Explosion injury condition assessment and image data analysis system based on deep learning

Abstract

The application discloses an explosion injury condition assessment and image data analysis system based on deep learning, and relates to the technical field of image data processing.The system comprises a body surface image analysis module, a feature extraction and injury analysis assessment module and an injury analysis assessment and feedback module.The application extracts body surface injury features by processing the body surface images of each target, obtains corresponding emergency early warning determination results according to the body surface injury features, then dynamically assesses the injury conditions of all targets in the order of emergency grades, selects an injury analysis strategy based on the image features of each target to improve the injury condition assessment precision and efficiency, finally performs injury condition assessment based on the selected injury analysis strategy, obtains the injury parameters of the corresponding target, obtains the injury condition state determination results based on the physiological parameters and the injury parameters of each target, and performs feedback, thereby solving the problem of low effectiveness of explosion injury condition assessment in the prior art.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, and in particular to a deep learning-based system for assessing blast injuries and analyzing image data. Background Technology

[0002] Blast injuries, a complex type of trauma caused by the instantaneous release of enormous energy from explosives, require rapid and accurate assessment. This assessment is crucial for determining the success rate of patient treatment, optimizing medical resource allocation, and enhancing disaster emergency response capabilities. With the development of medical imaging technology, computed tomography (CT) and ultrasound have provided objective evidence for the diagnosis of blast injuries. These techniques non-invasively acquire visual information about the internal structures of the human body, providing doctors with intuitive information for assessing the injury. Deep learning technology, especially convolutional neural networks, has made groundbreaking progress in medical image analysis. Utilizing deep learning models to automatically analyze medical images of blast injury patients can assist or even partially replace manual interpretation, significantly improving the speed and objectivity of assessment.

[0003] In existing technologies, drones equipped with tri-light pods (visible light, infrared, and laser rangefinders) are used to collect on-site image data in order to quickly obtain macroscopic injury information. Simultaneously, to supplement and validate data, some studies design warhead explosion experiments in laboratory settings to collect fragmentation images and use image analysis software to digitize the shape and size of the fragments. In the clinical setting, existing technologies typically integrate patient imaging data (such as CT scans and ultrasound), physiological parameters, and clinical text records to construct a multi-source database. Then, convolutional neural networks, such as ResNet or U-Net, are used to process the image data for segmenting the injury area and classifying injury severity. This is combined with recurrent neural networks to analyze time-series data (such as changes in physiological parameters) to predict potential injury areas and severity.

[0004] For example, Chinese patent application CN118229708A discloses a multi-task image segmentation method and system for burn injury assessment, which includes: inputting a standardized burn image into a convolutional neural network segmentation model based on multi-task learning; outputting segmentation results for different parts of the human body and burn wounds of different grades through the convolutional neural network segmentation model; extracting hierarchical feature representations from the original RGB burn image through a shared encoder; completing the segmentation tasks of burn wounds and human body parts through two decoupled task branches respectively; in the burn injury assessment system, combining the segmentation results of human body parts and burn wounds, the severity of the wound can be graded and its location can be located; combining the patient's age, gender, and other information, the burn area can be further assessed.

[0005] For example, Chinese patent application CN111696085B discloses a method and device for rapid on-site ultrasound assessment of lung impact injury, comprising: sequentially scanning each lung region with a handheld ultrasound probe to acquire ultrasound image video streams and transmitting them to a host computer; the host computer inputs the acquired ultrasound image video streams into an artificial intelligence model, which dynamically estimates the score based on the input ultrasound image video streams and outputs the maximum score value for each lung region after scanning; and the maximum score values ​​for each lung region are added together to calculate the total lung ultrasound score.

[0006] The above-mentioned technology has at least the following technical problems: Blast injuries can present as damage areas of varying scales. For example, fractures may appear as fragmented pieces or cracks, while burns may be extensive, localized areas of redness, swelling, or charring. Different types of injuries (fracture fragments, soft tissue burns, hematomas, etc.) exhibit multi-scale heterogeneous features on images, such as high-density punctate bone fragments or large areas of low-density necrosis. Existing image aggregation analyses typically use global averaging or total volume calculations, which fail to preserve local spatial variation information at the image level.

[0007] The distribution of blast injuries is often extremely uneven. Existing technologies typically perform global aggregation analysis on multiple injuries identified in medical images (such as CT scans), calculating factors like the total volume, average density, or overall score of the injured area, ultimately outputting a single quantitative indicator representing the overall injury severity. This fails to reflect the spatial heterogeneity of the injury. Moreover, over time, CT or ultrasound images show changes in density, blurred boundaries, and expansion of edema areas. Traditional assessment models only use images from a single time point and lack modeling of the evolution of time-series images.

[0008] Meanwhile, blast injuries are a dynamic pathological process. For example, secondary bleeding and progressive edema can continue to evolve for hours or even days after the injury, making it impossible to capture the dynamic development trend of the injury.

[0009] In addition, images from the explosion site may be affected by factors such as smoke, dust, and fire, making the images more blurry or noisy. Medical imaging equipment will also experience a decrease in signal-to-noise ratio in high interference environments, further affecting injury assessment and resulting in low effectiveness of explosion injury assessment. Summary of the Invention

[0010] To address the low effectiveness of existing technologies in assessing blast injuries, this invention provides a deep learning-based system for assessing blast injuries and analyzing image data. The specific technical solution is as follows:

[0011] A deep learning-based system for assessing blast injuries and analyzing image data includes: a body surface image analysis module, a feature extraction and damage analysis assessment module, and a damage analysis assessment and feedback module. The body surface image analysis module is used to: acquire body surface images of all targets through sensors mounted on the UAV, process the body surface images of each target based on the distribution of pixel values ​​in the body surface images to extract body surface damage features reflecting blast injuries in the body surface images, and obtain emergency warning judgment results based on the body surface damage features. The feature extraction and damage analysis assessment module is used to: if the received emergency warning judgment result is that an emergency warning has been triggered, then the target that triggered the emergency warning will be given priority for dynamic injury assessment; otherwise, all targets will be dynamically assessed for injury according to the difference in the extracted surface damage features. Based on the image features of each target, it will be determined whether to select global damage analysis to improve the accuracy and efficiency of injury assessment. The damage analysis assessment and feedback module is used to: if so, obtain the injury status judgment result based on the acquired damage area of ​​each target and provide feedback; otherwise, assess the injury of each target based on the damage analysis strategy of multi-scale convolutional neural network, obtain the corresponding damage parameters of the target, obtain the injury status judgment result based on the acquired physiological parameters and damage parameters of each target, and provide feedback.

[0012] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. By acquiring surface images of all targets using sensors mounted on drones, the high mobility and wide field of view of drones can quickly cover the explosion site, solving the problem that rescuers cannot immediately contact all targets during an explosion. This enables wide-area, rapid, and contactless initial information collection, laying the data foundation for subsequent batch processing and rapid triage. The extracted surface injury features are assigned corresponding emergency levels, achieving automated and standardized on-site triage. This allows for timely prioritization of a large number of targets, identifying individuals most in need of emergency rescue, greatly improving rescue efficiency and resource utilization. Then, the injuries of all targets are dynamically assessed according to the order of emergency level, enabling dynamic and intelligent allocation of computing resources. By selecting damage analysis strategies based on the image features of each target, an intelligent balance between assessment accuracy and efficiency is achieved. This avoids the waste of resources caused by directly using the most complex but time-consuming model for simple image analysis, and also avoids the problem of insufficient accuracy caused by using a simple model to analyze complex images. Finally, the selected damage analysis strategy is used to analyze the image to assess the injury, which improves the accuracy of image analysis and thus makes the injury assessment more accurate. The injury status judgment result is obtained based on the physiological and damage parameters of each target. The overall physiological status of the target is comprehensively judged, realizing the fusion assessment of multimodal information. This enables a more comprehensive and accurate judgment of the target's true state, ensuring that abnormal situations can be detected and dealt with in the first time.

[0013] 2. By segmenting local windows and calculating standard deviation, local variations in the image, such as edges and textures, can be highlighted, which helps to more accurately identify and analyze damaged areas in body surface images. Image normalization of the body surface image eliminates differences between different local windows caused by factors such as brightness and contrast, improving the accuracy and reliability of image analysis. Then, by comparing the average variance with the reference average variance, it is possible to accurately determine whether there is a lot of noise in the image, thereby selecting targeted denoising processing or scale consistency judgment, avoiding unnecessary processing steps, and improving processing efficiency and accuracy. Compared with the fixed process of processing all images in the existing technology, it improves the accuracy of image analysis and can flexibly adjust the processing path according to the actual quality of the image, achieving optimal resource allocation. Finally, through a denoising autoencoder network, image noise is effectively removed. This method effectively removes various types of noise, improves the signal-to-noise ratio of the image, and makes the image clearer, facilitating accurate identification and analysis of the damaged area. Adaptive histogram equalization adjusts the brightness distribution of the image, enhances the contrast, and highlights details such as the damaged area, avoiding overexposure or underexposure problems that may occur with global equalization, thus further improving image quality. Finally, by calculating the scale consistency index, the consistency of the damaged area in the body surface image can be quantified, providing a basis for selecting an appropriate feature extraction method. It can take into account feature information at different scales, and the extracted features are more comprehensive and accurate, providing a reliable feature basis for confirming the emergency level. Compared with existing technologies that rely on fixed scales or templates for feature extraction, this method improves the objectivity and adaptability of the entire image processing workflow, and can extract hierarchical features at different scales, adapting to damage of different sizes.

[0014] 3. By calculating the damage area of ​​each damaged region in the image, the density difference of the local area is obtained to quantify the difference in damage severity between different damaged areas. This provides a key basis for selecting appropriate analysis strategies. Compared with existing technologies that cannot adaptively adjust the analysis strategy according to the heterogeneity characteristics of the damaged area, this method solves the problem of poor adaptability to complex damage patterns. Then, the calculated density difference of the local area is compared with the density difference of the reference local area to select the analysis strategy in a targeted manner, thereby improving the effectiveness of image analysis and thus improving the accuracy and efficiency of injury assessment. When the density difference of the local area is large, spatial heterogeneity analysis can be used to deeply explore local and global information. Otherwise, global damage analysis can quickly obtain the overall injury assessment result. Furthermore, through spatial heterogeneity analysis, local details and global structure of the image can be captured simultaneously to gain a more comprehensive understanding of the injury and avoid assessment bias caused by focusing only on local or global image analysis results. This allows for a more accurate judgment of the severity and development trend of the injury. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the structure of the deep learning-based explosion injury assessment and image data analysis system provided in an embodiment of the present invention; Figure 2 This is a flowchart of image processing provided in an embodiment of the present invention; Figure 3 This is a flowchart of the damage analysis strategy selection provided in the embodiments of the present invention; Figure 4 This is a model architecture diagram of the image feature extraction model provided in the embodiments of the present invention; Figure 5 This is a flowchart of segmentation accuracy determination provided in an embodiment of the present invention. Detailed Implementation

[0017] This invention provides a deep learning-based system for assessing and analyzing blast injuries and image data, addressing the problem of low effectiveness in existing blast injury assessment technologies. The system processes surface images of each target to extract surface damage features, obtains corresponding emergency warning judgments based on these features, and then dynamically assesses the injuries of all targets according to their emergency levels. Damage analysis strategies are selected based on the image features of each target to improve the accuracy and efficiency of injury assessment. Finally, injury assessment is performed based on the selected damage analysis strategies to obtain the corresponding damage parameters. Injury status judgments are obtained based on the physiological and damage parameters of each target, and feedback is provided, thus improving the effectiveness of blast injury assessment.

[0018] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0019] This invention provides a deep learning-based system for assessing blast injuries and analyzing image data. For example... Figure 1 As shown, the system includes: a body surface image analysis module, a feature extraction and damage analysis and assessment module, and a damage analysis, assessment, and feedback module.

[0020] The body surface image analysis module acquires body surface images of all targets using sensors mounted on the UAV (such as high-definition cameras and infrared thermal imagers), overcoming the difficulties of manual data collection and improving the efficiency and completeness of information acquisition. It processes the body surface images of each target to extract surface injury features reflecting blast injuries. Compared to traditional manual identification methods, this significantly improves the accuracy and efficiency of identification, reduces human error, and provides an emergency warning judgment based on the body surface injury features. This provides a scientific basis for the subsequent target processing order, ensuring that the most severely injured targets receive priority treatment, thus improving the targeting and effectiveness of rescue efforts. Body surface injury features include bleeding characteristics (including the color, area, and shape of the bleeding area), burn characteristics (including the color, area, and depth of the burn area), and limb deformity characteristics (including morphological changes such as abnormal limb angles, shortened lengths, and joint dislocations, which can be quantified through key point detection and skeletal analysis). Red areas are extracted using a color space (such as HSV), and segmentation models (such as U-Net and Mask) are used. R-CNN extracts bleeding features; burn areas are identified through image color (e.g., red, black) and texture features (e.g., charred, bubbly); limb shapes are identified through pose estimation models (e.g., OpenPose) combined with morphological analysis (e.g., edge detection, contour analysis); and emergency levels are used to measure the severity of the target's injury. For example, they can be divided into different levels such as Level 1 (most severe), Level 2, and Level 3, to determine the order of target treatment.

[0021] In the feature extraction and damage analysis assessment module, if the received emergency warning judgment result indicates that an emergency warning has been triggered, priority is given to dynamically assessing the injuries of the targets that have triggered the emergency warning. Otherwise, all targets are dynamically assessed according to their emergency level to accurately determine the emergency warning situation, rationally arrange the order of dynamic injury assessment, prioritize the handling of emergency targets, optimize the allocation of rescue resources, and improve overall rescue efficiency. After analyzing the body surface images, the targets are sent to the preset locations according to their emergency level, while the image features of the injured are acquired. Damage analysis strategies are selected based on the image features of each target, which can more accurately analyze the target injuries, avoid unnecessary complex analysis, and improve the accuracy and efficiency of injury assessment. Image features refer to the features extracted from medical images (such as CT scans) that can reflect the internal structure and nature of the injury.

[0022] The damage analysis, assessment, and feedback module assesses the injuries of each target based on the selected damage analysis strategy, obtaining more accurate damage parameters and providing quantitative indicators for a comprehensive assessment of target injuries. Based on the acquired physiological and damage parameters of each target, the module determines the injury status. The physiological parameters reflect the real-time physiological state of the target's body. Combined with the damage parameters, the module can more comprehensively and accurately determine the target's injury status, promptly detect changes in target injuries, improve the effectiveness of target blast injury assessment, and provide feedback for emergency treatment and resource allocation.

[0023] like Figure 2 The image processing flowchart shown further illustrates the processing of the surface images of each target. The specific process is as follows: Image normalization is performed on the body surface image to obtain the processed body surface image. Image normalization means converting body surface images of different specifications (such as size, color range, etc.) into a unified standard form. For example, the image size is adjusted to a fixed size, and the pixel value range is unified to a specific interval (such as 0-255 normalized to 0-1). This allows subsequent processing to be based on a unified data foundation, reduces interference caused by differences in image specifications, and improves the processing efficiency of the entire image analysis process.

[0024] If the average variance of the processed body surface image of each target is greater than the average variance of the reference variance, then denoising processing is performed followed by scale consistency determination; otherwise, scale consistency determination is performed directly. The average variance of the processed body surface image represents the average variance of all local regions in the processed body surface image. The average variance, reference scale consistency index, and preset scale are set by preset personnel. By comparing the variance with the reference variance, it is possible to accurately determine whether there is noise in the image, thereby selectively choosing denoising processing or scale consistency determination, avoiding unnecessary processing steps, and improving processing efficiency and accuracy.

[0025] The denoising process specifically involves: employing a denoising autoencoder network for image denoising and adjusting brightness using adaptive histogram equalization. The denoising autoencoder network is a deep learning model consisting of an encoder and a decoder. The encoder compresses and encodes the input noisy body surface image, extracting its key features; the decoder then reconstructs an output close to the original noise-free body surface image based on the encoded features. Through continuous training and optimization, the goal of removing image noise is achieved, thereby improving the signal-to-noise ratio of the body surface image, making it clearer, and facilitating accurate identification and analysis of damaged areas. Adaptive histogram equalization is an image enhancement technique that adaptively adjusts the image contrast based on the histogram distribution of local image regions. Unlike ordinary histogram equalization, it better preserves local details of the body surface image, making the brightness distribution more uniform, enhancing the visual effect of the body surface image, and further improving image quality and analyzability.

[0026] Scale consistency determination is specifically as follows: The scale consistency index is obtained based on the area of ​​the damaged region in the processed body surface image (if no denoising is performed, the image is the processed body surface image; otherwise, the image is the denoised body surface image). In the formula, s represents the standard deviation of the area of ​​the damaged region in the processed body surface image, and the scale consistency index is used to quantify the consistency of the area of ​​the damaged region in the body surface image, providing a basis for selecting a suitable feature extraction method.

[0027] If the scale consistency index is greater than the reference scale consistency index, it indicates that the consistency of the damaged area is better. The body surface image is processed using a preset scale to extract the body surface damage features. Otherwise, the corresponding body surface damage features are extracted by combining the pyramid pooling structure and the convolution kernel size corresponding to each damaged area. The pyramid pooling structure is a pooling method used in deep learning to process features at different scales. The input image or feature map is divided according to different scales to form a pyramid-like structure. Then, pooling operations (such as max pooling, average pooling, etc.) are performed at each scale. Finally, the pooling results at different scales are fused to extract feature information at different scales.

[0028] The kernel size represents the result obtained by mapping the area of ​​each damaged region into a kernel size mapping set. The kernel size mapping set is a collection obtained from a pre-set database that represents the mapping relationship between the area of ​​each damaged region and the corresponding kernel size. The kernel size mapping set is trained using kernel size training data, which includes the area of ​​each damaged region based on historical data for a historical time period, as well as kernel sizes set by professional technicians based on empirical rules. By using a pyramid pooling structure combined with the kernel size mapping set to extract surface damage features, the feature information at different scales can be taken into account, and the extracted features are more comprehensive and accurate, improving the effectiveness and robustness of blast injury assessment.

[0029] Furthermore, image normalization processing is performed on the body surface image. The specific process is as follows: The body surface image is divided according to the preset sliding window and preset step size set by the preset personnel to obtain the local window image. By dividing the image into local windows, local analysis of the image can be performed, focusing on the local features of the image rather than the overall features. This is very useful for processing body surface images of targets with blast injuries, and can capture local information more accurately.

[0030] The average value of all pixels within each local window is obtained by averaging the values ​​of all pixels in the corresponding local window image. In other words, the sum of all pixel values ​​within a local window divided by the number of pixels in that window reflects the overall brightness level of the pixels in that window and can eliminate brightness fluctuations in local areas.

[0031] The standard deviation of pixels in each local window is calculated by processing the standard deviation of all pixel values ​​and the corresponding average pixel value. This results in the pixel standard deviation of the corresponding local window image. Specifically, the difference between each pixel value and the average pixel value is calculated first, then the squares of these differences are summed, divided by the number of pixels, and finally the square root is taken. This reflects the dispersion of pixel value distribution within the window and can help identify the changes in pixel values ​​in local areas. It is of great significance for detecting features such as edges and textures in images. For example, in damaged areas, the pixel standard deviation may be large, reflecting the difference in color or texture between the lesion area and normal skin.

[0032] The regularization factor, pixel values, pixel average, and pixel standard deviation of each local window image are normalized to obtain normalized pixel values. The normalized pixel values ​​within each local window replace the corresponding pixel values ​​to obtain a normalized body surface image, thus eliminating local illumination unevenness. The regularization factor is a very small positive number set by the predefined personnel to avoid zero denominators, ensuring smooth normalization calculations. Normalization adjusts the original pixel values ​​to a relatively uniform range, eliminating differences between different local windows caused by factors such as brightness and contrast. This results in better comparability and stability of the body surface images, facilitating subsequent image analysis and processing, and ultimately improving the accuracy of blast injury image analysis.

[0033] The normalized body surface image is input into a guided filter for filtering to obtain a processed body surface image that achieves smooth lighting and preserves details. The body surface image is used as the guide image, and the normalized body surface image is used as the input image. The guided filter learns the edge structure of the guide image and uses it as a basis to filter the input image, so that the output image closely resembles the normalized image in flat areas, but strictly aligns with the edges of the original image at the edges. Furthermore, the guided filtering has a fast linear time complexity.

[0034] The specific constraint expression for normalized pixel values ​​is: ; In the formula, x represents the x-coordinate of a pixel in the body surface image, y represents the y-coordinate of a pixel in the body surface image, and I adaptive (x,y) represents the normalized pixel value within the window centered at (x,y), I(x,y) represents the pixel value within the window centered at (x,y), µ(x,y) represents the average pixel value within the window centered at (x,y), and σ(x,y) represents the standard deviation of the pixel value within the window centered at (x,y). This represents the regularization factor.

[0035] Furthermore, based on the urgency level, an emergency warning determination is obtained. The specific process is as follows: The characteristics of surface injuries of each target are fed back to the pre-selected personnel, providing them with detailed information support. The pre-selected personnel then confirm the emergency level, providing a scientific basis for subsequent rescue and treatment. If there is a target with an emergency level no less than the reference emergency level set by the pre-selected personnel, an emergency warning is triggered, prompting the pre-selected personnel to prepare resources in advance. The target that triggered the emergency warning is set as a priority injury assessment target, so that the dynamic injury assessment of the target that triggered the emergency warning is prioritized. Resources include, but are not limited to, medical equipment (such as ventilators, defibrillators, surgical instruments, etc.), medicines (such as hemostatic drugs, antibiotics, anesthetics, etc.), medical personnel (arranging a sufficient number of doctors and nurses with professional skills), beds (reserving emergency beds, intensive care beds, etc.), and logistical support resources (such as food, water, hygiene products, etc.). Dynamic injury assessment refers to the continuous and real-time observation and assessment of the target's injury to understand the changes in the injury in a timely manner. By determining the emergency level, emergency targets that require special attention can be quickly identified.

[0036] If there are no targets with an emergency level no lower than the reference emergency level, then all targets are sorted in descending order according to their emergency level and their injuries are dynamically assessed in sequence.

[0037] By following the steps above, medical resources can be allocated rationally, injury assessment can be carried out in an orderly manner, the efficiency of injury assessment can be improved, and timely and accurate assessment of the target injury can be ensured.

[0038] like Figure 3 The flowchart shown further illustrates that the damage analysis strategy is selected based on the image features of each target. The specific process is as follows: The standard deviation of the damage area in each damaged region in the image is processed to obtain the local area density difference. The local area density difference is used to assess the fluctuation of the damage degree in each damaged region, providing a key basis for selecting an appropriate analysis strategy. By processing the standard deviation, the dispersion of the area of ​​each damaged region is transformed into a comparable value, making the injury assessment more scientific and objective.

[0039] If the density difference in a local area is greater than the reference density difference in a local area set by preset personnel, spatial heterogeneity analysis is used for injury assessment; otherwise, global damage analysis is used for injury assessment. This allows for targeted selection of analysis strategies, improving the accuracy and efficiency of injury assessment.

[0040] Spatial heterogeneity analysis utilizes multi-scale convolutional neural networks (CNNs) to extract features from images layer by layer, ensuring the simultaneous capture of local details and global structure. Multi-scale CNNs employ convolutional layers of varying levels, each capable of extracting features at different scales. Lower-level convolutional layers capture local details such as edges and textures of damaged areas, while higher-level layers extract more abstract global structural information, such as the overall distribution patterns of multiple damaged areas. Through layer-by-layer feature extraction, the network can comprehensively utilize this information at different scales to more accurately assess injury severity, avoiding assessment biases caused by focusing only on local or global information.

[0041] Global damage analysis, specifically: By statistically analyzing the damage area of ​​all affected areas and summing the damage areas of all affected areas, or by using other appropriate statistical methods, a global quantitative indicator can be obtained to quantify the target's injury status. For example, it can be expressed as the proportion of the total damage area to the area of ​​the relevant body parts. This facilitates a rapid understanding of the target's injury severity and provides important reference for developing treatment plans and assessing prognosis. For instance, in emergency situations, the global quantitative indicator can be used to quickly determine whether the target needs immediate surgery or other critical treatment measures.

[0042] Specifically, such as Figure 4 The image feature extraction model architecture shown includes: an input layer, downsampling (encoder), a bottleneck layer (intermediate layer), upsampling (decoder), skip connections, and an output layer. In the input layer, the input is an image patch (medical image) with a size of 572×572 (or 570×570 or 568×568). These different sizes represent cropping or scaling operations that may be performed on the image during preprocessing. In downsampling, starting from the input image, it passes through a series of convolutional layers (blue boxes), each with a 3×3 kernel size, using the ReLU activation function to perform convolution operations. The number of output channels in each layer gradually increases (e.g., 64 at the input, reaching 256 or 512 after several convolutions), while the spatial dimension is reduced (from 572×572 to 140×140) through max pooling (downward black arrow). Each pooling operation halves the image size. In the bottleneck layer... Figure 4At the bottom layer, the spatial size of the image is reduced to its minimum, while the number of channels reaches its maximum (1024). This bottleneck layer represents the highly compressed feature representation of the input image by the network. In upsampling, the upper part of the network is the decoder, which performs spatial upsampling through upconvolution (black arrow pointing upwards), gradually restoring the spatial resolution of the image. The upconvolution operation uses a 2×2 convolution kernel to expand the spatial size of the image, and with each upsampling layer, the spatial size of the image is gradually restored to a size close to the input. In skip connections, skip connections copy and crop (middle gray arrow) the preceding layers and concatenate them with the decoder layers. This helps the network retain more detailed features while restoring spatial information. For example, as can be seen in the figure, some layers of the decoder (e.g., 128 and 256 channels) are directly connected to the corresponding layers of the encoder. In the output layer, the final output is a segmentation result of an image (labeled image), with a size of 392×392 or 388×388, containing the probability of each pixel belonging to different categories; the output layer has a convolution kernel size of 1×1, and this step compresses the feature map to the final number of segmentation categories (e.g., 2 or more categories).

[0043] Spatial heterogeneity analysis, the specific process is as follows: The target image is input into an image feature extraction model that extracts features at different resolutions to obtain feature maps containing local details and global structure. This is achieved using a U-Net network with a preset number of downsampling modules (set by a predefined user). The U-Net network is a deep learning network architecture commonly used for image segmentation tasks, featuring a symmetrical encoder-decoder structure. The encoder gradually reduces the image size through downsampling operations (such as convolution and pooling) while extracting features at different levels, from low-level edge and texture features to high-level semantic features. The decoder gradually restores the image size through upsampling operations (such as transposed convolution) and integrates the features from the encoder with the corresponding levels of the downsampling modules. The image segmentation module is composed of convolutional and pooling layers. The convolutional layer extracts and transforms local features by sliding the convolutional kernel across the input image or feature map. The pooling layer downsamples the feature map output by the convolutional layer, usually using max pooling or average pooling, to reduce the size of the feature map while retaining the main features, allowing the model to focus on a wider range of image information. Through multi-resolution feature extraction, the image feature extraction model can understand image information at different scales, paying attention to both subtle damage features and the location and distribution of damaged areas in the overall image, laying the foundation for accurate identification and segmentation of damaged areas in the future.

[0044] Spatial attention processing is applied to the feature map to generate a spatial attention weight map that enhances the importance of key injury regions defined by pre-defined personnel. A spatially weighted feature map is then obtained based on the spatial attention weight map and the feature map. Injury regions include fracture areas, soft tissue injury areas, and hematoma areas. The spatial attention processing mechanism generates the spatial attention weight map by analyzing the correlation between different locations in the feature map and these key injury regions. Key injury regions in the weight map have higher weight values, while non-key regions have lower weight values. The spatial attention weight map is then weighted with the original feature map to obtain a spatially weighted feature map. This weighting enhances the features of key injury regions, improving the image feature extraction model's ability to identify key injury features.

[0045] The spatially weighted feature map is input into the self-attention module to obtain a self-attention weighted feature map used to capture global contextual information. By calculating the similarity between different locations in the feature map, the degree of association between each location and other locations is determined, and corresponding attention weights are generated. The feature map is then weighted according to these attention weights to obtain the self-attention weighted feature map. The self-attention module can capture long-distance dependencies between different locations in the feature map, thereby obtaining global contextual information. This helps the image feature extraction model better understand the overall structure and semantics of the image, further improving the model's ability to understand and process complex image situations, and helping to more accurately identify and segment damaged areas.

[0046] The self-attention weighted feature map is input into the image feature extraction model to obtain image segmentation results containing the location and area of ​​each damage region in the image. The decoder part of the U-Net network gradually restores the image size through upsampling and fuses it with the features of the corresponding level of the encoder, finally outputting the image segmentation result. This result clearly marks the location and area information of each damage region (fracture region, soft tissue injury region, and hematoma region) in the image. By inputting it again into the image feature extraction model, combined with the previously extracted multi-level features, key region information emphasized by spatial attention, and global context information captured by self-attention, the model can generate accurate image segmentation results, improving the effectiveness of blast injury assessment. The self-attention mechanism helps the model capture global information, especially for dynamically changing regions in the image, enabling the model to effectively model on a global scale.

[0047] Specifically, the spatial attention processing workflow is as follows: Average pooling and max pooling operations are performed on the feature maps to obtain the average pooling spatial descriptor and the max pooling spatial descriptor, respectively. Average pooling means averaging all values ​​within a local region (pooling window) of the input feature map, which is then used as the value corresponding to that region on the output feature map. Average pooling can smooth the feature map, reduce data variance, and retain overall feature information. Max pooling means selecting the maximum value within a local region (pooling window) of the input feature map, which is then used as the value corresponding to that region on the output feature map. Max pooling can highlight salient features in the feature map and enhance the robustness of the model. Average pooling can retain the overall information of the feature map, has a certain smoothing effect on noise, and can capture common features in the feature map. Max pooling highlights salient features in the feature map, such as edges and corners, enhancing the model's ability to perceive important features. By using average pooling and max pooling simultaneously, spatial information of the feature map can be extracted from different perspectives, providing rich feature representations for generating more accurate attention weights.

[0048] The average pooling spatial descriptor and the max pooling spatial descriptor are concatenated to obtain a concatenated feature map. Concatenation means merging two spatial descriptors along a specific dimension. For example, if two spatial descriptors are the same size in other dimensions but differ only in one dimension (such as the channel dimension), they can be concatenated along that dimension to form a new feature map. The number of channels in the new feature map is the sum of the number of channels in the original two spatial descriptors. The concatenation operation combines different types of spatial information extracted by average pooling and max pooling, making the new feature map contain a more comprehensive spatial feature representation. This helps subsequent convolutional operations to better capture the complex spatial relationships in the feature map and improves the accuracy of spatial attention weights.

[0049] The concatenated feature maps are convolved using convolution kernels to generate intermediate feature maps. A suitable convolution kernel (e.g., 3×3) is selected and slid across the concatenated feature maps with a preset stride. At each slid position, the convolution kernel performs a dot product with a local region on the feature map, and the results are summed to obtain the corresponding value on the intermediate feature map. Through multiple convolution operations (multiple convolutional layers can be set), spatial information in the feature maps can be further extracted and integrated. Convolution operations can automatically learn local patterns and spatial relationships in the feature maps. The intermediate feature map contains spatial information processed by convolution operations, better reflecting the importance of different locations in the feature map.

[0050] The intermediate feature map is activated using the Sigmoid activation function to obtain a spatial attention weight map. The activation process involves inputting each value of the intermediate feature map into the Sigmoid activation function, which maps the input values ​​to the (0,1) interval. Each value in the weight map represents the attention weight of the corresponding position in the feature map. The closer the value is to 1, the more important the position is; the closer the value is to 0, the less important the position is. This allows for dynamic adjustment of the contribution of different positions in the feature map, enabling the model to focus more on important spatial regions and ignore irrelevant regions, thereby improving the performance and generalization ability of the image feature extraction model.

[0051] like Figure 5 The flowchart shown, illustrating the segmentation accuracy determination, further explains that the self-attention weighted feature map is input into the image feature extraction model to obtain image segmentation results containing the location and area of ​​each damage region in the image. This process also includes: The segmentation accuracy, or Intersection over Union (IoU), is the ratio of the intersection area to the union area of ​​the predicted segmentation region and the actual segmentation region. The closer the ratio is to 1, the higher the segmentation accuracy. Segmentation accuracy is used to measure the degree of overlap between the predicted segmentation region and the actual segmentation region (i.e., the result of the preset segmentation by the personnel), providing a quantitative basis for subsequent segmentation optimization, so as to carry out targeted optimization.

[0052] If the segmentation accuracy is not greater than the first reference segmentation accuracy, then the first segmentation optimization is performed to effectively solve the problem of loss of detail information caused by downsampling in deep networks.

[0053] If the segmentation accuracy is greater than the first reference segmentation accuracy but not greater than the second reference segmentation accuracy, then a second segmentation optimization is performed to improve the accuracy of the segmentation boundary.

[0054] If the segmentation accuracy is greater than the second reference segmentation accuracy, no segmentation optimization is performed. The first reference segmentation accuracy is less than the second reference segmentation accuracy. The first and second reference segmentation accuracies are set by preset personnel. Segmentation optimization includes first segmentation optimization and second segmentation optimization.

[0055] Hierarchical optimization based on segmentation accuracy allows for appropriate optimization measures to be taken for segmentation results of different quality, avoiding blind optimization; the first segmentation optimization solves the problem of loss of detail information caused by downsampling in deep networks through structural optimization and parameter optimization, improving the model's ability to recognize details; the second segmentation optimization improves the accuracy of segmentation boundaries by fusing edge feature maps.

[0056] Further, the first segmentation optimization is as follows: The image feature extraction model is structurally optimized to obtain the first optimized model. Structural optimization means introducing a dense connection module, so that each layer in the model receives the feature maps of all its predecessor layers as input. Unlike the traditional neural network structure where each layer only receives the output of the previous layer, dense connections can make fuller use of the features extracted from different layers, reduce information loss, and help to segment images more accurately.

[0057] The first optimization model is used to segment the self-attention weighted feature map to obtain the first optimized segmentation result and the corresponding optimized segmentation accuracy. By using the first optimization model to segment the self-attention weighted feature map, the quality of the segmentation result can be further improved, making the segmentation result closer to the real annotation.

[0058] If the optimized segmentation accuracy is not greater than the first reference segmentation accuracy, then the parameters of the first optimized model are optimized to obtain the second optimized model, and the second optimized model is used for segmentation. Parameter optimization means increasing the number of convolutional layers and the number of channels in the convolutional layers through cross-validation to minimize the model complexity while meeting the requirements of the first reference segmentation accuracy. This improves the performance of the image feature extraction model and avoids the waste of computational resources and overfitting caused by an overly complex image feature extraction model. Increasing the number of convolutional layers can enhance the nonlinear expression capability of the model, enabling the model to learn more complex feature patterns. Increasing the number of channels in the convolutional layers can expand the dimension of features and improve the model's ability to distinguish different features.

[0059] The second segmentation optimization is as follows: The acquired edge feature maps and self-attention weighted feature maps are fused to obtain a fused feature map. Fusion means directly concatenating the edge feature maps and self-attention weighted feature maps along the channel dimension of the feature maps. By increasing the number of channels, richer feature representations are provided for subsequent model processing, which helps the model learn more comprehensive image information and improves the segmentation accuracy of damaged areas. The edge feature maps highlight the edge information of objects in the image, so that the fused feature map can simultaneously possess important features and edge information focused on by self-attention, which can provide the model with more comprehensive information and help the model more accurately identify the boundaries and details of objects during segmentation.

[0060] The fused feature map is input into the image feature extraction model to obtain the second optimized segmentation result, which makes the model pay more attention to details during segmentation and improves segmentation accuracy. The edge feature map represents the result obtained by edge detection operators (such as Sobel operator, Canny operator, etc.). It combines the overall features of the image and edge detail information, which enables the model to more accurately determine the boundaries of objects during segmentation, improves the accuracy of segmentation boundaries, and thus obtains more accurate segmentation results.

[0061] Furthermore, the injury status is determined based on the acquired physiological and injury parameters of each target. The specific process is as follows: The system acquires the target's physiological parameters in real time. If a target's physiological parameter is outside the reference range (set by a preset person), the preset person is immediately alerted that the target's condition is abnormal. Physiological parameters include blood pressure, heart rate, blood oxygen saturation, and body temperature. The reference range includes reference blood pressure range, reference blood oxygen saturation range, and reference body temperature range. If any physiological parameter is found to be outside its corresponding reference range, subsequent prompts and judgment processes are triggered. This reflects the target's physical condition from different perspectives, avoiding the omission of important information due to judgment based on a single parameter, and improving the accuracy of assessing the target's injury status.

[0062] If the target's physiological parameters are within the reference range, the damage area is determined. Specifically, damage area determination involves the following steps: if the damage area of ​​each affected region is greater than the initial damage area after a preset monitoring period (set by preset personnel) (i.e., the area of ​​each affected region recorded at the start of target monitoring, serving as a benchmark for subsequent assessment of injury deterioration), the preset personnel are immediately alerted to assess the deterioration. Otherwise, they are prompted to perform wound examination according to a preset cycle (set by preset personnel). This quantitative approach to assessing injury deterioration provides an objective basis for injury evaluation, offering greater accuracy and reliability compared to purely subjective judgment. It clearly distinguishes between deteriorating injury and abnormal wound condition, enabling preset personnel to take more targeted measures. For example, deteriorating injury requires emergency treatment, while abnormal wound condition requires initial wound examination to determine the appropriate course of action. From physiological parameter monitoring to damage area determination and wound condition management, a complete closed-loop injury assessment system is formed.

[0063] In summary, this invention, through sensors mounted on a drone, acquires surface images of all targets. Utilizing the drone's high maneuverability and wide field of view, it can rapidly cover the explosion site, solving the problem that rescuers cannot immediately contact all targets during an explosion. This achieves wide-area, rapid, and contactless initial information collection, laying the data foundation for subsequent batch processing and rapid triage. The extracted surface injury features are assigned corresponding emergency levels, enabling automated and standardized on-site triage. This allows for timely prioritization of a large number of targets, identifying individuals most in need of emergency rescue, greatly improving rescue efficiency and resource utilization. Then, dynamic injury assessment is performed on all targets according to their emergency levels, realizing dynamic utilization of computational resources. Intelligent allocation selects damage analysis strategies based on the image features of each target, achieving an intelligent balance between assessment accuracy and efficiency. This avoids the waste of resources caused by directly using the most complex but time-consuming models for simple image analysis, and also avoids the problem of insufficient accuracy caused by using simple models to analyze complex images. Finally, the selected damage analysis strategy is used to analyze the image to assess the injury, improving the accuracy of image analysis and thus enabling more accurate injury assessment. The injury status judgment result is obtained based on the acquired physiological and damage parameters of each target, and the overall physiological state of the target is comprehensively judged. This achieves multimodal information fusion assessment, which can more comprehensively and accurately judge the true state of the target, ensuring that abnormal situations can be detected and dealt with in a timely manner.

[0064] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0065] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0066] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0067] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0068] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0069] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0070] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0071] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0072] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0073] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0074] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0075] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0076] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0077] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A deep learning-based system for assessing blast injuries and analyzing image data, characterized in that, The system includes: a body surface image analysis module, a feature extraction and damage analysis and assessment module, and a damage analysis, assessment, and feedback module; The body surface image analysis module is used to: acquire body surface images of all targets through sensors mounted on the UAV, process the body surface images of each target based on the distribution of pixel values ​​in the body surface images to extract body surface damage features reflecting blast injuries in the body surface images, and obtain emergency warning judgment results based on the body surface damage features. The feature extraction and damage analysis assessment module is used to: if the received emergency warning judgment result is that an emergency warning has been triggered, then prioritize the dynamic assessment of the injury of the target that triggered the emergency warning; otherwise, perform dynamic assessment of the injury of all targets according to the difference of the extracted surface damage features, and determine whether to select global damage analysis based on the image features of each target to improve the accuracy and efficiency of injury assessment. The damage analysis, assessment, and feedback module is used to: if so, obtain the injury status determination result based on the acquired damage area of ​​each target and provide feedback; otherwise, assess the injury of each target based on the damage analysis strategy of a multi-scale convolutional neural network, obtain the corresponding damage parameters of the target, obtain the injury status determination result based on the acquired physiological parameters and damage parameters of each target, and provide feedback.

2. The deep learning-based blast injury assessment and image data analysis system according to claim 1, characterized in that, The specific process for processing the surface images of each target is as follows: The body surface image is normalized to obtain the processed body surface image; If the average variance of the processed body surface image of each target is greater than the average variance of the reference, then denoising is performed and scale consistency is determined; otherwise, scale consistency is determined directly. The scale consistency determination is specifically as follows: The area of ​​the damaged region in the processed body surface image is used to calculate the consistency, and the scale consistency index is obtained to quantify the degree of consistency of the area of ​​the damaged region in the body surface image. If the scale consistency index is greater than the reference scale consistency index, the preset scale is used to extract the surface damage features; otherwise, the corresponding surface damage features are extracted by combining the pyramid pooling structure and the convolution kernel size corresponding to each damage region. The kernel size represents the result obtained by mapping the area of ​​each damaged region into the kernel size mapping set.

3. The deep learning-based blast injury assessment and image data analysis system according to claim 2, characterized in that, The specific process for image normalization of the body surface image is as follows: The body surface image is divided according to a preset sliding window and a preset step size to obtain the local window image; The pixel values ​​within each local window are averaged to obtain the average pixel value of the corresponding local window image. The standard deviation of all pixel values ​​and the corresponding average pixel value within each local window is processed to obtain the pixel standard deviation of the corresponding local window image. The pixel standard deviation reflects the degree of dispersion of pixel value distribution within the corresponding local window. The regularization factor, the pixel values ​​of each local window image, the pixel mean, and the pixel standard deviation are normalized to obtain normalized pixel values. The normalized pixel values ​​in each local window are replaced with the corresponding pixel values ​​to obtain a normalized surface image, thereby eliminating local illumination unevenness.

4. The deep learning-based blast injury assessment and image data analysis system according to claim 1, characterized in that, The specific process for obtaining emergency warning determination results based on body surface injury characteristics is as follows: The surface injury characteristics of each target are fed back to the designated personnel, who then confirm the emergency level. If there is a target with an emergency level no less than the reference emergency level, an emergency warning will be triggered, and the preset personnel will be prompted to prepare resources in advance. The target that triggered the emergency warning will be set as the priority target for injury assessment, so that the dynamic injury assessment of the target with the higher emergency level will be prioritized. If there are no targets with an emergency level no lower than the reference emergency level, then all targets are sorted in descending order according to their emergency level and their injuries are dynamically assessed in sequence.

5. The deep learning-based blast injury assessment and image data analysis system according to claim 1, characterized in that, The process for determining whether to select global damage analysis based on the image features of each target is as follows: The standard deviation of the damage area in each damaged region in the image is processed to obtain the local region density difference to assess the fluctuation of the damage degree in each damaged region. If the density difference in a local area is greater than that in a reference local area, spatial heterogeneity analysis is used to assess the injury; otherwise, global damage analysis is used to assess the injury. The spatial heterogeneity analysis involves using a multi-scale convolutional neural network to extract features from the image layer by layer, ensuring that both local details and global structure are captured simultaneously. The global damage analysis specifically involves: statistically analyzing the damage area of ​​all damaged regions to obtain a global quantitative index to quantify the target injury status.

6. The deep learning-based blast injury assessment and image data analysis system according to claim 5, characterized in that, The spatial heterogeneity analysis process is as follows: The image images of each target are input into an image feature extraction model for extracting features at different resolutions to obtain feature maps containing local details and global structure. The image feature extraction model is a neural network model containing a preset number of downsampling modules. Spatial attention processing is performed on the feature map to generate a spatial attention weight map. The spatial attention weight map is used to increase the importance of key damage areas set by preset personnel, and a spatial weighted feature map is obtained based on the spatial attention weight map and the feature map. The spatially weighted feature map is input into the self-attention module to obtain a self-attention weighted feature map used to capture global contextual information; The self-attention weighted feature map is input into the image feature extraction model to obtain image segmentation results containing the location and area of ​​each damaged region in the image.

7. The deep learning-based blast injury assessment and image data analysis system according to claim 6, characterized in that, The spatial attention processing flow is as follows: The feature maps are subjected to average pooling and max pooling operations respectively to obtain the average pooling space descriptor and the max pooling space descriptor. The average pooling space descriptor and the max pooling space descriptor are concatenated to obtain the concatenated feature map; The concatenated feature maps are convolved using convolution kernels to generate intermediate feature maps. The intermediate feature map is activated using the Sigmoid activation function to obtain the spatial attention weight map.

8. The deep learning-based blast injury assessment and image data analysis system according to claim 6, characterized in that, The process of inputting the self-attention weighted feature map into the image feature extraction model to obtain image segmentation results containing the location and area of ​​each damaged region in the image image, further includes: Obtain the segmentation accuracy used to measure the degree of overlap between the predicted segmented region and the actual segmented region; If the segmentation accuracy is not greater than the first reference segmentation accuracy, then the first segmentation optimization is performed to effectively solve the problem of loss of detail information caused by downsampling in deep networks; If the segmentation accuracy is greater than the first reference segmentation accuracy but not greater than the second reference segmentation accuracy, then a second segmentation optimization is performed to improve the accuracy of the segmentation boundary. If the segmentation accuracy is greater than the second reference segmentation accuracy, no segmentation optimization is performed. The segmentation optimization includes first segmentation optimization and second segmentation optimization. The first reference segmentation accuracy is less than the second reference segmentation accuracy.

9. The deep learning-based blast injury assessment and image data analysis system according to claim 8, characterized in that, The first segmentation optimization is as follows: The image feature extraction model is structurally optimized to obtain a first optimized model. The structural optimization means introducing a dense connection module so that each layer in the model receives the feature maps of all its predecessor layers as input to share and propagate more fully and reduce information loss. The self-attention weighted feature map is segmented using the first optimization model to obtain the first optimized segmentation result, and the corresponding optimized segmentation accuracy is obtained. If the optimized segmentation accuracy is not greater than the first reference segmentation accuracy, then the parameters of the first optimized model are optimized to obtain the second optimized model, and the second optimized model is used for segmentation. The parameter optimization means increasing the number of convolutional layers and the number of channels in the convolutional layers through cross-validation to minimize the model complexity while meeting the requirements of the first reference segmentation accuracy. The second segmentation optimization is as follows: The acquired edge feature map and the self-attention weighted feature map are fused to obtain the fused feature map; The fused feature map is input into the image feature extraction model to obtain the second optimized segmentation result, which makes the model pay more attention to details during segmentation and improves the segmentation accuracy. The edge feature map represents the result obtained by using the edge detection operator.

10. The deep learning-based blast injury assessment and image data analysis system according to claim 1, characterized in that, The injury status assessment result is obtained based on the acquired physiological and injury parameters of each target. The specific process is as follows: If the target's physiological parameters are outside the reference range, the pre-selected personnel will be immediately alerted. If the target's physiological parameters are within the reference physiological parameter range, the damage area is determined. Specifically, if the damage area of ​​each damaged area is greater than the initial damage area after the preset monitoring period, the preset personnel are immediately notified; otherwise, the preset personnel are notified to perform wound inspection according to the preset cycle.

Images