A soft quantification evaluation method and evaluation system for muscle fibrosis caused by nerve injury

Abstract

This invention relates to the fields of neurorehabilitation and medical artificial intelligence, specifically disclosing a soft-quantification assessment method and system for muscle fibrosis caused by nerve injury. The method first acquires continuous spectral data of the target muscle tissue in the visible to near-infrared bands, and calculates pixel-level physical prior scores based on the characteristic absorption and scattering properties of collagen. Then, using these prior scores as physical constraints, a deep learning segmentation network is trained through a joint loss function that incorporates spectral constraint losses. Furthermore, a piecewise soft-threshold mapping function is used to perform continuous weighted integration on the probability map to accurately calculate the equivalent area and proportion of fibrosis, overcoming the underestimation of latent lesions with ambiguous boundaries inherent in traditional hard segmentation. Finally, automatic grading is achieved based on the proportion of fibrosis, and electromyographic features are integrated to differentiate between neurogenic spasticity and myogenic contracture. This invention internalizes optical physical laws into algorithmic constraints, significantly improving quantification accuracy and cross-scenario robustness.

Description

Technical Field

[0001] This invention relates to the fields of neurorehabilitation and medical artificial intelligence, and more specifically, to a soft quantitative assessment method and system for assessing muscle fibrosis caused by nerve injury. Background Technology

[0002] In clinical practice of neurorehabilitation, central nervous system injuries (such as cerebral hemorrhage, cerebral infarction, spinal cord injury, etc.) often lead to limb motor dysfunction. As the disease progresses, the muscle tissue of the damaged limbs often undergoes secondary structural changes, mainly manifested as muscle fiber atrophy, fatty infiltration, and muscle fibrosis caused by abnormal deposition of extracellular matrix. Muscle fibrosis causes muscle tissue to lose its original viscoelasticity and extensibility, manifested as decreased joint range of motion and abnormally increased muscle tone, which seriously affects the reconstruction of the patient's motor function.

[0003] Traditionally, clinical assessment of muscle condition and the degree of spasticity relies primarily on the following four methods:

[0004] First, there are clinical scale assessment methods, such as the modified Ashworth Spasticity Scale and the Tardieu Scale. These methods involve therapists passively moving the patient's joints manually and grading the resistance based on the perceived magnitude. However, scale scoring is highly subjective, and the consistency of assessments among different therapists is often low. More importantly, the scales only reflect the overall mechanical performance of the muscle and cannot distinguish whether the resistance originates from neurogenic spasticity caused by upper motor neuron damage or from myogenic contracture caused by structural changes in the muscle itself. This etiological identification directly determines whether a nerve block (such as botulinum toxin injection) or physical release treatment is used.

[0005] Second, conventional musculoskeletal ultrasound (B-mode ultrasound). Ultrasound is widely used for muscle morphology observation due to its advantages of being non-invasive, real-time, and low-cost. Physicians usually make a rough judgment on the degree of tissue degeneration by visually observing changes in muscle echo intensity (fibrotic areas usually appear as hyperechoic areas). However, the quality of ultrasound images is greatly affected by probe pressure, coupling agent, and beam incidence angle, and speckle noise in ultrasound can easily mask early subtle lesions. In addition, current quantitative ultrasound analysis mostly uses texture analysis based on gray-level histograms or hard threshold image segmentation. These methods simply binarize the blurred boundaries of biological tissue, ignoring the pathological characteristics of fibrosis infiltrating muscles in a gradient manner, leading to a serious underestimation of the extent of the lesion.

[0006] Third, ultrasound elastography, including strain elastography and shear wave elastography, reflects stiffness by measuring the Young's modulus of tissue, achieving quantification to some extent. However, strain elastography is highly dependent on the operator's pressure application technique and has poor repeatability; while shear wave elastography is more objective, it measures the average mechanical properties within the ROI (region of interest) and cannot provide pixel-level information on the spatial distribution of fibrosis. Furthermore, tissue edema and congestion can also increase stiffness, resulting in a lack of specific biomarkers in elastography.

[0007] Fourth, magnetic resonance imaging (MRI) and water / lipid quantification techniques; MRI can assess changes in collagen and water in muscles through T2 mapping or ultrashort echo time series, and is currently recognized as the most accurate non-invasive assessment method; however, its equipment is bulky, the examination cost is high, and it takes a long time, making it impossible to carry out high-frequency, bedside dynamic follow-up assessments in rehabilitation departments, and it is difficult to popularize it as a routine rehabilitation assessment tool.

[0008] In recent years, deep learning technology has made significant progress in medical image analysis; semantic segmentation networks such as U-Net and U-Net++ have been used to delineate muscle contours in ultrasound or MRI images. However, most existing deep learning models belong to a purely data-driven paradigm, meaning that the model only learns the mapping relationship between the image's appearance features such as color and texture and the labels from massive amounts of labeled data. This paradigm has two fatal flaws: first, it lacks physical interpretability, and the model is prone to learning background artifacts unrelated to lesions (such as fascial interfaces and acoustic shadows), leading to a sharp drop in generalization ability when applied across centers and devices; second, the output results are hard classification labels (0 or 1), which cannot express the continuous transition state of tissue from healthy to diseased, and does not conform to the actual biological evolution in clinical practice.

[0009] On the other hand, hyperspectral and multispectral imaging technologies have demonstrated powerful material identification capabilities in agriculture, cultural relic preservation, and remote sensing. Different substances possess unique spectral fingerprint characteristics in specific wavelength bands due to differences in their molecular structure and chemical composition. Studies have shown that collagen, as a major structural protein in the human body, exhibits quantifiable physical differences in its light absorption and scattering properties in the visible to near-infrared band (especially 600–900 nm) compared to normal muscle fibers.

[0010] However, there is currently no technology to systematically introduce this spectral physical property as prior knowledge into the assessment system of rehabilitation medicine, nor has the problem of how to use this property to guide deep learning models to achieve continuous quantization of fuzzy boundaries been solved. Summary of the Invention

[0011] In view of the shortcomings of existing technologies, the purpose of this invention is to provide a muscle fibrosis assessment technology that can overcome the limitations of subjective scales, eliminate dependence on large equipment, possess physical interpretability, and accurately quantify early microscopic lesions.

[0012] To achieve the above objectives, the present invention provides a soft quantitative assessment method for muscle fibrosis caused by nerve injury, comprising the following steps:

[0013] Acquire multispectral image data D∈R of the target muscle tissue H×W×B Where H is the number of pixels in image height, W is the number of pixels in image width, and B is the number of bands. The multispectral image data covers the visible to near-infrared band range.

[0014] Based on the optical response characteristics of collagen in this wavelength range, the pixel-level physical prior score q is calculated. i The q i Used to quantify the degree to which the tissue characteristics at pixel i conform to the pathological features of fibrosis;

[0015] The multispectral image data is input into a deep learning segmentation network, and the joint loss function L is used. total The deep learning segmentation network is trained, and the joint loss function consists of the standard segmentation loss L. seg and spectral constraint loss L spec The weighted composition is expressed as: L total =L seg +γL spec Where γ∈(0,1) is the physical regularization coefficient, used to control the physical prior constraints within a reasonable convergence range;

[0016] The trained deep learning segmentation network outputs a soft segmentation probability map P∈R corresponding to the multispectral image data. H×W The pixel value P in the probability map P x,y ∈[0,1] represents the predicted probability that the tissue at coordinate (x,y) is fibrotic tissue; the pixel value P in the soft segmentation probability map P x,y Due to the physical prior score q during the training phase i The consistency constraint is essentially equivalent to a continuous measure of the physical properties of tissue fibrosis at the corresponding coordinate, rather than a simple classification confidence level.

[0017] Based on the preset piecewise soft threshold mapping function f(P) x,y The equivalent area of ​​fibrillation, S, is calculated based on the soft segmentation probability map P. fib And according to the fiber equivalent area S fib Determine the percentage of fibrosis R fibThis is to enable a quantitative assessment of the degree of muscle fibrosis.

[0018] Furthermore, the calculation of pixel-level physical prior score q i The steps specifically include:

[0019] Calculate the normalized relative absorption area of ​​pixel i within the preset first band interval [λ1, λ2]. The first band corresponds to the characteristic absorption peak of collagen;

[0020] Calculate the normalized reflectance rise slope of pixel i within the preset second band interval [λ3, λ4]. The second band interval corresponds to the scattering characteristic interval of collagen fibers;

[0021] The physical prior score q is calculated using a weighted fusion formula. i : Wherein, α and β are the relative absorption areas, respectively. and the rising slope The weighting coefficients satisfy α+β=1;

[0022] The normalized relative absorption area The calculation formula is: ;

[0023] Among them, R i (λ) represents the reflectance of pixel i at wavelength λ, b i The baseline reflectance of pixel i in the first band interval [λ1, λ2];

[0024] The normalized reflectance rise slope The calculation formula is: ;

[0025] in, Let be the average reflectance of pixel i in the high-reflectivity sub-interval within the second band interval [λ3, λ4]. Δλ is the average reflectance of pixel i in the low reflectance sub-interval within the second band interval [λ3,λ4], and Δλ is the wavelength difference between the corresponding sub-intervals.

[0026] Furthermore, the spectral constraint loss L spec Including sorting loss L rank And soft consistency loss L soft ;

[0027] The sorting loss L rank The expression used to constrain pixels predicted as fibrous regions to have a higher average physical prior score than pixels predicted as background regions is:

[0028] ;

[0029] in, The set of pixels predicted to be fibrous regions. Let m be the set of pixels predicted to be the background region, where m>0 is a preset interval threshold parameter. To avoid the minimum value of division by zero;

[0030] The soft consistency loss L soft The predicted probability p used to constrain the output of the deep learning segmentation network i Approximating the physical prior score q i Its expression is: ;

[0031] Where N = H × W is the total number of pixels. The Euclidean norm is used; this constraint ensures that the model's output in the region of fuzzy boundaries is no longer a black-box confidence level, but rather an interpretable physical equivalent of fibrous density, providing a legal basis for calculating the equivalent area.

[0032] The spectral confinement loss L spec The expression is: ;

[0033] in, δ and δ are respectively the sorting loss L rank and the soft consistency loss L soft The weight coefficients, and satisfying + =1.

[0034] Furthermore, the piecewise soft threshold mapping function f(P) x,y The expression for ) is:

[0035] ;

[0036] in, and The preset probability threshold is 0 < < <1;

[0037] The equivalent area of ​​fibrillation The calculation formula is:

[0038] The percentage of fibrosis The calculation formula is:

[0039] Furthermore, the probability threshold The probability threshold is 0.3. The preferred value is 0.7;

[0040] When P x,y When the value is less than 0.3, the pixel is determined to be either healthy tissue or background noise.

[0041] When 0.3≤P x,y When the value is ≤0.7, the pixel is determined to be in the fibrous transition wetting zone, and the probability value itself is used as the area weight.

[0042] When P x,y When the value is greater than 0.7, the pixel is determined to be a dense fibrous core region and counted as a complete equivalent pixel.

[0043] Furthermore, the physical regularization coefficient γ is set to 0.4; the deep learning segmentation network is a UNet++ architecture, which fuses multi-scale features through nested dense skip connections to adapt to the blurred boundaries of muscle fibrosis.

[0044] Furthermore, the assessment method also includes basing the assessment on the fibrosis percentage R. fib The steps for grading the degree of muscle fibrosis are as follows:

[0045] When R fib When it is less than 5%, it is judged as grade 0, corresponding to normal or very mild fibrosis;

[0046] When 5%≤R fib When the percentage is less than 15%, it is classified as Grade 1, corresponding to mild fibrosis.

[0047] When 15%≤R fib When the percentage is less than 30%, it is classified as Grade 2, corresponding to moderate fibrosis.

[0048] When 30%≤R fib When the percentage is less than 50%, it is classified as Grade 3, corresponding to severe fibrosis.

[0049] When R fib When the percentage is ≥50%, it is classified as Grade 4, corresponding to extremely severe fibrosis.

[0050] Furthermore, this assessment method also includes steps for differential diagnosis of the etiology of spasticity:

[0051] Obtain electromyographic signals or clinical spasticity scale scores from the target muscle tissue;

[0052] Combined with the aforementioned fibrosis ratio R fib Based on electromyography signals or clinical spasticity scale scores, dystonia can be classified into neurogenic spasticity, myogenic contracture, and mixed type.

[0053] The neurogenic spasm is characterized by low R fib The value is accompanied by abnormal electromyography discharges; the characteristic of the myogenic contracture is high R...fib The value is accompanied by low electromyography abnormal discharge or constant passive movement resistance.

[0054] On the other hand, the present invention also provides a soft quantitative assessment system for muscle fibrosis caused by nerve injury, comprising:

[0055] The multispectral imaging module is used to acquire multispectral image data D∈R of the target muscle tissue. H×W×B ;

[0056] The physical prior calculation module is used to calculate pixel-level physical prior scores q based on the optical response characteristics of collagen. i ;

[0057] The model inference module includes a built-in deep learning segmentation network and joint loss function L. total =L seg +γL spec , used to receive the multispectral image data and output a soft segmentation probability map P;

[0058] The quantitative analysis module is used for analysis based on the piecewise soft threshold mapping function f(P) x,y Calculate the equivalent area S of fibrosis. fib and the proportion of fibrosis R fib ;

[0059] The report generation module is used to generate a report containing the fiberization percentage R. fib An evaluation report of grading results and differential diagnostic recommendations.

[0060] Furthermore, the system is integrated into a portable bedside multispectral imaging device or a pushbroom hyperspectral probe.

[0061] Compared with existing technologies, the soft quantitative assessment method and system for nerve injury-induced muscle fibrosis provided by the present invention have at least the following beneficial technical effects:

[0062] First, this invention models the process from an optical physics perspective, transforming the molecular absorption and scattering characteristics of collagen into mathematical constraints to guide the training of deep learning models. This gives the model output a solid biophysical foundation, overcoming the drawbacks of traditional medical AI operating as a black box and relying on apparent features. It significantly improves the model's anti-interference ability and generalization performance in complex clinical settings. In conventional understanding, hyperspectral imaging is mostly used for the classification of static substances. However, human muscles exhibit involuntary micro-movements and significant changes in water content. Those skilled in the art typically lack the motivation and evidence to directly constrain the deep learning training of dynamic medical images using static spectral absorption features. This invention cleverly resolves this adaptation challenge through soft consistency loss, achieving unexpected technical results.

[0063] Secondly, this invention abandons the hard classification approach that causes information truncation, and uses soft probability maps and segmented mapping mechanisms to scientifically preserve information about the fibrotic transition infiltration zone. This allows for the recovery of the area of ​​latent lesions that has been underestimated by traditional methods, making the assessment of rehabilitation potential more objective and forward-looking.

[0064] Third, this invention provides objective R... fib The fibrosis index can clearly distinguish whether joint stiffness is caused by abnormal nerve conduction or connective tissue hyperplasia, thereby avoiding patient suffering and waste of resources due to misdiagnosis and mistreatment, reflecting the concept of precision rehabilitation.

[0065] Fourth, this invention can be deployed based on a portable multispectral device without relying on large-scale radiation equipment, and has the advantages of being non-invasive, real-time, and repeatable, making it suitable for large-scale application in rehabilitation departments. Attached Figure Description

[0066] Figure 1 This is an overall flowchart of a soft quantitative assessment method for muscle fibrosis caused by nerve injury, provided in an embodiment of the present invention.

[0067] Figure 2 This is a block diagram of a soft quantitative assessment system for muscle fibrosis caused by nerve injury, provided as an embodiment of the present invention. Detailed Implementation

[0068] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0069] Example 1:

[0070] This embodiment uses the assessment of calf muscle fibrosis in stroke patients during the recovery period as a specific application scenario, fully presenting the implementation method and outstanding effects of each step.

[0071] Step S1: Multispectral data acquisition.

[0072] Data was acquired using a pushbroom hyperspectral imaging system; the spectral coverage was set to the visible to near-infrared band, specifically 400nm~1000nm, for a total of B=224 continuous bands, to obtain a three-dimensional data cube D∈R of the target muscle region. H×W×B In this embodiment, the image space size is set to H=512 and W=512 pixels, and the single pixel space resolution is about 78μm, which is sufficient to distinguish the microscopic connective tissue texture.

[0073] To eliminate sensor dark current and light source inhomogeneity, rigorous radiometric calibration was performed: dark current images (DN) were acquired. dark Image DN of a standard PTFE whiteboard with a reflectivity of 99% wbConvert the original digital quantization value to absolute physical reflectance:

[0074] ; Let be the reflectance of pixel i at wavelength λ. The original digital count value of the image at wavelength λ is given by element i.

[0075] Subsequently, a Savitzky-Golay filter (quadratic polynomial, window width 11) was used to smooth and denoise the spectral curve of each pixel, and the 37th band (approximately 540nm), the 81st band (approximately 680nm), and the 112th band (approximately 850nm) were selected and mapped to the red (R), green (G), and blue (B) channels respectively to synthesize a three-channel pseudo-color image IRGB.

[0076] Physical calibration can eliminate imaging differences under different lighting and devices, making subsequent network inputs comparable across scenarios; pseudo-color synthesis preserves the spatial structure information of muscle movement, providing a friendly input format for deep learning.

[0077] Step S2: Calculation of pixel-level physical prior scores.

[0078] This embodiment does not directly rely on image appearance features, but rather calculates the physical prior score q pixel by pixel based on the biophysical and optical properties of collagen. i .

[0079] First, calculate the normalized relative absorption area. :

[0080] The first band interval [λ1, λ2] is set to [640nm, 720nm]; the baseline b of pixel reflectance within this interval is calculated by linear interpolation. i (λ)=R i (λ1)+k(λ−λ1) (where k is the slope of the line connecting the start and end points); then calculate the relative area of ​​the absorption valley and normalize it:

[0081] ;

[0082] .

[0083] Then, the normalized scattering slope is calculated. :

[0084] The second band interval [λ3, λ4] is set as [750nm, 860nm]. The mean difference ratio of the high and low reflectivity sub-regions in this interval is calculated and normalized to obtain the result. .

[0085] Finally, weighted fusion is used to construct q i :

[0086] In this embodiment, the weights α=0.55 and β=0.45 were determined through a preliminary grid search, and the following calculations were performed:

[0087] q i =0.55⋅ +0.45⋅ .

[0088] Generate a physical prior graph Q of the same size as the image, where q i The closer the value is to 1, the more significant the collagen deposition optical features of that pixel.

[0089] Independent samples t-test was used to determine the average q of calf muscles in healthy volunteers. i The mean value was 0.11 ± 0.04, while the mean value for clinically diagnosed fibrosis patients was 0.11 ± 0.04. i The value reached 0.76 ± 0.08 (p < 0.001); indicating that q i It establishes a mapping between visual features and pathological essence from the physical level, avoiding the blindness of traditional imaging that relies solely on grayscale for judgment.

[0090] Step S3: Training a deep learning network that incorporates physical priors.

[0091] This embodiment uses UNet++ with nested dense skip connections as the backbone segmentation network, and the encoder uses ResNet-34 pre-trained on ImageNet. During the training phase, not only is the standard segmentation loss used, but the joint loss function and spectral constraint mechanism are also strictly implemented.

[0092] Basic segmentation loss L seg A combination of binary cross-entropy (BCE) and Dice loss, each with a weight of 50%, is used to ensure the geometric accuracy of the segmentation mask.

[0093] Set the interval threshold m=0.4, and use the network to predict the probability p. i >0.5 Generate a mask for the predicted target area Using p i <0.2 Generate background mask ,calculate:

[0094] .

[0095] When the average physical score difference between the target area and the background area is less than 0.4, the loss function imposes a strong penalty, which in turn forces the network to suppress the response to pseudo-highlighted areas such as fascia and subcutaneous veins.

[0096] Apply a spatial weight of 3 to the transition boundary region (pixels with reflectivity gradient > 0.1) to force the network to predict probabilities closer to the physical score:

[0097] .

[0098] Set the physical regularization coefficient γ = 0.4, with L total =L seg +0.4(L rank +L soft To optimize the process, the AdamW optimizer was trained for 300 epochs.

[0099] Ablation experiments showed that without physical constraints, the model's mIoU was only 81.3%, and it was prone to missegmenting the intermuscular fascia. After introducing the joint loss in this step, the mIoU increased to 89.7%, and the IoU at the fibrosis boundary increased by 13.2%. More importantly, the probability of the model output no longer fluctuated drastically with equipment changes, and the coefficient of variation (CV) in cross-center tests decreased from 22% to 5.1%.

[0100] Step S4: Output the soft segmentation probability map.

[0101] After model training converges, only the pseudo-color image (IRGB) of the test subject is input during the inference phase, eliminating the need to repeatedly calculate the physical prior score (the physical knowledge has been internalized into the network parameters during training). Through forward propagation, the output is a soft-segmented probability map P∈R of the same size as the input. 512×512 P for each coordinate (x, y) in the graph x,y ∈[0,1] continuously distributed.

[0102] At this time, P x,y This is no longer the traditional classification confidence level; due to the L... (The sentence is incomplete and requires more context to translate accurately). soft Constraint, P x,y =0.6 indicates that 60% of the physical properties at this location conform to the collagen infiltration characteristics, providing a measurement basis for subsequent soft quantization.

[0103] During the model training phase, soft consistency loss L is used. soft Forced network prediction probability p i Approximating the physical prior score q i , making q i The physical laws governing collagen absorption and scattering are encoded into the network weight parameters; therefore, although q is no longer explicitly calculated during the inference phase... i However, the continuous probability P output by the network x,y It is already an implicit mapping of the physical prior during the training period (i.e., P). x,y ≈qi inferred This mechanism ensures that the subsequent quantification process is always based on the underlying optical physical laws, rather than random guesses of apparent visual features.

[0104] Step S5: Segmented soft threshold mapping and equivalent area quantization.

[0105] Since each value in the aforementioned probability graph P carries the physical prior q i The meaning of the attribute metric, and the piecewise soft threshold mapping function f(P) described in this embodiment. x,y It is not a simple mathematical truncation, but a continuous aggregation of the degree of physical infiltration:

[0106] When 0.3≤P x,y When the value is ≤0.7, its probability value itself is retained as the fractional pixel weight, which is essentially a factor of the physical prior score q. i Quantitative integration in the fibrotic transition zone (microscopic spore / collagen diffusion zone);

[0107] When P x,y When the value is greater than 0.7, the region is considered to have significant physical characteristics and is equivalent to a dense fibrotic core region.

[0108] Based on this, the accumulated equivalent area of ​​fibrosis S fib In a physical sense, it is equivalent to the physical prior fraction q. i The soft quantization summation within the entire observation field overcomes the underestimation bias of latent lesions caused by traditional hard segmentation.

[0109] Set the first probability threshold =0.3, second probability threshold =0.7, construct the mapping function:

[0110] ;

[0111] Traverse the entire probability map and accumulate the fibrosis equivalent area. And calculate the percentage:

[0112] .

[0113] Taking a stroke patient with a 10-month disease course as an example, the fibrosis area calculated by traditional hard segmentation (P>0.5 truncation) was 14.1%. Using the soft thresholding method of this embodiment, after probability-weighted inclusion of 12453 transition zone pixels between 0.3 and 0.7, the final R... fib =16.6%, recovering approximately 17.7% of the occult lesion area; subsequent micro-tissue biopsy (hydroxyproline content determination) confirmed that the patient's true collagen infiltration rate was approximately 16.9%, proving that this soft quantification method is significantly superior to hard segmentation and is closer to the biological true value.

[0114] Step S6: Grading of fibrosis and differential diagnosis of spasticity etiology.

[0115] System reads R fib Numerical values, strictly adhering to gradient criteria. For example, when R...fib When the collagen content is 34.2%, it is automatically classified as Grade 3 (severe fibrosis), and the system will link to a pre-set clinical knowledge base in the background to output the following advice: "Extensive collagen deposition has been detected. It is recommended to perform ultrasound-guided botulinum toxin injection and shockwave release. Do not perform forceful traction."

[0116] In addition, it can be simultaneously connected to a surface electromyography (sEMG) device to collect integrated electromyography (iEMG) values ​​during passive ankle dorsiflexion. Combined with R fib Execute authentication logic:

[0117] The patient R fib >30% and iEMG increase <10% during passive activity (no abnormal nerve discharge), the system clearly diagnoses predominantly myogenic contracture (Type A); if R fib If the rate is 8% and the iEMG increases sharply by >50%, the diagnosis is predominantly neurogenic spasm (Type B).

[0118] In the past, myogenic contractures were often misdiagnosed as spasms in clinical practice, leading to the indiscriminate injection of botulinum toxin, with an effectiveness rate of less than 40%. After introducing this procedure, the treatment efficacy rate in pilot cases increased to 82%, avoiding waste of medical resources and patient suffering.

[0119] Example 2:

[0120] This embodiment provides a soft quantitative assessment system for muscle fibrosis caused by nerve injury. It is a physical system corresponding to the assessment method shown in Embodiment 1. The hardware composition, software logic and collaborative process of each module are described in detail below.

[0121] The assessment system is integrated into a portable bedside rehabilitation assessment cart, which includes both hardware and software components.

[0122] The hardware components include:

[0123] Multispectral imaging module 101: It adopts a customized line push-broom hyperspectral camera (CMOS sensor, spatial resolution 1280×1024) and is equipped with a dual-band cold light source with wavelengths of 650K and 850K; the module uploads the raw data cube in real time through a gigabit network port.

[0124] Edge computing host: Configured with NVIDIA Jetson AGXOrin module (32GB memory, 200TOPS computing power), with physical prior computing module, model inference module and quantization analysis module deployed inside, responsible for all algorithm operations.

[0125] Touchscreen display terminal: Medical touchscreen used to display pseudo-color original image, soft probability heat map and final report.

[0126] The software modules include:

[0127] Physics Prior Calculation Module 102: Utilizing C++ and OpenCV, an efficient parallel algorithm can be written to calculate the absorption area within 45ms after receiving a 512×512×224 data cube. With scattering slope Pixel-by-pixel calculation to generate q i The matrix is ​​used for training.

[0128] Model inference module 103: It uses TensorRT to convert the trained UNet++ model into an FP16 precision engine; after receiving the pseudo-color image, the single inference time is short, and the output probability map is sent to the front end via the Web Socket protocol.

[0129] Quantitative Analysis Module 104: A microservice written in Python that receives probability graph JSON data and executes... =0.3, The soft threshold accumulation logic with a value of 0.7 is used to calculate R. fib Numerical value and corresponding hierarchical code (0~4 levels).

[0130] Report generation module 105: Utilizes Vue.js + html2canvas technology to automatically generate patient information, R... fib Numerical values, grading conclusions, identification types, and muscle images overlaid with Jet pseudo-color heatmaps are combined to form an assessment report that conforms to medical record writing standards, supporting direct printing or uploading to the hospital's HIS system.

[0131] In addition, a computer-readable storage medium can be provided, which can be an industrial-grade SSD solid-state drive, storing a pre-trained UNet++ network weight file (best_unetpp.pth), a physical prior calculation algorithm library, and a soft threshold quantization program as shown in Examples 1 and 2; when the processor of the edge computing host calls the program in the storage medium, all the method steps of Example 1 are implemented.

[0132] Finally, it should be noted that this article uses specific examples to illustrate the principles and implementation methods of the present invention. The above description of the embodiments is only for the purpose of helping to understand the core ideas of the present invention. Without departing from the principles of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the protection scope of the present invention.

[0133] For example, those skilled in the art should understand that the selected band range [640nm, 720nm], weight α=0.55, and UNet++ architecture in this embodiment are all preferred ways to implement the present invention. Without changing the physical prior constraints and the core concept of soft segmentation quantization, replacing multispectral imaging with photoacoustic imaging, replacing UNet++ with SegFormer or Swin-Unet, and fine-tuning the specific values ​​of τ1 / τ2 all fall within the scope of protection of equivalent substitutions in the claims of the present invention. Those skilled in the art can implement these substitutions by combining common knowledge without any creative effort.

Claims

1. A soft quantitative assessment method for muscle fibrosis caused by nerve injury, characterized in that, Includes the following steps: Acquire multispectral image data D∈R of the target muscle tissue H×W×B Where H is the number of pixels in image height, W is the number of pixels in image width, and B is the number of bands. The multispectral image data covers the visible to near-infrared band range. Based on the optical response characteristics of collagen in this wavelength range, the pixel-level physical prior score q is calculated. i The q i Used to quantify the degree to which the tissue characteristics at pixel i conform to the pathological features of fibrosis; The multispectral image data is input into a deep learning segmentation network, and the joint loss function L is used. total The deep learning segmentation network is trained, and the joint loss function consists of the standard segmentation loss L. seg and spectral constraint loss L spec The weighted composition is expressed as: L total =L seg +γL spec Where γ∈(0,1) is the physical regularization coefficient; The trained deep learning segmentation network outputs a soft segmentation probability map P∈R corresponding to the multispectral image data. H×W The pixel value P in the probability map P x,y ∈[0,1] represents the predicted probability that the tissue at coordinate (x,y) is fibrotic tissue; the pixel values ​​in the soft segmentation probability map P are affected by the physical prior score q. i The physical measurement meaning of fibrous density is derived from the constraints of fibrous density. Based on the preset piecewise soft threshold mapping function f(P) x,y The equivalent area of ​​fibrillation, S, is calculated based on the soft segmentation probability map P. fib And according to the fiber equivalent area S fib Determine the percentage of fibrosis R fib This is to enable a quantitative assessment of the degree of muscle fibrosis.

2. The soft quantitative assessment method for muscle fibrosis caused by nerve injury according to claim 1, characterized in that, The calculation of pixel-level physical prior score q i The steps specifically include: Calculate the normalized relative absorption area of ​​pixel i within the preset first band interval [λ1, λ2]. The first band corresponds to the characteristic absorption peak of collagen; Calculate the normalized reflectance rise slope of pixel i within the preset second band interval [λ3, λ4]. The second band interval corresponds to the scattering characteristic interval of collagen fibers; The physical prior score q is calculated using a weighted fusion formula. i : Wherein, α and β are the relative absorption areas, respectively. and the rising slope The weighting coefficients satisfy α+β=1; The normalized relative absorption area The calculation formula is: ; Among them, R i (λ) represents the reflectance of pixel i at wavelength λ, b i The baseline reflectance of pixel i in the first band interval [λ1, λ2]; The normalized reflectance rise slope The calculation formula is: ; in, Let be the average reflectance of pixel i in the high reflectance sub-interval within the second band interval [λ3, λ4]. Δλ is the average reflectance of pixel i in the low reflectance sub-interval within the second band interval [λ3,λ4], and Δλ is the wavelength difference between the corresponding sub-intervals.

3. The soft quantitative assessment method for muscle fibrosis caused by nerve injury according to claim 1, characterized in that, The spectral confinement loss L spec Including sorting loss L rank And soft consistency loss L soft ; The sorting loss L rank The expression used to constrain pixels predicted as fibrous regions to have a higher average physical prior score than pixels predicted as background regions is: ； in, The set of pixels predicted to be fibrous regions. Let m be the set of pixels predicted to be the background region, where m>0 is a preset interval threshold parameter. To avoid the minimum value of division by zero; The soft consistency loss L soft The predicted probability p used to constrain the output of the deep learning segmentation network i Approximating the physical prior score q i Its expression is: ; Where N = H × W is the total number of pixels. It is the Euclidean norm; The spectral confinement loss L spec The expression is: ; in, δ and δ are respectively the sorting loss L rank and the soft consistency loss L soft The weight coefficients, and satisfying + =1.

4. The soft quantitative assessment method for muscle fibrosis caused by nerve injury according to claim 1, characterized in that, The piecewise soft threshold mapping function f(P) x,y The expression for ) is: ； in, and The preset probability threshold is 0 < < <1; The equivalent area of ​​fibrillation The calculation formula is:

5. The percentage of fibrosis The calculation formula is:

6. The soft quantitative assessment method for muscle fibrosis caused by nerve injury according to claim 4, characterized in that, The probability threshold The probability threshold is 0.

3. The preferred value is 0.7; When P x,y When the value is less than 0.3, the pixel is determined to be either healthy tissue or background noise. When 0.3≤P x,y When the value is ≤0.7, the pixel is determined to be in the fibrous transition wetting zone, and the probability value itself is used as the area weight. When P x,y When the value is greater than 0.7, the pixel is determined to be a dense fibrous core region and counted as a complete equivalent pixel.

7. The soft quantitative assessment method for muscle fibrosis caused by nerve injury according to claim 1, characterized in that, The physical regularization coefficient γ is 0.4; the deep learning segmentation network is a UNet++ architecture, which fuses multi-scale features through nested dense skip connections to adapt to the blurred boundaries of muscle fibrosis.

8. The soft quantitative assessment method for muscle fibrosis caused by nerve injury according to claim 1, characterized in that, The assessment method also includes assessment based on the fibrosis percentage R. fib The steps for grading the degree of muscle fibrosis are as follows: When R fib When it is less than 5%, it is judged as grade 0, corresponding to normal or very mild fibrosis; When 5%≤R fib When the percentage is less than 15%, it is classified as Grade 1, corresponding to mild fibrosis. When 15%≤R fib When the percentage is less than 30%, it is classified as Grade 2, corresponding to moderate fibrosis. When 30%≤R fib When the percentage is less than 50%, it is classified as Grade 3, corresponding to severe fibrosis. When R fib When the percentage is ≥50%, it is classified as Grade 4, corresponding to extremely severe fibrosis.

9. A soft quantitative assessment method for muscle fibrosis caused by nerve injury according to claim 1, characterized in that, The assessment method also includes steps for differential diagnosis of the etiology of spasticity: Obtain electromyographic signals or clinical spasticity scale scores from the target muscle tissue; Combined with the aforementioned fibrosis ratio R fib Based on electromyography signals or clinical spasticity scale scores, dystonia can be classified into neurogenic spasticity, myogenic contracture, and mixed type. The neurogenic spasm is characterized by low R fib The value is accompanied by abnormal electromyography discharges; the characteristic of the myogenic contracture is high R... fib The value is accompanied by low electromyography abnormal discharge or constant passive movement resistance.

10. A soft quantitative assessment system for muscle fibrosis caused by nerve injury, characterized in that, include: The multispectral imaging module is used to acquire multispectral image data D∈R of the target muscle tissue. H×W×B ; The physics prior calculation module is used to calculate pixel-level physics prior scores q based on the optical response characteristics of collagen. i ; The model inference module includes a built-in deep learning segmentation network and joint loss function L. total =L seg +γL spec , used to receive the multispectral image data and output a soft segmentation probability map P; The quantitative analysis module is used for analysis based on the piecewise soft threshold mapping function f(P) x,y Calculate the equivalent area S of fibrosis. fib and the proportion of fibrosis R fib ; And a report generation module, used to generate a report containing the fiber percentage R. fib An evaluation report of grading results and differential diagnostic recommendations.

11. A soft quantitative assessment system for muscle fibrosis caused by nerve injury according to claim 9, characterized in that, The system is integrated into a portable bedside multispectral imaging device or a pushbroom hyperspectral probe.

Images