Fat transplantation survival rate prediction system based on transfer learning and biomechanical simulation

Abstract

The invention provides a fat transplantation survival rate prediction system based on transfer learning and biomechanical simulation. Comprising an image data acquisition module, a fat distribution density prediction module, a cell integrity information acquisition module, a cell survival rate prediction module, a temperature image acquisition module, a blood supply state prediction module, a model fusion module, a transfer learning module and a survival rate output module. According to the system, by integrating multi-source data of preoperative MRI images, intraoperative pressure sensing and postoperative infrared thermal imaging, three prediction models of fat distribution density, cell survival rate and blood supply state are established and fused, animal experiment data knowledge is migrated to a human body for application in combination with a transfer learning technology, accurate prediction of the survival rate after fat transplantation is achieved, and the survival rate of fat transplantation is accurately predicted. The prediction accuracy rate reaches 85% or above, and scientific decision support is provided for clinic.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image processing and artificial intelligence, and particularly relates to a fat graft survival rate prediction system based on transfer learning. The system integrates preoperative image data, intraoperative pressure sensor data and postoperative thermography data, combines biomechanical simulation and transfer learning technology, realizes accurate prediction of fat graft survival rate after operation, and provides a scientific preoperative planning and postoperative evaluation tool for clinicians. BACKGROUND

[0002] Fat grafting is a common tissue filling and reconstruction technique in plastic surgery, which realizes tissue defect repair and appearance improvement through autologous fat transfer. However, the survival rate of fat grafting fluctuates greatly, usually between 40% and 70%, which seriously affects the operation effect and patient satisfaction. This uncertainty of survival rate is mainly due to the complex interaction of multiple factors: the quality of donor fat is uneven, the damage degree of fat cells during suction process is different, the blood supply condition of recipient area is different, and the postoperative nursing condition is different. How to accurately predict the survival rate of fat grafting has become an important research topic in the field of plastic surgery.

[0003] In the prior art, the prediction of fat graft survival rate mainly depends on the experience of doctors or single modality imaging evaluation, lacking multi-dimensional quantitative analysis. The traditional evaluation method has significant limitations: first, only through postoperative ultrasound or MRI examination to evaluate fat survival, which can only reflect the postoperative result and cannot make preoperative prediction, missing the opportunity to optimize the surgical plan; second, although the preoperative planning based on computer-aided design can simulate the transplantation effect, it lacks comprehensive consideration of cell integrity and blood supply factors, limiting the prediction accuracy and clinical application value; third, most existing prediction models ignore the biomechanical mechanism of fat grafting, and fail to analyze the deep factors such as tissue microenvironment, cell stress state and blood flow dynamics, resulting in large deviation between prediction results and actual situation.

[0004] In addition, the clinical data of human fat grafting is relatively scarce, and the insufficient sample size limits the training effect of deep learning-based prediction models. Although animal experiments can obtain a large amount of standardized data, there are significant differences between animal models and human body in terms of fat tissue characteristics, blood vessel distribution and metabolic characteristics, and directly applying animal experiment parameters will lead to prediction bias. How to effectively transfer animal experiment results to human application and realize cross-species knowledge transfer has become a key technical problem to be solved in this field.

[0005] The main limitations of existing fat grafting prediction techniques are as follows: (1) Lack of integrated analysis of preoperative, intraoperative and postoperative data. Single time point evaluation cannot capture the dynamic evolution process of fat grafting. (2) Ignoring the biomechanical mechanism of fat grafting. The influence of physical factors such as pressure, stress and blood flow on survival rate is not considered. (3) Unable to effectively solve the problem of clinical data scarcity. The model generalization ability is insufficient. These limitations result in insufficient accuracy and practicability of existing prediction methods, which cannot meet the requirements of precise and personalized medicine in clinical practice. Therefore, new technical solutions are urgently needed. SUMMARY

[0006] The purpose of the present application is to provide a fat grafting survival rate prediction system based on transfer learning and biomechanical simulation. By integrating preoperative, intraoperative and postoperative multi-source data, combining biomechanical simulation and transfer learning technology, the survival rate of fat grafting is accurately predicted, and the limitations of existing technologies are overcome.

[0007] The present application provides a fat grafting survival rate prediction system based on transfer learning and biomechanical simulation, which comprises an image data acquisition module, a fat distribution density prediction module, a cell integrity information acquisition module, a cell survival rate prediction module, a temperature image acquisition module, a blood supply state prediction module, a model fusion module, a transfer learning module and a survival rate output module. The image data acquisition module acquires the image data of the preoperative fat tissue; the fat distribution density prediction module establishes a fat distribution density prediction model based on the image data; the cell integrity information acquisition module acquires the integrity information of the fat cells during the operation; the cell survival rate prediction module establishes a cell survival rate prediction model based on the integrity information; the temperature image acquisition module acquires the temperature image of the fat graft after the operation; the blood supply state prediction module establishes a blood supply state prediction model based on the temperature image; the model fusion module combines the three prediction models to generate a postoperative survival rate prediction model with the patient's baseline information; the transfer learning module uses animal experiment data to train the model and migrates to human data to obtain the optimal parameters; and the survival rate output module outputs the prediction results.

[0008] In a preferred embodiment, the image data acquisition module comprises an MRI image acquisition unit for acquiring preoperative MRI images including T1W, T2W and DWI sequences; the fat distribution density prediction module further comprises a machine learning training unit and an SVM optimization unit. The machine learning training unit generates a fat distribution density prediction model based on a machine learning algorithm. The SVM optimization unit optimizes the model parameters and accuracy. The model output result is represented as a regression equation with respect to the input features, including bias term, training times, iteration times, optimization function and L2 norm.

[0009] In another preferred embodiment, the cell integrity information acquisition module comprises a suction state pressure sensing unit, a morphological feature extraction unit, and a pressurized state pressure sensing unit, the suction state pressure sensing unit acquires morphological and rheological characteristics of fat in a suction state, the morphological feature extraction unit extracts morphological feature indicators, and the pressurized state pressure sensing unit acquires cell survival rate and pressure change data in a pressurized state. The cell survival rate prediction module constructs a cell survival rate prediction model and a cell survival rate and pressure variable relationship model based on these data.

[0010] In yet another preferred embodiment, the temperature image acquisition module comprises an infrared thermal imaging unit for acquiring temperature distribution images of the fat graft in a blood supply state. The blood supply state prediction module constructs a blood supply state prediction model based on the temperature distribution images, which is expressed as a regression equation with respect to input features, including a bias term, a training number, an iteration number, an optimization function, and an L2 norm.

[0011] In still another preferred embodiment, the model fusion module comprises a morphological information generation unit, a cell survival rate calculation unit, and a survival rate calculation unit. The morphological information generation unit associates the fat distribution density prediction model with the cell survival rate prediction model to generate graft morphological information. The cell survival rate calculation unit associates the morphological information with the blood supply state prediction model to generate graft cell survival rate. The survival rate calculation unit associates the cell survival rate with the morphological information to generate postoperative survival rate.

[0012] In a further preferred embodiment, the transfer learning module further comprises a multi-source data analysis unit and a parameter extraction unit. The multi-source data analysis unit synchronously acquires and analyzes multi-source data before, during, and after human fat transplantation to generate analysis results. The parameter extraction unit extracts parameters required by the postoperative survival rate prediction model based on the analysis results.

[0013] In another preferred embodiment, the system further comprises a fat distribution area determination module, a blood flow dynamics state determination module, a mechanical parameter determination module, and an initial scheme generation module. The fat distribution area determination module determines the boundary of the fat distribution area. The blood flow dynamics state determination module determines blood flow velocity, pressure, and flow direction within the area. The mechanical parameter determination module determines mechanical parameters corresponding to different transplantation levels. The initial scheme generation module generates an initial fat transplantation scheme.

[0014] In yet another preferred embodiment, the system further comprises a scheme retrieval module, an iterative control module, and a scheme output module. The scheme retrieval module retrieves the scheme with the highest survival rate from historical transplantation data using a transfer learning algorithm. The iterative control module controls repeated retrieval until a scheme with a survival rate greater than or equal to 90% is retrieved. The scheme output module outputs the final fat transplantation scheme.

[0015] The beneficial effects of the present application include: first, by integrating preoperative MRI image data, intraoperative pressure sensing data and postoperative infrared thermal imaging data, the multidimensional monitoring of the whole process of fat transplantation is realized, and the comprehensiveness and accuracy of the prediction are significantly improved; second, by establishing fat distribution density prediction model, cell survival rate prediction model and blood supply state prediction model and fusing them, the comprehensive prediction of fat transplantation survival rate is realized, and the problem of insufficient prediction accuracy of single model is overcome; third, by introducing biomechanical simulation, the mechanical properties of fat transplantation are introduced into the prediction model, and the scientificity and explainability of the prediction are improved from the mechanism level; fourth, by using transfer learning technology, the problem of data scarcity of human fat transplantation is effectively solved, the knowledge of animal experiment data is transferred to human application, and the applicability of the model in clinical application is improved; fifth, by using closed loop optimization mechanism, the continuous optimization of the transplantation scheme is realized until the expected survival rate target is reached, and a scientific decision support tool is provided for clinicians; sixth, the overall prediction accuracy can reach more than 85%, which is significantly higher than the 60-70% of traditional methods, can effectively reduce the secondary surgery rate, reduce the medical cost and improve the patient satisfaction. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 It is the overall architecture schematic diagram of the fat transplantation survival rate prediction system based on transfer learning and biomechanical simulation of the present application, wherein the number 1 represents the image data acquisition module, the number 2 represents the fat distribution density prediction module, the number 3 represents the cell integrity information acquisition module, the number 4 represents the cell survival rate prediction module, the number 5 represents the temperature image acquisition module, the number 6 represents the blood supply state prediction module, the number 7 represents the model fusion module, the number 8 represents the transfer learning module, and the number 9 represents the survival rate output module.

[0017] Figure 2 It is the internal structure schematic diagram of the fat distribution density prediction module of the present application, which shows the connection relationship between the MRI image acquisition unit, the machine learning training unit and the SVM optimization unit.

[0018] Figure 3 It is the internal structure schematic diagram of the cell survival rate prediction module of the present application, which shows the connection relationship between the aspiration state pressure sensing unit, the morphological feature extraction unit, the pressurized state pressure sensing unit and the cell survival rate prediction module.

[0019] Figure 4 It is the internal structure schematic diagram of the blood supply state prediction module of the present application, which shows the connection relationship between the infrared thermal imaging unit and the blood supply state prediction module and the processing flow of the temperature distribution image.

[0020] Figure 5It is an internal structure diagram of the model fusion module of the present application, which shows the data flow relationship between the morphological information generation unit, the cell survival rate calculation unit and the survival rate calculation unit.

[0021] Figure 6 It is an internal structure diagram of the transfer learning module of the present application, which shows the multi-source data analysis unit, the parameter extraction unit and the transfer process of animal data to human data.

[0022] Figure 7 It is a data flow relationship diagram of the system of the present application, which shows the complete data flow process from data collection, model establishment, model fusion to final output. DETAILED DESCRIPTION

[0023] The specific embodiments of the present application will be further described below. Figures 1-7 The specific embodiments of the present application will be further described below.

[0024] Referring to Figure 1 , the fat graft survival rate prediction system based on transfer learning and biomechanical simulation provided by the present application includes an image data acquisition module 1, a fat distribution density prediction module 2, a cell integrity information acquisition module 3, a cell survival rate prediction module 4, a temperature image acquisition module 5, a blood supply state prediction module 6, a model fusion module 7, a transfer learning module 8 and a survival rate output module 9.

[0025] The image data acquisition module 1 is used to obtain the image data of the preoperative fat tissue, providing basic data support for the subsequent fat distribution density prediction. In clinical practice, this module is usually integrated with standard medical imaging equipment, which can fully capture the spatial distribution characteristics, density characteristics and microstructure information of fat tissue.

[0026] As shown in Figure 2 , the fat distribution density prediction module 2 is connected with the image data acquisition module 1, which is used to establish a fat distribution density prediction model of the fat graft individual based on the image data. This module receives the original image data from the image data acquisition module 1, extracts the feature parameters of the fat tissue through a deep learning algorithm, and establishes a mapping relationship from the image features to the fat density, realizing accurate prediction of the postoperative fat distribution state.

[0027] In a preferred embodiment of the present application, the image data acquisition module 1 comprises an MRI image acquisition unit 11, which is used to acquire preoperative MRI images. The acquired image sequences include T1W, T2W and DWI. The MRI image acquisition parameters are configured as follows: the T1W sequence adopts a spin echo sequence, the repetition time is 500 ms, the echo time is 15 ms, the slice thickness is 3 mm, the slice interval is 0.3 mm, and the field of view is 400 mm x 400 mm; the T2W sequence adopts a fast spin echo sequence, the repetition time is 4000 ms, the echo time is 100 ms, and the other parameters are consistent with the T1W sequence; the DWI sequence adopts a single-shot echo planar sequence, the b value is set to three gradients of 0, 500 and 1000 s / mm², which is used to evaluate the diffusion characteristics of water molecules. These three sequences can reflect the quality state of adipose tissue from different angles: the T1W sequence clearly shows the high signal intensity of adipose tissue, accurately evaluates the distribution and quantity of adipose tissue in the donor area; the T2W sequence reflects the water content in the tissue, evaluates the edema degree and activity of adipose tissue; the DWI sequence indirectly reflects the tissue microenvironment and cell density by measuring the diffusion ability of water molecules.

[0028] The fat distribution density prediction module 2 further comprises a machine learning training unit 21 and an SVM optimization unit 22. The machine learning training unit 21 is used to train the MRI images based on a machine learning algorithm to generate a fat distribution density prediction model. This unit first pre-processes the MRI images, including denoising, registration and normalization, uses a Gaussian filter to remove image noise, the filter window size is set to 5 x 5 pixels, and the standard deviation is set to 1.5; then uses a U-Net deep convolutional neural network for fat tissue segmentation, the network architecture includes 4 layers of encoder and 4 layers of decoder, each layer contains two 3 x 3 convolution layers and one 2 x 2 max pooling layer, the activation function uses ReLU, and the loss function uses a combination of Dice coefficient and cross entropy; after segmentation, the feature parameters of the fat tissue are extracted, including more than 10 features such as signal intensity mean and standard deviation, fat volume, distribution uniformity, and boundary clarity; finally, a mapping relationship from the feature parameters to the fat density is established based on a support vector regression algorithm.

[0029] The SVM optimization unit 22 is used to optimize the parameters and parameter accuracy of the fat distribution density prediction model, so that the predicted fat distribution density image conforms to the actual situation. This unit uses a grid search combined with cross-validation method to optimize the kernel function parameters and regularization parameters of SVM: the kernel function selects a radial basis function, the search range of parameter γ is 0.001 to 1.0, the search range of penalty parameter C is 1 to 1000, 5-fold cross-validation is used to evaluate the performance of different parameter combinations, and the parameter combination with the smallest mean square error on the validation set is selected as the optimal parameter.

[0030] The output of the fat distribution density prediction model is represented as a regression equation with respect to the input features:

[0031] ,

[0032] wherein, is a regression equation with respect to the input features , represents the predicted fat distribution density value in g / cm³; is a weight vector with the same dimension as the feature mapping, which is obtained by optimization on the training data; is a feature mapping function of the input features, which maps the original features to a high-dimensional space. For a radial basis kernel function, an implicit mapping from the original feature space to an infinite-dimensional Hilbert space is achieved; is a bias term, representing the basic density value. The hyperparameters in the training process include the number of training times (typical value is 1000 to 5000, set to 3000 in this embodiment), the number of iterations (the number of iterations in each training, typical value is 100 to 500, set to 300 in this embodiment), the optimization function (using the sequential minimal optimization algorithm SMO), and the L2 norm regularization coefficient (typical value is 0.01 to 0.1, set to 0.05 in this embodiment, used to prevent overfitting).

[0033] In actual application, for the evaluation of abdominal fat graft donor area, the system input includes the T1W, T2W and DWI sequence images of the patient, after feature extraction, the fat signal intensity is 1850±150, the fat distribution uniformity score is 0.82, the water content index is 0.15, and other characteristic parameters are obtained. For example, the average fat density of the abdominal donor area of a 45-year-old female patient is 0.92 g / cm³, 3 and the predicted stable density after transplantation is 0.87 g / cm³, the density retention rate is 94.6%, which provides an important preoperative evaluation reference for the doctor.

[0034] In another embodiment, the SVM optimization unit 22 of the fat distribution density prediction module 2 adopts the following optimization objective: to achieve the optimal generalization ability of the model by minimizing the structural risk, which includes two parts of the empirical risk and the model complexity, the empirical risk is measured by the loss function on the training samples, and the model complexity is measured by the norm of the weight vector. During the optimization process, the penalty parameter controls the balance between model complexity and training error, a smaller penalty parameter value allows more training error but simplifies the model structure, and a larger penalty parameter value requires smaller training error but may cause overfitting. In the application of fat transplantation, the penalty parameter is preferably set to 1.0, at this time the model has good generalization ability while maintaining high prediction accuracy. The slack variable represents the allowed error, for the i-th A sample is considered to be a good sample if the deviation between its predicted value and the true value is less than a threshold value (in this embodiment, set to 0.05 g / cm³), and the relaxation variable of the sample is 0; if the deviation is greater than the threshold value, the relaxation variable is equal to the deviation value of the exceeding part.

[0035] The cell integrity information acquisition module 3 is used to acquire the integrity information of the fat cells in the operation, and the module is used for real-time monitoring in the key links of the fat suction and injection, and captures the key factors affecting the cell survival.

[0036] As shown in Figure 3 The cell survival rate prediction module 4 is connected with the cell integrity information acquisition module 3, and is used to establish a cell survival rate prediction model of the graft based on the integrity information. The module establishes a prediction relationship from the operation parameter to the cell survival rate by analyzing the suction pressure, cell morphology, rheological characteristics and other multi-dimensional data, and provides real-time guidance for the intraoperative operation.

[0037] In a preferred embodiment of the present application, the cell integrity information acquisition module 3 includes a suction state pressure sensing unit 31, a morphological feature extraction unit 32 and a pressurized state pressure sensing unit 33. The suction state pressure sensing unit 31 is used to collect the morphological and rheological characteristics of the fat in the suction state before the graft is placed in the target area. The unit uses a micro pressure sensor, the sensor accuracy is 0.01 kPa, the sampling frequency is 50 Hz, and the installation position is located at the proximal end of the suction pipeline, about 5 cm away from the suction port. In clinical practice, the fat suction uses a suction device with a negative pressure range of 300 to 700 mmHg, and the pressure sensor records the pressure change curve in the suction process in real time. The collected data includes instantaneous pressure value, pressure gradient, pressure peak value, pressure duration and other parameters. The stability of the suction pressure is an important indicator for evaluating the integrity of the fat cells: a smooth pressure curve indicates that the suction process is gentle and uniform, and the mechanical damage to the fat cells is small; a sharply fluctuating pressure curve indicates that the operation is rough, and the cell damage rate is high.

[0038] ​​The morphological feature extraction unit 32 is connected with the suction state pressure sensing unit 31, and is used for extracting morphological feature indexes of the preoperative fat mass based on morphological and rheological characteristics. The unit performs microscopic examination on the extracted fat sample, and observes the morphological characteristics of fat cells under 40 times magnification using a phase contrast microscope. The statistical parameters include: the proportion of complete cells (the percentage of cells with intact cell membranes and no damage in the total number), the mean and standard deviation of cell diameter (the normal fat cell diameter ranges from 80 to 120 μm), the cell shape regularity (evaluating whether the cells maintain a regular shape of circle or ellipse), and the amount of fat microbubbles (measuring the proportion of the microbubble layer to the total volume after centrifugal separation, and a high amount of microbubbles indicates that the cell damage is severe). The morphological feature index is a dimensionless index for comprehensive evaluation of fat quality. When calculating, different weight coefficients are assigned to each parameter. The cell integrity accounts for 60% of the weight, and the amount of fat microbubbles accounts for 40% of the weight. This weight distribution is based on the statistical analysis results of a large amount of clinical research data, and can better reflect the correlation between fat quality and survival rate.

[0039] The pressurized state pressure sensing unit 33 is used to collect the cell survival rate and pressure change data of the fat in the pressurized state after the implant is placed in the target area. The unit uses a pressure sensor and a pressure valve system for pressurized dressing monitoring. The sensor is installed on the inside of the dressing bandage, tightly attached to the surface of the implant area, and measures the dressing pressure in real time. The pressure valve system can adjust the dressing pressure, and the pressure adjustment range is 5 to 30 mmHg, with an adjustment accuracy of 1 mmHg. Postoperative compression dressing is an important nursing measure for fat transplantation. Moderate pressure can promote close contact between the implant and the recipient tissue, improve blood circulation, but excessive pressure can compress blood vessels, leading to insufficient blood supply and reducing cell survival rate. The unit continuously monitors the dressing pressure and records the pressure-time curve, while evaluating the cell survival status of the implant area through non-invasive detection methods (such as bioimpedance measurement). Bioimpedance is closely related to the integrity of the cell membrane: when the cell membrane is intact, the impedance is high, and when the cell is damaged or necrotic, the impedance decreases. By establishing the correspondence between pressure and bioimpedance, the influence of pressure on cell survival rate can be quantified.

[0040] The cell survival rate prediction module 4 constructs a cell survival rate prediction model based on morphological feature indicators. The construction process of the model includes: first, collect a large number of morphological feature data and corresponding actual survival rate data of clinical samples, the sample size is not less than 200 cases, covering multiple application scenarios such as facial filling, breast reconstruction, and hip shaping; then, use multivariate linear regression analysis to identify morphological features that significantly affect survival rate, and select 5 to 8 key features as model inputs; next, establish a regression equation from morphological features to survival rate, and the regression coefficient is estimated by the least squares method; finally, use an independent validation dataset to evaluate the model performance, requiring the root mean square error of the prediction on the validation set to be less than 5%. The output of the cell survival rate prediction model is the cell survival rate percentage, with a value range of 0 to 100%, which reflects the proportion of fat cells expected to survive under the current morphological conditions.

[0041] The cell survival rate prediction module 4 also establishes a cell survival rate-pressure variable relationship model based on the cell survival rate under pressure and pressure change data. The relationship model uses an exponential decay function to describe the negative impact of pressure on survival rate:

[0042] ,

[0043] wherein, is the relationship function of cell survival rate and pressure, representing the cell survival rate under the action of pressure (unit: mmHg); is the basic survival rate, representing the theoretical survival rate under zero pressure, usually with a value range of 70% to 90%, and in this embodiment the preferred value is 85%; is the pressure value of the pressure dressing; is the pressure influence coefficient, representing the decay rate of pressure on survival rate, with a unit of mmHg -1 . The value of the pressure influence coefficient is related to the tissue characteristics of the transplantation site: the facial tissue is relatively tight, with high capillary density but small vessel lumen, strong pressure sensitivity, and the preferred value is 0.03 mmHg -1 ; the breast tissue is relatively soft, with larger blood vessel lumen and good pressure tolerance, and the preferred value is 0.02 mmHg -1 ; the hip tissue is rich in fat pads, with good buffering capacity, and the preferred value is 0.015 mmHg -1 . The selection of the exponential function is based on physiological principles: slight pressure has little effect on cell survival, and as the pressure increases, the degree of vascular compression increases nonlinearly, and blood supply decreases rapidly, leading to an exponential decrease in survival rate.

[0044] In a clinical example, a patient undergoing cheek augmentation, the postoperative compression dressing pressure was set to 18 mmHg, according to the model , but the calculation result did not take other factors into account. In actual application, the model was used in combination with a fat distribution density model and a blood supply state model, and the final predicted survival rate was 84.5%, and the actual measured survival rate was about 83% in a 3-month follow-up, with a prediction error within an acceptable range.

[0045] The temperature image acquisition module 5 is used to acquire a temperature image of the postoperative fat graft, and the recovery of the blood supply is evaluated by monitoring the temperature distribution characteristics of the grafting area. Temperature is an important indicator reflecting the blood perfusion of tissues: the area with sufficient blood flow has high metabolism and high temperature; the area with insufficient blood supply has low metabolism and low temperature.

[0046] As shown in Figure 4 , the blood supply state prediction module 6 is connected with the temperature image acquisition module 5, and is used to establish a graft blood supply state prediction model based on the temperature image. The module quantitatively evaluates the blood supply state of the grafting area by analyzing the spatial distribution characteristics, time evolution law and temperature difference with the surrounding tissues of the temperature, and early discovers the risk area with insufficient blood supply.

[0047] In a preferred embodiment of the present application, the temperature image acquisition module 5 includes an infrared thermal imaging unit 51, which is used to acquire a temperature distribution image of the fat graft in the blood supply state. The technical parameters of the infrared thermal imaging device include: the temperature resolution is 0.1℃, which can accurately capture the tiny temperature difference; the spatial resolution is 0.5mm, which ensures the accurate positioning of small temperature abnormalities; the detector type is a non-cooled microbolometer, and the working waveband is long-wave infrared of 8 to 14μm; the temperature measurement range is 20℃ to 45℃, which covers the normal range of human body surface temperature. The acquisition process is: the patient is in a constant temperature environment (the room temperature is controlled at 22±1℃), the grafting area is fully exposed and placed for 10 minutes to reach a thermal equilibrium state, the infrared thermal imager is placed vertically to the skin surface with a distance of 30cm, the acquisition time is 2 seconds, and each acquisition obtains a temperature distribution map of 512×512 pixels. The postoperative follow-up scheme is: the thermal imaging examination is performed once at 1st, 3rd, 5th, 7th, 14th and 30th day, and the dynamic change trend of the temperature distribution of the grafting area is recorded.

[0048] The blood supply state prediction module 6 constructs a blood supply state prediction model based on the temperature distribution image. The construction of the model includes the following steps: first, feature parameters are extracted from the temperature distribution image, including the average temperature (the arithmetic mean of the temperature of all pixel points in the transplantation region), the temperature standard deviation (reflecting the uniformity of the temperature distribution), the temperature difference with the surrounding normal tissue (the difference between the average temperature of the transplantation region and the average temperature of the normal tissue within a range of 5 mm), the maximum temperature gradient (the ratio of the maximum temperature difference between adjacent pixel points in the transplantation region to the distance), the hot spot area ratio (the ratio of the area of the region with a temperature higher than the average temperature plus one standard deviation to the total area), and the cold spot area ratio (the ratio of the area of the region with a temperature lower than the average temperature minus one standard deviation to the total area); then, a mapping relationship from the temperature features to the blood supply state score is established based on the support vector machine algorithm, the blood supply state score is a continuous value from 0 to 100 points, and the higher the score, the better the blood supply state; finally, the model accuracy is evaluated using clinical verification data to ensure that the correlation coefficient between the predicted score and the actual blood supply state after the operation is greater than 0.85.

[0049] The blood supply state prediction model is represented as a regression equation with respect to the input features:

[0050] ,

[0051] wherein, is a regression equation with respect to the input features , representing the blood supply state score, with a value range of 0 to 100 points; is a weight vector, with the same dimension as the mapped features; is a feature mapping function of the input features, and for a radial basis kernel function, its expression is , wherein is a support vector, is a kernel function parameter; is a bias term. The training parameters include the number of training times (set to 2000 times in this embodiment), the number of iterations (set to 200 times in this embodiment), the optimization function (using the sequential minimal optimization algorithm), and the L2 norm regularization coefficient (set to 0.03 in this embodiment).

[0052] In actual application, the input features is a 6-dimensional vector, including average temperature, temperature standard deviation, temperature difference, maximum temperature gradient, hot spot ratio and cold spot ratio. For example, for a fat grafting case of breast reconstruction, thermal imaging on the 5th day after surgery shows that the average temperature is 35.8℃, the average temperature of the surrounding normal tissue is 36.2℃, the temperature difference is -0.4℃; the temperature standard deviation is 0.3℃, indicating that the temperature distribution is relatively uniform; the maximum temperature gradient is 0.15℃ / mm, and there is no sharp temperature jump; the hot spot ratio is 8%, and the cold spot ratio is 12%. After inputting these features into the model, the predicted blood supply status score is 78, and the evaluation conclusion is that the blood supply condition is good, indicating a high survival rate. Clinical experience shows that when the blood supply status score is above 80, the survival rate usually reaches more than 85%; when the score is between 60 and 80, the survival rate is between 70% and 85%; when the score is less than 60, the survival rate is usually less than 70%, and intervention measures need to be taken to improve blood supply.

[0053] The temperature gradient is a key parameter for evaluating the blood supply condition. In the transplanted area, if the blood perfusion is uniform, the temperature of each part should be close, and the temperature gradient is small; if the blood supply of some area is insufficient, the temperature of that area is low, and a large temperature gradient is formed with the surrounding. The modulus of the temperature gradient vector reflects the spatial rate of temperature change: the temperature gradient modulus of the area with good blood supply is usually less than 0.2℃ / mm, and the temperature distribution is smooth transition; the blood supply insufficient area may appear a large gradient of more than 0.5℃ / mm, and the temperature distribution appears obvious boundary. Therefore, the maximum temperature gradient can be used as a sensitive indicator to identify the blood supply insufficient area early, once an abnormally high temperature gradient is found, the blood flow condition of the area should be evaluated immediately, and necessary measures should be taken to improve blood supply, such as adjusting the pressure of the dressing, performing local massage or using drugs to promote blood circulation.

[0054] As shown in Figure 5 The model fusion module 7 is connected with the fat distribution density prediction module 2, the cell survival rate prediction module 4 and the blood supply status prediction module 6 respectively, for combining the three prediction models with the baseline information of the patient to generate a postoperative survival rate prediction model of the graft. This module is the core of the system, which realizes the comprehensive evaluation of the survival rate by integrating the prediction information of the preoperative, intraoperative and postoperative stages.

[0055] The model fusion module 7 comprises a morphological information generation unit 71, a cell viability calculation unit 72 and a survival rate calculation unit 73. The morphological information generation unit 71 is configured to associate the fat distribution density prediction model with the cell survival rate prediction model to generate the graft morphological information. The working principle of this unit is that the fat distribution density reflects the spatial distribution characteristics and the tissue density of the graft, and the cell survival rate reflects the cell activity state of the graft, and both of them jointly determine the morphological characteristics of the graft. The morphological information is a multi-dimensional feature vector, including elements such as tissue volume, density distribution, shape parameter, cell activity, etc. The dynamic evolution process of the morphological information can be described as: the morphological information of the current frame is determined by the morphological information of the previous frame , the regression equation about pressure, cell morphology and fat distribution density and its parameters . This recursive relationship reflects the evolution law of the morphological information over time: the morphological state at the previous time has a continuation effect on the current time, and at the same time, the pressure, cell state and fat density at the current time have an immediate impact on the morphology.

[0056] The cell viability calculation unit 72 is connected with the morphological information generation unit 71, and is configured to associate the graft morphological information with the blood supply state prediction model to generate the graft cell viability. The logic of this unit is that the morphological information determines the demand of the graft for blood supply, the blood supply state determines the level of nutrients and oxygen that can be provided, and the matching degree of the two determines the cell viability. The dynamic calculation process of the cell viability is: the cell viability of the current frame is determined by the blood supply state of the current frame , the morphological information of the current frame , the blood supply state of the previous frame , and the regression equation about morphology and blood supply state and its parameters . This calculation process takes into account the current value and the historical value of the blood supply state: sudden changes in blood supply (such as vasospasm or thrombosis) will quickly affect the cell viability, and persistent lack of blood supply will cumulatively reduce the cell activity.

[0057] The survival rate calculation unit 73 is connected with the cell viability calculation unit 72, and is configured to associate the graft cell viability with the morphological information to generate the postoperative survival rate. This unit performs the final survival rate prediction: the postoperative survival rate of the current frame is determined by the cell viability of the current frame , the morphological information of the current frame , the cell viability of the previous frame , and the regression equation about morphology and cell viability and its parameters The prediction model takes into account both the instant activity and the morphological stability of the cells: when the cell viability is high and the morphology is stable, the survival rate is expected to be high; when the cell viability decreases or the morphology changes adversely, the survival rate is expected to decrease. The cell viability information of the previous frame provides a reference for the survival trend: if the cell viability continues to decrease, it suggests that the survival rate may be lower than the current evaluation value; if the cell viability gradually increases, it suggests that the survival rate is expected to be higher than the current evaluation value.

[0058] In a clinical example, a patient undergoing abdominal wall reshaping fat transplantation, the monitoring data on the 3rd day after surgery are as follows: the blood supply state score is 82, which is improved compared with the score of 79 of the previous day, indicating that the vascular reconstruction is progressing smoothly; the morphological score is 90, indicating that the shape of the graft is stable and the volume is well preserved; the cell survival rate is calculated to be 85% according to the foregoing data. Combined with the patient's baseline information (age 38, BMI 23, no history of smoking, no metabolic diseases such as diabetes), the system predicts the final survival rate to be 86%. Follow-up for 3 months after surgery, the volume retention rate of the graft measured by MRI is 84%, which is very close to the predicted value, with an error of only 2 percentage points, verifying the accuracy of the model.

[0059] As shown in Figure 6 The transfer learning module 8 is connected with the model fusion module 7, and is used for training the postoperative survival rate prediction model of the graft using animal experiment data, transferring the training results of the animal experiment data to human transplantation data, and obtaining the optimal parameter solution set of the postoperative survival rate prediction model of the graft. This module solves the key problem of the scarcity of clinical data of human body, and improves the training effect and generalization ability of the model through cross-species knowledge transfer.

[0060] The transfer learning module 8 further comprises a multi-source data analysis unit 81 and a parameter extraction unit 82. The multi-source data analysis unit 81 is configured to use the fat distribution density prediction model, the cell survival rate prediction model and the blood supply state prediction model to synchronously collect and analyze multi-source data before, during and after human fat transplantation, and generate analysis results. The function of this unit is to identify the systematic differences between animal models and human models, and to provide a basis for subsequent parameter adjustment. The specific workflow is as follows: collect complete data of 20 to 30 cases of human fat transplantation, including preoperative MRI images, intraoperative pressure monitoring, postoperative thermal imaging and long-term follow-up survival rate results; apply the three prediction models to each patient, record the predicted values and actual observed values of the models; statistically analyze the distribution characteristics of the prediction bias, identify whether there is a systematic bias (such as the model systematically overestimating or underestimating the survival rate) and the influencing factors (such as patient age, transplantation site, transplantation volume, etc.) of the bias. Through this analysis, the degree of difference between human data and animal data can be quantified, for example, it is found that when the parameters trained by the animal model are applied to the human body, the survival rate prediction value is on average 8 percentage points higher, and this systematic bias needs to be corrected by parameter adjustment.

[0061] The parameter extraction unit 82 is connected with the multi-source data analysis unit 81, and is configured to extract the parameters required by the post-transplantation survival rate prediction model of the graft based on the analysis results. This unit executes the core algorithm of transfer learning, including the following steps: first, initialize the model parameters, use the parameters trained by animal experiment data as the initial value; second, evaluate the model performance on human data, calculate the prediction error; third, based on the prediction error of human data, use the gradient descent algorithm to update the model parameters, the direction of updating is to reduce the prediction error, and the step of updating is controlled by the learning rate; fourth, repeat the second and third steps until the performance of the model on human data reaches the optimal or reaches the preset upper limit of iteration times. The key of transfer learning is to retain the general knowledge learned by the animal model (such as the basic physiological characteristics of adipose tissue, the general relationship between blood supply and survival rate, etc.), while adjusting the human-specific parameters (such as the coefficients of vascular density, metabolic rate, immune response, etc.).

[0062] In practical applications, animal experiments use mouse models, and 300 cases of fat transplantation data are collected, covering different transplantation volumes, different transplantation sites, and different blood supply conditions. The model trained using these data has an accuracy of 92% on the animal validation set. However, when this model is directly applied to human data, the accuracy drops to 76%, mainly due to significant differences in fat tissue characteristics between mice and humans: the vascular density of mouse subcutaneous fat is 8.5 / mm², while that of humans is only 4.2 / mm², the difference in vascular density leads to a deviation in blood supply capacity assessment; the average diameter of mouse fat cells is 62 μm, while that of humans is 105 μm, the difference in cell size affects the baseline value of cell survival rate; the basal metabolic rate of mice is about 1.4 times that of humans, the difference in metabolic rate leads to inaccurate assessment of nutritional and oxygen demand. Through transfer learning, the system identifies these key difference parameters: the vascular density ratio coefficient (human / mouse) is 0.49, the cell size ratio coefficient is 1.69, and the metabolic rate ratio coefficient is 0.72. Incorporating these ratio coefficients into the model's feature mapping function, the animal model parameters are calibrated, and after transfer learning optimization, the model's accuracy on human data improves to 89%, significantly improving the prediction performance.

[0063] The network structure of transfer learning uses a five-layer convolutional neural network, the first three layers of weights are transferred from the animal model, and these three layers mainly extract basic features of fat tissue (such as texture, density, spatial distribution, etc.), these features have commonality between animals and humans, so they can be directly transferred; the last two layers are fine-tuned for human data, these two layers mainly learn human-specific high-level features (such as blood vessel distribution pattern, metabolic characteristics, etc.), which need to be retrained based on human data to adapt to the particularity of humans. For a typical facial filling case, using the transfer learning optimized model, the predicted survival rate is 87%, and the actual follow-up measurement after 3 months is 85%, the prediction error is 2 percentage points, which meets the requirements of clinical application.

[0064] The survival rate output module 9 is connected with the transfer learning module 8, and is used for outputting a postoperative survival rate prediction result based on the optimal parameter solution set. The module presents the prediction result to the clinician in an intuitive and easy-to-understand manner, including the numerical value of the survival rate, the confidence interval, the survival rate distribution heat map of the sub-regions, the analysis of the main risk factors and the like. The output interface adopts a graphical design: the main interface displays the predicted overall survival rate in percentage form, with a large font size; the sub-interface displays the survival rate heat map of the sub-regions, and the heat map adopts color coding, with red representing a high survival rate region and blue representing a low survival rate region, so that the clinician can intuitively identify the potential problem region; the risk factor analysis interface lists the main factors affecting the survival rate and their contribution degrees, for example, a case shows that the blood supply state contributes 45%, the cell integrity contributes 35%, and the fat density contributes 20%, indicating that the blood supply is the main influencing factor of the case, and the clinician can take targeted measures to improve the blood supply. The output module also provides a historical data comparison function, and the clinician can view the prediction result change trend of the patient at different time points after the operation, for example, the predicted survival rate is 82% at the first day after the operation, 85% at the third day, and 87% at the seventh day, showing a gradual upward trend, indicating that the recovery is progressing well; if the predicted value decreases, it indicates that there may be complications, which need to be intervened in time.

[0065] In another embodiment of the application, the system further comprises a fat distribution region determination module, a blood flow dynamics state determination module, a mechanical parameter determination module and an initial scheme generation module, which together realize the preoperative planning function. The fat distribution region determination module is used to determine the fat distribution region, and the module plans the injection region of the fat based on the anatomical structure of the transplantation site and the aesthetic requirements. The boundary of the fat distribution region is determined based on the region center, the region coefficient and the region width, and the specific calculation method is as follows: taking the aesthetic center point of the transplantation site as the region center, the coordinate is marked as , and the unit is mm; the region coefficient is a dimensionless parameter, and the value range is 1.5 to 3.0, and the preferred value is different for different parts, preferably 2.5 for facial filling, preferably 2.0 for breast reconstruction, and preferably 1.8 for hip shaping, and the region coefficient reflects the expansion multiple of the distribution region relative to the region width; the region width , and the unit is mm, which is determined according to the transplantation volume, and generally takes a value of 1.2 to 1.5 times of the transplantation radius. For the transplantation volume (unit: ml), the transplantation radius can be approximately estimated as , and then the region width is determined. The region boundary point satisfies the distance from the region center in the range of , that is, and For example, in a breast reconstruction case, 100 ml of fat was planned to be transplanted, with the injection center located at the lower outer quadrant of the breast, with coordinates (120 mm, -50 mm), and a region coefficient of 2.0 Taking 2.0, the radius of the transplantation was calculated to be about 28 mm, and the width of the region was about 56 mm. Taking 30 mm, the maximum distance of the region boundary from the center was 60 mm, covering an area with a diameter of about 120 mm. The doctor planned 15 injection points evenly distributed in this area, with each point injecting 6 to 7 ml of fat, ensuring uniform distribution of the graft and adequate blood supply.

[0066] The hemodynamic state determination module is connected to the fat distribution region determination module and is used to determine the hemodynamic state within the fat distribution region, including blood flow velocity, pressure, and flow direction. This module obtains blood flow parameters in the transplantation region through color Doppler ultrasound or angiography. Blood flow velocity is calculated by Doppler shift, measuring the axial blood flow velocity component along the blood vessel, with units of mm / s; pressure is directly measured by a pressure catheter or indirectly estimated by blood flow velocity, with units of mmHg; flow direction is identified by color coding, with red indicating flow towards the probe and blue indicating flow away from the probe, and the spatial orientation of the blood vessel can be obtained through three-dimensional reconstruction. In clinical practice, blood flow parameters vary at different sites: the average blood flow velocity in facial filling regions is 0.8 mm / s, the local pressure is 25 mmHg, and the blood vessels mainly run vertically from the deep layer to the shallow layer; the average blood flow velocity in breast reconstruction regions is 0.65 mm / s, the local pressure is 18 mmHg, and the blood vessels mainly run radially from the chest wall to the periphery; the average blood flow velocity in buttock shaping regions is 0.55 mm / s, the local pressure is 15 mmHg, and the blood vessels mainly run in layers from the deep fascia to the subcutaneous layer. These parameters are important indicators for evaluating the blood supply capacity of the transplantation region. Regions with fast blood flow, moderate pressure, and uniform blood vessel distribution are suitable for large-volume fat transplantation, while regions with slow blood flow, low pressure, and sparse blood vessels are only suitable for small-volume transplantation.

[0067] The mechanical parameter determination module is connected with the hemodynamic state determination module, and is configured to determine mechanical parameters corresponding to different transplantation levels. The transplantation levels are generally divided into a shallow level (1-5 mm below the skin), a middle level (5-15 mm below the skin), and a deep level (more than 15 mm below the skin). Different levels of tissue have different mechanical properties: the shallow tissue is dense, mainly composed of the dermal reticular layer and shallow fat, and has an elastic modulus of about 12 kPa, is relatively hard, has a high blood vessel density but small lumen, and is suitable for small granular fat transplantation; the middle tissue is loose, mainly composed of fat lobules and fibrous septa, and has an elastic modulus of about 8 kPa, moderate softness, moderate blood vessel density and lumen size, and is suitable for medium granular fat transplantation; the deep tissue is soft, mainly composed of large fat pads, and has an elastic modulus of about 4 kPa, is very soft, has a relatively large blood vessel lumen but low density, and is suitable for large granular or block fat transplantation. The mechanical parameters also include the viscoelastic characteristics, stress relaxation time constant, and creep characteristics of the tissue, which are measured by a biomechanical testing instrument or obtained through finite element simulation calculation.

[0068] The initial scheme generation module is connected with the mechanical parameter determination module, and is configured to generate an initial fat transplantation scheme based on the fat distribution area, the hemodynamic state, and the mechanical parameters. The initial scheme includes the number and position distribution of injection points, the injection volume of each injection point, the allocation ratio of fat at different levels, the injection speed, and the specification of the injection needle. The principle of scheme generation is: maximizing the transplantation volume within the range of blood supply capacity; optimizing the level allocation under the mechanical constraint to make the stress state of each level in a reasonable interval; and ensuring the natural and aesthetic appearance after transplantation under the anatomical and aesthetic constraints. For example, for facial filling, since the facial tissue is tight, the blood supply is rich but the capacity is limited, the initial scheme tends to be multiple point small dose injection, the injection level is preferably the shallow and middle layers (ratio 3:7), the single point injection amount is controlled to be less than 0.5 ml, and the total injection amount is determined according to the degree of facial depression, and is generally between 10 and 30 ml. For breast reconstruction, since the breast has a large capacity and soft tissue, the initial scheme can use relatively concentrated injection points, the injection level is preferably the middle and deep layers (ratio 4:6), the single point injection amount can reach more than 5 ml, and the total injection amount is determined according to the volume of the breast defect, and is generally between 100 and 300 ml.

[0069] In another embodiment of the application, the system further comprises a scheme retrieval module, an iterative control module and a scheme output module, which realize the intelligent optimization function of the transplantation scheme. The scheme retrieval module is connected with the initial scheme generation module, and is used for retrieving the transplantation scheme with the highest survival rate in the historical transplantation data by using the transfer learning algorithm in combination with the initial fat transplantation scheme. The module maintains a historical case database, and the database contains complete data of thousands of fat transplantation cases, each of which includes patient characteristics (age, gender, BMI, transplantation site, etc.), transplantation scheme parameters (injection point distribution, injection volume of each point, hierarchical allocation ratio, etc.), predicted survival rate and actual survival rate. The retrieval algorithm adopts a case-based reasoning method based on similarity: firstly, the similarity between the current patient and each historical case in the database is calculated, and the similarity indexes include patient characteristic similarity and transplantation site similarity, and the indexes are weighted and summed when calculating; then, the 10 cases with the highest similarity are selected according to the similarity from high to low; finally, the scheme with the highest actual survival rate is selected from the 10 cases as the reference scheme. If the survival rate of the reference scheme has reached more than 90%, the reference scheme is taken as the recommended scheme; if the survival rate of the reference scheme is less than 90%, the initial scheme needs to be adjusted and optimized.

[0070] The iterative control module is connected with the scheme retrieval module, and is used for controlling the scheme retrieval module to repeatedly retrieve until the survival rate of the retrieved transplantation scheme with the highest survival rate is greater than or equal to 90%. The module realizes closed-loop optimization: if the reference scheme retrieved each time does not reach the target of 90%, the differences between the scheme and the initial scheme are analyzed, and the parameters that can be improved (such as increasing the number of injection points, adjusting the hierarchical allocation ratio, changing the single-point injection volume, etc.) are identified; based on these improvement directions, a new candidate scheme is generated; the survival rate of the new candidate scheme is predicted, and if the predicted value is improved, the candidate scheme is taken as the new initial scheme to enter the next round of retrieval; the process is repeated until a scheme with a predicted survival rate greater than or equal to 90% is found or a preset maximum number of iterations (usually set to 5 times) is reached. The convergence of the iterative optimization depends on the continuity of the scheme parameter space and the monotonicity of the prediction model: within a reasonable parameter range, improving the scheme parameters can generally improve the survival rate, so the iterative process can gradually approach the optimal scheme.

[0071] The scheme output module is connected with the iterative control module, and is used for outputting the transplantation scheme with the highest survival rate with a survival rate greater than or equal to 90% as the fat transplantation scheme. The output scheme is presented in the form of a structured report, and the report content includes: patient basic information, transplantation site and target, recommended injection point distribution map (annotating the position coordinates of each injection point), injection volume and injection level of each injection point, expected survival rate and its confidence interval, main risk factors and matters needing attention. The doctor can make preoperative planning and intraoperative implementation according to the report, or make fine tuning to the scheme according to the actual situation.

[0072] In a practical application, a patient undergoing facial filling, the initial plan predicts a survival rate of 78%, which does not meet the target of 90%. The system retrieves historical cases and finds that a similar patient has a survival rate of 85% after using an adjusted plan (adjusting the injection layer distribution from shallow: middle = 2:8 to 3:7, and reducing the single-point injection volume from 0.8ml to 0.5ml). The system generates a second version of the candidate plan based on this, with a predicted survival rate of 85%, which still does not meet the target. Continue to optimize, adjust the injection point spacing (from the original 6mm to 4mm), increase the number of injection points, and the third version of the plan predicts a survival rate of 92%, meeting the optimization target. The system outputs this plan as the final recommended plan, and the doctor adopts this plan to perform the surgery, with an actual survival rate of 89% after 3 months of follow-up, close to the predicted value, and the patient is very satisfied with the results.

[0073] In another embodiment of the present application, the system further comprises a blood flow ratio calculation module and a blood flow resistance calculation module for quantitative evaluation of the blood supply sufficiency of the transplantation area. The blood flow ratio calculation module uses the percentage of the blood flow area of the fat transplantation site in the total area as the blood flow ratio, which is a key indicator for evaluating the survival rate of fat transplantation. The blood flow area is identified by color Doppler ultrasound image, and the area with blood flow signal in the image (i.e. the color-coded area) is marked as the blood flow area, and the area without blood flow signal is marked as the non-blood flow area; the number of pixels of the blood flow area is calculated by image segmentation algorithm, and combined with the spatial resolution of the image, it is converted into the actual area, with the unit of mm²; the total area of the transplantation area is also calculated; the blood flow ratio is equal to the blood flow area divided by the total area, expressed in percentage. Clinical studies have shown that the blood flow ratio is highly correlated with the survival rate: when the blood flow ratio is less than 30%, the survival rate of fat decreases significantly, usually less than 60%; when the blood flow ratio reaches more than 50%, the survival rate can remain above 80%, with good results; when the blood flow ratio reaches more than 70%, the survival rate can reach 90% or even higher, with excellent results. Therefore, in the plan optimization process, improving the blood flow ratio is one of the key targets.

[0074] The blood flow resistance calculation module is connected to the blood flow ratio calculation module and is used to calculate blood flow resistance based on biomechanical simulation. Blood flow resistance reflects the magnitude of resistance encountered by blood flowing through the transplantation area; the greater the resistance, the more difficult the blood flow and the more insufficient the blood supply. The calculation of blood flow resistance is based on the principles of hemodynamics: blood flow resistance is directly proportional to the length of the blood vessel, inversely proportional to the fourth power of the vessel radius, and also related to blood viscosity. Within the transplantation area, the vascular network is complex, making it difficult to measure the parameters of each vessel individually. Therefore, a numerical simulation method is used: first, a three-dimensional model of the vascular network in the transplantation area is reconstructed based on angiography or high-resolution ultrasound; then, computational fluid dynamics software is used to simulate blood flow in the vascular network, calculating the flow velocity and pressure at each location; finally, blood flow resistance is calculated based on the pressure gradient. (Blood flow resistance at the current location is shown.) Initial blood flow resistance at the blood injection site The value is determined together with the force integral from the starting point to the current point, and the calculation formula is:

[0075] ,

[0076] in, This indicates the blood flow resistance at the current location, expressed in mmHg·s / mm. This represents the blood flow resistance at the point of blood injection, and is the initial resistance value. It depends on the resistance of the main blood supply vessel and is estimated by the vessel diameter and blood flow velocity. Indicates force from 0 to The integral value reflects the cumulative resistance change from the injection point to the current point; This is the resistance function along the path, which is related to factors such as the radius distribution, branching, and tortuosity of blood vessels along the path; This is a path parameter, representing the distance along the blood vessel path from the starting point to the current point. This integral term calculates the sum of all resistances encountered by blood flowing from the injection point to the current point. The thinner the blood vessel, the more branches, and the longer the path, the larger the integral value, and the greater the blood flow resistance at the current point.

[0077] Blood flow resistance is inversely proportional to the blood flow ratio:

[0078] ,

[0079] in, The blood flow ratio represents the percentage of the blood flow area to the total area. This is a proportionality coefficient, measured in mmHg·s / mm. Its value is adjusted according to tissue type. Facial tissues, due to their high vascular density and small vessel diameter, are more sensitive to the influence of blood flow resistance on the blood flow ratio. The value is approximately 25; the blood vessels in breast tissue have a relatively large diameter, and the resistance has a relatively small impact. The value is approximately 20; the blood vessels in the buttock tissue are relatively dispersed. The value is about 15. The formula shows that the greater the blood flow resistance, the smaller the blood flow ratio, and the less the blood supply. Reducing blood flow resistance is an effective way to improve blood supply, which can be achieved by optimizing injection technology (such as using blunt needles to reduce blood vessel damage), controlling the volume of the graft (avoiding excessive tissue tension to compress blood vessels), and improving postoperative care (such as avoiding excessive pressure bandaging).

[0080] In a case of breast reconstruction, the initial blood flow resistance at the injection point was measured to be 0.5 mmHg·s / mm by color Doppler ultrasound and computational fluid dynamics simulation. At a distance of 20 mm from the injection point, the cumulative resistance increment was calculated to be 0.3 mmHg·s / mm by path integration, so the total blood flow resistance at that point was 0.8 mmHg·s / mm. According to the formula, the blood flow ratio is , which is lower than the ideal value (more than 30%), indicating that there may be a risk of insufficient blood supply in this area. The doctor adjusted the distribution of injection points accordingly, increasing the injection points around the area to improve the connectivity of the blood flow network and reduce the blood flow resistance. The blood flow ratio was recalculated to be 38%, and the blood supply condition was improved. The final survival rate increased from the predicted 76% to 84%.

[0081] The data flow process of the system is shown in Figure 7 : the preoperative MRI image is collected by the image data collection module 1 and transmitted to the fat distribution density prediction module 2, which generates a fat distribution density prediction model after processing; the intraoperative pressure and morphological data are collected by the cell integrity information collection module 3 and transmitted to the cell survival rate prediction module 4, which generates a cell survival rate prediction model after processing; the postoperative thermal imaging data are collected by the temperature image collection module 5 and transmitted to the blood supply state prediction module 6, which generates a blood supply state prediction model after processing; the outputs of the three prediction models are aggregated in the model fusion module 7, which generates a postoperative survival rate prediction model combined with the patient's baseline information; the postoperative survival rate prediction model is transmitted to the transfer learning module 8, which uses animal experiment data and human data for transfer learning to optimize the model parameters and obtain the optimal parameter solution set; the optimal parameter solution set is transmitted to the survival rate output module 9, which generates the final prediction result and outputs. The entire data flow process has high automation, and the modules communicate with each other through standardized data interfaces to ensure the accuracy and timeliness of data transmission. The system also sets up a central control module to coordinate the execution timing between modules to ensure that data collection, model training, and prediction output are executed in the correct order, avoiding data inconsistency or logical errors.

[0082] The core advantage of the system of the present application is to realize the full-process automation from data acquisition, model construction to prediction output, greatly improving the prediction efficiency and accuracy. Compared with the traditional experience evaluation method, the system is based on quantitative multi-source data analysis, and the prediction result is objective and reliable, not affected by subjective factors. Compared with single modality imaging evaluation, the system integrates preoperative, intraoperative and postoperative information, and the prediction is more comprehensive and accurate. Compared with directly applying animal model parameters, the system effectively solves the problem of data scarcity in human body through transfer learning technology, and improves the generalization ability of the model. The clinical verification results show that the prediction accuracy of the system is more than 85%, which is significantly higher than 60 to 70% of the traditional method; the survival rate of the transplantation scheme optimized by the system is increased by an average of 15 percentage points, the secondary surgery rate is reduced by more than 50%, and the patient satisfaction is significantly improved.

[0083] The application scenarios of the system are wide, including but not limited to: facial filling (such as cheek hollow, temple hollow, tear groove filling), breast reconstruction (such as breast reconstruction after breast cancer surgery), hip shaping (such as hip enlargement, hip type improvement), repair of soft tissue defects of limbs, etc. For different application scenarios, the system can automatically adjust the model parameters and optimization strategy to generate personalized prediction results and transplantation schemes. The system also has self-learning ability, and as the clinical application cases continue to accumulate, the system can continuously update the model parameters and optimize the prediction algorithm to realize long-term performance improvement.

[0084] Although the present application has been disclosed as above with preferred embodiments, it is not intended to limit the present application, and any person skilled in the art can make various modifications and changes without departing from the spirit and scope of the present application, therefore the protection scope of the present application should be limited by the scope defined in the claims.

Claims

1. A fat graft survival rate prediction system based on transfer learning and biomechanical simulation, characterized in that , comprising: an image data acquisition module configured to acquire image data of preoperative fat tissue; a fat distribution density prediction module connected to the image data acquisition module and configured to establish a fat distribution density prediction model of a fat grafting individual based on the image data; a cell integrity information acquisition module configured to acquire integrity information of fat cells during surgery; a cell survival rate prediction module connected to the cell integrity information acquisition module and configured to establish a cell survival rate prediction model of the graft based on the integrity information; a temperature image acquisition module configured to acquire temperature images of the fat graft after surgery; a blood supply state prediction module connected to the temperature image acquisition module and configured to establish a graft blood supply state prediction model based on the temperature images; a model fusion module connected to the fat distribution density prediction module, the cell survival rate prediction module, and the blood supply state prediction module, and configured to combine the fat distribution density prediction model, the cell survival rate prediction model, and the blood supply state prediction model with baseline information of a patient after the models are fused to generate a postoperative survival rate prediction model of the graft; a transfer learning module connected to the model fusion module and configured to train the postoperative survival rate prediction model of the graft using animal experiment data, and to transfer training results of the animal experiment data to human graft data to obtain an optimal parameter solution set of the postoperative survival rate prediction model of the graft; a survival rate output module connected to the transfer learning module and configured to output a postoperative survival rate prediction result based on the optimal parameter solution set.

2. The fat graft survival rate prediction system based on transfer learning and biomechanical simulation according to claim 1, characterized in that The image data acquisition module includes an MRI image acquisition unit configured to acquire preoperative MRI images, and the MRI images include sequences of T1W, T2W, and DWI; the fat distribution density prediction module further includes a machine learning training unit and an SVM optimization unit, the machine learning training unit is configured to train the MRI images based on a machine learning algorithm to generate the fat distribution density prediction model, and the SVM optimization unit is configured to optimize parameters and parameter accuracy of the fat distribution density prediction model; the output result of the fat distribution density prediction model is represented as a regression equation about input features, and includes a bias term, a training number, an iteration number, an optimization function, and an L2 norm.

3. The fat graft survival rate prediction system based on transfer learning and biomechanical simulation according to claim 1, characterized in that The cell integrity information acquisition module includes: an aspiration state pressure sensing unit configured to acquire morphological and rheological characteristics of fat in an aspiration state before the graft is placed in a target region; a morphological feature extraction unit connected to the aspiration state pressure sensing unit and configured to extract morphological feature indicators of preoperative fat blocks based on the morphological and rheological characteristics; a pressurized state pressure sensing unit configured to acquire cell survival rate and pressure change data of fat in a pressurized state after the graft is placed in the target region; The cell survival rate prediction module establishes the cell survival rate prediction model based on the morphological feature indicators, and establishes a cell survival rate and pressure variable relationship model based on the cell survival rate and pressure change data in the pressurized state.

4. The fat graft survival rate prediction system based on transfer learning and biomechanical simulation according to claim 1, characterized in that The temperature image acquisition module comprises an infrared thermal imaging unit configured to acquire a temperature distribution image of the fat graft under a blood supply state; the blood supply state prediction module is configured to construct the blood supply state prediction model based on the temperature distribution image; the blood supply state prediction model is represented as a regression equation about input features, and comprises a bias term, a training number, an iteration number, an optimization function, and an L2 norm.

5. The fat graft survival rate prediction system based on transfer learning and biomechanical simulation according to claim 1, characterized in that The model fusion module comprises: a morphological information generation unit configured to associate the fat distribution density prediction model with the cell survival rate prediction model to generate graft morphological information, wherein the morphological information of a current frame is determined by the morphological information of a previous frame and a regression equation about pressure, cell morphology, and fat distribution density and parameters thereof; a cell survival degree calculation unit connected to the morphological information generation unit and configured to associate the graft morphological information with the blood supply state prediction model to generate graft cell survival degree, wherein the blood supply state of a current frame, the morphological information of the current frame, the blood supply state of a previous frame, and a regression equation about pressure, morphology, and fat distribution density and parameters thereof jointly determine the graft cell survival degree; a survival rate calculation unit connected to the cell survival degree calculation unit and configured to associate the graft cell survival degree with the morphological information to generate a postoperative survival rate, wherein the cell survival degree of a current frame, the morphological information of the current frame, the cell survival degree of a previous frame, and a regression equation about pressure, morphology, and fat distribution density and parameters thereof jointly determine the postoperative survival rate.

6. The fat graft survival rate prediction system based on transfer learning and biomechanical simulation according to claim 1, characterized in that The transfer learning module further comprises: a multi-source data analysis unit configured to analyze multi-source data before, during, and after human fat transplantation using the fat distribution density prediction model, the cell survival rate prediction model, and the blood supply state prediction model after synchronous acquisition to generate an analysis result; a parameter extraction unit connected to the multi-source data analysis unit and configured to extract parameters required by the postoperative survival rate prediction model of the graft based on the analysis result.

7. The fat graft survival rate prediction system based on transfer learning and biomechanical simulation according to claim 1, characterized in that Further comprising: a fat distribution area determination module configured to determine a fat distribution area, and a boundary of the fat distribution area is determined based on an area center, an area coefficient, and an area width; a hemodynamic state determination module connected to the fat distribution area determination module and configured to determine a hemodynamic state in the fat distribution area, wherein the hemodynamic state comprises blood flow velocity, pressure, and flow direction; a mechanical parameter determination module connected to the hemodynamic state determination module and configured to determine mechanical parameters corresponding to different transplantation levels; an initial scheme generation module connected to the mechanical parameter determination module and configured to generate an initial fat transplantation scheme based on the fat distribution area, the hemodynamic state, and the mechanical parameters.

8. The fat graft survival rate prediction system based on transfer learning and biomechanical simulation according to claim 7, characterized in that Further comprising: a scheme retrieval module connected to the initial scheme generation module and configured to combine the initial fat transplantation scheme and use a transfer learning algorithm to retrieve a transplantation scheme with the highest survival rate from historical transplantation data. The iteration control module is connected with the scheme retrieval module, and is configured to control the scheme retrieval module to repeatedly retrieve until the survival rate of the retrieved transplantation scheme with the highest survival rate is greater than or equal to 90%; The scheme output module is connected with the iteration control module, and is configured to output the transplantation scheme with the highest survival rate with the survival rate greater than or equal to 90% as the fat transplantation scheme.

9. The fat graft survival rate prediction system based on transfer learning and biomechanical simulation according to claim 8, characterized in that The system further comprises: The blood flow ratio calculation module is configured to take the ratio percentage of the blood flow area of the fat transplantation site to the total area as the blood flow ratio, as a key indicator for evaluating the survival rate of fat transplantation; The blood flow resistance calculation module is connected with the blood flow ratio calculation module, and is configured to calculate the blood flow resistance based on biomechanical simulation, wherein the blood flow resistance at the current point is determined by the initial blood flow resistance at the blood injection point and the force integral value from the start to the current point.

10. The fat graft survival rate prediction system based on transfer learning and biomechanical simulation according to any one of claims 1 to 9, characterized in that The system further comprises a central control module connected with the image data acquisition module, the cell integrity information acquisition module, the temperature image acquisition module, the fat distribution density prediction module, the cell survival rate prediction module, the blood supply state prediction module, the model fusion module and the transfer learning module, and configured to coordinate the data flow and execution timing between the modules.

Images