Intelligent identification system for osteoporosis area

Abstract

The invention discloses an intelligent identification system for an osteoporosis area, belongs to the field of image processing calculation, and aims to solve the problems of limited technical coverage and lack of a dynamic optimization mechanism. In a feature extraction stage, a system combines traditional image processing and deep learning technologies in parallel, extracts bone trabecula multidirectional texture features by using a Gabor filter and a local binary pattern algorithm, analyzes bone contour curvature by combining Sobel edge detection and morphological operation, constructs geometric morphological parameters, and performs feature extraction on the bone trabecula. The local branch focuses on the porosity and arrangement rule of the bone trabecula by adopting a 3D convolutional network, the global branch is embedded into a compression excitation module based on an improved MobileNetV3 network to strengthen the overall morphological expression of the bone, the response intensity of a local microstructure is enhanced by space attention, the weight of global bone topological characteristics is calibrated by channel attention, and a multi-scale splicing strategy is combined, so that the overall morphological expression of the bone trabecula is optimized. And finally, outputting a fusion feature matrix after noise suppression, and remarkably improving the expression ability of pathological features.

Description

Technical Field

[0001] The present invention relates to the field of image processing and computing, and in particular to an intelligent identification system for osteoporosis areas. Background Art

[0002] Osteoporosis is a systemic bone metabolic disease characterized by decreased bone mass and destruction of bone microstructure, leading to increased bone brittleness. It is common in the hips, spine and wrists, and is mainly manifested by easy fractures, back pain, hunchback and shortened height. In severe cases, it can cause respiratory disorders. The disease is mainly caused by physiological factors, genetics, unhealthy lifestyles, diseases or drugs, and is divided into primary, secondary and idiopathic categories. Life cycle management is the core of prevention and treatment. High-risk groups should undergo dual-energy X-ray bone density testing for early diagnosis and treatment.

[0003] Published patent: An intelligent identification system for osteoporosis areas (publication number: CN116993736A). The system obtains the distribution parameters of the bone scan image and the sliding window, and compares the local and overall distribution parameters to obtain the first sliding window. The window weight of the first sliding window is obtained according to the difference in the distribution parameters. The center window and the neighborhood window are obtained in combination with the positional relationship of the sliding window, and the confidence of the center window is further obtained. The window weight of the center window and the weight factor of the corresponding neighborhood window are used to adjust the weight factor of the center window to obtain the weighted confidence of the center window. The spine window is obtained according to the size of the weighted confidence, and the osteoporosis area is obtained using the spine window. The present invention accurately extracts the osteoporosis area in the bone scan image by extracting and analyzing the grayscale features of windows at different positions in the bone scan image.

[0004] The above patent has the following defects: the system mainly relies on image brightness changes and fixed window movement to identify problems, does not use advanced intelligent algorithms or traditional image texture analysis, and has difficulty capturing complex bone lesion characteristics. During the detection process, it cannot automatically adjust parameters according to actual needs, nor can it automatically update the model, resulting in difficulty in balancing detection efficiency and accuracy. There is a lack of technology for automatically generating data or borrowing other data experience, resulting in the system's ability to adapt to different patient conditions being weak. It can only mark osteoporosis areas and cannot generate detailed bone density reports that meet international standards or provide personalized risk warnings based on the patient's age and gender. It is difficult for doctors to use it directly for diagnosis. The results lack color-coded heat maps and multi-angle viewing functions. Doctors cannot quickly locate problem areas or verify the credibility of the results. It cannot run efficiently on medical equipment, has slow processing speeds, and cannot meet the hospital's real-time diagnosis needs. Summary of the Invention

[0005] The purpose of the present invention is to provide an intelligent identification system for osteoporosis areas, which solves the problems of limited technical coverage and lack of dynamic optimization mechanism.

[0006] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: an intelligent identification system for osteoporosis areas, comprising: an image acquisition and preprocessing module, a candidate area positioning and optimization module, a multimodal feature extraction module, a feature fusion and weighting module, a classification decision and clinical report module, a data enhancement and model optimization module and a real-time reasoning and feedback module, wherein the image acquisition and preprocessing module comprises a multimodal data interface, a resolution calibration unit, a contrast optimization unit, a noise suppression unit and a preprocessing unit, the candidate area positioning and optimization module comprises a dynamic adaptive sliding window unit, a region growing algorithm unit and a dynamic threshold segmentation unit, the feature fusion and weighting module comprises a traditional feature unit and a deep learning feature unit. The traditional feature unit includes a grayscale statistics subunit, a texture analysis subunit and a shape description subunit, the deep learning feature unit includes a local branch subunit, a global branch subunit and a topology modeling subunit, the multimodal feature extraction module includes a channel-space attention unit, a compression excitation unit and a multi-scale fusion unit, the classification decision and clinical report module includes a cascade classifier unit, a personalized report generation unit and a visualization output unit, the data enhancement and model optimization module includes an unsupervised data generation unit, a transfer learning unit, a model compression unit and a dynamic hot update unit, and the real-time reasoning and feedback module includes an embedded deployment unit, a dynamic parameter adjustment unit and a visualization interaction unit.

[0007] Furthermore, the image acquisition and preprocessing module supports multimodal medical image input and standardized processing, the multimodal data interface supports fully automatic access to CT, MRI, and X-ray data, the resolution calibration unit uses a bicubic interpolation algorithm to calibrate the image resolution to a standard pixel density, the contrast optimization unit dynamically adjusts the contrast through adaptive histogram equalization, the noise suppression unit removes Gaussian noise and equipment artifacts based on non-local mean filtering, and the morphological preprocessing unit combines morphological opening operations to eliminate interference from non-skeletal structures such as blood vessels and soft tissues.

[0008] Furthermore, the candidate region positioning and optimization module efficiently screens and optimizes suspected osteoporosis areas, the dynamic adaptive sliding window unit is based on prior knowledge of bone anatomy, the region growing algorithm unit is based on four-neighborhood and eight-neighborhood connectivity analysis, and the dynamic threshold segmentation unit adaptively selects segmentation strategies based on local grayscale variance differences.

[0009] Furthermore, the multimodal feature extraction module extracts traditional and deep learning features in parallel. The traditional feature unit includes a grayscale statistics subunit, a texture analysis subunit and a shape description subunit. The grayscale statistics subunit calculates the pixel mean, variance and grayscale distribution histogram of the candidate area. The texture analysis subunit analyzes the multi-directional texture of trabeculae through a Gabor filter. The shape description subunit extracts the bone contour curvature and connectivity features based on Sobel edge detection and morphological operations, and constructs geometric morphological parameters. The deep learning feature unit includes a local branch subunit, a global branch subunit and a topology modeling subunit. The local branch subunit extracts the microstructural texture features of trabecular porosity and arrangement regularity through a 3D convolutional network. The global branch subunit embeds a compressed incentive attention mechanism into the improved MobileNetV3 backbone network. The topology modeling subunit uses a graph convolutional network to model the spatial topology of key bones.

[0010] Furthermore, the feature fusion and weighting module dynamically weights and fuses multimodal features to suppress noise interference, the channel-spatial attention unit assigns spatial weights to local feature maps, the compression excitation unit dynamically calibrates the channel dimension of global features, and the multi-scale fusion unit fuses the local microstructure features of 3D convolution with the global topological features of graph convolution through channel splicing and element-by-element addition strategies.

[0011] Furthermore, the classification decision and clinical reporting module finely identifies osteoporosis areas and generates clinical reports. The first-level classifier support vector machine of the cascade classifier unit quickly eliminates low-confidence interference areas through kernel function mapping and hyperplane partitioning. The second-level classifier quantitatively evaluates the degree of bone degeneration through a lightweight residual network combined with the WHO bone density T-value standard. The personalized report generation unit generates a PDF / graphic report based on the patient's age and gender. The visual output unit generates a probability heat map and embeds it into a DICOM image.

[0012] Furthermore, the data enhancement and model optimization module solves the problems of data scarcity and model lightweighting. The unsupervised data generation unit synthesizes osteoporosis images of multiple age groups and multiple disease degrees based on CycleGAN. The transfer learning unit uses the ImageNet pre-trained model to fine-tune the classifier, greatly reducing the demand for labeled data. The model compression unit compresses the model through knowledge distillation and quantitative pruning. The dynamic hot update unit realizes seamless iterative update of the model through OTA technology.

[0013] Furthermore, the real-time reasoning and feedback module realizes low-latency reasoning and dynamic parameter optimization, the embedded deployment unit deploys the compression model on the Jetson Nano platform, the dynamic parameter adjustment unit dynamically adjusts the sliding window step size and Gabor filter kernel size according to the confidence score, and the visualization interaction unit generates a probability heat map and superimposes it on the original image.

[0014] Furthermore, the system workflow is as follows:

[0015] S1. Multimodal Image Input and Preprocessing

[0016] Input: CT, MRI, X-ray medical images;

[0017] Output: normalized preprocessed image;

[0018] 1. Multimodal data access: automatic recognition and analysis of CT, MRI, and X-ray imaging data;

[0019] 2. Standardization preprocessing: unify image resolution to standard pixel density, and optimize image contrast through adaptive histogram equalization;

[0020] 3. Interference suppression: remove Gaussian noise and equipment artifacts, and eliminate interference from non-skeletal structures such as blood vessels and soft tissues;

[0021] S2. Dynamic positioning of candidate regions

[0022] Input: preprocessed normalized image;

[0023] Output: optimized candidate region coordinates and image slices;

[0024] 1. Dynamic sliding window initial screening: Prioritize detection of high-incidence areas such as the spine and femoral neck, and dynamically adjust the window size and step length based on the image signal-to-noise ratio;

[0025] 2. Region growing algorithm optimization: merge spatially adjacent candidate regions to reduce redundant calculations;

[0026] 3. Dynamic threshold segmentation: Optimize the segmentation threshold based on local grayscale variance to exclude low-contrast interference areas;

[0027] S3. Multimodal feature extraction

[0028] Input: optimized candidate region coordinates and image slices;

[0029] Output: traditional feature vector and deep learning feature vector;

[0030] 1. Traditional feature extraction: Grayscale statistics: calculate the mean, variance and grayscale distribution histogram of the candidate region; Texture analysis: Gabor filtering and LBP algorithm are used to extract trabecular texture; Shape description: Sobel edge detection and morphological operations are used to calculate the curvature of the bone contour;

[0031] 2. Deep learning feature extraction: Local branch: 3D convolutional network extracts trabecular microstructure texture, global branch: MobileNetV3 extracts global bone features, and graph convolutional network models the topological relationship of key bones;

[0032] S4. Feature fusion and noise suppression

[0033] Input: traditional feature vector and deep learning feature vector;

[0034] Output: fused multimodal feature matrix;

[0035] 1. Channel-spatial attention weighting: Assign spatial weights to local features of 3D convolution output to enhance key area responses;

[0036] 2. SE module enhances global features: dynamically weights the channel features output by MobileNetV3 to suppress irrelevant background;

[0037] 3. Multi-scale fusion strategy: cross-scale splicing of local texture features and global topological features;

[0038] S5. Cascade classification and clinical decision making

[0039] Input: fused multimodal features;

[0040] Output: Clinical report recording T value, regional coordinates, risk level results and corresponding confidence score data, probability heat map with confidence score annotated, and PDF / graphic report;

[0041] 1. First-level classifier: Screening high-confidence candidate regions based on SVM;

[0042] 2. Second-level classifier: Use lightweight ResNet to further determine the degree of osteoporosis;

[0043] 3. Output clinically applicable test results: WHO standard grading: output bone density T value to assess the degree of osteoporosis; personalized risk assessment: generate PDF / graphic report based on patient age and gender, marking high-risk areas;

[0044] S6. Real-time Reasoning and Dynamic Feedback

[0045] Input: classification results and confidence scores;

[0046] Output: real-time detection results and interactive interface;

[0047] 1. Embedded deployment: compress the model to adapt to embedded devices;

[0048] 2. Dynamic parameter optimization: Automatically adjust the sliding window step size and Gabor filter kernel scale based on the confidence score;

[0049] 3. Visual interaction: Mark the probability of osteoporosis areas in the heat map and embed confidence scores in DICOM images;

[0050] S7. Closed-loop model optimization

[0051] Run in the background

[0052] Input: false positive / missed positive samples annotated by doctors;

[0053] Output: iteratively optimized model version;

[0054] 1. Unsupervised data augmentation: Use CycleGAN to generate synthetic images of multiple age groups and disease severity;

[0055] 2. Transfer learning fine-tuning: Based on the pre-trained ImageNet model, fine-tune the classifier using a small amount of labeled data;

[0056] 3. Model compression and update: Through knowledge distillation, the teacher model ResNet50 is compressed into the student model MobileNetV3;

[0057] 4. Dynamic hot update: Deploy the optimized model to the terminal device via OTA.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] 1. The present invention supports the input of various medical images such as CT, MRI, and X-ray, and is compatible with the international formats of DICOM and NIfTI. It extracts the image pixel matrix and metadata through a fully automatic parsing engine to achieve seamless access to data across devices. To improve image quality, the system uses a bicubic interpolation algorithm to calibrate the resolution to the standard pixel density, and dynamically adjusts the contrast through adaptive histogram equalization to enhance the visibility of details in the trabecular and cortical bone areas, while suppressing local overexposure or underexposure problems. In the feature extraction stage, the system combines traditional image processing and deep learning technology in parallel, quantifies the bone density distribution through grayscale statistics, uses Gabor filters and local binary pattern algorithms to extract multi-directional texture features of trabecular bone, and combines Sobel edge detection with morphological operations to analyze the curvature of bone contours and construct geometric morphological parameters. The local branch uses a 3D convolutional network to focus on the porosity and arrangement rules of trabecular bone, and the global branch is based on the improved Mobi The leNetV3 network embeds a compression excitation module to enhance the overall morphological expression of the skeleton. At the same time, it uses a graph convolutional network to model the topological relationships of key bones such as the vertebral body and femoral head. Spatial attention enhances the response strength of local microstructures, and channel attention calibrates the weights of global bone topological features. Combined with a multi-scale splicing strategy, it finally outputs a fusion feature matrix after noise suppression, significantly improving the expressiveness of pathological features.

[0060] 2. This invention balances detection efficiency and accuracy through a real-time feedback mechanism. Based on the confidence score output by the classifier, the sliding window step size and Gabor filter kernel size are dynamically adjusted. The step size is reduced to improve accuracy when the confidence level is low, and the kernel size is increased to accelerate processing when the confidence level is high. To meet clinical real-time requirements, the system uses knowledge distillation and quantization pruning to compress the model, adapting it to embedded device deployment. CUDA parallel computing accelerates inference and enables simultaneous multimodal image processing. Based on the CycleGAN framework, osteoporosis images covering different age groups, genders, and lesion severity are synthesized. Trabecular density distribution constraints are used to ensure the authenticity of the synthesized data, increasing the data volume to several times the original. The top-level classifier is fine-tuned using an ImageNet pre-trained model, combined with focal loss to address class imbalance, achieving high-precision classification with a small amount of labeled data. Knowledge distillation is used to transfer knowledge from the teacher model to a lightweight student model, preserving discriminative power while reducing computational overhead. The optimized model is transmitted to the terminal device via encrypted differential packets, supporting over-the-air updates and version rollbacks. Blockchain hash verification ensures update security, continuously reducing false positive rates without compromising real-time inference efficiency.

[0061] 3. The present invention adopts cascade classifiers to achieve accurate discrimination. The first-level support vector machine quickly screens high-confidence candidate areas and eliminates interference such as calcification points. The second-level lightweight residual network combines the WHO bone density T-value standard to quantitatively evaluate the degree of bone degeneration and integrate the patient's age and gender to generate a personalized risk level report. The result output includes a structured report and visual interaction. The system continuously improves performance through closed-loop optimization. The false detection / missed detection samples marked by doctors are fed back to the background, triggering unsupervised data enhancement and model fine-tuning. The visual interface provides multi-dimensional interactive functions to enhance the interpretability of the results, helping doctors to formulate diagnosis and treatment plans while providing high-quality data support for algorithm iteration. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 is a system flow chart of the present invention;

[0063] Figure 2 FIG. 1 is a diagram of the module architecture of the present invention. FIG. DETAILED DESCRIPTION

[0064] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0065] In order to solve the technical problems of lack of systematic methods and tools for product innovation and service optimization, which are difficult to meet the needs of enterprises, such as Figure 1-2 As shown, the following preferred technical solutions are provided:

[0066] 1. An intelligent identification system for osteoporosis areas, characterized by comprising: an image acquisition and preprocessing module, a candidate area positioning and optimization module, a multimodal feature extraction module, a feature fusion and weighting module, a classification decision and clinical reporting module, a data enhancement and model optimization module, and a real-time reasoning and feedback module. The image acquisition and preprocessing module includes a multimodal data interface, a resolution calibration unit, a contrast optimization unit, a noise suppression unit, and a preprocessing unit. The candidate area positioning and optimization module includes a dynamic adaptive sliding window unit, a region growing algorithm unit, and a dynamic threshold segmentation unit. The feature fusion and weighting module includes a traditional feature unit and a deep learning feature unit. The traditional feature unit includes a grayscale statistics subunit, a texture analysis subunit and a shape description subunit; the deep learning feature unit includes a local branch subunit, a global branch subunit and a topology modeling subunit; the multimodal feature extraction module includes a channel-space attention unit, a compression excitation unit and a multi-scale fusion unit; the classification decision and clinical report module includes a cascade classifier unit, a personalized report generation unit and a visualization output unit; the data enhancement and model optimization module includes an unsupervised data generation unit, a transfer learning unit, a model compression unit and a dynamic hot update unit; the real-time reasoning and feedback module includes an embedded deployment unit, a dynamic parameter adjustment unit and a visualization interaction unit.

[0067] The image acquisition and preprocessing module supports multimodal medical image input and standardized processing. The multimodal data interface supports fully automatic access to CT, MRI, and X-ray data, is compatible with DICOM and NIfTI format parsing, extracts pixel matrices and metadata, and ensures cross-device compatibility. The resolution calibration unit uses a bicubic interpolation algorithm to calibrate the image resolution to the standard pixel density. The contrast optimization unit dynamically adjusts the contrast through adaptive histogram equalization to enhance the visibility of details in the trabecular and cortical bone areas and suppress local overexposure and underexposure problems. The noise suppression unit removes Gaussian noise and equipment artifacts based on non-local mean filtering to preserve the integrity of bone microstructure. For pixel i in the medical image, its filtered pixel value u(i) is calculated as follows: Where Ω is the set of all pixels in the image, v(j) is the original grayscale value of pixel j, is the normalization factor, and the weight w(i,j) is: In medical image processing, N i and N j It is a local neighborhood centered on pixels i and j, and the neighborhood selection here takes into account the characteristics of the bone microstructure. It is the weighted Euclidean distance between two neighborhoods, a is the weighting parameter, which is adjusted according to the characteristics of the bone image, and h is the parameter that controls the smoothness of the filter, which will be adjusted according to the image noise level and the need to retain bone details. The morphological preprocessing unit combines morphological opening operations to eliminate interference from non-bone structures such as blood vessels and soft tissues, and separates the bone body from residual noise through connected domain analysis, and finally outputs a high-purity standardized bone image, providing an accurate data basis for subsequent osteoporosis detection. It adapts to most medical equipment through multimodal compatibility, greatly improves the bone area retention rate, and the output image quality meets international medical image analysis standards.

[0068] The candidate region positioning and optimization module efficiently screens and optimizes suspected osteoporosis areas from pre-processed medical images through the collaborative work of dynamic adaptive sliding window, region growing algorithm and dynamic threshold adjustment. The dynamic adaptive sliding window unit prioritizes high-incidence areas such as spine and femoral neck as initial candidate areas based on prior knowledge of bone anatomy, and dynamically adjusts the size and step length of the sliding window according to the real-time image signal-to-noise ratio. When the signal-to-noise ratio is low, the window size is increased to enhance robustness, and when the signal-to-noise ratio is high, the step length is reduced to improve detection accuracy. The pre-trained model is used to guide the initial screening of candidate areas to reduce invalid calculation overhead. The long algorithm unit is based on the connectivity analysis of four-neighborhood and eight-neighborhood, merges the candidate areas with similar pixel intensities and spatially adjacent spaces. By setting the grayscale difference threshold, the fragmented areas are aggregated into complete lesion areas, which significantly reduces the amount of redundant calculations. The dynamic threshold segmentation unit adaptively selects the segmentation strategy according to the difference in local grayscale distribution. The Otsu algorithm is used to generate the threshold for the low-contrast area. Assuming that the grayscale level of the image is 0 to L-1, the probability p(i) of each grayscale level i is calculated. The probability calculation here is based on the pixel statistics in the low-contrast area, dividing the image into foreground and background. The threshold is t, and the probability of the foreground pixel is Probability of background pixels Mean value of foreground pixels The mean value of background pixels Between-class variance By traversing all possible thresholds t∈[0,L-1], find the between-class variance The maximum threshold t * ,Right now This is used to segment the trabecular and cortical bone structures in low-contrast areas, and a fixed threshold is used for segmentation in high-contrast areas. At the same time, by excluding areas of uniform soft tissue interference, the trabecular and cortical bone structures are accurately separated to ensure that the detection sensitivity and specificity meet clinical requirements. Finally, the optimized candidate area coordinates and image slices are output to provide accurate bone structure segmentation results for the subsequent feature extraction module.

[0069] The multimodal feature extraction module achieves multi-dimensional feature representation of osteoporosis areas by extracting traditional image features and deep learning features in parallel, combining local microstructure analysis with global bone topology modeling. The traditional feature unit includes a grayscale statistics subunit, a texture analysis subunit, and a shape description subunit. The traditional feature unit quantifies the physical properties of bone density, texture, and shape based on grayscale statistics, texture filtering, and morphological operations. The grayscale analysis subunit calculates the mean, variance, and grayscale distribution histogram of pixels in the candidate area to characterize the bone density distribution. The texture analysis subunit analyzes the multi-directional texture of trabeculae through Gabor filters. The two-dimensional Gabor filter formula is: When analyzing trabecular bone texture, is the position vector of the trabecular bone area in the medical image, is the frequency vector of the filter, which is set according to the direction and frequency characteristics of the trabecular texture. σ is the standard deviation of the Gaussian function, which is used to control the bandwidth of the filter and is also adjusted according to the thickness of the trabecular texture. is a constant term to remove the DC component, and the local binary pattern algorithm is used to capture the microstructural heterogeneity, with the center pixel g c The 3*3 neighborhood centered on the neighborhood pixel g p , LBP code LBP P,R Calculated as: Where P is the number of neighborhood pixels, usually P = 8, R is the neighborhood radius, which can be adjusted according to the size of the structure when analyzing trabecular microstructure, and s(x) is the sign function. The heterogeneous characteristics of microstructures are captured by calculating LBP codes. The shape description subunit extracts bone contour curvature and connectivity features based on Sobel edge detection and morphological operations, and constructs geometric morphological parameters. The deep learning feature unit consists of a local branch subunit, a global branch subunit, and a topology modeling subunit. A local and global branching strategy is adopted. The local branch subunit extracts microstructural texture features such as trabecular porosity and arrangement regularity through a 3D convolutional network. The global branch embeds a compressed excitation attention mechanism based on an improved MobileNetV3 backbone network to enhance the saliency of overall bone morphological features. The topology modeling subunit models the spatial topological relationships of key bones through a graph convolutional network. The two branches achieve collaborative analysis through a multi-scale fusion strategy. The microstructural features of the local branch and the bone topological features of the global branch are dynamically weighted and fused through a channel-spatial attention unit. Channel attention is used to calibrate the global feature channel weights. Spatial attention enhances the response strength of trabecular and cortical bone edges in the local feature map, suppresses noise interference, and enhances key pathological features. The final fused multimodal feature matrix provides a high-dimensional feature representation that combines local details and global structural information for subsequent classification and decision-making.

[0070] The feature fusion and weighting module combines multimodal features through dynamic weighting to suppress noise interference and enhance the expression of key pathological features. The channel-spatial attention unit works in conjunction with the compression excitation module to assign spatial weights to the trabecular microstructure and cortical bone edges in the high-value areas of the local feature map. is the input feature map, σ is the Sigmoid activation function, δ is the ReLU activation function, To perform element-by-element multiplication, enhance its response strength and suppress noise interference in low signal-to-noise ratio areas, the compression excitation unit dynamically calibrates the channel dimension of the global feature, compresses the channel dimension of the global feature map to generate a weight vector, and enhances the significant expression of vertebral morphology and femoral head spatial arrangement topological structure. The multi-scale fusion unit fuses the local microstructure texture features extracted by 3D convolution with the global bone topology features modeled by the graph convolution network across scales, fusing 3D convolution and graph convolution features: is element-by-element addition, α is the balance coefficient, and Concat is channel splicing. A multimodal fusion feature matrix is constructed by combining channel splicing with element-by-element addition to achieve collaborative analysis of local details and global structures, significantly improving the sensitivity and specificity of osteoporosis area detection. Finally, the optimized fusion feature matrix is output for use by subsequent classification decision modules.

[0071] The classification decision and clinical reporting module uses cascade classifiers and multidimensional data analysis technology to achieve refined identification of osteoporosis areas and clinical decision support. The first-level classifier support vector machine of the cascade classifier unit performs preliminary screening of the fusion feature matrix, quickly eliminates low-confidence interference areas through kernel function mapping and hyperplane partitioning, and strictly controls the false detection rate. The first-level classifier: K(x,x i ) is the kernel function, RBF kernel: K(x,x i )=exp(-γ‖xx i ‖ 2 )), α i is the Lagrange multiplier, y i ∈{-1,1} is the category label. The second-level classifier uses a lightweight residual network to perform in-depth analysis on the candidate area after the initial screening, and combines the World Health Organization bone density T value standard to quantitatively evaluate the degree of bone degeneration. The second-level classifier: y=x+F(x,W i )(skip connection), F is the convolution operation, W iis the weight, x is the input feature, and the personalized report generation unit integrates clinical data such as patient age and gender, optimizes feature selection capabilities through the channel attention mechanism, and generates personalized osteoporosis low-risk, medium-risk, and high-risk levels. The visualization output unit generates a probability heat map and embeds it into DICOM images, supports coronal and sagittal multi-plane reconstruction views and dynamic threshold adjustment, and finally outputs a bone density T-value quantitative report that meets WHO standards, including structured data area coordinates, risk level, confidence score, and visualization heat map. High-risk areas are marked in PDF and DICOM formats, supporting doctors to quickly locate lesions and formulate diagnosis and treatment plans, achieving high precision and high interpretability in clinical decision-making.

[0072] The data enhancement and model optimization module solves the problem of scarce medical image annotation data and realizes lightweight model deployment through unsupervised data generation, transfer learning and model compression technology. The unsupervised data generation unit synthesizes multi-dimensional osteoporosis images covering different age groups, genders and disease severity based on the CycleGAN framework. The loss function is: adversarial loss + cycle consistency loss + pathological feature constraint.

[0073] G is the generator, synthesizing osteoporosis images, D X is the discriminator, φ is the pathological feature extractor, x patho For real pathological samples, the synthetic data is ensured to be consistent with real cases in microstructure and macromorphology through pathological feature constraints. The data volume is increased to several times the original data. The transfer learning unit uses the ImageNet pre-trained model for domain adaptive fine-tuning, freezes the underlying parameters of the backbone network, and only optimizes the top-level classifier. Combined with focal loss to solve the problem of category imbalance, the model can still achieve high-accuracy classification with a small amount of labeled data. The model compression unit uses knowledge distillation technology to transfer the knowledge of the teacher model to the lightweight student model. Model compression: teacher model T to student model S:

[0074] L CE is the cross entropy loss, T is the temperature parameter, and the pseudo-quantization formula is: b is the number of quantization bits, △ is the quantization step size, and quantization-aware training and channel pruning are combined to compress the model and adapt it to embedded device deployment, reducing accuracy loss. The dynamic hot update unit uses OTA technology to achieve seamless iterative updates of the model, ensuring update security through encrypted differential packet transmission and blockchain hash verification, supporting incremental upgrades and version rollbacks, and continuously optimizing system performance.

[0075] The real-time reasoning and feedback module realizes real-time efficient reasoning and closed-loop feedback of medical images through lightweight model deployment, dynamic parameter optimization and visual interaction technology. The embedded deployment unit compresses the model based on knowledge distillation and quantitative pruning technology, and the pruning standard Sum of absolute values of weights, based on channel importance score S c , retain the top k% high-scoring channels, delete redundant channels, adapt to embedded devices, use CUDA parallel computing to achieve low-latency inference, support multi-modal image synchronous processing, the dynamic parameter adjustment unit dynamically adjusts the detection parameters according to the confidence score output by the classifier, reduces the sliding window step size to improve detection accuracy when the confidence is low, and increases the Gabor filter kernel size to accelerate feature extraction when the confidence is high, balancing efficiency and accuracy, and reducing redundant calculations. The visualization interaction unit generates a probability heat map of the osteoporosis area and superimposes it on the original image, embeds the confidence score through the DICOM standard interface, supports coronal and sagittal multi-plane reconstruction views and dynamic threshold adjustment functions, assists doctors to quickly verify the diagnosis results, and ultimately achieves low-latency and high-throughput real-time processing capabilities, dynamically balanced detection strategies and multi-dimensional visualization interaction interfaces, significantly improving the practicality and interpretability of AI-assisted diagnosis in clinical scenarios, and combines wireless update technology to achieve seamless hot updates of the model and continuously optimize system performance.

[0076] The system workflow is as follows:

[0077] S1. Multimodal Image Input and Preprocessing

[0078] Input: CT, MRI, X-ray medical images;

[0079] Output: normalized preprocessed image;

[0080] 1. Multimodal data access: The system supports fully automatic access to CT, MRI, and X-ray multimodal imaging data. It achieves cross-device data compatibility through DICOM and NIfTI format parsing engines. A built-in decoder extracts pixel matrices and metadata, including slice thickness and scan parameters, ensuring efficient access and unified management of multi-source heterogeneous images, providing standardized input for subsequent processing.

[0081] 2. Standardization preprocessing: The input image is resolution-calibrated, and the image is unified to a standard pixel density using a bicubic interpolation algorithm. The contrast is dynamically adjusted using an adaptive histogram equalization algorithm to enhance the visibility of details in the trabecular and cortical bone areas while suppressing local overexposure or underexposure, ensuring that the image quality meets analysis requirements.

[0082] 3. Interference suppression: Based on non-local mean filtering, Gaussian noise and equipment artifacts are removed, while preserving bone texture features. Morphological opening operations, i.e., corrosion followed by dilation, are combined to eliminate interference from non-bone structures such as blood vessels and soft tissue. Connected domain analysis is used to separate the main bone from residual noise, ultimately outputting a high-purity image of the bone region, providing an accurate data foundation for subsequent osteoporosis testing.

[0083] S2. Dynamic positioning of candidate regions

[0084] Input: preprocessed normalized image;

[0085] Output: optimized candidate region coordinates and image slices;

[0086] 1. Dynamic sliding window initial screening: Based on anatomical prior knowledge that the spine and femoral neck are high-incidence areas of osteoporosis, the system prioritizes the initial window position and range, guides the positioning of candidate areas through a pre-trained model, and dynamically adjusts the window size based on the real-time image signal-to-noise ratio. Specifically, when the signal-to-noise ratio is below the set threshold, the window size is increased to improve robustness, and when the signal-to-noise ratio is above the set threshold, the step size is reduced to enhance detection accuracy. This implements an adaptive traversal strategy and significantly reduces the computational overhead of invalid areas.

[0087] 2. Region growing algorithm optimization: Based on the initial sliding window screening, the region growing algorithm is used to merge spatially adjacent candidate regions. Through pixel intensity similarity and connectivity analysis, namely thresholded distance and grayscale difference, fragmented regions are aggregated into complete lesion areas to avoid repeated detection. In combination with eight-neighborhood or four-neighborhood growing strategies, merging efficiency is optimized, redundant calculations are reduced, and the overall system operation speed is improved.

[0088] 3. Dynamic Threshold Segmentation: Based on the grayscale variance within the sliding window, the segmentation threshold is adjusted in real time to address differences in local image grayscale distribution. The adaptive threshold of the Otsu algorithm is used in low-contrast areas, while a fixed threshold is used in high-contrast areas (where the variance is greater than or equal to 50). By excluding low-variance interference areas (i.e., uniform soft tissue), trabecular and cortical bone structures are accurately segmented, ensuring significantly improved sensitivity and specificity in detecting osteoporotic areas.

[0089] S3. Multimodal feature extraction

[0090] Input: optimized candidate region coordinates and image slices;

[0091] Output: traditional feature vector and deep learning feature vector;

[0092] 1. Traditional feature extraction: The system calculates the mean, variance, and grayscale distribution histogram of candidate regions through grayscale statistics to quantify bone density distribution characteristics. It then uses Gabor filters and local binary pattern algorithms to extract multi-directional texture features of trabeculae, capturing microstructural changes. It then combines Sobel edge detection with morphological operations, such as erosion and dilation, to extract bone contour curvature and construct shape description parameters, providing quantitative indicators for osteoporosis diagnosis based on traditional image processing.

[0093] 2. Deep Learning Feature Extraction: The local branch uses a three-dimensional convolutional network to extract the microstructural texture features of trabecular bone, focusing on the porosity and arrangement of local trabecular bone. The global branch uses an improved MobileNetV3 backbone network embedded with a compression-incentive attention mechanism to extract the overall morphological features of the bone. A graph convolutional network is used to model the spatial topological relationships of key bones such as the vertebral body and femoral head, enabling the coordinated analysis of local details and global structure, enhancing the ability to express the pathological characteristics of osteoporosis.

[0094] S4. Feature fusion and noise suppression

[0095] Input: traditional feature vector and deep learning feature vector;

[0096] Output: fused multimodal feature matrix;

[0097] 1. Channel-Spatial Attention Weighting: Based on the spatial attention mechanism, dynamic weights are assigned to the local feature maps output by the three-dimensional convolution. A learnable attention mask is used to enhance the response strength of high-value areas such as trabecular microstructure and cortical bone edges, while suppressing noise interference in low signal-to-noise ratio areas, allowing the model to focus on key local features related to osteoporosis.

[0098] 2. Compression-excitation module enhances global features: The compression-excitation module is embedded in the MobileNetV3 backbone network. It dynamically calibrates global features through a channel-attention mechanism. This module first compresses the channel dimension of the global feature map, i.e., global average pooling. After generating a channel weight vector, the original features are weighted and scaled in the channel dimension. This enhances the salience of skeletal topology, i.e., vertebral morphology and femoral head spatial arrangement, while suppressing interference from irrelevant information such as soft tissue background.

[0099] 3. Multi-scale fusion strategy: This method integrates the local microstructural texture features (i.e., trabecular pore distribution) extracted by 3D convolution with the global skeletal topology features (i.e., spatial relationships between vertebrae) modeled by graph convolutional networks. A multi-dimensional fusion feature matrix is constructed by combining channel splicing and element-by-element addition. This enables the coordinated analysis of local details and global structure, significantly improving the sensitivity and specificity of osteoporosis area detection.

[0100] S5. Cascade classification and clinical decision making

[0101] Input: fused multimodal features;

[0102] Output: Clinical report recording T value, regional coordinates, risk level results and corresponding confidence score data, probability heat map with confidence score annotated, and PDF / graphic report;

[0103] 1. First-level classifier: The system uses a support vector machine to initially screen candidate areas, selecting high-probability osteoporosis areas through weighted confidence scoring. Based on the multimodal feature matrix output by the feature fusion module, the support vector machine uses kernel function mapping and hyperplane partitioning to quickly eliminate low-confidence interference areas, such as calcification points or artifacts, significantly improving subsequent processing efficiency.

[0104] 2. Second-level classifier: A lightweight residual network is used to refine the candidate areas after the initial screening. Combined with the World Health Organization's bone density T-score grading standard, the network quantitatively analyzes the degree of bone microstructure degeneration. The network introduces a channel attention mechanism to optimize feature selection capabilities. At the same time, it integrates clinical data such as patient age and gender to achieve osteoporosis degree grading based on multi-dimensional information.

[0105] 3. Clinical test result output: The system generates a quantitative bone density T-value report that meets the World Health Organization's standards. It also generates a personalized risk assessment graphic report in PDF and DICOM formats based on the patient's age and gender, marking high-risk areas, namely those at risk of vertebral compression fractures. The output includes structured data, namely T-values, regional coordinates, risk levels, and visual heat maps, enabling doctors to quickly locate lesions and formulate treatment plans.

[0106] S6. Real-time Reasoning and Dynamic Feedback

[0107] Input: classification results and confidence scores;

[0108] Output: real-time detection results and interactive interface;

[0109] 1. Embedded deployment: The system compresses the model through knowledge distillation and quantization pruning technology, adapts to embedded device deployment, achieves efficient inference speed, and supports real-time processing of multimodal images based on lightweight models and CUDA acceleration optimization, meeting the clinical scenario requirements of low latency and high throughput.

[0110] 2. Dynamic parameter optimization: Dynamically adjust detection parameters based on the confidence score output by the classifier: When the confidence score is lower than the set threshold, the sliding window step size is reduced to improve detection accuracy. When the confidence score is higher than the set threshold, the Gabor filter kernel size is increased to accelerate feature extraction. Through a real-time feedback mechanism, a dynamic balance between detection efficiency and accuracy is achieved, reducing redundant calculations.

[0111] 3. Visual Interaction: Mark osteoporosis areas (red high-probability areas) in the probability heat map and embed confidence scores into the original medical images through the DICOM standard interface, supporting interactive verification by doctors. Combined with multi-view overlay display (coronal and sagittal planes) and dynamic threshold adjustment functions, it assists clinical decision-making and improves diagnostic interpretability.

[0112] S7. Closed-loop model optimization

[0113] Input: false positives and missed positives annotated by doctors;

[0114] Output: iteratively optimized model version;

[0115] 1. Unsupervised Data Enhancement: The system uses a recurrent generative adversarial network framework for unsupervised data generation. Through adversarial training strategies, it synthesizes osteoporosis images covering different age groups, genders, and lesion severity. This significantly increases the amount of data and effectively alleviates the scarcity of medical image annotations. The generation process introduces pathological feature constraints, namely the distribution of trabecular density, to ensure that the synthesized data maintains a high degree of consistency with real cases in terms of microstructure and macromorphology, providing diverse sample support for model training.

[0116] 2. Transfer learning fine-tuning: Based on a pre-trained deep network model, domain-adaptive fine-tuning is performed using minimal annotated data. The underlying parameters of the backbone network, namely the edge detection filter, are frozen, and only the top-level classifier and osteoporosis-related feature layer are optimized. This is combined with focal loss to address class imbalance, enabling the model to achieve extremely high accuracy even with minimal annotated data, significantly reducing the cost of clinical data annotation.

[0117] 3. Model Compression and Update: Knowledge distillation technology is used to transfer the knowledge of the teacher model to a lightweight student model. The teacher model's soft labels, or probability distributions, are used to guide student model training, preserving the ability to discriminate osteoporosis microstructural features. Combining quantization-aware training with channel pruning, the model is compressed to achieve a balance between inference speed and accuracy on embedded devices.

[0118] 4. Dynamic hot update: Wireless update technology is used to achieve seamless hot update of the model: the optimized model is transmitted to the terminal device, namely the hospital imaging workstation, through encrypted differential packets, completing version iteration while ensuring service continuity. The update process supports incremental upgrade and rollback mechanisms to ensure that the clinical diagnosis system continues to achieve performance improvements, that is, the false detection rate is reduced without affecting real-time reasoning efficiency.

[0119] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0120] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An intelligent identification system for osteoporosis areas, characterized in that: include: Image acquisition and preprocessing module, candidate region positioning and optimization module, multimodal feature extraction module, feature fusion and weighting module, classification decision and clinical report module, data enhancement and model optimization module and real-time reasoning and feedback module, the image acquisition and preprocessing module includes a multimodal data interface, a resolution calibration unit, a contrast optimization unit, a noise suppression unit and a preprocessing unit, the candidate region positioning and optimization module includes a dynamic adaptive sliding window unit, a region growing algorithm unit and a dynamic threshold segmentation unit, the feature fusion and weighting module includes a traditional feature unit and a deep learning feature unit, the traditional feature unit includes a grayscale statistics sub unit, texture analysis subunit and shape description subunit, the deep learning feature unit includes a local branch subunit, a global branch subunit and a topology modeling subunit, the multimodal feature extraction module includes a channel-space attention unit, a compression excitation unit and a multi-scale fusion unit, the classification decision and clinical report module includes a cascade classifier unit, a personalized report generation unit and a visualization output unit, the data enhancement and model optimization module includes an unsupervised data generation unit, a transfer learning unit, a model compression unit and a dynamic hot update unit, and the real-time reasoning and feedback module includes an embedded deployment unit, a dynamic parameter adjustment unit and a visualization interaction unit.

2. The intelligent osteoporosis area identification system according to claim 1, characterized in that: The image acquisition and preprocessing module supports multimodal medical image input and standardized processing. The multimodal data interface supports fully automatic access to CT, MRI, and X-ray data. The resolution calibration unit uses a bicubic interpolation algorithm to calibrate the image resolution to a standard pixel density. The contrast optimization unit dynamically adjusts the contrast through adaptive histogram equalization. The noise suppression unit removes Gaussian noise and equipment artifacts based on non-local mean filtering. The morphological preprocessing unit combines morphological opening operations to eliminate interference from non-skeletal structures such as blood vessels and soft tissue.

3. The intelligent osteoporosis area identification system according to claim 1, characterized in that: The candidate region positioning and optimization module efficiently screens and optimizes suspected osteoporosis areas, the dynamic adaptive sliding window unit is based on prior knowledge of bone anatomy, the region growing algorithm unit is based on four-neighborhood and eight-neighborhood connectivity analysis, and the dynamic threshold segmentation unit adaptively selects segmentation strategies based on local grayscale variance differences.

4. The intelligent osteoporosis area identification system according to claim 1, characterized in that: The multimodal feature extraction module extracts traditional and deep learning features in parallel. The traditional feature unit includes a grayscale statistics subunit, a texture analysis subunit and a shape description subunit. The grayscale statistics subunit calculates the pixel mean, variance and grayscale distribution histogram of the candidate area. The texture analysis subunit analyzes the multi-directional texture of trabeculae through a Gabor filter. The shape description subunit extracts the bone contour curvature and connectivity features based on Sobel edge detection and morphological operations, and constructs geometric morphological parameters. The deep learning feature unit includes a local branch subunit, a global branch subunit and a topology modeling subunit. The local branch subunit extracts the microstructural texture features of trabecular porosity and arrangement regularity through a 3D convolutional network. The global branch subunit embeds a compressed incentive attention mechanism into the improved MobileNetV3 backbone network. The topology modeling subunit uses a graph convolutional network to model the spatial topological relationship of key bones.

5. The intelligent osteoporosis area identification system according to claim 4, characterized in that: The feature fusion and weighting module dynamically weights and fuses multimodal features to suppress noise interference. The channel-spatial attention unit allocates spatial weights to local feature maps. The compression excitation unit dynamically calibrates the channel dimension of global features. The multi-scale fusion unit fuses the local microstructure features of 3D convolution with the global topological features of graph convolution through channel splicing and element-by-element addition strategies.

6. The osteoporosis area intelligent identification system according to claim 1, characterized in that: The classification decision and clinical reporting module finely identifies osteoporosis areas and generates clinical reports. The first-level classifier support vector machine of the cascade classifier unit quickly eliminates low-confidence interference areas through kernel function mapping and hyperplane partitioning. The second-level classifier quantitatively assesses the degree of bone degeneration through a lightweight residual network combined with the WHO bone density T-value standard. The personalized report generation unit generates a PDF / graphic report based on the patient's age and gender. The visual output unit generates a probability heat map and embeds it into a DICOM image.

7. The intelligent osteoporosis area identification system according to claim 1, characterized in that: The data enhancement and model optimization module solves the problems of data scarcity and model lightweighting. The unsupervised data generation unit synthesizes osteoporosis images of multiple age groups and multiple disease degrees based on CycleGAN. The transfer learning unit uses the ImageNet pre-trained model to fine-tune the classifier, greatly reducing the demand for labeled data. The model compression unit compresses the model through knowledge distillation and quantitative pruning. The dynamic hot update unit realizes seamless iterative update of the model through OTA technology.

8. The osteoporosis area intelligent identification system according to claim 1, characterized in that: The real-time reasoning and feedback module realizes low-latency reasoning and dynamic parameter optimization. The embedded deployment unit deploys the compression model on the Jetson Nano platform. The dynamic parameter adjustment unit dynamically adjusts the sliding window step size and Gabor filter kernel size according to the confidence score. The visualization interaction unit generates a probability heat map and superimposes it on the original image.

9. The intelligent osteoporosis area identification system according to claim 1, characterized in that: The system workflow is as follows: S1. Multimodal Image Input and Preprocessing Input: CT, MRI, X-ray medical images; Output: normalized preprocessed image; 1. Multimodal data access: automatic recognition and analysis of CT, MRI, and X-ray imaging data; 2. Standardization preprocessing: unify image resolution to standard pixel density, and optimize image contrast through adaptive histogram equalization; 3. Interference suppression: remove Gaussian noise and equipment artifacts, and eliminate interference from non-skeletal structures such as blood vessels and soft tissues; S2. Dynamic positioning of candidate regions Input: preprocessed normalized image; Output: optimized candidate region coordinates and image slices; 1. Dynamic sliding window initial screening: Prioritize detection of high-incidence areas such as the spine and femoral neck, and dynamically adjust the window size and step length based on the image signal-to-noise ratio; 2. Region growing algorithm optimization: merge spatially adjacent candidate regions to reduce redundant calculations; 3. Dynamic threshold segmentation: Optimize the segmentation threshold based on local grayscale variance to exclude low-contrast interference areas; S3. Multimodal feature extraction Input: optimized candidate region coordinates and image slices; Output: traditional feature vector and deep learning feature vector; 1. Traditional feature extraction: Grayscale statistics: calculate the mean, variance and grayscale distribution histogram of the candidate region; Texture analysis: Gabor filtering and LBP algorithm are used to extract trabecular texture; Shape description: Sobel edge detection and morphological operations are used to calculate the curvature of the bone contour; 2. Deep learning feature extraction: Local branch: 3D convolutional network extracts trabecular microstructure texture, global branch: MobileNetV3 extracts global bone features, and graph convolutional network models the topological relationship of key bones; S4. Feature fusion and noise suppression Input: traditional feature vector and deep learning feature vector; Output: fused multimodal feature matrix; 1. Channel-spatial attention weighting: Assign spatial weights to local features of 3D convolution output to enhance key area responses; 2. SE module enhances global features: dynamically weights the channel features output by MobileNetV3 to suppress irrelevant background; 3. Multi-scale fusion strategy: cross-scale splicing of local texture features and global topological features; S5. Cascade classification and clinical decision making Input: fused multimodal features; Output: Clinical report recording T value, regional coordinates, risk level results and corresponding confidence score data, probability heat map with confidence score annotated, and PDF / graphic report; 1. First-level classifier: Screening high-confidence candidate regions based on SVM; 2. Second-level classifier: Use lightweight ResNet to further determine the degree of osteoporosis; 3. Output clinically applicable test results: WHO standard grading: output bone density T value to assess the degree of osteoporosis; personalized risk assessment: generate PDF / graphic report based on patient age and gender, marking high-risk areas; S6. Real-time Reasoning and Dynamic Feedback Input: classification results and confidence scores; Output: real-time detection results and interactive interface; 1. Embedded deployment: compress the model to adapt to embedded devices; 2. Dynamic parameter optimization: Automatically adjust the sliding window step size and Gabor filter kernel scale based on the confidence score; 3. Visual interaction: Mark the probability of osteoporosis areas in the heat map and embed confidence scores in DICOM images; S7. Closed-loop model optimization Run in the background Input: false positive / missed positive samples annotated by doctors; Output: iteratively optimized model version; 1. Unsupervised data augmentation: Use CycleGAN to generate synthetic images of multiple age groups and disease severity; 2. Transfer learning fine-tuning: Based on the pre-trained ImageNet model, fine-tune the classifier using a small amount of labeled data; 3. Model compression and update: Through knowledge distillation, the teacher model ResNet50 is compressed into the student model MobileNetV3; 4. Dynamic hot update: Deploy the optimized model to the terminal device via OTA.

Images