features, structures, or characteristics in conjunction with other embodiments (whether or not explicitly described).

[0029] In general, terms can be understood, at least in part, from their use in context. For example, depending at least in part on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending at least in part on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0030] See Figure 1 shown

[0031] The AI-assisted fracture X-ray diagnosis system proposed in this paper uses a deep learning-based convolutional neural network architecture to automatically identify and diagnose fractures in DICOM or JPG format X-ray images. The overall system architecture includes a data preprocessing module, a deep learning model module, a diagnostic output module, and a user interface module.

[0032] The data preprocessing module undertakes the core function of image quality optimization. Considering that X-rays are two-dimensional grayscale images with relatively low contrast, bone boundaries are easily aliased with soft tissue or shadows. This system uses edge detection technology based on edge deep learning for preprocessing. The specific processing flow includes denoising and smoothing, data enhancement, normalization, and edge detection algorithm application. Denoising and smoothing removes image noise and smoothes image details through filtering algorithms. Data enhancement includes operations such as rotation, scaling, and contrast enhancement to expand the diversity of training samples. Normalization covers brightness normalization and pixel value normalization to unify data distribution. Edge detection algorithms are used to highlight bone contour features.

[0033] like Figure 2As shown in the figure, the deep learning model module adopts a convolutional neural network architecture, which consists of multiple functional layers. The convolution layer is responsible for extracting fracture features from the image, calculating and generating a feature map by sliding the convolution kernel across the image. The activation function layer introduces nonlinear transformations, enabling the model to fit complex feature patterns. The pooling layer reduces the spatial resolution to compress the data volume while retaining the key information, improving computational efficiency. After multiple layers of convolution and pooling, the flattening layer converts the two-dimensional feature map into a one-dimensional vector, providing input for the subsequent fully connected layer. The fully connected layer integrates the features extracted by each layer and learns the discriminative features that distinguish fractures from normal conditions. The output layer outputs the probability distribution of each category through the softmax function, making the final classification decision.

[0034] The system interface is designed around the actual hospital workflow. The homepage integrates multiple functional portals, including full-body posture assessment, custom measurements, local ROM assessment, and AI-assisted fracture X-ray diagnosis, providing doctors with a one-stop diagnostic tool. The customer management module supports creating new customer profiles and viewing customer lists, facilitating case management and tracking. The image upload module includes clear prompts, requiring users to upload high-resolution and appropriately sized X-rays to ensure the accuracy of AI analysis.

[0035] The detailed implementation steps of the above fracture X-ray AI-assisted diagnosis system are as follows:

[0036] Data preprocessing detailed process

[0037] Data preprocessing is a fundamental component of the system, and its specific steps have a decisive impact on the final diagnostic accuracy. First, upon receiving the raw X-ray image, the system immediately executes an image quality check program, automatically identifying key parameters such as image resolution, contrast, and noise level. This quality check program calculates the image's signal-to-noise ratio and contrast to determine whether quality enhancement processing is necessary.

[0038] Denoising and smoothing are achieved using an adaptive median filter algorithm combined with bilateral filtering techniques. The adaptive median filter dynamically adjusts the filter window size based on local pixel distribution characteristics, effectively removing salt-and-pepper noise while preserving bone edge detail. The bilateral filter performs filtering operations simultaneously in the spatial and grayscale domains, controlling the smoothing intensity through a weighting function to ensure that important features such as fracture lines are preserved while denoising. Filter parameters are automatically adjusted based on the type of image noise, with stronger filtering applied to noisy images and lighter filtering applied to high-quality images to avoid over-smoothing.

[0039] Data augmentation processing includes two dimensions: geometric transformation and photometric transformation. In terms of geometric transformation, the system performs a random rotation operation on each fracture sample image, and the rotation angle is randomly selected in the range of negative 15 degrees to positive 15 degrees to simulate X-ray images under different shooting angles. The scaling transformation is randomly performed in the range of 0.9 to 1.1 times to reflect the influence of different patient body shapes and shooting distances. The translation transformation is randomly performed in the range of 5% of the image width and height to simulate slight deviations in the shooting position. The horizontal flip transformation is applied with a probability of 50% to increase the diversity of the sample. Photometric transformation includes brightness adjustment, contrast enhancement and gamma correction. Brightness adjustment varies randomly in the range of 0.8 to 1.2 times the original brightness, contrast enhancement uses histogram equalization technology, and the gamma correction value is randomly selected between 0.8 and 1.2.

[0040] Normalization is performed in two stages. First, brightness normalization maps image pixel values ​​to a standard range of 0 to 1, eliminating brightness variations caused by different devices and shooting conditions. This is achieved by calculating the minimum and maximum values ​​of the image and then performing a linear mapping transformation. Next, pixel normalization is performed, standardizing pixel values ​​using a precalculated mean and standard deviation. Normalization parameters are derived from statistics of a large-scale X-ray image dataset to ensure consistent data distribution and stable model training.

[0041] The edge detection algorithm utilizes a modified Canny edge detector combined with a deep learning edge enhancement network. The traditional Canny algorithm first removes noise using Gaussian filtering, then calculates the image gradient magnitude and direction, performing non-maximum suppression and double-threshold detection. Building on the Canny algorithm, the deep learning edge enhancement network uses a convolutional neural network to learn the characteristic patterns of bone edges in X-ray images, enabling more accurate identification of key edge information such as fracture lines. This network utilizes a U-Net architecture, with an encoder extracting multi-scale features and a decoder restoring edge details, ultimately outputting an enhanced edge image.

[0042] Model training detailed process

[0043] The model training process uses a phased training strategy to ensure that the network effectively learns fracture characteristics and exhibits good generalization capabilities. The training dataset is split using an 8:1:1 ratio: 80% for training, 10% for validation, and 10% for testing. This data partitioning ensures that normal and fracture samples are distributed equally across all subsets, avoiding data skew.

[0044] During the network initialization phase, a transfer learning strategy based on pre-trained models was employed. The system selected a ResNet-50 model pre-trained on the ImageNet dataset as the backbone feature extraction network, using its learned general image features as initialization parameters. The pre-trained weights provide a sound feature representation foundation for the model, significantly reducing training time and improving final performance. The final classification layer of the network was redesigned for the fracture diagnosis task, with the number of neurons in the output layer set to 2, corresponding to the normal and fracture categories.

[0045] The training process uses a progressive learning rate adjustment strategy. The initial learning rate is set to 0.001 and remains unchanged for the first three training cycles, allowing the model to quickly converge to a relatively optimal solution. From the fourth to sixth training cycles, the learning rate is reduced to 0.0005 for fine-tuning. From the seventh to tenth training cycles, the learning rate is further reduced to 0.0001 to ensure that the model converges to the optimal solution. After each training cycle, the system automatically evaluates performance metrics on the validation set, including accuracy, precision, recall, and F1 score.

[0046] During batch training, the system monitors the trends of training and validation loss in real time. If validation loss fails to improve for three consecutive cycles, the system automatically triggers early stopping to prevent overfitting. Furthermore, the system employs a model checkpointing mechanism that automatically saves the current model parameters whenever validation set performance reaches a new optimal value. This ensures that even if training is interrupted, the model can be restored to its optimal state.

[0047] The loss function calculation process takes sample imbalance into account. The system uses a weighted cross-entropy loss function, assigning higher weights to fracture samples to compensate for the impact of insufficient sample numbers. Weights are calculated based on the inverse ratio of the number of samples in each category, ensuring that the model fully learns fracture characteristics without being dominated by normal samples. The optimizer uses the Adam algorithm, whose adaptive learning rate mechanism automatically adjusts the parameter update step size based on historical gradient information, improving training efficiency and stability.

[0048] Detailed steps of image processing and AI recognition

[0049] The image upload module implements multi-format compatibility and intelligent preprocessing. When a doctor uploads an X-ray image through the client, the system first performs a file format verification process, automatically identifying supported formats such as DICOM and JPG. For DICOM files, the system extracts the image data and metadata, including imaging parameters, patient information, and device information. This metadata is used for subsequent image standardization processing and diagnostic report generation.

[0050] The image normalization process utilizes an intelligent scaling algorithm to maintain the original image's aspect ratio while resizing it to a standard 224×224 pixel size. The system first calculates the original image's aspect ratio and then determines a scaling factor to ensure the image fits perfectly within the standard size. For any mismatches in aspect ratio, the system employs an edge-filling strategy, using pixel values ​​from the image's edges to minimize the impact of image distortion on diagnostic results.

[0051] The AI ​​recognition process consists of two stages: feature extraction and classification decision-making. During the feature extraction stage, the convolutional neural network's multiple layers of convolution kernels perform sliding convolutions on the image, with each kernel specifically detecting specific image feature patterns. Shallow convolution layers primarily extract basic features such as edges and textures, while deeper convolution layers learn more complex semantic features, such as the shape, orientation, and continuity of fracture lines. Pooling layers reduce the spatial resolution of feature maps through maximum or average pooling operations, reducing computational effort while preserving key feature information.

[0052] During the classification decision phase, the fully connected layer maps the extracted feature vectors into the decision space. The network learns the mapping between features and fracture states through linear transformations of the weight matrix and bias vector. The activation function introduces nonlinear transformations, enabling the network to learn complex decision boundaries. Finally, the output layer uses a softmax function to convert the network output into a probability distribution, representing the confidence level that the image belongs to the normal or fracture category.

[0053] The fracture line annotation process uses gradient-weighted class activation mapping (GLM) technology to achieve visual localization. This technology analyzes the gradient information of the convolutional feature map in the network's final layer to calculate the contribution of each pixel to the final classification decision. The system generates a heat map to display the area of ​​interest of the network. Then, through threshold segmentation and morphological processing, the precise location of the fracture line is extracted. The annotation results are superimposed on the original X-ray image as a color overlay, providing intuitive visual assistance to doctors.

[0054] Detailed process for generating a diagnostic report

[0055] The diagnostic report generation module integrates AI analysis results, a medical knowledge base, and standardized report templates. The system first determines the fracture type and location based on the AI ​​recognition results. Fracture classifications include complete, incomplete, and comminuted fractures, and location information includes specific anatomical locations. The system then matches the pre-built orthopedic knowledge base to provide corresponding diagnostic and treatment recommendations and precautions for each fracture type.

[0056] The report structured content generation process includes standardized sections such as patient basic information, examination items, imaging findings, AI analysis results, diagnostic opinions, and treatment recommendations. Basic patient information, including name, age, and gender, is extracted from uploaded metadata or user input. The examination items section records the specific X-ray site and projection direction. The imaging findings section uses standardized medical terminology to describe image features, including skeletal alignment, soft tissue status, and abnormal signs.

[0057] The AI ​​Analysis Results section details the algorithm's recognition process and confidence score. The system provides a numerical representation of the fracture probability, typically presented as a percentage, along with the level of recognition accuracy. For cases with low confidence, the system adds a warning message to the report, prompting the physician to perform a manual review. The Diagnostic Opinion section integrates AI analysis results with medical knowledge, providing preliminary diagnostic conclusions and differential diagnosis recommendations.

[0058] The treatment recommendation module generates a personalized treatment plan based on fracture type, severity, and patient characteristics. The system's built-in clinical decision support algorithm considers factors such as patient age, fracture location, and degree of displacement to automatically generate a comprehensive treatment plan, including conservative treatment, surgical indications, and rehabilitation recommendations. The report also includes follow-up recommendations and prognosis assessments, providing guidance for subsequent treatment.

[0059] Report output supports multiple standard formats, including PDF documents, Word documents, and structured data formats. PDF format reports use standardized medical report templates, including necessary information such as hospital identification, report number, and generation time. Word format facilitates further editing and supplementation by doctors, while structured data format supports seamless integration with hospital information systems. The system also provides report printing and electronic signature functions to meet the actual needs of clinical work.

[0060] Example 1

[0061] The following describes the implementation of the present invention in detail through a specific embodiment. In practical application, a doctor first uploads a patient's X-ray image file via a mobile application or desktop client. After receiving the image, the system automatically performs format detection and compatibility verification, supporting the DICOM medical standard format and the common JPG image format.

[0062] During image preprocessing, the system resizes the input image to a standard size of 224×224 pixels and converts it into a tensor format to meet the input requirements of deep learning models. Normalization standardizes pixel values, setting the mean and standard deviation to [0.5, 0.5, 0.5] to ensure consistency in data distribution. To address sample imbalance, the system uses a data augmentation strategy to expand the number of fracture samples. Specific methods include geometric transformation operations such as flipping, rotation, scaling, and translation, as well as lighting adjustments and random masking techniques. Traditional data augmentation techniques have been shown to have better stability and medical authenticity than generative adversarial network methods.

[0063] The model training process uses supervised learning, utilizing a labeled dataset containing approximately 1,000 normal samples and 200 fracture samples. Regarding training configuration, the system prioritizes GPU (CUDA) for accelerated computing, automatically switching to CPU mode in a GPU-less environment. The loss function uses cross-entropy loss to measure the difference between the predicted result and the true label. The optimizer uses the Adam algorithm with a learning rate of 0.001. The training process is set to 10 training cycles with a batch size of 32 to balance training efficiency and memory usage. The detailed model training process is as follows:

[0064] The model training process uses a phased training strategy to ensure that the network effectively learns fracture characteristics and exhibits good generalization capabilities. The training dataset is split using an 8:1:1 ratio: 80% for training, 10% for validation, and 10% for testing. This data partitioning ensures that normal and fracture samples are distributed equally across all subsets, avoiding data skew.

[0065] During the network initialization phase, a transfer learning strategy based on pre-trained models was employed. The system selected a ResNet-50 model pre-trained on the ImageNet dataset as the backbone feature extraction network, using its learned general image features as initialization parameters. The pre-trained weights provide a sound feature representation foundation for the model, significantly reducing training time and improving final performance. The final classification layer of the network was redesigned for the fracture diagnosis task, with the number of neurons in the output layer set to 2, corresponding to the normal and fracture categories.

[0066] The training process uses a progressive learning rate adjustment strategy. The initial learning rate is set to 0.001 and remains unchanged for the first three training cycles, allowing the model to quickly converge to a relatively optimal solution. From the fourth to sixth training cycles, the learning rate is reduced to 0.0005 for fine-tuning. From the seventh to tenth training cycles, the learning rate is further reduced to 0.0001 to ensure that the model converges to the optimal solution. After each training cycle, the system automatically evaluates performance metrics on the validation set, including accuracy, precision, recall, and F1 score.

[0067] During batch training, the system monitors the trends of training and validation loss in real time. If validation loss fails to improve for three consecutive cycles, the system automatically triggers early stopping to prevent overfitting. Furthermore, the system employs a model checkpointing mechanism that automatically saves the current model parameters whenever validation set performance reaches a new optimal value. This ensures that even if training is interrupted, the model can be restored to its optimal state.

[0068] The loss function calculation process takes sample imbalance into account. The system uses a weighted cross-entropy loss function, assigning higher weights to fracture samples to compensate for the impact of insufficient sample numbers. Weights are calculated based on the inverse ratio of the number of samples in each category, ensuring that the model fully learns fracture characteristics without being dominated by normal samples. The optimizer uses the Adam algorithm, whose adaptive learning rate mechanism automatically adjusts the parameter update step size based on historical gradient information, improving training efficiency and stability.

[0069] During the model inference phase, the system feeds the preprocessed image into a trained CNN model and calculates fracture probability through forward propagation. The system automatically annotates the identified fracture lines on the image and generates a structured report that includes fracture type, severity assessment, and treatment recommendations. Physicians can review the AI-annotated results and make final diagnostic decisions based on clinical experience. The system also offers a one-click diagnostic report generation function, integrating AI analysis results, image annotations, and relevant medical data into a standardized report format.

[0070] The statistical analysis module records system usage and diagnostic results, providing data support for hospital management and algorithm optimization. In the current implementation, the system achieves a diagnostic accuracy of 78% to 80% on a test set. This accuracy is expected to increase further with increasing sample sizes and continued algorithm optimization. The system's primary optimization focus is on increasing the sensitivity of AI recognition, reducing the false-negative rate, and ensuring the safety and reliability of clinical applications.

[0071] The model preservation mechanism ensures persistent storage of model parameters after training, supporting subsequent model deployment and version management. The system design adopts a modular architecture to facilitate functional expansion and performance optimization, laying the technical foundation for the subsequent integration of more medical imaging AI functions.

[0072] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0073] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. An AI-assisted diagnosis system for fracture X-rays, characterized in that: It includes an image preprocessing module, a deep learning recognition module, a diagnosis result output module and a user interaction interface module; the image preprocessing module includes an image format conversion unit, a size standardization unit and a data enhancement unit, the image format conversion unit supports X-ray image input in DICOM and JPG formats, the size standardization unit adjusts the input image to a standard size of 224×224 pixels and converts it into a tensor format, and the data enhancement unit performs geometric transformation and photometric transformation operations on the fracture sample; the deep learning recognition module adopts a deep learning model based on a convolutional neural network, including multiple convolution layers, activation function layers, pooling layers, flattening layers, fully connected layers and output layers, the convolution layers extract fracture features through convolution kernels, the pooling layers use maximum pooling or average pooling to reduce spatial resolution, the fully connected layers learn the discriminant features between fractures and normal states, and the output layer outputs the fracture probability distribution through the softmax function; the diagnosis result output module uses gradient weighted class activation mapping technology to generate fracture line annotations and automatically generates a structured diagnosis report containing fracture type, severity assessment and treatment recommendations.

2. The fracture X-ray AI-assisted diagnosis system according to claim 1, characterized in that: The image preprocessing module also includes a denoising and smoothing processing unit and an edge detection unit. The denoising and smoothing processing unit uses an adaptive median filtering algorithm combined with bilateral filtering technology to remove image noise, and the edge detection unit uses an improved Canny edge detector combined with a deep learning edge enhancement network to identify bone edge features.

3. The fracture X-ray AI-assisted diagnosis system according to claim 1, characterized in that: The geometric transformation of the data enhancement unit includes random rotation in the range of negative 15 degrees to positive 15 degrees, random scaling in the range of 0.9 to 1.1 times, random translation in the range of 5% of the image width and height, and horizontal flipping with a probability of 50%. The photometric transformation includes brightness adjustment in the range of 0.8 to 1.2 times the original brightness, histogram equalization contrast enhancement, and gamma correction with a gamma value between 0.8 and 1.

2.

4. The fracture X-ray AI-assisted diagnosis system according to claim 1, characterized in that: The deep learning recognition module uses a pre-trained ResNet-50 model as the feature extraction backbone network, uses the pre-trained weights of the ImageNet dataset for transfer learning, and redesigns the output layer into two neurons corresponding to the normal and fracture categories respectively.

5. The fracture X-ray AI-assisted diagnosis system according to claim 1, characterized in that: The training process of the deep learning recognition module adopts a weighted cross-entropy loss function to assign higher weights to fracture samples to solve the sample imbalance problem. The optimizer adopts the Adam algorithm, and the learning rate adopts a progressive adjustment strategy. The initial learning rate is set to 0.001 and gradually reduced to 0.0001 during the training process.

6. The fracture X-ray AI-assisted diagnosis system according to claim 1, characterized in that: The user interaction interface module includes a homepage function entrance, a customer management module and an image upload module. The homepage integrates full-body posture assessment, custom measurement, local ROM assessment and fracture X-ray AI-assisted diagnosis function entrances. The customer management module supports the creation of new customer files and customer list viewing. The image upload module sets an image quality detection program and provides upload prompt information.

7. The fracture X-ray AI-assisted diagnosis system according to claim 1, characterized in that: The diagnosis result output module includes a confidence assessment unit and a report format conversion unit. The confidence assessment unit calculates the probability value of fracture identification and provides accuracy level annotation. The report format conversion unit supports the output of PDF documents, Word documents and structured data formats.

8. The fracture X-ray AI-assisted diagnosis system according to claim 1, characterized in that: The system also includes a statistical analysis module that records system usage and diagnostic results data. The training set, validation set, and test set are divided into 8:1:1 ratios. The training cycle is set to 10 times, the batch size is set to 32, and an early stopping mechanism and a model checkpoint saving mechanism are used to prevent overfitting.

9. The fracture X-ray AI-assisted diagnosis system according to claim 2, characterized in that: The deep learning edge enhancement network adopts a U-Net architecture, including an encoder and a decoder. The encoder extracts multi-scale features, and the decoder restores edge details and outputs an enhanced edge image, which is combined with the traditional Canny algorithm to achieve accurate recognition of fracture lines.

Images