An infrared thermal imaging diabetic foot early risk assessment system based on biological heat conduction constraint
By using an infrared thermal imaging system constrained by biothermal conduction, combined with thermal sensing region segmentation and adaptive feature fusion, the problem of insufficient accuracy in the risk assessment of diabetic foot in existing technologies has been solved. This enables precise localization and efficient assessment of diabetic foot lesions, and improves the interpretability and stability of the assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HEFEI INSTITUTE OF PHYSICAL SCIENCE CHINESE ACADEMY OF SCIENCES
- Filing Date
- 2026-01-23
- Publication Date
- 2026-06-09
AI Technical Summary
Current technologies for assessing the risk of diabetic foot rely on single-modal infrared thermal imaging information, which lacks precise localization of the lesion area, resulting in low assessment accuracy. Furthermore, they lack comprehensive analysis of the biological heat conduction process and blood perfusion.
An infrared thermal imaging system based on biothermal conduction constraints is employed, combining thermal sensing region segmentation, feature extraction, and adaptive feature fusion. Through end-to-end automatic extraction of the foot region, inversion of blood perfusion rate using the Pennes biothermal conduction equation, and multi-source feature analysis using deep learning methods, a comprehensive evaluation of foot temperature, blood perfusion, and thermal texture patterns is achieved.
It improves the accuracy of localization and assessment of diabetic foot lesions, enhances the interpretability and stability of assessments, provides visualized risk assessment results, and facilitates clinical diagnosis.
Smart Images
Figure CN122175864A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence-assisted diagnosis, and in particular to an infrared thermal imaging early risk assessment system for diabetic foot based on biological thermal conduction constraints, which can be applied to early screening of diabetic foot, primary health service centers, and follow-up monitoring by family doctors. Background Technology
[0002] Diabetic foot is a condition characterized by foot infections, ulcers, and / or deep tissue damage caused by distal lower extremity nerve abnormalities and varying degrees of peripheral vascular disease in diabetic patients. It is a leading cause of disability and death in diabetic patients. The International Diabetic Foot Working Group's Guidelines for the Prevention and Management of Diabetes-Related Foot Diseases (2023 Edition) states that early identification of high-risk foot risk factors can effectively prevent 50%-70% of foot ulcers.
[0003] Existing methods for assessing the risk of diabetic foot mainly include the following categories:
[0004] 1. Screening based on physical stimulation involves preliminary assessment of diabetic foot risk through methods such as palpation of the dorsalis pedis artery and acupuncture. This type of method relies on the experience of professional physicians and is highly subjective (e.g., Chinese patent application "CN202510144430.9 - A Diabetic Foot Screening Device").
[0005] 2. Based on piezoelectric or optical sensors, some studies use multi-dimensional parameters such as plantar pressure and blood flow detection to assess local blood flow. However, the equipment is complex to deploy and costly, making it difficult to promote widely at the grassroots level (e.g., Chinese patent application "CN202411595210.X-Diabetic foot risk assessment method and system based on big data" and Chinese patent application "CN202510998599.0-A diabetic foot monitoring system").
[0006] Infrared thermal imaging is a non-contact, non-invasive, and radiation-free technology that can convert changes in surface thermal radiation related to microcirculatory disorders and nerve damage in the foot into visualized, digital information. This helps to identify high-risk feet earlier and provide risk warnings, and it has become a recommended technology in the "International Diabetes Foot Working Group Guidelines for the Prevention and Management of Diabetes-Related Foot Diseases (2023 Edition)". However, existing methods generally have the following shortcomings:
[0007] (1) Some of the foot regions are extracted by manual drawing or simple threshold segmentation, which has low accuracy and efficiency and is difficult to adapt to complex lighting and background environments. Others are extracted by visible light and infrared registration, which has problems such as leakage of patient privacy and increased equipment costs.
[0008] (2) Most analyses are based solely on surface temperature statistical characteristics, lacking modeling of internal physiological parameters such as biological heat conduction processes and blood perfusion, resulting in insufficient physical interpretability;
[0009] (3) Insufficient utilization of thermal texture information and lack of dedicated network structures designed for the low contrast and high noise characteristics of infrared.
[0010] (4) Features from different sources (temperature, physical parameters, texture) are generally simply spliced or fused with fixed weights, and the importance of features cannot be adaptively adjusted according to the specific pathological state;
[0011] Therefore, there is an urgent need for an early risk assessment system for diabetic foot that can combine infrared thermal imaging data, foot anatomy, and biological heat conduction equations, and integrate deep learning methods with physical priors, to achieve comprehensive analysis of foot temperature, blood perfusion rate, and thermal texture patterns, thereby improving the accuracy, stability, and interpretability of early risk assessment. Summary of the Invention
[0012] This invention provides an infrared thermal imaging early risk assessment system for diabetic foot based on biological thermal conduction constraints. It aims to solve the problem that existing technologies rely solely on single-modal infrared thermal imaging information, lack precise localization of lesion areas, and thus have low accuracy in identifying the progression of diabetic foot disease.
[0013] To achieve the above objectives, the present invention provides the following technical solution:
[0014] An infrared thermal imaging early risk assessment system for diabetic foot based on bio-thermal conduction constraints includes:
[0015] The data acquisition module is used to acquire pseudo-color infrared images of the subject's soles and the raw temperature matrix;
[0016] The thermal sensing region segmentation module is used to automatically extract the foot region from the pseudo-color infrared image end-to-end, and output a pseudo-color infrared image containing only the foot region and a foot temperature matrix.
[0017] The feature extraction module is used to extract temperature features and physical parameter features with pathological basis based on the plantar temperature matrix and pseudo-color infrared image, and to extract high-dimensional thermal texture features from the pseudo-color infrared image.
[0018] The dual-channel adaptive feature fusion module is used to dynamically generate weights based on pathological indicators and adaptively fuse temperature features, physical parameter features and thermal texture features.
[0019] The risk assessment module is used to perform classification reasoning based on the fused feature vectors and output the risk assessment level of diabetic foot.
[0020] The human-computer interaction module is used to deploy the entire system and show users the risk assessment reasoning process and results, and visualize high-risk areas.
[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0022] End-to-end automatic foot region extraction: The thermal sensing ConvNeXt encoder and boundary-aware upsampling decoder, designed for the low contrast and high noise characteristics of infrared, enable accurate foot segmentation without preprocessing the original image under complex backgrounds and diverse shooting conditions, improving the reliability and automation of subsequent analysis.
[0023] Integrating biological heat conduction constraints to improve physical interpretability: The Pennes biological heat conduction equation is embedded as a physical constraint into the neural network loss function. By inverting the blood perfusion rate distribution through PINN, the output results not only fit the observed temperature data, but also satisfy the thermophysiological laws, thereby improving the physiological interpretability and credibility of risk assessment.
[0024] Multi-source feature deep mining combined with anatomical priors: Based on the anatomical division of the foot, the comprehensive use of temperature statistical features, bi-foot asymmetric temperature difference features, blood perfusion rate features and deep thermal texture features enables a multi-angle characterization of pathological changes in diabetic foot, which is beneficial for identifying early minor abnormalities.
[0025] Pathology-gated adaptive feature fusion: The dual-channel adaptive feature fusion module can automatically adjust the weights of physical and texture features according to the pathological state of the current sample. When physical abnormalities are significant, the physical channel is emphasized, and when texture abnormalities are obvious, the texture channel is emphasized. This is superior to simple splicing or fixed weighting strategies, and improves classification accuracy and generalization ability.
[0026] Visual and clinically friendly human-computer interaction: The system can intuitively display the distribution of plantar temperature, the heat map of blood perfusion rate, and the outline of high-risk areas, and provide a comprehensive risk score, which makes it easier for doctors to understand the basis of the system's decision-making and assists in clinical diagnosis and follow-up management. Attached Figure Description
[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.
[0028] Figure 1 This is an overall structural diagram of an infrared thermal imaging early risk assessment system for diabetic foot based on biological thermal conduction constraints, according to an embodiment of the present invention.
[0029] Figure 2 This is a schematic diagram of the network structure of the thermal sensing region segmentation module of the present invention, including a thermal sensing ConvNeXt encoder, a boundary sensing upsampling decoder, and a multi-scale fusion module.
[0030] Figure 3 This is a schematic diagram of the foot region partitioning of the present invention;
[0031] Figure 4 This is a schematic diagram of the feature extraction module of the present invention;
[0032] Figure 5 This is a schematic diagram of the dual-channel adaptive feature fusion module structure of the present invention. Detailed Implementation
[0033] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
[0034] like Figure 1 As shown, this invention proposes an infrared thermal imaging early risk assessment system for diabetic foot based on bio-thermal conduction constraints, for achieving accurate and efficient early risk assessment of diabetic foot. The system automatically and accurately extracts the plantar region; extracts various temperature statistics and asymmetric features based on foot anatomical divisions; introduces the Pennes bio-thermal conduction equation to construct a Physical Information Neural Network (PINN) to invert the blood perfusion rate distribution of the foot; combines a convolutional neural network structure adapted to infrared characteristics to extract deep thermal texture features; achieves adaptive dual-channel fusion of temperature / physical features and texture features through a pathological gating network; and displays the risk assessment results and high-risk areas in a visual manner, thereby achieving an objective, precise, and interpretable assessment of early diabetic foot risk. The system includes: a data acquisition module, a thermal sensing region segmentation module, a feature extraction module, a dual-channel adaptive feature fusion module, a risk assessment module, and a human-computer interaction module.
[0035] The data acquisition module is used to acquire pseudo-color infrared images of the subject's soles and the raw temperature matrix. The module includes an infrared thermal imager with a thermal sensitivity of less than 20 mK, capable of simultaneously outputting pseudo-color infrared thermal images and temperature matrices. The infrared thermal imager possesses high thermal sensitivity (less than 20 mK), ensuring accurate temperature measurements.
[0036] The thermal sensing region segmentation module is used to automatically extract the foot region end-to-end from the infrared image output by the data acquisition module, outputting only a pseudo-color infrared image of the foot region and a plantar temperature matrix. The thermal sensing region segmentation module includes: a custom encoder based on the ConvNeXt architecture adapted to the low contrast and high noise characteristics of infrared data, a boundary-aware upsampling decoder, and a multi-scale fusion module. The thermal sensing region segmentation module is as follows... Figure 2 As shown, this module receives the raw infrared image and its temperature matrix from the data acquisition module. It employs a customized thermal encoder based on the ConvNeXt architecture, a boundary-aware upsampling decoder, and a multi-scale fusion module to automatically segment the foot region from end to end in the infrared thermal image, outputting only the temperature matrix and / or mask of the foot region. The customized encoder based on the ConvNeXt architecture specifically includes an early high-resolution Stem layer and a thermal sensing ConvNeXt module.
[0037] The early high-resolution Stem layer is used to perform a 3×3 convolution with a stride of 2 on the input infrared image, mapping the input to an initial feature map. Compared with the original ConvNeXt design, it retains a higher spatial resolution, which is beneficial for capturing fine edges in subsequent stages. The specific implementation is as follows:
[0038] Given an input infrared image The initial feature projection is:
[0039] ,
[0040] in, The number of channels in the input image. Represents the height of the input image. Represents the width of the input image. This represents a 3×3 convolution with a stride of 2.
[0041] The thermal sensing ConvNeXt module specifically introduces Adaptive Normalization (ACN), Depth Dynamic Convolution (DyDWConv), and Channel Reweighting mechanisms. The overall mathematical expression of the thermal sensing ConvNeXt module is as follows:
[0042] ,
[0043] in, As input features, For output features, and These are 1×1 convolutions for channel expansion and contraction, respectively. This represents the channel reweighted attention function. This indicates a reweighting operation. Represents the Gaussian activation function. This indicates a depthwise dynamic convolution operation using a 1×1 convolution kernel. This represents the feature map obtained by adaptively normalizing the input feature X.
[0044] Adaptive Normalization (ACN) calibrates features using statistically aware affine parameters, and is implemented as follows:
[0045] ,
[0046] in, , These are the mean and standard deviation of the instances, and the dynamic scale. and It uses a lightweight multilayer perceptron to obtain global context Generated in, It is a global average pooling operation. It is an element-wise multiplication operation.
[0047] Depth-dynamic convolution (DyDWConv) transforms the original static weights into a dynamic form that depends on the input. The specific implementation is as follows:
[0048] ,
[0049] in, For input, It is a set of learnable static kernels. It is based on global features Predicted attention weights. It is the number of basic convolutional kernels. Is with The corresponding input feature map is used to perform convolution operations with dynamically combined convolution kernels. This indicates a depthwise dynamic convolution operation.
[0050] Channel reweighting is a reweighting mechanism inserted after the activation function to prioritize the retention of channels containing rich information and suppress noisy channels. The specific implementation is as follows:
[0051] ,
[0052] in, This represents the intermediate features after activation by the Gaussian Error Linear Unit (GELU). Indicates the feature The channel statistics vector obtained by performing global average pooling. This indicates a multilayer sensor used for channel recalibration. This indicates a channel reweighting operation.
[0053] The boundary-aware upsampling decoder specifically includes four stages, in the first stage... stage( The boundary-aware upsampling decoder receives decoded features from the previous (deeper) layer. and corresponding skip connection features from the encoder The specific implementation is as follows:
[0054] ,
[0055] in, Robust upsampling of responsible features Responsible for extracting boundaries from encoder features, Responsible for dynamically merging the two, This refers to the decoding features of the current layer. The specific submodule implementation is as follows:
[0056] ,
[0057] ,
[0058] ,
[0059] in, This represents a 3×3 convolution with a stride of 2. Indicates the decoding features Perform a 2x upsampling interpolation operation. This represents a 3×3 dilated convolution operation with a dilation rate of 2. This represents a gradient-based attention module. Indicates passage The obtained upsampled feature map, Indicates passage The obtained boundary enhancement feature map, This represents the fusion weight coefficient for the i-th decoding stage; This represents a 1×1 convolution.
[0060] The multi-scale fusion module is used to map the features output from each decoding stage to a unified channel dimension and align them to the same spatial resolution. The specific implementation is as follows:
[0061] ,
[0062] in, The scalar weights are obtained during network training. This represents a 1×1 convolution operation on the features of the i-th layer. This represents the bilinear interpolation operation. It is the output of the multi-scale fusion module.
[0063] The feature extraction module is used to extract pathologically relevant temperature features and physical parameter features based on the plantar temperature matrix and pseudo-color infrared image, and to extract high-dimensional thermal texture features from the pseudo-color infrared image. The feature extraction module is as follows: Figure 4 As shown. Based on the aforementioned plantar temperature matrix and pseudo-color infrared image, extracting temperature features with pathological basis may include: dividing the foot into multiple functional areas based on foot anatomy, such as... Figure 3 As shown, the temperature statistical features of each region, the asymmetric temperature difference features of the corresponding regions of both feet, and the temperature gradient features of adjacent blood vessels are extracted. A physical information neural network with the Pennes equation as the physical constraint is constructed to invert the blood flow perfusion rate distribution of the foot through infrared thermal sequences and extract the perfusion rate statistical features. Extracting high-dimensional thermal texture features from pseudo-color infrared images can include: using an improved convolutional neural network and texture-aware pooling strategy to extract deep thermal texture features from pseudo-color infrared thermal images.
[0064] The specific execution process of the feature extraction module includes:
[0065] Based on foot anatomy, the foot is divided into six independent regions, including the big toe region, the little toe region, the medial forefoot region, the lateral forefoot region, the arch region, and the heel region;
[0066] Calculate the temperature statistical characteristics of each region, including the region's average temperature, maximum temperature, minimum temperature, standard deviation, etc., and calculate the temperature difference asymmetry characteristics between the same region of both feet and the temperature difference gradient characteristics between adjacent blood vessels.
[0067] A Physical Information Neural Network (PINN) is constructed, which takes the spatial coordinates or space-time coordinates of each sampling point in the sole area as input and outputs the predicted temperature at the corresponding location. and blood perfusion rate The Pennes biological heat conduction equation was embedded as a physical constraint into the loss function, and infrared thermal imaging sequences were used as observation data to predict temperature. The distribution map of blood perfusion rate in the foot was inverted, and its statistical characteristics were extracted as physical features;
[0068] An improved convolutional neural network is used to encode pseudo-color infrared images, and a texture-aware pooling strategy is employed to extract deep thermal texture features.
[0069] The process of constructing a physical information neural network includes: defining the Penns biological heat conduction equation as:
[0070] ,
[0071] in, , , These are tissue density, specific heat capacity, and thermal conductivity, respectively. The blood perfusion rate to be inverted, Arterial blood temperature, For metabolic heat production, the temperature T is a predicted temperature output by a physical information neural network. express; Represents the gradient and divergence operators with respect to spatial coordinates. For time, It refers to blood;
[0072] Construct the total loss function :
[0073] ,
[0074] For data error loss, As a physical constraint loss, the network parameters are optimized by minimizing this loss function, so that the output temperature field approximates the observed data and satisfies the physical constraints of the Penns equation, thereby obtaining a blood perfusion rate that conforms to thermophysiological laws. . It is a weighting coefficient that balances the relative importance of data error loss and physical constraint loss.
[0075] The dual-channel adaptive feature fusion module dynamically generates weights based on pathological indicators and adaptively fuses temperature features, physical parameter features, and thermal texture features. The dual-channel adaptive feature fusion module is as follows: Figure 5 As shown: A pathological gating network is constructed, and anatomical and physical feature vectors (temperature features and blood perfusion statistical features, respectively) and thermal texture feature vectors are used as two-channel inputs. Multilayer perceptron and Softmax are used to generate dynamic gating weights to achieve adaptive weighted fusion of physical and texture channels to obtain risk feature vectors.
[0076] The specific implementation of the dual-channel adaptive feature fusion module is as follows: This module contains a pathology gating network to generate dynamic fusion weights;
[0077] The input consists of two parts: the first channel is the physical eigenvector from the temperature matrix. The second channel is the thermal texture feature vector from the pseudo-color infrared image. ;
[0078] The fused risk feature vector is calculated using the following formula. :
[0079] ,
[0080] ,
[0081] in, To achieve adaptive gating weights, the network automatically increases the weights when a significant anomaly in physical parameters is detected. The weight of the weight increases if the weight of ... The weight. This indicates that the two input vectors are concatenated along the feature dimension. This represents a multilayer perceptron. This represents the normalized exponential function.
[0082] The risk assessment module is used to perform classification reasoning based on the fused feature vectors and output the risk assessment level of diabetic foot.
[0083] The risk assessment module includes a fully connected classification network; this module receives the risk feature vector. Through multiple fully connected layers and Dropout layers, the final output of the probability distribution of categories "normal", "low risk" and "high risk" is obtained through the Softmax function.
[0084] The human-computer interaction module is used to deploy the entire infrared thermal imaging early risk assessment system for diabetic foot, and to display the risk assessment reasoning process and results to the user, visualizing high-risk areas. The human-computer interaction module includes a visual interface, and its specific functions include:
[0085] Real-time monitoring and display is used to display the infrared images collected by the data acquisition module and the foot region segmentation results obtained by the thermal sensing region segmentation module in real time;
[0086] Visualization of physical parameters is used to invert blood perfusion rate obtained from physical information through neural networks. Data such as these are mapped onto heat maps for display, to visually reflect areas of microcirculation obstruction in the feet;
[0087] High-risk area early warning is used to overlay the high-risk areas identified by the risk assessment module with a highlighted outline on the original infrared image or foot area segmentation results, and display a comprehensive risk score.
[0088] Experiments show that the system has an overall parameter count of 10M, a foot region extraction accuracy of 97.5%, and a three-class risk assessment accuracy of 88.4%. It provides a visual interface to realize real-time display of infrared thermograms and foot segmentation results, display of blood perfusion rate heatmaps, and display of high-risk area warnings and comprehensive risk scores.
[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. An infrared thermal imaging early risk assessment system for diabetic foot based on biological thermal conduction constraints, characterized in that, include: The data acquisition module is used to acquire pseudo-color infrared images of the subject's soles and the raw temperature matrix; The thermal sensing region segmentation module is used to automatically extract the foot region from the pseudo-color infrared image end-to-end, and output a pseudo-color infrared image containing only the foot region and a foot temperature matrix. The feature extraction module is used to extract temperature features and physical parameter features with pathological basis based on the plantar temperature matrix and pseudo-color infrared image, and to extract high-dimensional thermal texture features from the pseudo-color infrared image. The dual-channel adaptive feature fusion module is used to dynamically generate weights based on pathological indicators and adaptively fuse temperature features, physical parameter features and thermal texture features. The risk assessment module is used to perform classification reasoning based on the fused feature vectors and output the risk assessment level of diabetic foot. The human-computer interaction module is used to deploy the entire infrared thermal imaging early risk assessment system for diabetic foot and to show users the risk assessment reasoning process and results, and to visualize high-risk areas.
2. The infrared thermal imaging early risk assessment system for diabetic foot based on biological thermal conduction constraint according to claim 1, characterized in that, The data acquisition module includes an infrared thermal imager with a thermal sensitivity of less than 20 mK and capable of simultaneously outputting pseudo-color infrared images and temperature matrices.
3. The infrared thermal imaging early risk assessment system for diabetic foot based on biological thermal conduction constraint according to claim 1, characterized in that, The thermal sensing region segmentation module includes: an encoder based on the ConvNeXt architecture, adapted to the low contrast and high noise characteristics of infrared data, a boundary-aware upsampling decoder, and a multi-scale fusion module.
4. The infrared thermal imaging early risk assessment system for diabetic foot based on biological thermal conduction constraint according to claim 3, characterized in that, The encoder based on the ConvNeXt architecture specifically includes: an early high-resolution Stem layer and a thermal sensing ConvNeXt module; The early high-resolution Stem layer is used to perform a 3×3 convolution with a stride of 2 on the input infrared image, mapping the input to an initial feature map: Given an input thermal image The initial feature projection is: ， in, The number of channels in the input image. Represents the height of the input image. Represents the width of the input image. This represents a 3×3 convolution with a stride of 2; The thermal sensing ConvNeXt module introduces adaptive normalization, depthwise dynamic convolution, and channel reweighting; the overall mathematical expression of the thermal sensing ConvNeXt module is as follows: ， in, As input features, For output features, and These are 1×1 convolutions for channel expansion and contraction, respectively; This represents the channel reweighting attention function, which generates reweighting coefficients based on the importance of each channel and weights the features along the channel dimension. Represents the Gaussian error linear unit activation function. This indicates a depthwise dynamic convolution operation using a 7×7 convolution kernel. This represents the feature map obtained by adaptively normalizing the input feature X. Adaptive normalization uses statistically aware affine parameters to calibrate features, as specifically implemented below: ， in, , These are the mean and standard deviation of the instances, and the dynamic scale. and It uses a lightweight multilayer perceptron to obtain global context Generated in, It is global average pooling. It is an element-wise multiplication operation; Depth dynamic convolution replaces traditional static convolution weights with dynamic convolution weights that adapt to the input features. The specific implementation is as follows: ， in, For input, For dynamically combined convolutional kernels, where It is a set of learnable static kernels. It is based on global features Predicted attention weights The number of basic convolutional kernels, For convolution operations, To and The corresponding input feature map is used to perform convolution operations with dynamically combined convolution kernels; This represents a depthwise dynamic convolution operation; Channel reweighting is a reweighting mechanism inserted after the activation function, implemented as follows: ， in, This represents the intermediate features after activation by Gaussian error linear units. GAP(H) represents global average pooling, where GAP(H) represents the sum of features. The channel statistics vector obtained by performing global average pooling. This indicates a multilayer sensor used for channel recalibration. The sigmoid activation function is used to compress channel weights to the range of 0 to 1. This indicates a channel reweighting operation.
5. The infrared thermal imaging early risk assessment system for diabetic foot based on biological thermal conduction constraint according to claim 3, characterized in that, The boundary-aware upsampling decoder specifically includes four stages, in the first stage... stage, The boundary-aware upsampling decoder receives decoded features from the previous layer. and corresponding skip connection features from the encoder The specific implementation is as follows: ， in, Robust upsampling of responsible features Responsible for extracting boundaries from encoder features, Responsible for dynamically integrating the two; This refers to the decoding features of the current layer; the specific submodule implementation is as follows: ， ， ， in, This represents a 3×3 convolution with a stride of 2. Indicates the decoding features Perform a 2x upsampling interpolation operation. This represents a 3×3 dilated convolution operation with a dilation rate of 2. This indicates a gradient-based attention module that generates a boundary attention map by calculating the spatial gradient magnitude of features. Indicates passage The obtained upsampled feature map, Indicates passage The obtained boundary enhancement feature map, This represents the fusion weight coefficient for the i-th decoding stage; This represents a 1×1 convolution.
6. The infrared thermal imaging early risk assessment system for diabetic foot based on biological thermal conduction constraint according to claim 3, characterized in that, The multi-scale fusion module is used to map the features output from each decoding stage to a unified channel dimension and align them to the same spatial resolution. The specific implementation is as follows: ， in, The scalar weights are obtained during network training. This represents a 1×1 convolution operation on the features of the i-th layer. This represents the bilinear interpolation operation. It is the output of the multi-scale fusion module.
7. The infrared thermal imaging early risk assessment system for diabetic foot based on biological thermal conduction constraint according to claim 1, characterized in that, The feature extraction module is used for: Based on foot anatomy, the foot is divided into six independent regions, including the big toe region, the little toe region, the medial forefoot region, the lateral forefoot region, the arch region, and the heel region; Calculate the temperature statistical characteristics of each region, including the region's average temperature, maximum temperature, minimum temperature, and standard deviation, and calculate the temperature difference asymmetry characteristics between the same-name regions of both feet and the temperature difference gradient characteristics between adjacent blood vessels. A physical information neural network is constructed, which takes the spatial coordinates or space-time coordinates of each sampling point in the sole area as input and outputs the predicted temperature at the corresponding location. and blood perfusion rate The process includes: embedding the Pennes biological heat conduction equation as a physical constraint into a loss function, using infrared thermal imaging sequences as observation data, and based on the predicted temperature. The distribution map of blood perfusion rate in the foot was inverted, and its statistical characteristics were extracted as physical features; An improved convolutional neural network is used to encode pseudo-color infrared images, and a texture-aware pooling strategy is employed to extract deep thermal texture features.
8. The infrared thermal imaging early risk assessment system for diabetic foot based on biological thermal conduction constraint according to claim 7, characterized in that, The process of constructing a physical information neural network includes: defining the Pennes biological heat conduction equation as: ， in, , , These are tissue density, specific heat capacity, and thermal conductivity, respectively. The blood perfusion rate to be inverted, Arterial blood temperature, For metabolic heat production, the temperature T is a predicted temperature output by a physical information neural network. express; Represents the gradient and divergence operators with respect to spatial coordinates. For time, It refers to blood; Construct the total loss function : ， For data error loss, As a physical constraint loss, the network parameters are optimized by minimizing this loss function, so that the output temperature field approximates the observed data and satisfies the physical constraints of the Penns equation, thereby obtaining a blood perfusion rate that conforms to thermophysiological laws. , It is a weighting coefficient that balances the relative importance of data error loss and physical constraint loss.
9. The infrared thermal imaging early risk assessment system for diabetic foot based on biological thermal conduction constraint according to claim 1, characterized in that, The specific implementation of the dual-channel adaptive feature fusion module is as follows: This module includes a pathological gating network for generating dynamic fusion weights; The input consists of two parts: the first channel is the physical eigenvector from the temperature matrix. The second channel is a depth thermal texture feature vector from a pseudo-color infrared image. ; The fused risk feature vector is calculated using the following formula. : ， ， in, To achieve adaptive gating weights, the network automatically increases the weights when a significant anomaly in physical parameters is detected. The weight of the weight increases if the weight of ... The weight, This indicates that the two input vectors are concatenated along the feature dimension. This represents the structure of a multilayer perceptron network. This represents the normalized exponential function.
10. The infrared thermal imaging early risk assessment system for diabetic foot based on biological thermal conduction constraint according to claim 9, characterized in that, The risk assessment module includes a fully connected classification network; this module receives the risk feature vector. Through multiple fully connected layers and Dropout layers, the final output of the probability distribution of normal, low-risk, and high-risk categories is obtained through the Softmax function.
11. The infrared thermal imaging early risk assessment system for diabetic foot based on biological thermal conduction constraint according to claim 1, characterized in that, The human-computer interaction module includes a visual interface, and its specific functions include: Real-time monitoring and display is used to display the pseudo-color infrared image collected by the data acquisition module and the foot region segmentation result obtained by the thermal sensing region segmentation module in real time; Visualization of physical parameters is used to invert blood perfusion rate obtained from physical information through neural networks. The data is mapped to a heat map for display, to visually reflect the areas of microcirculation obstruction in the feet; High-risk area early warning is used to overlay the high-risk areas identified by the risk assessment module with a highlighted outline on the original pseudo-color infrared image or segmentation results, and display a comprehensive risk score.