Spinal cord injury assessment method and system based on optical technology

Abstract

The invention relates to the field of biological medicine, and discloses a spinal cord injury assessment method and system based on an optical technology. The method comprises the following steps: synchronously acquiring a hyperspectral reflection image, a linear polarization scattering image and a time-resolved autofluorescence decay curve; respectively generating three types of feature maps representing tissue oxygenation, collagen microstructure disorder and mitochondrial dysfunction; a three-dimensional space-time coupling feature tensor is constructed through pixel-level registration, and key region response is enhanced through a damage sensitivity weighted convolution kernel; and finally, outputting 0-100 spinal cord injury quantitative indexes, and automatically judging the injury level according to a preset threshold value. The system integrates a multi-mode optical acquisition device and an intelligent analysis module, and supports real-time, objective and accurate evaluation in an operation. According to the method, through a sensitivity weighting mechanism of multi-scale quantitative characterization and data driving, the evaluation consistency, sensitivity and clinical interpretability are improved.

Description

Technical Field

[0001] This invention belongs to the field of biomedicine, specifically relating to a method and system for assessing spinal cord injury based on optical technology. Background Technology

[0002] With the increasing application of optical imaging and spectral analysis technologies in neuromedicine, early assessment of spinal cord injury based on non-invasive optical methods has become a research hotspot. Optical technology can reflect physiological parameters such as local blood flow, oxygenation status, and cellular metabolic levels in the spinal cord by detecting the absorption, scattering, and fluorescence characteristics of tissues to specific wavelengths of light, providing potential evidence for the localization and assessment of the extent of injury.

[0003] However, in current clinical practice, the interpretation of such optical images or spectral data is highly dependent on the doctor's subjective experience and lacks a unified and reproducible quantitative standard, resulting in differences in the damage grading results of different doctors for the same case.

[0004] Optical assessment of spinal cord injury primarily focuses on acquiring bio-optical characteristics of the injury site using techniques such as near-infrared spectroscopy, Raman spectroscopy, or optical coherence tomography (OCT), and attempting to establish a correlation between these characteristics and neurological functional status. The core objective of this approach is to capture the dynamic changes in the spinal cord microenvironment in real time through non-invasive methods, thereby achieving an objective assessment of the severity of the injury. However, existing methods generally treat optical signals as isolated indicators, failing to integrate multidimensional clinical parameters such as neurophysiological and motor function scores, making it difficult to construct a comprehensive assessment model with clinical interpretability.

[0005] In existing technologies, optical assessment systems typically only provide raw spectral curves or pseudo-color images, requiring physicians to rely on experience to determine abnormal areas. This approach is neither sensitive enough to identify mild or subclinical lesions nor can it differentiate between different pathological stages such as inflammatory responses, edema, and parenchymal nerve necrosis. Furthermore, the lack of an automated analysis engine results in a time-consuming and unrepeatable assessment process, which is particularly limiting in emergency or intraoperative rapid decision-making scenarios.

[0006] The aforementioned deficiencies result in significant shortcomings in the consistency, sensitivity, and clinical applicability of existing optical assessment methods in terms of damage grading. There is an urgent need for an intelligent assessment system that can deeply couple optical features with neurological functional parameters and has automatic grading capabilities. Summary of the Invention

[0007] This invention provides a method and system for assessing spinal cord injury based on optical technology, aiming to solve the problems of low consistency in injury grading and missed diagnosis of mild injuries caused by the current reliance on doctors' subjective interpretation of spectral images and the lack of objective quantitative indicators in optical assessment.

[0008] This method establishes a quantitative mapping model between the microstructure and metabolic state of spinal cord tissue by constructing a multimodal optical signal acquisition architecture and integrating hyperspectral imaging, polarization scattering characteristic analysis and autofluorescence dynamic response data.

[0009] Based on this, a spatiotemporal coupling feature extraction mechanism and a damage sensitivity weighting strategy are introduced to generate a clinically interpretable quantitative index of damage severity. The standardized damage level judgment result is automatically output according to the preset grading threshold, thereby realizing an objective, accurate, and repeatable automated assessment of spinal cord injury.

[0010] This invention provides a method for assessing spinal cord injury based on optical technology, comprising:

[0011] Hyperspectral reflectance images, linear polarization scattering images, and time-resolved autofluorescence decay curves of the target spinal cord region are acquired simultaneously using a multi-channel optical probe.

[0012] The hyperspectral reflectance image is subjected to spectral normalization and tissue oxygenation index calculation to obtain a first feature map characterizing the distribution of hemoglobin oxygen saturation.

[0013] Muller matrix decomposition is performed on the linearly polarized scattering image to extract the depolarization coefficient and birefringence phase delay, forming a second feature map reflecting the degree of disorder in collagen fiber arrangement; a double exponential decay model is fitted to the time-resolved autofluorescence decay curve to obtain the ratio of fast and slow component lifetimes and the average fluorescence lifetime, forming a third feature map indicating the level of mitochondrial dysfunction.

[0014] The first feature map, the second feature map and the third feature map are registered at the pixel level in the spatial coordinate system to construct a three-dimensional spatiotemporal coupled feature tensor.

[0015] Based on the aforementioned three-dimensional spatiotemporal coupling feature tensor, a damage-sensitive weighted convolution kernel is used for feature enhancement processing. The weight distribution of the damage-sensitive weighted convolution kernel is obtained by training based on historical pathological annotation data, with a focus on enhancing the response intensity to areas of mild edema, axonal fracture, and early inflammatory infiltration.

[0016] Global pooling and fully connected mapping are performed on the enhanced feature tensor to output a spinal cord injury quantification index between 0 and 100.

[0017] The spinal cord injury quantification index is compared with a preset grading threshold range to determine the injury status. When the index is less than 30, it is determined to be no injury; when the index is greater than or equal to 30 and less than 60, it is determined to be mild injury; when the index is greater than or equal to 60 and less than 80, it is determined to be moderate injury; and when the index is greater than or equal to 80, it is determined to be severe injury.

[0018] As one embodiment of the present invention, the multi-channel optical probe includes a coaxially integrated hyperspectral imaging unit, a polarization-modulated illumination unit, and a time-correlated single-photon counting detection unit;

[0019] The hyperspectral imaging unit covers the 400 nm to 1000 nm wavelength band, with a spectral resolution of 5 nm and a spatial resolution of 10 μm. The polarization-controlled illumination unit includes a broadband light source, a polarizer, an electronically controlled liquid crystal phase retarder, and an objective lens arranged sequentially. The electronically controlled liquid crystal phase retarder switches the fast axis direction according to a preset sequence to generate four sets of incident light with different polarization states. The time-correlated single-photon counting detection unit uses a microchannel plate photomultiplier tube with a time resolution of 50 picoseconds and a gated acquisition window of 0-5 nanoseconds.

[0020] As one embodiment of the present invention, the spectral normalization processing specifically includes: dividing the original spectral curve of each pixel by its reflectance value at 550 nanometers to eliminate the influence of light source intensity fluctuations; the tissue oxygenation index calculation adopts a modified Beer-Lambert law model, the expression of which is:

[0021] ;

[0022] in and These represent the reflectance values ​​of the normalized spectrum at 540 nm and 580 nm, respectively.

[0023] In one embodiment of the present invention, the Mueller matrix decomposition obtains the complete 4×4 Mueller matrix by solving 16 independent polarization measurement equations, and then calculates the depolarization coefficient. Phase retardation with birefringence The calculation formulas are as follows:

[0024] ;

[0025] ;

[0026] , , , These are the sub-block elements of the Mueller matrix.

[0027] As one embodiment of the present invention, the expression for the double exponential decay model is:

[0028] ;

[0029] in and For amplitude coefficient, , These represent the lifetimes of the fast and slow fluorescence components, respectively.

[0030] Calculate the lifetime ratio of the fast and slow components With average fluorescence lifetime .

[0031] In one embodiment of the present invention, the pixel-level registration is achieved through a rigid transformation matrix, which is pre-calibrated by a non-rigid registration algorithm guided by preoperative magnetic resonance images to ensure that the spatial error of the three types of feature maps is less than two pixels.

[0032] In one embodiment of the present invention, the damage sensitivity weighted convolution kernel is a 3×3×3 three-dimensional convolution kernel, and its weights... satisfy:

[0033] ;

[0034] in This represents the probability density of pathologically positive markers appearing at the corresponding positions in the historical training set. This is the weighting coefficient, with a value of 0.7.

[0035] As one embodiment of the present invention, the preset grading threshold range is determined based on large sample clinical follow-up data. The lower limit of the mild injury threshold of 30 corresponds to the mean of the minimum quantitative index of Grade B of the American Spinal Cord Injury Association Disability Scale minus one standard deviation. The lower limit of the moderate injury threshold of 60 corresponds to the mean of the minimum quantitative index of Grade C. The lower limit of the severe injury threshold of 80 corresponds to the mean of the maximum quantitative index of Grade D plus half a standard deviation.

[0036] This invention also provides a spinal cord injury assessment system based on optical technology, comprising:

[0037] A multimodal optical signal acquisition device is used to simultaneously acquire hyperspectral reflectance images, linear polarization scattering images, and time-resolved autofluorescence decay curves of the target spinal cord region.

[0038] The feature map generation module is used to process the hyperspectral reflectance image, linear polarization scattering image and time-resolved autofluorescence decay curve respectively to generate a first feature map, a second feature map and a third feature map;

[0039] The spatiotemporal feature fusion module is used to perform pixel-level registration of the first feature map, the second feature map and the third feature map to construct a three-dimensional spatiotemporal coupled feature tensor.

[0040] The damage sensitivity enhancement module is used to enhance the features of the three-dimensional spatiotemporal coupled feature tensor using a pre-trained damage sensitivity weighted convolution kernel;

[0041] The quantization index calculation module is used to perform global pooling and fully connected mapping on the enhanced feature tensor and output the spinal cord injury quantization index.

[0042] The injury level determination module is used to compare the spinal cord injury quantification index with a preset grading threshold range and output a standardized injury level determination result.

[0043] As one embodiment of the present invention, the multimodal optical signal acquisition device is integrated into the optical path system of the surgical microscope. Its working distance is 200 mm and its field of view diameter is 15 mm, which can acquire optical data of the exposed spinal cord surface in real time during surgery.

[0044] As one embodiment of the present invention, the feature map generation module has a built-in dedicated image processor that performs parallel computation pipelines for spectral normalization, Mueller matrix decomposition and fluorescence attenuation fitting, with a single frame processing latency of less than 200 milliseconds.

[0045] As one embodiment of the present invention, the convolution kernel parameters of the damage sensitivity enhancement module are stored in non-volatile memory, supporting the loading of new weight parameters validated by multi-center clinical trials through a cloud update mechanism.

[0046] As one embodiment of the present invention, the output results of the damage level determination module are synchronously displayed on the main control interface of the surgical navigation system and a structured electronic report is generated, which includes quantitative index values, damage level, heat map of suspicious lesion areas and confidence score.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] 1. This invention constructs a multi-scale quantitative characterization system for spinal cord injury from three dimensions—tissue oxygenation, collagen microstructure, and cell metabolism—by simultaneously acquiring three types of optical signals: hyperspectral, polarization scattering, and time-resolved fluorescence, thus overcoming the limitations of single-modal optical imaging information.

[0049] 2. By adopting a damage sensitivity weighting mechanism trained based on historical pathological data, the detection sensitivity of mild injury areas is improved, and the false negative rate is reduced. The generated spinal cord injury quantitative index is continuous and comparable, eliminating the difference in subjective interpretation by doctors, and making the injury grading results highly consistent and repeatable.

[0050] 3. The system is integrated into the optical path of the surgical microscope, supporting real-time intraoperative assessment, providing neurosurgeons with immediate and objective decision-making basis, shortening the diagnostic cycle and improving the accuracy of surgical intervention;

[0051] 4. The grading thresholds are set based on large-sample clinical follow-up data to ensure that the assessment results have a good correspondence with internationally accepted functional scales, which facilitates clinical promotion and multi-center research applications. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the overall technical architecture of the spinal cord injury assessment and evaluation method and system based on optical technology proposed in this invention;

[0053] Figure 2 This is a schematic diagram of the core principle framework of multimodal optical signal fusion and damage sensitivity weighted feature enhancement in this invention;

[0054] Figure 3 This is a flowchart illustrating the logical flow of data acquisition and feature map generation for the three modes of hyperspectral, polarization scattering, and autofluorescence in this invention.

[0055] Figure 4 This is a flowchart illustrating the logical flow framework of the construction of three-dimensional spatiotemporal coupled feature tensors and pixel-level registration in this invention.

[0056] Figure 5 This is a flowchart illustrating the logical process of calculating and classifying the quantitative index of spinal cord injury in this invention.

[0057] Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the multimodal optical acquisition device and the surgical navigation system in this invention. Detailed Implementation

[0058] Please refer to Figures 1 to 6 This invention provides a method and system for assessing spinal cord injury based on optical technology, aiming to solve the problems of low consistency in injury grading and high rate of missed diagnosis of mild injury caused by the reliance on doctors' subjective interpretation of spectral images in current clinical practice.

[0059] This method constructs a multimodal optical signal synchronous acquisition architecture, integrating three types of physical signals: hyperspectral reflectance, linear polarization scattering, and time-resolved autofluorescence, to characterize the blood oxygenation status, collagen microstructure integrity, and mitochondrial metabolic function of spinal cord tissue, respectively.

[0060] Based on this, a spatiotemporal coupled feature tensor is established, and a damage sensitivity weighting mechanism trained based on historical pathological annotation data is introduced to enhance the features of mild injury areas such as early edema, axonal breakage, and inflammatory infiltration. Finally, a clinically interpretable spinal cord injury quantitative index is output, and the injury level is automatically determined according to a preset threshold, so as to achieve objective, accurate and repeatable intraoperative assessment.

[0061] The method includes the following steps:

[0062] S1, which simultaneously acquires hyperspectral reflectance images, linear polarization scattering images, and time-resolved autofluorescence decay curves of the target spinal cord region through a multi-channel optical probe;

[0063] S2, Perform spectral normalization processing and tissue oxygenation index calculation on the hyperspectral reflectance image to obtain the first feature map;

[0064] S3, Perform Mueller matrix decomposition on the linearly polarized scattering image to extract the depolarization coefficient and birefringence phase delay to form a second feature map;

[0065] S4. Fit a double exponential decay model to the time-resolved autofluorescence decay curve to obtain the ratio of the lifetime of the fast and slow components and the average fluorescence lifetime, and form a third feature map.

[0066] S5, perform pixel-level registration of the first feature map, the second feature map and the third feature map in the spatial coordinate system to construct a three-dimensional spatiotemporal coupled feature tensor;

[0067] S6. Based on the three-dimensional spatiotemporal coupled feature tensor, feature enhancement processing is performed using damage-sensitive weighted convolution kernels;

[0068] S7 performs global pooling and fully connected mapping on the enhanced feature tensor, and outputs a spinal cord injury quantification index between 0 and 100.

[0069] S8. The spinal cord injury quantification index is compared with a preset grading threshold range to determine the injury level.

[0070] In step S1, the multi-channel optical probe is a coaxial integrated structure, comprising a hyperspectral imaging unit, a polarization-modulated illumination unit, and a time-correlated single-photon counting detection unit.

[0071] The hyperspectral imaging unit covers the 400 nm to 1000 nm band, with a spectral resolution of 5 nm and a spatial resolution of 10 μm. It adopts pushbroom or snapshot imaging modes to acquire complete spectral cubic data of the target area in a single exposure.

[0072] The polarization-controlled illumination unit consists of a broadband white light source, a polarizer, an electronically controlled liquid crystal phase retarder, and an objective lens in sequence. The electronically controlled liquid crystal phase retarder switches the fast axis direction according to a preset sequence to generate four incident light states: horizontal linear polarization, vertical linear polarization, +45-degree linear polarization, and circular polarization. In each state, a high-resolution polarization camera synchronously records the backscattered image, resulting in a total of four original polarization images.

[0073] The time-correlated single-photon counting detection unit uses a microchannel plate photomultiplier tube as the core detector, with a time resolution of 50 picoseconds and a gated acquisition window set to 0-5 nanoseconds. It excites endogenous fluorophores (such as nicotinamide adenine dinucleotide and flavin adenine dinucleotide) in spinal cord tissue by pulsed laser and records the histogram of fluorescence intensity decaying over time by single-photon counting, forming a time-resolved autofluorescence decay curve.

[0074] The acquisition of the three types of signals is strictly synchronized and triggered by a unified clock source to ensure that the time alignment error is less than 1 millisecond and the spatial field of view overlap is greater than 95%.

[0075] In step S2, the hyperspectral reflectance image is subjected to spectral normalization processing. Specifically, this involves processing the original spectral curve of each pixel. Divide by its reflectivity value at 550 nanometers The normalized spectral curve was obtained. This operation eliminates the overall reflectance shift caused by fluctuations in light source intensity, changes in probe distance, or curvature of tissue surface, while preserving spectral shape characteristics.

[0076] Subsequently, the tissue oxygenation index was calculated based on the modified Beer-Lambert law model. Its expression is:

[0077] ;

[0078] in and These represent the reflectance values ​​of the normalized spectrum at 540 nm and 580 nm, respectively.

[0079] The wavelength selection is based on the difference in absorption peaks between oxyhemoglobin and deoxyhemoglobin in the visible light band. 540 nm is the strong absorption peak of deoxyhemoglobin, and 580 nm is the absorption trough of oxyhemoglobin. The ratio between the two can reflect the local hemoglobin oxygen saturation.

[0080] After the calculation is completed, each pixel is assigned a value. The numerical values ​​form a two-dimensional distribution map, namely the first feature map, whose grayscale or pseudo-color encoding intuitively presents the spatial heterogeneity of the oxygenation state on the spinal cord surface.

[0081] In step S3, Mueller matrix reconstruction is performed on the four original polarization images. Mueller matrix A 4×4 real matrix is ​​used to fully describe the polarization state transformation after light interacts with tissue. By solving 16 independent polarization measurement equations, the Mueller matrix elements corresponding to each pixel are fitted using the least squares method. .

[0082] Subsequently, two key parameters were extracted from the Mueller matrix: the depolarization coefficient. Phase retardation with birefringence The depolarization coefficient reflects a tissue's ability to depolarize incident polarized light; a higher value indicates a more disordered scattering center, commonly seen in areas of collagen fiber breakage or edema. The birefringence phase retardation characterizes the tissue's birefringence properties and is positively correlated with the ordered arrangement of collagen fibers. Their calculation formulas are as follows:

[0083] ;

[0084] ;

[0085] in , , , These are the sub-block elements of the Mueller matrix. The calculated values ​​are... and Two two-dimensional images are generated separately, and after weighted fusion or channel stitching, a second feature map is formed to characterize the integrity and disorder of collagen microstructure in spinal cord white matter.

[0086] In step S4, the time-resolved autofluorescence decay curve is analyzed. Fit the model using a double exponential decay model.

[0087] Endogenous fluorescence in spinal cord tissue mainly originates from mitochondrial coenzymes, and its decay behavior is influenced by cellular metabolic state, exhibiting both rapid and slow kinetic components. The fitted model expression is:

[0088] ;

[0089] in and For amplitude coefficient, , These represent the lifetimes of the fast and slow fluorescence components, respectively; the unit is nanoseconds.

[0090] The fitting process employs a nonlinear least squares optimization algorithm. Initial parameters are set empirically, and the iteration converges to minimize the sum of squared residuals. After fitting, the lifetime ratio of the fast and slow components is calculated. With average fluorescence lifetime The former reflects the degree of metabolic imbalance, while the latter comprehensively characterizes overall mitochondrial function. Each pixel corresponds to a set of... and The values ​​form a dual-channel third feature map, used to indicate the level of energy metabolism disorders in neurons and glial cells.

[0091] In step S5, the first feature map, the second feature map, and the third feature map are spatially registered at the pixel level.

[0092] Due to differences in optical paths, detector resolutions, and sampling methods among the three imaging modalities, subpixel-level geometric distortions and displacements exist between the original images.

[0093] The registration process uses a rigid transformation matrix. Its form is a combination of two-dimensional translation and rotation transformations.

[0094] The transformation matrix is ​​pre-calibrated using a non-rigid registration algorithm guided by preoperative magnetic resonance images:

[0095] Before the operation, T2-weighted images of the patient's spinal cord were acquired. During the operation, the optical field of view was mapped to the magnetic resonance coordinate system by registering the positioning markers on the optical probe with the navigation system.

[0096] Subsequently, during the offline phase, the trimodal optical data of historical cases were densely matched with the corresponding magnetic resonance lesion areas to learn universal cross-modal registration parameters.

[0097] In practical applications, the system calls the pre-stored transformation matrix. Bilinear interpolation resampling is performed on the second and third feature maps respectively to make them perfectly aligned with the spatial grid of the first feature map. After registration, the spatial error of the three types of feature maps is controlled within two pixels, thus constructing a system with a size of [missing information]. The three-dimensional spatiotemporal coupled feature tensor, in which and For the image height and width, The number of feature channels is typically five to seven dimensions. wait).

[0098] In step S6, the damage sensitivity weighted feature enhancement is performed on the three-dimensional spatiotemporal coupling feature tensor.

[0099] The enhancement operation is achieved through 3D convolution with a kernel size of 3×3×3, covering the local spatial neighborhood and feature dimension.

[0100] The weights of the convolution kernel It is not fixed, but rather a damage sensitivity-weighted convolution kernel trained based on historical pathological annotation data.

[0101] Specifically, in a large-scale, multicenter clinical trial, intraoperative optical data and postoperative pathological sections from hundreds of spinal cord injury patients were collected and annotated to the gold standard. The annotations included edema boundaries, axonal fracture areas, and inflammatory cell infiltration foci. Statistical analysis was used to calculate the spatial-feature locations. The probability density is obtained by normalizing the frequency of occurrence in pathologically positive samples. The convolution kernel weights are defined as follows:

[0102] ;

[0103] This is a weighting factor with a value of 0.7, ensuring that the basic response is not completely suppressed.

[0104] This design enables convolution operations to automatically enhance feature responses in pathologically prevalent regions (such as white matter margins and around the central canal), while maintaining linear propagation in normal tissue regions. The enhanced feature tensors not only retain the original multimodal information but also highlight subtle structural changes highly correlated with damage, improving the detection sensitivity for mild lesions.

[0105] In step S7, a global average pooling operation is performed on the enhanced feature tensor, and... Tensor compression to The feature vector is used to eliminate spatial location dependence and focus on the overall injury load. Subsequently, this feature vector is input into a fully connected neural network containing two hidden layers with 64 and 32 nodes respectively. The activation function is a modified linear unit, and the output layer is a single-node linear regression head, which is mapped to the 0-100 range after being constrained by the Sigmoid function, outputting the spinal cord injury quantification index.

[0106] This index is a continuous variable; a higher value indicates a more severe injury. It exhibits good linear separability and clinical interpretability.

[0107] In step S8, the spinal cord injury quantification index is compared with a preset grading threshold range. The threshold is determined based on statistical analysis of large-sample clinical follow-up data.

[0108] The lower limit of the mild injury threshold of 30 corresponds to the mean of the minimum quantitative index of patients in the American Spinal Injury Association Disability Scale (ALS Scale) minus one standard deviation, ensuring coverage of the vast majority of ALS cases;

[0109] The lower limit of the moderate injury threshold of 60 corresponds to the mean minimum quantitative index of grade C patients, reflecting the critical state of partial preservation of motor function;

[0110] The lower limit of the severe injury threshold of 80 corresponds to the mean of the maximum quantitative index of grade D patients plus half a standard deviation, covering complete injury and severe incomplete injury. The judgment rule is as follows:

[0111] When the index is less than 30, it is considered as no damage; when the index is greater than or equal to 30 and less than 60, it is considered as minor damage.

[0112] When the index is greater than or equal to 60 and less than 80, it is judged as moderate damage;

[0113] A score of 80 or higher is considered severe impairment. This grading system correlates well with the International Functional Scale, facilitating clinical decision-making and multicenter studies.

[0114] The system includes a multimodal optical signal acquisition device, a feature map generation module, a spatiotemporal feature fusion module, a damage sensitivity enhancement module, a quantization index calculation module, and a damage level determination module.

[0115] The multimodal optical signal acquisition device is integrated into the coaxial optical path of the surgical microscope. It has a working distance of 200 mm and a field of view diameter of 15 mm. It can perform non-contact scanning directly after the spinal cord is exposed in neurosurgery without the need for additional incisions or dye injection.

[0116] The feature map generation module has a built-in dedicated image processor and adopts a collaborative architecture of field-programmable gate array and graphics processor to achieve parallel pipelined processing of spectral normalization, Mueller matrix decomposition and fluorescence attenuation fitting. The processing latency of a single frame of three-modal data is less than 200 milliseconds, which meets the real-time requirements of the operation.

[0117] The convolution kernel parameters of the damage sensitivity enhancement module are stored in non-volatile memory, and can support the download of new copyright refactoring parameters validated by multi-center clinical trials from the cloud server through a secure encrypted channel, enabling continuous model evolution.

[0118] The output of the damage level determination module is synchronously pushed to the main control interface of the surgical navigation system and presented in the form of a structured electronic report. It includes quantitative index values, textual descriptions of damage levels, heat maps of suspicious lesion areas superimposed on the surgical field image, and scores calculated based on a Bayesian confidence model, which helps doctors quickly identify high-risk areas and adjust surgical strategies.

[0119] During intraoperative operation, the surgeon first places the exposed spinal cord area at the center of the microscope's field of view and initiates the acquisition program. The system completes simultaneous acquisition and preprocessing of three modal data within three seconds; feature map generation and registration within five seconds; enhancement, quantification, and grading within two seconds; and finally, a complete assessment report is output within 10 seconds. This workflow seamlessly integrates into existing surgical workflows without adding extra operational burden, while providing microscopic functional information unavailable through traditional visual observation, thus improving the objectivity and accuracy of intraoperative assessment of spinal cord injuries.

[0120] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0121] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for assessing and evaluating spinal cord injury based on optical technology, characterized in that, include: Hyperspectral reflectance images, linear polarization scattering images, and time-resolved autofluorescence decay curves of the target spinal cord region were acquired simultaneously using a multi-channel optical probe. The hyperspectral reflectance image is subjected to spectral normalization and tissue oxygenation index calculation to obtain a first feature map characterizing the distribution of hemoglobin oxygen saturation. Mueller matrix decomposition is performed on the linearly polarized scattering image to extract the depolarization coefficient and birefringence phase delay, forming a second feature map that reflects the degree of disorder in collagen fiber arrangement. A double exponential decay model was fitted to the time-resolved autofluorescence decay curve to obtain the ratio of fast and slow component lifetimes and the average fluorescence lifetime, which constituted a third feature map indicating the level of mitochondrial dysfunction. The first feature map, the second feature map and the third feature map are registered at the pixel level in the spatial coordinate system to construct a three-dimensional spatiotemporal coupled feature tensor. Based on the aforementioned three-dimensional spatiotemporal coupling feature tensor, a damage-sensitive weighted convolution kernel is used for feature enhancement processing. The weight distribution of the damage-sensitive weighted convolution kernel is obtained by training based on historical pathological annotation data, with a focus on enhancing the response intensity to areas of mild edema, axonal fracture, and early inflammatory infiltration. Global pooling and fully connected mapping are performed on the enhanced feature tensor to output a spinal cord injury quantification index between 0 and 100. The spinal cord injury quantification index is compared with a preset grading threshold range to determine the extent of the injury.

2. The spinal cord injury assessment method based on optical technology according to claim 1, characterized in that, The process of performing spectral normalization processing and calculating the tissue oxygenation index on the hyperspectral reflectance image to obtain a first feature map characterizing the distribution of hemoglobin oxygen saturation includes: Divide the original spectral curve of each pixel by its reflectance value at 550 nm to eliminate the influence of light source intensity fluctuations; The tissue oxygenation index is calculated based on the modified Beer-Lambert law model, and its expression is as follows: ； in and These represent the reflectance values ​​of the normalized spectrum at 540 nm and 580 nm, respectively. The tissue oxygenation index value corresponding to each pixel is used to construct a two-dimensional distribution map as the first feature map.

3. The spinal cord injury assessment method based on optical technology according to claim 2, characterized in that, The process of performing Mueller matrix decomposition on the linearly polarized scattering image to extract the depolarization coefficient and birefringence phase retardation, forming a second feature map reflecting the degree of disorder in collagen fiber arrangement, includes: The complete 4×4 Mueller matrix was obtained by solving 16 independent polarization measurement equations; The depolarization coefficient is calculated based on the Mueller matrix. Phase retardation with birefringence The calculation formulas are as follows: ； ； , , , These are the sub-block elements of the Mueller matrix; The depolarization coefficient and the birefringence phase delay are respectively generated into two-dimensional images and then fused into the second feature map.

4. The spinal cord injury assessment method based on optical technology according to claim 3, characterized in that, The process of fitting a double exponential decay model to the time-resolved autofluorescence decay curve, obtaining the ratio of fast to slow component lifetimes and the average fluorescence lifetime, constitutes a third feature map indicating the level of mitochondrial dysfunction, including: The time-resolved autofluorescence decay curves were analyzed using a nonlinear least squares optimization algorithm. Fitting a double exponential decay model: ； in and For amplitude coefficient, , These represent the lifetimes of the fast and slow fluorescence components, respectively. Calculate the lifetime ratio of the fast and slow components With average fluorescence lifetime ; The ratio of the fast and slow component lifetimes corresponding to each pixel and the average fluorescence lifetime are used to construct a dual-channel image, which is then used as the third feature map.

5. The spinal cord injury assessment method based on optical technology according to claim 4, characterized in that, The step of performing pixel-level registration of the first feature map, the second feature map, and the third feature map in a spatial coordinate system to construct a three-dimensional spatiotemporally coupled feature tensor includes: Call the rigid transformation matrix pre-calibrated by the non-rigid registration algorithm guided by preoperative magnetic resonance images; The second and third feature maps are resampled using bilinear interpolation to align them with the spatial grid of the first feature map. The three registered feature maps are stacked along the feature dimension to form a shape of size [size missing]. The three-dimensional spatiotemporal coupled feature tensor, in which and For the image height and width, This represents the number of feature channels.

6. The spinal cord injury assessment method based on optical technology according to claim 5, characterized in that, The feature enhancement process using damage-sensitive weighted convolution kernels includes: The three-dimensional spatiotemporal coupled feature tensor is convolved using a 3×3×3 three-dimensional convolution kernel; The weights of the 3D convolution kernel satisfy: ； in This represents the probability density of pathologically positive markers appearing at the corresponding positions in the historical training set. These are weighting coefficients; Output the enhanced feature tensor.

7. The spinal cord injury assessment method based on optical technology according to claim 6, characterized in that, The process of performing global pooling and fully connected mapping on the enhanced feature tensor to output a spinal cord injury quantification index between 0 and 100 includes: Perform a global average pooling operation on the enhanced feature tensor to obtain eigenvectors; The feature vector is input into a fully connected neural network containing two hidden layers, with the activation function being a modified linear unit. The spinal cord injury quantification index is output by combining a single-node linear regression head with a Sigmoid function mapping.

8. A spinal cord injury assessment system based on optical technology, characterized in that, include: A multimodal optical signal acquisition device is used to simultaneously acquire hyperspectral reflectance images, linear polarization scattering images, and time-resolved autofluorescence decay curves of the target spinal cord region. The feature map generation module is used to process the hyperspectral reflectance image, linear polarization scattering image and time-resolved autofluorescence decay curve respectively to generate a first feature map, a second feature map and a third feature map; The spatiotemporal feature fusion module is used to perform pixel-level registration of the first feature map, the second feature map and the third feature map to construct a three-dimensional spatiotemporal coupled feature tensor. The damage sensitivity enhancement module is used to enhance the features of the three-dimensional spatiotemporal coupled feature tensor using a pre-trained damage sensitivity weighted convolution kernel; The quantization index calculation module is used to perform global pooling and fully connected mapping on the enhanced feature tensor and output the spinal cord injury quantization index. The injury level determination module is used to compare the spinal cord injury quantification index with a preset grading threshold range and output a standardized injury level determination result.

9. The spinal cord injury assessment system based on optical technology according to claim 8, characterized in that, The feature map generation module has a built-in dedicated image processor that performs parallel computational pipelines for spectral normalization, Mueller matrix decomposition, and fluorescence attenuation fitting.

10. The spinal cord injury assessment system based on optical technology according to claim 9, characterized in that, The convolution kernel parameters of the damage sensitivity enhancement module are stored in non-volatile memory, and new copyright weight parameters validated by multi-center clinical trials can be loaded via a cloud update mechanism. The output of the damage level determination module is simultaneously displayed on the main control interface of the surgical navigation system and generates a structured electronic report, which includes quantitative index values, damage level, heat map of suspicious lesion areas, and confidence score.

Images