A post-traumatic terahertz and optical multi-modal assessment system

Abstract

The application aims to provide a post-traumatic terahertz and optical multi-modal evaluation system, comprising: an imaging probe assembly, the imaging probe assembly comprising a terahertz emission unit, a terahertz receiving unit, a visible light imaging unit and an infrared thermal imaging unit arranged around a common optical axis, and a sterile isolation window arranged at the front end of the imaging probe assembly; an imaging control and data acquisition module, the imaging control and data acquisition module being electrically connected with the imaging probe assembly, used for controlling the terahertz emission unit to emit terahertz waves and collecting corresponding terahertz signals, and collecting image data output by the visible light imaging unit and the infrared thermal imaging unit; an emergency trauma multi-modal terahertz / optical intelligent evaluation system and method which can be used in an emergency environment, and improvements are made in multi-modal imaging, integrated scanning control, environment and individual difference correction, multi-injured person resource scheduling and safety stability guarantee.

Description

Technical Field

[0001] This invention belongs to the field of emergency trauma care, specifically relating to a post-traumatic terahertz and optical multimodal assessment system. Background Technology

[0002] Trauma is one of the most common and leading causes of death and disability in various emergencies. Medical personnel often need to conduct initial assessments and triage of multiple injured individuals within a very short timeframe to determine if there are critical conditions such as deep burns, severe crush injuries, or complex shrapnel injuries, and based on this, make decisions regarding immediate or priority evacuation. Traditional methods of visual inspection, palpation, and simple rating scales are significantly affected by factors such as ambient lighting, contamination, clothing obstruction, and differences in physician experience, easily leading to underestimation or overestimation of the severity of trauma, and failing to meet the needs of rapid, objective, and standardized emergency assessment.

[0003] To improve the objectivity of trauma assessment, various techniques based on single-modal imaging have been proposed in the current technology. For example, visible light cameras are used to photograph the wound surface, and the severity of the injury is estimated through color, shape, and area; infrared thermal imaging devices are used to obtain the temperature distribution of the wound surface, thereby inferring local blood flow and tissue perfusion; ultrasound, CT, or MRI imaging techniques are used to assess deep tissue damage; and in recent years, multispectral or hyperspectral optical imaging has emerged, attempting to distinguish burns of different depths by different reflections across multiple wavelengths. Furthermore, terahertz imaging technology, due to its sensitivity to tissue water content and microstructural changes, has also been used to study the depth of skin burns and the degree of tissue edema. While the aforementioned technologies can provide richer information than the naked eye in certain scenarios, they generally suffer from the following problems: First, most are single-modal devices, making it difficult to simultaneously consider surface morphology, surface temperature, and depth damage information, which can easily lead to incomplete diagnostic evidence; second, the devices are mostly used in environments such as hospitals, and are relatively large and complex in structure, making them unsuitable for carrying and rapid deployment in emergency situations; third, most imaging processes rely on manual handheld probes or fixed-path scanning, lacking adaptive scanning path planning and time resource optimization, making it difficult to achieve "maximum information" scanning in time-constrained emergency environments.

[0004] With the development of artificial intelligence and image processing technologies, some studies have attempted to apply machine learning or deep learning methods to the automatic grading and prediction of trauma images, providing auxiliary diagnosis for lesions such as burns and skin ulcers. However, these methods are usually based on single-modal visible light or thermal images, and the input data is easily affected by factors such as skin color, ambient lighting, ambient temperature, and humidity. Furthermore, existing methods focus more on "how to classify based on existing images" and less on "how to acquire the most valuable image data within a limited time." In emergency situations, environmental factors such as temperature, humidity, wind speed, and smoke fluctuate drastically, and there are significant differences in skin color and body fat thickness among injured individuals. Without correcting for these environmental parameters and individual differences, it is difficult to uniformly interpret image features obtained from different tasks and different battlefields. In addition, existing equipment and algorithms mostly assess individual wounded individuals, lacking a unified scheduling mechanism for scanning resources in scenarios where multiple wounded individuals are simultaneously assessed, and rarely considering safety constraints during the contact between the imaging probe and the wound, such as risks of excessive contact pressure and excessively high local temperatures.

[0005] In summary, existing technologies in the field of emergency trauma assessment still have the following shortcomings: First, there is a lack of systems that coaxially integrate multiple modalities such as terahertz imaging, visible light imaging, and infrared thermal imaging on the same portable probe, making it difficult to acquire depth damage information, surface temperature perfusion information, and fine edge morphology information of the trauma area in a single scan. Second, there is a lack of adaptive scanning path planning mechanisms for emergency environments. Existing solutions typically employ manual or fixed-path scanning, failing to incorporate the confidence level of trauma severity assessment, information uncertainty, and time and energy constraints into a unified optimization framework. Third, there is a lack of technical solutions for systematically calibrating multimodal data using parameters such as ambient temperature, humidity, wind speed, and the injured person's skin color and subcutaneous fat thickness, making the assessment results susceptible to interference from environmental and individual differences. Fourth, there is a lack of scanning time quota allocation and priority control strategies for multiple injured persons, making it impossible to reasonably queue and prioritize multiple injured persons for assessment under limited imaging time and power conditions. Fifth, existing systems do not adequately consider probe contact safety and acquisition stability under emergency conditions, lacking safety constraints and stability detection mechanisms based on pressure, temperature, tilt angle, and acceleration feedback. Therefore, it is necessary to propose a compact emergency trauma multimodal terahertz / optical intelligent assessment system and method that can be used in emergency environments, and to improve aspects such as multimodal imaging, integrated scanning control, environmental and individual difference correction, multi-patient resource scheduling, and safety and stability assurance, so as to overcome the shortcomings of the existing technologies. Summary of the Invention

[0006] The purpose of this invention is to provide a dynamic comprehensive training system to enhance high-altitude work capabilities. This invention can improve the multimodal terahertz / optical intelligent assessment system and method for trauma in emergency environments in terms of multimodal imaging, integrated scanning control, environmental and individual difference correction, multi-casualty resource scheduling, and safety and stability assurance.

[0007] To achieve the above objectives, the present invention provides the following technical solution, comprising: a post-traumatic terahertz and optical multimodal assessment system, comprising:

[0008] An imaging probe assembly, comprising a terahertz emitting unit, a terahertz receiving unit, a visible light imaging unit, and an infrared thermal imaging unit arranged around a common optical axis, wherein a sterile isolation window is provided at the front end of the imaging probe assembly.

[0009] An imaging control and data acquisition module is electrically connected to the imaging probe assembly. It is used to control the terahertz emitting unit to emit terahertz waves and acquire the corresponding terahertz signals, as well as to acquire the image data output by the visible light imaging unit and the infrared thermal imaging unit.

[0010] A multimodal registration and feature fusion module, which is connected to the imaging control and data acquisition module, is used to perform spatial registration and intensity calibration on the terahertz signal, visible light image and infrared thermal image to generate multidimensional feature vectors at each location of the trauma area;

[0011] The scanning path planning and execution module includes an actuator mechanically connected to the imaging probe assembly and a reinforcement learning control unit connected to the multimodal registration and feature fusion module. The actuator is used to drive the imaging probe assembly to move relative to the trauma area, and the reinforcement learning control unit is used to output scanning path control commands based on the current multidimensional feature vector and the current scanning state.

[0012] An emergency environment and individual parameter acquisition module is used to acquire environmental parameters at the trauma scene and individual parameters of the injured object.

[0013] A trauma severity assessment and grading module, connected to the multimodal registration and feature fusion module and the emergency environment and individual parameter acquisition module, is used to intelligently assess the severity of trauma based on the multidimensional feature vector, environmental parameters, and individual parameters, and output grading results and subsequent priority suggestions. The reward function of the reinforcement learning control unit includes a reward term to improve the confidence of the trauma severity assessment results and a penalty term to constrain scan time and energy consumption, thereby improving the reliability of the trauma severity assessment results under predetermined scan time constraints.

[0014] Furthermore, the terahertz transmitting unit is used to transmit terahertz pulses in multiple frequency bands, including a first frequency band and a second frequency band, wherein the center frequency of the first frequency band is lower than the center frequency of the second frequency band;

[0015] The imaging probe assembly has a spatial coding structure at its front end, which is used to spatially modulate incident or reflected terahertz waves.

[0016] The multimodal registration and feature fusion module includes a compressed sensing reconstruction unit, which is used to reconstruct the terahertz intensity distribution of the trauma area based on spatial coding measurement data.

[0017] Furthermore, the state space of the reinforcement learning control unit includes: the current position of the imaging probe assembly relative to the trauma area, the distribution of areas that have been scanned, and the distribution of graded confidence output by the trauma severity assessment and grading module; the reward function of the reinforcement learning control unit is also used to increase the penalty value when repeatedly scanning areas that already have high graded confidence, so as to reduce repeated scanning of areas with high confidence.

[0018] Furthermore, the imaging probe assembly is equipped with a contact pressure sensor and a temperature sensor at its front end, and the scanning path planning and execution module is equipped with a safety constraint unit. The safety constraint unit is connected to the reinforcement learning control unit and is used to: restrict the scanning path control command output by the reinforcement learning control unit when the contact pressure or local temperature exceeds a preset threshold, and to impose a penalty on the corresponding action in the reward function when the reinforcement learning control unit attempts to execute an action that exceeds the safety threshold.

[0019] Furthermore, the multimodal registration and feature fusion module includes a geometric registration unit and an edge-guided reconstruction unit;

[0020] The geometric registration unit is used to perform spatial registration of the terahertz intensity distribution based on the visible light image and the infrared thermal image;

[0021] The edge-guided reconstruction unit is used to reconstruct the terahertz intensity distribution and the infrared thermal image using the trauma edge information in the visible light image, so as to improve the spatial resolution of the trauma boundary region.

[0022] Furthermore, the emergency environment and individual parameter acquisition module includes a temperature sensor, a humidity sensor, and a wind speed sensor, as well as an individual information acquisition unit for acquiring skin color information and body fat thickness information of the injured person;

[0023] The trauma severity assessment and grading module includes a calibration unit, which is used to correct the terahertz signal intensity, infrared temperature value, and visible light brightness in the multidimensional feature vector based on the temperature, humidity, wind speed, skin color information, and body fat thickness information.

[0024] Furthermore, the trauma severity assessment and grading module includes a multi-casualty resource allocation unit, which is used for:

[0025] After completing the initial rapid scan of multiple injured individuals, a fine scan time quota is allocated to each injured individual based on the initial grading results and grading confidence level. The scanning path planning and execution module is then controlled to perform fine scans on each injured individual sequentially according to the fine scan time quota.

[0026] Furthermore, the sterile isolation window adopts a double-layer structure, which includes an outer heating anti-condensation layer and an inner optical layer. The heating anti-condensation layer is used to reduce water vapor condensation on the window surface in a high-humidity environment, and the optical layer is used to provide a stable transmission path for terahertz waves, visible light, and infrared light while maintaining optical performance. The imaging probe assembly is equipped with a tilt sensor and an acceleration sensor, and the scanning path planning and execution module is equipped with a stability detection unit. The stability detection unit is used to: pause the acquisition of the imaging control and data acquisition module and issue a prompt message to the operator when the tilt angle change detected by the tilt sensor or the acceleration detected by the acceleration sensor exceeds a preset threshold.

[0027] Furthermore, the following steps are included:

[0028] S1, simultaneously acquire terahertz signals, visible light images and infrared thermal images of the trauma area through the imaging probe assembly, and send the terahertz signals, visible light images and infrared thermal images to the imaging control and data acquisition module;

[0029] S2, input the terahertz signal, visible light image and infrared thermal image into the multimodal registration and feature fusion module, and obtain multidimensional feature vectors of each location in the trauma area after geometric registration and intensity calibration;

[0030] S3, through the emergency environment and individual parameter acquisition module, collects environmental temperature, environmental humidity, wind speed, as well as skin color information and body fat thickness information of the injured person, and based on the environmental temperature, environmental humidity, wind speed, skin color information and body fat thickness information, corrects the terahertz signal intensity, infrared temperature value and visible light brightness in the multidimensional feature vector;

[0031] S4, the corrected multidimensional feature vector and the current scanning status are input into the reinforcement learning control unit in the scanning path planning and execution module. The reinforcement learning control unit outputs scanning path control commands to drive the imaging probe assembly to scan the trauma area according to the scanning path.

[0032] S5 inputs the collected multimodal features into the trauma severity assessment and grading module to assess the severity of the trauma and output the grading results and subsequent priority suggestions.

[0033] Furthermore, in step S4, the reward function used by the reinforcement learning control unit increases the reward value when the confidence of the trauma severity classification increases, increases the penalty value when repeatedly scanning areas that already have a high classification confidence, and increases the penalty value when the scanning time or energy consumption exceeds a preset threshold, so as to improve the reliability of the trauma severity classification results within the predetermined total scanning time constraint.

[0034] When there are multiple injured individuals, the procedure further includes: performing an initial rapid scan as described in steps S1 to S5 on each injured individual to obtain an initial grading result and its grading confidence level; allocating a fine scan time quota to each injured individual based on the initial grading result and its grading confidence level; and sequentially performing supplementary scans and assessments on each injured individual according to the fine scan time quota.

[0035] Compared with the prior art, the emergency trauma multimodal terahertz / optical intelligent assessment system and method of the present invention have at least the following beneficial effects:

[0036] Multimodal coaxial imaging provides more comprehensive information and more reliable judgment. This invention integrates a terahertz transmitting / receiving unit, a visible light imaging unit, and an infrared thermal imaging unit around a common optical axis within a single imaging probe assembly. Through a multimodal registration and feature fusion module, spatial registration and intensity calibration of the terahertz signal, visible light image, and infrared thermal image are performed to generate multidimensional feature vectors for each location in the wound area. Compared to existing technologies that only use visible light or infrared single-modal imaging, this invention can simultaneously acquire information on deep water content / tissue damage, surface temperature and perfusion information, as well as wound morphology and edge information in a single scan. This enables a three-dimensional, comprehensive assessment of the wound area, significantly reducing the risk of misclassification or missed diagnosis due to reliance on visual inspection or a single device.

[0037] This invention combines terahertz dual-band imaging with spatial coding compression reconstruction to improve imaging speed and depth diagnostic capabilities. It transmits terahertz pulses in both the first and second frequency bands using a terahertz transmitting unit, and incorporates a spatial coding structure at the front end of the imaging probe. A compressed sensing reconstruction unit then reconstructs the spatially coded measurement data, significantly reducing the sampling volume while maintaining image quality. The low-frequency band facilitates penetration of deeper tissues, while the high-frequency band provides higher resolution. Combined with compressed sensing, this allows for depth imaging of large wound areas within a limited time. Compared to traditional point-scan or line-scan terahertz imaging, this invention enables high-value terahertz imaging in a shorter time, even under conditions of limited emergency time and power, improving the ability to identify deep injury areas.

[0038] This invention employs a reinforcement learning-driven adaptive scanning path to maximize diagnostic value under time constraints. The scanning path planning and execution module utilizes a reinforcement learning control unit to incorporate the current probe position, the distribution of scanned areas, and the distribution of grading confidence levels output by the trauma severity assessment and grading module into the state space. The reward function includes both a reward for increasing assessment confidence and a penalty for repeated scans and high time / energy consumption. Based on this, the system can prioritize detailed scanning of uncertain areas, reduce repeated scanning of high-confidence areas, and adaptively adjust the scanning strategy to meet the predetermined total scanning time constraint. Compared to existing fixed-path or human-experience-based scanning methods, this invention achieves "acquiring more useful information in the same amount of time" without increasing hardware burden, significantly improving the reliability and efficiency of trauma grading results, and is particularly suitable for rapid emergency assessment scenarios.

[0039] Environmental and individual parameter calibration enhances the robustness and comparability of multimodal assessment. This invention includes an emergency rescue environment and individual parameter acquisition module to collect temperature, humidity, wind speed, and skin color and subcutaneous fat thickness information of the injured subject. A calibration unit is incorporated into the trauma severity assessment and grading module to correct the terahertz signal intensity, infrared temperature value, and visible light brightness in the multidimensional feature vector based on these parameters. By actively introducing environmental and individual differences, this invention reduces the interference of different war zones, climates, and skin color / body type conditions on imaging results, enabling the same assessment model to maintain relatively stable interpretation thresholds and diagnostic accuracy in various complex environments, thereby improving the system's generalization ability and practicality.

[0040] This invention features dual protection through safety constraints and stability detection, adapting to harsh emergency conditions and reducing usage risks. A contact pressure sensor and a temperature sensor are positioned at the front end of the imaging probe. A safety constraint unit limits the output of the reinforcement learning control unit when the pressure or local temperature exceeds a preset threshold, and penalizes actions that attempt to exceed the safety threshold in the reward function. Simultaneously, a tilt sensor and an acceleration sensor are installed in the imaging probe. A stability detection unit pauses data acquisition and issues a warning when attitude changes or motion acceleration exceed a threshold. Furthermore, the sterile isolation window employs a double-layer structure with an outer heating anti-condensation layer and an inner optical layer, reducing fogging and maintaining optical performance in high-humidity environments. These structures, working in conjunction with the control logic, ensure safe contact and sterile isolation for wounded personnel in field environments, while minimizing the impact of shaking and condensation on image quality, thus improving the overall reliability of the system in harsh environments.

[0041] Intelligent allocation of scanning resources for multiple casualties improves the overall efficiency of emergency medical treatment. This invention incorporates a multi-casualty resource allocation unit within the trauma severity assessment and grading module. After initial rapid scanning of multiple injured individuals, a detailed scanning time quota is allocated to each casualty based on their initial grading results and confidence levels. The scanning path planning and execution module then sequentially performs the detailed scans. In situations involving mass casualties or multiple casualties arriving simultaneously, the system can rationally allocate scanning resources within limited scanning time and power constraints. Prioritizing more imaging and assessment resources for those with severe injuries and uncertain assessments optimizes the overall priority of emergency medical treatment and evacuation decisions, thereby improving overall treatment efficiency and success rate.

[0042] In summary, this invention overcomes the shortcomings of existing emergency trauma assessment technologies in terms of limited imaging information, poor environmental adaptability, low scanning efficiency, insufficient capacity to handle multiple casualties, and weak safety control through collaborative design in hardware structure, imaging mode, intelligent scanning control, environmental and individual correction, multi-casualty resource scheduling, and safety and stability assurance. It has significant substantive features and remarkable technological progress. Attached Figure Description

[0043] Figure 1 This is a block diagram of the overall structure of the emergency trauma terahertz and optical multimodal assessment system of the present invention. The diagram illustrates the overall composition of the system and the data flow between modules, including: imaging probe assembly 1, imaging control and data acquisition module 2, multimodal registration and feature fusion module 3, reinforcement learning-based scan path planning and execution module 4, emergency environment and individual parameter acquisition module 5, trauma severity assessment and grading module 6, and user interface and display module 7. The modules interact through a central controller (processor and memory), forming a closed-loop structure of "multimodal imaging—feature fusion—adaptive scanning—intelligent assessment".

[0044] Figure 2 This is a structural block diagram of the imaging control and data acquisition module of the present invention. The diagram illustrates the flow of various signals output from the imaging probe assembly 1 to the imaging control and data acquisition module 2, as well as the functional division within the module. This module includes a THz control and high-speed A / D acquisition unit, used to trigger THz transmission according to a predetermined timing sequence and acquire time-domain or frequency-domain signals; it also includes a visible light / infrared synchronous acquisition and preprocessing unit, used to receive synchronous trigger signals, acquire image frames, and perform preliminary processing such as noise reduction and non-uniformity correction; simultaneously, it outputs multimodal raw data and auxiliary sensor readings with a unified timestamp to the multimodal registration and feature fusion module 3.

[0045] Figure 3 This is a structural block diagram of the multimodal registration and feature fusion module of the present invention. The diagram shows the process by which the module receives THz images / signals, infrared thermal images, visible light images, and environmental temperature, humidity, wind speed, skin color, and fat thickness information output by the emergency environment and individual parameter acquisition module 5, and processes this information through a geometric registration unit, an intensity calibration unit, an edge-guided reconstruction unit, and a feature vector construction unit. The geometric registration unit is responsible for resampling the THz image to align it with the visible light / infrared coordinate system; the intensity calibration unit performs environmental compensation for THz intensity and infrared temperature; the edge-guided reconstruction unit uses visible light edge information to perform super-resolution or interpolation optimization on the THz / IR image; and the feature vector construction unit generates a multidimensional feature vector containing components at each pixel / voxel location and outputs a multimodal feature map of the trauma area.

[0046] Figure 4 This is a schematic diagram of the emergency environment and individual parameter acquisition module of the present invention. The diagram illustrates the components, including an ambient temperature sensor, a relative humidity sensor, a wind speed sensor, a skin color information acquisition unit, and a subcutaneous fat thickness acquisition unit. The temperature sensor measures the current ambient temperature, the humidity sensor measures the relative humidity, and the wind speed sensor measures the wind speed. The skin color information acquisition unit images uninjured skin areas using a visible light camera to estimate skin color type. The fat thickness acquisition unit obtains subcutaneous fat thickness at different locations through ultrasound measurement or BMI-based estimation. These parameters are uniformly transmitted to the multimodal registration and feature fusion module 3 for environmental and individual difference correction of THz, IR, and visible light characteristics.

[0047] Figure 5This is a schematic diagram of the scanning path planning and execution module based on reinforcement learning in this invention. The diagram illustrates the interaction between the reinforcement learning agent (RL Control Unit), the environment (trauma area and probe / sensor), and the actuator (motion platform or robotic arm). The agent selects the next action (up / down / left / right movement, local fine scanning, end of scan) from the action space based on the current state (including probe position, scan coverage, and evaluation confidence distribution). The actuator drives the probe movement, and the environment returns new multimodal observations and immediate rewards. The reward function includes confidence enhancement rewards, repetitive scan penalties, and time / energy consumption penalties. The strategy trained through reinforcement learning can prioritize fine scanning of areas with high information content or uncertain evaluation within a limited time, achieving adaptive scanning path optimization.

[0048] Figure 6 This is a structural block diagram of the trauma severity assessment and multi-casualty resource allocation module of the present invention. The diagram shows that the trauma severity assessment and grading module 6 receives input from a multimodal feature map and feature vector set, and outputs pixel-level severity distribution (severity heatmap) and overall trauma severity level, along with subsequent priority suggestions, through a deep neural network or other machine learning model. Simultaneously, the diagram also shows the multi-casualty resource allocation unit: when multiple casualties are present at the scene, this unit makes optimization decisions regarding subsequent fine-scale scanning time quotas and scanning order based on the initial assessment results and confidence distribution of each casualty, thereby achieving a reasonable allocation of limited scanning resources among multiple casualties.

[0049] Figure 7 This is a schematic diagram of the overall workflow of the system of the present invention. The main steps of the system are shown in the flowchart form: multimodal data acquisition (S1), multimodal registration and feature fusion (S2), environmental and individual parameter correction (S3), adaptive scanning path planning based on reinforcement learning (S4), trauma severity assessment and grading, and subsequent priority determination (S5). The steps are connected by data flow arrows, reflecting the complete workflow of the present invention from the acquisition of THz / visible light / infrared information by the imaging probe to feature fusion, environmental correction, adaptive scanning, and intelligent grading. Detailed Implementation

[0050] This embodiment provides a post-traumatic terahertz and optical multimodal assessment system. As described above, it integrates multiple sensing technologies such as terahertz (THz) imaging, visible light imaging, and infrared thermal imaging, and combines them with artificial intelligence algorithms to achieve rapid and accurate assessment of the severity of trauma in emergency patients. The system uses a portable imaging probe for non-contact or touchless scanning of the wound, and significantly improves diagnostic accuracy through a neural network model (e.g., the accuracy of burn severity diagnosis can be increased to approximately 93%). Multimodal data fusion utilizes the complementary advantages of information from different wavelengths: THz imaging is highly sensitive to changes in tissue water content and can be used to detect wound edema and tissue damage depth; infrared thermal imaging acquires the real-time temperature distribution of the wound surface, reflecting local blood flow and inflammation, which is very useful for early assessment of burn depth and monitoring of healing progress; visible light imaging provides information such as wound morphology, color, and edge details, helping to locate the wound extent and identify foreign objects. Through spatial registration and feature fusion of the above multi-source information, the system can generate multidimensional feature vectors for various locations in the trauma area, serving as input to the intelligent assessment model, thereby achieving objective and consistent severity grading and subsequent priority determination. Meanwhile, this embodiment introduces an adaptive scanning strategy using reinforcement learning control, enabling the imaging probe to concentrate more scanning resources on the most important areas of information (such as wound edges or areas of diagnostic uncertainty), thereby improving the efficiency of acquiring key diagnostic information within a limited time. The specific implementation of this invention will be described in detail below, in conjunction with the system structure and workflow.

[0051] Imaging probe assembly

[0052] The imaging probe assembly includes a terahertz emitting unit, a terahertz receiving unit, a visible light imaging unit, and an infrared thermal imaging unit arranged coaxially. Preferably, the terahertz emitting unit comprises pulsed terahertz sources in two frequency bands, for example, a first band with a center frequency of 0.3 THz and a second band with a center frequency of 0.8 THz (the first band's center frequency is lower than the second band), to accommodate both deeper tissue detection and superficial high-resolution imaging. For example, the first band (low-frequency THz, e.g., 0.2–0.5 THz) has greater penetration depth and can assess changes in subcutaneous tissue water content; the second band (high-frequency THz, e.g., 0.5–0.9 THz) provides higher spatial resolution, aiding in imaging details of surface wounds. The terahertz receiving unit can employ a terahertz antenna or detector (such as a THz time-domain spectrometer based on a photoconductive antenna) corresponding to the emitting unit to receive terahertz signals reflected or transmitted from wound tissue.

[0053] The visible light imaging unit can be a high-definition CMOS camera (e.g., 2 megapixels, visible light band 400–700nm) to acquire visible light color images of the wound area; the infrared thermal imaging unit can be an uncooled focal plane infrared camera (e.g., 640×480 pixels, 8–14μm wavelength thermal infrared) to measure the temperature distribution of the wound surface in real time, with a preferred temperature resolution of 0.1℃. The optical axes of the above imaging units are mutually calibrated and substantially collinear to ensure that the imaging fields of different modalities coincide. The front end of the imaging probe assembly is provided with a sterile isolation window, which adopts a double-layer structure, including an outer heating anti-condensation layer and an inner optical transmission layer. The heating anti-condensation layer is preferably glass or a thin film with a transparent conductive coating, which can maintain the window surface at a temperature slightly higher than the ambient dew point (e.g., ≈40℃) through a built-in heating element, thereby reducing water vapor condensation in high humidity environments. The optical layer is made of a material with high transmittance to THz waves, visible light, and infrared light, such as high-density polyethylene (HDPE) or polymer film a few millimeters thick, ensuring sterile isolation and providing a stable multimodal transmission path. The dual-layer sterile window structure ensures that the probe is not affected by ambient humidity when imaging close to the wound surface, and maintains good optical performance and sterile barrier.

[0054] The imaging probe assembly also integrates multiple auxiliary sensors for safety constraints and attitude stability detection. Specifically, the probe tip is equipped with a contact pressure sensor (such as a miniature strain gauge) and a temperature sensor to monitor the pressure and temperature when the probe comes into contact with the wound surface. The probe also contains a tilt sensor (attitude gyroscope) and an accelerometer to detect changes in the probe's spatial attitude and acceleration. The signals provided by these sensors will be used for safety threshold monitoring and stability detection during subsequent scanning processes to protect the injured person and ensure the quality of imaging data.

[0055] Furthermore, to improve the efficiency of terahertz imaging, a spatial coding structure can be provided at the front end of the imaging probe assembly. This structure can be a programmable spatial light modulator or a fixed coding mask, used to spatially modulate and code the incident or reflected terahertz waves. For example, a two-dimensional coding mask array (such as a metal plate with randomly patterned openings) can be placed in the terahertz transmitting / receiving optical path, changing the spatial transmission pattern each time it is emitted to code and modulate the terahertz field. Through this compressed sensing imaging scheme, the probe can acquire the terahertz intensity distribution of the trauma area with a sampling quantity far less than the number of pixels, thereby accelerating the imaging speed. Under the compressed sensing principle, the measurement signal y can be expressed as... Where x is the terahertz intensity image vector to be reconstructed, Φ is the measurement matrix determined by the spatial coding pattern, and n is noise. The compressed sensing reconstruction unit will solve for x in the reconstruction module, for example, by solving the following optimization problem to obtain the reconstructed image:

[0056] Under constraints Under the condition of minimizing ||x||1. In a specific implementation example, 1000 measurements with different spatial codes can be performed on a 100×100 pixel THz image to acquire data (compression sampling rate of 10%). The terahertz image can still be accurately restored by the l1 norm regularization reconstruction algorithm, realizing rapid imaging of a wound area of ​​about 10cm².

[0057] Imaging control and data acquisition module

[0058] The imaging control and data acquisition module controls the coordinated operation of each unit of the imaging probe and performs synchronous acquisition and preliminary processing of multimodal raw data. This module can consist of a high-speed data acquisition card, an embedded processor, and control circuitry. During operation, the imaging control module triggers the terahertz transmitting unit to radiate THz pulses according to a predetermined timing sequence and acquires the corresponding terahertz receiving signals through a high-speed A / D converter. For the THz time-domain spectral system, the time waveform of each pulse signal can be recorded, and then the amplitude and phase information in the desired frequency band can be extracted through a fast Fourier transform. If a multi-band discrete source (such as two center frequencies) is used, the control unit can switch between different frequency bands of THz sources to emit pulses sequentially and acquire the received signals separately. Simultaneously, the control module acquires image frame data from the visible light camera and infrared camera via gigabit Ethernet or a high-speed interface. To achieve synchronization of multimodal data, this module can send synchronization trigger signals to the visible light imaging unit and the infrared imaging unit, ensuring that the visible light image and thermal image at the corresponding moment are recorded simultaneously during each terahertz scan or each frame sampling. The collected data includes: terahertz time-domain or frequency-domain signal sequences, visible light image frames at corresponding locations, infrared thermal image frames, and readings from various auxiliary sensors (environmental sensors, pressure / tilt sensors, etc.).

[0059] During data acquisition, the imaging control module can also perform preliminary preprocessing on the data. For example, it applies non-uniformity correction algorithms to infrared temperature data, performs denoising and white balance processing on visible light images, and filters out high-frequency noise in terahertz signals in real time. The preprocessed multimodal data and information such as the probe's current position are packaged and sent to the multimodal registration and feature fusion module for further processing.

[0060] Multimodal registration and feature fusion module

[0061] This module receives terahertz signals, visible light images, and infrared thermal images from the imaging control and data acquisition module, and performs operations such as spatial alignment, intensity calibration, and multi-dimensional feature extraction and fusion. First, the geometric registration unit performs spatial registration of the terahertz intensity distribution based on the visible light and infrared thermal images. Since there may be differences in the field of view and installation position deviations among the imaging units, geometric registration requires estimating the mapping relationship between the terahertz imaging coordinates and the visible / infrared image coordinate system, for example, by performing coordinate transformation using a calibrated affine transformation matrix or a binocular correction model. Specifically, using the marker light spots or natural texture features projected onto the trauma area, image registration algorithms (such as ORB / SIFT feature matching) can be used to calculate the feature correspondence between the visible light and THz images, solving for the planar homography matrix, thereby resampling and aligning the terahertz intensity image to the visible / infrared image coordinate system. After geometric registration, it can be ensured that data from different modalities correspond to the same location in the trauma area under the same spatial reference.

[0062] Next, the intensity calibration process performs cross-modal intensity scale unification and environmental impact correction for physical quantities of different modes. The calibration unit uses parameters such as ambient temperature, humidity, wind speed, skin color, and subcutaneous fat thickness provided by the emergency environment and individual parameter acquisition module to correct and compensate for terahertz signal intensity, infrared temperature value, and visible light brightness.

[0063] Taking the terahertz channel as an example, terahertz waves are easily affected by water vapor absorption when propagating in air. The calibration unit calculates the attenuation factor η corresponding to the water vapor absorption coefficient based on the current relative humidity (RH) and the distance d from the probe to the wound surface. humidity ∈(0,1], and for the received original intensity value I THz,raw Compensation is performed to obtain the corrected terahertz intensity I. THz,cal :

[0064]

[0065] Where, η humidity The attenuation of terahertz signals decreases with increasing ambient humidity and propagation distance. When the relative humidity is high (e.g., RH=90%), the attenuation of the terahertz signal is more pronounced, at which point η... humidity <1, the system is multiplied by η −1 humidity To achieve compensation for signal amplitude; conversely, in a drier environment, η humidity ≈1, the correction factor has little effect on the intensity.

[0066] Similarly, ambient temperature and wind speed can affect the difference between the actual temperature of the wound surface and the infrared measurement: in low ambient temperatures and high wind speeds, the wound surface may cool rapidly, causing the infrared camera reading to be lower than the true temperature of the deeper tissue. The calibration unit can use a linear model to compensate for the infrared temperature.

[0067] Among them, T IR,raw T represents the raw measurements from the infrared camera. IR,cal The corrected infrared temperature; T env T represents the current ambient temperature. body,nom Normal human body surface temperature (approximately 37°C) ∘ C), v wind denoted as wind speed; k1 and k2 are empirical coefficients obtained from experimental calibration, used to characterize the influence of ambient temperature and wind speed on infrared temperature measurement error.

[0068] For the brightness value of the visible light channel, a corresponding calibration function can also be established based on factors such as ambient light intensity, skin color, and subcutaneous fat thickness, to adjust the original brightness L. VIS,raw Mapped to corrected brightness L VIS,cal This allows for comparability with terahertz intensity and infrared temperature on a unified intensity scale. For example, the brightness of skin or wounds in visible light images is significantly affected by skin color. The calibration unit estimates skin color type by acquiring the average color / brightness of the undamaged skin area of ​​the injured person, and then standardizes the brightness and color of visible light imaging accordingly (e.g., setting different correction parameter matrices for Caucasians, Asians, and Africans) to reduce the impact of skin color differences on image feature extraction. Furthermore, the thickness of body fat also significantly affects infrared and terahertz signals: areas with thicker fat layers have relatively lower infrared surface temperatures and shallower terahertz penetration depth. The calibration unit can compensate for data in corresponding areas based on a pre-established fat thickness influence model (e.g., measured by ultrasound or estimated based on BMI). Through the above environmental and individual parameter calibration, the consistency and comparability of multimodal data under different conditions can be improved.

[0069] After geometric registration and intensity calibration, multimodal data from various locations within the trauma region are integrated into a multidimensional feature vector. For example, for each pixel / voxel location p in the trauma image segmentation, a feature vector is constructed. , where I （1） THz (p),

[0070] I （2） THz(p) represents the terahertz reflection intensity of the first and second frequency bands (calibrated), TIR represents the infrared temperature, Lvis represents the visible light intensity, and Evis represents the wound edge features extracted from the visible light image. These vectors can be considered as multi-dimensional pixel features that fuse "terahertz-infrared-visible" information at this location. Furthermore, to enhance the resolution of terahertz and infrared imaging at wound boundary details, this module includes an edge-guided reconstruction unit: utilizing the clear wound edge contour in the visible light image, super-resolution reconstruction or interpolation correction is performed on the terahertz intensity distribution image and the infrared thermal image.

[0071] In one implementation, edge detection (e.g., using the Canny operator) is first performed on the visible light image to obtain a binary edge map B. VIS (x,y), where B is the location where the edge of the trauma is detected. VIS (x,y)=1, the rest of region B VIS (x,y)=0. When interpolating, upscaling, or reconstructing terahertz / infrared images, a constraint term based on this edge map is introduced to ensure that the reconstructed THz / IR image is within the B... VIS =1 indicates a strong gradient (corresponding to the trauma boundary), while unnecessary gradient changes are suppressed in non-edge regions.

[0072] Taking terahertz image reconstruction as an example, let the low-resolution terahertz image be I. THz,low (x,y), the high-resolution image to be obtained is I THz,up (x,y). Edge-guided reconstruction can be achieved by minimizing the following cost function.

[0073]

[0074] I ↓ THz,up (x,y) represents the high-resolution image I THz,up The values ​​taken on the low-resolution grid after downsampling are used to ensure that the reconstruction results are consistent with the original observations. THz,low Consistent;

[0075] ∇I THz,up (x,y) represents the spatial gradient of the high-resolution image at the position (x,y);

[0076] First item Constrain data fidelity;

[0077] Second item This is used to penalize excessive gradients in non-edge regions, making the image as smooth as possible in non-edge locations, while in B... VIS Edge positions where the gradient is equal to 1 allow for either gradient preservation or enhancement.

[0078] λ>0 is a weighting coefficient used to adjust the balance between data fidelity and edge smoothing terms.

[0079] By iteratively optimizing the cost function J (e.g., using gradient descent or iterative reconstruction algorithms), a high-resolution terahertz image I that balances data fidelity and edge structure constraints can be obtained. THz,up Similarly, this can be applied to infrared thermal images T... IR Using a similar edge-guided reconstruction model, leveraging the same B VIS (x,y) constrains the gradient structure of the infrared image at the wound boundary.

[0080] Through the aforementioned edge-guided reconstruction, this invention can significantly improve the spatial clarity of terahertz and infrared imaging in detailed areas such as wound boundaries, enabling subsequent multimodal fusion features to more accurately reflect the true shape and boundary location of the wound, and providing more reliable input for wound depth assessment, necrosis area prediction, and other tasks.

[0081] After completing the above processing, the multimodal registration and feature fusion module will output a set of multidimensional feature vectors for each location in the trauma area, as well as an overall fused feature map. These fused features will be passed to the scan path planning and execution module and the trauma severity assessment and grading module, respectively, for adaptive scan decision-making and intelligent assessment and classification.

[0082] Scan Path Planning and Execution Module

[0083] The scan path planning and execution module includes an actuator mechanically connected to the imaging probe assembly and a reinforcement learning-based control unit. The actuator, used to drive the imaging probe to move relative to the wound area, can be implemented in various ways. For example, a three-axis motorized motion platform can be used to move the probe along a planned path in a two-dimensional plane; or a lightweight robotic arm can be used to hold the probe and position it for scanning in three-dimensional space. For emergency portability, the actuator can be a battery-powered servo motor drive capable of moving the probe above the wound at, for example, a speed of 50 mm per second with a positioning accuracy of 1 mm. The actuator receives path control commands from the control unit, such as movement direction, step distance, and scanning speed, to precisely execute the scanning action.

[0084] The reinforcement learning control unit is the core of the scan path planning, responsible for making real-time decisions on the next scanning action based on the current multidimensional features and scan state. In this embodiment, the control unit models the scanning problem as a Markov Decision Process (MDP), where an agent (imaging probe) acquires maximum information through trial actions in the environment (trauma area). First, the state space S is defined, including the following elements:

[0085] spos: The current position coordinates of the imaging probe relative to the wound area (e.g., two-dimensional position (x,y) and probe tilt angle, etc.);

[0086] scov: The distribution of scanned regions, i.e., the scan coverage matrix, marking which locations have acquired multiple modal features;

[0087] sconf: The distribution of grading confidence scores output by the current trauma severity assessment and grading module, representing the confidence score of the severity assessment of each part of the trauma area or the overall grading result.

[0088] First, the motion space A is defined as the set of next motion / operation commands for the probe, for example:

[0089] A = {Move up Δd, move down Δd, move left Δd, move right Δd, perform local fine scan, end scan}

[0090] Where Δd is the single-step translation distance, "local fine scan" means to perform a higher resolution or higher frequency measurement near the current position, and "end scan" is used to terminate the scanning process in advance when the evaluation accuracy requirements are met.

[0091] At each decision time t, the agent determines the current state s based on the current state. t (Including probe location, distribution of scanned areas, current severity assessment results and their confidence levels, remaining time / energy budget, etc.) Select action a t ∈∈A, the environment transitions to a new state s t+1 The probe obtains new observations at the new location (the multimodal characteristics of that location and the updated evaluation results), and then calculates the results based on the reward function r(s). t ,a t ) Receive instant rewards t .

[0092] To improve the confidence of trauma severity assessment results and control scanning costs, the instantaneous reward function designed in this invention mainly consists of the following parts:

[0093] 1. Confidence assessment boost reward Rc

[0094] If an action increases the overall confidence level of the trauma severity grading, a positive reward is given, the magnitude of which is proportional to the increase in confidence level. Let Conf prev With Conf new Let each represent a certain aggregated confidence index before and after the action (e.g., the average confidence level of the entire trauma area, or the average confidence level of the current key suspicious area), then define...

[0095] Here, α>0 is the weighting coefficient, used to control the contribution of confidence improvement to the total reward. When the action does not lead to confidence improvement, Rc=0.

[0096] 2. Penalty Rr for repeated scans

[0097] To avoid excessive re-scanning of regions that already have high confidence, this invention includes a re-scanning penalty. Let n be the number of times the target region for the current action has already been scanned. scan The current confidence level for this region is Conf. target If the condition "the region has been scanned multiple times and the confidence level is higher than the threshold θ" is met, a fixed negative reward is given. For example, define an indicator function.

[0098] Other situations

[0099] Where θ is the confidence threshold (e.g., 0.9), n min If the minimum number of repetitions threshold is used, then the penalty for repeated scans can be written as:

[0100]

[0101] Where β>0 represents the penalty intensity. That is, when scanning a region that has reached high confidence and has been scanned multiple times, the agent will be subject to a fixed penalty, thus tending to switch to regions with higher information content.

[0102] 3. Time / Energy Consumption Penalty R t

[0103] To encourage high-quality assessments within a limited total time or energy budget, this invention introduces a time / energy consumption penalty. Let Δt be the time consumed for each scan step, then we can define...

[0104] Where γ>0 is the unit time penalty coefficient. If the cumulative scan time T cum Exceeding the preset budget T budget You can also add timeout penalties, for example...

[0105]

[0106] Where γ over >0 represents the timeout penalty coefficient. Similarly, corresponding energy consumption penalty items can be set for battery power consumption or probe terahertz emission dose, which will not be elaborated here.

[0107] 4. Comprehensive forms of instant rewards

[0108] In summary, in one implementation, the instantaneous reward r at time t is... t It can be represented as:

[0109] in I high,tThis is an indicator function, indicating whether the current action targets a high-confidence region. The ellipsis "..." can be used to represent other penalty terms such as energy consumption and probe dose. By appropriately setting α, β, γ, and γ... over These parameters can guide the agent to prioritize the scanning action with the highest information gain, while meeting the constraints of total scanning time and energy, thereby maximizing the reliability of the trauma severity assessment results.

[0110] 5. Strengthen learning strategies and training methods

[0111] The goal of the reinforcement learning control unit is to learn the optimal policy π under the above reward design. ∗ (s) makes the cumulative return The expected value is maximized. Training can be performed using value iteration or policy gradient algorithms. In one implementation, a Deep Q-Network (DQN) is used to approximate the state-action value function Q(s,a). Through repeated interactive training on a large amount of trauma model data or simulation scenarios, the network parameters converge to near-optimal levels. In another implementation, an Actor-Critic architecture can be used to simultaneously learn the policy network (Actor) and the value network (Critic), further improving training stability.

[0112] After training, the resulting reinforcement learning agent can make online decisions about scanning paths in practical applications: when an uneven distribution of the wound surface is detected or the current assessment is uncertain, the agent tends to choose actions such as "local fine scanning" to concentrate more measurements on that area to improve local confidence; for areas that have been scanned multiple times and whose severity assessment is stable, it tends to take actions such as "rapid translation" or "end scanning" to avoid repeated measurements. Experimental results show that compared with the uniform scanning strategy across the entire area, the adaptive scanning strategy of this invention can significantly reduce the total number of measurements under the same time budget, while improving the overall confidence of the severity assessment results for key areas.

[0113] The stability detection unit uses tilt and acceleration sensor data from the probe to determine the system's operational stability. If an external impact or drastic change in probe attitude is detected (e.g., a tilt angle change exceeding 5° or a momentary acceleration exceeding 0.5g), the current acquisition conditions are considered unstable. The stability detection unit immediately pauses data acquisition from the imaging control and data acquisition modules and notifies the operator to adjust the probe posture or wait for the environment to stabilize. The system only resumes scanning after the tilt angle and acceleration return to a safe range. Simultaneously, the stability detection unit also feeds this information back to the control unit so that reinforcement learning decisions can account for invalid data due to shaking (e.g., discarding interfered frames). Through these safety constraints and stability controls, the system can maintain the safety of the wounded in complex field environments and ensure the quality of acquired data for reliable subsequent assessments.

[0114] Emergency Environment and Individual Parameter Acquisition Module

[0115] This module is used to collect environmental parameters at the trauma scene and individual information of the injured person to assist the system in calibrating itself according to external conditions and individual differences. In this embodiment, the following sensing and acquisition units are included:

[0116] Temperature sensor: detects ambient temperature Tenv (e.g., using a digital thermometer with a range of −20 to 50°C and an accuracy of ±0.5°C).

[0117] Humidity sensor: detects the relative humidity (RH) of the environment (e.g., capacitive humidity sensor, range 0–100%RH, accuracy ±3%RH).

[0118] Wind speed sensor: detects the wind speed at the scene (e.g., hot wire anemometer, range 0-20m / s, accuracy ±0.1m / s). In windy emergency environments, wind has a significant impact on wound heat dissipation and sensor readings.

[0119] Skin color information acquisition unit: Acquires skin color information of the injured person. This can be achieved by capturing images of uninjured skin areas with a visible light camera, analyzing their RGB color values ​​or Laplace color parameters to determine skin color type (e.g., fair, medium, dark), or by using a dedicated skin color recognition sensor.

[0120] Subcutaneous fat thickness acquisition unit: This unit acquires information on the thickness of subcutaneous fat near the injury site. Subcutaneous fat layer thickness can be measured using an ultrasound probe, or its general distribution can be estimated based on the patient's height, weight, and BMI. In emergency situations where equipment is limited, medical personnel can also estimate the thickness based on the patient's body type and wound location (e.g., abdominal fat is approximately 10mm thick, while limbs have more muscle and less fat, approximately 5mm thick).

[0121] The ambient temperature, humidity, wind speed, and individual parameters such as skin color type and fat thickness obtained by the aforementioned acquisition module are sent to the calibration unit in the multimodal registration and feature fusion module for the multidimensional feature correction described above. By acquiring environmental and individual information in real time, the system can dynamically adjust the assessment model, exhibiting robustness to different environments and patient conditions. For example, in a hot and humid jungle environment, the system automatically corrects for THz signal attenuation and infrared baseline temperature; for patients with dark skin or obesity, the system adjusts the thresholds for visible / infrared feature interpretation accordingly to ensure fair and accurate assessment.

[0122] Trauma Severity Assessment and Grading Module

[0123] The trauma severity assessment and grading module intelligently assesses trauma severity based on multimodal fusion feature vectors and environmentally and individually corrected data, outputting grading results and re-transfer priority suggestions. This module typically includes a machine learning model or expert rule system to classify and make decisions based on the input features. In a preferred embodiment, a pre-trained deep neural network is used as the assessment model. For example, a multi-layer convolutional neural network or an integrated Transformer structure is constructed, taking a multi-dimensional feature map of the trauma area as input, and outputting pixel-level severity classification results (such as a trauma grading heatmap) and the overall severity level of the entire wound. The model can grade according to predefined standards, such as referring to clinical trauma scoring or burn grading standards, classifying injuries into mild, moderate, and severe levels, and providing corresponding re-transfer (medical evacuation) priorities (e.g., Level I requires immediate re-transfer, Level II can be delayed, Level III requires simple on-site treatment and re-transfer, etc.).

[0124] In burn scenarios, the grading module can be refined to identify wound depth categories (superficial first degree, superficial second degree, deep second degree, third degree, etc.) and assess the overall severity of the injury based on wound area. For complex traumas commonly encountered in emergency care, grading can be based on factors such as bleeding volume, extent of tissue necrosis, and infection risk. Regardless of the trauma type, the assessment model provides a confidence level for its output. For example, if using probability output, the model might give a 90% probability (high confidence) for "severe" vs. a 10% probability for "moderate," indicating high confidence; conversely, if the two most likely probabilities are close, the confidence level is low. The module can calculate the average confidence level or the lowest local confidence level for the entire trauma assessment, providing a reference for scanning decisions (i.e., the aforementioned sconf).

[0125] Furthermore, to accommodate situations where multiple injured individuals may be assessed simultaneously during emergency care, this module includes a multi-injury resource allocation unit. When multiple injured individuals require assessment at the scene, the system can perform a preliminary rapid scan on each individual (e.g., acquiring rough features of the main injury areas within 10 seconds of the initial scan). The multi-injury resource allocation unit intelligently allocates time for subsequent detailed scans based on each individual's initial classification and confidence level. The principle is to prioritize allocating resources to those with severe injuries and low initial assessment confidence to maximize overall treatment effectiveness. For example, after the initial scan, if Injury A is assessed as critically injured with only 60% confidence, Injury B as moderately injured with 90% confidence, and Injury C as slightly injured with 80% confidence, the system can allocate more detailed scan time to Injury A (e.g., giving him an additional 20 seconds for detailed scanning), require only a small supplementary scan for Injury B (e.g., 5 seconds to confirm key details), and Injury C, due to his minor injury and relatively certain assessment, does not require additional scanning. The resource allocation unit outputs the detailed scanning time quota and sequence plan for each injured person, and controls the scanning path planning and execution module to perform detailed scans on the injured persons in sequence according to the plan. Through this multi-injury scheduling strategy, the system can make reasonable use of limited equipment time in large-scale injury events, quickly complete the initial screening of multiple people and conduct in-depth assessments of key individuals, thereby optimizing emergency rescue decisions.

[0126] After the triage is completed, the trauma severity assessment and grading module will generate a triage result report, including the trauma severity level, description of the trauma location, model confidence level, and corresponding priority recommendations for evacuation. Priority recommendations can be automatically generated based on the triage results; for example, severe trauma corresponds to "priority evacuation (immediate medical attention)," moderate trauma corresponds to "secondary priority (brief on-site treatment followed by evacuation)," and mild trauma corresponds to "can be postponed (only simple treatment required or to be left for evacuation later)," etc. This result can be displayed on the terminal to provide emergency medical personnel with an objective basis for on-site treatment and subsequent evacuation arrangements.

[0127] Finally, the above implementation methods will be summarized and explained in conjunction with the workflow of this system.

[0128] System Workflow

[0129] The specific workflow of the system of this invention can be summarized as follows (refer to the illustrated workflow sequence):

[0130] Multimodal data acquisition (S1): The imaging probe assembly simultaneously acquires terahertz signals, visible light images, and infrared thermal images of the wound area. The probe can rapidly scan the wound surface along a predetermined path to obtain initial data, and send the acquired THz time-domain signals, visible light and infrared image frames to the imaging control and data acquisition module for synchronous recording.

[0131] Feature registration and fusion (S2): The terahertz signal, visible light image, and infrared thermal image acquired above are input into the multimodal registration and feature fusion module. After geometric registration and intensity calibration, the spatial position and amplitude values ​​of each modality data are corrected, and multidimensional feature vectors at each location of the trauma area are generated by fusion.

[0132] Environmental and Individual Correction (S3): The emergency environmental and individual parameter acquisition module measures the current ambient temperature, humidity, wind speed, as well as the injured person's skin color and subcutaneous fat thickness. Based on these environmental and individual parameters, the terahertz signal intensity, infrared temperature value, and visible light brightness in the multidimensional feature vector obtained in step S2 are corrected and compensated to eliminate the influence of environmental interference and individual differences on the features, thus obtaining a more accurate feature vector set.

[0133] Adaptive Scanning and Path Planning (S4): The corrected multidimensional feature vector and the current scan status (scanned area, current probe position, etc.) are input into the reinforcement learning control unit in the scan path planning and execution module. Based on this, the reinforcement learning unit outputs the next scan path control command (e.g., movement direction and distance), driving the actuator to move the imaging probe assembly to further scan the trauma area according to the command. During this process, if the reinforcement learning algorithm determines that a certain area has received a high confidence assessment, it reduces repeated scanning of that area; conversely, it intensifies scanning for areas with uncertain assessments. Simultaneously, it ensures that the total scan time does not exceed a predetermined constraint. If probe movement would cause an exceedance of a safety threshold (e.g., excessive pressure or temperature), the safety constraint unit adjusts or prevents the movement and applies penalty feedback to the control unit, thus preventing similar movements in the reinforcement learning strategy. The scanning process continues iteratively until a stopping condition is met (e.g., covering the entire trauma, confidence convergence, or timeout).

[0134] Intelligent Severity Assessment (S5): The final acquired multimodal fusion features of the entire trauma area are input into the trauma severity assessment and grading module. The model assesses the severity of the trauma and outputs the grading result and priority recommendations for subsequent treatment. If it is a single injured person, the process is complete at this point; if it is multiple injured persons, the system determines the scanning priority and time for the next injured person based on the preliminary assessment results through the multi-injury resource allocation unit, and returns to the loop of executing steps S1–S5 to complete the assessment of all injured persons in sequence. Finally, each injured person will receive an objective injury grading and corresponding treatment / transfer recommendations for decision-making reference.

[0135] Through the above steps, the system provided by this invention can perform rapid, objective, and highly accurate intelligent assessments of the trauma conditions of multiple injured individuals in complex environments such as emergency care. Actual testing shows that when using this system to scan and assess different types of trauma, such as burns and penetrating injuries, multimodal data fusion significantly improves the accuracy of the assessment. For example, for a 10cm × 10cm burn wound, the initial scan was completed within 30 seconds, and the system determined it to be a deep second-degree burn with only 70% confidence. Subsequently, the reinforcement learning control unit guided the probe to perform an additional 10 seconds of detailed scanning on the edge area suspected of being a third-degree burn, increasing the assessment confidence of that area to over 90%, ultimately determining the injury level as severe and recommending priority transfer. The entire assessment took less than one minute, while the accuracy rate of traditional manual visual assessment is only about 60–75%. In another scenario involving multiple injured individuals, the system quickly scanned three injured individuals sequentially, completing the initial screening in a total of two minutes. Additional time was automatically allocated for a focused re-examination of one of the suspected cases, and the final assessment results and transfer priority arrangements for the three individuals were highly consistent with the subsequent expert assessment. These specific embodiments verify the effectiveness of the present invention and demonstrate its great application value in emergency rescue.

[0136] In summary, the post-traumatic terahertz and optical multimodal assessment system and method of this invention achieves rapid, high-confidence assessment and intelligent classification of trauma in emergency patients by fusing THz, visible light, and infrared multi-source information and employing a reinforcement learning adaptive scanning strategy. Its detailed structural design, control flow, algorithm model, and implementation parameters are as described above, demonstrating sufficient disclosure and feasibility, and providing strong technical support for emergency medical decision-making.

Claims

1. A post-traumatic terahertz and optical multimodal assessment system, characterized in that, include: An imaging probe assembly, comprising a terahertz emitting unit, a terahertz receiving unit, a visible light imaging unit, and an infrared thermal imaging unit arranged around a common optical axis, wherein a sterile isolation window is provided at the front end of the imaging probe assembly. An imaging control and data acquisition module is electrically connected to the imaging probe assembly. It is used to control the terahertz emitting unit to emit terahertz waves and acquire the corresponding terahertz signals, as well as to acquire the image data output by the visible light imaging unit and the infrared thermal imaging unit. A multimodal registration and feature fusion module, which is connected to the imaging control and data acquisition module, is used to perform spatial registration and intensity calibration on the terahertz signal, visible light image and infrared thermal image to generate multidimensional feature vectors at each location of the trauma area; The scanning path planning and execution module includes an actuator mechanically connected to the imaging probe assembly and a reinforcement learning control unit connected to the multimodal registration and feature fusion module. The actuator is used to drive the imaging probe assembly to move relative to the trauma area, and the reinforcement learning control unit is used to output scanning path control commands based on the current multidimensional feature vector and the current scanning state. An emergency environment and individual parameter acquisition module is used to acquire environmental parameters at the trauma scene and individual parameters of the injured object. A trauma severity assessment and grading module, connected to the multimodal registration and feature fusion module and the emergency environment and individual parameter acquisition module, is used to intelligently assess the severity of trauma based on the multidimensional feature vector, environmental parameters, and individual parameters, and output grading results and subsequent priority suggestions. The reward function of the reinforcement learning control unit includes a reward term to improve the confidence of the trauma severity assessment results and a penalty term to constrain scan time and energy consumption, thereby improving the reliability of the trauma severity assessment results under predetermined scan time constraints.

2. The post-traumatic terahertz and optical multimodal assessment system according to claim 1, characterized in that, The terahertz transmitting unit is used to transmit terahertz pulses in multiple frequency bands, including a first frequency band and a second frequency band, wherein the center frequency of the first frequency band is lower than the center frequency of the second frequency band. The imaging probe assembly has a spatial coding structure at its front end, which is used to spatially modulate incident or reflected terahertz waves. The multimodal registration and feature fusion module includes a compressed sensing reconstruction unit, which is used to reconstruct the terahertz intensity distribution of the trauma area based on spatial coding measurement data.

3. The post-traumatic terahertz and optical multimodal assessment system according to claim 1, characterized in that, The state space of the reinforcement learning control unit includes: the current position of the imaging probe assembly relative to the trauma area, the distribution of areas that have been scanned, and the distribution of graded confidence output by the trauma severity assessment and grading module; the reward function of the reinforcement learning control unit is also used to increase the penalty value when repeatedly scanning areas that already have high graded confidence, so as to reduce repeated scanning of areas with high confidence.

4. The post-traumatic terahertz and optical multimodal assessment system according to claim 3, characterized in that, The imaging probe assembly is equipped with a contact pressure sensor and a temperature sensor at its front end. The scanning path planning and execution module is equipped with a safety constraint unit, which is connected to the reinforcement learning control unit. The safety constraint unit is used to: restrict the scanning path control command output by the reinforcement learning control unit when the contact pressure or local temperature exceeds a preset threshold, and to apply a penalty to the corresponding action in the reward function when the reinforcement learning control unit attempts to execute an action that exceeds the safety threshold.

5. The post-traumatic terahertz and optical multimodal assessment system according to claim 2, characterized in that, The multimodal registration and feature fusion module includes a geometric registration unit and an edge-guided reconstruction unit; The geometric registration unit is used to perform spatial registration of the terahertz intensity distribution based on the visible light image and the infrared thermal image; The edge-guided reconstruction unit is used to reconstruct the terahertz intensity distribution and the infrared thermal image using the trauma edge information in the visible light image, so as to improve the spatial resolution of the trauma boundary region.

6. The post-traumatic terahertz and optical multimodal assessment system according to claim 1, characterized in that, The emergency environment and individual parameter acquisition module includes a temperature sensor, a humidity sensor, and a wind speed sensor, as well as an individual information acquisition unit for acquiring skin color information and body fat thickness information of the injured person. The trauma severity assessment and grading module includes a calibration unit, which is used to correct the terahertz signal intensity, infrared temperature value, and visible light brightness in the multidimensional feature vector based on the temperature, humidity, wind speed, skin color information, and body fat thickness information.

7. The post-traumatic terahertz and optical multimodal assessment system according to claim 6, characterized in that, The trauma severity assessment and grading module includes a multi-casualty resource allocation unit, which is used for: After completing the initial rapid scan of multiple injured individuals, a fine scan time quota is allocated to each injured individual based on the initial grading results and grading confidence level. The scanning path planning and execution module is then controlled to perform fine scans on each injured individual sequentially according to the fine scan time quota.

8. The post-traumatic terahertz and optical multimodal assessment system according to claim 1, characterized in that, The sterile isolation window adopts a double-layer structure, which includes an outer heating and anti-condensation layer and an inner optical layer. The heating and anti-condensation layer is used to reduce water vapor condensation on the window surface in a high-humidity environment, and the optical layer is used to provide a stable transmission path for terahertz waves, visible light, and infrared light while maintaining optical performance. The imaging probe assembly is equipped with a tilt sensor and an acceleration sensor, and the scanning path planning and execution module is equipped with a stability detection unit. The stability detection unit is used to: pause the acquisition of the imaging control and data acquisition module and issue a prompt message to the operator when the tilt angle change detected by the tilt sensor or the acceleration detected by the acceleration sensor exceeds a preset threshold.

9. A method for emergency trauma assessment based on multimodal terahertz / optical imaging, characterized in that, Includes the following steps: S1, simultaneously acquire terahertz signals, visible light images and infrared thermal images of the trauma area through the imaging probe assembly, and send the terahertz signals, visible light images and infrared thermal images to the imaging control and data acquisition module; S2, input the terahertz signal, visible light image and infrared thermal image into the multimodal registration and feature fusion module, and obtain multidimensional feature vectors of each location in the trauma area after geometric registration and intensity calibration; S3, through the emergency environment and individual parameter acquisition module, collects environmental temperature, environmental humidity, wind speed, as well as skin color information and body fat thickness information of the injured person, and based on the environmental temperature, environmental humidity, wind speed, skin color information and body fat thickness information, corrects the terahertz signal intensity, infrared temperature value and visible light brightness in the multidimensional feature vector; S4, the corrected multidimensional feature vector and the current scanning status are input into the reinforcement learning control unit in the scanning path planning and execution module. The reinforcement learning control unit outputs scanning path control commands to drive the imaging probe assembly to scan the trauma area according to the scanning path. S5 inputs the collected multimodal features into the trauma severity assessment and grading module to assess the severity of the trauma and output the grading results and subsequent priority suggestions.

10. The emergency trauma assessment method based on multimodal terahertz / optical imaging according to claim 9, characterized in that, In step S4, the reward function used by the reinforcement learning control unit increases the reward value when the confidence of the trauma severity classification increases, increases the penalty value when repeatedly scanning areas that already have a high classification confidence, and increases the penalty value when the scanning time or energy consumption exceeds a preset threshold, so as to improve the reliability of the trauma severity classification results within the predetermined total scanning time constraint. When there are multiple injured persons, the procedure further includes: performing an initial rapid scan in steps S1 to S5 on each injured person to obtain an initial grading result and its grading confidence level; allocating a fine scan time quota to each injured person based on the initial grading result and its grading confidence level; and performing supplementary scans and assessments on each injured person in sequence according to the fine scan time quota.

Images