Diabetic foot ulcer risk intelligent early warning system based on multi-mode sensing

Abstract

The invention relates to a diabetic foot ulcer risk intelligent early warning system based on multi-mode sensing, and relates to the technical field of medical health, and the diabetic foot ulcer risk intelligent early warning system comprises the following steps: carrying out pressure and temperature data acquisition on a foot sole when a user walks, and generating a dynamic pressure image and temperature map; and identifying a pressure concentration area through image segmentation, and finding out temperature abnormal hot spots through threshold analysis. The two are spatially registered to determine a foot risk overlap region, and a time sequence accumulation is calculated to obtain an accumulated pressure heat map. Morphological analysis and gradient calculation are carried out based on the graph, and the geometric features and the pressure change rate of the risk area are extracted. A foot risk level is determined according to the characteristics, a risk early warning instruction is sent out after a preset early warning rule is matched, and the technical problems that continuous and objective monitoring means for key physiological parameters such as plantar pressure distribution and temperature change are lacked, and effective intervention is difficult to achieve in the early stage of tissue damage are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical health, in particular to a diabetic foot ulcer risk intelligent early warning system based on multi-modal sensing. BACKGROUND

[0002] With the continuous rise of the prevalence of diabetes, diabetic foot ulcer (DFU) as one of its serious complications has become an important clinical problem that threatens the quality of life and even the life safety of patients. Due to peripheral neuropathy and vascular disease caused by long-term hyperglycemia, diabetic patients often have foot sensory reduction and blood circulation disorders, making the foot prone to local tissue damage when bearing daily mechanical load, but the patient is difficult to detect. At present, the clinical monitoring of diabetic foot mainly relies on regular manual inspection and subjective evaluation, and there is a lack of continuous and objective monitoring means for key physiological parameters such as plantar pressure distribution and temperature change, making it difficult to achieve effective intervention in the early stage of tissue damage. SUMMARY

[0003] The present application aims to at least partly solve one of the problems in the prior art.

[0004] To achieve the above-mentioned purpose, the present application provides a diabetic foot ulcer risk intelligent early warning system based on multi-modal sensing, comprising the following steps: A collection module is used to collect pressure and temperature data of the plantar during walking, and obtain dynamic plantar pressure images and plantar surface temperature maps; A segmentation module is used to perform image segmentation processing based on the dynamic plantar pressure images, and obtain a pressure concentration area, and perform threshold analysis processing based on the plantar surface temperature maps, and obtain a temperature abnormal hot spot; A registration module is used to perform spatial position registration on the pressure concentration area and the temperature abnormal hot spot, and obtain a foot risk overlap area, and perform time sequence accumulation calculation on the foot risk overlap area, and obtain a cumulative pressure heat map; A calculation module is used to perform morphological analysis and gradient calculation based on the cumulative pressure heat map, and obtain risk area geometric features and pressure change rate features; A generation module is used to generate a foot risk level according to the risk area geometric features and the pressure change rate features, and match a preset early warning rule based on the foot risk level, and obtain a risk early warning instruction.

[0005] Further, the spatial position registration on the pressure concentration area and the temperature abnormal hot spot to obtain the foot risk overlap area comprises: The second clustering module is configured to perform spatial coordinate transformation on the stress concentration area and the temperature abnormal hot spot to obtain a unified coordinate feature map, and perform geometric consistency correction based on the unified coordinate feature map to obtain a corrected multi-modal distribution map. The mapping module is configured to perform feature vector field mapping on the foot bottom area based on the corrected multi-modal distribution map to obtain a fused feature vector field, and perform local density clustering through the fused feature vector field to obtain clustered multi-modal blocks. The second fusion module is configured to perform time sequence correlation analysis on the clustered multi-modal blocks to obtain a time sequence collaborative feature map, and perform weighted feature fusion based on the time sequence collaborative feature map to obtain a fused risk feature area. The positioning module is configured to perform boundary optimization integration on the foot bottom area through the fused risk feature area to obtain an optimized risk overlap area, and perform risk area positioning based on the optimized risk overlap area to obtain a foot risk overlap area.

[0006] Further, the foot risk overlap area is subjected to time sequence accumulation calculation to obtain an accumulated stress heat map, including: The accumulation module is configured to perform time dimension weighted accumulation on the foot risk overlap area to obtain a time sequence risk accumulation field, and perform dynamic trend extraction based on the time sequence risk accumulation field to obtain a risk evolution trend map. The smoothing module is configured to perform trend cluster segmentation on the foot bottom area based on the risk evolution trend map to obtain a trend cluster segmentation area, and perform time sequence gradient smoothing through the trend cluster segmentation area to obtain a smoothed risk trend area. The third fusion module is configured to perform multi-dimensional time sequence feature fusion on the smoothed risk trend area to obtain a fused trend feature map, and perform risk accumulation mapping based on the fused trend feature map to obtain the accumulated stress heat map.

[0007] Further, morphological analysis and gradient calculation are performed based on the accumulated stress heat map to obtain risk area geometric features and pressure change rate features, including: The denoising module is configured to perform noise filtering and edge enhancement processing on the accumulated stress heat map through a multi-scale morphological dilation-erosion operator to obtain a denoised and enhanced stress heat map, and perform adaptive threshold segmentation based on the denoised and enhanced stress heat map to obtain an initial risk area mask. The second extraction module is configured to perform contour extraction and polygon approximation processing based on the initial risk area mask to obtain a risk area approximation contour, and perform geometric parameter calculation on the risk area approximation contour to obtain risk area geometric features. a decomposition module configured to perform dynamic gradient extraction on the cumulative pressure heat map by time-difference gradient field calculation to obtain a time gradient distribution field, and perform rate vector decomposition based on the time gradient distribution field to obtain a rate vector feature map; a quantization module configured to perform time-series rate clustering on the plantar region based on the rate vector feature map to obtain a clustered rate feature region, and perform weighted rate quantization through the clustered rate feature region to obtain a pressure change rate feature.

[0008] Further, generating a foot risk level based on the risk region geometric feature and the pressure change rate feature comprises: a transformation module configured to perform coordinate system transformation on the risk region geometric feature and the pressure change rate feature to obtain a foot fusion risk vector field, and perform local density gradient aggregation based on the foot fusion risk vector field to obtain an aggregated risk density distribution; a second tracking module configured to perform peak cluster spatiotemporal propagation tracking on the aggregated risk density distribution to obtain a propagation risk cluster dynamic map, and perform multi-level threshold vector quantization based on the propagation risk cluster dynamic map to obtain a foot risk level.

[0009] Further, performing peak cluster spatiotemporal propagation tracking on the aggregated risk density distribution to obtain a propagation risk cluster dynamic map comprises: a discretization processing module configured to perform grid discretization processing on the aggregated risk density distribution to obtain a discretized risk density grid, and perform local extreme point detection based on the discretized risk density grid to obtain an initial risk peak point set; a correlation analysis module configured to perform spatiotemporal correlation analysis based on the initial risk peak point set to construct a peak point spatiotemporal correlation matrix, obtain an associated risk peak cluster, and perform dynamic diffusion range calculation on the associated risk peak cluster to obtain a peak cluster diffusion boundary; a drawing module configured to perform time-series evolution trajectory drawing based on the peak cluster diffusion boundary to obtain a risk peak cluster time-series trajectory map, and perform dynamic feature extraction on the risk peak cluster time-series trajectory map to obtain a propagation risk cluster dynamic map.

[0010] Further, performing time-series evolution trajectory drawing based on the peak cluster diffusion boundary to obtain a risk peak cluster time-series trajectory map comprises: an interpolation calculation module configured to perform key frame coordinate extraction on the peak cluster diffusion boundary to obtain a boundary key frame coordinate set, and perform neighborhood frame interpolation calculation based on the boundary key frame coordinate set to obtain an interpolated completed boundary coordinate; A fitting module is configured to perform boundary contour fitting based on the interpolation-completed boundary coordinates to obtain a continuous boundary contour curve, and to perform curvature mutation point detection on the continuous boundary contour curve to obtain a boundary feature mutation point set; An anchoring module is configured to perform time sequence trajectory anchoring based on the boundary feature mutation point set to obtain a feature point time sequence trajectory line, and to perform overall contour splicing on the feature point time sequence trajectory line to obtain a risk peak cluster time sequence trajectory graph.

[0011] The application provides a diabetic foot ulcer risk intelligent early warning system based on multi-modal sensing, which comprises the following steps: collecting pressure and temperature data of the foot bottom during user walking to obtain a dynamic foot bottom pressure image and a foot bottom surface temperature atlas; performing image segmentation processing based on the dynamic foot bottom pressure image to obtain a pressure concentration area, and performing threshold analysis processing based on the foot bottom surface temperature atlas to obtain a temperature abnormal hot spot; performing spatial position registration on the pressure concentration area and the temperature abnormal hot spot to obtain a foot risk overlap area, and performing time sequence accumulation calculation on the foot risk overlap area to obtain an accumulated pressure heat map; performing morphological analysis and gradient calculation based on the accumulated pressure heat map to obtain risk area geometric features and pressure change rate features; generating a foot risk grade according to the risk area geometric features and the pressure change rate features, and matching a preset early warning rule based on the foot risk grade to obtain a risk early warning instruction, so that the technical problem that there is a lack of continuous and objective monitoring means for key physiological parameters such as foot bottom pressure distribution and temperature change, and it is difficult to realize effective intervention in the early stage of tissue damage is solved, the geometric features (such as area and shape factor) and the pressure change rate features of the risk area are extracted by performing morphological analysis and gradient calculation on the accumulated pressure heat map, deep-level quantitative evaluation from 'abnormality or not' to 'abnormality degree and development trend' is realized, and the technical effect that multi-dimensional and interpretable discrimination basis is provided for risk grade division is realized. BRIEF DESCRIPTION OF DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0013] Figure 1 is a step schematic diagram of the diabetic foot ulcer risk intelligent early warning system based on multi-modal sensing in an embodiment of the present application; The object implementation, functional characteristics and advantages of the present application will be further described with reference to the embodiments and the drawings. DETAILED DESCRIPTION

[0014] Embodiments of the present application are described below in detail with reference to the accompanying drawings, wherein the same or similar components are denoted by the same or similar reference numerals throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application and should not be construed as limiting the present application. For the step numbers in the following embodiments, they are only set for the convenience of description and do not limit the order between the steps, and the execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0015] A multi-modal sensing-based diabetic foot ulcer risk intelligent early warning system according to an embodiment of the present application is described in detail below with reference to the accompanying drawings. First, a multi-modal sensing-based diabetic foot ulcer risk intelligent early warning system according to an embodiment of the present application will be described with reference to the accompanying drawings.

[0016] Figure 1 is a multi-modal sensing-based diabetic foot ulcer risk intelligent early warning system in an embodiment of the present application, comprising the following steps: The acquisition module is configured to collect pressure and temperature data of the foot bottom during user walking to obtain a dynamic foot bottom pressure image and a foot bottom surface temperature map. The segmentation module is configured to perform image segmentation processing based on the dynamic foot bottom pressure image to obtain a pressure concentration area, and perform threshold analysis processing based on the foot bottom surface temperature map to obtain a temperature abnormal hot spot. The registration module is configured to perform spatial position registration on the pressure concentration area and the temperature abnormal hot spot to obtain a foot risk overlap area, and perform time sequence accumulation calculation on the foot risk overlap area to obtain a cumulative pressure heat map. The calculation module is configured to perform morphological analysis and gradient calculation based on the cumulative pressure heat map to obtain risk area geometric features and pressure change rate features. The generation module is configured to generate a foot risk level according to the risk area geometric features and the pressure change rate features, and match a preset early warning rule based on the foot risk level to obtain a risk early warning instruction.

[0017] Specifically, in the implementation of a scheme for a diabetic foot ulcer risk intelligent early warning system based on multi-modal sensing, first, the user's foot bottom during daily walking is continuously subjected to pressure and temperature data collection by a collection module. This process relies on the cooperative work of a high-density pressure sensor array embedded in the insole or sock and an infrared thermal sensor network, thereby obtaining the pressure distribution changes and surface temperature distribution of the foot bottom at different stages of the gait cycle in real time. The collected data, after signal conditioning and digital processing, are organized into a two-dimensional matrix form of time series, and then form dynamic foot bottom pressure images and foot bottom surface temperature maps. For example, when the user is walking normally on flat ground, the heel and forefoot areas show obvious pressure peaks at the moment of touching the ground, accompanied by a slight local temperature rise. All this information is recorded and built into an image data set with spatial and temporal resolution. Subsequently, the processing flow enters the segmentation module. This module uses image segmentation algorithms such as threshold segmentation, region growing or deep learning semantic segmentation models for the dynamic foot bottom pressure images generated above. According to the preset pressure threshold or statistical distribution characteristics, the areas in the image above a certain pressure level are identified as pressure concentration areas. At the same time, threshold analysis is performed on the foot bottom surface temperature map, setting a temperature upper limit based on the group mean plus standard deviation (for example, above 35.5°C is considered abnormal), and marking areas exceeding this threshold as temperature abnormal hotspots. Thus, suspicious sites that may have mechanical overload and inflammatory response are extracted, for example, a patient's right foot arch lateral side continuously appears pressure points exceeding 200 kPa during multiple walks, and the corresponding area temperature reaches 36.2°C. This area is identified as a pressure concentration area and a temperature abnormal hotspot, respectively. Next, the registration module performs spatial position registration. Since the pressure image and temperature map may come from different sensor layouts or have sampling coordinate deviations, spatial transformation algorithms such as affine transformation or elastic registration techniques are used to map the two images to a unified spatial reference coordinate system, ensuring that the pressure concentration area and the temperature abnormal hotspot are accurately aligned in the anatomical structure. The overlapping area after registration is defined as the foot risk overlap area, indicating that this part is simultaneously subjected to high pressure and high temperature pathological factors, with a higher risk of tissue damage. On this basis, the registration module further performs time series accumulation calculation on the foot risk overlap area, i.e., repeatedly appearing risk overlap areas in multiple gait cycles or continuous hours of monitoring are superimposed and counted, with time weight given to reflect their frequency and duration. Finally, a cumulative pressure heat map reflecting the long-term load accumulation effect is generated, for example, a diabetic patient's left foot bottom area under the first metatarsal head is in a risk overlap state for 80% of the time during walking activities in a day. This area presents high-intensity bright color coding in the cumulative pressure heat map, directly showing that it is a chronic high-risk area.Then, the analysis phase of the calculation module is entered, which carries out morphological analysis based on the generated cumulative pressure heat map, removes noise and extracts connected domains by using mathematical morphological operations such as dilation, erosion, and opening and closing operations, and then calculates the geometric characteristic parameters of each risk area, including area size, perimeter, circularity, boundary irregularity, and other indicators, for evaluating the extension trend of the lesion; at the same time, gradient calculation is performed in combination with the cumulative heat map difference between adjacent time frames to obtain the pressure change rate characteristic, i.e., the growth slope of the cumulative pressure intensity per unit time, for judging the risk deterioration speed, for example, if an area has a large area and a jagged boundary with a continuous three-day upward trend in the cumulative pressure gradient, it indicates that the tissue in this area is in a rapid degeneration process. Finally, the risk level determination and early warning output are completed by the generation module, which inputs the geometric characteristics of the risk area and the pressure change rate characteristic obtained above into a pre-trained classification model or a rule-based decision engine, and generates a foot risk level by comprehensive scoring, such as dividing into three levels of low risk, medium risk, and high risk, and matching the level with the pre-set early warning rules in the system, for example, when the risk level is “high” and the pressure change rate exceeds the threshold, triggering a red early warning instruction and pushing it to the user's mobile phone APP or the medical staff management platform through Bluetooth or wireless network, reminding timely decompression, foot inspection or medical intervention. The whole process forms a closed loop from raw data collection to final risk early warning instruction output, realizing the whole process automation and fine monitoring and response of diabetic foot ulcer risk.

[0018] In specific embodiments, pressure and temperature data of the foot sole during walking of a user are collected to obtain dynamic foot sole pressure images and foot sole surface temperature maps, including: A sampling module is configured to perform high-frequency continuous sampling on the foot sole area of the user during walking through a pressure sensor array to obtain a foot sole pressure time sequence, and perform adaptive buffer fusion based on the foot sole pressure time sequence to obtain continuous pressure flow data. A fitting module is configured to perform spatio-temporal convolution processing on the foot sole area based on the continuous pressure flow data to obtain a pressure continuous distribution field, and perform dynamic surface fitting through the pressure continuous distribution field to obtain dynamic foot sole pressure images. A scanning module is configured to perform high-resolution thermal imaging scanning on the foot sole area of the user during walking through an infrared sensor array to obtain a foot sole temperature time sequence matrix, and perform dynamic region clustering based on the foot sole temperature time sequence matrix to obtain a temperature spatial distribution field. An extraction module is configured to perform multi-scale feature extraction on the foot sole area based on the temperature spatial distribution field to obtain a temperature distribution topology map, and perform curvature analysis through the temperature distribution topology map to obtain a foot sole surface temperature map.

[0019] Specifically, the process of collecting pressure and temperature data of the user's foot during walking to obtain dynamic foot pressure images and foot surface temperature maps is first performed by the sampling module. This module uses a high-density pressure sensor array deployed in the smart insole to continuously sample the pressure applied to each region of the foot at a frequency of no less than 100 Hz during the user's daily walking, obtaining raw pressure signals containing time and spatial position information, forming a foot pressure time sequence. For example, when the user transitions from standing to walking, the heel first touches the ground, generating an instantaneous impact force, and then the pressure gradually moves forward to the forefoot region. This series of dynamic changes is recorded by the sensor array as a time-ordered pressure value stream. To further improve data stability and eliminate sampling fluctuations caused by uneven gait rhythm, the system implements adaptive buffer fusion processing on the foot pressure time sequence. This involves dynamically adjusting the size of the sliding window based on the duration of the current gait cycle and the regularity of pressure peak occurrence, and using weighted averaging or Kalman filtering algorithms to smooth and integrate data at adjacent time points, thereby generating continuous pressure stream data with stronger continuity and less noise. The next step is the processing flow of the fitting module, which performs spatio-temporal convolution on the foot anatomical regions based on the continuous pressure stream data obtained earlier. This module uses a two-dimensional convolution kernel to capture the pressure propagation pattern between adjacent sensors in the spatial dimension, while modeling the trend and inertia of pressure changes in the time dimension. This results in a continuous pressure distribution field that reflects the evolution of pressure on the foot surface over time. For example, in normal gait, the wave-like progression of pressure from the heel to the first metatarsal head is accurately restored as a continuous spatial gradient change. Based on this, further dynamic surface fitting is performed using the pressure continuous distribution field. This uses bicubic spline interpolation or radial basis function methods to extend discrete sensor point pressure values to a continuous three-dimensional surface covering the entire foot surface. The height represents the pressure intensity, and the plane coordinates correspond to the foot position. Finally, a frame-by-frame dynamic foot pressure image is output, which intuitively presents the distribution pattern and transfer path of pressure in each step of the user.Meanwhile, the scanning module synchronously carries out the collection of temperature data. The scanning module uses an infrared sensor array embedded in the insole or the bottom of the shoe cavity to implement high-resolution thermal imaging scanning on the foot bottom area at a rate of more than 30 frames per second during the user's walking, capture the subtle temperature difference caused by the change of blood flow, inflammatory response or friction heat on the skin surface, and generate a foot bottom temperature time sequence matrix composed of multiple time slices, wherein each layer of the matrix corresponds to the temperature value measured by each sensing unit at a time; in order to extract stable heat distribution patterns from the time sequence matrix, the system performs dynamic region clustering analysis on the time sequence matrix, uses clustering algorithms such as DBSCAN or K-means, and divides the foot bottom into several statistically significant heat zones according to temperature similarity and spatial proximity, and then constructs a temperature spatial distribution field that can reflect the heat accumulation trend. For example, after a long time of walking, a diabetic patient's foot bottom under the fifth metatarsal head continuously appears an isolated hot spot higher than the surrounding area by more than 2°C, and this area is stably clustered as an independent high-temperature zone in multiple consecutive frames. Then the extraction module continues to process, which extracts multi-scale features of the foot bottom area based on the temperature spatial distribution field, decomposes the temperature structure at different resolutions by constructing a Gaussian pyramid or wavelet transform, identifies multi-level abnormal patterns from local tiny hot spots to large-scale temperature rise, and integrates to generate a temperature distribution topology map describing the complex structure of temperature distribution; further, by performing curvature analysis on the temperature distribution topology map, the bending degree of the isotherm and the temperature gradient discontinuity point are calculated, and the boundary area with sharp temperature gradient change is identified. These areas often correspond to early signs of active tissue metabolism or blood circulation disorders, and the result after curvature enhancement is finally organized into a foot bottom surface temperature map with clear structure and rich details, for example, the isotherm of the lateral side of the arch of a patient's foot is distorted and densely arranged, indicating the presence of a potential inflammatory edge, and this feature is accurately retained in the final map, providing reliable input for subsequent fusion analysis with other modal data. Through the collaborative operation of the sampling module, the fitting module, the scanning module and the extraction module, the complete transformation of the two types of physiological signals, pressure and temperature, from the original sensing data to the high-quality visualized atlas is realized, supporting the accurate implementation of the subsequent risk assessment link.

[0020] In specific embodiments, the dynamic foot bottom pressure image is subjected to image segmentation processing to obtain a pressure concentration area, including: A first fusion module is configured to calculate a gradient field of the dynamic foot bottom pressure image to obtain a pressure gradient vector map, and perform streamline tracking fusion based on the pressure gradient vector map to obtain a fused pressure streamline map. An expansion module is configured to perform region growing expansion on the foot bottom area based on the fused pressure streamline map to obtain an expanded pressure growing area, and perform boundary curvature optimization on the expanded pressure growing area to obtain an optimized pressure block. The marking module is configured to mark the communication components of the optimized pressure block to obtain marked pressure components, and correct the centroid offset based on the marked pressure components to obtain a pressure concentration area.

[0021] Specifically, the process of image segmentation based on the dynamic plantar pressure image is first performed by a first fusion module. The module performs gradient field calculation on the dynamic plantar pressure image, uses a differential operator such as Sobel or Scharr to calculate the pressure change rate of each pixel point in the image in the horizontal and vertical directions respectively, and thus constructs a pressure gradient vector diagram containing pressure gradient direction and intensity information. The vector diagram depicts the path and steepness of the transition of plantar pressure from low to high in the form of arrows or color coding. For example, during the walking process of a diabetic patient, the areas under the first and second metatarsal heads of the forefoot have been subjected to abnormally high load for a long time, and their pressure gradient vectors are obviously dense and point in a direction perpendicular to the longitudinal arch of the foot, reflecting that there is a significant pressure mutation in this area. On this basis, further streamline tracking fusion operations are performed based on the pressure gradient vector diagram, that is, the pressure propagation path is tracked along the gradient direction from multiple starting points to generate a series of continuous pressure streamlines, and the streamlines in multiple gait cycles are spatiotemporally fused by weighted superposition to form a fused pressure streamline diagram reflecting the dominant flow direction of pressure. This diagram not only retains the position of instantaneous pressure concentration, but also highlights the main channel of repeated stress. For example, in the continuous walking of a patient, the pressure streamlines always converge on the distal end of the third metatarsal bone, indicating that this part is a long-term stress concentration zone. Subsequently, the process enters the processing stage of an expansion module. The module performs region growing expansion on the plantar region based on the fused pressure streamline diagram, takes several seed points with the highest streamline density or the largest pressure value as the starting point, gradually expands the growing region outward according to the similarity threshold (such as a difference of less than 15 kPa) of the pressure values between adjacent pixels, and includes adjacent regions with similar mechanical properties into the same block, thereby obtaining an expanded pressure growth region. This process can effectively overcome the fragmentation problem caused by single threshold segmentation and ensure the integrity of the pressure concentration region. For example, when there are secondary high-pressure points around a hot spot but they do not reach the global threshold, they can still be reasonably included in the same risk zone through the region growing mechanism. To further improve the geometric reasonableness of the segmentation boundary, the system performs boundary curvature optimization processing on the expanded pressure growth region, uses an active contour model or a Level Set method to smooth and adjust the initial segmentation boundary, suppresses jagged artifacts, and strengthens the identification of recessed and protruding parts according to local curvature characteristics, and finally generates an optimized pressure block with continuous shape and clear boundary, making it more consistent with the actual stress distribution of the plantar anatomical structure.Next, the annotation module continues the subsequent processing. This module annotates the optimized pressure block with connected components, using a four-neighbor or eight-neighbor search algorithm to identify interconnected pixel sets, assigning a unique label to each independent connected component, thereby dividing the entire sole into several non-overlapping annotated pressure components. Each component corresponds to a potential pressure concentration unit. For example, in a diabetic patient with foot deformity, the system can annotate three independent high-pressure components: the medial longitudinal arch collapse area, the lateral heel area, and the area below the first metatarsal head. Based on this, centroid offset correction is further performed based on the annotated pressure components. The geometric center of each component is calculated, and its weighted centroid position is recalculated in combination with the pressure weight. If a significant deviation (e.g., more than 3 mm) is found between the geometric center of the annotated pressure component and its maximum pressure position, the component is finely adjusted according to the pressure gradient direction to ensure that the finally determined pressure concentration area accurately reflects the real high-load center and avoids positioning deviations caused by uneven sensor arrangement or foot slippage. The entire segmentation process, through the step-by-step processing of the first fusion module, the extension module, and the annotation module, accurately extracts pressure concentration areas with spatial continuity, reasonable boundaries, and accurate positioning from dynamic plantar pressure images. This provides high-quality structured input for subsequent spatial registration with temperature anomaly hotspots and risk overlap analysis, supporting the accurate discrimination capability of the diabetic foot ulcer risk early warning system.

[0022] In a specific embodiment, threshold analysis is performed based on the plantar surface temperature spectrum to obtain abnormal temperature hotspots, including: The first tracking module is used to adaptively divide the temperature spectrum of the foot surface into threshold regions to obtain an initial screening area of ​​temperature anomalies, and to perform time-series heat flow tracking based on the initial screening area of ​​temperature anomalies to obtain a heat flow trajectory map. The analysis module is used to perform spatial thermal cluster segmentation on the sole area based on the heat flow trajectory map to obtain thermal cluster segmentation blocks, and to perform gradient texture analysis on the thermal cluster segmentation blocks to obtain texture-enhanced thermal areas. The first clustering module is used to perform multi-feature weighted clustering on the texture-enhanced hot areas to obtain clustered inflammatory feature regions, and to locate abnormal points based on the clustered inflammatory feature regions to obtain temperature abnormal hot spots.

[0023] Specifically, the process of threshold analysis based on the plantar surface temperature map to obtain temperature anomaly hotspots is first executed by the first tracking module. This module performs adaptive threshold partitioning on the plantar surface temperature map, automatically determining the optimal threshold using algorithms such as the Otsu method or K-means clustering, and segmenting the image into background and potential temperature anomaly screening areas. The temperature anomaly screening areas represent local areas where the temperature is significantly higher than the surrounding environment. For example, in diabetic patients, due to sensory loss caused by neuropathy, certain areas may experience temperature increases due to repeated pressure. These areas are the initial screening areas. The selected temperature anomaly screening area is then further subjected to time-series heat flow tracing. By comparing the temperature change trends at consecutive time points, pixels with the same heating rate or cooling pattern are identified and connected to form a series of heat flow trajectory maps reflecting the heat diffusion path. This helps to capture the dynamic process of local temperature rise caused by inflammatory response. In practical applications, if a patient's forefoot lateral plantar surface shows a continuously rising temperature curve over several consecutive days, the heat conduction path of that area can be accurately depicted through time-series heat flow tracing, thus providing a basis for subsequent analysis. The system then enters the processing stage of the analysis module. Based on the heat flow trajectory map, this module performs spatial thermal cluster segmentation on the sole area. Using clustering algorithms such as DBSCAN and Mean Shift, the sole is divided into multiple independent spatial thermal cluster segments according to the density and directionality of the heat flow trajectory. Each segment corresponds to a relatively independent heat accumulation center. For example, in a diabetic patient with a risk of foot ulcers, the system can identify three main heat accumulation areas located on the inner and outer sides of the heel and below the first metatarsal head. To further improve the recognizability of thermal zone features, the system performs gradient texture analysis on the thermal cluster segments, calculates the temperature gradient distribution within each segment, and extracts texture features using the gray-level co-occurrence matrix (GLCM) to enhance the contrast of small temperature differences between different tissues. Finally, texture-enhanced thermal zones are generated, making early signs of inflammation hidden in complex backgrounds stand out. For example, during the detection process, a subtle but continuously increasing temperature gradient was found near the first metatarsal head of a high-risk patient. After texture enhancement, its boundary became clearer, which is beneficial for accurate diagnosis.Next, the first clustering module continues the subsequent processing. This module performs multi-feature weighted clustering on the texture-enhanced hot areas, comprehensively considering information from multiple dimensions such as temperature value, temperature gradient, and texture features. It uses methods such as fuzzy C-means (FCM) or hierarchical clustering to group pixels with similar features into one class, forming several clustered inflammatory feature areas with clear physiological significance. Each clustered inflammatory feature area may indicate a potential inflammatory lesion. For example, in a patient with long-term diabetes, the system can identify three clustered inflammatory feature areas located in the middle of the arch of the foot, the medial side of the heel, and between the second and fifth toes. Based on this, abnormal point localization is further performed based on the clustered inflammatory feature areas. Distance measurement or statistical methods are used to find the pixel point that deviates the farthest from the normal range as the temperature abnormality hotspot, ensuring that the located hotspot reflects the highest local temperature abnormality and meets the clinical diagnostic criteria. For example, when a patient's sole shows a significant temperature peak that exceeds the ipsilateral control area by more than 2°C, it can be identified as a true temperature abnormality hotspot. The entire processing flow, through the collaborative work of the first tracking module, the analysis module, and the first clustering module, enables the efficient and accurate extraction of pathologically significant temperature anomalies from the plantar surface temperature spectrum. This provides important reference information for the risk assessment of diabetic foot ulcers and supports the formulation and optimization of personalized medical plans.

[0024] In a specific embodiment, the spatial location of the pressure concentration area and the temperature anomaly hotspot is registered to obtain the foot risk overlap area, including: The second clustering module is used to perform spatial coordinate transformation on the pressure concentration area and the temperature anomaly hotspot to obtain a unified coordinate feature map, and to perform geometric consistency correction based on the unified coordinate feature map to obtain a corrected multimodal distribution map. The mapping module is used to map the feature vector field of the plantar region based on the corrected multimodal distribution map to obtain a fused feature vector field, and to perform local density clustering through the fused feature vector field to obtain clustered multimodal blocks; The second fusion module is used to perform temporal correlation analysis on the clustered multimodal blocks to obtain a temporal collaborative feature map, and to perform weighted feature fusion based on the temporal collaborative feature map to obtain a fusion risk feature region. The positioning module is used to optimize and integrate the boundaries of the sole area through the fused risk feature area to obtain an optimized risk overlap area, and to locate the risk area based on the optimized risk overlap area to obtain the foot risk overlap area.

[0025] Specifically, the process of spatially registering the pressure concentration area and the temperature anomaly hotspot to obtain the foot risk overlap area is first executed by the second clustering module. This module performs spatial coordinate transformation on the pressure concentration area and the temperature anomaly hotspot, using known sensor layout parameters or anatomical landmarks in the calibration image (such as the heel center, the positions of the first and fifth metatarsal heads) to map the two sets of data from the pressure sensing system and the infrared thermal imaging system to the same spatial reference coordinate system, forming a unified coordinate feature map containing pressure and temperature anomaly location information. For example, during continuous monitoring of a diabetic patient, the pressure concentration area below the first metatarsal head of the foot is located at (x1, y1) in the original coordinate system, while the corresponding temperature anomaly hotspot is located at (x2, y1) in the infrared image. After translation, rotation, and scaling corrections using an affine transformation matrix, the two are precisely aligned to the same or adjacent positions in a unified grid coordinate system. Based on this, geometric consistency correction is further performed based on the unified coordinate feature map. A deformation matching algorithm based on the foot contour template or a non-rigid registration technique is used to eliminate local misalignments caused by foot deformation, sensor fit differences, or gait posture changes. This ensures that the pressure and temperature features maintain morphological consistency in the anatomical structure, ultimately generating a corrected multimodal distribution map with geometrically aligned features and coexisting dual-modal information. This makes the spatial correspondence between the high-pressure area and the high-temperature area more accurate. The process then proceeds to the mapping module, which maps the plantar region with a feature vector field based on the corrected multimodal distribution map. It combines the pressure and temperature values ​​at each spatial location into a two-dimensional or multi-dimensional feature vector and introduces directional weights (such as pressure gradient direction and temperature change trend) to construct a directional fusion feature vector field. This vector field not only reflects the magnitude of the physical quantities at each point but also encodes their spatial evolution trend. For example, in the forefoot region, the pressure vector points towards the distal metatarsal head, while the temperature vector diffuses along the blood flow direction; together, they constitute a composite feature field. Based on this, local density clustering is performed using the fusion feature vector field. Algorithms such as DBSCAN or Mean Shift are used to identify high-risk clusters densely distributed in both the feature space and geographic space. Regions simultaneously exhibiting high pressure, high temperature, and consistent feature vector directions are clustered into several clustered multimodal blocks. These blocks represent the potential core damage areas resulting from the synergistic effect of pressure and thermal anomalies. For instance, a patient repeatedly exhibits a clustered multimodal block located below the third metatarsal head during multiple walks, suggesting that this area has been subjected to a long-term dual pathological load.Next, the second fusion module continues the processing. This module performs temporal correlation analysis on the clustered multimodal blocks, extracting their frequency of occurrence, duration, and trend during continuous gait cycles or long-term monitoring. It calculates the temporal synchronicity of pressure peaks and temperature increases (such as the Pearson correlation coefficient), identifies regions that are highly co-evolving in the time dimension, and generates a temporal co-evolution feature map reflecting dynamic coupling relationships. For example, if a block experiences a sudden rise in pressure and a rise in temperature simultaneously in more than 90% of walking cycles, it is identified as a strong temporal co-evolution zone. Based on this, weighted feature fusion is performed on the temporal co-evolution feature map. Different weights are assigned to each block according to its spatial overlap, temporal correlation strength, and physiological importance. Multidimensional indicators are fused using linear weighting or fuzzy integral methods to generate a fusion risk feature area with a high comprehensive score, highlighting the complex lesion areas with the highest ulcer risk. Finally, the positioning module completes the final output. This module optimizes and integrates the foot region boundary through the fusion risk feature area, and uses active contour model or morphological closing operation to smooth and close the edge of the fusion area, eliminating isolated noise points and gaps, forming a continuous and complete optimized risk overlap area. Based on this, the risk area is located based on the optimized risk overlap area, and its specific anatomical location and coverage are determined by combining the foot anatomical partition atlas (such as the Rothbart partition or Manchester classification system). Finally, the precisely labeled foot risk overlap area is output, for example, clearly indicating "1.5cm below the first metatarsal head of the right foot". 2 The area is designated as a "high-risk overlapping zone," providing a reliable spatial basis for subsequent cumulative calculations and early warning decisions. The entire registration process, through the hierarchical processing of the second clustering module, mapping module, second fusion module, and localization module, achieves deep fusion of pressure and temperature abnormality areas in spatial, temporal, and feature dimensions, ensuring the accuracy and clinical interpretability of the foot risk overlapping zone.

[0026] In a specific embodiment, a time-series cumulative calculation is performed on the overlapping risk area of ​​the foot to obtain a cumulative pressure heatmap, including: The accumulation module is used to perform time-weighted accumulation of the overlapping risk areas of the foot to obtain a time-series risk accumulation field, and to extract dynamic trends based on the time-series risk accumulation field to obtain a risk evolution trend map. The smoothing module is used to segment the plantar region into trend clusters based on the risk evolution trend map to obtain the trend cluster segmentation area, and to perform time-series gradient smoothing through the trend cluster segmentation area to obtain the smoothed risk trend area. The third fusion module is used to perform multi-dimensional time-series feature fusion on the smoothed risk trend area to obtain a fusion trend feature map, and to perform risk accumulation mapping based on the fusion trend feature map to obtain a cumulative pressure heatmap.

[0027] Specifically, the process of performing time-series cumulative calculations on the foot risk overlap areas to obtain a cumulative pressure heatmap is first executed by the accumulation module. This module performs time-weighted accumulation on the foot risk overlap areas, superimposing the foot risk overlap areas identified over multiple consecutive days or gait cycles frame by frame in a unified spatial coordinate system. The risk area at each time point is assigned a corresponding time weight based on its duration, frequency of occurrence, and physiological load intensity. For example, for diabetic patients, if a certain area repeatedly shows abnormal overlap of pressure and temperature during daily walking activities, its cumulative weight increases with the number of days, thereby generating a heatmap reflecting long-term exposure. The temporal risk accumulation field under dual pathological stimulation records the cumulative risk exposure at each location with pixel-level intensity values. Based on this, dynamic trend extraction is further performed on the temporal risk accumulation field. Sliding window regression analysis or wavelet transform methods are used to model the trend of the risk value sequence at each spatial location, identify the rising, falling or stable change patterns, and encode the slope information into a spatial distribution map to generate a risk evolution trend map. For example, the area below the fourth metatarsal head of a patient's left foot showed a continuous positive growth trend during seven consecutive days of monitoring, indicating that the risk of tissue damage in this area is gradually increasing. This change feature is clearly mapped onto the risk evolution trend map. The process then proceeds to the smoothing module, which segments the plantar region into trend clusters based on the risk evolution trend map. Clustering algorithms such as K-means or spectral clustering are used to divide spatially adjacent regions with similar trends (e.g., both showing rapid increases or slow fluctuations) into independent trend cluster segments. This achieves functional division of regions with different mechanical evolution behaviors of the plantar surface; for example, the heel, forefoot, and arch can be assigned to different trend clusters to distinguish their stress evolution characteristics. Based on this, temporal gradient smoothing is performed on the trend cluster segments. Low-pass filtering or Savitzky-Golay filtering is applied to the time-series data within each trend cluster to suppress non-physiological abrupt changes caused by short-term activity fluctuations or measurement noise, while preserving long-term stable load evolution characteristics. This results in a smoothed risk trend region with clear boundaries and continuous changes, ensuring that subsequent analysis is not affected by occasional interference.Next, the third fusion module continues the fusion operation. This module performs multi-dimensional time-series feature fusion on the smoothed risk trend area, integrating various time-series features including total risk accumulation, rate of change, fluctuation amplitude, and peak frequency. It constructs a high-dimensional feature vector and uses dimensionality reduction techniques such as principal component analysis (PCA) or autoencoders for effective compression and fusion, generating a fusion trend feature map that comprehensively expresses the multi-dimensional characteristics of risk development. This map not only reflects the current risk level but also contains complex structural information about the development trend. For example, in a diabetic patient with peripheral neuropathy, although the current cumulative value of a certain area on the sole of the forefoot is not high, the fusion trend feature map shows that its growth acceleration is significantly higher than that of other areas, indicating a potential risk of accelerated deterioration. Based on this, risk accumulation mapping is performed based on the fusion trend feature map, and the fused multi-dimensional features are reprojected back onto the sole spatial grid. Each pixel is assigned a value according to the comprehensive score to form the final cumulative pressure heatmap. This heatmap uses color gradients to intuitively display the comprehensive risk load accumulation of each area on the sole over a long time scale, with high-intensity areas corresponding to the location of chronic high-risk lesions. The entire calculation process, through the coordinated operation of the accumulation module, the smoothing module and the third fusion module, realizes the transformation from discrete foot risk overlap areas to continuous, stable and predictive cumulative pressure heat maps, providing a solid data foundation for subsequent morphological analysis and risk level determination, and supporting the system to accurately capture and provide early warning of the precursors of diabetic foot ulcers.

[0028] In a specific embodiment, morphological analysis and gradient calculation are performed based on the cumulative pressure heatmap to obtain the geometric features of the risk area and the pressure change rate features, including: The denoising module is used to perform noise filtering and edge enhancement processing on the cumulative pressure heatmap using a multi-scale morphological dilation-erosion operator to obtain a denoised and enhanced pressure heatmap, and to perform adaptive threshold segmentation based on the denoised and enhanced pressure heatmap to obtain an initial risk region mask. The second extraction module is used to perform contour extraction and polygon approximation processing based on the initial risk region mask to obtain the risk region approximation contour, and to calculate the geometric parameters of the risk region approximation contour to obtain the geometric features of the risk region. The decomposition module is used to perform dynamic gradient extraction on the cumulative pressure heatmap through time-series differential gradient field calculation to obtain a time-series gradient distribution field, and to perform rate vector decomposition based on the time-series gradient distribution field to obtain a rate vector feature map. The quantization module is used to perform time-series rate clustering on the plantar region based on the rate vector feature map to obtain a clustered rate feature region, and to perform weighted rate quantization on the clustered rate feature region to obtain the pressure change rate feature.

[0029] Specifically, morphological analysis and gradient calculation based on the accumulated pressure heatmap are first performed by the denoising module. This module uses a multi-scale morphological dilation-erosion operator to filter noise and enhance edges in the accumulated pressure heatmap. Specifically, for each pixel in the accumulated pressure heatmap, dilation and erosion operations are applied sequentially according to multiple preset scale parameters to remove isolated small noise points and enhance the boundaries of real risk areas. For example, in a diabetic foot monitoring system, by iteratively processing the accumulated pressure data over different time spans using gradually increasing structuring elements, a denoised and enhanced pressure heatmap that clearly shows the boundaries of high-risk areas is finally generated. Then, adaptive threshold segmentation is performed based on the denoised and enhanced pressure heatmap. The optimal segmentation threshold is automatically determined using the Otsu method or other dynamic thresholding algorithms, and the image is binarized to obtain an initial risk area mask, ensuring accurate capture of all potential risk areas. For example, for a patient who has been affected by uneven pressure for a long time, some parts of their sole may form new high-risk points due to small pressure changes, and these points can be accurately identified. Next, the second extraction module is used for processing. This module performs contour extraction and polygon approximation based on the initial risk area mask. It uses techniques such as chain code or Freeman coding to track and record coordinates pixel by pixel along the risk area boundary. Then, the Douglas-Peucker algorithm is used to simplify the complex contour into an approximate polygon, thereby obtaining the risk area approximation contour. This process can effectively describe the approximate shape characteristics of each risk area. For example, if there is an elliptical high-risk area at the heel of a patient, after approximation, its major axis, minor axis length, and orientation angle can be obtained. Then, the geometric parameters of the risk area approximation contour are calculated, including indicators such as area, perimeter, and roundness, to comprehensively quantify the spatial characteristics of each risk area. The decomposition module performs dynamic gradient extraction on the cumulative pressure heatmap through temporal differential gradient field calculation. First, it calculates the difference between adjacent time frames to construct a temporal differential gradient field, capturing the changing trend of pressure distribution over time. For example, during several days of gait monitoring, if the pressure value in a specific area shows an upward trend, the corresponding position in the temporal differential gradient field will show a positive gradient value, and vice versa. This yields a temporal gradient distribution field reflecting the dynamic evolution of pressure in various parts of the foot. Based on this, a velocity vector decomposition is performed on the temporal gradient distribution field, that is, the two-dimensional gradient field is converted into a velocity vector field, and the velocity components in the x-axis and y-axis directions are calculated respectively to form a velocity vector feature map. This process helps to deeply understand the directionality and intensity of local pressure changes. For example, during training, the pressure change rate in the forefoot area of ​​an athlete is significantly higher than that in other parts, indicating that this area has been subjected to a greater instantaneous load impact.Finally, the quantization module performs temporal rate clustering on the plantar region based on the rate vector feature map. Unsupervised learning algorithms such as K-means clustering or DBSCAN are used to group spatial locations with similar rate patterns into one class, forming clustered rate feature regions. This step can identify which regions exhibit consistent pressure change behavior, such as a group of intertoe regions that frequently experience simultaneous high-pressure peaks. Further weighted rate quantization is performed through these clustered rate feature regions, assigning corresponding weight coefficients based on the magnitude and direction of the rate vector within each region. This comprehensively evaluates the pressure change rate characteristics of the entire plantar region, allowing doctors or researchers to intuitively see the rate of pressure increase or decrease in different areas over long periods, providing strong support for developing personalized treatment plans. The entire process starts from the original cumulative pressure heatmap and, through multiple stages including denoising, edge enhancement, adaptive segmentation, contour extraction and approximation, gradient extraction, rate vector decomposition, and rate clustering quantization, gradually delves into the geometric characteristics and pressure change patterns of risk areas in the foot, providing detailed data support for clinical diagnosis and preventative measures.

[0030] In a specific embodiment, a foot risk level is generated based on the geometric features of the risk area and the pressure change rate features, including: The transformation module is used to perform a coordinate system transformation on the geometric features of the risk area and the pressure change rate features to obtain a plantar fusion risk vector field, and to perform local density gradient aggregation based on the plantar fusion risk vector field to obtain an aggregated risk density distribution. The second tracking module is used to perform spatiotemporal propagation tracking of peak clusters in the aggregated risk density distribution to obtain a dynamic map of propagation risk clusters, and to perform multi-level threshold vector quantization based on the dynamic map of propagation risk clusters to obtain the foot risk level.

[0031] Specifically, the process of generating foot risk levels based on the geometric features of the risk area and the pressure change rate features is first executed by the transformation module. This module performs a coordinate system transformation on the geometric features of the risk area and the pressure change rate features. Since the geometric features of the risk area originate from contour parameters obtained from morphological analysis (such as area, perimeter, and roundness), while the pressure change rate features come from the rate vector map extracted from gradient calculation, the two initially exist in different data spaces and coordinate systems. Therefore, they need to be unified to the standard foot anatomy grid coordinate system through spatial mapping and normalization to ensure that the geometric and dynamic attributes at each spatial location are accurately aligned. For example, in the long-term monitoring of diabetic patients, if a certain area shows both a large risk area and a high pressure increase rate, the two types of features at that location will be fused at the same pixel. On this basis, local density gradient aggregation is further performed based on the fused foot risk vector field, that is, the geometric features (such as regional irregularity) and rate features (such as pressure change slope) of each spatial point are combined into a multi-dimensional vector, and density estimation is performed by combining the similarity of other points in its neighborhood. Mean The Shift or kernel density estimation algorithm identifies high-density clusters, reflecting the core lesion area where multiple risk factors converge on the sole of the foot, thereby generating an aggregated risk density distribution. This distribution presents the density of comprehensive risk in various parts of the sole in grayscale or pseudo-color. For example, a patient has a large area of ​​abnormal morphology below the first metatarsal head on the forefoot, and multiple sampling points around it show a rapid increase in pressure. After local density gradient aggregation, a significantly bright aggregated peak is formed, indicating that this area is the current highest-risk site.The process then proceeds to the second tracking module, which tracks the spatiotemporal propagation of peak clusters in the aggregated risk density distribution. Using density distribution maps in a continuous time series, it identifies risk density peaks exceeding a set threshold in each frame. Through minimum Euclidean distance matching or Kalman filter prediction methods, it connects peaks at adjacent time points, constructing a propagation path of the risk cluster over time, forming a dynamic map of the propagating risk clusters. This map reveals whether high-risk areas migrate, spread, or remain stable on the sole of the foot. For example, in a patient with diabetic neuropathy, the risk density peak, initially concentrated at the third metatarsal head, gradually extended towards the distal fourth metatarsal over three days. The dynamic map of the propagating risk clusters clearly records this "hotspot." The "migration" process indicates that tissue damage is expanding along a specific mechanical path. Based on this, multi-level threshold vector quantification is performed using the dynamic map of the propagation risk clusters. Multiple risk level classification thresholds (e.g., low, medium, high, and extremely high) are set. A hierarchical vector space is constructed by combining multiple dimensions such as risk density intensity, propagation speed, coverage area growth rate, and duration. Each tracked risk cluster is projected into this space, and its similarity to the baseline vectors of each level is calculated. Finally, the foot risk level to which it belongs is determined. For example, when the density value of a risk cluster exceeds the 80th percentile, the propagation speed is greater than 0.5 mm / day, and it persists for more than five consecutive days, the system determines it to have reached the "high risk" level and marks it as an object requiring immediate intervention. The entire process, through the synergy of the transformation module and the second tracking module, achieves a complete transformation from static geometric and dynamic rate characteristics to interpretable and hierarchical foot risk levels. It not only considers the spatial distribution characteristics of the current risk but also incorporates its temporal evolution trend, improving the foresight and clinical applicability of risk assessment and supporting the system's accurate identification and hierarchical early warning of the precursors to diabetic foot ulcers.

[0032] In a specific embodiment, peak cluster spatiotemporal propagation tracking is performed on the aggregated risk density distribution to obtain a dynamic graph of propagation risk clusters, including: The discretization module is used to perform grid discretization processing on the aggregated risk density distribution to obtain a discretized risk density grid, and to perform local extreme point detection based on the discretized risk density grid to obtain an initial risk peak point set. The correlation analysis module is used to perform spatiotemporal correlation analysis based on the initial risk peak point set, construct the spatiotemporal correlation matrix of the peak points, obtain the associated risk peak cluster, and perform dynamic diffusion range calculation on the associated risk peak cluster to obtain the peak cluster diffusion boundary. The drawing module is used to draw the time-series evolution trajectory based on the diffusion boundary of the peak cluster, to obtain the time-series trajectory map of the risk peak cluster, and to extract dynamic features from the time-series trajectory map of the risk peak cluster to obtain the dynamic map of the propagation risk cluster.

[0033] Specifically, the process of tracking the spatiotemporal propagation of peak clusters in the aggregated risk density distribution to obtain a dynamic map of propagating risk clusters is first executed by the discretization module. This module performs grid discretization processing on the aggregated risk density distribution, mapping the continuous plantar risk density field to a regular spatial grid of a preset resolution. Each grid cell corresponds to a fixed position (e.g., 5mm × 5mm) in the plantar anatomical region and is assigned the average or maximum risk density value of its region, thereby generating a structured discretized risk density grid. This grid provides a unified data foundation for subsequent accurate spatial calculations. For example, in plantar monitoring of diabetic patients, the forefoot region is divided into several... The system uses a grid of cells, where multiple cells located below the first and second metatarsal heads exhibit values ​​significantly higher than those in the surrounding areas. Further, based on the discretized risk density grid, local extreme points are detected. A sliding window method is used to compare the values ​​within each cell and its eight neighboring areas, identifying the local maximum points that are spatially higher than all adjacent points. These points, along with their coordinates and corresponding density values, form an initial set of risk peak points. These points represent the core locations with the highest risk concentration in various parts of the foot at the current moment. For example, in a patient monitored for three consecutive days, the cell directly below the third metatarsal head was consistently identified as a local extreme point, indicating that this area is a long-term high-load center. The process then proceeds to the correlation analysis module, which performs spatiotemporal correlation analysis based on the initial risk peak point set. Using the initial risk peak point set collected at multiple consecutive time points (e.g., once daily), a spatiotemporal correlation matrix of peak points, containing both time and spatial coordinates, is constructed. This matrix records the position, intensity, and evolution path of all peak points in each time frame. A minimum distance matching algorithm is used to pair the peak point from the previous time moment with the nearest neighbor peak point of similar density from the next time moment. If the distance between two points is less than a preset threshold (e.g., 2 mm) and the density change rate is below a certain range, they are determined to be a continuation of the same risk cluster; otherwise, they are considered newly emerging. Independent events, or interruptions, form a cluster of associated risk peaks with temporal continuity. For example, a patient's initial risk peak point is located at the proximal end of the fourth metatarsal head on day 1, migrates to the middle on day 2, and continues to move forward to the distal end on day 3. The system confirms through spatiotemporal correlation analysis that these three points belong to the same risk cluster in the same propagation. Based on this, the dynamic diffusion range of the associated risk peak cluster is calculated. According to the risk density decay gradient of the area adjacent to the cluster at each time point, the influence boundary is determined by the contour encirclement method or the Gaussian half-width expansion method, generating a peak cluster diffusion boundary that changes over time, reflecting the expansion or contraction trend of the risk influence range.Next, the visualization and feature extraction module completes the visualization and feature extraction. This module draws the temporal evolution trajectory based on the diffusion boundary of the peak cluster, superimposing the diffusion boundary of each time point onto the same plantar anatomical background image in chronological order. Different colors or transparency are used to distinguish the coverage areas of different dates, forming a temporal trajectory map of risk peak clusters that continuously displays the spatial evolution of risk. For example, the process of a certain risk cluster gradually spreading from the medial side of the plantar surface towards the fifth metatarsal head can be clearly seen in the image. Based on this, dynamic feature extraction is performed on the temporal trajectory map of the risk peak clusters, calculating time-varying parameters such as its movement speed, direction angle, area growth rate, boundary irregularity, and overlap ratio with adjacent anatomical areas. Finally, these are integrated into a dynamic map of the propagation risk cluster that includes spatial expansion patterns and temporal evolution laws. This map not only shows "where" the risk is, but also reveals "how" it develops and "where it spreads," providing a key basis for subsequent multi-level risk quantification. The entire process, through the step-by-step execution of the discretization processing module, correlation analysis module, and drawing module, achieves complete modeling from static aggregated risk density to dynamic propagation behavior, supporting the foresight and accuracy of diabetic foot ulcer risk level determination.

[0034] In a specific embodiment, a time-series evolution trajectory is plotted based on the diffusion boundary of the peak cluster to obtain a time-series trajectory map of the risk peak cluster, including: The interpolation calculation module is used to extract key frame coordinates from the peak cluster diffusion boundary to obtain a boundary key frame coordinate set, and to perform neighborhood frame interpolation calculation based on the boundary key frame coordinate set to obtain interpolated and completed boundary coordinates. The fitting module is used to fit the boundary contour based on the interpolated and completed boundary coordinates to obtain a continuous boundary contour curve, and to detect curvature abrupt change points on the continuous boundary contour curve to obtain a set of boundary feature abrupt change points. An anchoring module is used to anchor the time-series trajectory based on the set of boundary feature mutation points, obtain the time-series trajectory line of the feature points, and stitch the overall contour of the time-series trajectory line of the feature points to obtain the time-series trajectory map of the risk peak cluster.

[0035] Specifically, the process of drawing the temporal evolution trajectory based on the peak cluster diffusion boundary to obtain the risk peak cluster temporal trajectory map is first executed by the interpolation calculation module. This module extracts the key frame coordinates of the peak cluster diffusion boundary and obtains the set of contour coordinates of the calculated peak cluster diffusion boundary at each time step from discrete time monitoring points (such as a daily plantar scan) to form a boundary key frame coordinate set. These coordinate points accurately mark the spatial range of the high-risk area at the current moment. For example, in the monitoring of diabetic patients for five consecutive days, the system extracts the diffusion boundary coordinates on the 1st, 3rd and 5th days respectively to characterize the expansion state of the risk in the forefoot region. Due to the limited monitoring frequency, the original boundary exhibits discrete jump characteristics in the time dimension, failing to reflect the continuous evolution of the risk. Therefore, further neighbor frame interpolation calculations are performed based on the boundary key frame coordinate set. Linear interpolation, spline interpolation, or deformation interpolation methods based on optical flow estimation are used to generate virtual boundary coordinates at several intermediate time points between two adjacent key frames, filling the time gaps and achieving a smooth transition of the boundary morphology. This results in interpolated and complete boundary coordinates. For example, the diffusion boundary of day 2 is inserted between day 1 and day 3, making the expansion process of the risk area from a smaller range to a larger range present a continuous and gradual effect, avoiding the abrupt artifacts caused by the sampling interval. The process then proceeds to the fitting module, which performs boundary contour fitting based on the interpolated boundary coordinates. Using least squares, B-spline curve fitting, or an active contour model, the discrete interpolated coordinate points are connected into a closed and smooth continuous boundary contour curve. This curve more realistically reflects the geometry of high-risk areas on the sole of the foot, eliminating jagged edges caused by sensor noise or interpolation errors, and improving visualization quality and subsequent analysis accuracy. For example, the risk area below the fifth metatarsal head of a patient shows a regular elliptical contour after fitting, rather than the original set of points. Irregular polygons; based on this, curvature abrupt change points are detected on the continuous boundary contour curve. By calculating the curvature value (i.e., the rate of change in the direction of unit arc length) of each point on the curve, local peak points with curvature significantly higher than those in the adjacent area are identified. These points usually correspond to sharp corners, depressions, or protrusions on the boundary, forming a set of boundary feature abrupt change points. For example, in a diabetic patient with foot deformity, the boundary of the risk area shows a significant depression on the inner side of the arch of the foot. The curvature value at this location is much higher than that in other areas, and it is accurately identified as a feature abrupt change point, indicating that the tissue is abnormally stressed at this location.Next, the anchoring module continues trajectory construction. This module performs temporal trajectory anchoring based on the set of boundary feature mutation points. It matches and connects feature mutation points with the same geometric semantics at different time points (such as all located at the foremost or innermost edge of the risk area) in chronological order. It establishes cross-frame correspondence using the spatial nearest neighbor principle or morphological similarity measurement and generates a series of feature point temporal trajectory lines that span multiple time points. Each trajectory line represents the movement path of a specific boundary feature in the time dimension. For example, if the farthest endpoint of a patient's forefoot plantar risk area moves anterolaterally for five consecutive days, the corresponding feature point temporal trajectory line will show a clear path. A clear, slanted extension path reflects the risk spreading towards the interdigital region. Based on this, the temporal trajectory lines of the feature points are stitched together as a whole, and the continuous boundary contour curves at all time points are superimposed onto the same plantar anatomical background image in chronological order. Gradient color coding or transparency overlay methods are used to distinguish the boundaries of different time periods, ultimately forming a complete temporal trajectory map of risk peak clusters that fully demonstrates the morphological evolution of the risk area. This map not only presents the overall migration trend of the risk area but also preserves the dynamic changes in its boundary morphology. For example, it allows for a direct observation of a high-risk area gradually evolving from a circle into an irregular polygon with boundary bifurcation. The entire process, through the collaborative processing of interpolation, fitting, and anchoring modules, achieves the construction of an evolutionary trajectory map from discrete peak cluster diffusion boundaries to a continuous, smooth, and temporally logical one, providing high-precision spatiotemporal visualization support for the dynamic monitoring and clinical intervention of diabetic foot ulcer risk.

Claims

1. A smart early warning system for diabetic foot ulcer risk based on multimodal sensing, characterized in that, Includes the following steps: The data acquisition module is used to collect pressure and temperature data of the user's feet during walking, and obtain dynamic plantar pressure images and plantar surface temperature maps. The segmentation module is used to perform image segmentation processing based on the dynamic plantar pressure image to obtain the pressure concentration area, and to perform threshold analysis processing based on the plantar surface temperature spectrum to obtain the temperature anomaly hotspots. The registration module is used to spatially register the pressure concentration area with the temperature anomaly hot spot to obtain the foot risk overlap area, and to perform time-series cumulative calculation on the foot risk overlap area to obtain a cumulative pressure heat map. The calculation module is used to perform morphological analysis and gradient calculation based on the cumulative pressure heat map to obtain the geometric features of the risk area and the pressure change rate features. The generation module is used to generate a foot risk level based on the geometric features of the risk area and the pressure change rate features, and to match the foot risk level with preset warning rules to obtain a risk warning instruction.

2. The intelligent early warning system for diabetic foot ulcer risk based on multimodal sensing according to claim 1, characterized in that, Spatial registration is performed between the pressure concentration area and the temperature anomaly hotspot to obtain the foot risk overlap area, including: The second clustering module is used to perform spatial coordinate transformation on the pressure concentration area and the temperature anomaly hotspot to obtain a unified coordinate feature map, and to perform geometric consistency correction based on the unified coordinate feature map to obtain a corrected multimodal distribution map. The mapping module is used to map the feature vector field of the plantar region based on the corrected multimodal distribution map to obtain a fused feature vector field, and to perform local density clustering through the fused feature vector field to obtain clustered multimodal blocks; The second fusion module is used to perform temporal correlation analysis on the clustered multimodal blocks to obtain a temporal collaborative feature map, and to perform weighted feature fusion based on the temporal collaborative feature map to obtain a fusion risk feature region. The positioning module is used to optimize and integrate the boundaries of the sole area through the fused risk feature area to obtain an optimized risk overlap area, and to locate the risk area based on the optimized risk overlap area to obtain the foot risk overlap area.

3. The intelligent early warning system for diabetic foot ulcer risk based on multimodal sensing according to claim 1, characterized in that, A time-series cumulative calculation is performed on the aforementioned foot risk overlap area to obtain a cumulative pressure heatmap, including: The accumulation module is used to perform time-weighted accumulation of the overlapping risk areas of the foot to obtain a time-series risk accumulation field, and to extract dynamic trends based on the time-series risk accumulation field to obtain a risk evolution trend map. The smoothing module is used to segment the plantar region into trend clusters based on the risk evolution trend map to obtain the trend cluster segmentation area, and to perform time-series gradient smoothing through the trend cluster segmentation area to obtain the smoothed risk trend area. The third fusion module is used to perform multi-dimensional time-series feature fusion on the smoothed risk trend area to obtain a fusion trend feature map, and to perform risk accumulation mapping based on the fusion trend feature map to obtain a cumulative pressure heatmap.

4. The intelligent early warning system for diabetic foot ulcer risk based on multimodal sensing according to claim 1, characterized in that, Based on the cumulative pressure heatmap, morphological analysis and gradient calculation are performed to obtain the geometric features of the risk area and the pressure change rate characteristics, including: The denoising module is used to perform noise filtering and edge enhancement processing on the cumulative pressure heatmap using a multi-scale morphological dilation-erosion operator to obtain a denoised and enhanced pressure heatmap, and to perform adaptive threshold segmentation based on the denoised and enhanced pressure heatmap to obtain an initial risk region mask. The second extraction module is used to perform contour extraction and polygon approximation processing based on the initial risk region mask to obtain the risk region approximation contour, and to calculate the geometric parameters of the risk region approximation contour to obtain the geometric features of the risk region. The decomposition module is used to perform dynamic gradient extraction on the cumulative pressure heatmap through time-series differential gradient field calculation to obtain a time-series gradient distribution field, and to perform rate vector decomposition based on the time-series gradient distribution field to obtain a rate vector feature map. The quantization module is used to perform time-series rate clustering on the plantar region based on the rate vector feature map to obtain a clustered rate feature region, and to perform weighted rate quantization on the clustered rate feature region to obtain the pressure change rate feature.

5. The intelligent early warning system for diabetic foot ulcer risk based on multimodal sensing according to claim 1, characterized in that, A foot risk level is generated based on the geometric features of the risk area and the pressure change rate features, including: The transformation module is used to perform a coordinate system transformation on the geometric features of the risk area and the pressure change rate features to obtain a plantar fusion risk vector field, and to perform local density gradient aggregation based on the plantar fusion risk vector field to obtain an aggregated risk density distribution. The second tracking module is used to perform spatiotemporal propagation tracking of peak clusters in the aggregated risk density distribution to obtain a dynamic map of propagation risk clusters, and to perform multi-level threshold vector quantization based on the dynamic map of propagation risk clusters to obtain the foot risk level.

6. The intelligent early warning system for diabetic foot ulcer risk based on multimodal sensing according to claim 5, characterized in that, The peak cluster spatiotemporal propagation of the aggregated risk density distribution is tracked to obtain a dynamic graph of the propagation risk clusters, including: The discretization module is used to perform grid discretization processing on the aggregated risk density distribution to obtain a discretized risk density grid, and to perform local extreme point detection based on the discretized risk density grid to obtain an initial risk peak point set. The correlation analysis module is used to perform spatiotemporal correlation analysis based on the initial risk peak point set, construct the spatiotemporal correlation matrix of the peak points, obtain the associated risk peak cluster, and perform dynamic diffusion range calculation on the associated risk peak cluster to obtain the peak cluster diffusion boundary. The drawing module is used to draw the time-series evolution trajectory based on the diffusion boundary of the peak cluster, to obtain the time-series trajectory map of the risk peak cluster, and to extract dynamic features from the time-series trajectory map of the risk peak cluster to obtain the dynamic map of the propagation risk cluster.

7. The intelligent early warning system for diabetic foot ulcer risk based on multimodal sensing according to claim 6, characterized in that, Based on the diffusion boundary of the peak cluster, a time-series evolution trajectory is plotted to obtain a time-series trajectory map of the risk peak cluster, including: The interpolation calculation module is used to extract key frame coordinates from the peak cluster diffusion boundary to obtain a boundary key frame coordinate set, and to perform neighborhood frame interpolation calculation based on the boundary key frame coordinate set to obtain interpolated and completed boundary coordinates. The fitting module is used to fit the boundary contour based on the interpolated and completed boundary coordinates to obtain a continuous boundary contour curve, and to detect curvature abrupt change points on the continuous boundary contour curve to obtain a set of boundary feature abrupt change points. An anchoring module is used to anchor the time-series trajectory based on the set of boundary feature mutation points, obtain the time-series trajectory line of the feature points, and stitch the overall contour of the time-series trajectory line of the feature points to obtain the time-series trajectory map of the risk peak cluster.

Images