Foot medical image multi-target rapid detection positioning system based on improved YOLO architecture

Abstract

This invention relates to the fields of computer vision, deep learning, and medical image processing technology, specifically to a rapid multi-target detection and localization system for foot medical images based on an improved YOLO architecture. The system comprises four main modules: processing, inference, verification, and output. The processing module performs image compliance matching and sharpness assessment; the inference module loads an optimized YOLOv11 pre-trained model and adapts it to scene parameters; the verification module constructs a complete process system of "confidence screening – anatomical compliance verification – repeatability verification"; and the output module accurately outputs standardized results or anomaly alerts. This invention achieves precise and stable localization of key anatomical points in the foot, with traceable and reusable localization data, meeting the stringent requirements of forensic medicine and judicial identification scenarios for core data in foot arch injury assessment.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, deep learning and medical image processing technology, and more specifically, to a rapid multi-target detection and localization system for foot medical images based on an improved YOLO architecture. Background Technology

[0002] In the fields of forensic medicine, judicial identification, and medical imaging, the assessment of the degree of foot arch injury relies on the precise localization of key anatomical points in foot medical images to obtain reliable coordinate data to support subsequent calculations of the arch angle. This is a core prerequisite for ensuring the scientific and impartial nature of the assessment results. With the application of deep learning technology in medical imaging, traditional physical measurements (such as manual measurement with a set square) are gradually evolving towards automated image localization. However, there are still technical gaps in the accuracy and reliability of multi-target localization in foot medical images.

[0003] Existing technologies for locating key anatomical points of the foot suffer from insufficient accuracy. The core reason lies in the lack of a comprehensive, multi-dimensional verification system in current solutions: Early physical measurement technologies, such as the "Measuring Device for Foot Feature Parts" in patent announcement CN102599915B, rely on manual operation, requiring insertion under the arch of the foot for mechanical measurement. This results in low automation and susceptibility to human error. Image-based detection technologies, such as the "Arch Shape Detection Method and System" in patent announcement CN104366906B, while obtaining arch coordinates from 3D images, only focus on calculating the arch index and do not verify the confidence level or anatomical compliance of key anatomical points. Even deep learning-based arch recognition technologies, such as the "Arch Type Recognition Method" in patent announcement CN110070077B, only determine the arch type based on the slope of the side view contour and do not design a repeatability verification mechanism for key anatomical points of the foot, making the positioning results susceptible to interference from image quality and shooting deviations.

[0004] Specifically, the key problem with existing technologies is that they have not established a comprehensive, multi-dimensional verification system for locating key anatomical points in foot medical imaging, which includes "confidence screening, anatomical compliance verification, and repeatability verification." They have neither screened the confidence level of the location results to eliminate low-confidence false positives, nor verified the relative positions and proportions of anatomical points according to foot anatomy rules, nor confirmed the stability of the location through repeated testing. Ultimately, this results in insufficient accuracy and reliability of the location results, failing to meet the stringent requirements for foot arch injury assessment data in forensic identification scenarios. Summary of the Invention

[0005] In view of this, the present invention proposes a rapid multi-target detection and localization system for foot medical images based on an improved YOLO architecture. It aims to solve the problem that the existing technology does not have a full-process multi-dimensional verification system for the localization of key anatomical points in foot medical images, which includes "confidence screening - anatomical compliance verification - repeatability verification". This results in insufficient localization accuracy and reliability, and fails to meet the stringent requirements for core localization data in forensic medicine and judicial identification scenarios for foot arch injury identification.

[0006] This invention provides a rapid multi-target detection and localization system for foot medical images based on an improved YOLO architecture, comprising: a processing module configured to receive target foot medical images, perform compliance matching judgment through a preset foot contour feature template, perform standardized processing on compliant images, and simultaneously determine whether the image quality meets the detection requirements through a sharpness assessment, thereby obtaining a preprocessed image that meets the model input conditions; The inference module is configured to load a pre-trained model optimized based on YOLOv11, determine the validity of the pre-trained weight file, configure inference parameters adapted to the foot medical image detection scenario, input the pre-processed image into the model to synchronously detect and locate key anatomical points of the foot and output coordinate information and confidence values, and determine whether the confidence of each key anatomical point reaches a preset threshold. The verification module is configured to filter valid positioning points based on the confidence level judgment results, and to judge whether the relative positions and proportions between key anatomical points are compliant based on foot anatomy. At the same time, it judges whether the repeatability of the positioning results meets the standard through multiple repeated tests, and judges whether the positioning results meet the preset accuracy requirements. The output module is configured to verify the positioning results, determine whether to generate a valid positioning report, output standardized valid positioning results or prompts containing abnormal reasons, and complete multi-target detection and positioning of foot medical images.

[0007] In some embodiments, the processing module is configured to receive a target foot medical image, perform compliance matching judgment using a preset foot contour feature template, standardize the compliant image, and simultaneously determine whether the image quality meets the detection requirements through a sharpness assessment, obtaining a preprocessed image that meets the model input conditions, including: Determine whether the format of the target foot medical image is a preset target format. If the format of the target foot medical image is not a preset target format, output a prompt signal and terminate the detection. When the target foot medical image is in the preset target format, the grayscale features and skeletal contour features of the target foot medical image are extracted and used to jointly determine whether the target foot medical image belongs to the foot X-ray image type. When the target foot medical image is not a foot X-ray image, a prompt signal is output and the detection is terminated. When the target foot medical image is a foot X-ray image, the initial matching degree is obtained by comparing the input image with a preset foot contour feature template. At the same time, the contour matching threshold is dynamically adjusted according to the type of image capturing device to obtain the final matching degree threshold. When the initial matching degree is greater than or equal to the final matching degree threshold, it is determined to be a compliant image. When the initial matching degree is less than the final matching degree threshold, it is determined to be a non-target image, and a prompt message containing the specific matching degree value is output and the detection process is terminated.

[0008] In some embodiments, the processing module is configured to receive a target foot medical image, perform compliance matching judgment using a preset foot contour feature template, standardize the compliant image, and simultaneously determine whether the image quality meets the detection requirements through a sharpness assessment. When obtaining a preprocessed image that meets the model input conditions, the module further includes: The edge gradient detection algorithm is used to analyze the skeletal edge details of the compliant image, and the overall gradient mean is calculated as a sharpness evaluation index. When the gradient mean is greater than or equal to a preset gradient threshold, the compliant image is determined to meet the quality standards. When the average gradient value is less than the preset gradient threshold, it is determined whether there is obvious noise in the compliant image. If there is noise in the compliant image, adaptive noise reduction processing is performed first, and then adaptive sharpening processing is started. After sharpening, the presence of local blurred areas in the compliant image is determined by the proportion of areas with local gradient values ​​below the standard threshold. If local blurred areas exist in the compliant image, the contrast and edge strength of those areas are increased. After noise reduction, sharpening, and local enhancement are completed, the gradient mean is calculated a second time. If the gradient mean is still less than the preset gradient threshold, a prompt message containing the coordinate range of the specific blurred area and a comparison of the gradient values ​​before and after processing is output and the process is terminated. When the average gradient value is greater than or equal to the preset gradient threshold, the image quality level is marked and all preprocessing steps are recorded.

[0009] In some embodiments, the inference module is configured to load a pre-trained model optimized based on YOLOv11, determine the validity of the pre-trained weight file, configure inference parameters adapted to the foot medical image detection scenario, input the pre-processed image into the model to simultaneously detect and locate key anatomical points of the foot and output coordinate information and confidence values, and determine whether the confidence of each key anatomical point reaches a preset threshold, including: The determination of the validity of the pre-trained weight file includes: verifying the completeness and format availability of the pre-trained weight file; if the weight file is complete and the format is available, the model is directly loaded and initialized. If the weight file is missing or corrupted, a local backup weight loading mechanism is activated. After loading, it is determined whether the model initialization was successful. If the initialization is successful, a weight switching log is recorded. If initialization fails, check if there are other versions of YOLOv11 pre-trained weights on the local machine. If they exist, try to load them in order of version update time from newest to oldest. After each loading, check the initialization status. If no available backup weights are available locally, an alert signal will be output and the detection will be terminated.

[0010] In some embodiments, during the training process of the pre-trained model, it is first determined whether the training cycle has entered the preset stage of disabling Mosaic data augmentation, and then the convergence of the loss function and the performance index of the validation set are combined for a comprehensive judgment. If the validation set mAP50(B) does not reach the preset threshold, the period for closing Mosaic enhancement will be extended; if the validation set mAP50(B) has reached the threshold, it will be closed according to the original preset stage. Simultaneously, monitor in real time whether there are signs of overfitting during the training process. When signs of overfitting are detected, Mosaic enhancement is turned off. After disabling Mosaic enhancement, determine whether the model's VFL loss has converged stably. If fluctuations still exist, adjust the decay rate of the cosine learning rate.

[0011] In some embodiments, the inference module is configured to load a pre-trained model optimized based on YOLOv11, determine the validity of the pre-trained weight file, configure inference parameters adapted to the foot medical image detection scenario, input the pre-processed image into the model to simultaneously detect and locate key anatomical points of the foot and output coordinate information and confidence values, and when determining whether the confidence of each key anatomical point reaches a preset threshold, it further includes: Determine whether the input preprocessed image needs to be standardized by Letterbox scaling and filling. If the original size of the image is inconsistent with the preset input size ratio of the model, then Letterbox scaling and filling is performed. Then, adjust the confidence threshold based on the noise detection results of the image. When the image noise is less than the preset value, the basic confidence threshold is used. When the image noise is greater than or equal to the preset value, the confidence threshold is increased. At the same time, the overlap of key anatomical points in the image is judged based on the foot anatomy. When there are potential overlapping targets, the NMS threshold is decreased. Otherwise, use the basic NMS threshold.

[0012] In some embodiments, the inference module is configured to load a pre-trained model optimized based on YOLOv11, determine the validity of the pre-trained weight file, configure inference parameters adapted to the foot medical image detection scenario, input the pre-processed image into the model to simultaneously detect and locate key anatomical points of the foot and output coordinate information and confidence values, and when determining whether the confidence of each key anatomical point reaches a preset threshold, it further includes: When the confidence level of a single key anatomical point is greater than or equal to a preset threshold, the coordinates of that point are directly retained and marked as a high-confidence location point. When the confidence level of a single key anatomical point is less than the preset threshold, a local area re-detection process is performed. During the re-detection, the coordinates of the anatomically adjacent points of the key anatomical point are combined for auxiliary localization. At the same time, it is determined whether the number of re-detections exceeds the preset limit. If the number of re-detections exceeds the preset limit, the manual correction interface is directly triggered. The manual correction interface automatically displays an anatomical reference location diagram of the key anatomical point, coordinate markers of adjacent points, and a description of the anatomical function of the node; After manual adjustment, the confidence level of the adjusted coordinates is checked again. If the confidence level is met, the coordinates are retained and the adjustment type is marked as manual correction. If the confidence level still does not meet the standard, mark the node as a location anomaly and record the anomaly node number and confidence level value.

[0013] In some embodiments, the verification module is configured to filter valid positioning points based on confidence level judgment results, and to determine whether the relative positions and proportional relationships between key anatomical points are compliant based on foot anatomy. Simultaneously, it determines whether the repeatability of the positioning results meets the standard through multiple repeated tests. When determining whether the positioning results meet preset accuracy requirements, the module includes: If the spacing ratio and relative position of all nodes are within the range allowed by the rules, it is considered compliant. If there are deviations in the spacing ratio or relative position of all nodes, step-by-step correction shall be adopted; First, correct the positional deviation of adjacent nodes, then correct the relative relationship of non-adjacent nodes. During the correction process, determine whether the adjustment range exceeds the preset safety range. When the adjustment range exceeds the preset safety range, stop automatic correction and trigger manual adjustment. After manual adjustment, the coordinates are checked again to see if they simultaneously meet both the spacing ratio and relative position constraints. If only one of the spacing ratio or relative position is met, a prompt signal is issued. The system is deemed compliant once both constraints are met.

[0014] In some embodiments, the verification module is configured to filter valid positioning points based on confidence level judgment results, and to determine whether the relative positions and proportional relationships between key anatomical points are compliant based on foot anatomy. Simultaneously, it determines whether the repeatability of the positioning results meets the standard through multiple repeated tests. When determining whether the positioning results meet preset accuracy requirements, the module includes: The detection and localization are performed continuously a preset number of times on the same preprocessed image, and the deviation between the coordinates of each key anatomical point in each detection and the average coordinates of multiple detections is calculated using Euclidean distance. When the deviation value of all nodes is less than or equal to the preset accuracy threshold, the repeatability is determined to be up to standard. When the deviation value of all nodes is greater than the preset accuracy threshold, it is determined that the deviation of a single node exceeds the standard or that multiple nodes exceed the standard simultaneously. If only a single node exceeds the deviation limit, then only that node will be restarted for local re-detection; When multiple nodes simultaneously exceed the deviation limit, it is determined whether the problem is a quality issue in a specific area of ​​the image. Preprocessing is then performed again for that area before a full re-inspection is conducted. After retesting, the deviation value is recalculated. If there are still nodes with excessive deviation, the node number with the larger deviation, the specific deviation value, and the influencing factors are output.

[0015] In some embodiments, the output module is configured to determine whether to generate a valid positioning report based on the positioning result verification conclusion, and output a standardized valid positioning result or a prompt message containing the cause of the abnormality. When completing multi-target detection and positioning of foot medical images, the following is included: A standardized and valid localization report is generated when the localization results simultaneously meet the three conditions of confidence level, compliance with foot anatomy, and repeatability. When the report is generated, it is determined whether the user needs a comparison view of the original image and the preprocessed image. If so, a comparison view with key anatomical point location markers is added to the report. And determine whether the purpose of the report is to calculate the arch angle; if so, organize the coordinate information according to the preset data structure. When used only for positioning detection, a basic coordinate report is output.

[0016] Compared with existing technologies, the advantages of this invention are as follows: through the collaborative design of the processing module, inference module, verification module, and output module, it specifically solves the problem of insufficient positioning accuracy and reliability caused by the lack of a full-process, multi-dimensional verification system in existing technologies, thus possessing significant technical advantages. The processing module uses compliance matching and clarity assessment of preset foot contour feature templates to pre-filter non-target images and low-quality images, providing a qualified data foundation for positioning and reducing subsequent interference factors. The inference module loads a pre-trained model optimized based on YOLOv11, ensuring stable model operation through weight validity judgment, and simultaneously detecting key anatomical points with scene-adaptive inference parameter configuration, achieving a dual improvement in positioning efficiency and initial accuracy. The core lies in the verification module's construction of a full-process system of "confidence screening - anatomical compliance verification - repeatability verification." It first screens high-confidence positioning points, then verifies the position and proportional relationship according to foot anatomy rules, and combines multiple repeated tests to verify stability, avoiding positioning deviations from multiple dimensions. The output module accurately outputs standardized results or specific anomaly prompts based on the verification conclusions, ensuring that the positioning data is traceable and reusable, and ultimately meeting the accuracy and reliability requirements of core positioning data for foot arch injury identification in forensic and judicial appraisal scenarios.

[0017] The above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure.

[0018] Other features and aspects of this disclosure will become clearer from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 Functional block diagram of a multi-target rapid detection and localization system for foot medical images based on an improved YOLO architecture provided in an embodiment of the present invention; Figure 2 A flowchart of a rapid multi-target detection and localization system for foot medical images based on an improved YOLO architecture, provided in an embodiment of the present invention; Figure 3 A schematic diagram of key components of a multi-target rapid detection and localization system for foot medical images based on an improved YOLO architecture, provided in an embodiment of the present invention. Detailed Implementation

[0021] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] See Figures 1-3 As shown, a multi-target rapid detection and localization system for foot medical images based on an improved YOLO architecture, according to an embodiment of this application, includes: The processing module is configured to receive medical images of the target foot, perform compliance matching judgment through a preset foot contour feature template, standardize the compliant images, and determine whether the image quality meets the detection requirements through a clarity assessment, thereby obtaining a preprocessed image that meets the model input conditions. The inference module is configured to load a pre-trained model optimized based on YOLOv11, determine the validity of the pre-trained weight file, configure inference parameters adapted to the foot medical image detection scenario, input the pre-processed image into the model to synchronously detect and locate key anatomical points of the foot and output coordinate information and confidence values, and determine whether the confidence of each key anatomical point reaches a preset threshold. The verification module is configured to filter valid positioning points based on the confidence level judgment results, and to judge whether the relative positions and proportions between key anatomical points are compliant based on foot anatomy. At the same time, it judges whether the repeatability of the positioning results meets the standard through multiple repeated tests, and judges whether the positioning results meet the preset accuracy requirements. The output module is configured to verify the positioning results, determine whether to generate a valid positioning report, output standardized valid positioning results or prompts containing abnormal reasons, and complete multi-target detection and positioning of foot medical images.

[0023] It should be understood that the improved YOLO architecture is based on targeted optimization of the YOLOv11 model to meet the feature extraction requirements of foot medical images and improve the detection sensitivity and positioning accuracy of key anatomical points. Key anatomical points of the foot: These are the core points in the arch structure used to calculate the arch angle (6 in total, denoted as P1-P6), including points with clearly defined anatomical features such as the junction of the metatarsals and tarsals and key points of the calcaneus; P1 is the lowest point of contact between the head of the first metatarsal and the horizontal plane; P2 is the lowest point of contact between the head of the fifth metatarsal and the horizontal plane; P3 is the lowest point of the first tarsometatarsal joint; P4 is the lowest point of the calcaneal joint; P5 is the lowest point of the head of the talus; and P6 is the lowest point of contact between the calcaneus and the horizontal plane. In the preprocessing stage, the system uses pre-defined anatomical feature templates for these six points, combined with grayscale features and skeletal contour features, to accurately locate candidate regions for each point. In the inference stage, the optimized YOLOv11 model, trained on a labeled dataset containing these six specific points, can selectively extract local anatomical features (such as joint spaces and bone end-point morphology) for each point, enabling simultaneous detection and localization of the six specific points, and outputting accurate coordinates and confidence values. P1-P6 represent red, orange, yellow, green, cyan, and blue, respectively. The specific anatomical definitions of P1-P6 are clearly defined, making the system's target location more precise and unambiguous, avoiding positioning deviations caused by vague point definitions. The combination of exclusive anatomical feature templates and targeted training of the model significantly improves the detection sensitivity and positioning accuracy of each point, ensuring that the coordinate data of each core point conforms to medical anatomy standards. It provides a precise and unified source of basic data for subsequent foot arch angle calculation, ensuring the scientific validity and reliability of the angle calculation results, and further meeting the stringent requirements of forensic medicine and judicial identification for core data.

[0024] Preprocessed images: Foot medical images that meet the model input format and quality requirements after compliance screening and quality optimization; Preset accuracy requirements: The positioning error threshold (such as coordinate deviation ≤ 0.5mm) set according to the judicial appraisal standards is the core basis for judging whether the positioning result is valid.

[0025] Processing module: Includes compliance matching unit, standardization processing unit, and sharpness assessment unit, which operate sequentially; first, qualified images are screened through preset templates, then the image size / format is standardized, and finally the image quality is optimized to ensure the reliability of the data input to the model; Inference module: Composed of model loading unit, parameter configuration unit, and localization calculation unit, it loads the optimized YOLOv11 pre-trained model, dynamically adjusts the detection parameters according to image characteristics, and outputs the coordinates of the anatomical points and the confidence level (reflecting the reliability of the localization results). The verification module includes a confidence screening unit, an anatomical compliance unit, and a repeatability verification unit, which verify the localization results from three dimensions: "localization reliability, anatomical rationality, and result stability". Output module: Includes report generation unit and exception prompt unit. Based on the verification conclusion, it selects to output a valid report or exception information to form a closed loop of the process.

[0026] In a forensic identification scenario, after staff upload an X-ray image of an injured person's foot, the system starts running: The processing module first determines that the image format is DICOM (preset target format), and confirms that it is a foot X-ray image through joint recognition of grayscale features and skeletal contours. After contour matching (initial matching degree 85%, dynamic threshold corresponding to DR imaging equipment 70%), it is determined to be a compliant image. Then, the gradient mean is calculated to be 0.82 (preset threshold 0.7) through edge gradient detection. After the quality meets the standard, a pre-processed image is generated. The inference module loads the optimized YOLOv11 pre-trained model, verifies that the original file is complete and usable, and configures Letterbox ablation. The parameters are set (the image is standardized to 640×640), and after inputting into the model, the coordinate information and confidence scores (both ≥0.6) of P1-P6 are output simultaneously. The verification module selects high-confidence positioning points and verifies that the distance ratio between each point conforms to the rules of foot anatomy (e.g., the distance ratio between P2-P3 and P3-P4 is 1.2:1, within the allowable range of 1.0-1.3:1). After three repeated tests, the coordinate deviation is calculated to be ≤0.3mm (preset accuracy threshold 0.5mm), and the positioning result is judged to meet the standard. The output module generates a standardized and valid positioning report, which includes the coordinates, confidence scores and image comparison views of each anatomical point, to meet the needs of subsequent foot arch angle calculation.

[0027] The four modules work together to build a fully automated detection system that can complete image screening, positioning and verification without human intervention, greatly improving detection efficiency and avoiding human error. With the improved YOLOv11 model as the core and combined with targeted parameter configuration, the detection rate and positioning accuracy of key anatomical points of the foot are improved to meet the stringent requirements of forensic identification for data accuracy. A multi-dimensional verification mechanism eliminates interference from low-quality images and low-confidence positioning points from the source, ensuring that the positioning results are both reliable and reasonable. The output module is adapted to different use cases (foot arch angle calculation, simple positioning detection), enhancing the system's practicality and flexibility; The overall technical solution enables the traceability and verification of positioning data, providing a scientific basis for the assessment of the degree of foot arch injury and ensuring the fairness and validity of the assessment results.

[0028] In this application, the training process of the pre-trained model based on YOLOv11 optimization is finely controlled throughout the entire lifecycle of the foot medical imaging detection scenario. Specifically, the following steps are taken: First, the dataset is customized and prepared by collecting foot X-ray images taken by different devices such as DR machines, CT machines, and mobile X-ray machines. After selecting valid samples according to the forensic identification image standards, pixel-level coordinate annotations are performed on 6 core foot anatomical points (P1-P6, covering key points of the metatarsals, tarsals, and calcaneus) using a unified annotation specification. The annotation format is adapted to the YOLOv11 txt tag format (including category number, center point X / Y coordinates, and aspect ratio). The training set, validation set, and test set are divided in a 7:2:1 ratio. At the same time, all images are preprocessed by format conversion (unified to DICOM to PNG) and pixel value normalization (reduced to the 0-1 range). Then, the training environment and base are initialized. The basic configuration uses an optimized YOLOv11 network structure built on the PyTorch framework. The original C2f feature fusion module is retained, but the number of channels in the shallow convolutional layers is reduced from 64 to 32 for lightweight computation. A new 160×160 small-scale feature map output branch is added to capture minute anatomical points. Anchor box sizes (10×10, 15×15, 20×20) adapted to foot anatomical points are recalculated based on labeled data to replace the original general anchor boxes. YOLOv11 official pre-trained weights are loaded as initial weights. Basic training hyperparameters are set as follows: initial cosine learning rate 0.001, batch size 16, total training epochs 100. The VFL loss function is selected, and the focal length coefficient is initialized to 1.5 (to balance positive and negative sample weights). The last 20% (80-100 epochs) of the training cycle is designated as the Mosaic data augmentation shutdown phase.The training process is dynamically optimized in stages: From 0 to 80 epochs (preset Mosaic enabled stage), Mosaic enhancement is enabled (four different foot images are randomly scaled, cropped, and stitched together into one training sample). After each epoch, mAP50(B) (the core indicator for foot anatomical point detection) is calculated on the validation set. If the accuracy of the training set continues to rise but the mAP50(B) of the validation set decreases for three consecutive epochs (indicating overfitting), Mosaic enhancement is immediately disabled. If the mAP50(B) of the validation set does not reach the preset threshold at 80 epochs... If (0.9), the Mosaic shutdown phase is extended to 70-100 epochs, while the VFL loss curve is monitored in real time. If the VFL loss fluctuation exceeds 10% within 5 epochs after Mosaic shutdown (not stably converged), the cosine learning rate decay rate is adjusted from 0.001 to 0.0005 (to slow down the learning rate decrease), and the VFL loss focal length coefficient is adjusted to 2.0 to further increase the attention weight of low-confidence positive samples (fuzzy dissection points); after 80 / 70 epochs, the later training phase with Mosaic shutdown begins, and the number of shutdowns... Enhanced fine-grained learning, focusing on anatomical point features, reduces the learning rate to 1 / 10 of the initial value, strengthens gradient update weights for small-scale feature branches, and ensures precise matching between anchor boxes and anatomical point coordinates. After each epoch of training, the model performance is comprehensively evaluated on the validation set. In addition to mAP50(B), the false negative rate, false positive rate, and average deviation of localization coordinates of anatomical points are calculated. If the indicators do not meet the standards, the anchor box size or VFL loss parameters are adjusted and training is iterated again. After training, the training weight files for each stage are saved, and the weights with the highest mAP50(B) on the validation set (usually ≥0.92) are selected. The main pre-trained weights are used as the primary weights, while the weights from three different training stages are backed up as backup weights. All saved weights are checked for integrity (compare file byte size and calculate MD5 checksum) and format availability (import the optimized network structure to verify parameter dimension matching) to ensure that there are no missing parameters or format incompatibility issues during loading. The resulting optimized pre-trained model improves the mAP50(B) of foot anatomical point detection by 10%-15% compared to the native YOLOv11, and reduces the false negative rate and false positive rate by 8% and 12% respectively, fully meeting the accuracy requirements of forensic identification scenarios.

[0029] In some specific embodiments, the processing module is configured to receive target foot medical images, perform compliance matching judgments using preset foot contour feature templates, standardize compliant images, and simultaneously determine whether the image quality meets detection requirements through sharpness assessment. When obtaining a preprocessed image that meets the model input conditions, the following steps are included: Determine whether the format of the target foot medical image is a preset target format. If the format of the target foot medical image is not a preset target format, output a prompt signal and terminate the detection. When the target foot medical image is in the preset target format, the grayscale features and skeletal contour features of the target foot medical image are extracted and used to jointly determine whether the target foot medical image belongs to the foot X-ray image type. When the target foot medical image is not a foot X-ray image, a prompt signal is output and the detection is terminated. When the target foot medical image is a foot X-ray image, the initial matching degree is obtained by comparing the input image with a preset foot contour feature template. At the same time, the contour matching threshold is dynamically adjusted according to the type of image capturing device to obtain the final matching degree threshold. When the initial matching degree is greater than or equal to the final matching degree threshold, it is determined to be a compliant image. When the initial matching degree is less than the final matching degree threshold, it is determined to be a non-target image, and a prompt message containing the specific matching degree value is output and the detection process is terminated.

[0030] It should be understood that the system has the function of simultaneously judging the left and right feet, and the bilateral detection process is completely consistent with the core function. In the compliance matching stage, the processing module adds a step of left and right foot contour feature recognition: through the preset left and right foot exclusive contour templates (constructed based on the differences in the left and right anatomical structures of the human foot, such as the symmetrical differences in the arch curvature and calcaneal position), combined with the joint analysis of grayscale features and skeletal contours, the system automatically identifies the input image as a left foot, right foot, or bilateral foot image; for bilateral images, the system adopts a parallel processing mechanism, independently processing the left and right foot images according to the same detection process (preprocessing, inference, and verification), loading the same optimized YOLOv11 pre-trained model, inference parameters (confidence threshold, NMS threshold), and verification standards (anatomical rules, repeatability accuracy threshold); the output module generates standardized positioning reports for the left and right feet for bilateral images, or displays the bilateral positioning results, angle data, and verification conclusions in the same report in sections; It achieves full coverage detection of images of the left foot, right foot, and both feet, eliminating the need for manual image splitting or switching detection modes, thus improving detection efficiency and ease of operation. The left and right feet employ completely consistent detection standards and procedures, ensuring the comparability and impartiality of bilateral positioning results and avoiding deviations caused by differences in detection parameters. It adapts to the needs of bilateral arch injury comparison assessment in forensic identification, providing unified standard core data for assessing the degree of bilateral injury, further expanding the system's application scope. The parallel processing mechanism ensures consistency in bilateral detection functions without reducing detection speed, balancing practicality and efficiency.

[0031] Preset target formats: The system pre-sets supported image formats, including the commonly used DICOM format for medical imaging and the common JPG and PNG formats, to ensure the compatibility of image data; Gray-scale features: the brightness distribution characteristics of different regions in an image. In foot X-ray images, the gray-scale value of the bone region is significantly higher than that of the soft tissue region, which is a key feature for distinguishing image types. Skeletal contour features: Skeletal edge morphological features extracted by edge detection algorithm. The foot bones have clear tarsal and metatarsal contours, which are significantly different from images of other parts of the body. Dynamically adjust the contour matching threshold: Set different matching thresholds according to the type of imaging equipment (such as DR machine, CT machine, mobile X-ray machine) to adapt to the imaging characteristics of different equipment (such as mobile X-ray machine with lower imaging clarity, the threshold is lowered by 10-15%).

[0032] Step 1: Format Verification. The compliance matching unit of the processing module reads the image file header information and determines whether the format belongs to the preset target format. If it is an unsupported format such as BMP or TIFF, the process is terminated directly. Step 2: Type Recognition. For images with valid formats, grayscale features are extracted through grayscale histogram analysis, and skeletal contour features are extracted through the Canny edge detection algorithm. The two are then jointly input into a classification model (such as a lightweight CNN model) to determine whether it is a foot X-ray image. Step 3: Contour Matching. Load a preset standard foot contour template (built based on anatomical data), calculate the initial matching degree between the input image and the template using a template matching algorithm (such as normalized cross-correlation algorithm), and simultaneously call a preset threshold library according to the shooting device type to obtain the final matching degree threshold. If the initial matching degree meets the standard, it is determined to be a compliant image; otherwise, the process is terminated.

[0033] A forensic appraisal institution uploaded an image file of an injured person. The processing module initiated compliance matching: First, it detected that the file format was DICOM (a preset target format), and then proceeded to the type identification stage. Through grayscale histogram analysis, it was found that the image contained obvious high-density skeletal areas (grayscale values ​​180-220) and low-density soft tissue areas (grayscale values ​​80-120), which matched the grayscale characteristics of foot X-ray images. Then, the complete contours of the tarsal and metatarsal bones were extracted using the Canny algorithm, and the matching degree with the foot contour feature database reached 92%, confirming that it was a foot X-ray image. The preset standard foot contour template was called for matching, and an initial matching degree of 88% was obtained. Since the image was taken by a DR machine (preset basic threshold 75%), no threshold adjustment was required, and the final matching degree of 88% ≥ 75% was determined to be a compliant image, which then proceeded to the subsequent standardization processing stage. If the uploaded image was a hand X-ray image, the matching degree between the skeletal contour features and the foot contour in the type identification stage was only 45%, and the system output a "non-foot X-ray image" prompt, terminating the detection.

[0034] The three-level screening mechanism accurately filters out non-target images with incorrect format, type, or outline, preventing invalid data from entering subsequent processes and improving system operating efficiency. The combined recognition of grayscale features and skeletal contour features significantly improves the recognition accuracy of foot X-ray images and reduces misjudgments (such as avoiding misjudging hand and ankle images as foot images). The matching threshold is dynamically adjusted based on the type of shooting device to adapt to the characteristics of different imaging devices, reduce compliance misjudgments caused by device differences, and enhance the system's adaptability. The output includes anomaly prompts with specific matching degree values, which helps staff quickly locate problems (e.g., a matching degree of 65% < 70% can be identified as an image shooting angle deviation), improving operational convenience. Compliance screening provides a high-quality data foundation for subsequent clarity assessment and location detection, ensuring the accuracy of location results from the source.

[0035] In some specific embodiments, the processing module is configured to receive target foot medical images, perform compliance matching judgments using preset foot contour feature templates, standardize compliant images, and simultaneously determine whether the image quality meets detection requirements through sharpness assessment. When obtaining a preprocessed image that meets the model input conditions, the module further includes: The edge gradient detection algorithm is used to analyze the skeletal edge details of the compliant image, and the overall gradient mean is calculated as a sharpness evaluation index. When the gradient mean is greater than or equal to a preset gradient threshold, the compliant image is determined to meet the quality standards. When the average gradient value is less than the preset gradient threshold, it is determined whether there is obvious noise in the compliant image. If there is noise in the compliant image, adaptive noise reduction processing is performed first, and then adaptive sharpening processing is started. After sharpening, the presence of local blurred areas in the compliant image is determined by the proportion of areas with local gradient values ​​below the standard threshold. If local blurred areas exist in the compliant image, the contrast and edge strength of those areas are increased. After noise reduction, sharpening, and local enhancement are completed, the gradient mean is calculated a second time. If the gradient mean is still less than the preset gradient threshold, a prompt message containing the coordinate range of the specific blurred area and a comparison of the gradient values ​​before and after processing is output and the process is terminated. When the average gradient value is greater than or equal to the preset gradient threshold, the image quality level is marked and all preprocessing steps are recorded.

[0036] It should be understood that edge gradient detection algorithms (such as the Sobel algorithm) are algorithms used to extract edge details of images. They reflect the edge sharpness by calculating the rate of change (gradient) of pixel gray values. The larger the gradient value, the sharper the edge. Gradient mean: The average gradient value of all pixels in an image. It is a core indicator for quantifying the overall sharpness of an image. The higher the gradient mean, the sharper the overall image. Adaptive noise reduction: A processing method that dynamically adjusts noise reduction parameters based on the noise distribution characteristics of the image (such as median filtering algorithm), which can remove Gaussian noise and salt-and-pepper noise while preserving edge details; Local enhancement processing: Targeted optimization of local blurred areas in the image (such as blurred areas at the edges of bones) to enhance the recognition of local details by improving regional contrast and edge strength.

[0037] Sharpness assessment: The Sobel algorithm is used to calculate the gradient value of compliant images, and the average gradient value of all pixels is calculated (gradient mean). This average value is then compared with a preset gradient threshold to determine whether the image quality meets the standard. Low-quality image optimization: If the gradient mean is lower than the preset threshold, first use the adaptive median filtering algorithm to detect noise areas (such as isolated high grayscale pixels with salt and pepper noise) and remove noise in a targeted manner; then use the Laplacian sharpening algorithm to enhance edge details (calculate the grayscale difference between the pixel and the neighboring pixels to enhance edge contrast); finally use the local histogram equalization algorithm to improve the contrast and edge strength of the local blurred areas. Secondary verification: After optimization, the gradient mean is recalculated. If it is still lower than the preset threshold, an error message is output and the process is terminated. If it meets the standard, the image quality level is marked (excellent, good, qualified) and the processing steps are recorded.

[0038] The processing module performs a clarity assessment on compliant foot X-ray images: the Sobel algorithm is used to calculate the image gradient value, and the average gradient value is 0.65, while the preset gradient threshold is 0.7, indicating that the image quality is substandard; further detection reveals slight salt-and-pepper noise (isolated pixels with abrupt changes in grayscale value), so adaptive median filtering is initiated, and the window size is dynamically adjusted to 3×3-5×5 to accurately remove noise while preserving bone edge details; then, the Laplacian sharpening algorithm is used to enhance the edges, increasing the edge gradient value of the tarsal and metatarsal bones by 30%; finally, the calcaneal region (coordinate range: X[200-300], Y

[15] ) is detected in the image. If the local gradient value is less than 0.5, the contrast of the area is increased by 40% and the edge intensity is increased by 25% by local histogram equalization. The average gradient value is 0.78, which is higher than the preset threshold of 0.7. The image quality level is marked as "good". The parameters and steps of noise reduction, sharpening and local enhancement are recorded to generate a preprocessed image. If the average gradient value is still 0.68 < 0.7, the system outputs the prompt "The calcaneal region (X[200-300], Y[150-250]) is blurred. The average gradient value before processing is 0.65 and after processing is 0.68. The preset threshold of 0.7 has not been reached" and the process is terminated.

[0039] A quantitative evaluation standard with gradient mean as its core makes the clarity assessment more objective and accurate, avoiding subjective judgment errors; The step-by-step optimization scheme of "noise reduction-sharpening-local enhancement" specifically addresses issues such as image noise, edge blurring, and local low contrast, significantly improving image quality. Adaptive noise reduction and local enhancement techniques optimize quality while preserving the original anatomical details to the maximum extent, avoiding feature distortion caused by over-processing; The second gradient mean verification ensures that the quality of the preprocessed image fully meets the standards, providing a clear feature basis for the accurate positioning of the subsequent inference module and reducing the positioning deviation caused by image blur. Marking image quality levels and recording processing steps facilitates staff in tracing the image optimization process and provides a reference for adjusting parameters in the inference module (e.g., the confidence threshold can be appropriately increased for images with a "qualified" level).

[0040] In some specific embodiments, the inference module is configured to load a pre-trained model optimized based on YOLOv11, determine the validity of the pre-trained weight file, configure inference parameters adapted to the foot medical image detection scenario, input the pre-processed image into the model to simultaneously detect and locate key anatomical points of the foot and output coordinate information and confidence values, and determine whether the confidence of each key anatomical point reaches a preset threshold, including: The determination of the validity of the pre-trained weight file includes: verifying the completeness and format availability of the pre-trained weight file; if the weight file is complete and the format is available, the model is directly loaded and initialized. If the weight file is missing or corrupted, a local backup weight loading mechanism is activated. After loading, it is determined whether the model initialization was successful. If the initialization is successful, a weight switching log is recorded. If initialization fails, check if there are other versions of YOLOv11 pre-trained weights on the local machine. If they exist, try to load them in order of version update time from newest to oldest. After each loading, check the initialization status. If no available backup weights are available locally, an alert signal will be output and the detection will be terminated.

[0041] It should be understood that the pre-trained weight file: the parameter file obtained after the YOLOv11 model is trained on the foot medical image dataset (such as yolo11m_foot.pt), contains key parameters for model feature extraction and localization prediction, which determine the model's detection performance; Weighted file integrity: The byte size and file checksum (such as MD5 value) of the weighted file are consistent with the preset standard, and there is no missing, corrupt or tampered file; Format availability: The parameter format of the weight file is compatible with the network layer structure of the YOLOv11 model (e.g., the parameter dimensions of convolutional layers and the output dimensions of fully connected layers match), and the model can be successfully loaded and initialized; Local backup weight loading mechanism: The system pre-stores at least one set of backup weight files (such as yolo11m_foot_backup.pt) that are functionally identical to the sovereign weight file in the local storage medium. When the sovereign weight fails, the fault tolerance mechanism is automatically triggered to load the backup weight.

[0042] Sovereignty Revalidation: The model loading unit of the inference module reads the sovereignty revalidation file, first verifying the file integrity (comparing the actual byte size with the preset size, calculating the MD5 value and comparing it with the standard value), and then verifying the format availability (importing the weight parameters into the YOLOv11 model and verifying whether the network layer initialization was successful). Backup weight loading: If the sovereign weight file is missing, corrupted, or incompatible in format, the local backup weight loading mechanism will be automatically started to load the first set of backup weights and verify the initialization result. If successful, the weight switching log will be recorded (including switching time, reason for sovereign weight failure, and backup weight version). Multiple version traversal: If the first set of backup weights fails to initialize, check if other versions of YOLOv11 pre-trained weights are stored locally, load and verify them in order of version update time from newest to oldest, until a weight file that can be successfully initialized is found. Termination mechanism: If all backup weights fail to complete initialization, an error message will be output and the detection will be terminated.

[0043] The inference module initiates the model loading process as follows: First, it reads the sovereign refactoring file yolo11m_foot.pt, verifies that the file size is 28.5MB (consistent with the preset size), calculates the MD5 value, and confirms the file integrity. Then, it imports the weight parameters into the YOLOv11 model, verifies that the parameter dimensions of the convolutional layer and the detection head layer match, and confirms successful model initialization, proceeding normally to the subsequent parameter configuration stage. If the sovereign refactoring file is damaged due to storage media corruption, resulting in a file size of only 10MB (inconsistent with the preset 28.5MB), the file is deemed corrupted, and the local backup weight loading mechanism is automatically activated, loading the first set of backup weights. Reload yolo11m_foot_backup.pt. After verification, the model initialization is successful, and the log is recorded as "2025-XX-XX 14:30, the primary weight file is corrupted, switch to backup weight version V2.0". If the first set of backup weights fails to initialize due to a format error, the system detects that there are also backup weights of versions V1.8 and V1.6 on the local machine. They are loaded in order of update time from newest to oldest. After version V1.8 is initialized successfully, the process continues. If all three sets of backup weights fail to initialize, the system outputs the message "Model weight loading error, no available valid weight file" and terminates the detection.

[0044] Double verification (integrity + format availability) ensures that the loaded weight file is valid, avoiding performance degradation or localization errors caused by weight issues. The local backup weight loading mechanism provides fault tolerance, significantly reducing the probability of detection process interruption due to sovereign weight failure and improving system stability; The multi-version weight traversal design further enhances fault tolerance and adapts to weight compatibility issues in different scenarios (such as some spare weights still being usable after model version upgrade). Weight switching logs facilitate technical personnel in tracing problems, promptly fixing the primary weight or updating the backup weight, and ensuring the long-term stable operation of the system. Clear anomaly alerts allow staff to quickly locate weight loading issues, reducing troubleshooting time and improving work efficiency; Stable model loading provides a foundation for subsequent accurate positioning, ensuring that the system can continue to output reliable results even in complex environments.

[0045] In some specific embodiments, during the training process of the pre-trained model, it is first determined whether the training cycle has entered the preset stage of disabling Mosaic data augmentation, and then the convergence of the loss function and the performance index of the validation set are combined for a comprehensive judgment. If the validation set mAP50(B) does not reach the preset threshold, the period for closing Mosaic enhancement will be extended; if the validation set mAP50(B) has reached the threshold, it will be closed according to the original preset stage. Simultaneously, monitor in real time whether there are signs of overfitting during the training process. When signs of overfitting are detected, Mosaic enhancement is turned off. After disabling Mosaic enhancement, determine whether the model's VFL loss has converged stably. If fluctuations still exist, adjust the decay rate of the cosine learning rate.

[0046] It should be understood that Mosaic data augmentation is an image stitching enhancement technique that stitches four different foot medical images together into a new image at random proportions, thereby expanding the diversity of the training dataset and improving the model's generalization ability (avoiding overfitting to images with specific shooting angles and sizes). Validation set mAP50(B): The average accuracy of the model on the validation set, where the IOU (Intersection over Union) threshold is set to 50%, and "B" specifically refers to the foot key anatomical point detection task, which is the core indicator for quantifying the model's localization accuracy (the higher the value, the more accurate the localization, ranging from 0 to 1). Signs of overfitting: The model's accuracy on the training set continues to improve, but the mAP50(B) on the validation set begins to decline, indicating that the model is overlearning noise in the training set and its generalization ability is decreasing. VFL loss: Variable Focal Loss is used to solve the problem of imbalance between positive and negative samples in object detection. The lower the loss value, the closer the model prediction result is to the true label. Stable convergence of VFL loss indicates that the model training effect has met the target. Cosine learning rate: The learning rate is adjusted according to the cosine function during the training cycle. The learning rate is higher in the early stage to accelerate parameter updates, and gradually decreases in the later stage to stabilize model performance. The decay rate determines how fast the learning rate decreases.

[0047] Mosaic enhancement dynamic shutdown: The last 20% of the preset training cycle is the Mosaic enhancement shutdown phase (e.g., if the total training is 100 epochs, the 80-100 epochs are the shutdown phase). During training, the validation set mAP50(B) is monitored in real time. If it does not reach the preset threshold (e.g., 0.9), the shutdown phase is extended (e.g., extended to 70-100 epochs); if it has reached the threshold, it is shut down according to the original preset phase. Overfitting monitoring and handling: Real-time comparison of training set accuracy and validation set mAP50(B). When the validation set mAP50(B) decreases for 3 consecutive epochs, it is determined that there is an overfitting sign, and Mosaic enhancement is immediately turned off to reduce the interference of data augmentation on the model. VFL Loss Adjustment: After disabling Mosaic enhancement, continuously monitor the VFL loss curve. If the loss value fluctuates by more than 10% within 5 epochs (not converging stably), adjust the decay rate of the cosine learning rate (e.g., from 0.001 to 0.0005) to reduce the learning rate decrease and allow the model more time to optimize parameters.

[0048] The training process of the improved YOLOv11 model was set to a total of 100 epochs, with the 80-100 epochs pre-defined as the Mosaic enhancement phase disabled. The preset threshold for the validation set mAP50(B) was 0.9, and the fluctuation threshold for stable convergence of the VFL loss was 10%. At 75 epochs, the validation set mAP50(B) had reached 0.92 (≥0.9), and Mosaic enhancement was disabled at 80 epochs as originally preset. At 60 epochs, the training set accuracy improved from 92% to 95%, but the validation set mAP50(B) decreased from 0.88 to 0.85, a decrease over three consecutive epochs, indicating overfitting. Mosaic enhancement was disabled. After disabling it, the VFL loss curve was monitored. It was found that the loss value fluctuated from 0.32 to 0.41 within 5 epochs, with a fluctuation range of 28.1% (>10%), and did not converge stably. The decay rate of the cosine learning rate was adjusted from 0.001 to 0.0005. After the adjustment, the loss value gradually stabilized between 0.30 and 0.33, with a fluctuation range of <10%, and the model training was successful. If the validation set mAP50(B) was only 0.87 (<0.9) at 79 epochs, the Mosaic enhancement disabled phase was extended to 70-100 epochs. Training continued until mAP50(B) reached 0.91 at 85 epochs, and the training was completed.

[0049] Dynamically disable the Mosaic enhancement strategy to balance the generalization ability and positioning accuracy of the model, avoiding premature disabling leading to insufficient generalization, or delayed disabling affecting accurate positioning; Real-time overfit monitoring and timely disabling of Mosaic enhancement effectively suppress model overfitting and improve the model's generalization ability in real-world detection scenarios (such as adapting to images under different shooting conditions). VFL loss adjustment optimizes the cosine learning rate decay rate to ensure stable convergence of model parameters and improve the model's sensitivity and accuracy in identifying key anatomical points of the foot. Refined training and control enable the model to be specifically adapted to the characteristics of foot medical images, significantly improving the mAP50(B) index and ensuring that the positioning results meet the accuracy requirements of forensic identification. Quantitative monitoring of training process metrics (mAP50(B), VFL loss) provides a clear basis for model optimization, avoids blind training, and improves training efficiency. The optimized model has stronger environmental adaptability and can meet the detection needs of different shooting equipment, shooting angles and image quality, thus expanding the application scope of the system.

[0050] In some specific embodiments, the inference module is configured to load a pre-trained model optimized based on YOLOv11, determine the validity of the pre-trained weight file, configure inference parameters adapted to the foot medical image detection scenario, input the pre-processed image into the model to simultaneously detect and locate key anatomical points of the foot and output coordinate information and confidence values, and when determining whether the confidence of each key anatomical point reaches a preset threshold, it further includes: Determine whether the input preprocessed image needs to be standardized by Letterbox scaling and filling. If the original size of the image is inconsistent with the preset input size ratio of the model, then Letterbox scaling and filling is performed. Then, adjust the confidence threshold based on the noise detection results of the image. When the image noise is less than the preset value, the basic confidence threshold is used. When the image noise is greater than or equal to the preset value, the confidence threshold is increased. At the same time, the overlap of key anatomical points in the image is judged based on the foot anatomy. When there are potential overlapping targets, the NMS threshold is decreased. Otherwise, use the basic NMS threshold.

[0051] It should be understood that Letterbox scaling fill is a standardized processing method that maintains the aspect ratio of an image. It first scales the image according to the input size of the model, and then fills the blank areas with gray pixels (to avoid image stretching and distortion), ensuring that the anatomical proportions of the image remain unchanged. Confidence threshold: The probability threshold for judging whether the location result is valid. If it is higher than the threshold, it is considered a real dissection point; if it is lower than the threshold, it is judged as a false detection (e.g., the basic threshold is 0.5, that is, the location probability is ≥50% to retain it). NMS threshold: The threshold for non-maximum suppression, used to filter overlapping bounding boxes. The lower the threshold, the stronger the filtering of overlapping boxes (to avoid the same anatomical point being detected multiple times). Potential overlapping targets: Due to the image shooting angle or anatomical structure characteristics, the location frames of different key anatomical points may partially overlap (e.g., P3 and P4 are close to each other, and the overlap area of ​​the location frames is >30%).

[0052] Letterbox Scaling Fill: Preset the model input size (e.g., 640×640), calculate the difference between the aspect ratio of the preprocessed image and the model input aspect ratio, scale the image by the minimum scaling ratio (e.g., image 800×600, scaling ratio 0.8, scaled to 640×480), and then fill the blank area (80 pixels at the top and bottom) with gray pixels (grayscale value 128) to obtain a 640×640 standardized image; Confidence threshold adjustment: The noise density (ratio of noise pixels) of the image is detected by Gaussian filtering algorithm. If the noise ratio is <5% (less than the preset value), the basic confidence threshold (e.g., 0.5) is used; if the noise ratio is ≥5% (greater than or equal to the preset value), the threshold is increased to 0.6-0.7 to reduce false detections caused by noise. NMS threshold adjustment: Based on foot anatomical data (the distance range between key anatomical points), predict the probability of overlap of anatomical points in the image. If there are potential overlapping targets (such as the expected overlap area of ​​the localization boxes of P3 and P4 > 30%), reduce the NMS threshold to 0.2-0.3; if there is no risk of overlap, use the basic NMS threshold (such as 0.4).

[0053] The inference module configures the parameters of the preprocessed image: the preset model input size is 640×640, the original size of the preprocessed image is 960×720 (aspect ratio 4:3), the model input aspect ratio is 1:1, the scaling ratio is calculated as 640 / 960=0.667, the scaled image size is 640×480, and 80 pixels of gray (grayscale value 128) are filled in the upper and lower blank areas to obtain a 640×640 standardized image; the noise density of the image is detected by Gaussian filtering, the noise pixel ratio is 3% (<5% preset value), and a basic confidence threshold of 0.5 is used; Based on foot anatomy data, the distance between P3 (midpoint of metatarsal bone) and P4 (midpoint of tarsal bone) is 15mm, the bounding box size is 10×10mm, the predicted overlap area is about 10% (<30%), there are no potential overlapping targets, and the basic NMS threshold of 0.4 is used; if the image noise density is 7% (≥5%), the confidence threshold is increased to 0.65; if due to the shooting angle, the distance between P3 and P4 is only 8mm, the overlap area of ​​the bounding box is expected to reach 40% (>30%), the NMS threshold is reduced to 0.25 to avoid the same anatomical point being detected multiple times or the overlapping point being misjudged.

[0054] Letterbox scaling fill maintains anatomical proportions while standardizing image size, avoiding positioning deviations caused by image stretching (such as positioning point shifts after metatarsal length is stretched). The confidence threshold is dynamically adjusted based on noise density. When there is less noise, more potential location points are retained, and when there is more noise, false detection points are filtered out, thus balancing detection sensitivity and accuracy. Adjusting the NMS threshold for overlapping targets effectively solves the problem of multiple detections or missed detections caused by overlapping dissection points, thus improving the accuracy of localization results; The dynamic adaptability enhancement model of parameter configuration improves the system's ability to adapt to images of different qualities and shooting conditions, thus expanding the system's application scenarios. Standardized parameter configuration processes reduce manual intervention, improve the level of automation in detection, and ensure the consistency of positioning results (unified detection standards for different images). Precise parameter control enables the model to output reliable positioning results stably even in complex scenarios (such as high noise and overlapping dissection points), meeting the stringent requirements of forensic identification.

[0055] In some specific embodiments, the inference module is configured to load a pre-trained model optimized based on YOLOv11, determine the validity of the pre-trained weight file, configure inference parameters adapted to the foot medical image detection scenario, input the pre-processed image into the model to simultaneously detect and locate key anatomical points of the foot and output coordinate information and confidence values, and when determining whether the confidence of each key anatomical point reaches a preset threshold, it further includes: When the confidence level of a single key anatomical point is greater than or equal to a preset threshold, the coordinates of that point are directly retained and marked as a high-confidence location point. When the confidence level of a single key anatomical point is less than the preset threshold, a local area re-detection process is performed. During the re-detection, the coordinates of the anatomically adjacent points of the key anatomical point are combined for auxiliary localization. At the same time, it is determined whether the number of re-detections exceeds the preset limit. If the number of re-detections exceeds the preset limit, the manual correction interface is directly triggered. The manual correction interface automatically displays an anatomical reference location diagram of the key anatomical point, coordinate markers of adjacent points, and a description of the anatomical function of the node; After manual adjustment, the confidence level of the adjusted coordinates is checked again. If the confidence level is met, the coordinates are retained and the adjustment type is marked as manual correction. If the confidence level still does not meet the standard, mark the node as a location anomaly and record the anomaly node number and confidence level value.

[0056] It should be understood that high-confidence localization points: the localization confidence of a single key anatomical point is ≥ a preset threshold (e.g., 0.5), indicating that the model has high reliability in locating the point and the coordinates can be directly retained; Local re-detection: For low-confidence positioning points, the detection algorithm is re-run only in a local area of ​​the point (such as a 50×50 pixel range centered on the initial positioning coordinates) to focus on key areas and improve positioning accuracy; Anatomically adjacent points: Key points that are anatomically adjacent to low-confidence localization points (e.g., P1 and P3 are the neighbors of P2), whose coordinates have a clear anatomical relationship (e.g., fixed spacing and angle), which can help correct the coordinates of low-confidence points. Manual correction interface: An interactive interface for staff to manually adjust the coordinates of positioning points, integrating anatomical reference data and visualization tools for easy and precise correction.

[0057] High confidence processing: After the model outputs the coordinates and confidence of each anatomical point, points with confidence ≥ preset threshold are selected and marked as high confidence positioning points, and their coordinates are directly retained; Low confidence handling: For points with confidence scores < preset thresholds, local re-detection is initiated; a local area is delineated centered on the initial positioning coordinates, and the detection algorithm is rerun in combination with the coordinate constraints of anatomically adjacent points (e.g., the distance between P2 and P1 should be between 12-14mm). If the confidence score meets the standard after re-detection, the coordinates are retained. Manual correction as a backup: If the number of local re-tests exceeds the preset limit (e.g., 3 times) and still fails to meet the standard, the manual correction interface is triggered. The interface displays an anatomical reference location diagram of the point (based on the standard foot anatomy diagram), coordinate markings of adjacent points, and an anatomical function description of the node (e.g., "P2 is the connection between the second metatarsal and the intermediate cuneiform, which is a key point for medial arch support"). After the staff manually adjusts the coordinates, the confidence level is checked again. If it meets the standard, it is marked as a manual correction positioning point; if it does not meet the standard, it is marked as a positioning abnormal point and the information is recorded.

[0058] The inference module outputs the coordinates and confidence scores of six key anatomical points on the foot. A pre-set confidence threshold of 0.5 and a maximum of three re-detections are allowed. The confidence scores for P1-P5 are 0.72, 0.68, 0.55, 0.63, and 0.58, respectively, all ≥0.5, and are marked as high-confidence points, with their coordinates directly retained. The confidence score for P6 is 0.42 < 0.5, triggering a local re-detection. A 50×50 pixel local region is defined centered on the initial coordinates of P6 (X320, Y280). Combined with the coordinates of the neighboring point P5 (X300, Y260), and based on anatomical constraints (the distance between P5 and P6 should be between 10-12 mm), a re-detection is performed. The confidence score for the first re-detection is 0.46, and for the second… The confidence level is 0.49 for the first test and 0.51 (≥0.5) for the third test, and the P6 coordinates are retained. If the confidence level of P6 is still 0.48 (<0.5) for the third retest, the manual correction interface is triggered. The interface displays the anatomical reference position of P6 (calcaneal tuberosity), the P5 coordinate marker, and the explanation that "P6 is a key point of the calcaneal tuberosity, participating in the posterior support of the foot arch, and is 10-12mm away from P5". The staff manually adjusts the P6 coordinates to (X318, Y290) and re-verifies the confidence level to 0.53 (≥0.5), marking it as a manually corrected positioning point. If the confidence level is still 0.47 (<0.5) after manual adjustment, P6 is marked as a positioning abnormal point, and "P6 node confidence level 0.47, not reaching the preset threshold of 0.5" is recorded.

[0059] The strategy of directly retaining high-confidence points and stratifying low-confidence points maximizes the effectiveness of location points while ensuring detection efficiency. Local re-detection focuses on key areas and combines anatomical proximity point constraints to specifically improve the localization accuracy of low-confidence points, avoiding the efficiency decline caused by full-map re-detection; The artificial correction interface provides anatomical references and visualization tools, reducing the difficulty of manual operation, ensuring the scientific nature of manual adjustments (conforming to anatomical rules), and avoiding blind adjustments; Recording abnormal locations facilitates subsequent analysis of problems by staff (e.g., P6 location abnormalities may be due to blurring of the calcaneal region in the image), and provides a basis for image re-taking or system optimization. The layered processing mechanism significantly reduces the impact of low-confidence false positives on the final result, improves the reliability of the location results, and meets the requirements of forensic identification for data accuracy. The manual correction fallback solution enhances the system's flexibility, adapts to extreme scenarios (such as extremely poor image quality or abnormal anatomical structures), and ensures that the system can still output complete positioning results.

[0060] In some specific embodiments, the verification module is configured to filter valid positioning points based on confidence level judgment results, and determine whether the relative positions and proportional relationships between key anatomical points are compliant based on foot anatomy. Simultaneously, it determines whether the repeatability of the positioning results meets the standard through multiple repeated tests. When determining whether the positioning results meet preset accuracy requirements, the following steps are included: If the spacing ratio and relative position of all nodes are within the range allowed by the rules, it is considered compliant. If there are deviations in the spacing ratio or relative position of all nodes, step-by-step correction shall be adopted; First, correct the positional deviation of adjacent nodes, then correct the relative relationship of non-adjacent nodes. During the correction process, determine whether the adjustment range exceeds the preset safety range. When the adjustment range exceeds the preset safety range, stop automatic correction and trigger manual adjustment. After manual adjustment, the coordinates are checked again to see if they simultaneously meet both the spacing ratio and relative position constraints. If only one of the spacing ratio or relative position is met, a prompt signal is issued. The system is deemed compliant once both constraints are met.

[0061] It should be understood that foot anatomy rules are: key anatomical point constraints based on human anatomical data, including the range of distance ratios between nodes (e.g., the distance ratio between P1-P2 and P2-P3 is 1.1-1.3:1) and relative positional relationships (e.g., the angle between the line connecting P1-P3 and the horizontal direction is 30-35°). Step-by-step correction: First, correct the positional deviation of adjacent nodes (adjacent nodes have a more direct anatomical relationship and are easier to correct), and then, based on the corrected adjacent nodes, correct the relative relationship of non-adjacent nodes (the constraints of non-adjacent nodes depend on the accuracy of adjacent nodes). Preset safety range: To avoid overcorrection causing the positioning point to deviate from its true position, the upper limit of the single adjustment range is set (e.g., the coordinate deviation should not exceed 2mm in a single adjustment). Two constraints: the spacing ratio constraint and the relative position constraint. Both must be satisfied for the anatomical compliance to be determined.

[0062] Compliance judgment: Extract the coordinates of valid positioning points after the verification module has filtered them, calculate the spacing ratio and relative position (angle, parallel relationship, etc.) between each node, and compare them with the preset foot anatomy rules. If all nodes meet the constraints, it is judged to be compliant. Step-by-step correction: If deviation exists, first correct adjacent nodes (such as P1-P2, P2-P3); calculate the deviation value (e.g., if the spacing ratio of P2-P3 is 1.5:1, exceeding the range of 1.1-1.3:1, the deviation value is 0.2), adjust the node coordinates proportionally, and each adjustment should not exceed the preset safety range (2mm); after the adjacent nodes are corrected, correct the non-adjacent nodes (such as P1-P4), recalculate the constraint relationship based on the corrected adjacent node coordinates, and adjust the position of the non-adjacent nodes; Manual adjustment and verification: If the adjustment range exceeds the safe range during automatic correction (e.g., the P3 coordinate needs to be adjusted by 3mm), automatic correction will be stopped and manual adjustment will be triggered. After manual adjustment by the staff with reference to the anatomical rules, the spacing ratio and relative position will be verified again to see if they meet the constraints. If only one of them is met, a specific prompt will be issued (e.g., "The spacing ratio of P2-P3 meets the standard, but the relative position angle is deviated"), until both constraints are met.

[0063] The verification module performs anatomical compliance verification on the valid positioning points, with preset foot anatomy rules: the distance ratio between P1-P2 and P2-P3 is 1.1-1.3:1, and the relative position angle between P2-P3 and P3-P4 is 30-35°. The calculated distance between P1-P2 is 13mm, and the distance between P2-P3 is 10mm, with a ratio of 1.3:1 (within the allowable range). The angle between P2-P3 and P3-P4 is 28° (below 30°, deviating by 2°), which is judged as a partial deviation. Step-by-step correction is initiated. First, the coordinates of adjacent nodes P3-P4 are confirmed. The calculation shows that P4 needs to be moved 0.8mm to the right horizontally to achieve an angle of 31°. The adjustment range is 0.8mm < 2mm (safe range). After automatic correction, the angle is 31°, satisfying the constraints. If P2 - The P3 spacing ratio is 1.6:1 (deviation of 0.3). The P3 coordinate needs to be adjusted by 2.5mm (exceeding the safe range of 2mm). Automatic correction stops and manual adjustment is triggered. The interface displays "P2-P3 spacing ratio 1.6:1, exceeding the range of 1.1-1.3:1, P3 coordinate needs to be adjusted by 2.5mm (exceeding the safe range)". The staff manually adjusts the P3 coordinate by 2mm with reference to the anatomical diagram. The ratio is checked again and found to be 1.4:1 (still deviating). Continue to fine-tune by 0.3mm. The final ratio is 1.3:1 with an angle of 32°. Both constraints are met and it is judged to be compliant. If the ratio meets the standard after manual adjustment but the angle is still 27°, the system prompts "P2-P3 relative position angle 27°, not within the 30-35° range", until the adjustment meets the standard.

[0064] Compliance verification based on foot anatomy rules ensures that the positioning points conform to human physiological structure and avoids positioning results that are "anatomically impossible" (such as the distance between P1 and P2 being far beyond the normal range). The step-by-step correction strategy starts with the easy nodes and progresses to the difficult ones. It utilizes the direct constraint relationship between adjacent nodes to improve the correction accuracy and then extends to non-adjacent nodes to ensure the logic and reliability of the correction process. Pre-set a safety range to avoid overcorrection and prevent the positioning point from deviating from the actual anatomical position due to correction, thus balancing the correction effect and positioning accuracy; The specific prompts for manual adjustments allow staff to quickly locate the problem and make precise adjustments by combining anatomical reference diagrams, thus reducing the difficulty of operation. The dual verification of the two constraints ensures the comprehensiveness of anatomical compliance, avoids the situation where a single constraint meets the standard but the overall structure is unreasonable, and improves the rationality of the localization results; Anatomical compliance verification provides physiological structural assurance for the localization results, ensuring that the data is not only accurate but also in line with medical common sense, thereby enhancing the persuasiveness of forensic identification results.

[0065] In some specific embodiments, the verification module is configured to filter valid positioning points based on confidence level judgment results, and determine whether the relative positions and proportional relationships between key anatomical points are compliant based on foot anatomy. Simultaneously, it determines whether the repeatability of the positioning results meets the standard through multiple repeated tests. When determining whether the positioning results meet preset accuracy requirements, the following steps are included: The detection and localization are performed continuously a preset number of times on the same preprocessed image, and the deviation between the coordinates of each key anatomical point in each detection and the average coordinates of multiple detections is calculated using Euclidean distance. When the deviation value of all nodes is less than or equal to the preset accuracy threshold, the repeatability is determined to be up to standard. When the deviation value of all nodes is greater than the preset accuracy threshold, it is determined that the deviation of a single node exceeds the standard or that multiple nodes exceed the standard simultaneously. If only a single node exceeds the deviation limit, then only that node will be restarted for local re-detection; When multiple nodes simultaneously exceed the deviation limit, it is determined whether the problem is a quality issue in a specific area of ​​the image. Preprocessing is then performed again for that area before a full re-inspection is conducted. After retesting, the deviation value is recalculated. If there are still nodes with excessive deviation, the node number with the larger deviation, the specific deviation value, and the influencing factors are output.

[0066] It should be understood that the preset number of tests is the number of tests set to verify repeatability (e.g., 3 times). The number of tests should take into account both the verification effect and efficiency. Too many tests will increase the time consumption, while too few tests will not be able to eliminate random errors. Euclidean distance: The straight-line distance between two points, used to calculate the deviation of the coordinates of the same anatomical point in multiple inspections (e.g., in the first inspection, P1 coordinates are (X1, Y1), and in the second inspection, they are (X2, Y2). Euclidean distance = √[(X2-X1)]. 2 +(Y2-Y1) 2]), is a common method for quantifying coordinate deviation; Preset accuracy threshold: The upper limit of allowable deviation (e.g., 0.5mm). If the deviation of multiple tests of all nodes is ≤ this threshold, the repeatability is judged to be up to standard. Local re-inspection: Only the single node with excessive deviation is re-inspected, eliminating the need for full map re-inspection and improving verification efficiency; Full Re-inspection: Re-inspects all nodes, suitable for scenarios where multiple nodes have excessive deviations (usually due to image quality issues).

[0067] Repeated detection: Perform localization detection on the same preprocessed image a preset number of times (3 times), and record the coordinates of each key anatomical point in each detection; Deviation calculation: The Euclidean distance formula is used to calculate the deviation between the coordinates of each anatomical point and the average coordinates in three detections (e.g., the three coordinates of P1 are (X1,Y1), (X2,Y2), and (X3,Y3), and the average coordinates are ((X1+X2+X3) / 3, (Y1+Y2+Y3) / 3). The deviation value is the maximum Euclidean distance between the three coordinates and the average coordinates). Case-by-case handling: If the deviation values ​​of all nodes are ≤ the preset accuracy threshold (0.5mm), the repeatability is considered to be up to standard; if there are nodes with deviations exceeding the standard, it is determined whether it is a single node or multiple nodes exceeding the standard; if a single node exceeds the standard, only the node is re-detected locally and the deviation is recalculated; if multiple nodes exceed the standard, it is determined to be a quality problem in a specific area of ​​the image (such as a blurred area), and the area is re-preprocessed (noise reduction, enhancement) and then a full re-detection is performed to recalculate the deviation; if there are still nodes exceeding the standard after re-detection, an abnormal information is output.

[0068] The verification module performs three repeated positioning checks on the preprocessed image, with a preset accuracy threshold of 0.5mm: After recording the three coordinates of P1-P6, the deviation value of each point from the average coordinate is calculated. The deviation values ​​of P1-P5 are 0.2mm, 0.3mm, 0.25mm, 0.35mm, and 0.4mm, respectively, all ≤0.5mm. The deviation value of P6 is 0.65mm >0.5mm, which is judged as a single node deviation exceeding the standard. Only P6 is subjected to local re-detection. After repositioning, the deviation value of P6 is 0.42mm ≤0.5mm, which is judged as repeatability meeting the standard. If P3, P... 4. The deviation values ​​of P5 are 0.6mm, 0.7mm, and 0.55mm, respectively, all > 0.5mm, indicating that multiple nodes exceed the standard. Detection revealed that these nodes are all located in the blurred calcaneal area of ​​the image. After re-performing local enhancement processing on this area, a full re-detection was performed. After the re-detection, the deviation values ​​of the three nodes are all ≤ 0.45mm, meeting the standard. If the deviation value of P4 is still 0.58mm > 0.5mm after the full re-detection, the system outputs "P4 node has a deviation of 0.58mm after multiple detections, exceeding the preset accuracy threshold of 0.5mm. Influencing factor: poor image quality in the calcaneal area".

[0069] Repeated detection and bias calculation effectively eliminate random errors (such as accidental fluctuations in model detection), ensure the stability of positioning results, and avoid the impact of accidental biases from a single detection on the final conclusion. The quantitative calculation of Euclidean distance makes the deviation assessment more objective and accurate, avoids subjective judgment errors, and meets the quantitative standard requirements of forensic identification. The case-specific retesting strategy (partial / full) improves verification efficiency while ensuring verification effectiveness and avoids unnecessary full-image retesting. To address image quality issues at multiple nodes that exceed standards, a closed loop of "re-verification - quality optimization - re-detection" is formed to resolve the root cause of the unstable positioning problem. Clear anomaly alerts allow staff to quickly locate unstable nodes and influencing factors, facilitating targeted optimization (such as image re-capture or adjustment of preprocessing parameters). Repeatability standards ensure that positioning results remain consistent across multiple tests, enhancing the reliability and credibility of the data and providing stable core data support for the identification of foot arch injuries.

[0070] In some specific embodiments, the output module is configured to determine whether to generate a valid positioning report based on the positioning result verification conclusion, and output a standardized valid positioning result or a prompt message containing the cause of the abnormality. When completing multi-target detection and positioning of foot medical images, this includes: A standardized and valid localization report is generated when the localization results simultaneously meet the three conditions of confidence level, compliance with foot anatomy, and repeatability. When the report is generated, it is determined whether the user needs a comparison view of the original image and the preprocessed image. If so, a comparison view with key anatomical point location markers is added to the report. And determine whether the purpose of the report is to calculate the arch angle; if so, organize the coordinate information according to the preset data structure. When used only for positioning detection, a basic coordinate report is output.

[0071] It should be understood that, based on the precise coordinate data of P1-P6, the system automatically calculates four key arch angles: the medial longitudinal arch angle Angle1 (the angle formed by the coordinates of points P1, P3, and P5), the lateral longitudinal arch angle Angle2 (the angle formed by the coordinates of points P2, P4, and P6), the anterior arch angle Angle3 (the angle formed by the coordinates of points P1, P2, and P3), and the posterior arch angle Angle4 (the angle formed by the coordinates of points P4, P5, and P6). During the calculation process, the system first converts pixel coordinates into actual physical coordinates through coordinate transformation, and then uses the vector angle calculation formula to accurately solve for each angle value. The output module determines the purpose of the report. If the requirement is for calculating the arch angle, it automatically adds the specific values ​​of the four angles, the calculation basis (coordinates of the points involved), and angle diagrams to the standardized and effective positioning report. It also supports exporting angle data according to preset data structures (such as CSV and Excel formats), which can be directly adapted to subsequent identification and analysis software. There is no need for manual angle calculation, realizing full automation from point positioning to angle output, which significantly shortens the identification cycle and reduces human calculation errors. The four key angles comprehensively cover the core assessment indicators of arch injury identification, providing identification personnel with direct and reliable quantitative data support, and improving the scientificity and impartiality of the identification results. The angle data is output in conjunction with point coordinates and image views, which makes it easy for identification personnel to trace the calculation process and enhances the verifiability of the data. It adapts to the data analysis needs of different identification scenarios, improving the practicality and adaptability of the system. Three criteria for achieving the target: These are necessary prerequisites for generating a valid localization report, namely, the confidence level of all key anatomical points must be ≥ the preset threshold (confidence level meets the target), the relative position and proportion must conform to the rules of foot anatomy (anatomical compliance), and the deviation of multiple repeated tests must be ≤ the preset accuracy threshold (repeatability meets the target). Image comparison view: This view includes a side-by-side view of the original image and the preprocessed image, and marks the location of key anatomical points (such as different colored dots marking P1-P6, with coordinate values), which makes it easy for staff to intuitively compare the image optimization effect and location. Preset data structure: Adapted to the coordinate data format for foot arch angle calculation, including the X / Y coordinates of each anatomical point, confidence value, and positioning reliability rating (high confidence / manual correction), which can be directly imported into foot arch angle calculation software (such as Matlab, professional medical image analysis system). Basic coordinates report: A simplified report containing only the coordinates of key anatomical points, confidence values, and verification results. It is suitable for simple localization detection scenarios that do not require angle calculations (such as anatomical point location confirmation).

[0072] Report generation judgment: The output module receives the conclusion of the verification module. If all three compliance conditions are met at the same time, the standardized effective location report generation is initiated; if any one condition is not met, a prompt message containing the specific reason for the anomaly is output. Personalized Adaptation: When generating the report, the system prompts the user for their needs through the interactive interface. If the user requires it, an image comparison view is added (the original image and the preprocessed image are side by side, with positioning markers). Then, the purpose of the report is determined. If it is used for foot arch angle calculation, the coordinate information is organized according to a preset data structure (such as CSV format, with fields including "node number, X coordinate, Y coordinate, confidence level, and reliability rating"). If it is only used for positioning detection, a basic coordinate report is generated (with fields including "node number, X coordinate, Y coordinate, and verification result"). Output results: Based on user needs and intended use, output the corresponding report file (PDF / Excel format) or error message to complete the detection process.

[0073] The output module receives the verification results: the confidence level of P1-P6 is ≥0.5 (confidence level meets the standard), the anatomical spacing ratio and relative position both conform to the rules (anatomical compliance), and the deviation of the three repeated tests is ≤0.45mm (repeatability meets the standard). Meeting these three conditions, the generation of a valid localization report is initiated. The system asks the user if they need an image comparison view; if the user selects "yes," a side-by-side view of the original and preprocessed images is added to the report, with the localization positions of P1-P6 marked with red dots and the coordinates of each point (e.g., P1: X300, Y250). Further determination of the report's purpose is then made. When the user selects "Arch Angle Calculation," the system organizes the data according to a preset data structure (CSV format), including fields such as "Node Number, X Coordinate, Y Coordinate, Confidence Level, and Reliability Rating," which can be directly imported into the arch angle calculation software. If the user selects "Simple Positioning Detection," a basic coordinate report is generated, which only includes the coordinates, confidence levels, and "Verification Passed" results for each node. If the verification conclusion is that the confidence level of P6 is 0.47 < 0.5 (confidence level not met), the system outputs the message "P6 node confidence level 0.47, not reaching the preset threshold of 0.5, unable to generate a valid positioning report."

[0074] The three compliance criteria strictly control the threshold for generating valid reports, ensuring that the output location data is accurate, compliant, and stable, providing a reliable basis for forensic identification. Image comparison views allow staff to intuitively verify the rationality of the location and the effect of image optimization, enhancing the traceability and persuasiveness of the report; The pre-defined data structure enables seamless integration between positioning data and foot arch angle calculation software, eliminating the need for manual formatting, improving work efficiency, and avoiding data errors caused by format conversion. Differentiated report outputs (angle calculation adaptation report / basic coordinate report) meet the needs of different use cases, enhancing the system's practicality and flexibility; The standardized report format (PDF / Excel) facilitates file storage, transmission, and archiving, and meets the document management requirements for forensic identification. Specific error messages allow staff to quickly pinpoint the root cause of the problem and take targeted optimization measures (such as retesting or manual correction) to improve problem-solving efficiency. The completeness and personalization of the report enhance the user experience of the system and adapt to the operational needs of different users (such as professional technicians who need detailed data and ordinary staff who need intuitive views).

[0075] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0076] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0077] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0078] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A rapid multi-target detection and localization system for foot medical images based on an improved YOLO architecture, characterized in that, include: The processing module is configured to receive medical images of the target foot, perform compliance matching judgment through a preset foot contour feature template, standardize the compliant images, and determine whether the image quality meets the detection requirements through a clarity assessment, thereby obtaining a preprocessed image that meets the model input conditions. The inference module is configured to load a pre-trained model optimized based on YOLOv11, determine the validity of the pre-trained weight file, configure inference parameters adapted to the foot medical image detection scenario, input the pre-processed image into the model to synchronously detect and locate key anatomical points of the foot and output coordinate information and confidence values, and determine whether the confidence of each key anatomical point reaches a preset threshold. The verification module is configured to filter valid positioning points based on the confidence level judgment results, and to judge whether the relative positions and proportions between key anatomical points are compliant based on foot anatomy. At the same time, it judges whether the repeatability of the positioning results meets the standard through multiple repeated tests, and judges whether the positioning results meet the preset accuracy requirements. The output module is configured to verify the positioning results, determine whether to generate a valid positioning report, output standardized valid positioning results or prompts containing abnormal reasons, and complete multi-target detection and positioning of foot medical images.

2. The foot medical imaging multi-target rapid detection and localization system based on the improved YOLO architecture according to claim 1, characterized in that, The processing module is configured to receive target foot medical images, perform compliance matching judgments using preset foot contour feature templates, standardize compliant images, and simultaneously determine whether the image quality meets detection requirements through sharpness assessment. When obtaining a preprocessed image that meets the model input conditions, the module includes: Determine whether the format of the target foot medical image is a preset target format. If the format of the target foot medical image is not a preset target format, output a prompt signal and terminate the detection. When the target foot medical image is in the preset target format, the grayscale features and skeletal contour features of the target foot medical image are extracted and used to jointly determine whether the target foot medical image belongs to the foot X-ray image type. When the target foot medical image is not a foot X-ray image, a prompt signal is output and the detection is terminated. When the target foot medical image is a foot X-ray image, the initial matching degree is obtained by comparing the input image with a preset foot contour feature template. At the same time, the contour matching threshold is dynamically adjusted according to the type of image capturing device to obtain the final matching degree threshold. When the initial matching degree is greater than or equal to the final matching degree threshold, it is determined to be a compliant image. When the initial matching degree is less than the final matching degree threshold, it is determined to be a non-target image, and a prompt message containing the specific matching degree value is output and the detection process is terminated.

3. The foot medical image multi-target rapid detection and localization system based on the improved YOLO architecture according to claim 2, characterized in that, The processing module is configured to receive target foot medical images, perform compliance matching judgments using preset foot contour feature templates, standardize compliant images, and simultaneously determine whether the image quality meets detection requirements through sharpness assessment. When obtaining a preprocessed image that meets the model input conditions, the module further includes: The edge gradient detection algorithm is used to analyze the skeletal edge details of the compliant image, and the overall gradient mean is calculated as a sharpness evaluation index. When the gradient mean is greater than or equal to a preset gradient threshold, the compliant image is determined to meet the quality standards. When the average gradient value is less than the preset gradient threshold, it is determined whether there is obvious noise in the compliant image. If there is noise in the compliant image, adaptive noise reduction processing is performed first, and then adaptive sharpening processing is started. After sharpening, the presence of local blurred areas in the compliant image is determined by the proportion of areas with local gradient values ​​below the standard threshold. If local blurred areas exist in the compliant image, the contrast and edge strength of those areas are increased. After noise reduction, sharpening, and local enhancement are completed, the gradient mean is calculated a second time. If the gradient mean is still less than the preset gradient threshold, a prompt message containing the coordinate range of the specific blurred area and a comparison of the gradient values ​​before and after processing is output and the process is terminated. When the average gradient value is greater than or equal to the preset gradient threshold, the image quality level is marked and all preprocessing steps are recorded.

4. The foot medical image multi-target rapid detection and localization system based on the improved YOLO architecture according to claim 3, characterized in that, The inference module is configured to load a pre-trained model optimized based on YOLOv11, determine the validity of the pre-trained weight file, configure inference parameters adapted to the foot medical image detection scenario, input the pre-processed image into the model to simultaneously detect and locate key anatomical points of the foot and output coordinate information and confidence values, and determine whether the confidence of each key anatomical point reaches a preset threshold, including: The determination of the validity of the pre-trained weight file includes: verifying the completeness and format availability of the pre-trained weight file; if the weight file is complete and the format is available, the model is directly loaded and initialized. If the weight file is missing or corrupted, a local backup weight loading mechanism is activated. After loading, it is determined whether the model initialization was successful. If the initialization is successful, a weight switching log is recorded. If initialization fails, check if there are other versions of YOLOv11 pre-trained weights on the local machine. If they exist, try to load them in order of version update time from newest to oldest. After each loading, check the initialization status. If no available backup weights are available locally, an alert signal will be output and the detection will be terminated.

5. A rapid multi-target detection and localization system for foot medical images based on an improved YOLO architecture according to claim 4, characterized in that, During the training process of the pre-trained model, it is first determined whether the training cycle has entered the preset stage of disabling Mosaic data augmentation, and then the convergence of the loss function and the performance index of the validation set are combined for a comprehensive judgment. If the validation set mAP50(B) does not reach the preset threshold, the period for closing Mosaic enhancement will be extended; if the validation set mAP50(B) has reached the threshold, it will be closed according to the original preset stage. Simultaneously, monitor in real time whether there are signs of overfitting during the training process. When signs of overfitting are detected, Mosaic enhancement is turned off. After disabling Mosaic enhancement, determine whether the model's VFL loss has converged stably. If fluctuations still exist, adjust the decay rate of the cosine learning rate.

6. A rapid multi-target detection and localization system for foot medical images based on an improved YOLO architecture according to claim 5, characterized in that, The inference module is configured to load a pre-trained model optimized based on YOLOv11, determine the validity of the pre-trained weight file, configure inference parameters adapted to the foot medical image detection scenario, input the pre-processed image into the model to simultaneously detect and locate key anatomical points of the foot and output coordinate information and confidence values, and when determining whether the confidence of each key anatomical point reaches a preset threshold, it also includes: Determine whether the input preprocessed image needs to be standardized by Letterbox scaling and filling. If the original size of the image is inconsistent with the preset input size ratio of the model, then Letterbox scaling and filling is performed. Then, adjust the confidence threshold based on the noise detection results of the image. When the image noise is less than the preset value, the basic confidence threshold is used. When the image noise is greater than or equal to the preset value, the confidence threshold is increased. At the same time, the overlap of key anatomical points in the image is judged based on the foot anatomy. When there are potential overlapping targets, the NMS threshold is decreased. Otherwise, use the basic NMS threshold.

7. A rapid multi-target detection and localization system for foot medical images based on an improved YOLO architecture according to claim 6, characterized in that, The inference module is configured to load a pre-trained model optimized based on YOLOv11, determine the validity of the pre-trained weight file, configure inference parameters adapted to the foot medical image detection scenario, input the pre-processed image into the model to simultaneously detect and locate key anatomical points of the foot and output coordinate information and confidence values, and when determining whether the confidence of each key anatomical point reaches a preset threshold, it also includes: When the confidence level of a single key anatomical point is greater than or equal to a preset threshold, the coordinates of that point are directly retained and marked as a high-confidence location point. When the confidence level of a single key anatomical point is less than the preset threshold, a local area re-detection process is performed. During the re-detection, the coordinates of the anatomically adjacent points of the key anatomical point are combined for auxiliary localization. At the same time, it is determined whether the number of re-detections exceeds the preset limit. If the number of re-detections exceeds the preset limit, the manual correction interface is directly triggered. The manual correction interface automatically displays an anatomical reference location diagram of the key anatomical point, coordinate markers of adjacent points, and a description of the anatomical function of the node; After manual adjustment, the confidence level of the adjusted coordinates is checked again. If the confidence level is met, the coordinates are retained and the adjustment type is marked as manual correction. If the confidence level still does not meet the standard, mark the node as a location anomaly and record the anomaly node number and confidence level value.

8. A rapid multi-target detection and localization system for foot medical images based on an improved YOLO architecture according to claim 7, characterized in that, The verification module is configured to filter valid positioning points based on confidence level judgment results, and to determine whether the relative positions and proportions between key anatomical points comply with foot anatomy. It also checks the repeatability of the positioning results through multiple repeated tests. When determining whether the positioning results meet preset accuracy requirements, the module includes: If the spacing ratio and relative position of all nodes are within the range allowed by the rules, it is considered compliant. If there are deviations in the spacing ratio or relative position of all nodes, step-by-step correction shall be adopted; First, correct the positional deviation of adjacent nodes, then correct the relative relationship of non-adjacent nodes. During the correction process, determine whether the adjustment range exceeds the preset safety range. When the adjustment range exceeds the preset safety range, stop automatic correction and trigger manual adjustment. After manual adjustment, the coordinates are checked again to see if they simultaneously meet both the spacing ratio and relative position constraints. If only one of the spacing ratio or relative position is met, a prompt signal is issued. The system is deemed compliant once both constraints are met.

9. A rapid multi-target detection and localization system for foot medical images based on an improved YOLO architecture according to claim 8, characterized in that, The verification module is configured to filter valid positioning points based on confidence level judgment results, and to determine whether the relative positions and proportions between key anatomical points comply with foot anatomy. It also checks the repeatability of the positioning results through multiple repeated tests. When determining whether the positioning results meet preset accuracy requirements, the module includes: The detection and localization are performed continuously a preset number of times on the same preprocessed image, and the deviation between the coordinates of each key anatomical point in each detection and the average coordinates of multiple detections is calculated using Euclidean distance. When the deviation value of all nodes is less than or equal to the preset accuracy threshold, the repeatability is determined to be up to standard. When the deviation value of all nodes is greater than the preset accuracy threshold, it is determined that the deviation of a single node exceeds the standard or that multiple nodes exceed the standard simultaneously. If only a single node exceeds the deviation limit, then only that node will be restarted for local re-detection; When multiple nodes simultaneously exceed the deviation limit, it is determined whether the problem is a quality issue in a specific area of ​​the image. Preprocessing is then performed again for that area before a full re-inspection is conducted. After retesting, the deviation value is recalculated. If there are still nodes with excessive deviation, the node number with the larger deviation, the specific deviation value, and the influencing factors are output.

10. A rapid multi-target detection and localization system for foot medical images based on an improved YOLO architecture, as described in claim 9, is characterized in that... The output module is configured to determine whether to generate a valid positioning report based on the positioning result verification conclusion, and output standardized valid positioning results or prompt information containing the cause of abnormality. When completing multi-target detection and positioning of foot medical images, it includes: A standardized and valid localization report is generated when the localization results simultaneously meet the three conditions of confidence level, compliance with foot anatomy, and repeatability. When the report is generated, it is determined whether the user needs a comparison view of the original image and the preprocessed image. If so, a comparison view with key anatomical point location markers is added to the report. And determine whether the purpose of the report is to calculate the arch angle; if so, organize the coordinate information according to the preset data structure. When used only for positioning detection, a basic coordinate report is output.

Images