A limb volume correction method based on one-dimensional convolutional neural network

Abstract

The application discloses a kind of limb volume correction methods based on one-dimensional convolutional neural network, it is related to medical technical field.The method includes: for multiple training sample limbs, the original length-width data sequence of each sample and reference volume are collected, dynamic feature matrix is established and training data set is constructed;With dynamic feature matrix as input, with its corresponding reference volume as supervision label, train one-dimensional convolutional neural network regression model;For the limb to be corrected, the original length-width data sequence thereof is collected and dynamic feature matrix is established;Finally, input into the model trained to obtain predicted volume, and the initial measurement volume data of the limb is corrected.The application accurately captures the local morphological change trend of the limb by dynamic characteristics, compensates the systematic error and local protrusion misjudgment of prior art by means of neural network end-to-end learning, effectively improves the volume measurement accuracy, reduces the false positive risk, and is suitable for clinical early lymphedema monitoring.

Description

Technical Field

[0001] This invention relates to the field of medical technology, specifically to a method for limb volume correction based on a one-dimensional convolutional neural network. Background Technology

[0002] Lymphedema is a common complication after tumor surgery and radiotherapy. Early detection is crucial to preventing the disease from progressing to irreversible tissue fibrosis and functional impairment. Clinically, accurate measurement of limb volume is a key indicator for screening early lymphedema. An increase of approximately 5% to 10% in volume compared to baseline is clinically significant. While the traditional water replacement method is considered the gold standard for volume measurement, it has limitations such as being cumbersome to perform, highly susceptible to environmental and patient positioning changes, and unsuitable for patients with open wounds. The circumference measurement method estimates volume using an approximate geometric model, but its accuracy is significantly affected by the measurement interval and model simplification assumptions, making it difficult to capture subtle local morphological changes.

[0003] To achieve automated measurement, existing technologies typically employ a motor-driven frame integrating an infrared measurement grating to scan along the limb's axial direction. The length and width data are calculated by capturing the number of photoelectric sensors obscured at each cross-section, and volume is calculated based on an elliptical cylinder accumulation algorithm. However, this approach suffers from two inherent flaws stemming from its physical structure and algorithm model: First, the fixed spacing between the grating sensors creates a measurement blind zone between adjacent sensors, resulting in a systematic underestimation of the acquired length and width data, leading to a persistent underestimation of the calculated volume. Second, the elliptical cylinder accumulation algorithm idealizes the volume elements between adjacent cross-sections as regular elliptical cylinders of equal width. This geometric simplification fails to effectively distinguish between localized bulges and uniform circumferential thickening, easily misinterpreting localized swelling as overall enlargement, leading to an abnormally high overestimation of volume change and introducing the risk of false positives.

[0004] Therefore, the systematic error problem in volume measurement caused by the physical blind zone of the sensor and the geometric simplification algorithm in the existing infrared grating measurement technology makes it difficult for the existing technology to ensure both the convenience and the accuracy of absolute volume measurement and the specificity of local morphological change recognition. As a result, it cannot provide reliable technical support for early and accurate lymphedema monitoring, which constitutes the core technical bottleneck that urgently needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a limb volume correction method based on a one-dimensional convolutional neural network to solve the systematic error problem in volume measurement caused by the physical blind zone of the sensor and the geometric simplification algorithm in the existing infrared grating measurement technology, and to provide reliable technical support for the accurate diagnosis of early lymphedema.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A limb volume correction method based on a one-dimensional convolutional neural network, the method comprising:

[0008] S1. Measure the limbs of multiple training samples, collect the original length and width data sequence and reference volume of each sample, establish a dynamic feature matrix based on the original length and width data sequence, and construct a training dataset containing the original length and width data sequence, dynamic feature matrix and reference volume label.

[0009] S2. Read the dynamic feature matrix from the training dataset as input, and use its corresponding reference volume as the supervision label to train a one-dimensional convolutional neural network regression model.

[0010] S3. For the limb to be corrected, use an infrared measurement grating to collect its original length and width data sequence, and establish a dynamic feature matrix of the limb to be corrected based on this.

[0011] S4. Input the dynamic feature matrix of the limb to be corrected into the trained model to obtain the predicted volume, and correct the initial measurement volume data of the limb according to the predicted volume. The initial measurement volume data is calculated based on the original length and width data sequence of the limb to be corrected by the elliptical cylinder accumulation algorithm.

[0012] Furthermore, the original length and width data sequence mentioned in S1 is obtained by continuously acquiring the length and width data of the cross-section along the limb axis using an infrared measurement grating, specifically as follows:

[0013] The training sample limb is placed in the measurement frame, and the measurement frame is driven to move along the axis of the training sample limb. The measurement frame is provided with at least two infrared measurement gratings. The infrared measurement gratings include an emitter and a receiver. The receiver is composed of multiple photoelectric sensors arranged in a linear fashion.

[0014] Light is emitted from the transmitter toward the limbs of the training sample;

[0015] The light receiver starts counting from the first photoelectric sensor that does not receive light until the last photoelectric sensor that does not receive light, thus obtaining the number of blocked photoelectric sensors.

[0016] Based on the number of obscured photoelectric sensors and the fixed spacing between adjacent photoelectric sensors, the original length and width data sequence of each sample is calculated.

[0017] Further, the calculation to obtain the original length and width data sequence of each sample includes:

[0018] The center distance between adjacent photoelectric sensors is obtained as the fixed spacing.

[0019] The length and width data are obtained by subtracting one from the number of blocked photoelectric sensors and multiplying the difference by the fixed spacing.

[0020] Furthermore, obtaining the reference volume described in S1 includes:

[0021] Within the same measurement interval of the original length and width data sequence for each sample, the data were obtained through actual measurement using the water replacement method.

[0022] When the water replacement method cannot be implemented, the perimeter measurement method combined with the ellipse fitting formula is used to calculate the perimeter.

[0023] Furthermore, the dynamic feature matrix described in S1 is used to characterize the trend of limb shape change along the axis, including at least one of the following: current value feature reflecting the current cross-sectional size, rate of change feature reflecting the rate of change of adjacent cross-sectional sizes, aspect ratio feature reflecting the cross-sectional shape, sliding mean feature reflecting the average level of local dimensions, and sliding standard deviation feature reflecting the degree of fluctuation of local dimensions.

[0024] Furthermore, the construction of the training dataset described in S1 includes:

[0025] The limbs of the multiple training samples are measured to collect their original length and width data sequences and reference volumes to obtain the basic dataset.

[0026] By attaching simulated objects of known volume to the limbs of the training samples, changes in limb volume are simulated, thereby enhancing the basic dataset and constructing a training dataset that includes various limb shapes and volume changes.

[0027] Furthermore, the additional known volume of the simulant includes:

[0028] The flexible simulant is made of medical-grade silicone and its volume is calibrated and ranges from 50 mL to 200 mL.

[0029] The simulated object is attached to typical locations on the limbs where edema is prone to occur;

[0030] After attaching the sample, allow it to stand for a preset time to allow the deformation of the simulated object to stabilize before taking measurements.

[0031] Furthermore, the correction method used in S4 for the initial measured volume data of the limb based on the predicted volume includes:

[0032] Direct replacement correction: The initial measured volume data calculated by the elliptic cylinder accumulation algorithm is replaced with the model predicted volume as the final limb volume correction result;

[0033] Coefficient correction: Calculate the correction coefficients ,in Predict the volume for the model. The initial measured volume data is used; the initial measured volume data is corrected using the correction coefficient α to obtain the corrected volume.

[0034] Another objective of this invention is to provide a limb volume correction system based on a one-dimensional convolutional neural network, wherein the system, when executed, implements the aforementioned limb volume correction method based on a one-dimensional convolutional neural network, comprising:

[0035] The training dataset construction module is configured to measure the limbs of multiple training samples, collect the original length and width data sequences and reference volumes, and construct a training dataset containing a dynamic feature matrix.

[0036] The model training module is configured to read the dynamic feature matrix and its corresponding reference volume labels from the training dataset and train a one-dimensional convolutional neural network regression model.

[0037] The volume correction module is configured to collect data on the limb to be corrected and establish a dynamic feature matrix, and then use the trained model to predict and correct the volume.

[0038] Another object of the present invention is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the limb volume correction method based on a one-dimensional convolutional neural network.

[0039] This invention provides a limb volume correction method based on a one-dimensional convolutional neural network. Through the synergistic innovation of one-dimensional convolutional neural networks and dynamic feature engineering, it has achieved a breakthrough in the field of limb volume measurement. Its beneficial effects are specifically reflected in the following aspects:

[0040] 1. Measurement accuracy achieves bidirectional precise correction. This invention successfully solves the bidirectional error problem inherent in traditional measurement methods. Addressing the measurement blind zone caused by the physical spacing of the infrared grating sensor, a nonlinear mapping relationship from morphological sequence to true volume is established through end-to-end neural network learning, effectively compensating for volume underestimation errors caused by the systematic underestimation of cross-sectional data. Simultaneously, the model accurately identifies morphological differences between local bulges and circumferential thickening through extracted dynamic features, overcoming the geometric simplification defects of the elliptical cylinder accumulation algorithm and significantly reducing the risk of false positives caused by misjudgments of local swelling.

[0041] 2. Significantly improved clinical diagnostic reliability. This method organically combines the convenience of automated measurement with the accuracy of the gold standard. By constructing a multi-dimensional dynamic feature system including rate of change, aspect ratio, and sliding statistics, the model possesses strong adaptability to irregular morphologies such as muscle undulations and postoperative scars, greatly enhancing the detection sensitivity for small volume changes (5%-10%) in early lymphedema. This data-driven calibration approach reduces reliance on operator experience, providing a reliable tool for standardized diagnosis across different medical institutions and effectively supporting early detection and intervention decisions for lymphedema.

[0042] 3. Technological innovation brings multiple additional benefits. This invention pioneers the application of one-dimensional convolutional neural networks to the field of limb volume correction, whose sliding convolution characteristics have a natural fit with limb axial sequence data. This end-to-end solution avoids dependence on complex physical models and automatically learns error compensation mechanisms through a data-driven approach, significantly improving the system's robustness and generalization ability. Furthermore, this method can be directly integrated into existing measurement equipment, achieving a performance leap through software upgrades without hardware modifications. It boasts significant advantages in low cost and ease of deployment, laying a solid foundation for large-scale clinical application. Attached Figure Description

[0043] Figure 1 This is a flowchart of the method of the present invention;

[0044] Figure 2 This is a schematic diagram of the training of the one-dimensional convolutional neural network model of the present invention;

[0045] Figure 3 This is a flowchart of the data acquisition process of the present invention;

[0046] Figure 4 This is a schematic diagram of the structure of a measuring instrument in an embodiment of the present invention;

[0047] Figure 5 This is a schematic diagram of the volume calculation model for the traditional elliptical cylinder accumulation algorithm;

[0048] Figure 6 This is a schematic diagram illustrating the infrared measurement grating measurement principle of the present invention;

[0049] Figure 7 This is a system architecture block diagram of the present invention;

[0050] Figure 8 This is a schematic diagram of the computer device structure of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0052] This embodiment provides a limb volume correction method based on a one-dimensional convolutional neural network, such as... Figure 1 As shown, the method includes:

[0053] S1. Measure the limbs of multiple training samples, collect the original length and width data sequence and reference volume of each sample, establish a dynamic feature matrix based on the original length and width data sequence, and construct a training dataset containing the original length and width data sequence, dynamic feature matrix and reference volume label.

[0054] S2. Read the dynamic feature matrix from the training dataset as input, and use its corresponding reference volume as the supervision label to train a one-dimensional convolutional neural network regression model.

[0055] S3. For the limb to be corrected, use an infrared measurement grating to collect its original length and width data sequence, and establish a dynamic feature matrix of the limb to be corrected based on this.

[0056] S4. Input the dynamic feature matrix of the limb to be corrected into the trained model to obtain the predicted volume, and correct the initial measurement volume data of the limb according to the predicted volume. The initial measurement volume data is calculated based on the original length and width data sequence of the limb to be corrected by the elliptical cylinder accumulation algorithm.

[0057] The above method process will be described in detail below with reference to the embodiments.

[0058] 1. Measure the limbs of multiple training samples, collect the original length and width data sequence and reference volume of each sample, establish a dynamic feature matrix based on the original length and width data sequence, and construct a training dataset containing the original length and width data sequence, dynamic feature matrix and reference volume labels.

[0059] 1. Measuring instrument structure

[0060] This process is performed using a measuring instrument. Figure 4 Its structure mainly consists of several parts, including an outer frame, a measuring frame, a measuring grating, a stepper motor, and a supporting structure. The measuring frame is connected to a pair of measuring gratings on the top and bottom, and on the left and right sides, forming a symmetrical structure. The measuring grating includes linearly arranged LED emitters and photoelectric sensor receivers, with a fixed center-to-center spacing between the sensors (e.g., 2.0 mm).

[0061] The measuring frame is a rigid mechanical structure, usually made of aluminum alloy or carbon fiber composite material, which has high dimensional stability and low thermal deformation coefficient. Its shape is a rectangular open frame, with a pair of infrared measuring gratings installed on the top and bottom edges and the left and right edges, forming an orthogonal dual-axis measuring layout. The measuring frame is coupled to the output shaft of the stepper motor through a guide rail-slider mechanism and driven by the controller according to a preset pulse frequency to achieve equidistant, uniform, and repeatable displacement along the longitudinal axis of the limb (i.e., the Z-axis direction). The displacement accuracy is, for example, 0.1 mm, and a single stroke covers the clinical measurement range (300–500 mm) of a typical upper limb (such as the forearm) or lower limb (such as the calf).

[0062] 2. The original length and width data sequence is obtained by continuously collecting the length and width data of the cross section along the limb axis using an infrared measurement grating.

[0063] like Figure 6 The diagram illustrates the measurement principle of the infrared measurement grating used in this invention. The grating system consists of a linear array of infrared light emitters (red squares in the diagram) and corresponding infrared light receivers (blue dots in the diagram), arranged opposite each other to form a light curtain. During measurement, the limb cross-section (gray circle in the diagram) enters the light curtain area, blocking some of the light. The receivers determine the limb edge position by detecting whether the light path is blocked. The "error" area marked in the diagram visually represents the measurement blind zone caused by the physical spacing of the sensors. This is the physical root cause of the systematic volume underestimation error generated by the traditional elliptical cylinder accumulation algorithm.

[0064] During measurement, the limb to be measured is placed on the supporting structure, and the measuring frame is driven by a stepper motor to translate at a constant speed along the limb axis. Each time it moves by a sampling step (within a range of, for example, 1.0 mm to 5.0 mm), the infrared measuring grating synchronously triggers a contour scan, outputting the length (l) and width (w) values ​​of the current cross-section, forming a two-dimensional sequence arranged in spatial order. Let this be the original length and width data sequence; this sequence is essentially a one-dimensional spatiotemporal sequence, and its index corresponds to the axial position. (h is the axial sampling step size), with a clear spatial topological relationship; this sequence not only contains the static dimensions of each section, but also implicitly contains the continuous evolution information of limb morphology along the axis, which is the basic dataset for subsequent dynamic feature construction.

[0065] When the limb is not placed in the measuring frame, the emitter emits light, which the receiver can receive smoothly. When the limb is placed in the measuring frame, the limb obstructs the light transmission. The receiver starts counting from the first photoelectric sensor that does not receive light, and continues to the last, obtaining the number of photoelectric sensors that are blocked. This number minus one and multiplied by a fixed interval gives the current cross-section length or width of the limb being measured. As the stepper motor drives the measuring frame to move along the limb, a set of length and width data is recorded at intervals, forming the original length and width data sequence. Figure 3 As shown, the specific process is as follows:

[0066] (1) The training sample limb is placed in the measurement frame and the measurement frame is driven to move along the axis of the training sample limb. The measurement frame is provided with at least two infrared measurement gratings. The infrared measurement gratings include an emitter and a receiver. The receiver is composed of multiple photoelectric sensors arranged in a linear fashion.

[0067] The infrared measuring gratings are arranged in the same cross section of the measuring frame, with no less than two sets of independently operating grating units. They can be selected as a pair of vertical (corresponding to the height direction of the limb) and a pair of horizontal (corresponding to the width direction of the limb), forming two orthogonal measuring planes XZ and YZ, so as to synchronously obtain the length (L) and width (W) at each axial position. Each set of gratings uses linearly arranged infrared light-emitting diodes (LEDs, center wavelength 850nm) as emitters and a photoelectric sensor array strictly aligned with them as receivers.

[0068] (2) Light is emitted to the training sample limb through the transmitter. The transmitter continuously emits modulated infrared light (e.g., carrier frequency 38kHz), and the front end of the receiver is equipped with a bandpass filter to suppress visible light and 50Hz power frequency interference. The light travels in a straight line in the air and undergoes diffuse reflection and partial absorption when it encounters the limb to be tested, resulting in a sharp drop in the light intensity received by the photoelectric sensor at the corresponding position behind it. This process does not depend on the reflective properties of the limb surface and is robust to common clinical variables such as skin color, hair, and slight sweat.

[0069] (3) The light receiver starts counting from the first photoelectric sensor that does not receive light until the last photoelectric sensor that does not receive light is recorded, thus obtaining the number of blocked photoelectric sensors.

[0070] The first photoelectric sensor that did not receive light refers to the sensor number that is the first to output a voltage lower than the determination threshold when scanning along the length of the grating (X or Y direction) from left to right (or from top to bottom) in the linear array of photodetectors; the last photoelectric sensor recorded refers to the last sensor number whose output voltage is lower than the determination threshold in the same scanning direction.

[0071] This counting strategy avoids the problem of gradual signal change at the end caused by diffraction and scattering at the limb edge, and avoids over-counting or under-counting that is easy to occur when using the fixed threshold method. For example, when the actual width of a certain cross section is 92.3mm and the sensor spacing is 2.0mm, the theoretical number of occlusions should be 46.15. The system rounds it to 46, but if it starts counting from the first fully dark sensor, it actually captures 46 consecutive dark units, corresponding to a width of 92.0mm, with an error of only 0.3mm.

[0072] (4) Based on the number of obstructed photoelectric sensors and the fixed spacing between adjacent photoelectric sensors, the original length and width data sequence of each sample is calculated. The fixed spacing is the physical distance between the centers of adjacent photoelectric sensors, which is guaranteed by the manufacturing process. For example, it can be 1.0 mm, 1.5 mm or 2.0 mm. It can be calibrated by a laser interferometer before leaving the factory, with an error ≤0.01 mm. This parameter is stored in the device firmware and is automatically called with each measurement.

[0073] The calculation to obtain the original length and width data sequence of each sample includes: obtaining the center distance between adjacent photoelectric sensors as the fixed spacing; subtracting one from the number of obscured photoelectric sensors, and multiplying the difference by the fixed spacing to obtain the length and width data. Specifically, subtracting 1 from the number of obscured sensors N and then multiplying by the fixed spacing d yields L = (N-1) × d (length) or W = (N-1) × d (width). This conversion implicitly assumes that the limb projection contour is a rigid boundary with vertical cutting in the grating direction, and that the sensor response has ideal step characteristics.

[0074] 2. The reference volume is obtained by measuring the water displacement method within the same measurement interval of the original length and width data sequence of each sample, i.e., at the exact same start and end positions and body position as the infrared measurement (e.g., upper limb abducted at 30°, palm facing upward, lower limb extended). When the water displacement method cannot be implemented, the perimeter measurement method combined with the ellipse fitting formula is used for calculation. The same measurement interval is the limb axial range that is completely consistent with the actual scan start and end positions of the infrared grating, and the reference volume is the true physiological volume measured by the clinical gold standard method. The processes of the water displacement method and the perimeter measurement method are as follows:

[0075] (1) Water replacement method: Immerse the limb to be tested in a graduated overflow container, collect the overflow water volume and convert it into volume;

[0076] (2) Perimeter measurement method: Manually measure the perimeter multiple times along the axial direction at fixed intervals (e.g., 2cm). Assuming each infinitesimal element is an elliptical cylinder, take the major semi-axis... Wide half-shaft ( This is an empirical correction factor, ranging from 0.8 to 1.2, and then... The total volume is obtained by summing the volumes.

[0077] 3. The dynamic feature matrix is ​​used to characterize the trend of limb shape change along the axis, including at least one of the following: current value feature reflecting the current cross-sectional size, rate of change feature reflecting the rate of change of adjacent cross-sectional sizes, aspect ratio feature reflecting the cross-sectional shape, sliding mean feature reflecting the average level of local dimensions, and sliding standard deviation feature reflecting the degree of fluctuation of local dimensions.

[0078] The specific calculation methods for various features in the dynamic feature matrix are as follows (for the i-th cross-section in the sequence, let the axial sampling interval be h, usually h=2.0mm):

[0079] (1) Current value characteristics: The measured value of this section is directly used. This includes:

[0080] (Current cross-sectional length value)

[0081] (Current cross-section width value)

[0082] (2) Rate of change characteristics: Reflects the rate of change of adjacent cross-sectional dimensions, calculated using first-order differences. Includes:

[0083] (Rate of change in length, unit: mm / mm)

[0084] (Rate of change in width)

[0085] (3) Length-to-width ratio: reflects the cross-sectional shape.

[0086] (Aspect ratio, dimensionless)

[0087] (4) Moving average characteristics: Reflects the average level of local dimensions, calculated using a moving average with a window size of K=5. ).include:

[0088] (Length moving average)

[0089] (Wide sliding mean)

[0090] (5) Sliding standard deviation characteristics: Reflects the degree of fluctuation in local dimensions, calculated within the same window. Includes:

[0091] (Sliding standard deviation of length)

[0092] (Sliding standard deviation of width)

[0093] For cross-sections at the ends of the sequence that are less than the window length (K=5), the moving mean and moving standard deviation are calculated using valid data. Specifically, for the two beginning cross-sections (i=1, 2) and the two end cross-sections (i=N-1, N) of the sequence, the sliding window is automatically reduced to include only the range of existing data (e.g., for i=1, the window is [1,2,3]; for i=2, the window is [1,2,3,4]), ensuring that all cross-sections have corresponding feature values ​​and avoiding information loss.

[0094] Calculation example: Assume the length values ​​(mm) of five consecutive cross-sections of a limb are: [100, 102, 105, 103, 101]. Then, for the middle cross-section (i=3):

[0095]

[0096]

[0097] Dynamic features are not static statistics, but a set of derived variables with clear physiological significance and spatial locality, generated by deterministic mathematical transformation based on the original sequence. The design principle is to explicitly encode the "trend" rather than the "transient" nature of limb morphology.

[0098] 4. The construction of the training dataset includes: measuring the multiple training sample limbs, collecting their original length and width data sequences and reference volumes to obtain a basic dataset; and enhancing the basic dataset by attaching a simulated object of known volume to the training sample limbs to simulate changes in limb volume, thereby constructing a training dataset containing various limb shapes and volume changes.

[0099] Multiple subjects can be no fewer than 20 individuals with different ages (18–85 years), sex, body mass index, limb circumference distribution characteristics (such as forearm / upper arm, lower leg / thigh) and underlying pathological conditions (including healthy volunteers, patients with upper limb lymphedema after breast cancer surgery, patients with lower limb edema after prostate cancer surgery, etc.) to cover the diversity of limb morphology in real clinical scenarios.

[0100] The training sample limb is placed in a rigid measurement frame equipped with dual orthogonal gratings. A high-precision stepper motor drives it to move at a constant speed along the limb's axis, simultaneously triggering two sets of infrared gratings (top and bottom, left and right) to emit modulated infrared light and capture the occlusion state in real time. The receiver employs a first-to-last dark element counting strategy to eliminate edge blurring interference and accurately obtain the discrete number of occlusions for each cross-section. Combined with a calibrated fixed sensor spacing, the length and width values ​​are directly converted to millimeter-level values. This process does not rely on the optical properties of the limb material, has strong resistance to ambient light interference, and through optional interpolation algorithms or LUT correction, the original discrete sampling error is compressed from ±1 sensor (±2.0mm) to within ±0.2mm. The resulting length and width data sequence has high spatial consistency, high temporal synchronization, and high geometric fidelity, providing a high-quality, low-noise, and reproducible input foundation for dynamic feature construction and one-dimensional convolutional neural network training. This fundamentally ensures the reliability of the data source for subsequent volume correction models, thereby supporting the solution to the systematic volume underestimation problem caused by sensor physical blind spots in the background technology.

[0101] The additional known-volume simulant includes: a flexible simulant made of medical-grade silicone, the volume of which is calibrated and ranges from 50mL to 200mL; the simulant is attached to a typical location on the limb prone to edema; after attachment, it is left to stand for a preset time to allow the simulant to stabilize its deformation before measurement. In this embodiment, the known-volume simulant is a skin-like flexible patch made of medical-grade silicone-based composite gel, with a thickness of, for example, 0.5–3.0mm, a density of, for example, 1.02–1.08g / cm³, and a surface friction coefficient of, for example, 0.4–0.6, which can closely conform to the curved surface of the limb without slipping; the known volume can be calibrated by three-dimensional laser scanning + voxel reconstruction; the simulant is designed according to the clinical lymphedema staging in a three-level incremental mode: Level 1 simulates early edema (volume increase of 2–5mL), Level 2 simulates mid-stage edema (volume increase of 10–30mL), and Level 3 simulates significant swelling (volume increase of 50– 120mL); the application site covers typical edema-prone areas of the limbs—the upper limbs include the dorsal side of the wrist joint, the radial side of the mid-forearm, and the medial side of the elbow crease; the lower limbs include the area above the medial malleolus, the posterior side of the mid-calf, and the medial side of the knee joint; each application is a double-blind procedure (neither the operator nor the subject knows the patch volume number), and the patch is left to stand for 5 minutes after application to allow the material deformation to stabilize before infrared measurement is initiated; the same subject can have 1–3 different volumes and positions of simulants applied sequentially to form a longitudinal sequence of “baseline – mild change – moderate change – severe change”, thereby artificially constructing a gradient of minute volume changes that is difficult to obtain densely under natural conditions.

[0102] In this embodiment, the training dataset ultimately consists of ≥200 independent samples. Each sample includes a set of original length and width sequences (dimension N×2, N∈[120,480]), a set of dynamic feature matrices (dimension N×5, including current value, rate of change, aspect ratio, moving mean, and moving standard deviation), a scalar reference volume label, and structured metadata fields (including subject ID, limb type, application location, simulated object volume, and measurement timestamp). This training dataset can be divided into a training set (70%), a validation set (15%), and a test set (15%). The division can be done by stratified sampling based on subject ID to ensure that data from the same subject does not cross sets and to avoid data leakage. Furthermore, the training dataset can be stored in HDF5 format, supporting streaming loading and online enhancement (such as adding ±0.2mm Gaussian noise to simulate sensor drift and randomly cropping 10% of the sequence along the axis to simulate positioning deviation).

[0103] 2. The dynamic feature matrix is ​​read from the training dataset as input, and its corresponding reference volume is used as the supervision label to train a one-dimensional convolutional neural network regression model. This section details the network architecture, forward propagation process, and hyperparameter configuration for model training in the limb volume correction method based on a one-dimensional convolutional neural network (1D-CNN).

[0104] 1. Network architecture and forward propagation

[0105] The one-dimensional convolutional neural network is a deep learning architecture specifically designed for processing sequential data. As shown in the attached figure, its core structure sequentially includes an input layer, at least one one-dimensional convolutional layer, at least one pooling layer, and a fully connected output layer. The one-dimensional convolutional layer is configured with convolutional kernels of different sizes to extract local discriminative patterns from the dynamic feature sequence of the input.

[0106] In this embodiment, the network input is an N×F matrix, where N represents the total number of cross-sections along the limb axis and F represents the feature dimension (e.g., F=5, corresponding to 5 types of dynamic features). This matrix is ​​first input to the first one-dimensional convolutional layer (specific parameters are: number of convolutional kernels). kernel length Local feature extraction is performed using a convolution-activation-pooling layer (with a stride of S=1 and padding of P=1), outputting a feature map of approximately length N. The feature map is then non-linearly transformed using the ReLU activation function and dimensionality reduced using a max-pooling layer (pooling window length of 2, stride of 2). This basic "convolution-activation-pooling" structure is repeated 2 to 4 times to progressively expand the model's receptive field and compress the sequence length.

[0107] As a preferred embodiment that achieves a good balance between model complexity and performance, the network structure comprises three sequentially connected convolutional-pooling modules, with detailed parameter configurations as follows:

[0108] Module 1: Convolutional layer (32 kernels, 3 kernel length) → ReLU activation function → Max pooling layer (2 windows, 2 strides).

[0109] Second module: Convolutional layer (64 kernels, 3 kernels) → ReLU activation function → Max pooling layer (2 windows, 2 strides).

[0110] Module 3: Convolutional layer (128 kernels, 3 kernels) → ReLU activation function → Max pooling layer (2 windows, 2 strides).

[0111] After processing by the above modules, the network is connected to a Global Average Pooling layer, which averages all activation values ​​of each channel of the feature map into a scalar, thus outputting a feature vector of dimension [C] (where C is the number of channels in the last convolutional layer, for example, 256). Finally, this feature vector is fed into a fully connected output layer with a single neuron, mapping to the final scalar prediction volume. (Unit: cm³)

[0112] 2. Model Training and Hyperparameter Configuration

[0113] The training objective of the model is to minimize the prediction volume. Compared with reference volume The difference between (obtained through gold standard methods such as water displacement) is used. The loss function employs Mean Squared Error (MSE), which is defined as: ,in This refers to the training batch size. During training, the mean absolute error (MAE) and coefficient of determination (R²) are monitored simultaneously as auxiliary evaluation metrics. The specific hyperparameters and configurations for model training are as follows:

[0114] Optimizer: The Adam optimizer is selected, and the initial learning rate is set to... .

[0115] Learning rate adjustment: An exponential decay strategy is adopted, that is, the learning rate is multiplied by a decay factor of 0.9 every 20 training epochs.

[0116] Batch size: The training batch size is set to 32.

[0117] Regularization: A Dropout layer is used after each fully connected layer, with the dropout rate set to 0.5 to prevent overfitting.

[0118] Weight initialization: The weights of all convolutional and fully connected layers are initialized using the He Normal method.

[0119] Input standardization: Before inputting data into the network, Z-score standardization is performed on each dynamic feature dimension (5 dimensions in total) of the training set. The mean (μ) and standard deviation (σ) required for standardization are calculated and saved separately from the training dataset during the training phase; in the subsequent prediction (volume correction) phase, μ and σ calculated using this training set must be used to perform the exact same standardization process on the dynamic feature matrix of the limb to be corrected.

[0120] Early stopping strategy: An early stopping strategy is used during training to prevent overfitting, with a patience value set to 10 epochs. This strategy monitors the loss (MSE) on the validation set. When the validation set loss no longer decreases for 10 consecutive epochs, training is automatically stopped, and the model weights are rolled back to the state with the lowest validation set loss.

[0121] After 50 to 200 rounds of training, the model tends to converge. Once training is complete, the model weights are fixed and deployed on an embedded processor or cloud service platform for inference.

[0122] This embodiment uses dynamic features as model input and the reference volume obtained by the water replacement method as reference label to construct a one-dimensional convolutional neural network regression model. Figure 2 This model performs end-to-end prediction of the overall volume of a limb. Specifically, it extracts local features by sliding convolutional kernels of multiple sizes across the input data, and combines this with a non-linear activation function to enhance its ability to model complex morphologies. The predicted volume is then fed into a fully connected layer after convolutional and max-pooling operations in intermediate layers. This predicted value is compared with a reference value obtained using a water-replacement method, and the network parameters are optimized using backpropagation and gradient descent, with mean squared error as the loss function. After training, the model can achieve more accurate volume predictions without relying on the limitations of traditional formulas.

[0123] Through the above process, this invention achieves the following: ensuring broad morphological coverage through subject diversity, providing precise and traceable labels for minute changes through controllable simulated volume, and covering the real clinical pathological evolution path through multidimensional dynamic patterns, thereby constructing a high-quality supervised learning dataset that combines statistical representativeness and pathological interpretability. Because this dataset explicitly encodes the mapping relationship between "normal morphology and pathological changes," especially strengthening the labeling density for early volume increments of 5 mL, the subsequently trained one-dimensional convolutional neural network can not only learn the static functional relationship between length, width, and volume, but also capture the nonlinear coupling law between morphological distortion and volume response. Therefore, it solves the problem of insufficient model generalization ability caused by the inability of a single subject or limited samples to cover diverse limb morphological and volume changes, and significantly improves the sensitivity and cross-individual adaptability of the calibration model in recognizing minute volume changes in early lymphedema.

[0124] Through steps one and two, continuous axial scanning with infrared gratings is employed to form a spatially ordered sequence, ensuring the integrity of morphological evolution information. Furthermore, gold-standard reference volumes, such as those obtained using the water displacement method, are acquired simultaneously, providing a realistic physical basis for model training. By constructing multi-dimensional dynamic features, the model can explicitly perceive and distinguish between "overall thickening" and "local protrusions." Finally, a one-dimensional convolutional neural network is used for end-to-end regression learning, enabling the system to automatically learn a nonlinear error compensation mechanism that traditional geometric algorithms cannot analyze. This solves the problem of systematic underestimation caused by sensor blind spots and the overestimation of local protrusions caused by the elliptical cylinder accumulation assumption. It significantly improves the absolute volume measurement accuracy and relative change detection sensitivity, reduces the risk of false positives in early lymphedema diagnosis, and enhances adaptability and robustness to irregular limb morphologies such as muscle undulations, bone protrusions, and postoperative scars.

[0125] Third, for the limb to be calibrated, its original length and width data sequence is acquired using an infrared measurement grating, and a dynamic feature matrix of the limb to be calibrated is established based on this. For the limb to be calibrated, the measurement conditions and feature construction process are exactly the same as in step one.

[0126] Fourth, the dynamic feature matrix of the limb to be corrected is input into the trained model to obtain the predicted volume, and the initial measurement volume data of the limb is corrected according to the predicted volume. The initial measurement volume data is calculated based on the original length and width data sequence of the limb to be corrected by the elliptic cylinder accumulation algorithm.

[0127] Specifically, the calculation process for the initial measurement volume data is as follows:

[0128] Using the elliptic cylinder accumulation algorithm ( Figure 5 Based on the original length and width data sequence, the volumetric elements between adjacent cross-sections are... Approximately an elliptical cylinder:

[0129]

[0130] in, Let be the semi-major axis of the ellipse at the i-th cross section. It is a semi-minor axis. This is the axial spacing (i.e., the sampling step size).

[0131] The initial measured volume is obtained by summing up the infinitesimal volumes of the entire limb:

[0132]

[0133] After feature construction is complete, the predicted volume is obtained by inputting it into the trained model. The initial measured volume data of the limb is then corrected based on this predicted volume. The correction process offers two modes: a direct replacement mode, which is simple and reliable and suitable for single measurements; and a coefficient correction mode, which can be used for batch processing of historical data to improve data consistency. The correction methods employed include:

[0134] Direct replacement correction: The initial measured volume data calculated by the elliptic cylinder accumulation algorithm is replaced with the model predicted volume as the final limb volume correction result;

[0135] Coefficient correction: Calculate the correction coefficients ,in Predict the volume for the model. The initial measurement volume data; using the correction coefficient The initial measured volume data is corrected to obtain the corrected volume.

[0136] The reliability of the calibration results is ensured through multiple verifications, including comparative verification with the gold standard method, consistency verification among different operators, and repeatability verification over long time intervals. These verification measures guarantee the accuracy and reliability of the calibration results.

[0137] This embodiment provides a limb volume correction system based on a one-dimensional convolutional neural network. Figure 7 The system, when executed, implements a limb volume correction method based on a one-dimensional convolutional neural network, including:

[0138] The training dataset construction module is configured to measure the limbs of multiple training samples, collect the original length and width data sequences and reference volumes, and construct a training dataset containing a dynamic feature matrix.

[0139] The model training module is configured to read the dynamic feature matrix and its corresponding reference volume labels from the training dataset and train a one-dimensional convolutional neural network regression model.

[0140] The volume correction module is configured to collect data on the limb to be corrected and establish a dynamic feature matrix, and then use the trained model to predict and correct the volume.

[0141] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the limb volume correction method based on a one-dimensional convolutional neural network.

[0142] like Figure 8 As shown, the computer device 400 includes a central processing unit (CPU) 401, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 402 or a program loaded from storage section 408 into random access memory (RAM) 403. The RAM 403 also stores various programs and data required for the operation of the computer device 400. The CPU 401, ROM 402, and RAM 403 are all connected to bus 404. An I / O interface 405 (input / output interface) is also connected to bus 404.

[0143] The following components are connected to I / O interface 405: an input section 406 including a keyboard, mouse, etc.; an output section 407 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 408 including a hard disk, etc.; and a communication section 409 including a network interface card such as a LAN card, modem, etc. The communication section 409 performs communication processing via a network such as the Internet. Drive 410 is also connected to I / O interface 405 as needed. Removable media 411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 410 as needed so that computer programs read from them can be installed into storage section 408 as needed.

[0144] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0145] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. 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 illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0146] 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.

[0147] 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.

[0148] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A limb volume correction method based on a one-dimensional convolutional neural network, characterized in that, The method includes: S1. Measure the limbs of multiple training samples, collect the original length and width data sequence and reference volume of each sample, establish a dynamic feature matrix based on the original length and width data sequence, and construct a training dataset containing the original length and width data sequence, dynamic feature matrix and reference volume label. S2. Read the dynamic feature matrix from the training dataset as input, and use its corresponding reference volume as the supervision label to train a one-dimensional convolutional neural network regression model. S3. For the limb to be corrected, use an infrared measurement grating to collect its original length and width data sequence, and establish a dynamic feature matrix of the limb to be corrected based on this. S4. Input the dynamic feature matrix of the limb to be corrected into the trained model to obtain the predicted volume, and correct the initial measurement volume data of the limb according to the predicted volume. The initial measurement volume data is calculated based on the original length and width data sequence of the limb to be corrected by the elliptic cylinder accumulation algorithm. The dynamic feature matrix described in S1 is used to characterize the trend of limb shape change along the axis, including: current value feature reflecting the current cross-sectional size, rate of change feature reflecting the rate of change of adjacent cross-sectional sizes, aspect ratio feature reflecting the cross-sectional shape, sliding mean feature reflecting the average level of local dimensions, and sliding standard deviation feature reflecting the degree of fluctuation of local dimensions.

2. The limb volume correction method based on a one-dimensional convolutional neural network according to claim 1, characterized in that, The original length and width data sequence mentioned in S1 is obtained by continuously acquiring the length and width data of the cross section along the limb axis using an infrared measurement grating, specifically as follows: The training sample limb is placed in the measurement frame, and the measurement frame is driven to move along the axis of the training sample limb. The measurement frame is provided with at least two infrared measurement gratings. The infrared measurement gratings include an emitter and a receiver. The receiver is composed of multiple photoelectric sensors arranged in a linear fashion. Light is emitted from the transmitter toward the limbs of the training sample; The light receiver starts counting from the first photoelectric sensor that does not receive light until the last photoelectric sensor that does not receive light, thus obtaining the number of blocked photoelectric sensors. Based on the number of obscured photoelectric sensors and the fixed spacing between adjacent photoelectric sensors, the original length and width data sequence of each sample is calculated.

3. The limb volume correction method based on a one-dimensional convolutional neural network according to claim 2, characterized in that, The calculation obtains the original length and width data sequence for each sample, including: The center distance between adjacent photoelectric sensors is obtained as the fixed spacing. The length and width data are obtained by subtracting one from the number of blocked photoelectric sensors and multiplying the difference by the fixed spacing.

4. The limb volume correction method based on a one-dimensional convolutional neural network according to claim 1, characterized in that, The acquisition of the reference volume mentioned in S1 includes: Within the same measurement interval of the original length and width data sequence for each sample, the data were obtained through actual measurement using the water replacement method. When the water replacement method cannot be implemented, the perimeter measurement method combined with the ellipse fitting formula is used to calculate the perimeter.

5. The limb volume correction method based on a one-dimensional convolutional neural network according to claim 1, characterized in that, The construction of the training dataset described in S1 includes: The limbs of the multiple training samples are measured to collect their original length and width data sequences and reference volumes to obtain the basic dataset. By attaching simulated objects of known volume to the limbs of the training samples, changes in limb volume are simulated, thereby enhancing the basic dataset and constructing a training dataset that includes various limb shapes and volume changes.

6. The limb volume correction method based on a one-dimensional convolutional neural network according to claim 5, characterized in that, The additional known volume of the simulants includes: The flexible simulant is made of medical-grade silicone and its volume is calibrated and ranges from 50 mL to 200 mL. The simulated object is attached to typical locations on the limbs where edema is prone to occur; After attaching the sample, allow it to stand for a preset time to allow the deformation of the simulated object to stabilize before taking measurements.

7. The limb volume correction method based on a one-dimensional convolutional neural network according to claim 1, characterized in that, The correction method described in S4 for the initial measured volume data of the limb based on the predicted volume includes: Direct replacement correction: The initial measured volume data calculated by the elliptic cylinder accumulation algorithm is replaced with the model predicted volume as the final limb volume correction result; Coefficient correction: Calculate the correction coefficients ,in Predict the volume for the model. The initial measured volume data is used; the initial measured volume data is corrected using the correction coefficient α to obtain the corrected volume.

8. A limb volume correction system based on a one-dimensional convolutional neural network, characterized in that, When the system is executed, it implements the limb volume correction method based on a one-dimensional convolutional neural network according to any one of claims 1-7, including: The training dataset construction module is configured to measure the limbs of multiple training samples, collect the original length and width data sequences and reference volumes, and construct a training dataset containing a dynamic feature matrix. The model training module is configured to read the dynamic feature matrix and its corresponding reference volume labels from the training dataset and train a one-dimensional convolutional neural network regression model. The volume correction module is configured to collect data on the limb to be corrected and establish a dynamic feature matrix, and then use the trained model to predict and correct the volume.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a limb volume correction method based on a one-dimensional convolutional neural network as described in any one of claims 1-7.

Images