Foot arch index calculation method, electronic equipment, storage medium and program product

Abstract

The invention discloses a foot arch index calculation method, electronic equipment and a storage medium, and relates to the field of biomedical engineering. The method comprises the following steps: acquiring sole three-dimensional point cloud data; pre-analyzing the data by using a first deep learning model to generate situation metadata representing scanning quality and foot physiological features; dynamically configuring calculation parameters of a parameterized geometric algorithm model according to the situation metadata, and calculating to obtain a first arch index and confidence thereof; meanwhile, a second deep learning model is used for regression to obtain a second arch index and the confidence of the second arch index; calculating a consistency coefficient of the two results; and when the confidence coefficients of the two are higher than the threshold value and the consistency coefficient meets the requirement, fusing and outputting a final arch index. According to the method, through double-algorithm path cross validation and conditional fusion, the interpretability of a geometric algorithm and the adaptability of a deep learning algorithm are integrated, the influence of individual differences and scanning condition changes is overcome, and the accuracy and stability of arch index calculation are improved.

Description

Technical Field

[0001] This invention relates to the field of biomedical engineering technology, and in particular to a method for calculating arch index, an electronic device, a storage medium, and a program product. Background Technology

[0002] The Arch Index (AI) is a core indicator used in foot examinations and retail applications to quantitatively assess arch shape. It reflects the arch condition by the ratio of the midfoot area to the total contact area, and is an important basis for assessing foot morphological characteristics, custom footwear design, and athletic performance analysis.

[0003] In related technologies, the arch index is mainly calculated using digital methods. A common calculation method involves acquiring plantar point cloud data using a 3D scanner and then applying a geometric algorithm with fixed parameters (such as the fixed slice area method). This method typically relies on the precise identification of specific anatomical landmarks in the plantar point cloud (such as the calcaneal point, metatarsal point, and navicular point), followed by calculation using predefined geometric formulas (such as angles, the ratio of height to length, etc.).

[0004] However, the accuracy of the above methods depends heavily on the quality of point cloud data and the accuracy of marker positioning. Under different individual foot types or scanning conditions, the stability and accuracy of the calculation results can no longer meet higher requirements. Summary of the Invention

[0005] To address the aforementioned technical problems and deficiencies, the purpose of this invention is to provide a method, electronic device, storage medium, and program product for calculating the arch index, which can alleviate the current problems of poor accuracy and stability in arch index calculation.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for calculating the arch index, comprising: acquiring three-dimensional point cloud data of the sole of a subject using a three-dimensional foot scanner; inputting the three-dimensional point cloud data of the sole into a preset first deep learning model for pre-analysis to generate contextual metadata, wherein the contextual metadata is structured data characterizing the scanning quality of the three-dimensional point cloud data of the sole itself and the physiological characteristics of the foot reflected therein, and the first deep learning model is a meta-analysis model used for quality assessment and feature classification of the input data; dynamically configuring the calculation parameters in a preset parametric geometric algorithm model according to the contextual metadata to obtain the configured parametric geometric algorithm model; and performing three-dimensional point cloud data analysis on the sole using the configured parametric geometric algorithm model. Point cloud data is processed to obtain the first arch index result, and the geometric calculation confidence of the first arch index result is calculated. The three-dimensional point cloud data of the foot is input into a preset second deep learning model to generate the second arch index result and the corresponding model inference confidence. The second deep learning model is a regression prediction model used to directly map the input data to the target arch index value. The consistency coefficient between the first arch index result and the second arch index result is calculated. When both the geometric calculation confidence and the model inference confidence are greater than the preset confidence threshold, and the consistency coefficient is greater than or equal to the preset consistency threshold, the first arch index result and the second arch index result are fused to obtain the subject's final arch index.

[0007] This invention employs the aforementioned method for calculating the arch index, addressing the issues of insufficient accuracy and stability caused by fixed-parameter geometric algorithms. By introducing a dual-path parallel computing architecture, this invention combines a meta-analysis-based dynamically configured parametric geometric algorithm with an end-to-end deep learning regression model. The first deep learning model pre-analyzes the input point cloud, generating contextual metadata to dynamically optimize the geometric algorithm's calculation parameters, thus overcoming the problem of poor adaptability to different individual foot types and scanning conditions inherent in traditional methods. Simultaneously, the second deep learning model provides an independent regression result for cross-validation. Finally, a rigorous multi-dimensional evaluation of the confidence and consistency of the two results determines whether weighted fusion is performed, ensuring that the final arch index is output only under high-confidence conditions. This mechanism significantly improves the accuracy, robustness, and reliability of the calculation results in complex scenarios.

[0008] In some implementations, computational parameters in a preset parametric geometric algorithm model are dynamically configured based on contextual metadata. This includes: extracting scan quality indicators characterizing scan noise levels and foot physiological feature indicators characterizing foot pronation and supination angles from the contextual metadata; adaptively determining the Gaussian filter kernel radius for point cloud smoothing by querying a preset nonlinear mapping function based on the scan quality indicators, wherein a relatively lower scan quality indicator corresponds to a relatively larger Gaussian filter kernel radius; constructing an affine transformation matrix for rotating the three-dimensional point cloud data of the sole from the original scan coordinate system to a target coordinate system aligned with the principal axis of foot anatomy based on the foot physiological feature indicators; smoothing the three-dimensional point cloud data of the sole based on the Gaussian filter kernel radius to obtain smoothed point cloud data, and performing coordinate transformation on the smoothed point cloud data based on the affine transformation matrix to generate the final input data to be processed by the configured parametric geometric algorithm model.

[0009] By employing the above technical solution, intelligent and personalized preprocessing of input data is achieved by extracting scanning noise and foot posture indices from contextual metadata and adaptively configuring the point cloud smoothing filter intensity and coordinate transformation matrix. This dynamic configuration not only effectively suppresses the interference of scanning noise on subsequent calculations but also eliminates systematic errors introduced by differences in the subject's standing posture through posture normalization, providing high-quality and standardized input for subsequent geometric algorithms, thereby significantly improving the calculation accuracy and stability of the first arch index result.

[0010] In some implementations, the three-dimensional point cloud data of the foot is processed by a configured parametric geometric algorithm model to obtain a first arch index result, including: generating posture-normalized point cloud data based on the three-dimensional point cloud data of the foot and an affine transformation matrix; determining the search range of the navicular tuberosity region based on contextual metadata; locating the navicular tuberosity point as the source point in the posture-normalized point cloud data based on the search range; calculating the geodesic distances from all other points to the source point on the surface formed by the posture-normalized point cloud data to generate a geodesic distance field; and inputting the geodesic distance field into a preset Gaussian decay function. The system calculates a soft-partition weight mask centered on the source point with weights that decrease smoothly outwards. Based on the height of each data point in the 3D foot point cloud data and the probability transition band width parameter dynamically configured by contextual metadata, the contact probability density is calculated. This contact probability density is multiplied by the weight value of the soft-partition weight mask at each data point to obtain the weighted contact probability of that data point. All weighted contact probabilities are integrated to obtain the weighted contact probability volume of the arch core area. This weighted contact probability volume of the arch core area is compared with the total contact probability volume without a weight mask to obtain the first arch index result.

[0011] The above-mentioned technical solution utilizes a geodesic distance field to generate a soft-partition weighted mask, and combines this with a probabilistic transition zone to calculate the contact probability density, replacing the rigid region division and height threshold judgment based on geometric coordinates in traditional methods. This "soft calculation" mode is more consistent with the biomechanical characteristics of the foot and can more accurately describe the complex contact relationship between the core arch area and the ground. Especially when dealing with non-standard foot types such as flat feet or high arches, its calculation results are more robust and anatomically significant.

[0012] In some implementations, before inputting the three-dimensional foot point cloud data into a preset first deep learning model for pre-analysis, the method further includes: generating a current state reference point cloud reflecting the current hardware drift state of the three-dimensional foot scanner through the digital twin system of the three-dimensional foot scanner, based on a preset virtual foot standard representing an ideal scanning state and combined with the current real-time operating parameters of the three-dimensional foot scanner; registering the current state reference point cloud with the ideal point cloud data of the virtual foot standard to calculate a three-dimensional real-time distortion correction field for compensating for systematic hardware drift; and performing point-by-point coordinate compensation on the three-dimensional foot point cloud data according to the three-dimensional real-time distortion correction field to obtain the hardware drift corrected three-dimensional foot point cloud data.

[0013] By employing the above technical solution and introducing a digital twin system, real-time compensation for hardware drift in a 3D foot scanner is achieved. This method generates a distortion correction field reflecting the current hardware state through virtual scanning before scanning and performs point-by-point coordinate compensation on the acquired raw point cloud data. This proactively eliminates systematic measurement errors caused by factors such as prolonged equipment operation and temperature changes, ensuring the absolute accuracy of the 3D foot point cloud data from the data source. This provides high-fidelity foundational data for all subsequent calculation steps and is a key guarantee for achieving high-precision measurement.

[0014] In some implementations, acquiring three-dimensional point cloud data of the subject's sole includes: dividing the subject's sole projection area into key physiological regions and non-key regions with different sampling priorities based on prior knowledge of foot and ankle anatomy; dynamically adjusting the encoding density and complexity of the structured light pattern projected onto the key physiological regions according to the expected value of the contour curvature of the key physiological regions; and controlling the optical projector and image sensor to scan the non-key regions at a baseline sampling rate while simultaneously scanning the key physiological regions at an adaptively enhanced sampling rate to generate three-dimensional point cloud data of the sole with adaptively enhanced information encoding and sampling density.

[0015] By employing the aforementioned technical solution, and dividing key and non-key physiological regions, and dynamically adjusting the structured light coding density and sampling rate based on the expected contour curvature, optimized configuration of scanning resources is achieved. This adaptive enhanced scanning strategy can significantly improve the point cloud density and quality of key morphological assessment regions such as the navicular bone and metatarsal heads while ensuring scanning efficiency. The resulting point cloud data has higher information content and richer details, providing a more reliable data foundation for subsequent landmark recognition and morphological analysis, thereby improving the accuracy of the entire calculation method.

[0016] In some implementations, the 3D foot scanner is equipped with a thermal imaging sensor; before dividing the subject's plantar projection area into key physiological regions and non-key regions with different sampling priorities based on prior knowledge of foot and ankle anatomy, the method further includes: acquiring a real-time thermal map of the subject's plantar surface using the thermal imaging sensor before performing the 3D scan; performing image analysis on the real-time thermal map to identify one or more thermally abnormal sub-regions with abnormal temperatures; performing a spatial union operation between the thermally abnormal sub-regions and the key physiological regions to generate an individualized list of enhanced sampling target regions; and replacing the key physiological regions with the individualized list of enhanced sampling target regions.

[0017] By employing the above-mentioned technical approach, general anatomical prior knowledge is combined with the real-time physiological state of the individual subject (such as abnormal thermal areas and high-pressure points on the sole of the foot), making the areas to be scanned more targeted. This not only more accurately captures potential abnormalities or stress characteristics that may affect the arch shape of the foot, but also provides an additional dimension for subsequent biomechanical analysis, making the assessment of the arch index more comprehensive and clinically valuable.

[0018] In some implementations, calculating the consistency coefficient between the first arch index result and the second arch index result includes: calculating a dynamic consistency weighting factor based on scan quality indicators and foot physiological characteristic indicators in contextual metadata; calculating the absolute difference value between the first arch index result and the second arch index result; inputting the absolute difference value into a normalization function with the dynamic consistency weighting factor as a parameter for processing to obtain a preliminary consistency score; and performing a nonlinear mapping on the preliminary consistency score based on the foot physiological characteristics represented by the contextual metadata to obtain the consistency coefficient.

[0019] Using the above technical solution, a dynamic consistency coefficient calculation method based on multi-dimensional information was designed. This method not only considers the absolute difference between the two arch index results, but also innovatively introduces a dynamic weighting factor and nonlinear mapping determined by contextual metadata (scan quality, foot type features). This mechanism enables the consistency evaluation criteria to be intelligently adjusted according to the specific circumstances of the input data. Stricter requirements are applied when the data quality is high and the foot type is typical, while requirements are appropriately relaxed when the situation is complex. This avoids misjudgments that may result from a "one-size-fits-all" approach, making the final fusion decision more scientific and reasonable.

[0020] In a second aspect, the present invention provides an electronic device comprising: one or more processors and a memory; the memory being coupled to the one or more processors, the memory being used to store computer program code including computer instructions, the one or more processors invoking the computer instructions to cause the electronic device to perform the methods described in the first aspect and any possible implementation thereof.

[0021] Thirdly, the present invention provides a computer-readable storage medium storing instructions that, when executed on an electronic device, cause the electronic device to perform the method described in the first aspect and any possible implementation thereof.

[0022] Fourthly, the present invention provides a computer program product comprising instructions that, when the computer program product is run on an electronic device, cause the electronic device to perform the method described in the first aspect and any possible implementation thereof.

[0023] Understandably, the electronic device provided in the second aspect, the storage medium provided in the third aspect, and the computer program product provided in the fourth aspect are all used to execute the method provided by this invention. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here. Attached Figure Description

[0024] Figure 1 This is a flowchart of a method for calculating the arch index according to an embodiment of the present invention; Figure 2 This is a technical path diagram of a foot arch index calculation method according to an embodiment of the present invention; Figure 3 This is a real-time thermal image of the sole of the foot in an embodiment of the present invention; Figure 4 This is a schematic diagram of the hardware architecture of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0025] like Figure 1As shown, this embodiment of the invention provides a method for calculating the arch index, including the following steps: S101 acquires three-dimensional point cloud data of the subject's sole using a three-dimensional foot scanner.

[0026] In this embodiment, the arch index calculation system (hereinafter referred to as the system) controls a three-dimensional foot scanner to scan the soles of the subject's feet. The three-dimensional foot scanner can be any non-contact three-dimensional scanning device based on laser scanning technology, structured light scanning technology, or photogrammetry technology.

[0027] The scanning process requires the subject to stand naturally, placing one or both feet on the scanner's platform. The scanner emits a specific pattern of light (such as laser lines or structured light grids) onto the sole surface, and a camera captures the deformation of the light caused by the contours of the sole surface. Using triangulation principles or other relevant 3D reconstruction algorithms, the system can resolve the 3D spatial coordinates (x, y, z) of thousands of points on the sole surface.

[0028] The collection of these points constitutes the 3D point cloud data of the foot. The 3D point cloud data of the foot is a raw, discrete dataset describing the geometry of the foot. To facilitate subsequent processing, the system may also perform some preprocessing on the raw 3D point cloud data of the foot, such as: removing outlier noise points generated by the scanning environment, unifying the point cloud data to a standard coordinate system (e.g., using the lowest point of the foot as the origin, the length direction of the foot as the X-axis, and the width direction of the foot as the Y-axis), and downsampling the data to reduce the amount of computation while retaining key morphological features.

[0029] S102, input the three-dimensional point cloud data of the foot into the preset first deep learning model for pre-analysis to generate contextual metadata. The contextual metadata is structured data that characterizes the scanning quality of the three-dimensional point cloud data of the foot itself and the physiological characteristics of the foot it reflects. The first deep learning model is a meta-analysis model used to perform quality assessment and feature classification of the input data.

[0030] In this embodiment, a pre-analysis step is introduced to address the problem of poor adaptability of traditional geometric algorithms to different data quality and foot types. The core of this step is to use a deep learning model to "understand" the input 3D point cloud data of the foot, rather than directly calculating it.

[0031] The first deep learning model can be a meta-analysis model. It is defined as: a deep learning model used for quality assessment and feature classification of input data. The purpose of this model is not to output the final arch index, but to analyze and label the data itself, providing contextual information for subsequent computational steps.

[0032] The first deep learning model can adopt a network architecture that performs well in the field of point cloud processing, such as PointNet, PointNet++, or Dynamic Graph Convolutional Network (DGCNN).

[0033] The training process of the first deep learning model is as follows: A large-scale 3D foot point cloud dataset is prepared, and domain experts annotate each sample in the dataset. The annotations include two categories: Scan quality labels: such as "high quality", "medium quality (e.g., sparse point cloud)" and "low quality (e.g., data holes or severe noise)".

[0034] Foot physiological feature labels include, for example, "typical flat foot," "typical high arch," "normal arch," and "slight pronation." Then, this labeled dataset is used to train the first deep learning model using supervised learning, enabling it to automatically identify these quality and physiological features from the input 3D foot point cloud data.

[0035] In practical applications, when a new 3D point cloud data of the foot is input into the first trained deep learning model, the model will perform forward propagation calculations and output contextual metadata.

[0036] The contextual metadata can be a JSON object or a similar data structure, containing at least the scan quality level (e.g., sparse) and the foot physiological feature type (e.g., high_arch). For example: {"scan_quality":"sparse","physiological_feature":"high_arch","noise_level":"low"}, or {"scan_quality":"high","physiological_feature":"flat_foot","noise_level":"none"}.

[0037] This contextual metadata provides a crucial basis for adaptive adjustments in subsequent steps.

[0038] In some embodiments, the first deep learning model may employ the PointNet++ network architecture. Its output layer is designed as a multi-task output head: one classification head outputs the scan quality level (e.g., a confidence scalar in the [0,1] interval), and another regression head outputs foot feature vectors (e.g., a vector containing normalized arch height values ​​and pronation angle values). The training dataset includes manually labeled noise labels and foot type classification labels.

[0039] S103, Based on the context metadata, dynamically configure the calculation parameters in the preset parametric geometric algorithm model to obtain the configured parametric geometric algorithm model.

[0040] Traditional geometric algorithms typically use fixed parameters, which leads to inconsistent performance when faced with diverse inputs. This step aims to leverage the contextual metadata generated in the previous step to make the geometric algorithm "intelligent" and "adaptive."

[0041] A parametric geometric algorithm model refers to an algorithm model whose core computational logic is based on geometric principles, but some of its key parameters are variable and can be externally configured. A specific example is the arch index calculation method based on footprint segmentation, whose configurable computational parameters may include the following process: (1) Anatomical landmark search area: For example, when searching for the navicular point, if the contextual metadata indicates "high arch", the system can dynamically adjust the search area upward (i.e., in the positive direction of the Z-axis); if it indicates "flat foot", it will be adjusted downward.

[0042] (2) Foot arch index division ratio: The foot arch index calculation requires dividing the foot into three parts: forefoot, midfoot, and hindfoot. This division ratio (for example, the traditional fixed ratio of 1:1:1) can be dynamically adjusted according to the foot type characteristics in the contextual metadata. For example, for an individual with a short heel, the division ratio of the hindfoot region can be appropriately reduced.

[0043] (3) Point cloud slice threshold: In some methods that assess the arch by calculating the area under different height levels, the height or number of slices can be used as a parameter. If the contextual metadata indicates "high arch", the starting height of the slices can be set higher to better capture the overhanging part of the arch.

[0044] (4) Data smoothing and filtering strength: If the context metadata indicates that the scanning quality is "low quality" or "noise exists", the system can dynamically increase the strength or window size of the smoothing filter (such as Gaussian filter) in the data preprocessing stage to suppress the interference of noise on the localization of the marker point.

[0045] The dynamic configuration process can be based on a pre-defined rule base or lookup table. For example: Rule 1: IF context metadata.physiological_feature IS "high arch foot" THEN navicular point search Z-axis range = [Z_high_min, Z_high_max].

[0046] Rule 2: IF context metadata.scan_quality IS "sparse" THEN marker neighborhood search radius = R_large.

[0047] In this way, the original "one-size-fits-all" geometric algorithm is upgraded to a configured parametric geometric algorithm model that is "tailor-made" according to specific circumstances.

[0048] S104. The three-dimensional point cloud data of the foot is processed by the configured parametric geometric algorithm model to obtain the first arch index result, and the geometric calculation confidence of the first arch index result is calculated.

[0049] This step involves performing the dynamically configured geometric calculations and quantitatively evaluating the reliability of the results.

[0050] First, the system invokes the configured parametric geometric algorithm model to process the input 3D foot point cloud data. Specifically, this may include: Based on the configured search strategy, key anatomical landmarks such as the calcaneal point, the first metatarsal head, and the fifth metatarsal head are located on the point cloud.

[0051] Based on these landmarks, the footprint area (which can be the projection of the point cloud onto the XY plane) is divided into the hind foot area (A), the midfoot area (B), and the forefoot area (C).

[0052] Calculate the ratio of the area of ​​the midfoot region to the total footprint area excluding the toes, i.e., the first arch index result = area (B) / (area (A) + area (B) + area (C)).

[0053] After the calculation is completed, a value is obtained, which is the result of the first arch index.

[0054] Simultaneously, the system also needs to calculate the geometric confidence level of the first arch index result. The geometric confidence level is a value between 0 and 1, used to quantify the reliability of the calculation result. Its calculation can take into account one or more of the following factors: 1. Landmark Location Stability: If the algorithm can quickly converge to a location with extremely high probability density as the landmark within the configured search area, the confidence level is high. Conversely, if there are multiple candidate points around the landmark, or if the landmark is located in a sparse / missing data region, the confidence level is low.

[0055] 2. Geometric Constraint Compliance: Whether the calculated relative positions and distances between landmarks conform to common sense of human anatomy. For example, whether the height of the scaphoid point is within a reasonable range, and whether the metatarsal points form a reasonable transverse arch shape, etc. If abnormalities occur, the confidence level decreases.

[0056] 3. Sensitivity to input quality: This confidence level can be linked to the contextual metadata generated in step two. If the input data quality itself is rated as "low quality," then even if the computation process goes smoothly, the upper limit of its geometric computation confidence level should be limited to a low level.

[0057] For example, the geometric confidence score C_geom can be modeled as: C_geom=w1*S_landmark+w2*S_anatomy+w3*S_quality; C_geom represents the final calculated confidence level of the geometric calculation, which is a comprehensive evaluation value. Its value ranges from 0 to 1. The higher the value, the more reliable the first arch index result obtained based on the geometric algorithm.

[0058] w1, w2, and w3 represent the weighting coefficients for the three sub-scores. These weighting coefficients are pre-set constants used to adjust the importance of different influencing factors in the final confidence assessment. For example, based on a large amount of experimental data, they can be set to w1=0.5, w2=0.3, and w3=0.2. The sum of these weighting coefficients can be 1 to ensure that the final confidence value is within a reasonable range.

[0059] S_landmark represents the stability score for landmark localization. This score quantifies the determinism of the algorithm in locating anatomical landmarks (such as the calcaneal and navicular points) on 3D foot point cloud data. For example, if the probability distribution of candidate point locations identified by the algorithm within a predetermined search area for a landmark exhibits a sharp single peak, the S_landmark score is high; conversely, if the probability distribution is flat or has multiple peaks, it indicates ambiguity in localization, and the S_landmark score is low.

[0060] S_anatomy represents the score for geometric constraint compliance. This score assesses whether the spatial relationships between all located landmarks as a whole conform to standard human anatomy. For example, the system checks whether the calculated distance from the calcaneus to the metatarsal points is within a reasonable range for normal human foot length, or whether the height of the navicular point is a physiologically possible value. If geometric relationships deviate significantly from anatomical common sense, the S_anatomy score will be significantly reduced.

[0061] S_quality represents a score indicating the quality of the input data. This score is directly correlated with the evaluation of scan quality in the contextual metadata generated by the first deep learning model. If the contextual metadata indicates that the input 3D point cloud data is of "high quality," the S_quality score is high (e.g., 1.0); if it indicates "medium quality (e.g., sparse point cloud)," the score is medium (e.g., 0.6); and if it indicates "low quality (e.g., data holes exist)," the score is very low (e.g., 0.1). This score ensures that the quality of the raw data directly impacts the confidence level of the final geometric calculations.

[0062] S105, input the three-dimensional point cloud data of the foot to the preset second deep learning model to generate the second arch index result and the corresponding model inference confidence. The second deep learning model is a regression prediction model used to directly map the input data to the target arch index value.

[0063] This step employs a completely different approach in parallel—deep learning end-to-end regression—to calculate the arch index, serving as an alternative result that validates the geometric method.

[0064] The second deep learning model can be a regression prediction model, defined as: a deep learning model for directly mapping input data (3D point cloud of the foot) to a target arch index value. Unlike the first deep learning model, which performs a classification task, this model performs a regression task. Its network architecture can also use PointNet, PointNet++, etc., but its last layer can be one or more fully connected layers, ultimately outputting a continuous scalar value, namely the arch index.

[0065] The training process of the model is as follows: A large-scale 3D foot point cloud dataset is prepared, in which each point cloud sample is paired with a "gold standard" arch index value. This gold standard value can be repeatedly confirmed by multiple senior orthopedic surgeons or foot and ankle specialists based on clinical experience and combined with CT or MRI images to ensure its accuracy. Then, using the point cloud as input and the gold standard arch index as the target output, the second deep learning model is trained through end-to-end supervised learning. The model learns the complex nonlinear mapping relationship from point cloud morphology to arch index by minimizing loss functions such as the mean squared error (MSE) between the predicted and true values.

[0066] In practical applications, the same three-dimensional point cloud data of the foot is input into a trained second deep learning model, and the model directly outputs a predicted value, which is the result of the second arch index.

[0067] In addition, to evaluate the reliability of the deep learning model's prediction results, the system also calculates the corresponding model inference confidence score. Methods for calculating the model inference confidence score may include: 1. Variance Evaluation Based on Model Ensemble: Multiple (e.g., 5) second deep learning models with identical structures but different initialization parameters are trained simultaneously. For the same input point cloud, each model provides a prediction. The variance of this set of predictions can be used as a measure of confidence: the smaller the variance, the more consistent the models are, and the higher the confidence of the model's inference; conversely, the larger the variance, the lower the confidence. Confidence can be represented as 1 - normalized_variance.

[0068] 2. Based on Bayesian deep learning: A Bayesian neural network is used as the second deep learning model. This network not only outputs a predicted value, but also the uncertainty of that predicted value (i.e., a probability distribution). The smaller the variance or standard deviation of the distribution, the more certain the model is about the prediction result, and the higher the confidence of the model's inference.

[0069] 3. Uncertainty estimation based on Dropout: During the inference phase, the Dropout layer in the neural network is activated multiple times (e.g., 50 times) for forward propagation to obtain a set of prediction results. Similarly, the dispersion of this set of results can be used to measure the confidence of the model's inference.

[0070] S106, calculate the consistency coefficient between the first arch index result and the second arch index result.

[0071] The purpose of this step is to quantify the degree of agreement between results obtained from two completely different methods. High consistency means that the two methods arrive at the same conclusion from different perspectives, further enhancing the reliability of the results.

[0072] The consistency coefficient is an indicator that assesses the similarity between two numerical values, and its value ranges from 0 to 1. The calculation method is not specifically limited here; for example: Consistency coefficient = 1 - |first arch index result -second arch index result| / ((first arch index result + second arch index result) / 2); This formula calculates the ratio of the difference between two results to their mean, and subtracts that ratio from 1, so that the closer the results are, the higher the consistency coefficient.

[0073] Alternatively, a simpler normalized difference can be used: Consistency coefficient = max(0, 1-k*|first arch index result - second arch index result|), where k is the scaling factor.

[0074] S107. When both the confidence level of the geometric calculation and the confidence level of the model inference are greater than the preset confidence threshold, and the consistency coefficient is greater than or equal to the preset consistency threshold, the results of the first arch index and the second arch index are fused to obtain the final arch index of the subject.

[0075] This is the decision-making and fusion step of the method in this embodiment, and it is also the key to the final output of a high-quality arch index. The system will preset three thresholds: Preset confidence threshold for geometric calculation: for example, T_geom=0.8.

[0076] Preset model inference confidence threshold: for example, T_model=0.8.

[0077] Preset consistency threshold: for example, T_consistency=0.9.

[0078] The system will perform a logical judgment: IF (geometric calculation confidence > T_geom) AND (model inference confidence > T_model) AND (consistency coefficient >= T_consistency) It is only when this condition is met that both methods are confident in their internal computation processes, and their results corroborate each other, achieving a high degree of consistency. In this "ideal" situation, the system will perform the fusion operation.

[0079] The fusion operation aims to combine the advantages of two methods to obtain a final value that is more robust than any single result. A preferred fusion method is confidence-weighted averaging: Final Arch Index = (First Arch Index Result * Geometric Calculated Confidence + Second Arch Index Result * Model Inference Confidence) / (Geometric Calculated Confidence + Model Inference Confidence). In this way, the result with higher confidence has a larger proportion in the final result, achieving intelligent dynamic weighting.

[0080] If the above logical judgment is not true—that is, if neither the confidence score of the geometric calculation nor the confidence score of the model inference is greater than a preset confidence threshold, or if the consistency coefficient is less than a preset consistency threshold—it indicates at least one problem: the geometric calculation is unreliable, the deep learning model's prediction is uncertain, or the results of the two methods differ too much. In this case, the system will not hastily provide a fusion result but can trigger an exception handling process. For example: An alert is sent to the operator, indicating that "the calculation result is uncertain and manual verification is recommended."

[0081] The interface displays the two results, their respective confidence levels, and consistency coefficients for expert decision-making.

[0082] If the contextual metadata indicates a scan quality issue, the system can automatically prompt "Please rescan the feet." Alternatively, it can select the result with higher confidence from the two results as the reference result, along with a low-confidence label.

[0083] In some embodiments, when the consistency coefficient is less than the consistency threshold, the first deep learning model or the second deep learning model is retrained.

[0084] First, the system will automatically mark and store the data samples that produced the inconsistent results, including the input three-dimensional point cloud data of the foot, the two calculation results and their respective confidence levels.

[0085] Secondly, these labeled samples are manually reviewed by domain experts. The experts' task is to identify the reasons for inconsistencies and provide a correct "gold standard" label. If the first deep learning model is determined to have misclassified foot features, the expert will provide the correct classification label; if the second deep learning model is determined to have mispredicted the arch index, the expert will provide a correct arch index value.

[0086] The system then adds these expert-annotated new data pairs to the training datasets of the corresponding models. For example, the "point cloud - correct classification" data pair is used to update the first deep learning model, and the "point cloud - correct index value" data pair is used to update the second deep learning model.

[0087] Finally, after accumulating a sufficient amount of newly labeled data, the system will fine-tune or incrementally train the corresponding model. In this way, the model can learn from past errors and specifically improve its performance on challenging cases. The updated model will be validated before deployment to ensure that its overall performance is improved.

[0088] Through this closed-loop feedback and retraining mechanism, the arch index calculation system can continuously improve itself, making its ability to analyze and predict arch shape increasingly accurate and reliable.

[0089] refer to Figure 2 This embodiment employs the above-described method and steps to construct a technical path that integrates dual-model parallel computing and intelligent decision-making, achieving a significant improvement in the accuracy, robustness, and reliability of arch index calculation.

[0090] First, the parameterized geometric algorithm is dynamically configured using contextual metadata generated by the first deep learning model, enabling traditional geometric methods to have adaptive capabilities. This allows the calculation parameters to be optimized based on data quality and foot features, overcoming the performance degradation of fixed-parameter algorithms when faced with scanning noise or special foot types.

[0091] Secondly, an architecture that combines geometric algorithms and deep learning models in parallel processing is adopted to achieve complementary advantages; the geometric algorithm provides interpretable physical meaning support, while the deep learning model exerts powerful feature extraction capabilities, and the two mutually verify each other.

[0092] Finally, a fusion decision mechanism based on double-reset reliability and consistency coefficient is introduced to ensure that the final result is fused only when the calculation results of both parties are highly confident and consistent, effectively avoiding the risk of misjudgment by a single model on low-quality data or out-of-distribution samples.

[0093] This technical approach enables the method in this embodiment to adaptively balance computational efficiency and accuracy in complex real-world application scenarios, outputting an arch index with clear reliability guarantees, thus solving the problem of poor accuracy and stability in current arch index calculations.

[0094] In some optional embodiments of the present invention, in order to simplify specific application scenarios (such as pursuing ultimate computational efficiency), the parameterized geometric algorithm model in the first computation path can adopt a classic geometric algorithm with fixed parameters.

[0095] Specifically, this geometric algorithm slices the 3D point cloud data of the sole at a preset fixed height (e.g., 2.5 mm), and considers the projected area of ​​the point cloud below this slice as the sole contact area. By calculating the ratio of the midfoot area to the total contact area, the first arch index is obtained. This method can be regarded as a digital implementation of the traditional footprint method.

[0096] Simultaneously, a second deep learning model generates a second arch index result in parallel. The system then calculates the consistency coefficient (e.g., intragroup correlation coefficient, ICC) between the two results.

[0097] In the decision-making logic of this embodiment, when the consistency coefficient is higher than a preset threshold (e.g., 0.9), the system can selectively use the more adaptable second arch index result as the final output. If the consistency requirement is not met, a manual calibration process is initiated, and the calibrated data is used for iterative optimization of the second deep learning model to ensure that the system has the ability to learn independently and continuously improve.

[0098] The following further describes the method for calculating the arch index according to an embodiment of the present invention, including the following steps: S201: Before performing a 3D scan, a real-time thermal image of the subject's sole is obtained through a thermal imaging sensor set in a 3D foot scanner.

[0099] In this embodiment, the first step in data acquisition is to obtain temperature distribution information of the sole of the foot. In addition to standard optical scanning modules (such as structured light projectors and CMOS cameras), the 3D foot scanner can also integrate a thermal imaging sensor. A thermal imaging sensor is a device capable of detecting infrared radiation emitted by an object and converting it into a visible image; a microbolometer array can be used as the core component.

[0100] After the subject places their foot on the scanning platform and before the 3D scan officially begins, the arch index calculation system first activates the thermal imaging sensor. The thermal imaging sensor is aimed at the sole of the foot and, without contact, captures the infrared radiation intensity of different areas of the sole surface, converting it into a digital image—a real-time thermal map of the sole. Figure 3As shown in the real-time thermal map of the foot, the grayscale or pseudo-color value of each pixel corresponds to the temperature of a specific location on the sole surface. For example, areas with high temperatures may be displayed as red or white, while areas with low temperatures may be displayed as blue or black.

[0101] S202 performs image analysis on the real-time thermal map of the sole to identify one or more thermal anomaly sub-regions with abnormal temperatures.

[0102] In this step, the arch index calculation system calls an image analysis module to process the real-time thermal image of the sole. The purpose of this step is to automatically locate areas on the sole where there may be physiological abnormalities. Areas of abnormal foot temperature, poor blood circulation, or specific pressure points can cause significant differences in local temperature compared to surrounding tissues.

[0103] The image analysis process may include: 1. Baseline temperature calculation: Calculate the average or median temperature of the entire plantar real-time thermogram as the baseline for normal temperature.

[0104] 2. Threshold Segmentation: Set one or more temperature thresholds (e.g., 2°C higher or 1.5°C lower than the reference temperature). The image analysis module identifies all pixels in the heat map whose temperature values ​​exceed this threshold range.

[0105] 3. Region Growing and Morphological Processing: The identified anomalous pixels are clustered to form connected regions. Image morphological operations (such as erosion and dilation) are used to remove isolated noise points and smooth the region boundaries, ultimately resulting in one or more well-shaped thermal anomaly sub-regions.

[0106] Each thermal anomaly sub-region represents a potential area on the sole of the foot that requires close monitoring.

[0107] refer to Figure 3 This demonstrates an example of a real-time heat map of the foot. Figure 3 In the diagram, different colors correspond to different temperature levels. Figure 3 In the image, the real-time thermal image of the sole of the foot shows the temperature differences in different parts of the sole. Darker areas (such as dark blue) represent normal tissue areas with lower temperatures, while lighter or brighter areas (such as yellow, orange, or red) represent areas with relatively higher temperatures.

[0108] The system performs image analysis on real-time thermal maps of the sole to identify one or more thermally abnormal sub-regions. A thermally abnormal sub-region refers to a localized area on the sole surface where the temperature is significantly higher or lower than the average temperature of the surrounding normal tissues. This temperature abnormality is usually related to physiological conditions, such as localized thermally abnormal areas, high-pressure points, or abnormal blood circulation.

[0109] S203 performs a spatial union operation on the thermal anomaly sub-region and the key physiological regions divided based on prior knowledge of foot and ankle anatomy to generate an individualized list of enhanced sampling target regions.

[0110] To achieve efficient and accurate acquisition of the three-dimensional morphology of the foot, this embodiment predefines several regions that are of significant anatomical importance in the assessment of foot arch morphology. These key physiological regions are pre-defined based on prior knowledge of foot and ankle anatomy, including areas such as the navicular tuberosity, calcaneal tuberosity, and the first and fifth metatarsal heads. The morphology of these regions is crucial for calculating the arch index.

[0111] This step combines general human anatomy knowledge with the individual physiological condition of the subject. The arch index calculation system performs a union operation in spatial location. This operation merges the coordinate ranges of all thermal anomaly sub-regions identified in step S202 with the coordinate ranges of preset key physiological regions.

[0112] For example, if the preset key physiological regions are A and B, and the thermal anomaly sub-region identified from the thermogram is C, then the result of the union operation is a set of regions A, B, and C. This merged set is the individualized list of enhanced sampling target regions. This list is "tailor-made" for the current subject, including both universally important anatomical regions and potential abnormal regions specific to that individual.

[0113] S204, replace the key physiological regions with a personalized list of enhanced sampling target regions.

[0114] This step involves a replacement operation. In subsequent scan parameter configurations, the system will no longer use the original, universal list of key physiological regions, but will directly use the aforementioned more personalized and targeted list of enhanced sampling target regions. Through this replacement, the "focus" of the scan shifts from universally applicable anatomical key areas to dynamic target regions that combine universality and individuality.

[0115] S205, based on prior knowledge of foot and ankle anatomy, divides the subject's plantar projection area into key physiological regions (which have been replaced by an individualized list of enhanced sampling target regions) and non-key regions with different sampling priorities.

[0116] After the replacement operation is completed, the system officially divides the entire plantar scanning area into regions. At this point, all regions in the individualized enhanced sampling target region list are defined as high-sampling-priority regions. Other parts of the plantar projection region besides these high-priority regions (such as the area below most of the arch and relatively flat areas of the plantar skin) are defined as low-sampling-priority, non-critical regions. Thus, the entire plantar surface is divided into two categories: regions requiring fine scanning and regions requiring routine scanning.

[0117] S206, based on the expected value of the contour curvature of the key physiological region (which can be replaced by an individualized list of enhanced sampling target regions), dynamically adjust the encoding density and complexity of the structured light pattern projected onto the key physiological region.

[0118] This step aims to acquire higher-quality 3D data in key areas. Structured light patterns are gratings or stripes with coded information projected onto the sole surface by a 3D foot scanner; their coding density and complexity determine the accuracy of the 3D reconstruction.

[0119] The arch index calculation system queries a pre-set database that stores expected values ​​for the typical contour curvature of different anatomical regions of the foot. For example, the curvature of bony prominences such as the metatarsal heads varies greatly, resulting in a high expected value; while the mid-section of the plantar fascia is relatively flat, leading to a low expected value.

[0120] The system dynamically adjusts the structured light pattern projected onto each region in an individualized list of enhanced sampling target regions, based on the expected curvature value of that region. For regions with high expected curvature values, the system selects a more densely encoded, finer-stripe, and more complex structured light pattern to capture more surface details. For regions with low expected curvature values, a relatively sparse pattern can be used. This adjustment is achieved by generating and projecting different light patterns in real time using a digital micromirror device (DMD).

[0121] S207 controls the optical projector and image sensor to scan non-critical areas at a baseline sampling rate, while simultaneously scanning critical physiological areas at an adaptively enhanced sampling rate, in order to generate three-dimensional point cloud data of the foot with adaptively enhanced information encoding and sampling density.

[0122] This step describes the specific operations involved in performing the adaptive scan. The arch index calculation system controls the optical projector and image sensor of the 3D foot scanner to work together.

[0123] When the scanning beam sweeps across non-critical areas, the system acquires data at a standard, preset baseline sampling rate. When the scanning beam enters any region from the individualized list of enhanced sampling target regions, the system instantly increases its operating frequency, acquiring data at an adaptively enhanced sampling rate. This can be achieved by increasing the frame rate of the image sensor or by projecting and capturing more sets of differently coded structured light patterns in the same area.

[0124] After this scanning process, the final dataset generated is the three-dimensional point cloud data of the foot. The characteristics of this three-dimensional point cloud data of the foot are: the density and accuracy of the point cloud are significantly higher in key areas of the arch shape and individual-specific thermal anomaly areas than in other non-key areas, achieving optimized allocation of data acquisition resources.

[0125] S208, based on a preset virtual foot standard representing the ideal scanning state, combined with the current real-time working parameters of the 3D foot scanner, generates a current state reference point cloud reflecting the current hardware drift state of the 3D foot scanner through the digital twin system of the 3D foot scanner.

[0126] This step aims to proactively compensate for systematic errors that may arise from the scanner hardware itself.

[0127] The virtual foot standard is a computer-generated, geometrically precise, and known idealized 3D digital model of the foot. It is not a copy of a real foot, but rather designed using professional 3D modeling software (such as SOLIDWORKS and CATIA) based on standardized or idealized human foot and ankle anatomy data and biomechanical principles. During its construction, all geometric parameters, including length, width, arch height, and coordinates of key skeletal landmarks (such as the calcaneus and metatarsal heads), are precisely defined to ensure a smooth, flawless surface, representing the perfect foot shape that a 3D scanner should capture under ideal conditions. This standard serves as an unchanging reference point, its core value lying in providing a "gold standard" benchmark for measuring and calibrating the measurement deviations of physical scanning equipment.

[0128] A digital twin system for a 3D foot scanner is a virtual simulation model that is synchronized in real time with the physical scanner (i.e., the 3D foot scanner). The digital twin system can accurately simulate the behavior of each component of the physical scanner (such as the laser, camera, and motion mechanism) and can receive real-time operating parameters uploaded by the physical scanner, such as ambient temperature, laser power, and motor position.

[0129] This digital twin system is built on a professional physical simulation platform (such as Ansys or COMSOL Multiphysics). It is a highly realistic virtual copy that is synchronized in real time with the physical scanner. The generation process begins with a detailed digital model of the physical scanner. The physical characteristics, geometric dimensions, material properties, kinematic relationships, and optical parameters (such as lens distortion coefficient and light source spectrum) of all its key hardware components, including the structured light projector (DLP / LCoS), optical lens assembly, CMOS / CCD image sensor, and motor-driven moving guide rail, are input into the simulation environment. Subsequently, control logic is written to simulate the complete workflow of the physical scanner. The most critical step is to establish a real-time bidirectional data communication link between the physical device and the virtual model. Various sensors on the physical scanner (such as temperature sensors and encoders) transmit their current operating parameters (such as ambient temperature, laser power, and precise motor position) to the digital twin system in real time. The digital twin system then dynamically adjusts the state of its virtual components based on these real parameters, thereby accurately reproducing the actual performance of the physical scanner at the current moment due to factors such as hardware drift.

[0130] Hardware drift refers to the phenomenon where the scanner's hardware parameters deviate from their factory calibration values ​​due to factors such as prolonged operation, temperature changes, or mechanical wear. This deviation can lead to measurement errors.

[0131] In this step, the arch index calculation system takes a virtual foot standard as input and feeds it into the digital twin system of the 3D foot scanner, which has been updated with the physical scanner's current real-time operating parameters, to perform a virtual scan. The digital twin system simulates what kind of point cloud data would be obtained by scanning this ideal standard under the current hardware drift state. This simulated point cloud data is the current state reference point cloud.

[0132] S209, the current state reference point cloud is registered with the ideal point cloud data of the virtual foot standard to calculate the three-dimensional real-time distortion correction field used to compensate for systematic hardware drift.

[0133] The purpose of this step is to quantify the error caused by hardware drift. The arch index calculation system compares the current state reference point cloud with the ideal point cloud data of the original, perfect virtual foot standard.

[0134] The registration process can use the Iterative Closest Point (ICP) algorithm or its variants to align two point clouds. After alignment, the system calculates the spatial displacement vector between each pair of corresponding points in the two point clouds. All these displacement vectors together constitute a three-dimensional vector field, namely the three-dimensional real-time distortion correction field. The three-dimensional real-time distortion correction field describes the magnitude and direction of the coordinate offset caused by hardware drift at each location in three-dimensional space.

[0135] In some embodiments, to optimize user experience and further improve the real-time performance of hardware drift correction, this embodiment may employ an efficient mode combining offline modeling and online inference. The specific implementation of this mode is as follows: First, during the scanner's off-peak hours or preset maintenance cycles, the arch index calculation system executes an offline, comprehensive self-calibration procedure. In this procedure, the system controls the 3D foot scanner to repeatedly execute the entire process of generating a current-state reference point cloud and registering it with a virtual foot standard at multiple preset key operating parameter points (e.g., different ambient temperatures, cumulative laser operating time, etc.). Through this offline calibration process, the system obtains a series of precise 3D real-time distortion correction field samples describing the scanner under different hardware conditions.

[0136] Secondly, the system uses the collected sample data to train a parameterized hardware drift correction model. This model is preferably a lightweight machine learning regression model. The input of the model is the current real-time operating parameters of the 3D foot scanner (which can be acquired in real time by the built-in sensor), and the output of the model is a set of key parameters (e.g., a set of B-spline surface control points or a set of polynomial function coefficients) that can represent the entire 3D real-time distortion correction field in low dimension.

[0137] Finally, during real-time foot scanning, the arch index calculation system will no longer perform virtual scanning and point cloud registration. Instead, it will directly input the collected real-time operating parameters into a pre-trained parameterized hardware drift correction model. The model will then output the corresponding key parameters through a single rapid forward inference calculation. The system will then use these key parameters to quickly reconstruct an approximate three-dimensional real-time distortion correction field in the current state, which will then be used for point-by-point coordinate compensation in step S210.

[0138] By adopting this strategy of "slow offline calibration and fast online inference," this embodiment can reduce the time required for hardware drift correction from seconds to milliseconds without sacrificing correction accuracy, thereby effectively solving the real-time problem and ensuring a smooth user scanning experience.

[0139] S210 performs point-by-point coordinate compensation on the three-dimensional point cloud data of the foot based on the three-dimensional real-time distortion correction field, and obtains the three-dimensional point cloud data of the foot after hardware drift correction.

[0140] This is the process of applying the correction field. The arch index calculation system iterates through every point in the 3D point cloud data of the foot obtained in step S207. For each point, the system finds the corresponding displacement vector in the 3D real-time distortion correction field at its coordinate position, and then subtracts (or adds, depending on the definition of the distortion field) this displacement vector from the original coordinates of the point to obtain a new compensated coordinate.

[0141] After performing this operation on all points, we obtain the hardware-drift-corrected 3D point cloud data of the foot. This data eliminates systematic errors introduced by changes in the scanner's own state, and its coordinate values ​​are closer to the true physical shape of the measured foot. All subsequent calculations will be based on this high-quality point cloud data.

[0142] S211, input the hardware drift-corrected 3D point cloud data of the foot into the preset first deep learning model for pre-analysis to generate contextual metadata.

[0143] This step can be referred to the description in the previous embodiments, and will not be repeated here.

[0144] S212, extract scan quality indicators that characterize the level of scan noise and foot physiological characteristic indicators that characterize the internal and external rotation angles of the foot from contextual metadata.

[0145] The arch index calculation system parses contextual metadata. In this step, the contextual metadata can specifically be a data structure containing multiple key-value pairs, implemented by parsing predefined key-value pairs. This step specifically extracts two specific pieces of information from it: Scan quality metric for scan noise level: This is a numerical value or level, such as from 0 (no noise) to 1 (strong noise), used to quantify the severity of random noise in point cloud data. A relatively low scan quality metric indicates a high level of scan noise.

[0146] Foot physiological characteristic index of internal and external rotation angle: This is an angle value or classification label (such as "internal rotation", "external rotation", "neutral") used to describe the rotational posture of the foot relative to the standard forward direction when the subject is standing.

[0147] Specifically, the scan quality index is read from fields related to noise level (such as "noise_level" or "scan_quality") in the context metadata, which quantifies the noise level of the point cloud data; the foot physiological feature index is obtained from fields characterizing foot posture (such as "foot_rotation_angle" or "physiological_feature"), whose values ​​reflect the angle of internal and external rotation of the foot or the classification label.

[0148] The entire extraction process is based on the metadata tags generated by the first deep learning model in the pre-analysis stage. It does not require complex calculations and can be completed with only simple data access and mapping.

[0149] S213, based on the scan quality index, adaptively determines the Gaussian filter kernel radius for point cloud smoothing by querying a preset nonlinear mapping function.

[0150] This step aims to intelligently determine the strength of data smoothing based on the noise level of the point cloud. The Gaussian filter kernel radius is a key parameter in the Gaussian smoothing algorithm; a larger radius results in a stronger smoothing effect but may lose more details. The arch index calculation system internally stores a preset nonlinear mapping function that establishes the relationship between the scan quality index and the Gaussian filter kernel radius. The design principle of this function is: a relatively low scan quality index (meaning a high noise level, i.e., poor point cloud data quality) corresponds to a relatively large Gaussian filter kernel radius to achieve a stronger denoising effect; conversely, high-quality, low-noise point clouds correspond to a smaller radius to preserve surface details to the greatest extent. For example, the function could be Radius = R_min + (R_max - R_min) * (1 - Quality_Score), where Radius is the final calculated Gaussian filter kernel radius used for point cloud smoothing; R_min is the preset minimum filter kernel radius to ensure a lower limit on smoothing intensity; R_max is the preset maximum filter kernel radius to prevent over-smoothing from causing loss of detail; and Quality_Score is the scan quality score, where a higher value indicates better point cloud data quality, and its value ranges from 0 to 1.

[0151] S214, Based on foot physiological characteristic indicators, construct an affine transformation matrix to rotate the three-dimensional point cloud data of the sole from the original scanning coordinate system to the target coordinate system aligned with the principal axis of foot anatomy.

[0152] To eliminate calculation errors caused by non-standard posture of the test subject (such as "pigeon-toed" or "out-toed"), the foot point cloud needs to be rotated to a standard posture. The principal axis of foot anatomy can be defined as a straight line connecting the center point of the heel and the second metatarsal head. The arch index calculation system constructs a rotation matrix based on foot physiological characteristics (such as pronation and supination angles). This rotation matrix rotates the point cloud data around the Z-axis (the axis perpendicular to the ground) by a specific angle, aligning the principal axis of foot anatomy of the point cloud with the Y-axis (or X-axis) of the target coordinate system. This rotation matrix is ​​part of the affine transformation matrix (which is the affine transformation matrix when there is no translation or scaling).

[0153] S215: Based on the Gaussian filter kernel radius, the hardware drift-corrected 3D foot point cloud data is smoothed to obtain smoothed point cloud data. Then, based on the affine transformation matrix, the smoothed point cloud data is subjected to coordinate transformation to generate the final input data to be processed by the configured parametric geometric algorithm model. Thus, the configured parametric geometric algorithm model is obtained.

[0154] This step integrates the configuration results from the previous steps, completing the final preparation of the input data for the geometric algorithm. Specifically, it includes the following processing operations: 1. Smoothing: The system calls the Gaussian smoothing algorithm and uses the Gaussian filter kernel radius mentioned above to process the 3D point cloud data of the foot after hardware drift correction, so as to obtain a smoothed point cloud data with a lower noise level.

[0155] 2. Coordinate Transformation: The system applies an affine transformation matrix to each point in the smoothed point cloud data, rotating it to a standard pose.

[0156] The data obtained after the above two processes is the final input data to be processed by the configured parametric geometry algorithm model. In a broad sense, this entire data preprocessing process, which includes dynamically determining smoothing parameters and attitude correction parameters, is itself a "configuration" of the traditional fixed-parameter geometry algorithm. Therefore, the completion of this series of operations signifies that the configured parametric geometry algorithm model has been obtained.

[0157] S216 processes the hardware-drift-corrected three-dimensional point cloud data of the foot using a configured parametric geometric algorithm model. First, based on the hardware-drift-corrected three-dimensional point cloud data of the foot and the affine transformation matrix, pose-normalized point cloud data is generated.

[0158] This step begins the core calculations of the geometric algorithm. Although the final input data for the calculations has already been generated in S215, for the convenience of certain subsequent calculations (such as geodesic distance calculations), it is explicitly stated here that the first step of the geometric algorithm is to formally generate a complete copy of the point cloud, with all points having undergone attitude correction; that is, attitude-normalized point cloud data. This data is the result of the original hardware-corrected point cloud after affine transformation.

[0159] S217. Determine the search range of the navicular tuberosity region based on the contextual metadata. Locate the navicular tuberosity point as the source point in the attitude-normalized point cloud data based on the search range. Calculate the geodesic distances from all other points to the source point on the surface formed by the attitude-normalized point cloud data to generate a geodesic distance field.

[0160] Specifically, this step can be performed in three stages.

[0161] In the first stage, contextual metadata is used to define a three-dimensional search box to limit the search range of the navicular tuberosity point.

[0162] The navicular tuberosity is anatomically located on the medial side of the foot, near the highest point of the arch. The arch index calculation system first establishes an initial, relatively large search range based on general anatomical prior knowledge. For example, in a posture-normalized coordinate system (foot length as the Y-axis, foot width as the X-axis, and height as the Z-axis), the initial range might be defined as: between 1 / 3 and 2 / 3 of the total foot length along the Y-axis; and between the medial edge of the foot and the midline along the X-axis.

[0163] Next, the system dynamically adjusts the search range, especially the height (Z-axis) range, based on the foot physiological characteristic indicators in the contextual metadata. Specifically: if the contextual metadata indicates "High Arch," the system determines that the position of the navicular tuberosity will be relatively high. Therefore, the system will shift the lower limit (Z_min) and upper limit (Z_max) of the search range upwards. If the contextual metadata indicates "Flat Foot," the system determines that the position of the navicular tuberosity will be very close to the plantar support surface. Therefore, the system will shift the lower limit and upper limit of the search range downwards. For normal arches, a standard Z-axis range is used.

[0164] In this way, a fixed and broad search range is dynamically adjusted to a smaller and more relevant search range for the navicular tuberosity region, which is specific to the current subject's foot type.

[0165] Then, in the second stage, the arch index calculation system uses geometric feature analysis to precisely locate the tuberosity of the navicular bone within the search range determined in the previous stage.

[0166] The system first filters out all points whose coordinates are within the search range of the navicular tuberosity region from the complete pose-normalized point cloud data, forming a candidate point set.

[0167] Subsequently, the navicular tuberosity, as a bony prominence, appears as a locally convex region on the point cloud surface, exhibiting a large local surface curvature. Therefore, the system calculates the principal curvature of the local surface for each point in the candidate point set. The principal curvature is calculated by analyzing the eigenvalues ​​of the covariance matrix formed by the point and its neighboring points. For each point, two principal curvature values ​​(k_max and k_min) are obtained.

[0168] The system searches for the point with the largest positive curvature value (i.e., the largest k_max value) in the candidate point set. Geometrically, this point is the most "convex" point in the local region, which best matches the morphological characteristics of the navicular tuberosity.

[0169] The point with the maximum positive curvature that has been located is then formally designated as the source point for subsequent geodesic distance calculations.

[0170] Finally, we move to the third stage, which aims to calculate the shortest path distance from all other points on the point cloud surface to the source point, i.e., the geodesic distance.

[0171] To simulate a continuous surface on a discrete point cloud, the system first needs to construct an adjacency graph from the pose-normalized point cloud data. In this graph, each point cloud data point is a graph node. Using the k-Nearest Neighbors (k-NN) algorithm or a fixed-radius search algorithm, the system finds its spatial neighbors for each node and establishes edges between them. The weight of each edge is set to the Euclidean distance between the two nodes it connects. This weighted adjacency graph can be viewed as a discretized approximation of the three-dimensional surface of the foot.

[0172] With a weighted adjacency graph and the specified source vertex, calculating the geodesic distance is transformed into a graph theory problem: the single-source shortest path problem. The arch index calculation system can solve this problem using Dijkstra's algorithm.

[0173] The algorithm initializes a distance array, setting the distance from the source point to itself to 0 and the distance to all other nodes to infinity.

[0174] Starting from the source node, the algorithm employs a breadth-first search strategy, progressively expanding and updating the shortest distance from each node to the source node. In each step, the algorithm selects the unvisited node with the shortest known distance and uses this node to update the distance values ​​of all its neighboring nodes.

[0175] Repeat this process until the shortest distance to all nodes in the graph has been calculated and determined.

[0176] After Dijkstra's algorithm completes, the distance array stores the geodesic distance from each node (i.e., each data point) in the point cloud to the source point. This set of data, which associates each data point with its corresponding geodesic distance value, constitutes the geodesic distance field. This geodesic distance field is a scalar field that intuitively reflects the distance relationship between the entire plantar surface and the central arch of the foot (the tuberosity of the navicular bone).

[0177] S218 inputs the geodesic distance field into a preset Gaussian decay function to obtain a soft partition weight mask with the weight value smoothly decaying outwards from the source point.

[0178] The Gaussian decay function can be a function with geodesic distance as the independent variable; the specific function formula is not specifically limited in this embodiment.

[0179] The system substitutes the distance value of each point in the geodesic distance field into this function to calculate a corresponding weight value. The closer a point is to the source point (the tuberosity of the navicular bone), the closer its weight value is to 1; the farther away a point is, the more smoothly the weight value decays to 0. All these weight values ​​form another scalar field, which is the soft-partition weight mask. Unlike traditional methods that abruptly change the weight from 1 to 0 at a certain boundary, this mask provides a smooth transition.

[0180] S219, based on the height value of each data point in the three-dimensional point cloud data of the foot after hardware drift correction, and the probability transition zone width parameter dynamically configured by the context metadata, the contact probability density is calculated.

[0181] The probability transition band width parameter is also dynamically configured by context metadata. Specifically, it is a dynamically adjusted value that determines how "fuzzy" or "smooth" the change in the foot point cloud from "completely in contact with the ground" to "completely not in contact with the ground" is.

[0182] In this embodiment, the arch index calculation system extracts physiological characteristics of the subject's foot type, such as "flat feet" or "high arches," from contextual metadata. The system has a pre-defined lookup table or function that establishes a mapping relationship between foot type features and the width parameter of the probability transition zone. For example, for the ambiguous "flat feet," the system configures a relatively large width parameter; while for the well-defined "high arches," it configures a smaller width parameter.

[0183] Next, the system selects an S-shaped function (e.g., the Sigmoid function), which can be expressed as P(z) = 1 / (1 + exp(k*z)), where z is the height value of the data point. The key parameter k, the kurtosis coefficient, is uniquely determined by the probability transition band width parameter; the narrower the band, the larger the absolute value of k, and the steeper the function curve. Finally, the system iterates through each data point in the hardware-drift-corrected 3D foot point cloud data, substituting its height value z into the dynamically configured logistic function with parameter k, to obtain a probability value P(z) between 0 and 1. This probability value is the contact probability density at that point, and the set of these probability values ​​for all points together constitutes the final contact probability density field.

[0184] S220, multiply the contact probability density by the weight value of the soft partition weight mask corresponding to the data point to obtain the weighted contact probability of the data point.

[0185] This step integrates information from two aspects: "where is the core area of ​​the arch" and "whether the point is in contact with the ground." For each data point in the point cloud, the arch index calculation system multiplies the contact probability density value corresponding to that data point with the soft partition weight mask value. The result is the weighted contact probability of the data point. This value reflects both the degree to which the point belongs to the core area of ​​the arch and the likelihood of it contacting the ground.

[0186] S221, integrate all weighted contact probabilities to obtain the weighted contact probability volume of the core area of ​​the foot arch.

[0187] Specifically, the system sums the weighted contact probabilities of all data points (for discrete point clouds, this is an approximation of integration), and the resulting sum is the weighted contact probability volume of the core area of ​​the arch, which can be understood as the total "probability" of contact between the core area of ​​the arch and the ground.

[0188] In this embodiment, the weighted contact probability volume of the core arch area is a comprehensive value that represents the total probability that the parts defined as the "core arch area" will actually make contact with the ground. Its calculation incorporates information from two dimensions: First, the weight of each point belonging to the core area of ​​the foot arch is obtained by smoothing and decaying the geodesic distance from the point to the navicular tuberosity using a Gaussian function. The closer the distance, the higher the weight. Second, the contact probability of each point due to its height; the lower the height, the higher the contact probability.

[0189] By multiplying the weight of each data point in the point cloud by the contact probability, and then summing the results of these products for all points, the weighted contact probability volume is obtained. Therefore, this value not only considers whether the sole of the foot is in contact with the ground, but also intelligently assigns greater importance to the central area of ​​the arch, thereby more accurately quantifying the load-bearing state of the most critical part of the arch.

[0190] S222, compare the weighted contact probability volume of the core area of ​​the arch with the total contact probability volume without using a weighted mask to obtain the first arch index result.

[0191] In this step, to obtain a standardized index, the system also needs to calculate a reference value. The total contact probability volume without using a weighted mask refers to the result obtained by integrating (summing) the contact probability density of all points directly without considering the soft partition weighted mask (i.e., treating the weight of all points as 1). This value represents the total "probability" of the entire foot's contact with the ground. Finally, by comparing the two (e.g., by division), the first arch index result is obtained. The first arch index result = weighted contact probability volume of the core arch area / total contact probability volume without using a weighted mask. This result essentially reflects the proportion of contact in the core arch area to the entire foot contact, and is a more refined and robust definition of the arch index.

[0192] In this embodiment, the total contact probability volume without weighted masking is a standardized benchmark value representing the total probability of contact between the entire plantar area (regardless of core or non-core region) and the ground. Its calculation method is as follows: First, similar to calculating the weighted contact probability volume, an S-shaped function can be used to calculate the probability of contact between each point and the ground based on the height value of each point in the point cloud data, resulting in a contact probability density map covering the entire sole. Then, the contact probability values ​​of all data points in this map are directly summed without multiplying by any weight values ​​from the soft partition weight mask (equivalent to treating all points as having a weight of 1). The sum obtained in this way is the total contact probability volume without using a weight mask, reflecting the total weight-bearing situation of the entire sole. Using this as the denominator allows the final calculated arch index to become a standardized relative proportion reflecting the weight-bearing ratio of the core area. S223, calculate the geometric confidence level of the first arch index result.

[0193] This step can be referred to the description in the previous embodiments, and will not be repeated here.

[0194] S224, input the hardware drift-corrected three-dimensional point cloud data of the foot into the preset second deep learning model to generate the second arch index result and the corresponding model inference confidence.

[0195] This step can be described with reference to the previous embodiment. In parallel with the geometric method, the system inputs the same hardware drift-corrected 3D foot point cloud data into an end-to-end second deep learning model (regression prediction model) to directly obtain the second arch index result and the model inference confidence score that measures the reliability of the result.

[0196] S225, based on the scan quality index and foot physiological characteristic index in the context metadata, calculates the dynamic consistency weighting factor.

[0197] When comparing the consistency of the two results, this embodiment introduces a dynamic adjustment strategy. The arch index calculation system again utilizes scan quality indicators and foot physiological characteristic indicators from the contextual metadata. For example, the system may follow the rule that if the scan quality is low, or the foot type is very extreme (such as severe flat feet), then the possibility of differences between the two methods is greater, and the consistency requirement can be appropriately relaxed. The system converts these indicators into a numerical value, namely the dynamic consistency weighting factor, through a preset mapping relationship.

[0198] In some embodiments, step S225 may specifically include the following steps: S2251 calculates the Euclidean norm between the scan quality index and the preset standard quality benchmark to obtain the quality deviation.

[0199] In this embodiment, the system first quantifies the gap between the quality of the currently scanned point cloud data and the ideal state. The scan quality index is a set of values ​​extracted from contextual metadata, used to describe the quality of the point cloud from multiple dimensions. For example, the scan quality index can be a vector Q_current=[d,n,h], where d represents the average density of the point cloud, n represents the estimated noise level, and h represents the area ratio of data holes.

[0200] Meanwhile, the system has a preset standard quality benchmark, which is a vector representing the "perfect" or "ideal" scan quality, such as Q_standard=[d_ideal,0,0], where d_ideal is the ideal point cloud density, and the noise level and hole ratio are both 0.

[0201] This step involves calculating the Euclidean norm (i.e., Euclidean distance) between the scan quality index vector Q_current and the standard quality reference vector Q_standard. The Euclidean norm is a commonly used method to measure the straight-line distance between two points in vector space.

[0202] The calculated quality deviation is a non-negative real number. The larger the quality deviation value, the greater the gap between the current point cloud data quality and the ideal state, and the worse the data quality.

[0203] S2252, calculate the cosine similarity between foot physiological characteristic indicators and preset standard physiological characteristic vectors to obtain the feature fit.

[0204] Next, the system assesses how close the current subject's foot type is to the "standard" or "most typical" foot type. Foot physiological characteristic indicators are another set of values ​​extracted from contextual metadata, used to describe key anatomical features of the foot. For example, foot physiological characteristic indicators can be encoded as a vector F_current=[a1,a2,a3], where a1 represents the arch height, a2 represents the inversion / pronation angle, a3 represents the calcaneal width, etc.

[0205] Similarly, the system has a pre-defined standard physiological feature vector, F_standard, which represents the most "standard" or "average" foot shape anatomically. This standard vector can be obtained through statistical analysis of foot data from a large-scale healthy population.

[0206] This step uses cosine similarity to measure the directional consistency between the foot physiological feature vector F_current and the standard physiological feature vector F_standard. The formula for calculating cosine similarity is: Feature fit = (F_current · F_standard) / (||F_current|| * ||F_standard||) The calculated feature fit is a value between -1 and 1 (between 0 and 1 when all features are non-negative). The closer the feature fit value is to 1, the closer the current subject's foot type is to the standard foot type, belonging to a "normal" or "typical" case; the lower the value, the more special or extreme the foot type is.

[0207] S2253, determine the initial weighting base based on the quality deviation and the characteristic matching degree.

[0208] This step combines the evaluation results from the previous two steps to generate a basic adjustment factor. The initial weighted base is a numerical value that comprehensively reflects the difficulty of processing the current data. Its determination rule can be based on a two-dimensional lookup table or a preset function.

[0209] In some embodiments, the initial weighting base can be calculated according to the following formula: Initial weighted base = BaseValue * (1 / (1 + w_q * quality deviation)) * (w_f * feature fit) in: BaseValue is a baseline coefficient that represents the benchmark value under ideal conditions (quality deviation of 0 and feature fit of 1).

[0210] w_q and w_f are preset weighting coefficients used to adjust the influence of the two factors, quality and features.

[0211] In this formula: the greater the quality deviation (the worse the data quality), the smaller the value of the term 1 / (1+w_q*quality deviation); the lower the feature fit (the less standard the foot type), the smaller the value of the term w_f*feature fit. Therefore, the worse the data quality or the more unusual the foot type, the smaller the initial weighted base will be.

[0212] S2254 transforms the initial weighting base using a predefined exponential decay function to obtain a dynamic consistency weighting factor.

[0213] The system uses a nonlinear transformation to convert the initial weighting base into the final parameters used for consistency evaluation. This step aims to make the adjustment effect more sensitive and significant. A predefined exponential decay function is a common nonlinear mapping function, with a form such as: Dynamic Consistency Weighting Factor = A * exp(-k * Initial Weighting Base) Here, A and k are the shape parameters of the function, which can be calibrated using experimental data.

[0214] The characteristic of this function is that when the initial weighting base is large (corresponding to high-quality data and standard foot types), the resulting dynamic consistency weighting factor will be small; while when the initial weighting base is small (corresponding to low-quality data or special foot types), the resulting dynamic consistency weighting factor will be large.

[0215] The resulting dynamic consistency weighting factor will be used in subsequent consistency coefficient calculations. A smaller dynamic consistency weighting factor means lower tolerance for differences between the results of the two algorithms, requiring stricter consistency; conversely, a larger factor means higher tolerance, allowing for greater differences between results. Through this series of refined calculations, the system achieves intelligent and dynamic adjustment of the consistency evaluation criteria.

[0216] In some embodiments, the dynamic consistency weighting factor can also be calculated using the following formula:

[0217] in: U c : Conventional perturbation group factor, characterizing the combined perturbation intensity introduced by data quality and conventional foot features.

[0218] U n : Unconventional perturbation group factor, characterizing the combined perturbation intensity introduced by individual biomechanical state and internal algorithm uncertainty.

[0219] U effEffective perturbation strength, calculated by smoothly selecting the dominant perturbation group, is used for the final decision.

[0220] k : Dynamic consistency weighting factor, an adaptive coefficient used to control the tolerance of differences between the two algorithm results. The smaller the value, the higher the system's tolerance to scanning noise or foot anomalies (i.e., allowing for larger differences between the two results).

[0221] D q The data quality deviation index is a normalized index used to quantify the degree to which the quality of three-dimensional foot point cloud data deviates from the ideal state.

[0222] A f Anatomical feature deviation index is a normalized index used to quantify the degree to which the foot shape of the subject deviates from the typical anatomical range.

[0223] H L Location information entropy is a normalized index that characterizes the degree of dispersion in keypoint localization. The more dispersed the candidate region is when a geometric algorithm locates the scaphoid point, the more significant the dispersion. H L The larger the value, the lower the weight of the result for that path.

[0224] σ th Thermal stress index is a physical quantity calculated based on plantar thermograms. It represents the intensity of the abnormal gradient of the surface temperature field in a local area of ​​the sole caused by inflammation, injury, or excessive weight-bearing. It is a normalized index.

[0225] λ q , λ f 、α、β : These are dimensionless sensitivity coefficients used to adjust the influence weights of the four core disturbance indicators mentioned above.

[0226] k max : Maximum basic factor, representing the upper limit of the most stringent consistency judgment criterion that the system can apply under ideal, perturbation-free conditions (i.e., when the scan data is perfect and the foot type is standard).

[0227] γ A gain coefficient that adjusts the sensitivity of the system, used to control the dynamic consistency weighting factor. k The rate at which it decreases as the total disturbance intensity increases.

[0228] δ: A bias parameter used to filter out small background disturbances, ensuring that the relaxation of the consistency criteria is only triggered when the disturbance intensity accumulates to a certain level.

[0229] It should be noted that in the dynamic consistency weighting factor calculation formula of this embodiment, all parameters will be normalized or standardized to eliminate the problem of inconsistency in units that cannot be calculated.

[0230] The calculation method described in this embodiment has the following beneficial effects: 1. Intelligent adaptive assessment criteria: Through the LogSumExp mechanism, this method can automatically and smoothly shift the assessment focus to the most significant risk source. For example, for a subject with very standard foot type, even if the scan quality has slight flaws ( U c (Dominant) The system can also make reasonable judgments; and for a plantar area with severe thermal anomalies ( U n For rare foot types (dominant), the system automatically increases its tolerance for inconsistencies. This intelligent adaptive capability makes this method far more robust than fixed threshold or simple linear adjustment methods when facing various complex and changing real-world situations.

[0231] 2. Improved system robustness and numerical stability: tanh The function's range is bounded (-1, 1), which ensures that the final value of k is always strictly limited to (0, 1). k max Within a reasonable range, it will not cause program crashes due to abnormal inputs (such as an extremely large disturbance factor) such as numerical overflow or negative values, which greatly enhances the stability and reliability of the entire computing system.

[0232] 3. Provides more refined and interpretable control capabilities: in the formula γ and δ The parameters have a clear physical meaning. γ This can be viewed as adjusting the gain of an amplifier, which can control the system's "sensitivity" to risk; δ Just like setting a threshold, it can filter out insignificant background noise. This decoupled parameter design allows technicians to intuitively and accurately calibrate and fine-tune the model according to different application scenarios (such as strict requirements for physical examinations, while retail recommendations can be more lenient), achieving a high degree of flexibility and configurability.

[0233] S226, calculate the absolute difference between the first arch index result and the second arch index result.

[0234] Specifically, the following mathematical operation can be used: Absolute difference value = |First arch index result -Second arch index result|.

[0235] S227. The absolute difference value is input into a normalization function with a dynamic consistency weighting factor as a parameter for processing to obtain a preliminary consistency score.

[0236] This step transforms absolute differences into a standardized score. The normalization function can be designed as: Initial Consistency Score = exp(-k * absolute difference value), where k is the dynamic consistency weighting factor calculated in step S225. When the dynamic consistency weighting factor k is large (corresponding to high-quality data and standard foot type), even small absolute differences will cause the initial consistency score to drop rapidly, indicating stricter requirements; conversely, when k is small, the function decreases more gradually, showing a higher tolerance for differences.

[0237] S228. The initial consistency score is nonlinearly mapped based on the foot physiological characteristics represented by contextual metadata to obtain the consistency coefficient.

[0238] This step further refines the consistency assessment. Certain specific foot physiology characteristics may systematically cause a particular bias in the predictions of one method (geometric or deep learning). The arch index calculation system uses a pre-defined nonlinear mapping (such as a small lookup table or a simple polynomial function) to fine-tune the initial consistency score based on foot physiology characteristics in contextual metadata (such as "presence of hallux valgus"), resulting in a final consistency coefficient. For example, if it is known that there is a systematic bias of about 5% between the two methods for "haunch valgus" foot type, this can be compensated for when assessing consistency.

[0239] S229, when the confidence level of geometric calculation and the confidence level of model inference are both greater than the preset confidence threshold, and the consistency coefficient is greater than or equal to the preset consistency threshold, the results of the first arch index and the second arch index are fused to obtain the final arch index of the subject.

[0240] This final decision-making and fusion step has been described in previous embodiments. The system performs a ternary logical judgment: only when the geometric method is confident, the deep learning method is confident, and the results of the two methods corroborate each other, will the two results be fused through a confidence-weighted average or other methods to output a highly reliable final arch index. Otherwise, an exception handling mechanism will be triggered to ensure that unreliable results are not output.

[0241] The following describes an exemplary arch index calculation system provided by an embodiment of the present invention. Figure 4 This is a schematic diagram of an exemplary hardware structure of the arch index calculation system provided in an embodiment of the present invention.

[0242] In some embodiments, the arch index calculation system is an electronic device, or the arch index calculation system includes an electronic device. The electronic device includes a processor, memory, and a network interface connected via a system bus. The processor of the electronic device provides computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the electronic device stores data. The network interface of the electronic device is used to communicate with other external terminals or servers via a network connection. In some embodiments, the network interface can be a wired network interface; in some embodiments, the network interface can also be a wireless network interface. When the computer program is executed by the processor, it implements the method in the embodiments of the present invention.

[0243] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the electronic device to which the present invention is applied. A specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0244] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

[0245] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as meaning "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0246] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on an electronic device, all or part of the processes or functions described in the embodiments of the present invention are generated. The electronic device may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0247] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for calculating the arch index, characterized in that, include: Three-dimensional point cloud data of the subject's soles were obtained using a three-dimensional foot scanner; The three-dimensional point cloud data of the foot is input into a preset first deep learning model for pre-analysis to generate contextual metadata. The contextual metadata is structured data that characterizes the scanning quality of the three-dimensional point cloud data of the foot itself and the physiological characteristics of the foot it reflects. The first deep learning model is a meta-analysis model used to perform quality assessment and feature classification of the input data. Based on the context metadata, the calculation parameters in the preset parametric geometric algorithm model are dynamically configured to obtain the configured parametric geometric algorithm model; The configured parametric geometric algorithm model is used to process the three-dimensional point cloud data of the foot to obtain the first arch index result, and the geometric calculation confidence of the first arch index result is calculated. The three-dimensional point cloud data of the foot is input into a preset second deep learning model to generate a second arch index result and the corresponding model inference confidence. The second deep learning model is a regression prediction model used to directly map the input data to the target arch index value. Calculate the consistency coefficient between the first arch index result and the second arch index result; When both the geometric calculation confidence and the model inference confidence are greater than a preset confidence threshold, and the consistency coefficient is greater than or equal to a preset consistency threshold, the first arch index result and the second arch index result are fused to obtain the subject's final arch index.

2. The method according to claim 1, characterized in that, The step of dynamically configuring the calculation parameters in the preset parametric geometric algorithm model based on the context metadata includes: From the contextual metadata, scan quality indicators characterizing the scan noise level and foot physiological characteristic indicators characterizing the internal and external rotation angles of the foot are extracted. Based on the scan quality index, the Gaussian filter kernel radius for point cloud smoothing is adaptively determined by querying a preset nonlinear mapping function, wherein a relatively low scan quality index corresponds to a relatively large Gaussian filter kernel radius. Based on the foot physiological characteristic indicators, an affine transformation matrix is ​​constructed to rotate the three-dimensional point cloud data of the foot from the original scanning coordinate system to a target coordinate system aligned with the principal axis of foot anatomy. The three-dimensional point cloud data of the foot is smoothed based on the Gaussian filter kernel radius to obtain smoothed point cloud data. The smoothed point cloud data is then subjected to coordinate transformation based on the affine transformation matrix to generate the final input data to be processed by the configured parametric geometric algorithm model.

3. The method according to claim 2, characterized in that, The process of processing the three-dimensional point cloud data of the foot using the configured parametric geometric algorithm model to obtain the first arch index result includes: Based on the three-dimensional point cloud data of the foot and the affine transformation matrix, pose-normalized point cloud data is generated. The search range for the navicular tuberosity region is determined based on the aforementioned contextual metadata. Based on the search range, the navicular tuberosity is located as the source point in the pose-normalized point cloud data. On the surface formed by the attitude-normalized point cloud data, the geodesic distances from all other points to the source point are calculated to generate a geodesic distance field. The geodesic distance field is input into a preset Gaussian decay function to obtain a soft partition weight mask with the weight value smoothly decaying outwards from the source point. The contact probability density is calculated based on the height value of each data point in the three-dimensional point cloud data of the foot and the probability transition band width parameter dynamically configured by the context metadata. The contact probability density is multiplied by the weight value of the soft partition weight mask corresponding to the data point to obtain the weighted contact probability of the data point; Integrating all the weighted contact probabilities yields the weighted contact probability volume for the core area of ​​the foot arch. The weighted contact probability volume of the core area of ​​the foot arch is compared with the total contact probability volume without using a weighted mask to obtain the first foot arch index result.

4. The method according to claim 1, characterized in that, Before inputting the three-dimensional point cloud data of the foot sole into a preset first deep learning model for pre-analysis, the method further includes: Based on a preset virtual foot standard representing the ideal scanning state, and combined with the current real-time working parameters of the 3D foot scanner, a current state reference point cloud reflecting the current hardware drift state of the 3D foot scanner is generated through the digital twin system of the 3D foot scanner. The current state reference point cloud is registered with the ideal point cloud data of the virtual foot standard to calculate a three-dimensional real-time distortion correction field for compensating for systematic hardware drift. Based on the three-dimensional real-time distortion correction field, point-by-point coordinate compensation is performed on the three-dimensional point cloud data of the foot to obtain the three-dimensional point cloud data of the foot after hardware drift correction.

5. The method according to claim 1, characterized in that, The acquisition of the subject's three-dimensional point cloud data of the soles includes: Based on prior knowledge of foot and ankle anatomy, the plantar projection area of ​​the subject was divided into key physiological areas and non-key areas with different sampling priorities. Based on the expected value of the contour curvature of the key physiological region, the encoding density and complexity of the structured light pattern projected onto the key physiological region are dynamically adjusted. The optical projector and image sensor are controlled to scan the non-critical areas at a baseline sampling rate, while scanning the critical physiological areas at an adaptively enhanced sampling rate, to generate the three-dimensional point cloud data of the foot with adaptively enhanced information encoding and sampling density.

6. The method according to claim 5, characterized in that, The three-dimensional foot scanner is equipped with a thermal imaging sensor; before dividing the subject's plantar projection area into key physiological areas and non-key areas with different sampling priorities based on prior knowledge of foot and ankle anatomy, the method further includes: Before performing a 3D scan, the thermal imaging sensor is used to acquire a real-time thermal image of the subject's sole. Image analysis is performed on the real-time thermal map of the foot to identify one or more thermal anomaly sub-regions with abnormal temperatures; Perform a spatial union operation between the thermal anomaly sub-region and the key physiological region to generate an individualized list of enhanced sampling target regions; Replace the key physiological regions with the individualized list of enhanced sampling target regions.

7. The method according to claim 1, characterized in that, The calculation of the consistency coefficient between the first arch index result and the second arch index result includes: Based on the scan quality indicators and foot physiological characteristic indicators in the contextual metadata, a dynamic consistency weighting factor is calculated. Calculate the absolute difference between the first arch index result and the second arch index result; The absolute difference value is input into a normalization function with the dynamic consistency weighting factor as a parameter for processing to obtain a preliminary consistency score; The initial consistency score is subjected to a nonlinear mapping based on the foot physiological characteristics represented by the contextual metadata to obtain the consistency coefficient.

8. An electronic device, characterized in that, Includes one or more processors and memory; The memory is coupled to the one or more processors, the memory being used to store computer program code, the computer program code including computer instructions, the one or more processors invoking the computer instructions to cause the electronic device to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium storing computer instructions, characterized in that, When the computer instructions are executed on an electronic device, the electronic device causes the electronic device to perform the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on an electronic device, it causes the electronic device to perform the method as described in any one of claims 1-7.

Images