A method, apparatus and device for predicting a wound healing process

Abstract

The embodiment of the application discloses a kind of prediction method, device and equipment of wound healing process, the method comprises: obtaining target image sequence;Using the first module of pre-trained prediction model, respectively on each target image in the target image sequence Local feature extraction and global feature extraction processing are carried out, to obtain local feature vector and global feature vector;Using the second module of the pre-trained prediction model, based on different preset extraction period, respectively on the local feature vector and the global feature vector Time series feature extraction processing is carried out, to obtain multiple time series feature vectors;Using the pre-trained prediction model, based on the multiple time series feature vectors, generate healing condition prediction information for the wound site;According to the healing condition prediction information, determine treatment plan.

Description

Technical Field

[0001] This invention relates to the field of life science technology, and in particular to a method, apparatus and device for predicting the wound healing process. Background Technology

[0002] Currently, in primary care hospitals, rehabilitation facilities, and home care settings, wound healing prediction primarily relies on the experience of healthcare workers and visual observation, supplemented by measuring wound dimensions. This method of prediction is not only time-consuming and labor-intensive but also carries the risks of subjective bias and cross-infection. Therefore, there is a need for a technical solution that utilizes electro-digital data processing technology to objectively process wound images, thereby improving the accuracy of wound healing prediction and ultimately enhancing wound care outcomes. Existing patents such as CN115426939A "Machine Learning System and Method for Wound Assessment, Healing Prediction, and Treatment" and CN116798024A "A Rapid Grading and Recognition Method and System for Burn Images Based on Artificial Intelligence" both attempt to process and predict wound images using machine learning or artificial intelligence, reflecting the application trend of electro-digital data processing technology in this field. Summary of the Invention

[0003] The purpose of this invention is to provide a technical solution that improves wound care by increasing the accuracy of wound healing prediction.

[0004] To solve the above-mentioned technical problems, the embodiments of the present invention are implemented as follows:

[0005] In a first aspect, embodiments of the present invention provide a method for predicting the wound healing process, the method comprising:

[0006] A target image sequence is acquired, which consists of target images at different time points containing the user's trauma site and a standard reference object;

[0007] Using the first module of the pre-trained prediction model, local feature extraction and global feature extraction are performed on each target image in the target image sequence to obtain local feature vectors and global feature vectors.

[0008] Using the second module of the pre-trained prediction model, based on different preset extraction periods, temporal feature extraction processing is performed on the local feature vector and the global feature vector respectively to obtain multiple temporal feature vectors;

[0009] Using the pre-trained prediction model, based on the multiple temporal feature vectors, predictive information on the healing status of the wound site is generated;

[0010] Based on the healing prediction information, a treatment plan is determined for the user's wound site.

[0011] Secondly, embodiments of the present invention provide a device for predicting the wound healing process, the device comprising:

[0012] The sequence acquisition module is used to acquire a target image sequence, which consists of target images at different time points containing the user's trauma site and a standard reference object;

[0013] The first extraction module is used to perform local feature extraction and global feature extraction processing on each target image in the target image sequence using the first module of the pre-trained prediction model, to obtain local feature vectors and global feature vectors respectively.

[0014] The second extraction module is used to perform temporal feature extraction processing on the local feature vector and the global feature vector respectively based on different preset extraction periods using the second module of the pre-trained prediction model to obtain multiple temporal feature vectors.

[0015] The information prediction module is used to generate prediction information on the healing status of the wound site based on the pre-trained prediction model and the multiple temporal feature vectors.

[0016] The treatment plan determination module is used to determine a treatment plan for the user's wound site based on the healing status prediction information.

[0017] Thirdly, embodiments of the present invention provide a wound healing process prediction device, including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the wound healing process prediction method provided in the above embodiments.

[0018] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the wound healing process prediction method provided in the above embodiments.

[0019] Fifthly, embodiments of the present invention provide a computer program product, including a computer program that, when executed by a processor, implements the steps of the wound healing process prediction method provided in the above embodiments. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating a method for predicting the wound healing process according to the present invention.

[0022] Figure 2 This is a flowchart illustrating another method for predicting the wound healing process according to the present invention.

[0023] Figure 3 This is a schematic diagram of the structure of a wound healing process prediction device according to the present invention;

[0024] Figure 4 This is a schematic diagram of the structure of a wound healing process prediction device according to the present invention. Detailed Implementation

[0025] This invention provides a method, apparatus, and device for predicting the wound healing process.

[0026] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0027] This specification provides a method, apparatus, and device for predicting the wound healing process. Currently, in primary hospitals, rehabilitation institutions, and home care environments, wound healing prediction mainly relies on the experience of medical staff and visual observation, supplemented by measuring wound size with a ruler. This method of predicting healing is not only time-consuming and labor-intensive but also carries the risk of subjective bias and cross-infection. Therefore, a technical solution is needed to improve the accuracy of wound healing prediction and thus enhance wound care effectiveness. In this solution, a target image sequence is acquired, consisting of target images at different time points containing the user's wound site and a standard reference object. Using a first module of a pre-trained prediction model, local and global feature extraction processes are performed on each target image in the target image sequence to obtain local and global feature vectors. Using a second module of the pre-trained prediction model, temporal feature extraction processes are performed on the local and global feature vectors based on different preset extraction cycles to obtain multiple temporal feature vectors. Based on these multiple temporal feature vectors, the pre-trained prediction model generates healing prediction information for the wound site. Based on this healing prediction information, a treatment plan for the user's wound site is determined. In this way, on the one hand, since the target image contains a standard reference object, the accuracy of wound site identification can be improved by comparing the wound site with the standard reference object. On the other hand, through the prediction model, based on both local and global features, temporal features can be extracted through multiple extraction cycles to obtain a temporal feature vector that can characterize the change pattern of the wound site over time. Furthermore, the temporal feature vector can be used to accurately predict the healing status of the wound site, allowing for the generation of targeted treatment plans based on the predicted healing status information, thereby improving wound care effectiveness. Specific processing details can be found in the following embodiments.

[0028] like Figure 1 As shown, this embodiment of the invention provides a method for predicting the wound healing process. The execution subject of this method can be a terminal device or a server. The terminal device can be a mobile terminal device such as a mobile phone, tablet computer, or smartwatch, or a terminal device such as a computer. The server can be an independent server or a server cluster composed of multiple servers. The method specifically includes the following steps:

[0029] In step S102, the target image sequence is obtained.

[0030] The target image sequence can be composed of target images at different time points, including the user's wound site and a standard reference object. The wound site can be any part of the body, such as the arm, palm, or face. The standard reference object can be any object of a size that can be obtained, such as a one-yuan coin, a band-aid, or a medical ruler.

[0031] In practice, the server can acquire target images at multiple time points and construct a target image sequence based on the time point of each target image. For example, medical staff or users can take photos of the wound site using portable devices such as mobile phones. The server can acquire photos taken of the user's wound site and standard reference objects every day within the past week and determine the acquired photos as target images.

[0032] In addition, the terminal device can guide the user to place a standard reference object during shooting.

[0033] In step S104, the first module of the pre-trained prediction model is used to perform local feature extraction and global feature extraction on each target image in the target image sequence to obtain local feature vectors and global feature vectors.

[0034] In practice, the pre-trained prediction model can automatically detect the outline of a standard reference object and, combined with its known physical dimensions, achieve the mapping transformation from image pixels to actual length, thereby accurately calculating the actual area and edge contour of the wound site. This can solve problems such as size errors caused by the lack of scale information in photographic shooting.

[0035] Subsequently, the pre-trained prediction model can perform image preprocessing and wound condition recognition, outputting evaluation information including wound type, area, and color features.

[0036] Then, the first module of the pre-trained prediction model can perform local feature extraction and global feature extraction based on the evaluation information obtained from image preprocessing and wound state recognition processing, respectively.

[0037] In step S106, the second module of the pre-trained prediction model is used to perform temporal feature extraction processing on local feature vectors and global feature vectors based on different preset extraction periods, so as to obtain multiple temporal feature vectors.

[0038] In practice, the server can determine multiple preset extraction periods based on the time points corresponding to the target images contained in the target image sequence. For example, if the target image sequence contains target images from the past 15 days, the server can divide these 15 days into periods to obtain multiple different preset extraction periods. Specifically, the preset extraction periods can include four extraction periods: the past 3 days, the past 5 days, the past 10 days, and the past 15 days.

[0039] Since wound healing is a dynamic temporal process, it is difficult to grasp the whole picture by assessing the wound healing status at a single point in time. Therefore, the server performs temporal feature extraction processing through multiple preset extraction cycles, which can obtain temporal feature vectors at multiple time scales that can characterize the wound change pattern.

[0040] In step S108, a pre-trained prediction model is used to generate prediction information on the healing status of the wound site based on multiple temporal feature vectors.

[0041] In practice, the pre-trained prediction model may also include a fourth module. The server can use the fourth module to perform prediction processing on the healing status based on multiple temporal feature vectors to obtain prediction information on the healing status of the wound site.

[0042] The healing progress prediction information may include a healing cycle prediction (e.g., it is estimated that the wound will take X days to fully heal), an area reduction prediction (e.g., the area is expected to decrease by Y% next week), and risk warnings for infection or healing stagnation.

[0043] Wound healing is a complex and dynamic process, and accurately predicting its trend is crucial for timely intervention. By using time series analysis and deep learning techniques, based on continuously acquired wound image data, it is possible to accurately predict the wound healing status and potential risks.

[0044] In step S110, a treatment plan is determined for the user's wound site based on the healing prediction information.

[0045] In implementation, the server can utilize a large language model to determine a treatment plan for the user's wound site based on healing status prediction information. Alternatively, the server can determine a healing status score for the wound site based on the healing status prediction information and determine the treatment plan corresponding to the healing status score based on the preset correspondence between the score and the plan. Or, the server can send the healing status prediction information to preset medical personnel and receive the treatment plan feedback from the preset medical personnel. In addition, there can be various other methods for determining the treatment plan, and different methods can be selected according to different application scenarios. This specification does not specifically limit these methods in the embodiments.

[0046] Furthermore, in primary care hospitals, many doctors are not wound care experts. Systems that generate healing progress predictions and corresponding treatment plans (i.e., intelligent monitoring and decision support systems for the entire wound healing process) can act as "AI assistants" to provide second opinions. When patients visit, doctors can use tablets or mobile phones to photograph the wound. The system can quickly determine the wound's type and severity, and indicate whether there are signs of infection or if referral to a specialist hospital is necessary. Through AI-standardized scoring, primary care doctors can more confidently formulate treatment plans or decide on referrals, improving the accuracy and safety of wound care at the primary care level. In rehabilitation institutions (such as chronic wound clinics and rehabilitation nursing centers), nursing staff can use the aforementioned system to regularly monitor patients with long-term chronic wounds. The healing progress predictions generated by the system can help determine the effectiveness of the current treatment plan: if the healing speed is found to be slower than expected, doctors can adjust the treatment plan accordingly (such as changing medications or debridement methods); if the prediction shows that the wound is still difficult to heal after two weeks, more aggressive interventions can be considered in advance.

[0047] This data-driven decision support helps in the individualized management of patients with chronic wounds, avoiding the risk of complications from prolonged wound healing. In home care scenarios, the system can also provide services to users via a mobile app. For example, patients or their families can take photos of their wounds daily and upload them; the predictive model can automatically analyze changes in wound color to determine the healing progress. If the system detects an abnormality (such as a sudden increase in redness and swelling, potentially indicating infection), it will immediately issue a warning via the app and recommend a follow-up medical examination, significantly reducing the time spent on manual interpretation by medical staff. The system can incorporate a wide range of wound types during model training, and through its healing prediction function, it can become a universal remote wound management tool. Through the mobile app, patients can also easily view their healing scores and progress charts, which not only increases their motivation for self-care but also allows professional medical staff to monitor the patient's condition at any time, achieving true telemedicine and continuous care.

[0048] Based on the above process, an intelligent monitoring and decision support system for the entire wound healing process can be built. In terms of system deployment and user experience, HealScope-AI emphasizes the integration of "high precision and low barrier to entry." Medical staff or patients can upload wound images via mobile phone, and the system automatically completes image processing, recognition analysis, trend prediction, and visualization. Users can view healing indices, tissue structure atlases, heat maps of redness and swelling areas, and export reports according to their professional needs. They can also switch to the patient side to view graphical scores, emoticon prompts, and operation suggestions, truly achieving a two-way adaptation between "professionals" and "patients."

[0049] Furthermore, the system utilizes the ONNX export mechanism for model deployment, ensuring compatibility with various edge devices and platforms. It can be embedded in primary healthcare equipment and clinic terminals, or deployed on dedicated hospital servers for remote management, significantly enhancing its engineering capabilities and clinical application potential. The system also supports online model updates and automatic learning, continuously optimizing recognition accuracy and predictive capabilities as data accumulates, demonstrating strong self-evolutionary abilities.

[0050] In summary, HealScope-AI, through its multi-dimensional innovations in image recognition, temporal modeling, intelligent intervention suggestion generation, and interactive visualization, has constructed a complete and intelligent integrated platform for wound assessment, prediction, and intervention. It not only fills the gap between static assessment and suggestion output in traditional systems but also greatly expands the practical application boundaries of AI in skin wound management through its universal model and highly user-friendly design.

[0051] This invention provides a method for predicting the wound healing process. The method involves acquiring a target image sequence, which consists of target images at different time points containing the user's wound site and a standard reference object. A first module of a pre-trained prediction model is used to extract local and global features from each target image in the sequence, resulting in local and global feature vectors. A second module of the same pre-trained prediction model is then used to extract temporal features from the local and global feature vectors based on different preset extraction periods, resulting in multiple temporal feature vectors. Based on these multiple temporal feature vectors, the pre-trained prediction model generates prediction information regarding the wound healing status. Finally, based on this prediction information, a treatment plan for the user's wound site is determined. In this way, on the one hand, since the target image contains a standard reference object, the accuracy of wound identification can be improved by comparing the wound site with the standard reference object. On the other hand, through the prediction model, based on both local and global features, temporal features can be extracted through multiple extraction cycles to obtain a temporal feature vector that can characterize the change pattern of the wound site over time. Then, through the temporal feature vector, the healing status of the wound site can be accurately predicted, and corresponding treatment plans can be generated in a targeted manner based on the healing status prediction information, thereby improving the wound care effect.

[0052] In practical applications, the first module may include a first sub-module constructed by a residual neural network and a second sub-module constructed by a visual Transformer. The first sub-module is used to extract local features from each target image in the target image sequence to obtain a local feature vector, and the second sub-module is used to extract local features from each target image in the target image sequence to obtain a global feature vector.

[0053] Convolutional Neural Networks (CNNs) have become a standard tool in the field of medical image analysis due to their excellent performance in image feature extraction. This system uses a pre-trained ResNet residual neural network to extract local detail features from the target image through multi-level convolution and pooling operations, thereby obtaining local feature vectors.

[0054] While CNNs excel at local feature extraction, their ability to capture overall image information is relatively limited. To compensate for this deficiency, this system introduces a Visual Transformer (ViT) to construct a second module. ViT can comprehensively analyze the global contextual information of wound images through a self-attention mechanism, and can be used to assess the condition of the skin around the wound.

[0055] By combining the local feature extraction capability of CNN with the global information analysis capability of ViT, a comprehensive characterization of wound features can be achieved, which can improve the system's adaptability to complex wound scenarios.

[0056] In practical applications, local feature vectors are used to characterize the clarity of the wound edges, the texture changes of the tissue, and the color distribution at the wound site, while global feature vectors are used to characterize the condition of the skin surrounding the wound site.

[0057] For example, by analyzing the degree of blurring or color abnormalities (such as redness or blackening) at the wound edges, the system can identify early signs of infection; by observing changes in texture, it can determine the degree of necrosis of wound tissue or the regeneration of granulation tissue. The extraction of these local features plays a crucial role in distinguishing different types of wounds (such as surgical wounds and diabetic foot ulcers) and assessing wound severity.

[0058] For example, when signs of inflammation spreading around a wound appear (such as redness or swelling of the skin), ViT is able to detect these subtle but important changes.

[0059] In practical applications, the second module includes a third sub-module constructed from a convolutional neural network and a fourth sub-module constructed from a long short-term memory network. The multiple temporal feature vectors include those obtained by the third sub-module performing temporal feature extraction processing on local feature vectors and global feature vectors according to a preset first extraction period, and those obtained by the fourth sub-module performing temporal feature extraction processing on local feature vectors and global feature vectors according to a preset second extraction period. The first extraction period is shorter than the second extraction period.

[0060] Because the wound healing process involves changes at different time scales, such as short-term changes in color or exudate daily, and long-term trends of weekly area reduction, a second module incorporating multi-scale convolutional layer structures can be designed to comprehensively capture these dynamic features. Short-scale convolutional layers can focus on extracting short-term change features, such as subtle improvements in the wound surface (e.g., lighter tissue color or reduced exudate); long-scale convolutional layers can analyze long-term trends, such as the overall progress and boundary changes of wound closure. This multi-scale feature extraction method ensures a comprehensive understanding of the wound healing process, enabling the identification of subtle changes, such as the early appearance of signs of infection or the potential risk of healing stagnation.

[0061] In practical applications, the temporal feature vector extracted by the third submodule can be used to characterize the spatial structural features of the wound site, while the temporal feature vector extracted by the fourth submodule can be used to characterize the changes in wound features at the wound site.

[0062] Because Convolutional Long Short-Term Memory (ConvLSTM) networks combine the advantages of Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs), a fourth submodule can be constructed based on ConvLSTM. CNNs can handle the spatial structural features of the wound image (i.e., the target image), such as the shape of the wound area and the distribution of tissue; while LSTMs can capture the evolution of these features over time, such as the daily reduction of the wound area or the gradual process of tissue regeneration. Through ConvLSTM, the system can learn the wound closure speed, tissue regeneration patterns, and the development trends of potential complications, and output specific healing prediction information, such as the estimated number of days required for complete healing, the extent of wound area reduction in the coming week, and risk warnings of infection or stagnation in healing.

[0063] In predicting wound healing trends, the classic Convolutional Long Short-Term Memory (ConvLSTM) network structure was fully utilized and optimized, combined with a multi-scale convolutional feature extraction strategy, to form a temporal modeling module specifically for dynamic changes in wounds. ConvLSTM has been widely used in video analysis and spatial sequence modeling, and it can simultaneously capture the spatial features and temporal evolution patterns of images. Based on this, and considering the need for stage-specific feature changes in the wound healing process, a "multi-scale temporal feature extraction module" was innovatively designed to extract short-term changes (such as the spread of redness and swelling, and the increase or decrease of exudation within 12 days) and medium- to long-term trends (such as the closure rate and edge tissue growth within 714 days), and to improve the accuracy and robustness of healing status prediction through fusion modeling.

[0064] In experiments, this module demonstrated good generalization ability and stability in various wound trend prediction tasks, achieving a healing time prediction accuracy of over 85%. It is particularly valuable for identifying high-risk healing abnormalities such as infection stagnation. By integrating ConvLSTM with a temporal structure optimized for wound data characteristics, HealScope-AI not only overcomes the limitations of static analysis in traditional wound assessment systems but also provides more forward-looking auxiliary support for intelligent wound monitoring.

[0065] In practical applications, step S106 utilizes the second module of the pre-trained prediction model to perform temporal feature extraction on local and global feature vectors based on different preset extraction periods. The processing methods for obtaining multiple temporal feature vectors can vary. One optional processing method is provided below, such as... Figure 2 As shown, the specific process may include the following steps S1062 to S1066.

[0066] In step S1062, the third module of the pre-trained prediction model is used to determine the weight values ​​corresponding to the local feature vector and the global feature vector based on the image features of each target image in the target image sequence.

[0067] In implementation, to further optimize image recognition performance, an adaptive local-global attention module, i.e., the third module of the prediction model, can be constructed. This third module can dynamically adjust the model's attention to local and global features based on the image characteristics of the wound image.

[0068] For example, for small wounds, the system will pay more attention to local details, such as subtle changes in the wound edges; while for complex wounds with obvious inflammation, the system will prioritize analyzing global features, such as the color distribution of the surrounding skin and the overall extent of inflammation.

[0069] Furthermore, the third module of the adaptive mechanism can be implemented through a lightweight network structure, which enables the system to automatically select to focus on local details or global context information based on the complexity of the specific wound type (such as abrasions, burns, diabetic foot, etc.). This not only improves the accuracy of recognition but also ensures the robustness of the system, especially in scenarios with significant differences in lighting conditions or patient skin color.

[0070] In step S1064, the target feature vector is determined based on the local feature vector and its corresponding weight value, and the global feature vector and its corresponding weight value.

[0071] In practice, the server can perform weighted processing based on the local feature vector and its corresponding weight value, and the global feature vector and its corresponding weight value, to obtain the target feature vector.

[0072] The target feature vector can be used to characterize the wound type (e.g., surgical incision, abrasion, diabetic foot ulcer, pressure ulcer), severity level (mild, moderate, severe), accurate wound region segmentation results, and size measurement data of the trauma site. Through the first, second, and third modules constructed based on deep learning algorithms, the wound area and its changes at the trauma site can be automatically calculated, providing high-quality, quantitative basic data support for subsequent trend prediction and intervention decisions.

[0073] In step S1066, the second module of the pre-trained prediction model is used to perform temporal feature extraction processing on the target feature vector based on different preset extraction periods, thereby obtaining multiple temporal feature vectors.

[0074] In implementation, since wound healing is a dynamic temporal process, a second module can be constructed using a deep learning framework combining multi-scale temporal feature extraction and Convolutional Long Short-Term Memory (ConvLSTM) networks. To more accurately capture subtle changes in the wound healing process, a third sub-module, including multi-scale convolutional feature extraction layers, extracts scale features from time-series wound images (i.e., target images contained within a target image sequence) for multiple extraction periods, including short-term (e.g., daily changes in tissue color, increases / decreases in exudate, etc.) and long-term (e.g., weekly changes in wound area and boundaries). These multi-scale features are then fed as input into a fourth sub-module constructed with ConvLSTM, enabling accurate prediction of wound healing trends. The second model integrates the advantages of Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, effectively capturing the relationship between spatial structure and temporal features. Through the second module, the system can automatically learn and predict temporal feature vectors that characterize the changes in wounds over time, such as the rate of wound area reduction, the regeneration of marginal tissue, and potential signs of infection.

[0075] This innovative system deeply integrates two mainstream model architectures in computer vision—Vision Transformer (ViT) and Residual Neural Network (ResNet)—to construct an image recognition module with local-global perception capabilities. While most existing research still relies on a single CNN architecture, this system pioneers the use of ViT to process skin wound images, capturing global wound features such as overall distribution, symmetry of peripheral redness and swelling, and the spread of exudate from surrounding skin. Meanwhile, ResNet precisely focuses on local details in the image, such as tissue texture, microbleeds, and the color and distribution of granulation tissue. The two technologies are fused through a "local-global adaptive attention mechanism," dynamically adjusting attention allocation based on wound complexity. This effectively avoids redundant or missing feature information, significantly improving the model's adaptability and accuracy in recognizing different types of wounds and grading tissues. Experimental results show that the module achieves an average accuracy of 98.2% in image classification tasks and controls the error in area measurement to within ±4.5%, both outperforming existing open-source models.

[0076] In practical applications, the prediction model can also be trained. There are various ways to train a prediction model; one optional method is provided below. Figure 2 As shown, the specific process may include the following steps S202 to S208.

[0077] In step S202, a sample image sequence is obtained.

[0078] In step S204, data augmentation processing is performed on the sample images contained in the sample image sequence according to the preset data augmentation rules to obtain the augmented sample image sequence.

[0079] The preset data augmentation rules include at least image rotation, color jitter, brightness adjustment, and contrast adjustment.

[0080] In practice, the quality and features of wound images can vary significantly due to factors such as lighting conditions, patient skin color, and shooting angle, posing a challenge to model training and performance. To address this, data augmentation techniques can be used to increase the diversity of training data. Specifically, the server can apply data augmentation rules such as image rotation, color jitter, brightness adjustment, and contrast adjustment to the sample image sequence, resulting in an enhanced sample image sequence.

[0081] In step S206, the first prediction model is obtained.

[0082] The first prediction model can be trained based on the first sample data, where the amount of data in the first sample data is greater than the amount of data in the sample image sequence. The first sample data can be a general dataset (such as ImageNet).

[0083] In step S208, the first prediction model is trained based on the enhanced sample image sequence to obtain a pre-trained prediction model.

[0084] In implementation, the server can utilize transfer learning to leverage a pre-trained prediction model on a large, general-purpose dataset (such as ImageNet) and then fine-tune it for specific features of wound images to obtain a pre-trained prediction model. This overcomes the problem of data scarcity in the medical field and enhances the model's generalization ability, ensuring stable performance across different scenarios.

[0085] In this way, by employing techniques such as data augmentation and transfer learning, pre-training the model using public databases, and then fine-tuning it on data from specific application scenarios, we can ensure that the model can still work effectively even when the amount of data is insufficient. This multi-scale time series analysis method can effectively improve prediction accuracy and sensitivity, providing more reliable decision support for clinical wound management.

[0086] In addition, to further enhance the model's generalization ability and prediction accuracy, the system can incorporate a wide variety of wound cases during training, including samples of normal healing and delayed healing, which enables the model to accurately distinguish different healing trajectories and provide early warnings of abnormal situations.

[0087] HealScope-AI is built around the principles of "generality" and "scenario adaptability." Unlike many models that focus solely on single chronic wounds such as diabetic foot ulcers and pressure ulcers, this system's architecture is designed from the outset to address a wide range of wound types, including postoperative wounds, external injuries, burns, and animal trauma. Through data augmentation and cross-wound knowledge transfer technologies, it enhances the model's ability to identify rare wound types, constructing a universally applicable model framework. Furthermore, the system supports switching interfaces and interaction modes based on the task environment, adapting to various application scenarios such as primary healthcare, rehabilitation institutions, home care, and disaster relief.

[0088] In practical applications, the treatment plan determined in step S110 based on the healing prediction information for the user's wound site can take many forms. One possible treatment method is provided below, such as... Figure 2 As shown, the specific process may include the following steps S1102 to S1106.

[0089] In step S1102, wound care rules corresponding to the user's wound site are determined based on the healing prediction information.

[0090] In implementation, the server can build a rule engine that generates intervention recommendations based on internationally recognized clinical guidelines for wound care. For example, when the system detects signs of wound infection (such as redness or abnormal exudation), the rule engine will recommend wound care rules for using antibacterial dressings or topical antibiotics; if slow healing is detected, it may recommend increasing the frequency of dressing changes or using wound care rules for promoting healing (such as hydrocolloid dressings). This rule-based approach ensures the medical rationality and clinical operability of the recommendations, conforming to current best practice standards.

[0091] In step S1104, the multimodal user data corresponding to the user is obtained.

[0092] Multimodal data can include users' individual characteristics, medical history, laboratory test results, and lifestyle data.

[0093] In practice, in addition to image data, the server can also acquire additional user data, including individual characteristics (such as age, comorbidities, etc.), medical history information (such as history of diabetes, history of hypertension, etc.), laboratory test results (such as blood glucose levels, inflammatory markers, etc.), and lifestyle data (such as smoking status, activity level, etc.). By comprehensively analyzing this multimodal user data, the system can identify key factors affecting wound healing. For example, malnutrition may lead to delayed healing, and smoking may increase the risk of infection, thus generating more accurate recommendations. For instance, if a patient's blood glucose control is poor, the system will prioritize recommending blood glucose management strategies; if a high risk of infection is detected, it will recommend the use of antibacterial drugs. This multimodal approach fully reflects the scientific principle that wound healing is influenced by multiple systemic factors.

[0094] In step S1106, a model is determined using a pre-trained scheme, and a treatment plan for the user's wound site is determined based on wound care rules and multimodal user data.

[0095] The proposed model is one obtained through reinforcement learning training.

[0096] To further enhance the flexibility and personalization of decision-making during implementation, reinforcement learning techniques can be introduced. Reinforcement learning can continuously optimize treatment pathways by simulating a trial-and-error process, adjusting wound care rules based on the user's multimodal data to arrive at a treatment plan. For example, for diabetic patients, the system might prioritize blood sugar control supportive measures (such as dietary adjustments) to promote healing; for malnourished patients, it might suggest protein or vitamin supplementation. In this process, the reinforcement learning agent learns from historical treatment outcome data, evaluates the efficacy of different interventions, and dynamically adjusts strategies to maximize healing speed and minimize the risk of complications.

[0097] Based on healing prediction information derived from image recognition and healing trend prediction, a personalized intervention suggestion generation mechanism can be introduced, enabling a shift from simple monitoring to proactive intervention. By combining a rule engine with reinforcement learning technology, the system can fully utilize the patient's specific wound condition, predicted healing trends, and background data (such as age, chronic disease history, and lifestyle habits) to dynamically generate precise, personalized treatment plans. Wound management provides a scientific and dynamic decision-making framework that ensures the personalization, precision, and clinical effectiveness of intervention plans.

[0098] In practice, the process begins with an initial assessment by a built-in rule engine. This engine, based on the latest clinical guidelines and professional treatment consensus, ensures that the generated wound care rules are clinically effective and medically sound. The rule engine automatically identifies specific wound conditions (such as high infection risk or significantly delayed healing) and initially recommends appropriate interventions (i.e., wound care rules), such as increasing wound debridement frequency or using specific antibacterial or healing-promoting dressings. Subsequently, a reinforcement learning algorithm further optimizes these recommendations, continuously assessing and refining the effectiveness of the intervention strategies by learning from historical treatment outcome data. Reinforcement learning can identify and recommend specific measures that best improve healing, achieving progressive precision and personalization of the treatment plan.

[0099] Furthermore, by collecting real-time user feedback (such as treatment effectiveness data and patient satisfaction), the model can continuously learn and adjust its recommended strategies to maximize treatment outcomes. Ultimately, it can output specific and actionable intervention recommendations (i.e., treatment plans), including recommended debridement frequency, dressing types, medication regimens, and other supportive care measures. By providing personalized and precise treatment guidance, it can effectively improve wound care quality, shorten healing cycles, reduce the risk of complications, and enhance the patient's recovery experience.

[0100] In terms of functional design, this system breaks through the limitations of traditional AI assessment tools that "only identify, but do not make decisions," and innovatively introduces a "personalized intervention suggestion generation module." Based on reinforcement learning strategies and a clinical rule engine, this module combines individualized patient characteristics (such as basic medical history, previous medications, and immune status) with wound characteristics (such as infection risk level, tissue structure proportions, and exudate type) to output highly actionable nursing suggestions, such as "It is recommended to change dressings twice daily, and silver ion antibacterial dressings are recommended" and "Specialist consultation is required," achieving an intelligent upgrade from "passive assessment" to "proactive suggestion." This function not only enhances the system's clinical guidance value but also significantly reduces treatment discrepancies caused by insufficient experience among primary care physicians, promoting the standardization and scientification of wound management decisions.

[0101] In practical applications, healing prediction information can include healing trend curves. Step S108 utilizes a pre-trained prediction model, based on multiple temporal feature vectors, to generate healing prediction information for the wound site. Various processing methods can be employed; the following provides one optional method: Figure 2 As shown, the specific process may include the following steps S1082 to S1084.

[0102] In step S1082, the healing status corresponding to each predicted time point is determined using a pre-trained prediction model based on the time point corresponding to each target image in the target image sequence and multiple temporal feature vectors.

[0103] In step S1084, a healing trend curve is generated based on the healing status corresponding to each predicted time point.

[0104] In practice, the variable time intervals between image acquisition pose additional challenges to time series prediction. For example, patients may be unable to upload images on certain days, resulting in data gaps. To address this issue, a time-aware mechanism can be introduced, using time embedding technology to incorporate the differences in acquisition time intervals into the model. This approach allows the system to adapt to variations in observation gaps, ensuring the accuracy and reliability of prediction results. For instance, the system can infer potential changes within missing time periods based on historical data, thereby generating a continuous healing trend curve.

[0105] The aforementioned dynamic modeling technology can provide medical staff with personalized treatment reference data and provide early warnings of abnormal situations, such as infection risk or delayed healing.

[0106] Because intelligent monitoring and decision support systems for the entire wound healing process incorporate medical science principles of wound healing, they can ensure the clinical relevance of the system. This is specifically reflected in the following aspects:

[0107] Wound healing stages identification: Wound healing is divided into four key stages: hemostasis, inflammation, proliferation, and remodeling. This system uses the first module to analyze images and identify visual markers for each stage, such as the characteristics of exudate in the inflammation stage, granulation tissue formation in the proliferation stage, and scar maturation in the remodeling stage, ensuring accurate monitoring of the healing process.

[0108] Visual signs of complications: The system incorporates visual signs of wound complications, such as redness and swelling due to infection, tissue necrosis due to delayed closure, or irregular edges. These signs are automatically identified by an AI model, providing a basis for early intervention.

[0109] Reference to Standard Nursing Practices: The systematic recommendations reference standard wound care practices, including debridement techniques (removal of necrotic or infected tissue), dressing selection principles (selecting appropriate dressings based on wound moisture and exudate levels, such as alginate or hydrocolloid dressings), and infection control strategies (monitoring for signs of infection and recommending antimicrobial therapy when necessary). These practices are based on clinical guidelines, ensuring the scientific validity and operability of the recommended protocols.

[0110] By deeply integrating with medical science, the clinical applicability of the technology output can be ensured, providing reliable decision support for medical staff and patients.

[0111] In addition, to verify the scientific validity and practicality of the system, the following measures can be taken to ensure its credibility and effectiveness in reality:

[0112] Clinical trials: The accuracy of the system's identification, the reliability of its predictions, and the clinical effectiveness of the intervention recommendations are tested by comparing the results with expert evaluations. Indicators may include the accuracy of wound type identification, the sensitivity and specificity of healing trend prediction, and the improvement in healing time and infection rate by the intervention.

[0113] Outcome evaluation: Randomized controlled trials (RCTs) will be used to verify the clinical benefits of the system, such as shortened healing time, reduced infection rate, and reduced medical costs. These outcome evaluations will provide a scientific basis for the widespread application of the system.

[0114] Interdisciplinary collaboration: Working closely with medical experts, technical teams, and clinical institutions to ensure seamless integration of technology with clinical workflows. For example, collaborating with primary care hospitals to validate the system's usability and with rehabilitation facilities to optimize chronic wound management programs.

[0115] The above verification process ensures the scientific rigor and clinical feasibility of the system, setting a benchmark for the application of medical AI.

[0116] To ensure the system's training quality and broad adaptability to practical applications, a comprehensive, diverse, and standardized multi-source heterogeneous wound image database can be constructed. This database should cover different wound types, healing stages, and usage scenarios, possessing good representativeness, generalization ability, and clinical transferability. The database mainly includes the following three data sources:

[0117] 1. International publicly available wound image database

[0118] This system fully utilizes existing standardized public database resources, including the Medetec Wound Database, Foot Ulcer Challenge Dataset (FUCD), and CURA Wound Dataset. These datasets cover common chronic and acute wounds such as diabetic foot ulcers, pressure ulcers, and postoperative incisions, providing clear image quality and corresponding professional annotation information, suitable for tasks such as image classification, region segmentation, and healing trend modeling. As an important foundation for training the model, the use of public datasets enhances the model's initial recognition ability and cross-domain versatility.

[0119] 2. Real case images collected in conjunction with clinical institutions

[0120] A collaborative mechanism for acquiring real wound images can be established. All acquired images undergo anonymization and are collected and used only under ethical review and compliance conditions. The data can cover different wound types, such as postoperative wounds, burns, and lacerations, and some may also include basic patient information, medical history, treatment records, and wound development stage tags.

[0121] The clinical image data were annotated by multiple wound repair specialists to ensure the accuracy and professionalism of the label information. It is mainly used for: localization and optimization of the model; generation and training of personalized intervention suggestions; and modeling and prediction of dynamic changes in multi-stage wound healing.

[0122] 3. Model-aided generation and simulation data expansion

[0123] For tasks involving rare clinical wound types or severely limited sample sizes, such as small-area burns, complex wounds, or extremely infected wounds, various enhancement mechanisms can be introduced to expand the data, including: image enhancement (rotation, scale, noise, crop, etc.); medical image style transfer (StyleGAN, CycleGAN); multimodal joint construction and AI-assisted simulation generation.

[0124] By combining clinical expert review mechanisms, extended images undergo multi-stage quality screening to ensure that they meet modeling requirements in terms of image authenticity, structural logic, and consistency with medical interpretation, making them particularly suitable for trend prediction and small-sample learning tasks.

[0125] The above-mentioned intelligent system, which integrates wound image recognition, healing trend prediction and intervention decision support, is constructed based on artificial intelligence (AI) image recognition and generation technology. The system can include four core functional modules: (1) Wound image intelligent recognition module, which uses deep learning technology to automatically analyze wound images, identify wound types, measure wound area and tissue condition; (2) Healing trend intelligent prediction module, which accurately predicts the future healing process of wounds based on multi-scale temporal feature extraction technology and ConvLSTM algorithm; (3) Personalized intervention suggestion generation module, which combines rule engine and reinforcement learning to generate targeted treatment plans based on wound status, healing trend and patient background data; (4) Scoring and visualization feedback module, which transforms the above analysis results into intuitive and easy-to-read scoring index and visualization interface to support clinical decision-making and patient self-management.

[0126] The scoring and visualization feedback module displays all analysis results graphically on the user interface. This includes: a curve showing the change in wound area over time (unit: cm). 2 The system provides information such as healing speed trends, risk warning indicators, and system-generated intervention suggestions to help both doctors and patients keep abreast of wound recovery dynamics, enhance compliance, and improve remote management capabilities. Through a closed-loop process of "image analysis + trend prediction + intelligent intervention + visual feedback," a precise, efficient, and practical one-stop AI wound management platform has been built, suitable for primary healthcare, rehabilitation nursing, and home settings, providing a new technological solution for wound care.

[0127] With an objective assessment of the current wound condition and a prediction of future trends, the scoring and visualization feedback module translates the results into a user-friendly and easily understandable format. On one hand, the system generates a wound healing score for each wound, comprehensively considering wound size and depth, tissue health, and healing stage. This score is designed as a graded system; for example, A indicates expected complete healing, with lower scores indicating more severe wound problems or slower healing. The score calculation references commonly used clinical wound assessment scales and is automatically generated using AI-generated quantitative indicators, ensuring the score's medical significance and comparability. On the other hand, the system provides rich visualization feedback. For example, the application interface displays an analytical annotation of the wound photograph: the wound area is highlighted, and the system can use different colors to mark areas such as granulation tissue, new skin, and potential necrotic tissue, allowing users to intuitively understand the wound's composition. The system also plots healing progress curves, such as automatically drawing a graph of wound area changes over time based on follow-up visits, and using different colors to indicate predicted future trends on the curve. For potential abnormalities (such as a predicted risk of future infection), the interface will provide warning signs or information prompts. All of this feedback helps users understand the AI's assessment: it's not just an abstract judgment, but a data-driven and graphically supported evaluation. This intuitive presentation meets the different needs of healthcare professionals and patients—professionals can use it for case recording and efficacy assessment, while patients can more clearly see whether their wounds are improving or worsening, thus improving adherence. It's worth noting that the scoring and visual feedback module will also support report export functionality for archiving by medical institutions or remote consultations with experts.

[0128] This system enables real-time monitoring of wound status, risk warnings, and personalized guidance, outputting quantitative assessment reports and treatment recommendations. On one hand, by analyzing wound photographs, the system can automatically identify and quantitatively assess wound type, size, and healing stage, reducing reliance on human experience. On the other hand, by utilizing deep learning algorithms to extract features and perform temporal analysis on continuously acquired wound images, it can accurately determine the current healing progress, predict future trends, and anticipate the risk of wound deterioration or slow healing. Furthermore, it leverages reinforcement learning to fuse multimodal patient data to intelligently generate personalized intervention recommendations. In addition, the system will output unified, quantitative, and interpretable intelligent auxiliary assessment results, such as wound healing index and wound area change curves, helping doctors and patients intuitively understand the current wound condition and future trends. By integrating these functions into a unified platform, this application aims to improve the efficiency and quality of wound care: in primary hospitals, alleviating the assessment burden when specialized wound care departments are lacking; in rehabilitation institutions, providing scientific evidence to adjust rehabilitation plans; and in home care scenarios, providing technical support for patient self-management and telemedicine. Ultimately, this system is expected to reduce the incidence of wound complications, accelerate the healing process, reduce medical costs, and improve patients' quality of life.

[0129] This application not only overcomes the limitations of traditional diagnostic methods that rely on visual inspection and experience, but can also be effectively applied in various scenarios, generating tangible social and economic value.

[0130] Firstly, at the hospital clinical nursing level, this application can be used for postoperative follow-up and chronic wound management in multiple departments such as surgery, burn unit, dermatology, and rehabilitation. Currently, doctors or nurses typically rely on visual inspection and simple measurements (such as measuring wound size with a ruler) for wound monitoring, which is highly subjective and time-consuming. This application, however, utilizes AI recognition and time-series prediction models to automatically assess multiple indicators such as wound area, exudation level, and extent of tissue necrosis within seconds, and provides intuitive predictions of future healing progress. For example, the system can indicate whether the wound will heal successfully in a few days or weeks, or whether there is a potential risk of infection or scar hyperplasia. For inpatient wards requiring large-scale management, this intelligent assessment can significantly improve clinical nursing efficiency, reduce repetitive work for medical staff, and allow them to devote more time to high-value nursing interventions or communication.

[0131] Secondly, the value of this application is particularly evident in primary healthcare institutions. It is well known that primary healthcare units such as community health service centers and township hospitals generally suffer from a shortage of professional wound care personnel and insufficient medical equipment and technical support. When patients experience trauma or postoperative wounds requiring dressing changes and follow-up, primary healthcare workers often lack experience and are unable to promptly assess the risk of the wound, thus delaying treatment. This application deploys an AI assessment system in a lightweight manner on mobile phones or tablets. Patients only need to take a photo of the wound and input basic information; the system can quickly provide wound grading, healing trends, and intervention suggestions, helping primary care physicians make rapid judgments and decide whether referral is necessary. This alleviates the nursing burden on primary care facilities, effectively reduces the need for patients to travel to large hospitals, and conserves medical resources.

[0132] This application can also provide significant assistance in rehabilitation and nursing facilities such as chronic wound clinics, nursing homes, or rehabilitation hospitals. Chronic wounds typically require repeated care and observation for weeks or even months, and traditionally, each examination requires considerable manpower and time. This application provides nursing teams with quantified healing curves and risk alert mechanisms through regular photography and automated analysis. Once the system detects stagnant healing or an increased risk of infection, it immediately alerts healthcare professionals and provides intervention recommendations, such as changing dressings, increasing the frequency of dressing changes, or introducing new drug treatment regimens. This significantly reduces wound deterioration due to misjudgment or delayed observation, enabling better rehabilitation management and more rational allocation of resources.

[0133] Home care is another important application scenario for this application. Under the current medical model, many post-operative or chronically ill patients need to return to the hospital regularly for wound checkups after discharge. However, for patients with limited mobility or transportation, frequent trips to the hospital increase both financial burden and waste of medical resources. This application provides a mobile application where patients or their families can take photos of their wounds daily at home and record basic information. The system automatically assesses the current condition, predicts future changes, and provides daily care suggestions, such as whether pressure bandaging should be applied and when an in-person follow-up appointment is needed. If an acute deterioration trend is detected, the system can issue a timely warning, allowing the patient to seek professional help immediately. By enabling autonomous and intelligent wound management in the home environment, society as a whole can reduce the occupancy of hospital beds and outpatient resources, and significantly improve patient comfort and compliance.

[0134] The system also has significant potential in disaster response and special environmental scenarios. In natural disasters such as earthquakes and floods, or in battlefield rescue operations, there is often a situation where a large number of injured people appear in a short period of time, and medical resources are severely insufficient. Utilizing the intelligent recognition and prediction functions integrated into portable devices, rescue personnel can quickly assess the severity of wounds and the risk of infection in the field, assisting in triage and emergency treatment. Doctors can also remotely view wound images and, based on the system's automatic assessment, decide whether airlift or evacuation of patients is necessary, shortening valuable medical decision-making time and improving the overall efficiency and success rate of disaster emergency response.

[0135] Animal wound management is also a promising area for expansion. Pet hospitals and large-scale farms also need to identify wound types, predict healing progress, and intervene promptly in the care of animal trauma or post-operative care. For example, if hoof ulcers in dairy cows are not detected and properly cared for early, they can cause serious economic losses. This application, through cross-wound knowledge transfer and domain adaptation, can adapt to the characteristics of animal skin and tissue structure after training or fine-tuning, achieving wound management applicable to multiple species.

[0136] Finally, at a deeper level, this application also possesses research and educational value. The large amount of collected and annotated wound images and healing process data are of significant reference value for further research on tissue repair mechanisms, improvement of innovative dressings, and exploration of personalized treatment plans. Universities, research institutions, and enterprises can leverage this foundation to jointly develop more medical AI products, further catalyzing the development of the "Internet + Smart Healthcare" industry and injecting continuous momentum into public health.

[0137] The practical application value of this application lies in its long-term benefits across multiple levels, scenarios, and groups: it makes wound care in hospitals more standardized and efficient, enables primary healthcare institutions and rehabilitation centers to acquire professional assessment capabilities, makes home care and disaster emergency care more timely and accurate, and brings new development opportunities to the agricultural and pet healthcare sectors. With the continuous deepening of digitalization in the healthcare industry, this application will undoubtedly play an indispensable role in reducing healthcare costs, narrowing the urban-rural healthcare gap, and improving patients' quality of life, demonstrating genuine social welfare and industrial potential.

[0138] This specification provides a method for predicting the wound healing process. The method involves acquiring a target image sequence, which consists of target images at different time points containing the user's wound site and a standard reference object. A first module of a pre-trained prediction model is used to perform local and global feature extraction on each target image in the sequence, obtaining local and global feature vectors. A second module of the pre-trained prediction model is used to perform temporal feature extraction on the local and global feature vectors based on different preset extraction periods, obtaining multiple temporal feature vectors. Based on these multiple temporal feature vectors, the pre-trained prediction model generates prediction information on the healing status of the wound site. Based on this healing status prediction information, a treatment plan for the user's wound site is determined. In this way, on the one hand, since the target image contains a standard reference object, the accuracy of wound identification can be improved by comparing the wound site with the standard reference object. On the other hand, through the prediction model, based on both local and global features, temporal features can be extracted through multiple extraction cycles to obtain a temporal feature vector that can characterize the change pattern of the wound site over time. Then, through the temporal feature vector, the healing status of the wound site can be accurately predicted, and corresponding treatment plans can be generated in a targeted manner based on the healing status prediction information, thereby improving the wound care effect.

[0139] The above describes a method for predicting the wound healing process provided in the embodiments of this specification. Based on the same idea, embodiments of this specification also provide a device for predicting the wound healing process, such as... Figure 3 As shown.

[0140] The wound healing process prediction device includes: a sequence acquisition module 301, a first extraction module 302, a second extraction module 303, an information prediction module 304, and a scheme determination module 305, wherein:

[0141] The sequence acquisition module 301 is used to acquire a target image sequence, which consists of target images at different time points containing the user's wound site and a standard reference object;

[0142] The first extraction module 302 is used to perform local feature extraction and global feature extraction processing on each target image in the target image sequence using the first module of the pre-trained prediction model, so as to obtain local feature vectors and global feature vectors.

[0143] The second extraction module 303 is used to perform temporal feature extraction processing on the local feature vector and the global feature vector respectively based on different preset extraction periods using the second module of the pre-trained prediction model to obtain multiple temporal feature vectors.

[0144] The information prediction module 304 is used to generate prediction information on the healing status of the wound site based on the multiple temporal feature vectors using the pre-trained prediction model.

[0145] The treatment plan determination module 305 is used to determine a treatment plan for the user's wound site based on the healing prediction information.

[0146] In this embodiment of the specification, the first module includes a first sub-module constructed by a residual neural network and a second sub-module constructed by a visual Transformer. The first sub-module is used to extract local features from each target image in the target image sequence to obtain the local feature vector, and the second sub-module is used to extract local features from each target image in the target image sequence to obtain the global feature vector.

[0147] In this embodiment of the specification, the second module includes a third sub-module constructed from a convolutional neural network and a fourth sub-module constructed from a long short-term memory network. The plurality of temporal feature vectors include temporal feature vectors obtained by the third sub-module performing temporal feature extraction processing on the local feature vectors and the global feature vectors respectively according to a preset first extraction period, and temporal feature vectors obtained by the fourth sub-module performing temporal feature extraction processing on the local feature vectors and the global feature vectors respectively according to a preset second extraction period. The first extraction period is shorter than the second extraction period.

[0148] In the embodiments described in this specification, the second extraction module is used for:

[0149] Using the third module of the pre-trained prediction model, the weight values ​​corresponding to the local feature vector and the global feature vector are determined according to the image features of each target image in the target image sequence.

[0150] The target feature vector is determined based on the local feature vector and its corresponding weight value, and the global feature vector and its corresponding weight value.

[0151] Using the second module of the pre-trained prediction model, the target feature vector is subjected to temporal feature extraction processing based on different preset extraction periods to obtain the multiple temporal feature vectors.

[0152] In the embodiments described in this specification, the device further includes:

[0153] The sample acquisition module is used to acquire sample image sequences;

[0154] The data augmentation module is used to perform data augmentation processing on the sample images contained in the sample image sequence according to the preset data augmentation rules to obtain the enhanced sample image sequence. The preset data augmentation rules include at least image rotation, color jitter, brightness adjustment and contrast adjustment.

[0155] The model acquisition module is used to acquire a first prediction model, which is trained based on first sample data, and the amount of data in the first sample data is greater than the amount of data in the sample image sequence.

[0156] The model training module is used to train the first prediction model based on the enhanced sample image sequence to obtain the pre-trained prediction model.

[0157] In the embodiments of this specification, the scheme determination module is used for:

[0158] Based on the healing prediction information, determine the wound care rules corresponding to the user's wound site;

[0159] Obtain multimodal user data corresponding to the user; wherein, the multimodal data includes the user's individual characteristic data, medical history information, laboratory test results, and lifestyle data;

[0160] A model is determined using a pre-trained scheme. Based on the wound care rules and the multimodal user data, a treatment plan for the user's wound site is determined. The scheme determination model is a model trained through reinforcement learning.

[0161] In the embodiments of this specification, the healing prediction information includes a healing trend curve, and the information prediction module is used for:

[0162] Using the pre-trained prediction model, based on the time point corresponding to each target image in the target image sequence and the multiple temporal feature vectors, the healing status corresponding to each predicted time point is determined;

[0163] The healing trend curve is generated based on the healing status corresponding to each predicted time point.

[0164] In this embodiment of the specification, the local feature vector is used to characterize the clarity of the wound edge, the texture changes of the tissue, and the color distribution of the wound site, while the global feature vector is used to characterize the condition of the skin surrounding the wound site.

[0165] In this embodiment of the specification, the temporal feature vector extracted by the third submodule is used to characterize the spatial structural features of the trauma site, and the temporal feature vector extracted by the fourth submodule is used to characterize the changes in the wound features of the trauma site.

[0166] This specification provides an embodiment of a wound healing process prediction device. It acquires a target image sequence, which consists of target images at different time points containing the user's wound site and a standard reference object. Using a first module of a pre-trained prediction model, local and global feature extraction processes are performed on each target image in the target image sequence to obtain local and global feature vectors. Using a second module of the pre-trained prediction model, temporal feature extraction processes are performed on the local and global feature vectors based on different preset extraction periods to obtain multiple temporal feature vectors. Based on these multiple temporal feature vectors, the pre-trained prediction model generates prediction information on the healing status of the wound site. Based on the healing status prediction information, a treatment plan for the user's wound site is determined. In this way, on the one hand, since the target image contains a standard reference object, the accuracy of wound identification can be improved by comparing the wound site with the standard reference object. On the other hand, through the prediction model, based on both local and global features, temporal features can be extracted through multiple extraction cycles to obtain a temporal feature vector that can characterize the change pattern of the wound site over time. Then, through the temporal feature vector, the healing status of the wound site can be accurately predicted, and corresponding treatment plans can be generated in a targeted manner based on the healing status prediction information, thereby improving the wound care effect.

[0167] The above describes the wound healing process prediction device provided in the embodiments of this specification. Based on the same idea, the embodiments of this specification also provide a wound healing process prediction device, such as... Figure 4 As shown.

[0168] The device for predicting the wound healing process can be a terminal device or server, as described in the above embodiments.

[0169] The device for predicting the wound healing process can vary considerably depending on its configuration and performance. It may include one or more processors 401 and a memory 402, where one or more applications or data may be stored. The memory 402 may be temporary or persistent storage. The applications stored in the memory 402 may include one or more modules (not shown), each module including a series of computer-executable instructions for the wound healing process prediction device. Furthermore, the processor 401 may be configured to communicate with the memory 402 and execute the series of computer-executable instructions in the memory 402 on the wound healing process prediction device. The wound healing process prediction device may also include one or more power supplies 403, one or more wired or wireless network interfaces 404, one or more input / output interfaces 405, and one or more keyboards 406.

[0170] Specifically, in this embodiment, the wound healing process prediction device includes a memory and one or more programs, wherein one or more programs are stored in the memory, and one or more programs may include one or more modules, and each module may include a series of computer-executable instructions for the wound healing process prediction device, and is configured to be executed by one or more processors. The one or more programs include computer-executable instructions for performing the following:

[0171] A target image sequence is acquired, which consists of target images at different time points containing the user's trauma site and a standard reference object;

[0172] Using the first module of the pre-trained prediction model, local feature extraction and global feature extraction are performed on each target image in the target image sequence to obtain local feature vectors and global feature vectors.

[0173] Using the second module of the pre-trained prediction model, based on different preset extraction periods, temporal feature extraction processing is performed on the local feature vector and the global feature vector respectively to obtain multiple temporal feature vectors;

[0174] Using the pre-trained prediction model, based on the multiple temporal feature vectors, predictive information on the healing status of the wound site is generated;

[0175] Based on the healing prediction information, a treatment plan is determined for the user's wound site.

[0176] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for predicting the wound healing process are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0177] This specification provides an embodiment of a wound healing process prediction device. It acquires a target image sequence, which consists of target images at different time points containing the user's wound site and a standard reference object. Using a first module of a pre-trained prediction model, local and global feature extraction processes are performed on each target image in the target image sequence to obtain local and global feature vectors. Using a second module of the pre-trained prediction model, temporal feature extraction processes are performed on the local and global feature vectors based on different preset extraction periods to obtain multiple temporal feature vectors. Based on these multiple temporal feature vectors, the pre-trained prediction model generates prediction information on the healing status of the wound site. Based on the healing status prediction information, a treatment plan for the user's wound site is determined. In this way, on the one hand, since the target image contains a standard reference object, the accuracy of wound identification can be improved by comparing the wound site with the standard reference object. On the other hand, through the prediction model, based on both local and global features, temporal features can be extracted through multiple extraction cycles to obtain a temporal feature vector that can characterize the change pattern of the wound site over time. Then, through the temporal feature vector, the healing status of the wound site can be accurately predicted, and corresponding treatment plans can be generated in a targeted manner based on the healing status prediction information, thereby improving the wound care effect.

[0178] Furthermore, based on the above Figures 1 to 2 The method shown in this specification, along with one or more embodiments, also provides a storage medium for storing computer-executable instruction information. In one specific embodiment, the storage medium can be a USB flash drive, optical disc, hard disk, etc. When the computer-executable instruction information stored in the storage medium is executed by a processor, it can achieve the following process:

[0179] A target image sequence is acquired, which consists of target images at different time points containing the user's trauma site and a standard reference object;

[0180] Using the first module of the pre-trained prediction model, local feature extraction and global feature extraction are performed on each target image in the target image sequence to obtain local feature vectors and global feature vectors.

[0181] Using the second module of the pre-trained prediction model, based on different preset extraction periods, temporal feature extraction processing is performed on the local feature vector and the global feature vector respectively to obtain multiple temporal feature vectors;

[0182] Using the pre-trained prediction model, based on the multiple temporal feature vectors, predictive information on the healing status of the wound site is generated;

[0183] Based on the healing prediction information, a treatment plan is determined for the user's wound site.

[0184] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the above-described storage medium embodiment is basically similar to the method embodiment, so the description is relatively simple; relevant parts can be referred to the description of the method embodiment.

[0185] This specification provides a storage medium for acquiring a target image sequence, wherein the target image sequence consists of target images containing the user's wound site and a standard reference object at different time points. A first module of a pre-trained prediction model is used to perform local feature extraction and global feature extraction processing on each target image in the target image sequence, obtaining local feature vectors and global feature vectors. A second module of the pre-trained prediction model is used to perform temporal feature extraction processing on the local feature vectors and global feature vectors based on different preset extraction periods, obtaining multiple temporal feature vectors. Based on the multiple temporal feature vectors, the pre-trained prediction model generates prediction information on the healing status of the wound site. Based on the healing status prediction information, a treatment plan for the user's wound site is determined. In this way, on the one hand, since the target image contains a standard reference object, the accuracy of wound identification can be improved by comparing the wound site with the standard reference object. On the other hand, through the prediction model, based on both local and global features, temporal features can be extracted through multiple extraction cycles to obtain a temporal feature vector that can characterize the change pattern of the wound site over time. Then, through the temporal feature vector, the healing status of the wound site can be accurately predicted, and corresponding treatment plans can be generated in a targeted manner based on the healing status prediction information, thereby improving the wound care effect.

[0186] Furthermore, based on the above Figures 1 to 2 The method shown in this specification, along with one or more embodiments, also provides a computer program product including a computer program that, when executed by a processor, performs the following process:

[0187] A target image sequence is acquired, which consists of target images at different time points containing the user's trauma site and a standard reference object;

[0188] Using the first module of the pre-trained prediction model, local feature extraction and global feature extraction are performed on each target image in the target image sequence to obtain local feature vectors and global feature vectors.

[0189] Using the second module of the pre-trained prediction model, based on different preset extraction periods, temporal feature extraction processing is performed on the local feature vector and the global feature vector respectively to obtain multiple temporal feature vectors;

[0190] Using the pre-trained prediction model, based on the multiple temporal feature vectors, predictive information on the healing status of the wound site is generated;

[0191] Based on the healing prediction information, a treatment plan is determined for the user's wound site.

[0192] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the above-described embodiment of a computer program product is relatively simple in description because it is fundamentally similar to the method embodiment; relevant parts can be referred to the description of the method embodiment.

[0193] This specification provides a computer program product that acquires a target image sequence, wherein the target image sequence consists of target images containing the user's wound site and a standard reference object at different time points. Using a first module of a pre-trained prediction model, local and global feature extraction processes are performed on each target image in the target image sequence to obtain local and global feature vectors. Using a second module of the pre-trained prediction model, temporal feature extraction processes are performed on the local and global feature vectors based on different preset extraction periods to obtain multiple temporal feature vectors. Using the pre-trained prediction model, based on the multiple temporal feature vectors, prediction information on the healing status of the wound site is generated. Based on the healing status prediction information, a treatment plan for the user's wound site is determined. In this way, on the one hand, since the target image contains a standard reference object, the accuracy of wound identification can be improved by comparing the wound site with the standard reference object. On the other hand, through the prediction model, based on both local and global features, temporal features can be extracted through multiple extraction cycles to obtain a temporal feature vector that can characterize the change pattern of the wound site over time. Then, through the temporal feature vector, the healing status of the wound site can be accurately predicted, and corresponding treatment plans can be generated in a targeted manner based on the healing status prediction information, thereby improving the wound care effect.

[0194] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0195] In the 1990s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the methodology). However, with technological advancements, many methodological improvements today can be considered direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved methodology into the hardware circuit. Therefore, it cannot be said that a methodological improvement cannot be implemented using hardware physical modules. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is such an integrated circuit whose logic function is determined by the user programming the device. Designers can program and "integrate" a digital system onto a PLD themselves, without needing chip manufacturers to design and manufacture dedicated integrated circuit chips. Furthermore, nowadays, instead of manually manufacturing integrated circuit chips, this programming is mostly implemented using "logic compiler" software. Similar to the software compiler used in program development, the original code before compilation must also be written in a specific programming language, called a Hardware Description Language (HDL). There are many HDLs, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also understand that by simply performing some logic programming on the method flow using one of these hardware description languages ​​and programming it into an integrated circuit, the hardware circuit implementing the logical method flow can be easily obtained.

[0196] The controller can be implemented in any suitable manner. For example, it can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code form, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included therein for implementing various functions can also be considered as structures within the hardware component. Alternatively, the means for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0197] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0198] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, one or more embodiments of this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, one or more embodiments of this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0199] Embodiments in this specification are described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable parallel device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable parallel device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0200] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0201] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0202] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0203] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, one or more embodiments of this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, one or more embodiments of this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0204] One or more embodiments of this specification can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a particular task or implement a particular abstract data type. One or more embodiments of this specification can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0205] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0206] The above description is merely an embodiment of this specification and is not intended to limit this document. Various modifications and variations can be made to this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of the claims of this specification.

Claims

1. A device for predicting the wound healing process, characterized in that, The device comprises: a sequence acquisition module configured to acquire a target image sequence, the target image sequence comprising target images of a wound site of a user and a standard reference object at different time points; a first extraction module configured to perform local feature extraction and global feature extraction on each target image in the target image sequence by using a first module of a pre-trained prediction model, to obtain a local feature vector and a global feature vector; a second extraction module configured to perform time sequence feature extraction on the local feature vector and the global feature vector based on different preset extraction periods by using a second module of the pre-trained prediction model, to obtain a plurality of time sequence feature vectors; an information prediction module configured to generate healing condition prediction information for the wound site based on the plurality of time sequence feature vectors by using the pre-trained prediction model; a scheme determination module configured to determine a treatment scheme for the wound site of the user according to the healing condition prediction information. The second module comprises a third submodule constructed by a convolutional neural network and a fourth submodule constructed by a long short-term memory network, the plurality of time sequence feature vectors comprise time sequence feature vectors obtained by performing time sequence feature extraction on the local feature vector and the global feature vector according to a preset first extraction period by using the third submodule, and time sequence feature vectors obtained by performing time sequence feature extraction on the local feature vector and the global feature vector according to a preset second extraction period by using the fourth submodule, the first extraction period being less than the second extraction period.

2. The apparatus of claim 1, wherein, The first module comprises a first submodule constructed by a residual neural network and a second submodule constructed by a visual Transformer, the first submodule being configured to perform local feature extraction on each target image in the target image sequence to obtain the local feature vector, and the second submodule being configured to perform local feature extraction on each target image in the target image sequence to obtain the global feature vector.

3. The apparatus of claim 1, wherein, The device further comprises: a sample acquisition module configured to acquire a sample image sequence; a data enhancement module configured to perform data enhancement processing on sample images in the sample image sequence according to a preset data enhancement rule to obtain an enhanced sample image sequence, the preset data enhancement rule at least comprising image rotation, color jitter, brightness adjustment, and contrast adjustment; a model acquisition module configured to acquire a first prediction model, the first prediction model being trained according to first sample data, a data amount of the first sample data being greater than a data amount of the sample image sequence; a model training module configured to train the first prediction model according to the enhanced sample image sequence to obtain the pre-trained prediction model.

4. The apparatus of claim 1, wherein, The time sequence feature vectors extracted by the third submodule are used to represent spatial structure features of the wound site, and the time sequence feature vectors extracted by the fourth submodule are used to represent change conditions of wound features of the wound site.

5. A method of predicting a wound healing process, characterized by, The method comprises: acquire a target image sequence, the target image sequence being composed of target images containing a trauma site of a user and a standard reference object at different time points; perform local feature extraction and global feature extraction on each target image in the target image sequence respectively by using a first module of a pre-trained prediction model, to obtain a local feature vector and a global feature vector; perform time sequence feature extraction on the local feature vector and the global feature vector respectively based on different preset extraction periods by using a second module of the pre-trained prediction model, to obtain a plurality of time sequence feature vectors; generate healing condition prediction information for the trauma site based on the plurality of time sequence feature vectors by using the pre-trained prediction model; determine a treatment scheme for the trauma site of the user according to the healing condition prediction information; the second module includes a third submodule constructed by a convolutional neural network and a fourth submodule constructed by a long short-term memory network, the plurality of time sequence feature vectors include time sequence feature vectors obtained by performing time sequence feature extraction on the local feature vector and the global feature vector respectively according to a preset first extraction period by using the third submodule, and time sequence feature vectors obtained by performing time sequence feature extraction on the local feature vector and the global feature vector respectively according to a preset second extraction period by using the fourth submodule, the first extraction period being less than the second extraction period.

6. The method of claim 5, wherein, The determination of the treatment scheme for the trauma site of the user according to the healing condition prediction information includes: determining a wound care rule corresponding to the trauma site of the user according to the healing condition prediction information; acquiring multi-modal user data corresponding to the user; wherein the multi-modal user data includes individual feature data, medical history information, laboratory examination results and living habit data of the user; determining the treatment scheme for the trauma site of the user according to the wound care rule and the multi-modal user data by using a pre-trained scheme determination model, the scheme determination model being a model trained by reinforcement learning.

7. The method of claim 5, wherein, The healing condition prediction information includes a healing trend curve, and the generation of the healing condition prediction information for the trauma site based on the plurality of time sequence feature vectors by using the pre-trained prediction model includes: determining a healing condition corresponding to each prediction time point based on the time point corresponding to each target image in the target image sequence and the plurality of time sequence feature vectors by using the pre-trained prediction model; generating the healing trend curve according to the healing condition corresponding to each prediction time point.

8. The method of claim 5, wherein, The local feature vector is used to represent the definition of the wound edge of the trauma site, the texture change of the tissue and the distribution of the color, and the global feature vector is used to represent the skin condition around the trauma site.

Images