Limb circumference measurement method, system and storage medium based on image recognition

Abstract

The present invention discloses a limb circumference measurement method, system and storage medium based on image recognition, which solves the problem that when the existing method uses a ToF camera to collect data on a limb surface with corners and concave shapes, the ToF camera light will be reflected multiple times, thereby causing unnecessary reflections and measurement distortion, and affecting the measurement accuracy. The method comprises: obtaining time-series limb measurement related data, combining deep learning and the yo lov8 model to pre-build an edge detection model, and identifying and decomposing the limb measurement related data; the edge detection model extracts the limb circumference edge of the patient in the circumference data set, outputs the circumference measurement result, and synchronously uploads the circumference measurement result to the measurement database; the present invention generates a corrected circumference data set by fusing measurement sensor data and limb image data, and uses the edge detection model to perform corrected edge detection and automatic marking on the circumference data set, thereby ensuring the accurate positioning of the measurement point and guaranteeing the advantage of measurement accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of limb diameter measurement, and in particular relates to a limb circumference measurement method, system and storage medium based on image recognition. Background Art

[0002] Postoperative lymphedema is a common chronic complication in patients with gynecological tumors and thyroid and breast tumors. Its incidence is affected by the surgical method of breast cancer, the diagnostic criteria for lymphedema, and the follow-up time. Accurately assessing the severity of lymphedema has become the core of lymphedema treatment. Lymphedema manifests as swelling, heaviness, pain, and local fever.

[0003] Upper limb circumference measurement is currently the most direct and commonly used method for diagnosing upper limb lymphedema after breast cancer surgery. Chinese patent CN117883070B discloses a method for measuring human limb circumference using ToF technology, including the following steps: using a ToF camera for data collection and preprocessing steps; training the key point recognition steps of the limb joints; and selecting the points at the circumference positions to be measured to calculate the circumference of the human limbs. However, when the existing method uses a single ToF technology to train the PointNet model to measure the circumference of the human limb, when the ToF camera collects data on the limb surface with corners and concave shapes, the ToF camera light will be reflected multiple times, resulting in unnecessary reflections and measurement distortion, affecting the measurement accuracy. To address the above problems, we propose a limb circumference measurement method, system and storage medium based on image recognition. Summary of the invention

[0004] The purpose of the present invention is to address the shortcomings of the prior art and provide a limb circumference measurement method, system and storage medium based on image recognition, which solves the problem that when the existing method uses a single ToF technology to train a PointNet model to measure the circumference of a human limb, when the ToF camera collects data on the surface of a limb with corners and concave shapes, the ToF camera light will be reflected multiple times, resulting in unnecessary reflections and measurement distortion, affecting the measurement accuracy.

[0005] The present invention is achieved by a limb circumference measurement method based on image recognition, the limb circumference measurement method based on image recognition comprising:

[0006] Acquire time-series limb measurement-related data, and upload the limb measurement-related data to a measurement database, wherein the limb measurement-related data is acquired through an adjustable measurement device, and the limb measurement-related data includes limb image data and measurement sensor data;

[0007] Combine deep learning and yo lov8 model to pre-build edge detection model, measure the database in response to modeling instructions, generate sample sets, randomly divide the sample sets into training sets and validation sets, and train and verify the accuracy of edge detection model through training sets and validation sets;

[0008] Load limb measurement-related data in real time, identify and decompose the limb measurement-related data, and generate a corrected circumference data set;

[0009] The edge detection model is executed with the circumference data set as input, the edge detection model extracts the circumference edge of the patient's limb in the circumference data set, outputs the circumference measurement result, and synchronously uploads the circumference measurement result to the measurement database.

[0010] The measurement database is traversed, and the measurement database responds to the measurement result index instruction to generate a measurement result set. The risk prediction results of a single group of sampling points are calculated based on multiple domino effects. At least one group of risk prediction results is correlated and analyzed to generate an overall assessment of the patient's limb lymphedema risk.

[0011] Preferably, the pre-construction of the edge detection model by combining deep learning and the yo lov8 model specifically includes:

[0012] The yo lov8 model is used as the initial model of the edge detection model. The yo lov8 model consists of a backbone layer, a neck layer, and a head layer. The SPPCSPC module in the backbone layer is frozen, and the Canny operator structure is used to replace the MP pooling layer in the backbone layer.

[0013] The RFRNet recurrent feature inference network is used in the head layer to detect the edge responsiveness of the limb image data, and the SSNBDL network structure is added in the neck layer. The SSNBDL network structure is used to deconvolute the limb image data and extract the prominent areas in the limb image data.

[0014] Based on the improved whale algorithm combined with deep learning, the loss function, activation function, and matching strategy of the yo lov8 model are optimized, and an edge detection model combining deep learning and the yo lov8 model is constructed.

[0015] Preferably, the step of training and verifying the accuracy of the edge detection model through a training set and a verification set specifically includes:

[0016] The training set is obtained and the Mosaic-Mixup hybrid method is used to perform online data enhancement on the training set. In Mosaic data enhancement, eight sample images are randomly selected from the training set for cropping, and then the cropped samples are randomly spliced ​​into one image as a new sample image. In Mixup data enhancement, four groups of sample images and their corresponding labels are linearly combined to form new sample images and labels.

[0017] Initialize the pre-built edge detection model and add an encoder to the backbone layer of the edge detection model. The input sequence of the encoder is represented as:

[0018]

[0019] Among them, L en represents the length of the encoder input sequence, m is the dimension of the input sequence, N * is the size of the training set or test set, and t is the input time of the sequence;

[0020] The encoder input sequence and the training set prediction results are input into the Canny operator structure. The Canny operator structure performs Gaussian smoothing and denoising on the encoder input sequence, calculates the image gradient and direction, and obtains an edge enhancement set.

[0021] The neck layer upsamples the predicted results of the training set and performs element-level addition operations to fuse feature maps of different levels, realizing multi-scale learning of limb image data and measurement sensor data in the training set to obtain an edge detection set.

[0022] Obtain the edge detection set, and perform iterative training by alternately training the RFRNet cyclic feature inference network and the feature fusion layer. First, train the RFRNet cyclic feature inference network n times, and then train the feature fusion layer once, until the loss function of the head layer converges.

[0023] Use the RFRNet cyclic feature inference network to gradually reduce the area to be filled through multiple cyclic feature inference modules, perform repair correction on the edge detection set, and output the feature repair set;

[0024] Obtain a validation set, use the validation set to verify the accuracy of the edge detection model, and determine whether the accuracy of the edge detection model reaches a preset accuracy threshold;

[0025] If the preset accuracy threshold is met, the edge detection model is output. If the preset accuracy threshold is not met, the Adam optimizer is used to fine-tune the loss function and activation function hyperparameters, and the pre-built edge detection model is returned to initialize, and an encoder is added to the backbone layer of the edge detection model.

[0026] Preferably, the method of identifying and decomposing the limb measurement associated data to generate a corrected circumference data set specifically includes:

[0027] Loading measurement-related data, and dividing the measurement-related data into image data sets and measurement data sets based on data types;

[0028] Obtain a measurement data set, and perform missing value and standardization processing on the measurement data set;

[0029] Load the measurement data set after missing values ​​and standardization, decompose the data in the measurement data set into multiple groups of measurement sub-data based on the singular spectrum decomposition algorithm, introduce Lagrange multipliers and quadratic penalty terms to the measurement sub-data for secondary decomposition, and obtain measurement decomposition components;

[0030] The measurement decomposition components are calculated using the following formula:

[0031]

[0032] Among them, W ssa (x) is the measurement decomposition component output after the secondary decomposition of the measurement sub-data f(x), p x is the decomposition matrix after the measurement sub-data f(x) is mapped by singular spectrum decomposition, is the central decomposition matrix, k is the number of decomposition matrices, n is the number of groups into which the decomposition matrix is ​​divided during singular spectrum decomposition, α is the quadratic penalty term, ω is the noise tolerance, and γ is the Lagrange multiplier term;

[0033] Traverse the image data set, extract the key information of the low-resolution images in the image data set based on the iterative back-projection method, and merge them into the high-resolution images based on prior knowledge, filter out the low-resolution images after extracting the key information, and generate a reconstructed image set;

[0034] The measurements are decomposed into components, reconstructed image sets and combined into a circumference dataset.

[0035] Preferably, the method for extracting the circumference edge of a patient's limb in a circumference data set by the edge detection model specifically includes:

[0036] Obtain a circumference data set, and the encoder encodes the measurement decomposition component and the reconstructed image set respectively, so that the measurement decomposition component serves as a label corresponding to the reconstructed image set;

[0037] The convolution module of the backbone layer outputs the reconstructed image set containing labels through convolution;

[0038] The convolution output is max-pooled through 3×3 and 6×6 pooling layers to obtain significant image features;

[0039] The Canny operator structure performs Gaussian smoothing and denoising on significant image features, calculates image gradients and directions, and obtains edge enhancement sets;

[0040] Among them, Gaussian smoothing denoising is performed on the significant image features based on the Gaussian denoising function, and the Gaussian denoising function is:

[0041]

[0042] Where b is the Gaussian surround scale of the Gaussian denoising function, x i is the information data set of significant image features, η is the Gaussian denoising coefficient, and H(x) is the edge extraction coefficient;

[0043]

[0044] Among them, x i is the information data set of significant image features, η is the Gaussian denoising coefficient, q i Initially set weights for the current salient image features;

[0045] Use the RFRNet cyclic feature inference network to gradually reduce the area to be filled in the edge enhancement set through multiple cyclic feature inference modules, and perform repair and correction on the edge detection set;

[0046] The neck layer upsamples the edge detection set, obtains the marked edge contour data, and outputs the circumference measurement result.

[0047] Preferably, the method of calculating the risk prediction results of a single group of sampling points based on multiple domino effects and performing correlation analysis on at least one group of risk prediction results specifically includes:

[0048] Obtain at least one set of circumference measurement results, and calculate the swelling probability of a single set of sampling points based on multiple domino effects;

[0049] Among them, the swelling probability of a single group of sampling points is calculated by the following formula:

[0050]

[0051] Among them, J x is the swelling probability of a single group of sampling points, t is the sampling time, T is the total sampling period, y t is the circumference measurement result of the sampling point at the current moment, is the mean absolute error of the circumference measurement results;

[0052] Obtain the swelling probability of a single group of sampling points, and calculate the risk prediction value of a single group of sampling points based on the risk superposition principle;

[0053]

[0054] Performing correlation analysis on at least one set of risk prediction results, and outputting a comprehensive risk value of limb swelling;

[0055]

[0056] Among them, J all is the comprehensive risk value of limb swelling, J z is the risk prediction value of a single group of sampling points, c is the number of sampling points, and τ is the risk coupling parameter.

[0057] The present invention also provides a limb circumference measurement system based on image recognition, the limb circumference measurement system based on image recognition comprising:

[0058] An adjustable measuring device for acquiring sequential limb measurement-related data and uploading the limb measurement-related data to a measurement database;

[0059] The model building module combines deep learning and yolov8 models to pre-build edge detection models, measures the database in response to modeling instructions, generates sample sets, randomly divides the sample sets into training sets and validation sets, and trains and verifies the accuracy of the edge detection model through the training sets and validation sets;

[0060] A data correction module is used to load limb measurement-related data in real time, identify and decompose the limb measurement-related data, and generate a corrected circumference data set;

[0061] A result acquisition module, taking the circumference data set as input, executes the edge detection model, the edge detection model extracts the circumference edge of the patient's limb in the circumference data set, outputs the circumference measurement result, and synchronously uploads the circumference measurement result to the measurement database;

[0062] The risk warning module traverses the measurement database, which responds to the measurement result index instruction to generate a measurement result set, calculates the risk prediction results of a single group of sampling points based on multiple domino effects, performs correlation analysis on at least one group of risk prediction results, and generates an overall assessment of the patient's limb lymphedema risk.

[0063] Preferably, the adjustable measuring device comprises:

[0064] Equipment base;

[0065] A lifting adjustment part, the lifting adjustment part is arranged on the equipment base, the lifting adjustment part comprises an adjusting wheel, an adjusting screw, a lifting threaded seat and a lifting guide rod, the lifting guide rod is fixedly installed on the equipment base, the lifting threaded seat is slidably connected with the lifting guide rod, the lifting threaded seat is threadedly sleeved on the outside of the adjusting screw, one end of the adjusting screw is fixedly installed with an adjusting wheel, and the bottom of the adjusting screw is rotatably connected with the equipment base;

[0066] a measuring limiter connected to the lifting and lowering adjustment unit, and

[0067] A circumference measuring part, which is arranged in the measuring limit part and is used to measure the circumference of a patient's limb;

[0068] The measuring limit part includes an annular guide rail, one side of the annular guide rail is fixedly connected to the lifting threaded seat, an annular slide groove is provided in the annular guide rail, a linkage slider is slidably installed in the annular slide groove, the linkage slider is detachably connected to an adjustable fitting part, and the adjustable fitting part is connected to the circumference measuring part;

[0069] The adjustable fitting portion comprises:

[0070] A fitting limit seat, wherein the fitting limit seat is detachably connected to the linkage slider, a fitting limit groove is provided in the fitting limit seat, and an auxiliary lever is detachably mounted on the fitting limit seat;

[0071] A linkage screw, wherein the linkage screw passes through one end of the fitting limit seat, and the linkage screw is connected to the fitting limit seat through a thread, a linkage rotating wheel is fixedly installed on one end of the linkage screw, and an extrusion slider is rotatably connected to the other end of the linkage screw, and the extrusion slider is slidably installed in the fitting limit groove;

[0072] A fitting spring fixedly installed in the extrusion slider, one end of the fitting spring away from the extrusion slider is fixedly connected to a driven slider, one end of the driven slider away from the fitting spring is fixedly connected to a driven fitting rod, the driven fitting rod is slidably connected to the fitting limit seat, and the end of the driven fitting rod is fixedly connected to a circumference measuring part;

[0073] The circumference measuring unit comprises:

[0074] A measuring mounting seat, wherein the measuring mounting seat is fixedly mounted on the end of the driven fitting rod;

[0075] A data acquisition group, the data acquisition group comprising a displacement sensor and at least one group of acquisition cameras;

[0076] A camera adjustment group, the camera adjustment group is arranged in the measurement mounting seat, and the camera adjustment group is used to adjust the acquisition position of the acquisition camera;

[0077] The camera adjustment group includes:

[0078] An adjusting motor, wherein the adjusting motor is fixedly embedded in the measuring mounting seat;

[0079] An inner rotating disc fixedly connected to the output end of the regulating motor, wherein at least one set of inner linkage grooves is provided in the inner rotating disc;

[0080] An outer limit plate detachably mounted on the measuring mounting seat, wherein at least one set of outer linkage grooves is formed in the outer limit plate;

[0081] At least one set of adjusting sliders, wherein the adjusting sliders are respectively slidably connected with the inner linkage groove and the outer linkage groove, and a camera support seat is detachably mounted on the adjusting slider, and the acquisition camera is detachably mounted in the camera support seat.

[0082] The present invention also provides a computer-readable storage medium, which stores computer program instructions. When the computer program instructions are executed, the limb circumference measurement method based on image recognition is implemented.

[0083] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0084] In an embodiment of the present invention, a corrected circumference data set is generated by fusing measurement sensor data and limb image data, and the circumference data set is corrected for edge detection and automatic marking with the help of an edge detection model, thereby ensuring accurate positioning of the measurement points and the advantages of measurement accuracy. This overcomes the problem that when the existing method uses a single ToF technology to train a PointNet model to measure the circumference of a human limb, when the ToF camera collects data on the surface of a limb with corners and concave shapes, the ToF camera light will be reflected multiple times, thereby causing unnecessary reflections and measurement distortion, affecting the measurement accuracy.

[0085] In an embodiment of the present invention, an edge detection model is provided. When constructing the edge detection model, the yo lov8 model is used as the initial model, and the edge detection model combining deep learning and the yo lov8 model is constructed in combination with the RFRNet cyclic feature inference network and the Canny operator structure, thereby optimizing and improving the yo lov8 model, making the model more suitable for feature fusion and extraction of circumference data sets including measurement decomposition components and reconstructed image sets. The introduction of the RFRNet cyclic feature inference network effectively reduces the computational cost and improves the feature expression ability and recognition accuracy of the model.

[0086] In an embodiment of the present invention, the data in a measurement data set is decomposed into multiple groups of measurement sub-data through a singular spectrum decomposition algorithm, and Lagrange multipliers and quadratic penalty terms are introduced into the measurement sub-data to perform a secondary decomposition of the measurement sub-data, thereby enabling long-term trend prediction, periodic change feature extraction, and noise elimination for nonlinear and non-stationary signal data, so that the measurement data set can cooperate with the image data set to assist in constructing an edge detection model, thereby improving the robustness and generalization ability of the model.

[0087] In an embodiment of the present invention, a method is provided for calculating the risk prediction results of a single group of sampling points based on multiple domino effects, and for correlation analysis of at least one group of risk prediction results, thereby being able to evaluate the risk of circumferential limb edema of a patient during multiple sampling measurements, helping patients and doctors to promptly capture any subtle changes in the condition and providing timely data support for the adjustment of treatment plans.

[0088] In an embodiment of the present invention, an adjustable measuring device is provided, which consists of a circumference measuring part, a measuring limit part and a lifting and lowering adjustment part. The circumference measuring part, the measuring limit part and the lifting and lowering adjustment part work together to ensure flexible measurement of the circumference of different patients and different limb areas, thereby ensuring the stability and accuracy of the measurement data. At the same time, it is easy to operate and improves the efficiency of circumference measurement. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 This is a schematic diagram of measuring the circumference of some limbs in the prior art.

[0090] Figure 2 It is a schematic diagram of the implementation flow of the limb circumference measurement method based on image recognition provided by the present invention.

[0091] Figure 3 It is a schematic diagram of the implementation process of combining deep learning and yo lov8 model to pre-build an edge detection model provided by the present invention.

[0092] Figure 4 It is a schematic diagram of the implementation flow of the method for training and verifying the accuracy of an edge detection model through a training set and a verification set provided by the present invention.

[0093] Figure 5 It is a schematic diagram of the implementation flow of the method provided by the present invention for identifying and decomposing limb measurement associated data to generate a corrected circumference data set.

[0094] Figure 6 It is a schematic diagram of the implementation process of the edge detection model provided by the present invention for extracting the circumference edge of a patient's limb in a circumference data set.

[0095] Figure 7 It is a schematic diagram of the implementation flow of the method for correlation analysis of at least one group of risk prediction results based on multiple domino effects to calculate the risk prediction results of a single group of sampling points provided by the present invention.

[0096] Figure 8 It is a schematic diagram of the control principle of the limb circumference measurement system based on image recognition provided by the present invention.

[0097] Fig. 9 It is a structural schematic diagram of the adjustable measuring device provided by the present invention.

[0098] Fig.10 It is a top view of the adjustable measuring device provided by the present invention.

[0099] Fig.11 It is a structural schematic diagram of the adjustable fitting portion provided by the present invention.

[0100] Fig.12 It is an axonometric view of the adjustable fitting portion provided by the present invention.

[0101] Fig.13 It is a structural schematic diagram of the circumference measuring unit provided by the present invention.

[0102] Fig.14 It is an axonometric diagram of the circumference measuring part provided by the present invention.

[0103] Fig.15 It is a top view of the circumference measuring part provided by the present invention.

[0104] In the figure: 10-adjustable measuring device, 11-equipment base, 12-lifting adjustment part, 121-adjusting wheel, 122-adjusting screw, 123-lifting threaded seat, 124-lifting guide rod, 2-measuring limit part, 21-annular guide rail, 22-annular slide, 23-adjustable fitting part, 231-fitting limit seat, 232-fitting limit groove, 233-linked wheel, 234-linked screw, 235-extrusion slider, 236-fitting spring, 237-driven slider, 238 -driven fitting rod, 239-auxiliary lever, 24-linkage slider, 3-circumference measurement part, 31-measurement mounting seat, 32-data acquisition group, 321-displacement sensor, 322-acquisition camera, 33-camera adjustment group, 331-adjustment motor, 332-inner turntable, 333-outer limit plate, 334-outer linkage slot, 335-adjustment slider, 336-camera support seat, 20-model building module, 30-data correction module, 40-result acquisition module, 50-risk warning module. DETAILED DESCRIPTION

[0105] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of this application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0106] At present, in the ultramicrosurgery treatment of lymphedema caused by tumors (gynecological tumors, thyroid and breast tumors), each preoperative and postoperative follow-up requires measuring the circumference of certain points of the limb to evaluate the detumescence effect, which is very cumbersome and prone to errors. Figure 1 shown.

[0107] There are also some methods that use a single ToF technology to train the PointNet model to measure the circumference of human limbs. When the ToF camera collects data on the surface of limbs with corners and concave shapes, the ToF camera light will be reflected multiple times, resulting in unnecessary reflections and measurement distortion, affecting the measurement accuracy. To address the above problems, we proposed a limb circumference measurement method, system and storage medium based on image recognition. In short, when implementing the method, first obtain the time-series limb measurement related data, and upload the limb measurement related data to the measurement database. Then, combine deep learning and yo lov8 model to pre-build an edge detection model, identify and decompose the limb measurement related data, and generate a corrected circumference data set. The edge detection model extracts the patient's limb circumference edge in the circumference data set and outputs the circumference measurement result. Finally, based on multiple domino effects, the risk prediction results of a single group of sampling points are calculated, and at least one group of risk prediction results are correlated and analyzed to generate an overall assessment of the patient's limb lymphedema risk. In an embodiment of the present invention, a corrected circumference data set is generated by fusing measurement sensor data and limb image data, and the circumference data set is corrected for edge detection and automatic marking with the help of an edge detection model, thereby ensuring accurate positioning of the measurement points and the advantages of measurement accuracy. This overcomes the problem that when the existing method uses a single ToF technology to train a PointNet model to measure the circumference of a human limb, when the ToF camera collects data on the surface of a limb with corners and concave shapes, the ToF camera light will be reflected multiple times, thereby causing unnecessary reflections and measurement distortion, affecting the measurement accuracy.

[0108] Example 1

[0109] The embodiment of the present invention provides a method for measuring limb circumference based on image recognition. Figure 2 The figure shows a schematic diagram of the implementation process of the limb circumference measurement method based on image recognition, which specifically includes:

[0110] Step S10, acquiring time-series limb measurement-related data, and uploading the limb measurement-related data to a measurement database;

[0111] It should be noted that the limb measurement-related data is obtained through the adjustable measuring device 10, and the limb measurement-related data includes but is not limited to limb picture data and measurement sensor data; the measurement sensor data can be displacement sensor data, angle sensor data, and infrared sensor data.

[0112] Step S20, pre-building an edge detection model by combining deep learning and the yo lov8 model, measuring the database in response to the modeling instruction, generating a sample set, randomly dividing the sample set into a training set and a validation set, and training and verifying the accuracy of the edge detection model through the training set and the validation set;

[0113] Step S30, loading the limb measurement related data in real time, identifying and decomposing the limb measurement related data, and generating a corrected circumference data set;

[0114] Step S40, taking the circumference data set as input, executing the edge detection model, the edge detection model extracts the circumference edge of the patient's limb in the circumference data set, outputs the circumference measurement result, and synchronously uploads the circumference measurement result to the measurement database.

[0115] Step S50, traverse the measurement database, the measurement database responds to the measurement result index instruction, generates a measurement result set, calculates the risk prediction results of a single group of sampling points based on multiple domino effects, performs correlation analysis on at least one group of risk prediction results, and generates an overall assessment of the patient's limb lymphedema risk.

[0116] In this embodiment, in the embodiment of the present invention, a corrected circumference data set is generated by fusing measurement sensor data and limb image data, and the circumference data set is corrected for edge detection and automatic marking with the help of an edge detection model to ensure accurate positioning of the measurement points and the advantages of measurement accuracy. This overcomes the problem that when the existing method uses a single ToF technology to train the PointNet model to measure the circumference of human limbs, when the ToF camera collects data on the surface of a limb with corners and concave shapes, the ToF camera light will be reflected multiple times, resulting in unnecessary reflections and measurement distortion, affecting the measurement accuracy.

[0117] The embodiment of the present invention provides a pre-built edge detection model combining deep learning and yo lov8 model, Figure 3 The schematic diagram of the implementation process of combining deep learning and yo lov8 model to pre-build edge detection model is shown, and the pre-building edge detection model by combining deep learning and yo lov8 model specifically includes:

[0118] Step S101, using the yo lov8 model as the initial model of the edge detection model, wherein the yo lov8 model consists of a backbone layer, a neck layer, and a head layer, freezing the SPPCSPC module in the backbone layer, and using the Canny operator structure to replace the MP pooling layer in the backbone layer;

[0119] It should be noted that in the present invention, the yo lov8 model is used as the initial model of the edge detection model because the yo lov8 model adopts an anchor-free detection head. This improvement improves the accuracy of detection, especially when processing small-sized objects.

[0120] In this embodiment, the Canny operator structure realizes efficient and accurate edge detection of signals and data based on edge detection algorithm, non-maximum suppression, dual threshold processing and edge tracking. The Canny operator structure has the advantages of low error rate, high localization and minimum response, and has a positive effect on signal preprocessing, sensitive response, enhanced selection and advanced decision-making stages.

[0121] Step S102, using the RFRNet cyclic feature inference network in the head layer to detect the edge responsiveness of the limb image data, and adding the SSNBDL network structure in the neck layer, the SSNBDL network structure is used to deconvolute and restore the limb image data, and extract the significant salient areas in the limb image data;

[0122] Step S103, based on the improved whale algorithm combined with deep learning, the loss function, activation function, and matching strategy of the yo lov8 model are optimized to construct an edge detection model that combines deep learning and the yo lov8 model.

[0123] It should be noted that the loss function of the yo lov8 model can be a binary cross entropy loss function, and the loss function optimization can be adjusted and optimized by adjusting the weight factors of the three factors of overlapping area, center point distance and aspect ratio, and the activation function is set to the Leaky-ReLU activation function. The activation function can be optimized by dynamically adjusting the negative slope of Leaky-ReLU, and combining automatic learning to find the activation method that best suits the current task to achieve parameter optimization of the Leaky-ReLU activation function. The matching strategy is the Anchor-Free mechanism. In this embodiment, the adaptability of the network to multi-scale targets can be enhanced by adding more scale change layers.

[0124] In an embodiment of the present invention, an edge detection model is provided. When constructing the edge detection model, the yo lov8 model is used as the initial model, and the edge detection model combining deep learning and the yo lov8 model is constructed in combination with the RFRNet cyclic feature inference network and the Canny operator structure, thereby optimizing and improving the yo lov8 model, making the model more suitable for feature fusion and extraction of circumference data sets including measurement decomposition components and reconstructed image sets. The introduction of the RFRNet cyclic feature inference network effectively reduces the computational cost and improves the feature expression ability and recognition accuracy of the model.

[0125] The embodiment of the present invention provides a method for training and verifying the accuracy of an edge detection model through a training set and a verification set. Figure 4 The figure shows a schematic diagram of the implementation process of a method for training and verifying the accuracy of an edge detection model through a training set and a verification set, wherein the method of training and verifying the accuracy of an edge detection model through a training set and a verification set specifically includes:

[0126] Step S201, obtaining a training set, and performing online data enhancement on the training set using a Mosaic-Mixup hybrid method;

[0127] In this embodiment, during Mosaic data enhancement, eight sample images are randomly selected from the training set and cropped, and then the cropped samples are randomly spliced ​​into one image as a new sample image. During Mixup data enhancement, four groups of sample images and their corresponding labels are linearly combined to form new sample images and labels.

[0128] Step S202, initialize the pre-built edge detection model, add an encoder in the backbone layer of the edge detection model, and the input sequence of the encoder is expressed as:

[0129]

[0130] Among them, L en represents the length of the encoder input sequence, m is the dimension of the input sequence, N * is the size of the training set or test set, and t is the input time of the sequence;

[0131] Step S203, inputting the encoder input sequence and the training set prediction result into the Canny operator structure, the Canny operator structure performs Gaussian smoothing denoising on the encoder input sequence, calculates the image gradient and direction, and obtains an edge enhancement set;

[0132] Step S204, the neck layer upsamples the training set prediction results, and performs element-level addition operation to fuse feature maps of different levels, so as to realize multi-scale learning of limb image data and measurement sensor data in the training set, and obtain an edge detection set;

[0133] Step S205, obtaining an edge detection set, and performing iterative training by alternately training the RFRNet cyclic feature inference network and the feature fusion layer, wherein the RFRNet cyclic feature inference network is first trained n times, and then the feature fusion layer is trained once, until the loss function of the head layer reaches convergence;

[0134] Step S206, using the RFRNet cyclic feature inference network to gradually reduce the area to be filled through multiple cyclic feature inference modules, repair and correct the edge detection set, and output the feature repair set;

[0135] Step S207, obtaining a verification set, using the verification set to verify the accuracy of the edge detection model, and determining whether the accuracy of the edge detection model reaches a preset accuracy threshold. In this embodiment, the preset accuracy threshold may be 0.95-0.98;

[0136] Step S208: if the preset accuracy threshold is met, output the edge detection model;

[0137] If the preset accuracy threshold is not reached, the Adam optimizer is used to fine-tune the loss function and activation function hyperparameters, and the process returns to step S203 to execute the steps of initializing the pre-built edge detection model and adding an encoder to the backbone layer of the edge detection model.

[0138] In this embodiment, the use of the Adam optimizer to fine-tune the hyperparameters of the loss function and activation function has the advantages of adaptive learning rate adjustment, efficient processing of sparse gradients and noise, and improved stability and efficiency of model training. At the same time, the Adam optimizer has a natural advantage in processing sparse gradients and noise. In deep learning, especially when processing natural language processing or large-scale data sets, the problem of gradient sparsity is often encountered.

[0139] The embodiment of the present invention provides a method for identifying and decomposing limb measurement related data to generate a corrected circumference data set. Figure 5 The present invention shows a schematic diagram of the implementation process of a method for identifying and decomposing limb measurement-related data to generate a corrected circumference data set. The method for identifying and decomposing limb measurement-related data to generate a corrected circumference data set specifically includes:

[0140] Step S301, loading measurement-related data, and dividing the measurement-related data into an image data set and a measurement data set based on data type;

[0141] Step S302, obtaining a measurement data set, and performing missing value and standardization processing on the measurement data set;

[0142] In this embodiment, the missing value processing method for the measurement data set may be missing value deletion or filling processing, and the standardization processing for the measurement data set may be Min-Max standardization or Z-score standardization processing.

[0143] Step S303, loading the measurement data set after the missing values ​​and the normalization processing, decomposing the data in the measurement data set into multiple groups of measurement sub-data based on the singular spectrum decomposition algorithm, introducing Lagrange multipliers and quadratic penalty terms into the measurement sub-data to perform secondary decomposition on the measurement sub-data, and obtaining measurement decomposition components;

[0144] The measurement decomposition components are calculated using the following formula:

[0145]

[0146] Among them, W ssa (x) is the measurement decomposition component output after the secondary decomposition of the measurement sub-data f(x), p x is the decomposition matrix after the measurement sub-data f(x) is mapped by singular spectrum decomposition, is the central decomposition matrix, k is the number of decomposition matrices, n is the number of groups into which the decomposition matrix is ​​divided during singular spectrum decomposition, α is the quadratic penalty term, in this embodiment, the quadratic penalty term can be set to 1500-2000, ω is the noise tolerance, the noise tolerance can be set to 0.05-0.1, γ is the Lagrange multiplier term set to 0.03, in this embodiment, by introducing additional Lagrange multiplier terms, the constrained problem is converted into an unconstrained problem, thereby simplifying the solution process;

[0147] Step S304, traversing the image data set, extracting key information of low-resolution images in the image data set based on an iterative back-projection method and merging them into high-resolution images in combination with prior knowledge, filtering out the low-resolution images after extracting the key information, and generating a reconstructed image set;

[0148] Step S305, combining the measurement decomposition components and the reconstructed image set into a circumference data set.

[0149] In an embodiment of the present invention, the data in a measurement data set is decomposed into multiple groups of measurement sub-data through a singular spectrum decomposition algorithm, and Lagrange multipliers and quadratic penalty terms are introduced into the measurement sub-data to perform a secondary decomposition of the measurement sub-data, thereby enabling long-term trend prediction, periodic change feature extraction, and noise elimination for nonlinear and non-stationary signal data, so that the measurement data set can cooperate with the image data set to assist in constructing an edge detection model, thereby improving the robustness and generalization ability of the model.

[0150] The embodiment of the present invention provides a method for extracting the perimeter edge of a patient's limb in a perimeter data set using an edge detection model. Figure 6 The schematic diagram of the implementation process of the edge detection model for extracting the circumference edge of a patient's limb in the circumference data set is shown. The method for extracting the circumference edge of a patient's limb in the circumference data set by the edge detection model specifically includes:

[0151] Step S401, obtaining a circumference data set, and encoding the measurement decomposition component and the reconstructed image set respectively, so that the measurement decomposition component serves as a label corresponding to the reconstructed image set;

[0152] Step S402, the convolution module of the backbone layer outputs the reconstructed image set containing the label through convolution;

[0153] Step S403, performing maximum pooling processing on the convolution output through 3×3 and 6×6 pooling layers to obtain significant image features;

[0154] Step S404, the Canny operator structure performs Gaussian smoothing and denoising on the significant image features, calculates the image gradient and direction, and obtains an edge enhancement set;

[0155] Among them, Gaussian smoothing denoising is performed on the significant image features based on the Gaussian denoising function, and the Gaussian denoising function is:

[0156]

[0157] Wherein, b is the Gaussian surround scale of the Gaussian denoising function. In this embodiment, the Gaussian surround scale can be set to 85, x i is a set of information data of significant image features, η is a Gaussian denoising coefficient. In this embodiment, the Gaussian denoising coefficient may be 0.05, and H(x) is an edge extraction coefficient;

[0158]

[0159] Among them, x i is the information data set of significant image features, η is the Gaussian denoising coefficient, q i The initial weights of the current significant image features can be set by Xavier initialization or MSRA initialization.

[0160] Step S405, using the RFRNet cyclic feature inference network to gradually reduce the area to be filled in the edge enhancement set through multiple cyclic feature inference modules, and perform repair correction on the edge detection set;

[0161] Step S406: The neck layer upsamples the edge detection set, obtains the marked edge contour data, and outputs the circumference measurement result.

[0162] The embodiment of the present invention provides a method for calculating the risk prediction results of a single group of sampling points based on multiple domino effects, and performing correlation analysis on at least one group of risk prediction results. Figure 7 The present invention shows a schematic diagram of the implementation process of a method for correlation analysis of at least one group of risk prediction results by calculating the risk prediction results of a single group of sampling points based on multiple domino effects. The method for correlation analysis of at least one group of risk prediction results by calculating the risk prediction results of a single group of sampling points based on multiple domino effects specifically includes:

[0163] Step S501, obtaining at least one set of circumference measurement results, and calculating the swelling probability of a single set of sampling points based on multiple domino effects;

[0164] Among them, the swelling probability of a single group of sampling points is calculated by the following formula:

[0165]

[0166] Among them, J x is the swelling probability of a single group of sampling points, t is the sampling time, T is the total sampling period, y t is the circumference measurement result of the sampling point at the current moment, is the mean absolute error of the circumference measurement results;

[0167] It should be noted that during single sampling, at least four groups of sampling points are set on the surface of the limb to facilitate accurate measurement and evaluation of the limb circumference.

[0168] Step S502, obtaining the swelling probability of a single group of sampling points, and calculating the risk prediction value of the single group of sampling points based on the risk superposition principle;

[0169]

[0170] Step S503, performing correlation analysis on at least one set of risk prediction results, and outputting a comprehensive risk value of limb swelling;

[0171]

[0172] Among them, J all is the comprehensive risk value of limb swelling, J z is the risk prediction value of a single group of sampling points, c is the number of sampling points. In this embodiment, the number of sampling points can be 3-10, τ is the risk coupling parameter, and the risk coupling parameter is set to 0.05-0.3.

[0173] In an embodiment of the present invention, a method is provided for calculating the risk prediction results of a single group of sampling points based on multiple domino effects, and for correlation analysis of at least one group of risk prediction results, thereby being able to evaluate the risk of circumferential limb edema of a patient during multiple sampling measurements, helping patients and doctors to promptly capture any subtle changes in the condition and providing timely data support for the adjustment of treatment plans.

[0174] Example 2

[0175] The embodiment of the present invention provides a limb circumference measurement system based on image recognition. Figure 8 The control principle schematic diagram of the limb circumference measurement system based on image recognition is shown. The limb circumference measurement system based on image recognition specifically includes:

[0176] The adjustable measuring device 10 is used to obtain the time-series limb measurement-related data and upload the limb measurement-related data to the measurement database;

[0177] The model building module 20 combines deep learning and the yo lov8 model to pre-build an edge detection model, measures the database in response to the modeling instruction, generates a sample set, randomly divides the sample set into a training set and a validation set, and trains and verifies the accuracy of the edge detection model through the training set and the validation set;

[0178] The data correction module 30 is used to load the limb measurement related data in real time, identify and decompose the limb measurement related data, and generate a corrected circumference data set;

[0179] The result acquisition module 40 takes the circumference data set as input, executes the edge detection model, the edge detection model extracts the circumference edge of the patient's limb in the circumference data set, outputs the circumference measurement result, and synchronously uploads the circumference measurement result to the measurement database;

[0180] The risk warning module 50 traverses the measurement database, which responds to the measurement result index instruction to generate a measurement result set, calculates the risk prediction results of a single group of sampling points based on multiple domino effects, performs correlation analysis on at least one group of risk prediction results, and generates an overall assessment of the patient's limb lymphedema risk.

[0181] It can be understood that the limb circumference measurement system based on image recognition provided by the embodiment of the present invention corresponds to the above-mentioned limb circumference measurement method based on image recognition. The explanations, examples, beneficial effects and other parts of the relevant contents can refer to the corresponding contents in the limb circumference measurement method based on image recognition, and will not be repeated here.

[0182] In addition, it can be further understood that when the limb circumference measurement system based on image recognition is used to measure the limb circumference, the limb circumference measurement system based on image recognition needs to perform the measurement work according to the steps S10-S50.

[0183] In this embodiment, if Figure 9-10 As shown, the adjustable measuring device 10 comprises:

[0184] The equipment base 11 may be a rectangular base, a T-shaped base or a round base structure, and at least one set of anti-slip friction strips may be provided at the bottom of the equipment base 11;

[0185] A lifting and lowering adjustment part 12, the lifting and lowering adjustment part 12 is arranged on the equipment base 11, the lifting and lowering adjustment part 12 comprises an adjusting wheel 121, an adjusting screw 122, a lifting and lowering threaded seat 123 and a lifting and lowering guide rod 124, the lifting and lowering guide rod 124 is fixedly mounted on the equipment base 11, the lifting and lowering threaded seat 123 is slidably connected with the lifting and lowering guide rod 124, the lifting and lowering threaded seat 123 is threadedly sleeved on the outside of the adjusting screw 122, one end of the adjusting screw 122 is fixedly mounted with an adjusting wheel 121, and the bottom of the adjusting screw 122 is rotatably connected with the equipment base 11;

[0186] It should be noted that the bottom of the lifting guide rod 124 is fixedly installed on the equipment base 11 by means of a snap or thread. The lifting guide rod 124 can be a round rod, a square rod or a "T"-shaped rod structure, and the top of the adjusting screw 122 is fixedly connected to the adjusting wheel 121 by means of a snap or plug-in, and the bottom of the adjusting screw 122 is rotatably connected to the equipment base 11 by means of a bearing or a roller.

[0187] When the measuring height needs to be adjusted to facilitate switching of sampling points, the adjusting wheel 121 is manually rotated, and the adjusting wheel 121 drives the adjusting screw 122 to rotate, so that the adjusting screw 122 drives the lifting threaded seat 123, the measuring limit part 2 and the circumference measuring part 3 to move up and down, thereby facilitating switching of sampling points.

[0188] The measuring limit part 2 connected to the lifting and lowering adjustment part 12, and

[0189] The circumference measuring part 3 is arranged in the measuring limit part 2, and the circumference measuring part 3 is used to measure the circumference of the patient's limb.

[0190] In an embodiment of the present invention, an adjustable measuring device 10 is provided. The adjustable measuring device 10 is composed of a circumference measuring part 3, a measuring limit part 2 and a lifting and lowering adjustment part 12. The circumference measuring part 3, the measuring limit part 2 and the lifting and lowering adjustment part 12 cooperate with each other to ensure flexible measurement of the circumference of different patients and different limb areas, thereby ensuring the stability and accuracy of the measurement data. At the same time, it is easy to operate and improves the efficiency of circumference measurement.

[0191] The measuring limit part 2 includes an annular guide rail 21, one side of the annular guide rail 21 is fixedly connected to the lifting threaded seat 123, an annular slide groove 22 is provided in the annular guide rail 21, a linkage slider 24 is slidably installed in the annular slide groove 22, and the linkage slider 24 is detachably connected to an adjustable fitting part 23, and the adjustable fitting part 23 is connected to the circumference measuring part 3;

[0192] It should be noted that the annular guide rail 21 can be a round seat or a circular ring with a hollow interior. The annular guide rail 21 is fixedly connected to the lifting threaded seat 123 by fastening bolts or welding. The inner wall of the annular slide groove 22 is polished, and the linkage slider 24 can be a cylindrical slider or a round seat slider structure.

[0193] like Figure 11-Figure 12 As shown, the adjustable fitting portion 23 includes:

[0194] A fitting limit seat 231, wherein the fitting limit seat 231 is detachably connected to the linkage slider 24, a fitting limit groove 232 is provided in the fitting limit seat 231, and an auxiliary lever 239 is detachably mounted on the fitting limit seat 231;

[0195] In this embodiment, the fitting limit seat 231 can be a rectangular seat or a "T"-shaped seat structure with a hollow interior, the fitting limit groove 232 can be a rectangular groove, and the auxiliary lever 239 is fixedly installed on the fitting limit seat 231 by plugging or riveting.

[0196] A linkage screw 234, wherein the linkage screw 234 passes through one end of the fitting limit seat 231, and the linkage screw 234 and the fitting limit seat 231 are connected by threads, a linkage rotating wheel 233 is fixedly installed on one end of the linkage screw 234, and an extrusion slider 235 is rotatably connected to the other end of the linkage screw 234, and the extrusion slider 235 is slidably installed in the fitting limit groove 232;

[0197] It should be noted that one end of the linkage screw 234 is fixedly connected to the linkage rotating wheel 233 by plugging or clamping, and the extrusion slider 235 is rotationally connected to the end of the linkage screw 234 via a bearing.

[0198] A fitting spring 236 is fixedly installed in the extrusion slider 235, and one end of the fitting spring 236 away from the extrusion slider 235 is fixedly connected to a driven slider 237, and one end of the driven slider 237 away from the fitting spring 236 is fixedly connected to a driven fitting rod 238, the driven fitting rod 238 is slidably connected to the fitting limit seat 231, and the end of the driven fitting rod 238 is fixedly connected to the circumference measuring part 3;

[0199] In the embodiment of the present invention, an adjustable fitting portion 23 is provided, and the fitting spring 236 acting on the circumference measurement portion 3 can not only prevent deformation due to edema during measurement, but also ensure that the circumference measurement portion 3 fits the outside of the limb, thereby ensuring the accuracy of data acquisition. The tension of the fitting spring 236 can be flexibly adjusted by the linkage screw 234, further making the adjustable fitting portion 23 suitable for measurement work in different edema areas.

[0200] like Figure 13-Figure 15 As shown, the circumference measuring unit 3 includes:

[0201] A measuring mounting seat 31, wherein the measuring mounting seat 31 is fixedly mounted on the end of the driven fitting rod 238;

[0202] A data acquisition group 32, wherein the data acquisition group 32 includes a displacement sensor 321 and at least one group of acquisition cameras 322;

[0203] A camera adjustment group 33, wherein the camera adjustment group 33 is disposed in the measurement mounting seat 31, and the camera adjustment group 33 is used to adjust the acquisition position of the acquisition camera 322;

[0204] It should be noted that the measuring mount 31 can be an arc-shaped mount, a rectangular mount or a round mount, and the measuring mount 31 is fixedly connected to the end of the driven fitting rod 238 by a snap or a bolt, and the acquisition camera 322 can be a combination of one or more ToF cameras, infrared cameras, laser scanning cameras, and the displacement sensor 321 is fixedly mounted on the surface of the measuring mount 31 by bolts or snaps.

[0205] The camera adjustment group 33 includes:

[0206] An adjusting motor 331, wherein the adjusting motor 331 is fixedly embedded in the measuring mounting seat 31, and the adjusting motor 331 may be a servo motor, and the adjusting motor 331 is fixedly embedded in the measuring mounting seat 31 by means of fastening bolts or clamps;

[0207] An inner rotating disk 332 fixedly connected to the output end of the adjusting motor 331, wherein at least one set of inner linkage grooves is provided in the inner rotating disk 332;

[0208] An outer limiting plate 333 detachably mounted on the measuring mounting seat 31 , wherein at least one set of outer linkage grooves 334 is formed in the outer limiting plate 333 ;

[0209] The inner turntable 332 is fixedly connected to the output end of the adjustment motor 331 by means of bolts or interference fit, the inner linkage groove can be an arc groove or an "S"-shaped groove structure, the number of the inner linkage grooves and the outer linkage grooves 334 are consistent, and the number of the inner linkage grooves and the outer linkage grooves 334 can be 3-8 groups.

[0210] At least one set of adjusting sliders 335 , wherein the adjusting sliders 335 are respectively slidably connected to the inner linkage groove and the outer linkage groove 334 , and a camera support seat 336 is detachably mounted on the adjusting slider 335 , and the acquisition camera 322 is detachably mounted in the camera support seat 336 .

[0211] During operation, in order to improve the diversity of data samples, the adjustment motor 331 is started. The rotation of the adjustment motor 331 can drive the inner turntable 332 to rotate, and the inner turntable 332 drives the inner linkage groove to rotate, so that the inner linkage groove drives the adjustment slider 335 to slide along the outer linkage groove 334, and then the adjustment slider 335 drives the camera support seat 336 and the acquisition camera 322 to move, thereby realizing flexible adjustment of the sampling position of the acquisition camera 322.

[0212] Example 3

[0213] Another aspect of the embodiments of the present invention further provides a computer-readable storage medium, which stores computer program instructions, and the computer program instructions can be executed by a processor. When the computer program instructions are executed, the method of any one of the above embodiments is implemented.

[0214] Finally, it should be noted that the computer-readable storage medium (e.g., memory) herein may be a volatile memory or a nonvolatile memory, or may include both a volatile memory and a nonvolatile memory. As an example and not by way of limitation, a nonvolatile memory may include a read-only memory (ROM), a programmable ROM (PROM), an electrically programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), or a flash memory. A volatile memory may include a random access memory (RAM), which may act as an external cache memory. As an example and not by way of limitation, RAM may be obtained in a variety of forms, such as synchronous RAM (DRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), and direct Rambus RAM (DRRAM). The storage devices of the disclosed aspects are intended to include, but are not limited to, these and other suitable types of memory.

[0215] In summary, the present invention provides a limb circumference measurement method, system and storage medium based on image recognition. In an embodiment of the present invention, a corrected circumference data set is generated by fusing measurement sensor data and limb image data, and the circumference data set is corrected for edge detection and automatic marking with the help of an edge detection model to ensure accurate positioning of the measurement points and the advantages of measurement accuracy. This overcomes the problem that when the existing method uses a single ToF technology to train a PointNet model to measure the circumference of a human limb, when the ToF camera collects data on a limb surface with corners and concave shapes, the ToF camera light will be reflected multiple times, resulting in unnecessary reflections and measurement distortion, affecting the measurement accuracy.

[0216] In an embodiment of the present invention, an edge detection model is provided. When constructing the edge detection model, the yo lov8 model is used as the initial model, and the edge detection model combining deep learning and the yo lov8 model is constructed in combination with the RFRNet cyclic feature inference network and the Canny operator structure, thereby optimizing and improving the yo lov8 model, making the model more suitable for feature fusion and extraction of circumference data sets including measurement decomposition components and reconstructed image sets. The introduction of the RFRNet cyclic feature inference network effectively reduces the computational cost and improves the feature expression ability and recognition accuracy of the model.

[0217] In an embodiment of the present invention, the data in a measurement data set is decomposed into multiple groups of measurement sub-data through a singular spectrum decomposition algorithm, and Lagrange multipliers and quadratic penalty terms are introduced into the measurement sub-data to perform a secondary decomposition of the measurement sub-data, thereby enabling long-term trend prediction, periodic change feature extraction, and noise elimination for nonlinear and non-stationary signal data, so that the measurement data set can cooperate with the image data set to assist in constructing an edge detection model, thereby improving the robustness and generalization ability of the model.

[0218] In an embodiment of the present invention, a method is provided for calculating the risk prediction results of a single group of sampling points based on multiple domino effects, and for correlation analysis of at least one group of risk prediction results, thereby being able to evaluate the risk of circumferential limb edema of a patient during multiple sampling measurements, helping patients and doctors to promptly capture any subtle changes in the condition and providing timely data support for the adjustment of treatment plans.

[0219] In an embodiment of the present invention, an adjustable measuring device is provided, which consists of a circumference measuring part, a measuring limit part and a lifting and lowering adjustment part. The circumference measuring part, the measuring limit part and the lifting and lowering adjustment part work together to ensure flexible measurement of the circumference of different patients and different limb areas, thereby ensuring the stability and accuracy of the measurement data. At the same time, it is easy to operate and improves the efficiency of circumference measurement.

[0220] It should be noted that, for the above-mentioned embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described order of actions, because according to the present invention, some steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0221] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field can still combine, add, delete or make other adjustments to the features in the embodiments of the present invention according to the circumstances without conflict and without making creative work, so as to obtain different other technical solutions that do not deviate from the concept of the present invention in essence, and these technical solutions also belong to the scope of protection of the present invention.

Claims

1. A method for measuring limb circumference based on image recognition, characterized in that: The limb circumference measurement method based on image recognition comprises: Acquire time-series limb measurement-related data, and upload the limb measurement-related data to a measurement database, wherein the limb measurement-related data is acquired through an adjustable measurement device, and the limb measurement-related data includes limb image data and measurement sensor data; Combine deep learning and yolov8 model to pre-build edge detection model, measure the database in response to modeling instructions, generate sample sets, randomly divide the sample sets into training sets and validation sets, and train and verify the accuracy of edge detection model through training sets and validation sets; Load limb measurement-related data in real time, identify and decompose the limb measurement-related data, and generate a corrected circumference data set; Taking the circumference data set as input, executing the edge detection model, the edge detection model extracts the circumference edge of the patient's limb in the circumference data set, outputs the circumference measurement result, and synchronously uploads the circumference measurement result to the measurement database; The edge detection model pre-built by combining deep learning and yolov8 model specifically includes: The yolov8 model is used as the initial model of the edge detection model. The yolov8 model consists of a backbone layer, a neck layer, and a head layer. The SPPCSPC module in the backbone layer is frozen, and the Canny operator structure is used to replace the MP pooling layer in the backbone layer. The RFRNet recurrent feature inference network is used in the head layer to detect the edge responsiveness of the limb image data, and the SSNBDL network structure is added in the neck layer. The SSNBDL network structure is used to deconvolute the limb image data and extract the prominent areas in the limb image data. Based on the improved whale algorithm combined with deep learning, the loss function, activation function, and matching strategy of the yolov8 model are optimized, and an edge detection model combining deep learning and the yolov8 model is constructed.

2. The method for measuring limb circumference based on image recognition according to claim 1, characterized in that: The limb circumference measurement method based on image recognition also includes: The measurement database is traversed, and the measurement database responds to the measurement result index instruction to generate a measurement result set. The risk prediction results of a single group of sampling points are calculated based on multiple domino effects. At least one group of risk prediction results is correlated and analyzed to generate an overall assessment of the patient's limb lymphedema risk.

3. The method for measuring limb circumference based on image recognition according to claim 1, characterized in that: The edge detection model is trained and verified by the training set and the verification set, specifically including: The training set is obtained and the Mosaic-Mixup hybrid method is used to perform online data enhancement on the training set. In Mosaic data enhancement, eight sample images are randomly selected from the training set for cropping, and then the cropped samples are randomly spliced ​​into one image as a new sample image. In Mixup data enhancement, four groups of sample images and their corresponding labels are linearly combined to form new sample images and labels. Initialize the pre-built edge detection model and add an encoder to the backbone layer of the edge detection model. The input sequence of the encoder is represented as: Among them, L en represents the length of the encoder input sequence, m is the dimension of the input sequence, N * is the size of the training set or test set, and t is the input time of the sequence; The encoder input sequence and the training set prediction results are input into the Canny operator structure. The Canny operator structure performs Gaussian smoothing and denoising on the encoder input sequence, calculates the image gradient and direction, and obtains an edge enhancement set. The neck layer upsamples the predicted results of the training set and performs element-level addition operations to fuse feature maps of different levels, realizing multi-scale learning of limb image data and measurement sensor data in the training set to obtain an edge detection set. Obtain the edge detection set, and perform iterative training by alternately training the RFRNet cyclic feature inference network and the feature fusion layer. First, train the RFRNet cyclic feature inference network n times, and then train the feature fusion layer once, until the loss function of the head layer converges. Use the RFRNet cyclic feature inference network to gradually reduce the area to be filled through multiple cyclic feature inference modules, perform repair correction on the edge detection set, and output the feature repair set; Obtain a validation set, use the validation set to verify the accuracy of the edge detection model, and determine whether the accuracy of the edge detection model reaches a preset accuracy threshold; If the preset accuracy threshold is met, the edge detection model is output. If the preset accuracy threshold is not met, the Adam optimizer is used to fine-tune the loss function and activation function hyperparameters, and the pre-built edge detection model is returned to initialize, and an encoder is added to the backbone layer of the edge detection model.

4. The method for measuring limb circumference based on image recognition according to claim 1, characterized in that: The method for identifying and decomposing limb measurement-related data to generate a corrected circumference data set specifically includes: Loading measurement-related data, and dividing the measurement-related data into image data sets and measurement data sets based on data types; Obtain a measurement data set, and perform missing value and standardization processing on the measurement data set; Load the measurement data set after missing values ​​and standardization, decompose the data in the measurement data set into multiple groups of measurement sub-data based on the singular spectrum decomposition algorithm, introduce Lagrange multipliers and quadratic penalty terms to the measurement sub-data for secondary decomposition, and obtain measurement decomposition components; The measurement decomposition components are calculated using the following formula: Among them, W ssa (x) is the measurement decomposition component output after the secondary decomposition of the measurement sub-data f(x), p x is the decomposition matrix after the measurement sub-data f(x) is mapped by singular spectrum decomposition, is the central decomposition matrix, k is the number of decomposition matrices, n is the number of groups into which the decomposition matrix is ​​divided during singular spectrum decomposition, α is the quadratic penalty term, ω is the noise tolerance, and γ is the Lagrange multiplier term; Traverse the image data set, extract the key information of the low-resolution images in the image data set based on the iterative back-projection method, and merge them into the high-resolution images based on prior knowledge, filter out the low-resolution images after extracting the key information, and generate a reconstructed image set; The measurements are decomposed into components, reconstructed image sets and combined into a circumference dataset.

5. The method for measuring limb circumference based on image recognition as claimed in claim 4, characterized in that: The method for extracting the circumference edge of a patient's limb in a circumference data set using the edge detection model specifically includes: Obtain a circumference data set, and the encoder encodes the measurement decomposition component and the reconstructed image set respectively, so that the measurement decomposition component serves as a label corresponding to the reconstructed image set; The convolution module of the backbone layer outputs the reconstructed image set containing labels through convolution; The convolution output is max-pooled through 3×3 and 6×6 pooling layers to obtain significant image features; The Canny operator structure performs Gaussian smoothing and denoising on significant image features, calculates image gradients and directions, and obtains edge enhancement sets; Among them, Gaussian smoothing denoising is performed on the significant image features based on the Gaussian denoising function, and the Gaussian denoising function is: Where b is the Gaussian surround scale of the Gaussian denoising function, x i is the information data set of significant image features, η is the Gaussian denoising coefficient, and H(x) is the edge extraction coefficient; Among them, x i is the information data set of significant image features, η is the Gaussian denoising coefficient, q i Initially set weights for the current salient image features; Use the RFRNet cyclic feature inference network to gradually reduce the area to be filled in the edge enhancement set through multiple cyclic feature inference modules, and perform repair and correction on the edge detection set; The neck layer upsamples the edge detection set, obtains the marked edge contour data, and outputs the circumference measurement result.

6. The method for measuring limb circumference based on image recognition according to claim 2, characterized in that: The method of calculating the risk prediction results of a single group of sampling points based on multiple domino effects and performing correlation analysis on at least one group of risk prediction results specifically includes: Obtain at least one set of circumference measurement results, and calculate the swelling probability of a single set of sampling points based on multiple domino effects; Among them, the swelling probability of a single group of sampling points is calculated by the following formula: Among them, J x is the swelling probability of a single group of sampling points, t is the sampling time, T is the total sampling period, y t is the circumference measurement result of the sampling point at the current moment, is the mean absolute error of the circumference measurement results; Obtain the swelling probability of a single group of sampling points, and calculate the risk prediction value of a single group of sampling points based on the risk superposition principle; Performing correlation analysis on at least one set of risk prediction results, and outputting a comprehensive risk value of limb swelling; Among them, J all is the comprehensive risk value of limb swelling, J z is the risk prediction value of a single group of sampling points, c is the number of sampling points, and τ is the risk coupling parameter.

7. A limb circumference measurement system based on image recognition, used to implement a limb circumference measurement method based on image recognition as claimed in any one of claims 1 to 6, characterized in that: The limb circumference measurement system based on image recognition comprises: An adjustable measuring device for acquiring sequential limb measurement-related data and uploading the limb measurement-related data to a measurement database; The model building module combines deep learning and yolov8 models to pre-build edge detection models, measures the database in response to modeling instructions, generates sample sets, randomly divides the sample sets into training sets and validation sets, and trains and verifies the accuracy of the edge detection model through the training sets and validation sets; A data correction module is used to load limb measurement-related data in real time, identify and decompose the limb measurement-related data, and generate a corrected circumference data set; A result acquisition module, taking the circumference data set as input, executes the edge detection model, the edge detection model extracts the circumference edge of the patient's limb in the circumference data set, outputs the circumference measurement result, and synchronously uploads the circumference measurement result to the measurement database; The risk warning module traverses the measurement database, which responds to the measurement result index instruction to generate a measurement result set, calculates the risk prediction results of a single group of sampling points based on multiple domino effects, performs correlation analysis on at least one group of risk prediction results, and generates an overall risk assessment of limb lymphedema in patients; The edge detection model pre-built by combining deep learning and yolov8 model specifically includes: The yolov8 model is used as the initial model of the edge detection model. The yolov8 model consists of a backbone layer, a neck layer, and a head layer. The SPPCSPC module in the backbone layer is frozen, and the Canny operator structure is used to replace the MP pooling layer in the backbone layer. The RFRNet recurrent feature inference network is used in the head layer to detect the edge responsiveness of the limb image data, and the SSNBDL network structure is added in the neck layer. The SSNBDL network structure is used to deconvolute the limb image data and extract the prominent areas in the limb image data. Based on the improved whale algorithm combined with deep learning, the loss function, activation function, and matching strategy of the yolov8 model are optimized, and an edge detection model combining deep learning and the yolov8 model is constructed.

8. The limb circumference measurement system based on image recognition as claimed in claim 7, characterized in that: The adjustable measuring device comprises: Equipment base; A lifting adjustment part, the lifting adjustment part is arranged on the equipment base, the lifting adjustment part comprises an adjusting wheel, an adjusting screw, a lifting threaded seat and a lifting guide rod, the lifting guide rod is fixedly installed on the equipment base, the lifting threaded seat is slidably connected with the lifting guide rod, the lifting threaded seat is threadedly sleeved on the outside of the adjusting screw, one end of the adjusting screw is fixedly installed with an adjusting wheel, and the bottom of the adjusting screw is rotatably connected with the equipment base; a measuring limiter connected to the lifting and lowering adjustment unit, and A circumference measuring part, which is arranged in the measuring limit part and is used to measure the circumference of a patient's limb; The measuring limit part includes an annular guide rail, one side of the annular guide rail is fixedly connected to the lifting threaded seat, an annular slide groove is provided in the annular guide rail, a linkage slider is slidably installed in the annular slide groove, the linkage slider is detachably connected to an adjustable fitting part, and the adjustable fitting part is connected to the circumference measuring part; The adjustable fitting portion comprises: A fitting limit seat, wherein the fitting limit seat is detachably connected to the linkage slider, a fitting limit groove is provided in the fitting limit seat, and an auxiliary lever is detachably mounted on the fitting limit seat; A linkage screw, wherein the linkage screw passes through one end of the fitting limit seat, and the linkage screw is connected to the fitting limit seat through a thread, a linkage rotating wheel is fixedly installed on one end of the linkage screw, and an extrusion slider is rotatably connected to the other end of the linkage screw, and the extrusion slider is slidably installed in the fitting limit groove; A fitting spring fixedly installed in the extrusion slider, one end of the fitting spring away from the extrusion slider is fixedly connected to a driven slider, one end of the driven slider away from the fitting spring is fixedly connected to a driven fitting rod, the driven fitting rod is slidably connected to the fitting limit seat, and the end of the driven fitting rod is fixedly connected to a circumference measuring part; The circumference measuring unit comprises: A measuring mounting seat, wherein the measuring mounting seat is fixedly mounted on the end of the driven fitting rod; A data acquisition group, the data acquisition group comprising a displacement sensor and at least one group of acquisition cameras; A camera adjustment group, the camera adjustment group is arranged in the measurement mounting seat, and the camera adjustment group is used to adjust the acquisition position of the acquisition camera; The camera adjustment group includes: An adjusting motor, wherein the adjusting motor is fixedly embedded in the measuring mounting seat; An inner rotating disc fixedly connected to the output end of the regulating motor, wherein at least one set of inner linkage grooves is provided in the inner rotating disc; An outer limit plate detachably mounted on the measuring mounting seat, wherein at least one set of outer linkage grooves is formed in the outer limit plate; At least one set of adjusting sliders, wherein the adjusting sliders are respectively slidably connected with the inner linkage groove and the outer linkage groove, and a camera support seat is detachably mounted on the adjusting slider, and the acquisition camera is detachably mounted in the camera support seat.

9. A computer-readable storage medium, characterized in that: Computer program instructions are stored, and when the computer program instructions are executed, a limb circumference measurement method based on image recognition as described in any one of claims 1-6 is implemented.

Images