Wound analysis method and system based on multi-source perception and reinforcement learning

Abstract

The invention discloses a wound analysis method and system based on multi-source perception and reinforcement learning, and belongs to the field of wound analysis, and the method comprises the following steps: S1, respectively generating a three-dimensional wound model and an enhanced feature map; s2, determining a wound candidate area, and completing global coarse positioning of the wound; s3, generating a refined wound contour map in the wound candidate region; s4, the wound area is calculated, and time sequence stability verification is achieved through variance judgment; and S5, outputting a wound state evaluation result based on the three-dimensional wound model in combination with the wound area and the stability judgment result. By adopting the wound analysis method and system based on multi-source perception and reinforcement learning, the omnibearing, high-precision and intelligent analysis and decision support of the wound state are realized.

Description

Technical Field

[0001] This invention relates to the field of wound analysis technology, and in particular to a wound analysis method and system based on multi-source perception and reinforcement learning. Background Technology

[0002] Wound care is a core and critical task in clinical medicine and long-term care. Its core objective is to accurately assess the core condition of the wound (including key indicators such as area, depth, color, and tissue type) and to develop or dynamically adjust individualized treatment plans based on the assessment results.

[0003] With the rapid iteration of computer vision and artificial intelligence technologies, automated wound analysis technology based on machine vision has become a research hotspot in the field of medical intelligence. Its core objective is to provide medical staff with objective, efficient, and quantifiable wound assessment data and decision support, and reduce errors in human subjective judgment.

[0004] Currently, most mainstream automated wound analysis technologies rely on single or limited types of image sensors (such as ordinary RGB cameras) for research and development. Specifically, these technologies acquire wound images using mobile phones or fixed cameras, then use image segmentation algorithms such as U-Net and DeepLab to extract the wound region and calculate two-dimensional geometric parameters such as area and perimeter. Some studies introduce time-series data analysis methods to assess healing trends, fitting healing curves by comparing changes in image features at different time points to predict the healing process. At the recognition model level, there are also explorations into training dedicated deep learning classifiers or segmentation networks for specific wound types such as diabetic foot ulcers and burns.

[0005] However, the aforementioned existing technologies still face significant limitations and technical bottlenecks when applied in real and complex clinical scenarios, specifically in the following five core aspects: 1. Limited perceptual dimension and weak environmental resistance: Existing methods generally rely on single RGB visible light images for judgment. These images are easily affected by clinical environmental factors such as uneven lighting, shadow occlusion, bloodstains, and dressing coverage, leading to blurred wound boundaries and distorted color information, which severely restricts the accuracy of segmentation and feature extraction. More importantly, two-dimensional images cannot capture key three-dimensional information such as wound depth and volume, which plays an irreplaceable role in assessing wound severity (such as indentation depth) and healing progress.

[0006] 2. Lack of hierarchical analysis strategies, making it difficult to balance efficiency and accuracy: Existing algorithms mostly use a single end-to-end model for fine segmentation or recognition, resulting in a rigid processing flow. For high-resolution images or video streams, this strategy has a high computational load, making it difficult to meet the real-time requirements of clinical practice; moreover, when the image contains complex backgrounds or multiple wounds, the single model is easily affected by irrelevant regions, leading to segmentation errors or localization drift.

[0007] 3. Insufficient quantitative assessment accuracy and lack of temporal stability: Existing methods mostly calculate area based on pixel counting or simple contour fitting methods such as ellipse or rectangle approximation. This results in large approximation errors for irregular or blurred wound contours, leading to poor measurement stability. Furthermore, for the analysis results of the same wound across multiple consecutive frames or under different shooting conditions, existing technologies lack effective temporal consistency verification mechanisms, making it difficult to distinguish between actual wound changes and fluctuations caused by segmentation noise, significantly reducing the reliability of dynamic monitoring.

[0008] 4. Insufficient fusion of multi-source information, resulting in weak comprehensive decision-making ability: Wound assessment is a comprehensive judgment process that encompasses multiple dimensions, including edge clarity (contour), color (degree of inflammation, necrosis), depth (indentation), and tissue type (granulation tissue, necrotic tissue). Existing technologies typically extract only limited features from a single image modality, leading to unreliable judgment results.

[0009] 5. Rigid treatment strategies and lack of adaptive optimization capabilities: Current technologies cannot adjust treatment plans in a personalized and adaptive manner based on individual patient differences (such as physical condition and underlying diseases) and the dynamic evolution of the wound.

[0010] In summary, existing wound analysis technologies suffer from several core deficiencies, including limited perception dimensions, rigid analysis strategies, insufficient quantification accuracy, weak multi-source decision fusion capabilities, and a lack of adaptive decision optimization mechanisms. Summary of the Invention

[0011] The purpose of this invention is to provide a wound analysis method and system based on multi-source perception and reinforcement learning, thereby solving the aforementioned technical problems.

[0012] To achieve the above objectives, this invention provides a wound analysis method based on multi-source perception and reinforcement learning, comprising the following steps: S1. Simultaneously acquire binocular depth map, 3D depth map, multispectral image and RGB image of the wound area, perform pixel-level fusion of binocular depth map and 3D depth map, and feature-level fusion of multispectral image and RGB image to generate 3D wound model and enhanced feature map respectively. S2. Perform multi-level Haar wavelet transform on the enhanced feature map output by S1, extract low-frequency feature components, and run a lightweight target detection model to determine the candidate region of the wound and complete the global coarse localization of the wound. S3. Within the candidate wound region obtained in S2, extract the high-frequency feature components after wavelet transform, calculate the enhanced mask by combining low-frequency color correction information with the high-frequency image, multiply element-wise to enhance high-frequency details, and then use a refined sub-network based on a second-order Runge-Kutta module to optimize the contour and generate a refined wound contour map. S4. The refined contour output by S3 is processed by a splitting and merging algorithm based on sorting selection to extract key feature points, construct an approximate polygon and calculate the wound area using the shoelace formula; similarity matching is performed on the contours of multiple consecutive frames to generate a similarity sequence, and temporal stability is verified by variance judgment. S5. Based on the three-dimensional wound model generated by S1, combined with the wound area and stability judgment results output by S4, the wound status assessment result is output.

[0013] A system for a wound analysis method based on multi-source perception and reinforcement learning includes: The multi-source sensing module is used to simultaneously acquire binocular depth maps, 3D depth maps, multispectral images, and RGB images of the wound area. It performs pixel-level fusion of the binocular depth maps and 3D depth maps, and feature-level fusion of the multispectral images and RGB images to generate a 3D wound model and an enhanced feature map, respectively. The global localization module is used to perform multi-level Haar wavelet transform on the enhanced feature map, extract low-frequency feature components, run a lightweight target detection model, determine the candidate region of the wound, and complete the global coarse localization of the wound. The detail segmentation module is used to extract high-frequency feature components after wavelet transform in the candidate wound region. It calculates an enhanced mask by combining low-frequency color correction information with high-frequency image, and enhances high-frequency details by multiplying them element by element. Then, it uses a thinning sub-network based on the second-order Runge-Kutta module to optimize the contour and generate a refined wound contour map. The area calculation and correction module is used to extract key feature points from the refined contour using a sorting-based splitting and merging algorithm, construct an approximate polygon, and calculate the wound area using the shoelace formula; it performs similarity matching on contours of multiple consecutive frames to generate a similarity sequence, and verifies temporal stability through variance judgment. The decision fusion module is used to output wound status assessment results based on the three-dimensional wound model and the wound area and stability judgment results. The treatment strategy generation module is used to optimize the treatment strategy by constructing a Markov decision process and using the Soft Actor-Critic algorithm combined with priority experience playback and adaptive temperature parameter adjustment mechanism. The experience storage module stores the state, action, reward, and next state quadruple of each analysis into an experience library with a circular buffer structure. This continuously accumulates historical treatment data and patient feedback, which is used for iterative optimization of the reinforcement learning model to achieve adaptive upgrading of treatment strategies.

[0014] Therefore, the wound analysis method and system based on multi-source perception and reinforcement learning described above have the following beneficial effects: 1. Improve wound recognition and analysis accuracy: Through multi-source data fusion and a coarse-to-fine segmentation strategy, the system can achieve accurate positioning and contour refinement of the wound area. Combined with polygon approximation and stability verification, the error in wound area calculation can be controlled within 5%, which is significantly higher than traditional single-sensor or manual measurement methods. 2. Enhance the stability of dynamic monitoring and assessment: By adopting a temporal contour similarity matching and decision-level fusion mechanism, the system can identify stable changes in wound morphology during continuous monitoring, reducing misjudgments caused by changes in shooting angle and lighting, and is suitable for long-term wound healing tracking and remote nursing scenarios. 3. Achieve intelligent and personalized treatment decisions: Through a reinforcement learning framework, the system can automatically learn and optimize treatment plans based on historical wound data and patient feedback, ensuring healing effectiveness while taking into account treatment costs and patient comfort, improving the scientific nature and adaptability of clinical decision-making, and reducing the reliance on experience by medical staff. 4. Reduced manual intervention and operating costs: The system achieves fully automated analysis, reducing the workload of medical staff in manual measurement and recording. Combined with self-powered and low-power design, it is suitable for long-term, unattended wound monitoring scenarios, with significant economic and social benefits.

[0015] In summary, this invention integrates multi-source sensing data, adopts a hierarchical intelligent analysis strategy, achieves high-precision quantitative assessment, and possesses the ability to optimize decision-making based on historical data through self-learning, thereby comprehensively improving the intelligence, precision, and personalization of wound care and meeting the actual needs of clinical and remote care.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart of a wound analysis method based on multi-source perception and reinforcement learning according to the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0019] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0020] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] like Figure 1 As shown, a wound analysis method based on multi-source perception and reinforcement learning includes the following steps: S1. Simultaneously acquire binocular depth map, 3D depth map, multispectral image and RGB image of the wound area, perform pixel-level fusion of binocular depth map and 3D depth map, and feature-level fusion of multispectral image and RGB image to generate 3D wound model and enhanced feature map respectively. Step S1 specifically includes the following steps: S11. Simultaneously acquire binocular depth maps of the wound area using a binocular vision sensor, a TOF sensor, a multispectral sensor, and an RGB camera. 3D depth map Multispectral images and RGB images; S12, Integration; Weighted average method for stereo depth mapping With 3D depth map Perform registration and fusion: ; In the formula, This represents the depth map of the fused wound area. and These represent the weights corresponding to the binocular depth map and the 3D depth map, respectively. ,and , , Indicates the first The signal-to-noise ratio of data acquired by sensor-like devices Indicates the first Calibration error of sensor-like devices Indicates the first The signal-to-noise ratio of data acquired by sensor-like devices Indicates the first Calibration error of sensor-type sensors; Convolutional neural networks based on attention mechanism extract multispectral image features and RGB image features And perform weighted fusion: ; In the formula, This represents an enhanced feature map of the wound area. Represents attention weights, and , This represents the Sigmoid activation function. and All of these represent learning parameters.

[0022] S2. Perform multi-level Haar wavelet transform on the enhanced feature map output by S1, extract low-frequency feature components, and run a lightweight target detection model to determine the candidate region of the wound and complete the global coarse localization of the wound. In step S2, the enhanced feature map is... Multi-level Haar wavelet transform is performed to extract low-frequency feature components, resulting in a low-frequency image. : ; In the formula, Represents the transformation operator; Then in the low-frequency characteristic components Then, a lightweight object detection model is run to locate the candidate wound region, achieving global coarse localization.

[0023] S3. Within the candidate wound region obtained in S2, extract the high-frequency feature components after wavelet transform, calculate the enhanced mask by combining low-frequency color correction information with the high-frequency image, multiply element-wise to enhance high-frequency details, and then use a refined sub-network based on a second-order Runge-Kutta module to optimize the contour and generate a refined wound contour map. Step S3 specifically includes the following steps: S31, Based on the enhanced feature map in S2 The results of the multi-level Haar wavelet transform were used to extract high-frequency feature components from the candidate wound region, resulting in a high-frequency image. : ; in, ; ; ; In the formula, These represent the enhanced feature maps. High-frequency images in the horizontal, vertical, and diagonal directions obtained after Haar wavelet transform; These represent the enhancement feature maps respectively. When performing Haar wavelet transform, transform operators are used to extract high-frequency feature components in the horizontal, vertical, and diagonal directions. S32, call the low-frequency image obtained in S2. Simultaneously, color correction information is generated through low-frequency branch processing. and to and Perform upsampling operations separately to make the resolution of both images equal to that of the high-frequency image. Consistent; S33, Low-frequency image Color correction information With high-frequency images As input, through the feature transformation function Calculation yields high-frequency detail enhancement mask : ; In the formula, and These represent the low-frequency images respectively. and color correction information Perform upsampling; S34, transfer high-frequency images With enhancement mask Element-wise multiplication is performed to obtain the enhanced high-frequency detail image. : ; S35. Construct a refined sub-network based on a second-order Runge-Kutta module to enhance the high-frequency detail image. The input is a thinning subnetwork, which iteratively updates and optimizes the wound contour. The update expression is as follows: ; In the formula, and Representing the updated contour features and the first... Contour features of the next iteration; This represents a feature transformation function composed of convolutional layers and nonlinear activation functions; S36. After multiple rounds of iterative optimization and convergence, output the final refined wound contour map. .

[0024] S4. The refined contour output by S3 is processed by a splitting and merging algorithm based on sorting selection to extract key feature points, construct an approximate polygon and calculate the wound area using the shoelace formula; similarity matching is performed on the contours of multiple consecutive frames to generate a similarity sequence, and temporal stability is verified by variance judgment. Step S4 specifically includes the following steps: S41. Refining the wound outline Discretize the data to obtain a set of discrete points for the contour; S42. A splitting and merging algorithm based on sorting selection is used to filter feature points on the discretized contour curve, remove redundant points, and retain key feature points that can characterize the core shape of the contour to form a set of key feature points. S43. Sort the selected key feature points in a clockwise direction to construct an approximate polygon. , These represent the approximate polygons obtained by sorting the key feature points of the refined wound contour in a clockwise order. The 1st to the 1st vertices, and , Indicates the total number of discrete points on the contour; S44. Achieve a balance between polygon approximation and matching by optimizing the objective function; S441, Calculate polygons With refined wound contour Approximation error : ; In the formula, Representing a refined wound outline Any discrete point on; Representing a refined wound outline Discrete points to polygons The shortest distance; S442. Verify the approximation error Is it less than the set threshold? If not, return to S41; otherwise, execute S443. S443. Calculate the objective function. : ; In the formula, and They represent the first and Detailed wound contours collected at various time points; S444, Select to make The largest set of polygon vertices, which serves as the final approximating polygon; S45. Calculate the wound area using the shoelace formula based on the vertex coordinates of the final approximating polygon. : ; In the formula, and They represent the first and second digits of the approximating polygon. The coordinates of the first vertex, the first vertex The coordinates of the vertices; S46. Collect a refined contour sequence of the same wound at consecutive time points. And select the contour with the most points in the contour sequence as the initial matching reference contour; S47. Use dynamic programming to search for matching polygon vertices of adjacent contours. The goal is to maximize the similarity between adjacent polygons. The similarity is calculated by summing the dot products of the unit vectors of corresponding edges of adjacent polygons (the dot product value ranges from 1 to 2). (where 1 indicates a perfect match, and -1 indicates a complete reversal). ; in, ; In the formula, express and Matching similarity between them; Represents the approximation polygon Upper The unit vector corresponding to each edge; Indicates the first Approximate polygons at each time point Upper The unit vector corresponding to each edge; Represents the approximation polygon Upper The coordinates of the vertices; Represents the approximation polygon Upper The coordinates of the vertices; S48. Based on the similarity calculation results of each adjacent contour, generate a similarity sequence. ; S49. Calculate similarity sequences If the variance of the similarity is lower than a preset threshold, If the time sequence of the refined wound contour map is stable, it is considered that the time sequence of the refined wound contour map is stable; otherwise, it indicates that the refined wound contour map is abnormally changing.

[0025] S5. Based on the three-dimensional wound model generated by S1, combined with the wound area and stability judgment results output by S4, the wound status assessment result is output.

[0026] In step S5, the DS evidence theory is used to perform decision fusion on the three-dimensional wound model generated in S1 and the wound area and stability judgment results output in step S4: ; In the formula, In the DS evidence theory, the identification frame is represented. The basic probability assignment value of proposition A; and This represents the proposition corresponding to two independent pieces of evidence in the DS evidence theory, representing the judgment results of the sensor and the feature, respectively; and This indicates that the first piece of evidence supports the proposition. The basic probability distribution and the second piece of evidence for the proposition The basic probability allocation value; Represents the conflict coefficient, and ; Representation of recognition framework The empty set in; The output wound status includes wound type, severity, and healing stage, where wound type... Severity healing stage .

[0027] In this embodiment, in step S5, the multi-sensor outputs can also be nonlinearly fused based on the fuzzy integral fusion method and combined with Choquet integral: ; In the formula, Indicates the fusion result; This represents a fuzzy measure function that reflects the importance and interaction relationships of each sensor; Indicates the first Output data from each sensor; The integral variable representing the fuzzy integral; Step S5 is followed by: S6, by constructing a Markov decision process, the treatment strategy is optimized by using the Soft Actor-Critic algorithm combined with priority experience playback and adaptive temperature parameter adjustment mechanism. Specifically, it includes the following steps: S61. Define the wound state output by S5 as the state space. (Including features such as edge clarity and degree of inflammation), the treatment plan is motion space. (Treatment parameter combinations (drug type, dosage, dressing change frequency, etc.)) define healing effect Treatment costs With patient comfort The weighted combination is the reward function. : , , and Both represent weights, and ; S62. Initialize the network structure; Actor Network: Output Policy Distribution Specific treatment actions are generated through reparameterization. The formula is: ; In the formula, This represents the action generation function of the Actor network; Indicates the mean of the action; Indicates random noise, and ; Indicates the standard deviation of the action; Critic network: contains 2 independent Q networks and and the corresponding target Q-network and This is used to reduce the bias in motion value estimation; Target Value Network: Calculates the target state value using the following formula: ; In the formula, Indicates the target state The corresponding target state value; The parameter is The target Q-network for the target state and actions Value estimation; Indicates temperature parameter; The parameter is The Actor network in state Down Output Action The strategy distribution; Its state transition probability satisfies: ; In the formula, Indicates the current state is Execution of actions In this case, the system transitions to the next state. And receive a reward The state transition probability; express Always in a state of readiness Execution of actions At that time, the first Time to transition to state And receive a reward The state transition probability; S63. Employing the maximum entropy reinforcement learning objective function : ; In the formula, Represents the mathematical expectation operator; Discount factor of Power; express Instant rewards for each moment; Indicates temperature parameter; Represents policy entropy; express Moment State The strategy distribution under; S64. Calculate network loss; Strategy Loss : ; Q-value loss : ; Value loss: Based on the Critic network output and temperature parameters, the value estimation error is minimized through gradient descent to ensure that the target value is consistent with the Q-value network output; In the formula, Indicates the relationship with the experience base Mid-sampling Moment State Calculate the statistical average; The parameter is Actor networks in Moment State Down Output Action The strategy distribution; The parameter is The Critic network Moment State and actions Value estimation; Indicates the relationship with the experience base Mid-sampled quadruplets Calculate the statistical average; The parameter is The Critic network Moment State and actions Value estimation; express Moment State The corresponding target state value; S66. Initialize the experience base: an experience base for building circular buffer structures. Store historical interaction data quadruple This is used for non-policy learning to improve sample utilization. S67. Based on timing difference error Assign sampling priority: ; In the formula, Indicates the Critic network's... Moment State and actions Value estimation; S68. Train repeatedly according to the following process: S681, From the experience base The batch data is sampled according to the priority assigned in step S67; S682. Calculate Q-value loss, strategy loss, and value loss based on sampled data; S683. Update the parameters of the Critic network and Actor network using gradient descent. S684. Update the parameters of the target Q network and the target value network using a soft update strategy; S685. Determine whether the convergence condition is met. If yes, exit the loop; otherwise, return to step S681. S69. Based on the preset target entropy Optimize temperature parameters using a loss function: ; In the formula, Representing temperature parameters The loss function; S610. The optimal strategy distribution for the current wound state is output through the Actor network, and the combination of treatment parameters with the highest probability is selected as the final optimization scheme.

[0028] In step S6, the four-tuple of state, action, reward, and next state for each analysis is stored in the experience base of the circular buffer structure to continuously accumulate historical treatment data and patient feedback for iterative optimization of the reinforcement learning model and to achieve adaptive upgrading of the treatment strategy.

[0029] A system for performing a wound analysis method based on multi-source perception and reinforcement learning includes: The multi-source sensing module is used to simultaneously acquire binocular depth maps, 3D depth maps, multispectral images, and RGB images of the wound area. It performs pixel-level fusion of the binocular depth maps and 3D depth maps, and feature-level fusion of the multispectral images and RGB images to generate a 3D wound model and an enhanced feature map, respectively. The global localization module is used to perform multi-level Haar wavelet transform on the enhanced feature map, extract low-frequency feature components, run a lightweight target detection model, determine the candidate region of the wound, and complete the global coarse localization of the wound. The detail segmentation module is used to extract high-frequency feature components after wavelet transform in the candidate wound region. It calculates an enhanced mask by combining low-frequency color correction information with high-frequency image, and enhances high-frequency details by multiplying them element by element. Then, it uses a thinning sub-network based on the second-order Runge-Kutta module to optimize the contour and generate a refined wound contour map. The area calculation and correction module is used to extract key feature points from the refined contour using a sorting-based splitting and merging algorithm, construct an approximate polygon, and calculate the wound area using the shoelace formula; it performs similarity matching on contours of multiple consecutive frames to generate a similarity sequence, and verifies temporal stability through variance judgment. The decision fusion module is used to output wound status assessment results based on the three-dimensional wound model and the wound area and stability judgment results. The treatment strategy generation module is used to optimize the treatment strategy by constructing a Markov decision process and using the Soft Actor-Critic algorithm combined with priority experience playback and adaptive temperature parameter adjustment mechanism. The experience storage module stores the state, action, reward, and next state quadruple of each analysis into an experience library with a circular buffer structure. This continuously accumulates historical treatment data and patient feedback, which is used for iterative optimization of the reinforcement learning model to achieve adaptive upgrading of treatment strategies.

[0030] Experimental Example Hardware configuration: It adopts an NVIDIA RTX 4090 GPU, an Intel Core i9-13900K CPU, and 64GB of RAM to provide sufficient computing resources for experiments.

[0031] Software and Simulation Environment: Software Environment: Based on the Python 3.9 programming language, coupled with the PyTorch 2.0 deep learning framework and the OpenCV 4.8 computer vision library. The simulation environment is built using the Gym library. Sensor Simulation: The collaborative data acquisition process of a binocular vision sensor, a TOF depth sensor, a multispectral sensor, and an RGB camera is simulated using the Blender and PyBullet engines to generate a synthetic dataset that includes noise, illumination changes, and different shooting angles.

[0032] Experimental Datasets: (I) Synthetic Dataset: Using the simulation environment described above, a dataset of 5000 multi-source wound image sequences was generated, covering four common wound types: ulcers, burns, incisions, and pressure sores. Each wound type includes samples of three severity levels (mild, moderate, and severe) and three healing stages (inflammatory, proliferative, and maturation phases), comprehensively covering common clinical wound scenarios. (II) Validation on Public Datasets: To verify the generalization ability of this invention, additional validation tests were conducted on samples from the Medetec public wound image database and the Fiona Stanley Hospital diabetic foot ulcer dataset.

[0033] Comparison Methods: Benchmark Method 1 (Single RGB Segmentation): Employs a U-Net network for end-to-end wound segmentation and calculates the wound area using ellipse fitting. Benchmark Method 2 (Simple Multi-Sensor Fusion): Integrates multi-source data using early fusion (channel stitching) and performs the segmentation task using a DeepLabV3+ network. Benchmark Method 3 (No Reinforcement Learning Optimization): Uses the same multi-source perception and segmentation process as this invention, but generates treatment suggestions based on fixed rules (e.g., based on area change thresholds) without involving reinforcement learning optimization.

[0034] Evaluation Metrics: Segmentation Accuracy: Wound region segmentation accuracy is measured using the Dice coefficient (DSC) and Intersection over Union (IOU). Area Calculation Accuracy: Area calculation accuracy is evaluated using relative error (RE) based on measurements from a high-precision 3D scanner. Temporal Stability: Temporal stability is reflected by the standard deviation (Std) of area calculations between consecutive frames. Decision Fusion Performance: The effectiveness of decision fusion is evaluated by the classification accuracy for wound type, severity, and healing stage. Policy Optimization Performance: Policy optimization performance is measured by the cumulative reward during reinforcement learning training and the policy convergence speed.

[0035] Experiment 1: Comparison of wound region segmentation accuracy; The segmentation performance of the proposed method (multi-source fusion + coarse-to-fine segmentation) was compared with that of two benchmark methods on the test set.

[0036] Table 1 Comparison Results

[0037] As shown in Table 1, the average DSC and IOU of the method of this invention reach 92.9% and 87.0%, respectively, which are significantly better than the comparative methods. This result verifies the rich features provided by multi-source data fusion, as well as the effectiveness of the coarse-to-fine segmentation strategy of "low-frequency coarse localization + high-frequency detail enhancement + second-order Runge-Kutta module optimization" in dealing with complex and blurred wound boundaries, and can accurately extract the wound region.

[0038] Experiment 2: Analysis of the accuracy and temporal stability of wound area calculation; The area calculation accuracy and temporal stability of each method were evaluated in a dynamic simulation sequence (simulating the healing process of the same wound over 30 days, a total of 300 frames).

[0039] Table 2 Evaluation Results

[0040] As shown in Table 2, the method of this invention reduces the average error of area calculation to 4.3% and the standard deviation of temporal fluctuation to only 2.1%, which is significantly better than traditional methods. This is due to the high-precision approximation polygon constructed by the split-merge algorithm based on sorting selection, and the temporal stability verification mechanism achieved by continuous multi-frame contour similarity matching and variance judgment, which effectively filters out measurement noise and ensures the reliability and clinical reference value of long-term monitoring data.

[0041] Experiment 3: Performance evaluation of multi-source decision fusion; The accuracy of the decision-level fusion module based on DS evidence theory in judging wound type, severity, and healing stage was tested.

[0042] Table 3 Test Results

[0043] As shown in Table 3, the decision-level fusion method of this invention achieves an accuracy rate of over 90% in all assessment tasks. The DS evidence theory effectively handles the uncertainty and conflict between multi-sensor information (the conflict coefficient K is reduced by an average of 40%). Compared with single-modal decision-making or simple voting fusion, it significantly improves the robustness and reliability of the comprehensive judgment, providing accurate wound condition assessment results for clinical practice, thus verifying the effectiveness of this invention.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A wound analysis method based on multi-source perception and reinforcement learning, characterized in that: Includes the following steps: S1. Simultaneously acquire binocular depth map, 3D depth map, multispectral image and RGB image of the wound area, perform pixel-level fusion of binocular depth map and 3D depth map, and feature-level fusion of multispectral image and RGB image to generate 3D wound model and enhanced feature map respectively. S2. Perform multi-level Haar wavelet transform on the enhanced feature map output by S1, extract low-frequency feature components, and run a lightweight target detection model to determine the candidate region of the wound and complete the global coarse localization of the wound. S3. Within the candidate wound region obtained in S2, extract the high-frequency feature components after wavelet transform, calculate the enhanced mask by combining low-frequency color correction information with the high-frequency image, multiply element-wise to enhance high-frequency details, and then use a refined sub-network based on a second-order Runge-Kutta module to optimize the contour and generate a refined wound contour map. S4. The refined contour output by S3 is processed by a splitting and merging algorithm based on sorting selection to extract key feature points, construct an approximate polygon, and calculate the wound area using the shoelace formula. Similarity matching is performed on contours of multiple consecutive frames to generate a similarity sequence, and temporal stability is verified by variance judgment; S5. Based on the three-dimensional wound model generated by S1, combined with the wound area and stability judgment results output by S4, the wound status assessment result is output.

2. The wound analysis method based on multi-source perception and reinforcement learning according to claim 1, characterized in that: Step S1 specifically includes the following steps: S11. Simultaneously acquire binocular depth maps of the wound area using a binocular vision sensor, a TOF sensor, a multispectral sensor, and an RGB camera. 3D depth map Multispectral images and RGB images; S12, Integration; Weighted average method for stereo depth mapping With 3D depth map Perform registration and fusion: ； In the formula, This represents the depth map of the fused wound area. and These represent the weights corresponding to the binocular depth map and the 3D depth map, respectively. ,and , , Indicates the first The signal-to-noise ratio of data acquired by sensor-like devices Indicates the first Calibration error of sensor-like devices Indicates the first The signal-to-noise ratio of data acquired by sensor-like devices Indicates the first Calibration error of sensor-type sensors; Convolutional neural networks based on attention mechanism extract multispectral image features and RGB image features And perform weighted fusion: ； In the formula, This represents an enhanced feature map of the wound area. Represents attention weights, and , This represents the Sigmoid activation function. and All of these represent learning parameters.

3. The wound analysis method based on multi-source perception and reinforcement learning according to claim 2, characterized in that: In step S2, the enhanced feature map is... Multi-level Haar wavelet transform is performed to extract low-frequency feature components, resulting in a low-frequency image. : ； In the formula, Represents the transformation operator; Then in the low-frequency characteristic components Then, a lightweight object detection model is run to locate the candidate wound region, achieving global coarse localization.

4. The wound analysis method based on multi-source perception and reinforcement learning according to claim 3, characterized in that: Step S3 specifically includes the following steps: S31, Based on the enhanced feature map in S2 The results of the multi-level Haar wavelet transform were used to extract high-frequency feature components from the candidate wound region, resulting in a high-frequency image. : ； in, ; ; ; In the formula, These represent the enhanced feature maps. High-frequency images in the horizontal, vertical, and diagonal directions obtained after Haar wavelet transform; These represent the enhancement feature maps respectively. When performing Haar wavelet transform, transform operators are used to extract high-frequency feature components in the horizontal, vertical, and diagonal directions. S32, call the low-frequency image obtained in S2. Simultaneously, color correction information is generated through low-frequency branch processing. and to and Perform upsampling operations separately to make the resolution of both images equal to that of the high-frequency image. Consistent; S33, Low-frequency image Color correction information With high-frequency images As input, through the feature transformation function Calculation yields high-frequency detail enhancement mask : ； In the formula, and These represent the low-frequency images respectively. and color correction information Perform upsampling; S34, transfer high-frequency images With enhancement mask Element-wise multiplication is performed to obtain the enhanced high-frequency detail image. : ； S35. Construct a refined sub-network based on a second-order Runge-Kutta module to enhance the high-frequency detail image. The input is a thinning subnetwork, which iteratively updates and optimizes the wound contour. The update expression is as follows: ； In the formula, and Representing the updated contour features and the first... Contour features of the next iteration; This represents a feature transformation function composed of convolutional layers and nonlinear activation functions; S36. After multiple rounds of iterative optimization and convergence, output the final refined wound contour map. .

5. The wound analysis method based on multi-source perception and reinforcement learning according to claim 4, characterized in that: Step S4 Specifically, the following steps are included: S41. Refining the wound outline Discretize the data to obtain a set of discrete points for the contour; S42. A splitting and merging algorithm based on sorting selection is used to filter feature points on the discretized contour curve, remove redundant points, and retain key feature points that can characterize the core shape of the contour to form a set of key feature points. S43. Sort the selected key feature points in a clockwise direction to construct an approximate polygon. , These represent the approximate polygons obtained by sorting the key feature points of the refined wound contour in a clockwise order. The 1st to the 1st vertices, and , Indicates the total number of discrete points on the contour; S44. Achieve a balance between polygon approximation and matching by optimizing the objective function; S441, Calculate polygons With refined wound contour Approximation error : ； In the formula, Representing a refined wound outline Any discrete point on; Representing a refined wound outline Discrete points to polygons The shortest distance; S442. Verify the approximation error Is it less than the set threshold? If not, return to S41; otherwise, execute S443. S443. Calculate the objective function. : ； In the formula, and They represent the first and Detailed wound contours collected at various time points; S444, Select to make The largest set of polygon vertices, which serves as the final approximating polygon; S45. Calculate the wound area using the shoelace formula based on the vertex coordinates of the final approximating polygon. : ； In the formula, and They represent the first and second digits of the approximating polygon. The coordinates of the first vertex, the first vertex The coordinates of the vertices; S46. Collect a refined contour sequence of the same wound at consecutive time points. And select the contour with the most points in the contour sequence as the initial matching reference contour; S47. Use dynamic programming to search for matching polygon vertices of adjacent contours, with the goal of maximizing the similarity between adjacent polygons. The similarity is calculated by summing the dot products of the unit vectors of the corresponding sides of the adjacent polygons. ； in, ； In the formula, express and Matching similarity between them; Represents the approximation polygon Upper The unit vector corresponding to each edge; Indicates the first Approximate polygons at each time point Upper The unit vector corresponding to each edge; Represents the approximation polygon Upper The coordinates of the vertices; Represents the approximation polygon Upper The coordinates of the vertices; S48. Based on the similarity calculation results of each adjacent contour, generate a similarity sequence. ; S49. Calculate similarity sequences If the variance of the similarity is lower than a preset threshold, If the time sequence of the refined wound contour map is stable, it is considered that the time sequence of the refined wound contour map is stable; otherwise, it indicates that the refined wound contour map is abnormally changing.

6. The wound analysis method based on multi-source perception and reinforcement learning according to claim 5, characterized in that: In step S5, the DS evidence theory is used to perform decision fusion on the three-dimensional wound model generated in S1 and the wound area and stability judgment results output in step S4: ； In the formula, In the DS evidence theory, the identification frame is represented. The basic probability assignment value of proposition A; and This represents the proposition corresponding to two independent pieces of evidence in the DS evidence theory, representing the judgment results of the sensor and the feature, respectively; and This indicates that the first piece of evidence supports the proposition. The basic probability distribution and the second piece of evidence for the proposition The basic probability allocation value; Represents the conflict coefficient, and ; Representation of recognition framework The empty set in; The output wound status includes wound type, severity, and healing stage, where wound type... Severity healing stage .

7. The wound analysis method based on multi-source perception and reinforcement learning according to claim 5, characterized in that: Step S5 is followed by: S6, by constructing a Markov decision process, the treatment strategy is optimized by using the Soft Actor-Critic algorithm combined with priority experience playback and adaptive temperature parameter adjustment mechanism. Specifically, it includes the following steps: S61. Define the wound state output by S5 as the state space. The treatment plan is to improve the space of motion. Define the healing effect Treatment costs With patient comfort The weighted combination is the reward function. : , , and Both represent weights, and ; S62. Initialize the network structure; Actor Network: Output Policy Distribution Specific treatment actions are generated through reparameterization. The formula is: ； In the formula, This represents the action generation function of the Actor network; Indicates the mean of the action; Indicates random noise, and ; Indicates the standard deviation of the action; Critic network: contains 2 independent Q networks and and the corresponding target Q-network and This is used to reduce the bias in motion value estimation; Target Value Network: Calculates the target state value using the following formula: ； In the formula, Indicates the target state The corresponding target state value; The parameter is The target Q-network for the target state and actions Value estimation; Indicates temperature parameter; The parameter is The Actor network in state Down Output Action The strategy distribution; Its state transition probability satisfies: ； In the formula, Indicates the current state is Execution of actions In this case, the system transitions to the next state. And receive a reward The state transition probability; express Always in a state of readiness Execution of actions At that time, the first Time to transition to state And receive a reward The state transition probability; S63. Employing the maximum entropy reinforcement learning objective function : ； In the formula, Represents the mathematical expectation operator; Discount factor of Power; express Instant rewards for each moment; Indicates temperature parameter; Represents policy entropy; express Moment State The strategy distribution under; S64. Calculate network loss; Strategy Loss : ; Q-value loss : ; Value loss: Based on the Critic network output and temperature parameters, the value estimation error is minimized through gradient descent to ensure that the target value is consistent with the Q-value network output; In the formula, Indicates the relationship with the experience base Mid-sampling Moment State Calculate the statistical average; The parameter is Actor networks in Moment State Down Output Action The strategy distribution; The parameter is The Critic network Moment State and actions Value estimation; Indicates the relationship with the experience base Mid-sampled quadruplets Calculate the statistical average; The parameter is The Critic network Moment State and actions Value estimation; express Moment State The corresponding target state value; S66. Initialize the experience base: an experience base for building circular buffer structures. Store historical interaction data quadruple This is used for non-policy learning to improve sample utilization. S67. Based on timing difference error Assign sampling priority: ； In the formula, Indicates the Critic network's... Moment State and actions Value estimation; S68. Train repeatedly according to the following process: S681, From the experience base The batch data is sampled according to the priority assigned in step S67; S682. Calculate Q-value loss, strategy loss, and value loss based on sampled data; S683. Update the parameters of the Critic network and Actor network using gradient descent. S684. Update the parameters of the target Q network and the target value network using a soft update strategy; S685. Determine whether the convergence condition is met. If yes, exit the loop; otherwise, return to step S681. S69. Based on the preset target entropy Optimize temperature parameters using a loss function: ； In the formula, Representing temperature parameters The loss function; S610. The optimal strategy distribution for the current wound state is output through the Actor network, and the combination of treatment parameters with the highest probability is selected as the final optimization scheme.

8. The wound analysis method based on multi-source perception and reinforcement learning according to claim 7, characterized in that: In step S6, the four-tuple of state, action, reward, and next state for each analysis is stored in the experience base of the circular buffer structure to continuously accumulate historical treatment data and patient feedback for iterative optimization of the reinforcement learning model and to achieve adaptive upgrading of the treatment strategy.

9. A system for performing the wound analysis method based on multi-source perception and reinforcement learning as described in claim 8, characterized in that: include: The multi-source sensing module is used to simultaneously acquire binocular depth maps, 3D depth maps, multispectral images, and RGB images of the wound area. It performs pixel-level fusion of the binocular depth maps and 3D depth maps, and feature-level fusion of the multispectral images and RGB images to generate a 3D wound model and an enhanced feature map, respectively. The global localization module is used to perform multi-level Haar wavelet transform on the enhanced feature map, extract low-frequency feature components, run a lightweight target detection model, determine the candidate region of the wound, and complete the global coarse localization of the wound. The detail segmentation module is used to extract high-frequency feature components after wavelet transform in the candidate wound region. It calculates an enhanced mask by combining low-frequency color correction information with high-frequency image, and enhances high-frequency details by multiplying them element by element. Then, it uses a thinning sub-network based on the second-order Runge-Kutta module to optimize the contour and generate a refined wound contour map. The area calculation and correction module is used to extract key feature points from the refined contour using a sorting-selection splitting and merging algorithm, construct an approximate polygon, and calculate the wound area using the shoelace formula. Similarity matching is performed on contours of multiple consecutive frames to generate a similarity sequence, and temporal stability is verified by variance judgment; The decision fusion module is used to output wound status assessment results based on the three-dimensional wound model and the wound area and stability judgment results. The treatment strategy generation module is used to optimize the treatment strategy by constructing a Markov decision process and using the Soft Actor-Critic algorithm combined with priority experience playback and adaptive temperature parameter adjustment mechanism. The experience storage module stores the state, action, reward, and next state quadruple of each analysis into an experience library with a circular buffer structure. This continuously accumulates historical treatment data and patient feedback, which is used for iterative optimization of the reinforcement learning model to achieve adaptive upgrading of treatment strategies.

Images