Burn assessment method and system based on image recognition

Abstract

The invention discloses a burn assessment method and system based on image recognition, relates to the technical field of burn assessment, and realizes automatic calculation and grading judgment of pain grades by constructing a fusion index and a joint pain response index psi and mapping the fusion index and the joint pain response index psi into a standardized scoring index Ps. The scoring mechanism not only has standard uniformity and cross-patient comparison ability, but also has self-learning and error suppression ability through a subsequent correction mechanism, so that the system is more stable and credible. According to the method, a dynamic backtracking analysis mechanism of the psychological state is specially set, abnormal scores caused by psychological fluctuations such as anxiety and nervous tension can be recognized, and the score result is corrected through an emotion regulation function. The mechanism solves the problem that an existing AI evaluation system cannot distinguish'real pain 'and'emotional response', effectively prevents excessive response or misjudgment intervention, and improves the clinical applicability of an evaluation result.

Description

Technical Field

[0001] The present invention relates to the technical field of burn assessment, and in particular to a burn assessment method and system based on image recognition. Background Art

[0002] Traditional burn assessment methods rely primarily on medical staff's judgment and visual observation, which can lead to significant errors in scenarios characterized by subjectivity and inaccurate accuracy, particularly in children and patients with unconsciousness or limited expression. This method, through the construction of a fusion image recognition system, not only identifies burn tissue features but also incorporates models for analyzing facial muscle tension and expression to construct a comprehensive representation of the human pain response, further enabling refined assisted diagnosis and treatment in complex clinical scenarios.

[0003] In current burn clinical practice, physicians primarily rely on visualizing changes in burn area size, color, and morphology to assess injury severity. However, this process relies heavily on empirical judgment and lacks objective quantitative evidence. While some studies have used image recognition methods to analyze burn depth and area, these assessments are limited to physical injury characterization and fail to incorporate the patient's subjective pain perception, a key therapeutic consideration. This is particularly true for children, comatose patients, or those with communication disorders, making it difficult for healthcare professionals to understand their true feelings, potentially leading to ineffective pain management or delayed intervention.

[0004] The primary root cause of these issues is that medical image recognition systems generally employ single-region, single-view modeling logic, focusing solely on extracting pathological features within the burn area while ignoring the dynamic signal source of "human feedback behavior." In particular, they lack the ability to model non-verbal information. This single-minded structure easily leads to model outputs that deviate from the patient's actual experience. When a patient experiences severe pain but an image assessment only indicates "moderate burns," clinical intervention is delayed, and critical opportunities for analgesia, blood pressure reduction, and intravenous fluid administration may be missed. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the present invention provides a burn assessment method and system based on image recognition, which solves the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a burn assessment method based on image recognition, comprising the following steps:

[0007] S1. Acquire the original image Ty using a medical image acquisition device, perform preprocessing on it, and obtain the medical image It. Use an image segmentation algorithm to spatially separate the medical image It and obtain the facial image Ifa and the burn image Ibu.

[0008] S2. Extract features from the burn image Ibu using a deep neural network and fit it into a damage feature set Wb;

[0009] S3, extract features from the facial image Ifa by using the expression recognition model and fit it into the psychological feature set Wa;

[0010] S4, fusing the acquired injury feature set Wb and psychological feature set Wa to obtain the pain response index Ψ;

[0011] S5. Evaluate and analyze the obtained pain response index Ψ, use an exponential scaling function to obtain a scoring index Ps, and determine an abnormal state of the score;

[0012] S6. Analyze abnormal conditions that occur during the evaluation of the pain response index Ψ to obtain a final pain response score nPs.

[0013] Preferably, S1 includes S11 and S12;

[0014] S11, using a medical image acquisition device to collect an original image Ty, including burn area and facial information, and preprocessing the original image Ty, including noise filtering, color normalization, and image enhancement, to obtain a medical image It;

[0015] The noise filtering is performed by filtering the noise in the original image Ty using a median filtering method to remove the noise in the original image Ty;

[0016] Color normalization: By using color normalization methods, all images are unified into the same color space. This eliminates color shifts caused by different light sources, ensuring a consistent representation of burn colors in different scenes. This ensures consistent color contrast between facial skin and burn areas, providing a stable foundation for feature extraction.

[0017] Image enhancement uses histogram equalization to redistribute the grayscale histogram of the image and enhance the contrast between dark and bright areas.

[0018] Preferably, S12, performing structural recognition and partitioning on the medical image It by using an image segmentation algorithm to obtain a facial image Ifa and a burn image Ibu;

[0019] Among them, the facial image Ifa is used for expression and muscle tension analysis, and classification mask extraction is achieved through semantic segmentation;

[0020] The burn image Ibu is used for burn depth and pathological feature analysis, and classification mask extraction is achieved through semantic segmentation. The facial image Ifa is obtained using the following formula:

[0021]

[0022] Where Mfa represents the facial region mask function, Represents pixel-level dot multiplication operation; mask application;

[0023] The burn image Ibu is obtained by the following formula:

[0024]

[0025] Where Mbu represents the burn area mask function.

[0026] Preferably, S2 includes S21 and S22;

[0027] S21. Using a composite deep neural network model, we extract 3D convolutional features from the burn image Ibu, including the tissue edema heterogeneity index Bsu, the necrotic pigment concentration parameter Bcu, and the microtension disturbance factor Btu of the injured area.

[0028] The tissue edema heterogeneity index (Bsu) is obtained by calculating the reflectivity variation difference of the edema area using the structure tensor. The index significance is: the intensity of brightness fluctuation caused by the difference in water content within the tissue, which is used to reflect the uneven edema after burns. The characteristic area is: the boundary between the low-frequency reflection area and the high-brightness saturation area.

[0029] The tissue edema heterogeneity index Bsu is obtained by the following formula:

[0030]

[0031] Where Z represents the normalization factor, Ibu(i, j) represents the intensity value of the pixel in the i-th row and j-th column of the burn image Ibu, reflecting the brightness of the epidermal area, μloc(i, j) represents the local mean of the neighborhood of the pixel in the i-th row and j-th column of the burn image Ibu, representing the reference background value of the edema area, and Wed(i, j) represents the edge perception weight of the pixel in the i-th row and j-th column of the burn image Ibu, highlighting the tissue boundary and enhancing the structural response.

[0032] The tissue edema heterogeneity index Bsu measures the discrete degree of brightness fluctuations in different areas of the burn image through image processing, thereby quantifying the spatial distribution differences in water content changes between tissues. In burn tissue, edema is a typical manifestation of the body's acute inflammatory response. When a burn occurs, capillary permeability increases, and plasma exudation causes fluid accumulation in local tissue gaps. This phenomenon is often accompanied by increased brightness, blurred structure, and uneven reflection in the water-containing area. Therefore, the tissue edema heterogeneity index Bsu is needed;

[0033] The tissue edema heterogeneity index (Bsu) measures the brightness difference between each pixel in the image and its neighborhood, combined with edge structure sensitivity, to measure the unevenness of edema in the entire burn area.

[0034] The necrotic pigment concentration parameter Bcu constructs a color anomaly model based on the image color channel and extracts the non-physiological distribution cluster center in the RGB space. The indicator significance: measures the deposition intensity of black-brown pigment clusters after tissue epidermal necrosis. Characteristic area: low brightness, saturated color with a reddish-brown color.

[0035] The necrotic pigment concentration parameter Bcu is obtained by the following formula:

[0036]

[0037] Where N represents the total number of pixels in the calculation area, C(i, j) represents the color vector of the pixel in the i-th row and j-th column of the burn image Ibu in the RGB color space, and Cref represents the reference skin color vector, which represents the standard color of the healthy area.

[0038] The necrotic pigment concentration parameter Bcu is an overall quantitative expression of the degree of color abnormality in burn images. A larger value indicates a more significant color difference from healthy skin, a darker color of the suspected necrotic area, a denser distribution, and a deeper or more complex tissue abnormality.

[0039] The formula for determining the necrotic pigment concentration parameter, Bcu, is based on a difference analysis in color vector space. First, the system selects an undamaged healthy skin area from the burn image and calculates its color mean as the reference color vector, Cref. Then, for each pixel within the burn area, its color value, C(i, j), in RGB space is extracted and a Euclidean distance calculation is performed with the reference color. The squared distance represents the degree to which the pixel's color deviates from the healthy skin color.

[0040] The squared color differences of all pixels are summed and averaged to obtain the final necrotic pigment concentration parameter, Bcu, which represents the "abnormal pigment concentration" of the burn area as a whole. A larger value indicates a greater difference in color from normal skin, heavier pigmentation, and deeper or more extensive tissue necrosis. This formula combines physiological significance, computational simplicity, and image adaptability, making it suitable for automated assessment tasks across different skin tones and lighting conditions.

[0041] The microtension perturbation factor (Btu) of the injured area is acquired by constructing a microtension field using a Gabor ripple convolution kernel and a texture direction map. The indicator's significance: It assesses the directional deformation abnormalities of the subcutaneous muscle and connective tissue in burns. Characteristic areas: irregular curved morphology and high-frequency texture distortion areas.

[0042] The microtension disturbance factor Btu of the injured area is obtained by the following formula:

[0043]

[0044] Where Tga(i, j) represents the Gabor filter response value at the pixel in the i-th row and j-th column of the burn image Ibu, Ω represents the integration area, dA represents the integration unit area, Represents the gradient of the Gabor filter result in a specific direction θ;

[0045] The microtension perturbation factor (Btu) in the injured area refers to changes in the previously uniform and orderly collagen arrangement, vascular orientation, and epidermal tension distribution caused by burns, particularly those involving the dermis or even subcutaneous tissue. These changes typically manifest in images as disturbed texture orientation, fluctuating edge orientation, and regional structural tension distortion. These phenomena are difficult to detect with the naked eye, but can be effectively perceived and quantified using a directional filter.

[0046] Gabor filter response extraction Tga(i, j): Apply a Gabor filter to the burn image Ibu to obtain the response value of each pixel in a specific direction (such as 0°, 45°, 90°). The Gabor filter simulates the visual system's response to "texture direction and frequency" and is particularly suitable for analyzing image textures with directional changes. The filtered response map Tga can reflect the texture "intensity" and "directional regularity" of the area.

[0047] Gradient calculation At each pixel, the first-order gradient of the Gabor response map in the direction θ is calculated, that is, the rate of change of the response value; this represents the "disturbance degree" of the texture in the area - if the gradient is large, it means that the texture direction at this location has changed suddenly, and there may be fiber distortion or texture breakage.

[0048] S22. By constructing a nonlinear cross-response function Fcomp, the tissue edema heterogeneity index Bsu, the necrotic pigment concentration parameter Bcu, and the injury area microtension perturbation factor Btu are mapped into a structurally stable high-dimensional damage feature set Wb;

[0049] The damage feature set Wb is obtained by the following formula:

[0050] Wb=Fcomp(Bsu, Bcu, Btu).

[0051] Preferably, S3 includes S31 and S32;

[0052] S31, extracting facial action features from the facial image Ifa by using the expression recognition model, including the contraction intensity of the orbicularis oculi muscle Foa, the activity of the zygomatic major muscle Fza, and the tension fluctuation of the mandibular muscle Fda;

[0053] The contraction intensity of the orbicularis oculi muscle, Foa, is acquired through the orbicularis oculi muscle AU6. The indicator definition reflects the degree of eye tightening and is used to identify instantaneous eye closure reactions caused by pain.

[0054] The contraction intensity of the orbicularis oculi muscle Foa is obtained by the following formula:

[0055]

[0056] Where Aorb(Ifa) represents the orbicularis oculi action recognition function, M represents the number of continuous sampling frames, AU6t represents the activation intensity of the orbicularis oculi action unit in the t-th frame, and Weye(t) represents the weight function of the eye muscles.

[0057] Orbicularis oculi muscle contraction intensity (Foa) is a parameter used to quantify the intensity of the eye muscle response to pain or severe discomfort. It is specifically used to capture the tension, squeezing, and contraction of the orbicularis oculi muscle. This muscle movement is highly sensitive in non-verbal expressions of pain.

[0058] The formula for obtaining the orbicularis oculi muscle contraction intensity Foa is to extract the AU6 activation intensity from continuous facial image frames, and perform weighted averaging based on the quality and credibility of each frame image to calculate the overall intensity of eye muscle contraction during the entire time period.

[0059] The zygomatic major muscle activity Fza is obtained by collecting data from the zygomatic major muscle AU12. The indicator definition: reflects the strength of the mandibular lift during a smile or tense expression, reflecting the degree of body tension.

[0060] The zygomatic major muscle activity Fza is obtained by the following formula:

[0061] Fza=Azgy(Ifa)=max(t)(AU12t)*ηsym;

[0062] Where Azgy(Ifa) represents the zygomatic major muscle action extraction function, max(t) represents the maximum value of the zygomatic major muscle in all frames, and ηsym represents the facial muscle symmetry factor, which measures the consistency of the left and right facial muscle movements.

[0063] The zygomatic major muscle activity (Fza) is an indicator used to measure the degree of facial zygomatic major muscle activity over a specific time period. This muscle is a core muscle group that controls facial movements such as smiling, tensing and lifting, and muscle stretching. In burn pain assessment, its activity not only reflects emotional tension but may also reveal underlying muscle twitching, anxiety reactions, or involuntary expressions. It is particularly useful for identifying the psychological state of individuals who are nonverbal.

[0064] The formula for obtaining the zygomatic major muscle activity Fza is based on the activation intensity of AU12 (zygomatic major muscle action unit) in facial expression recognition, combined with the symmetry of muscle activity on the left and right sides of the face, to measure whether a person has a strong and credible zygomatic pull expression during the observation period, thereby reflecting their potential pain, anxiety or tension emotional response.

[0065] Mandibular muscle tension fluctuation Fda is acquired through mouth angle pull-down AU15 and mandibular opening AU26. Indicator definition: Quantifies the frequency and amplitude of fluctuations in mandibular movement, representing subjective tension, anxiety, or extreme pain reactions.

[0066] The mandibular muscle tension fluctuation Fda is obtained by the following formula:

[0067]

[0068] Where Adep(Ifa) represents the mandibular tension action evaluation function, T represents the total number of time frames, xAU15t represents the rate of change of the mouth corner pull-down activation in the t-th frame; xAU26t represents the rate of change of the mandibular pull-down activation in the t-th frame;

[0069] Mandibular muscle tension fluctuation (Fda) is a dynamic parameter used to assess the degree of muscle tension in the mandibular region during an individual's facial expressions, focusing specifically on nonverbal expressions of pain, emotional fluctuations, or neurological stress. It is calculated based on the combined rate of change of facial action units AU15 (mouth corner droop) and AU26 (jaw opening). It reflects frequent, intense, or unstable mandibular muscle activity and is a key indicator of emotional instability, pain sensitivity, or facial tremor.

[0070] The formula for obtaining the mandibular muscle tension fluctuation Fda is to capture the change speed of the two action units AU15 (pull down the corners of the mouth) and AU26 (open the mandible) in the time dimension, reflecting whether the facial muscles frequently contract, relax, and twitch in a short period of time, thereby judging whether the patient's emotional tension or pain expression is unstable or intense.

[0071] S32, inputting the acquired orbicularis oculi muscle contraction intensity Foa, zygomatic major muscle activity Fza, and mandibular muscle tension fluctuation Fda into the nonlinear modeling function Memo to construct a psychological feature set Wa;

[0072] The psychological feature set Wa is obtained by the following formula:

[0073] Wa=Memo(Foa,Fza,Fda).

[0074] Preferably, S4 includes S41 and S42;

[0075] S41, normalize the obtained damage feature set Wb and psychological feature set Wa, unify the dimensions and numerical ranges, and obtain a new damage feature set nWb and a new psychological feature set nWa;

[0076] The new damage feature set nWb is obtained by the following formula:

[0077]

[0078] Where nWb(c) represents the c-th data item in the new damage feature set nWb, Wb(c) represents the c-th data item in the damage feature set Wb, minWb(c) represents the valley value of the c-th data item in the damage feature set Wb, and maxWb(c) represents the peak value of the c-th data item in the damage feature set Wb.

[0079] The new psychological feature set nWa is obtained by the following formula:

[0080]

[0081] Where nWa(e) represents the e-th data item in the new psychological feature set nWa, Wa(e) represents the e-th data item in the psychological feature set nWa, minWa(e) represents the valley value of the e-th data item in the psychological feature set nWa, and maxWa(e) represents the peak value of the e-th data item in the psychological feature set nWa.

[0082] Preferably, S42, performing nonlinear fusion on the obtained new injury feature set nWb and the new psychological feature set nWa to calculate and obtain the pain response index Ψ;

[0083] The pain response index Ψ is obtained by the following formula:

[0084]

[0085] where α1, α2, and α3 represent the preset weight values ​​of the tissue edema heterogeneity index Bsu, the necrotic pigment concentration parameter Bcu, and the microtension disturbance factor Btu in the injured area, respectively; β1, β2, and β3 represent the preset weight values ​​of the orbicularis oculi muscle contraction intensity Foa, the zygomatic major muscle activity Fza, and the mandibular muscle tension fluctuation Fda, respectively.

[0086] Preferably, S5 includes S51 and S52;

[0087] S51. Evaluate the pain response index Ψ obtained to obtain a scoring index Ps;

[0088] The scoring index Ps is obtained by the following formula:

[0089] Ps=100*(1-e -k(Ψ-cx) );

[0090] Where e represents a constant, k represents an exponential rise rate control factor, and cx represents a response delay constant; the score is mapped to a standard score between 0 and 100.

[0091] Preferably, S52, comparing the obtained scoring index Ps with a preset scoring threshold Tps to determine an abnormal state of the scoring;

[0092] The abnormal status of the score is obtained by matching in the following ways:

[0093] When the scoring index Ps is less than the scoring threshold Tps*0.5, it indicates a normal score and normal pain detection;

[0094] When the scoring threshold Tps*0.5≤scoring index Ps≤scoring threshold Tps, it indicates a moderate score and the monitoring frequency is adjusted;

[0095] When the score index Ps exceeds the score threshold Tps, it indicates an abnormal score and needs to be corrected. Misjudgments may be caused by psychological fluctuations and excessive anxiety, and corrections need to be made and the score threshold Tps recalculated.

[0096] Preferably, S6, when the scoring index Ps is abnormal, a retrospective analysis is performed on the abnormal individual, and combined with the mandibular muscle tension fluctuation Fda, a psychological misjudgment factor psy is constructed;

[0097] The psychological misjudgment factor psy is obtained by the following formula:

[0098]

[0099] Where λ represents the weight coefficient, pFda represents the average value of mandibular muscle tension fluctuation, PL represents a positive number to avoid division by zero, and d represents the integral. represents the rate of change of pain score over time;

[0100] The final pain response score nPs was obtained by correcting the scoring index Ps using the psychological misjudgment factor psy;

[0101] The final pain response score nPs is obtained by the following formula:

[0102] nPs=Ps*[1-ω*tanh(psy)];

[0103] Where tanh() represents the hyperbolic tangent function, and ω represents the maximum callback coefficient.

[0104] A burn assessment system based on image recognition, comprising the following modules: an image preprocessing module, a burn area image recognition module, a facial expression recognition module, a joint pain response index construction module, a pain score normalization module, and a correction and optimization score module;

[0105] The image preprocessing module acquires the original image Ty through a medical image acquisition device, performs preprocessing, obtains the medical image It, and spatially separates the medical image It through an image segmentation algorithm to obtain the facial image Ifa and the burn image Ibu;

[0106] The burn area image recognition module extracts features from the burn image Ibu using a deep neural network and fits it into a damage feature set Wb;

[0107] The facial expression recognition module extracts features from the facial image Ifa using the expression recognition model and fits it into a psychological feature set Wa;

[0108] The joint pain response index construction module fuses the acquired injury feature set Wb and psychological feature set Wa to obtain the pain response index Ψ;

[0109] The pain score normalization module evaluates and analyzes the pain response index Ψ, uses the exponential scaling function to obtain the score index Ps, and determines the abnormal state of the score;

[0110] The correction and optimization scoring module analyzes abnormal conditions that occur during the evaluation of the pain response index Ψ to obtain the final pain response score nPs.

[0111] The present invention provides a burn assessment method and system based on image recognition, which has the following beneficial effects:

[0112] (1) The system extracts histopathological features from burn images, including the tissue edema heterogeneity index Bsu, the necrotic pigment concentration parameter Bcu, and the microtension perturbation factor Btu of the injured area. It also introduces facial expression features from facial image recognition, including the contraction intensity of the orbicularis oculi muscle Foa, the activity of the zygomatic major muscle Fza, and the fluctuation of the mandibular muscle tension Fda, to achieve multimodal information fusion modeling of pain perception. This improvement effectively compensates for the defects of existing assessment systems that only focus on lesion morphology and ignore the patient's subjective response, enabling the system to identify "invisible pain", which is particularly suitable for special groups such as children and comatose patients.

[0113] This method realizes the automatic calculation and grading judgment of pain levels by constructing a fusion index, combining the pain response index Ψ, and mapping it into a standardized scoring index Ps. The scoring mechanism not only has the standard uniformity and cross-patient comparison capabilities, but also has the self-learning and error suppression capabilities through the subsequent correction mechanism, making the system more stable and reliable. The invention specially sets up a dynamic retrospective analysis mechanism of psychological state, which can identify abnormal scores caused by psychological fluctuations, such as anxiety and nervous tension, and correct the scoring results through the emotion regulation function. This mechanism solves the problem that the current AI assessment system cannot distinguish between "real pain" and "emotional response", effectively prevents over-response or misjudgment of intervention, and improves the clinical applicability of the assessment results.

[0114] (2) Through image preprocessing technologies such as median filtering, color normalization and image enhancement, the problem of unstable image quality of burn images under different equipment, lighting and background conditions is effectively solved. The image processing chain provided by the present invention can uniformly convert the original image into medical image data with consistent visual quality and contrast clarity, ensuring the stability, repeatability and clinical usability of the subsequent model analysis results. This processing process greatly reduces the recognition error caused by environmental factors and improves the standardization of model input. By introducing the semantic segmentation algorithm, the facial image Ifa and the burn image Ibu are automatically extracted and partitioned, and no longer rely on the doctor or operator to manually circle or crop the target area, which significantly improves the operation efficiency and system intelligence level. In particular, for complex burn scenes, such as facial adjacent burns and neck extension burns, the structural partitioning algorithm can clearly divide the boundary area, avoid information confusion or misuse during feature extraction, and provide a precise input source for subsequent feature extraction.

[0115] (3) By using a composite deep neural network and texture direction perception modules such as Gabor filtering, stable recognition of key lesion areas in complex image backgrounds is achieved. This method significantly improves the sensitivity to image texture changes and the ability to maintain boundaries. Even when the boundaries between the burn area and normal tissue are blurred and the color gradient is complex, the model can still stably extract high-quality structural features, solving the inherent defects of traditional image recognition in terms of "unclear edge discrimination and weak recognition of morphological perturbations". Facial action units (AUs) are used to extract dynamic parameters of key muscle areas, especially for muscle groups closely related to pain and anxiety, such as the orbicularis oculi, zygomaticus major, and mandibular muscles. A quantitative expression mechanism is established, thereby achieving an objective reflection of the pain state of non-complaining patients. This mechanism breaks the previous assessment limitation of "only focusing on lesions and ignoring expressions" and truly introduces the individual's facial physiological feedback information when in pain, providing an important supplement to pain recognition and grading beyond the subjective level.

[0116] (4) Using a nonlinear fusion function, the injury feature set and the psychological feature set are cross-mapped to comprehensively generate a joint pain response index. This approach avoids the problem of over-reliance on a certain type of feature in the traditional weighted average model, and instead simulates the physiological-psychological linkage effect caused by burns through nonlinear interactions. This mechanism significantly improves the model's ability to perceive complex pain manifestations, especially in identifying atypical scenarios such as "mild injury but severe pain" or "severe injury but slow response".

[0117] By individually weighting each physiological and psychological trait, the model possesses an adjustable and configurable structure, enabling it to adapt to diverse assessment objectives for different burn types, patients of different age groups, or specific clinical needs. This mechanism also provides physicians and researchers with "structural transparency" of the scoring sources, allowing them to clearly track which factors drive specific scores, aiding clinical decision-making and model optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0118] Figure 1 This is a schematic diagram of the steps of a burn assessment method based on image recognition according to the present invention;

[0119] Figure 2 A schematic diagram of the process for final pain response scoring of the present invention;

[0120] Figure 3 is a bar graph of the pain response index of the present invention. DETAILED DESCRIPTION

[0121] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0122] Example 1

[0123] The present invention provides a burn assessment method based on image recognition, please refer to Figures 1 to 3 , including the following steps:

[0124] S1. Acquire the original image Ty using a medical image acquisition device, perform preprocessing on it, and obtain the medical image It. Use an image segmentation algorithm to spatially separate the medical image It and obtain the facial image Ifa and the burn image Ibu.

[0125] S2. Using a deep neural network, we extract features from the burn image Ibu, including the tissue edema heterogeneity index Bsu, the necrotic pigment concentration parameter Bcu, and the microtension perturbation factor Btu in the injured area, and fit them into the injury feature set Wb.

[0126] S3. By using the expression recognition model, the facial image Ifa is extracted, including the contraction intensity of the orbicularis oculi muscle Foa, the activity of the zygomatic major muscle Fza, and the fluctuation of the mandibular muscle tension Fda, and fitting them into the psychological feature set Wa;

[0127] S4, fusing the acquired injury feature set Wb and psychological feature set Wa, and constructing a joint pain response index Ψ through variable nonlinear mapping;

[0128] S5. Evaluate and analyze the pain response index Ψ obtained, use an exponential scaling function to obtain a scoring index Ps, and compare it with a preset scoring threshold Tps to determine whether the score is abnormal;

[0129] S6. Perform a modified retrospective analysis on abnormal conditions that occur during the evaluation of the pain response index Ψ to obtain a final pain response score nPs.

[0130] This example not only extracts histopathological features from burn images, including the tissue edema heterogeneity index (Bsu), the necrotic pigment concentration parameter (Bcu), and the injury area microtension perturbation factor (Btu), but also incorporates facial expression features from facial image recognition, including the contraction intensity of the orbicularis oculi muscle (Foa), the activity of the zygomatic major muscle (Fza), and the fluctuation of mandibular muscle tension (Fda). This enables multimodal information fusion modeling of pain perception. This improvement effectively overcomes the shortcomings of existing assessment systems that focus solely on lesion morphology and ignore the patient's subjective response. This system is capable of identifying "invisible pain," making it particularly suitable for special groups such as children and comatose patients.

[0131] Traditional burn assessment relies on the experience of doctors, is highly subjective, and has inconsistent assessment standards. This method constructs a fusion index, combines the pain response index Ψ, and maps it into a standardized scoring index Ps, thereby realizing the automatic calculation and grading judgment of pain levels. The scoring mechanism not only has the standard uniformity and cross-patient comparison capabilities, but also has the self-learning and error suppression capabilities through the subsequent correction mechanism, making the system more stable and reliable. The invention specifically sets up a dynamic retrospective analysis mechanism for psychological state, which can identify abnormal scores caused by psychological fluctuations, such as anxiety and nervous tension, and correct the scoring results through the emotion regulation function. This mechanism solves the problem that the current AI assessment system cannot distinguish between "real pain" and "emotional response", effectively prevents over-response or misjudgment of intervention, and improves the clinical applicability of the assessment results.

[0132] The data that this method relies on are only raw images, and no additional physiological detection hardware is required, which means that it can be widely used in scenarios such as telemedicine, mobile terminals, and self-service diagnosis and treatment. Especially in situations where conditions are limited, such as primary medical care, battlefield emergency treatment, and home care, this method can quickly provide structured pain indicators to assist doctors in remotely determining the degree of burns and intervention priorities, and promote the sinking and popularization of assessment standards. Unlike the static model of the existing system that "only outputs but does not feedback", the present invention forms a closed-loop processing flow of "detection-judgment-correction-re-output" through abnormal state identification and retrospective correction. This mechanism enhances the model's adaptability to individual differences, can continuously optimize assessment performance, and prevent the scoring system from failing under complex expressions or atypical injuries, reflecting strong intelligence and generalization capabilities.

[0133] Example 2

[0134] This embodiment is explained in Example 1, please refer to Figure 1 ,Specifically: S1 includes S11 and S12;

[0135] S11, using a medical image acquisition device to collect an original image Ty, including burn area and facial information, and preprocessing the original image Ty, including noise filtering, color normalization, and image enhancement, to obtain a medical image It;

[0136] The noise filtering is performed by filtering the noise in the original image Ty using a median filtering method to remove the noise in the original image Ty;

[0137] Color normalization: All images are unified into the same color space by using the color normalization method;

[0138] Image enhancement uses histogram equalization to redistribute the grayscale histogram of the image and enhance the contrast between dark and bright areas.

[0139] S12, performing structural recognition and partitioning on the medical image It by using an image segmentation algorithm to obtain a facial image Ifa and a burn image Ibu;

[0140] Among them, the facial image Ifa is used for expression and muscle tension analysis, and classification mask extraction is achieved through semantic segmentation;

[0141] The burn image Ibu is used for burn depth and pathological feature analysis, and classification mask extraction is achieved through semantic segmentation. The facial image Ifa is obtained using the following formula:

[0142]

[0143] Where Mfa represents the facial region mask function, Represents pixel-level dot multiplication operation;

[0144] The burn image Ibu is obtained by the following formula:

[0145]

[0146] Where Mbu represents the burn area mask function.

[0147] In this embodiment, the image preprocessing techniques such as median filtering, color normalization and image enhancement are used to effectively solve the problem of unstable image quality of burn images under different equipment, lighting and background conditions. The image processing chain provided by the present invention can uniformly convert the original image into medical image data with consistent visual quality and contrast clarity, ensuring the stability, repeatability and clinical usability of the subsequent model analysis results. This processing process greatly reduces the recognition error caused by interference from environmental factors and improves the standardization of model input. By introducing the semantic segmentation algorithm, the facial image Ifa and the burn image Ibu are automatically extracted and partitioned, and no longer rely on the doctor or operator to manually circle or crop the target area, which significantly improves the operational efficiency and system intelligence level. In particular, for complex burn scenes, such as facial adjacent burns and neck extension burns, the structural partitioning algorithm can clearly divide the boundary area, avoid information confusion or misuse during feature extraction, and provide a precise input source for subsequent feature extraction.

[0148] During the image segmentation process, this invention simultaneously extracts facial regions for pain expression analysis and burn regions for pathology analysis. While ensuring integrated image processing, it provides independent and non-interfering data inputs for the two model modules: the injury recognition module and the emotion recognition module. This mechanism significantly enhances the system's versatility, enabling it to simultaneously meet diagnostic and analysis requirements across multiple dimensions within a single image acquisition, making it particularly suitable for efficient scenarios such as telemedicine and mobile assessments.

[0149] In traditional methods, image cropping and region identification often rely on the operator's subjective judgment and manual processing, which not only increases the workload but also easily introduces "human interference errors". The present invention uses structured semantic segmentation to achieve fully automatic region identification and extraction, significantly reducing the need for human intervention, and providing a true "end-to-end closed-loop input channel" for subsequent feature extraction and multimodal fusion. The modular design of image preprocessing and region segmentation gives this method good cross-platform and multi-scene versatility, and even under image input conditions with different resolutions, large lighting differences, and complex backgrounds, it can ensure clear structure and complete feature retention. This is of great significance for promoting the widespread deployment and actual clinical implementation of intelligent burn assessment systems.

[0150] Example 3

[0151] This embodiment is explained in Example 2, please refer to Figure 1 and Figure 2 , specifically: S2 includes S21 and S22;

[0152] S21. Using a composite deep neural network model, we extract 3D convolutional features from the burn image Ibu, including the tissue edema heterogeneity index Bsu, the necrotic pigment concentration parameter Bcu, and the microtension disturbance factor Btu of the injured area.

[0153] Among them, the tissue edema heterogeneity index Bsu is obtained by calculating the reflectivity change difference of the edema area through the structure tensor;

[0154] The tissue edema heterogeneity index Bsu is obtained by the following formula:

[0155]

[0156] Where Z represents the normalization factor, Ibu(i, j) represents the intensity value of the pixel in the i-th row and j-th column of the burn image Ibu, μloc(i, j) represents the local mean of the neighborhood of the pixel in the i-th row and j-th column of the burn image Ibu, and Wed(i, j) represents the edge perception weight of the pixel in the i-th row and j-th column of the burn image Ibu;

[0157] The necrotic pigment concentration parameter Bcu is used to construct a color anomaly model based on the image color channel and extract the non-physiological distribution cluster centers in the RGB space;

[0158] The necrotic pigment concentration parameter Bcu is obtained by the following formula:

[0159]

[0160] Where N represents the total number of pixels in the calculation area, C(i, j) represents the color vector of the pixel in the i-th row and j-th column of the burn image Ibu in the RGB color space, and Cref represents the reference skin color vector, which represents the standard color of the healthy area.

[0161] The microtension disturbance factor Btu of the injured area is acquired by constructing the microtension field using Gabor ripple convolution kernel and texture direction map;

[0162] The microtension disturbance factor Btu of the injured area is obtained by the following formula:

[0163]

[0164] Where Tga(i, j) represents the Gabor filter response value at the pixel in the i-th row and j-th column of the burn image Ibu, Ω represents the integration area, dA represents the integration unit area, Represents the gradient of the Gabor filter result in a specific direction θ;

[0165] S22. By constructing a nonlinear cross-response function Fcomp, the tissue edema heterogeneity index Bsu, the necrotic pigment concentration parameter Bcu, and the injury area microtension perturbation factor Btu are mapped into a structurally stable high-dimensional damage feature set Wb;

[0166] The damage feature set Wb is obtained by the following formula:

[0167] Wb=Fcomp(Bsu, Bcu, Btu).

[0168] S3 includes S31 and S32;

[0169] S31, extracting facial action features from the facial image Ifa by using the expression recognition model, including the contraction intensity of the orbicularis oculi muscle Foa, the activity of the zygomatic major muscle Fza, and the tension fluctuation of the mandibular muscle Fda;

[0170] Among them, the contraction intensity of the orbicularis oculi muscle Foa is obtained by collecting the orbicularis oculi muscle AU6;

[0171] The contraction intensity of the orbicularis oculi muscle Foa is obtained by the following formula:

[0172]

[0173] Where Aorb(Ifa) represents the orbicularis oculi action recognition function, M represents the number of continuous sampling frames, AU6t represents the activation intensity of the orbicularis oculi action unit in the t-th frame, and Weye(t) represents the weight function of the eye muscles.

[0174] The activity of the zygomatic major muscle Fza was obtained by collecting AU12 of the zygomatic major muscle;

[0175] The zygomatic major muscle activity Fza is obtained by the following formula:

[0176] Fza=Azgy(Ifa)=max(t)(AU12t)*ηsym;

[0177] Where Azgy(Ifa) represents the zygomatic major muscle action extraction function, max(t) represents the maximum value of the zygomatic major muscle in all frames, and ηsym represents the facial muscle symmetry factor, which measures the consistency of the left and right facial muscle movements.

[0178] Mandibular muscle tension fluctuation Fda is acquired by pulling down the corner of the mouth AU15 and opening the mandible AU26;

[0179] The mandibular muscle tension fluctuation Fda is obtained by the following formula:

[0180]

[0181] Where Adep(Ifa) represents the mandibular tension action evaluation function, T represents the total number of time frames, xAU15t represents the rate of change of the mouth corner pull-down activation in the t-th frame; xAU26t represents the rate of change of the mandibular pull-down activation in the t-th frame;

[0182] S32, inputting the acquired orbicularis oculi muscle contraction intensity Foa, zygomatic major muscle activity Fza, and mandibular muscle tension fluctuation Fda into the nonlinear modeling function Memo to construct a psychological feature set Wa;

[0183] The psychological feature set Wa is obtained by the following formula:

[0184] Wa=Memo(Foa,Fza,Fda).

[0185] In this example, by constructing three pathological parameters—the tissue edema heterogeneity index Bsu, the necrotic pigment concentration parameter Bcu, and the wound area microtension perturbation factor Btu—the model goes beyond relying solely on low-level image features such as color, area, or edges. Instead, it delves into more pathologically meaningful mid- and high-level feature dimensions, such as changes in tissue fluid content, lesion pigment distribution, and abnormal texture structure. This parametric modeling approach significantly improves the ability to distinguish burns of varying depths and stages, particularly accurately distinguishing burn types with similar appearances but distinct pathological mechanisms, enhancing the model's adaptability for clinical applications.

[0186] By utilizing a composite deep neural network and texture-direction-aware modules like Gabor filtering, the method achieves stable recognition of key lesion areas against complex image backgrounds. This method significantly improves sensitivity to image texture changes and its ability to preserve boundaries. This allows the model to consistently extract high-quality structural features even when the boundaries between burn areas and normal tissue are blurred and color gradients are complex. This overcomes the inherent shortcomings of traditional image recognition, such as unclear edge discrimination and weak recognition of morphological perturbations.

[0187] By using facial action units (AUs) to extract dynamic parameters of key muscle areas, specifically targeting muscle groups closely associated with pain and anxiety, such as the orbicularis oculi, zygomaticus major, and mandibular muscles, a quantitative expression mechanism was established, enabling objective reflection of pain status in patients without complaints. This mechanism breaks the limitations of previous assessments that focused solely on the lesion and ignored facial expressions, truly incorporating physiological feedback from the individual's face during pain, providing a valuable supplement to pain identification and grading beyond the subjective level.

[0188] By combining the three physiological motion parameters nonlinearly, this embodiment effectively simulates the synergistic activation effect of different muscle groups in pain expression, and enhances the model's ability to recognize complex facial expression combinations. This structural design is particularly suitable for processing pain response recognition under multiple overlapping expressions, emotional masking or facial asymmetry, thereby greatly improving the accuracy and wide adaptability of psychological response modeling. By independently modeling the pathological feature parameter set Wb and the psychological feature set Wa, and then uniformly inputting them into the subsequent fusion module, a clear and interpretable dual-module structural path is constructed. This structure improves the transparency of model parameter adjustment, error tracing and clinical result interpretation, so that the evaluation system has modular maintenance capabilities, flexible expansion space and good standard compatibility, and is suitable for deployment on telemedicine platforms, mobile evaluation systems or hospital auxiliary diagnosis terminals.

[0189] Example 4

[0190] This embodiment is explained in Example 3, please refer to Figure 1 and Figure 3 , specifically: S4 includes S41 and S42;

[0191] S41, normalize the obtained damage feature set Wb and psychological feature set Wa, unify the dimensions and numerical ranges, and obtain a new damage feature set nWb and a new psychological feature set nWa;

[0192] The new damage feature set nWb is obtained by the following formula:

[0193]

[0194] Where nWb(c) represents the c-th data item in the new damage feature set nWb, Wb(c) represents the c-th data item in the damage feature set Wb, minWb(c) represents the valley value of the c-th data item in the damage feature set Wb, and maxWb(c) represents the peak value of the c-th data item in the damage feature set Wb.

[0195] The new psychological feature set nWa is obtained by the following formula:

[0196]

[0197] Where nWa(e) represents the e-th data item in the new psychological feature set nWa, Wa(e) represents the e-th data item in the psychological feature set nWa, minWa(e) represents the valley value of the e-th data item in the psychological feature set nWa, and maxWa(e) represents the peak value of the e-th data item in the psychological feature set nWa.

[0198] S42, performing nonlinear fusion on the acquired new injury feature set nWb and the acquired new psychological feature set nWa, and calculating and acquiring the pain response index Ψ;

[0199] The pain response index Ψ is obtained by the following formula:

[0200]

[0201] where α1, α2, and α3 represent the preset weight values ​​of the tissue edema heterogeneity index Bsu, the necrotic pigment concentration parameter Bcu, and the microtension disturbance factor Btu in the injured area, respectively; β1, β2, and β3 represent the preset weight values ​​of the orbicularis oculi muscle contraction intensity Foa, the zygomatic major muscle activity Fza, and the mandibular muscle tension fluctuation Fda, respectively.

[0202] Specific examples:

[0203] Table 1 Pain response index calculation table:

[0204]

[0205]

[0206] In this embodiment, because injury and psychological features originate from different image regions, recognition models, and physiological mechanisms, direct fusion can lead to weight imbalances due to differences in numerical ranges and inconsistent dimensions. This embodiment uses normalization to map all features to the same numerical scale, significantly improving the model's fusion stability and comparative reliability, providing a unified and controllable input foundation for subsequent indicator calculations and weight modeling. This mechanism enables the model to output stable and consistent assessment results across different populations, devices, or image conditions.

[0207] This embodiment uses a nonlinear fusion function to cross-map the injury feature set with the psychological feature set to comprehensively generate a joint pain response index. This approach avoids the problem of over-reliance on a single feature type in traditional weighted average models, instead simulating the physiological and psychological linkage effects caused by burns through nonlinear interactions. This mechanism significantly improves the model's ability to perceive complex pain manifestations, especially in identifying atypical scenarios such as "mild injury but severe pain" or "severe injury but slow response."

[0208] By individually weighting each physiological and psychological trait, the model possesses an adjustable and configurable structure, enabling it to adapt to diverse assessment objectives for different burn types, patients of different age groups, or specific clinical needs. This mechanism also provides physicians and researchers with "structural transparency" of the scoring sources, allowing them to clearly track which factors drive specific scores, aiding clinical decision-making and model optimization.

[0209] In actual applications, facial expressions are often affected by multiple factors, such as anxiety, panic, and side effects of drugs. This embodiment effectively prevents the occurrence of the phenomenon of "emotions misleading the main judge" by introducing expression parameters into the nonlinear fusion model and combining it with a weight adjustment mechanism. The model can adaptively reduce the scoring bias caused by abnormal expressions, so that the evaluation results can more truly reflect the physiological nature of pain and have strong robustness and generalization capabilities. Through the normalization and fusion design of S4, the model input is encapsulated as a standardized vector expression. This structure is not only suitable for the current model calculation, but also facilitates the subsequent use of large amounts of evaluation data for training, optimization, and transfer learning of machine learning models. This design provides a good data foundation for future iterations of the system, enabling this method to have the ability to naturally evolve from a rule-based algorithm to a deep learning algorithm, and has high scalability.

[0210] Example 5

[0211] This embodiment is explained in Example 4. Please refer to Figure 1 and Figure 2 ,Specifically: S5 includes S51 and S52;

[0212] S51. Evaluate the pain response index Ψ obtained to obtain a scoring index Ps;

[0213] The scoring index Ps is obtained by the following formula:

[0214] Ps=100*(1-e -k(Ψ-cx) );

[0215] Where, e represents a constant, k represents an exponential rise rate control factor, and cx represents a response delay constant.

[0216] S52: Compare the obtained scoring index Ps with the preset scoring threshold Tps to determine the abnormal state of the scoring;

[0217] The abnormal status of the score is obtained by matching in the following ways:

[0218] When the scoring index Ps is less than the scoring threshold Tps*0.5, it indicates a normal score and normal pain detection;

[0219] When the scoring threshold Tps*0.5≤scoring index Ps≤scoring threshold Tps, it indicates a moderate score and the monitoring frequency is adjusted;

[0220] When the scoring index Ps> the scoring threshold Tps, it indicates an abnormal score and the scoring index Ps needs to be corrected.

[0221] S6. When the scoring index Ps is abnormal, a retrospective analysis is conducted on the abnormal individual, and combined with the mandibular muscle tension fluctuation Fda, the psychological misjudgment factor psy is constructed;

[0222] The psychological misjudgment factor psy is obtained by the following formula:

[0223]

[0224] Where λ represents the weight coefficient, pFda represents the average value of mandibular muscle tension fluctuation, PL represents a positive number to avoid division by zero, and d represents the integral. represents the rate of change of pain score over time;

[0225] The final pain response score nPs was obtained by correcting the scoring index Ps using the psychological misjudgment factor psy;

[0226] The final pain response score nPs is obtained by the following formula:

[0227] nPs=Ps*[1-ω*tanh(psy)];

[0228] Where tanh() represents the hyperbolic tangent function, and ω represents the maximum callback coefficient.

[0229] In this example, an exponential scaling function is constructed to convert the fused pain response index into a standardized score. This scoring function exhibits nonlinear growth characteristics, making it highly sensitive to changes in values ​​within the moderate pain range and enabling better differentiation between mild and moderate pain responses. This scoring mechanism combines comparability and discrimination, enhancing the model's ability to quantify pain responses across patients. It is particularly applicable to complex burn scenarios where subtle reactions can be masked and scoring can be ambiguous.

[0230] This example establishes a three-stage pain classification mechanism: "normal-moderate-abnormal" by setting multi-level scoring thresholds and comparing them with the score values ​​in real time, enabling the model to automatically identify abnormal score states. This design provides clear intervention instructions for the model output, making it particularly suitable for hierarchical processing and resource prioritization in real-time clinical decision support or remote nursing scenarios, improving the system's practicality and scenario flexibility.

[0231] In response to the defect that traditional models cannot remove the "emotion-dominated score abnormalities", this embodiment constructs a behavior-psychology coupled score correction mechanism by introducing a psychological misjudgment factor with mandibular muscle tension fluctuations as the core. This mechanism can determine whether the score is amplified by non-pain factors such as anxiety and tension, and automatically adjust it to effectively solve the misleading problem of "non-pathological high scores" and enhance the system's immunity to subjective emotional noise. Once a traditional evaluation system generates a high score, it can only output it passively, and cannot confirm its rationality or perform correction processing. This embodiment automatically suppresses and adjusts abnormal score results by setting a psychological feedback adjustment function, so that the system has the ability of "self-correction and self-convergence". This closed-loop mechanism enables the model to not only identify pain, but also judge the credibility of the score and actively optimize the results. It is a key transition from a "static scoring system" to a "dynamic intelligent control system."

[0232] By comparing individual psychological indicators with group baseline values, the model adapts to individual differences, making it particularly suitable for deployment with children, the elderly, or those with poor expressive abilities. Furthermore, this mechanism allows for dynamic parameter adjustments based on application scenarios, such as postoperative monitoring and chronic pain management, ensuring the model consistently outputs assessment results with medical value and practical significance.

[0233] Example 6

[0234] A burn assessment system based on image recognition, please refer to Figure 1 ,Specifically: it includes the following modules: image preprocessing module, burn area image recognition module, facial expression recognition module, ,joint pain response index construction module, pain score normalization module, and ,correction and optimization score module;

[0235] The image preprocessing module acquires the original image Ty through a medical image acquisition device, performs preprocessing, obtains the medical image It, and spatially separates the medical image It through an image segmentation algorithm to obtain the facial image Ifa and the burn image Ibu;

[0236] The burn area image recognition module extracts features from the burn image Ibu using a deep neural network and fits it into a damage feature set Wb;

[0237] The facial expression recognition module extracts features from the facial image Ifa using the expression recognition model and fits it into a psychological feature set Wa;

[0238] The joint pain response index construction module fuses the acquired injury feature set Wb and psychological feature set Wa to obtain the pain response index Ψ;

[0239] The pain score normalization module evaluates and analyzes the pain response index Ψ, uses the exponential scaling function to obtain the score index Ps, and determines the abnormal state of the score;

[0240] The correction and optimization scoring module analyzes abnormal conditions that occur during the evaluation of the pain response index Ψ to obtain the final pain response score nPs.

[0241] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A burn assessment method based on image recognition, characterized by: The following steps are involved: S1. Acquire the original image Ty using a medical image acquisition device, perform preprocessing on it, and obtain the medical image It. Use an image segmentation algorithm to spatially separate the medical image It and obtain the facial image Ifa and the burn image Ibu. S2. Extract features from the burn image Ibu using a deep neural network and fit it into a damage feature set Wb; S3, extract features from the facial image Ifa by using the expression recognition model and fit it into the psychological feature set Wa; S4, fusing the acquired injury feature set Wb and psychological feature set Wa to obtain the pain response index Ψ; S5. Evaluate and analyze the obtained pain response index Ψ, use an exponential scaling function to obtain a scoring index Ps, and determine an abnormal state of the score; S6. Analyze abnormal conditions that occur during the evaluation of the pain response index Ψ to obtain a final pain response score nPs.

2. The burn assessment method based on image recognition according to claim 1, characterized in that: S1 includes S11 and S12; S11, using a medical image acquisition device to collect an original image Ty, including burn area and facial information, and preprocessing the original image Ty, including noise filtering, color normalization, and image enhancement, to obtain a medical image It; The noise filtering is performed by filtering the noise in the original image Ty using a median filtering method to remove the noise in the original image Ty; Color normalization: All images are unified into the same color space by using the color normalization method; Image enhancement uses histogram equalization to redistribute the grayscale histogram of the image and enhance the contrast between dark and bright areas.

3. The burn assessment method based on image recognition according to claim 2, characterized in that: S12, performing structural recognition and partitioning on the medical image It by using an image segmentation algorithm to obtain a facial image Ifa and a burn image Ibu; Among them, the facial image Ifa is used for expression and muscle tension analysis, and classification mask extraction is achieved through semantic segmentation; The burn image Ibu representation is used for burn depth and pathological feature analysis, and classification mask extraction is achieved through semantic segmentation; The facial image Ifa is obtained by the following formula: Where Mfa represents the facial region mask function, Represents pixel-level dot multiplication operation; The burn image Ibu is obtained by the following formula: Where Mbu represents the burn area mask function.

4. The burn assessment method based on image recognition according to claim 3, characterized in that: S2 includes S21 and S22; S21. Using a composite deep neural network model, we extract 3D convolutional features from the burn image Ibu, including the tissue edema heterogeneity index Bsu, the necrotic pigment concentration parameter Bcu, and the microtension disturbance factor Btu of the injured area. Among them, the tissue edema heterogeneity index Bsu is obtained by calculating the reflectivity change difference of the edema area through the structure tensor; The tissue edema heterogeneity index Bsu is obtained by the following formula: Where Z represents the normalization factor, Ibu(i, j) represents the intensity value of the pixel in the i-th row and j-th column of the burn image Ibu, μloc(i, j) represents the local mean of the neighborhood of the pixel in the i-th row and j-th column of the burn image Ibu, and Wed(i, j) represents the edge perception weight of the pixel in the i-th row and j-th column of the burn image Ibu; The necrotic pigment concentration parameter Bcu is used to construct a color anomaly model based on the image color channel and extract the non-physiological distribution cluster centers in the RGB space; The necrotic pigment concentration parameter Bcu is obtained by the following formula: Where N represents the total number of pixels in the calculation area, C(i, j) represents the color vector of the pixel in the i-th row and j-th column of the burn image Ibu in the RGB color space, and Cref represents the reference skin color vector, which represents the standard color of the healthy area. The microtension disturbance factor Btu of the injured area is acquired by constructing the microtension field using Gabor ripple convolution kernel and texture direction map; The microtension disturbance factor Btu of the injured area is obtained by the following formula: Where Tga(i, j) represents the Gabor filter response value at the pixel in the i-th row and j-th column of the burn image Ibu, Ω represents the integration area, and dA represents the integration unit area. Represents the gradient of the Gabor filter result in a specific direction θ; S22. By constructing a nonlinear cross-response function Fcomp, the tissue edema heterogeneity index Bsu, the necrotic pigment concentration parameter Bcu, and the injury area microtension perturbation factor Btu are mapped into a structurally stable high-dimensional damage feature set Wb; The damage feature set Wb is obtained by the following formula: Wb=Fcomp(Bsu, Bcu, Btu).

5. The burn assessment method based on image recognition according to claim 3, characterized in that: S3 includes S31 and S32; S31, extracting facial action features from the facial image Ifa by using the expression recognition model, including the contraction intensity of the orbicularis oculi muscle Foa, the activity of the zygomatic major muscle Fza, and the tension fluctuation of the mandibular muscle Fda; Among them, the contraction intensity of the orbicularis oculi muscle Foa is obtained by collecting the orbicularis oculi muscle AU6; The contraction intensity of the orbicularis oculi muscle Foa is obtained by the following formula: Where Aorb(Ifa) represents the orbicularis oculi action recognition function, M represents the number of continuous sampling frames, AU6t represents the activation intensity of the orbicularis oculi action unit in the t-th frame, and Weye(t) represents the weight function of the eye muscles. The activity of the zygomatic major muscle Fza was obtained by collecting AU12 of the zygomatic major muscle; The zygomatic major muscle activity Fza is obtained by the following formula: Fza=Azgy(Ifa)=max(t)(AU12t)*ηsym; Where Azgy(Ifa) represents the zygomatic major muscle action extraction function, max(t) represents the maximum value of the zygomatic major muscle in all frames, and ηsym represents the facial muscle symmetry factor, which measures the consistency of the left and right facial muscle movements. Mandibular muscle tension fluctuation Fda is acquired by pulling down the corner of the mouth AU15 and opening the mandible AU26; The mandibular muscle tension fluctuation Fda is obtained by the following formula: Where Adep(Ifa) represents the mandibular tension action evaluation function, T represents the total number of time frames, xAU15t represents the rate of change of the mouth corner pull-down activation in the t-th frame; xAU26t represents the rate of change of the mandibular pull-down activation in the t-th frame; S32, inputting the acquired orbicularis oculi muscle contraction intensity Foa, zygomatic major muscle activity Fza, and mandibular muscle tension fluctuation Fda into the nonlinear modeling function Memo to construct a psychological feature set Wa; The psychological feature set Wa is obtained by the following formula: Wa=Memo(Foa,Fza,Fda).

6. The burn assessment method based on image recognition according to claim 5, characterized in that: S4 includes S41 and S42; S41, normalize the obtained damage feature set Wb and psychological feature set Wa, unify the dimensions and numerical ranges, and obtain a new damage feature set nWb and a new psychological feature set nWa; The new damage feature set nWb is obtained by the following formula: Where nWb(c) represents the c-th data item in the new damage feature set nWb, Wb(c) represents the c-th data item in the damage feature set Wb, minWb(c) represents the valley value of the c-th data item in the damage feature set Wb, and maxWb(c) represents the peak value of the c-th data item in the damage feature set Wb. The new psychological feature set nWa is obtained by the following formula: Where nWa(e) represents the e-th data item in the new psychological feature set nWa, Wa(e) represents the e-th data item in the psychological feature set nWa, minWa(e) represents the valley value of the e-th data item in the psychological feature set nWa, and maxWa(e) represents the peak value of the e-th data item in the psychological feature set nWa.

7. The burn assessment method based on image recognition according to claim 6, characterized in that: S42, performing nonlinear fusion on the acquired new injury feature set nWb and the acquired new psychological feature set nWa, and calculating and acquiring the pain response index Ψ; The pain response index Ψ is obtained by the following formula: where α1, α2, and α3 represent the preset weight values ​​of the tissue edema heterogeneity index Bsu, the necrotic pigment concentration parameter Bcu, and the microtension disturbance factor Btu in the injured area, respectively; β1, β2, and β3 represent the preset weight values ​​of the orbicularis oculi muscle contraction intensity Foa, the zygomatic major muscle activity Fza, and the mandibular muscle tension fluctuation Fda, respectively.

8. The burn assessment method based on image recognition according to claim 7, characterized in that: S5 includes S51 and S52; S51. Evaluate the pain response index Ψ obtained to obtain a scoring index Ps; The scoring index Ps is obtained by the following formula: Ps=100*(1-e -k(Ψ-cx) ); Where, e represents a constant, k represents the exponential rise rate control factor, and cx represents the response delay constant; S52: Compare the obtained scoring index Ps with the preset scoring threshold Tps to determine the abnormal state of the scoring; The abnormal status of the score is obtained by matching in the following ways: When the scoring index Ps is less than the scoring threshold Tps*0.5, it indicates a normal score and normal pain detection; When the scoring threshold Tps*0.5≤scoring index Ps≤scoring threshold Tps, it indicates a moderate score and the monitoring frequency is adjusted; When the scoring index Ps> the scoring threshold Tps, it indicates an abnormal score and the scoring index Ps needs to be corrected.

9. The burn assessment method based on image recognition according to claim 8, characterized in that: S6. When the scoring index Ps is abnormal, a retrospective analysis is conducted on the abnormal individual, and combined with the mandibular muscle tension fluctuation Fda, the psychological misjudgment factor psy is constructed; The psychological misjudgment factor psy is obtained by the following formula: Where λ represents the weight coefficient, pFda represents the average value of mandibular muscle tension fluctuation, PL represents a positive number to avoid division by zero, and d represents the integral. represents the rate of change of pain score over time; The final pain response score nPs was obtained by correcting the scoring index Ps using the psychological misjudgment factor psy; The final pain response score nPs is obtained by the following formula: nPs=Ps*[1-ω*tanh(psy)]; Where tanh() represents the hyperbolic tangent function, and ω represents the maximum callback coefficient.

10. A burn assessment system based on image recognition, applied to the burn assessment method based on image recognition according to any one of claims 1 to 9, characterized in that: It includes the following modules: image preprocessing module, burn area image recognition module, facial expression recognition module, joint pain response index construction module, pain score normalization module and correction and optimization score module; The image preprocessing module acquires the original image Ty through a medical image acquisition device, performs preprocessing, obtains the medical image It, and spatially separates the medical image It through an image segmentation algorithm to obtain the facial image Ifa and the burn image Ibu; The burn area image recognition module extracts features from the burn image Ibu using a deep neural network and fits it into a damage feature set Wb; The facial expression recognition module extracts features from the facial image Ifa using the expression recognition model and fits it into a psychological feature set Wa; The joint pain response index construction module fuses the acquired injury feature set Wb and psychological feature set Wa to obtain the pain response index Ψ; The pain score normalization module evaluates and analyzes the pain response index Ψ, uses the exponential scaling function to obtain the score index Ps, and determines the abnormal state of the score; The correction and optimization scoring module analyzes the abnormal conditions that occur during the evaluation of the pain response index Ψ to obtain the final pain response score nPs.

Images