Endometrial tissue spectrum data processing method and equipment based on Raman spectrum

Abstract

The invention discloses a Raman spectrum-based endometrial tissue spectral data processing method and equipment, and relates to the technical field of data processing. The method comprises the following steps: firstly, calculating a gradient change sequence of to-be-processed spectral data, carrying out continuity verification in combination with wavelength and width, and accurately dividing a Raman characteristic peak interval and a background baseline interval; then, iterative baseline fitting processing based on a Raman peak protection mechanism is adopted, blood fluorescence interference is eliminated, and pure Raman spectrum data is obtained. And further, inputting the pure spectral data into a pre-constructed segmented fusion feature extraction network, extracting local wave band features, fusing the local wave band features with global features, and outputting a data classification result. According to the embodiment of the invention, the problem of difficult feature extraction caused by low signal-to-noise ratio of a bleeding sample and non-uniform background in related technologies can be effectively solved, the integrity of spectrum fingerprint information is ensured while strong fluorescence interference is thoroughly eliminated, and the accuracy of data feature extraction and the robustness of classification and recognition are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a method and device for processing endometrial tissue spectrum data based on Raman spectrum. BACKGROUND

[0002] Endometrial cancer is a common malignant tumor of the female reproductive tract, and the biochemical composition changes in the tissue can be detected by Raman spectrum technology. This technology uses the scattering effect of light and matter interaction to obtain molecular bond vibration information, and is suitable for ex vivo or in vivo analysis of tissues. In the clinical detection scene, endometrial tissue is often accompanied by bleeding. The broadband fluorescence background signal produced by hemoglobin in the blood will be superimposed on the Raman scattering signal. This broadband background presents a nonlinear distribution in the spectral morphology, which masks the characteristic peaks in the fingerprint region and changes the baseline morphology of the original spectrum, resulting in interference with the subsequent extraction of tissue lesion characteristics.

[0003] In order to eliminate the interference of the fluorescence background on the spectrum data, the related technology usually uses adaptive iteratively reweighted penalized least squares (airPLS) to perform baseline correction processing on the spectrum data. The core idea of this method is to dynamically adjust the weight in the iteration process, to give low weight to the suspected peak region and high weight to the baseline region, to fit a smooth fluorescence background curve by minimizing the objective function, and to subtract it from the original spectrum. The airPLS method can effectively separate the background and the signal by adjusting the smoothing parameter (lambda) when processing ideal spectrum data with good signal-to-noise ratio, and is the most widely used baseline subtraction algorithm in the field of spectrum preprocessing.

[0004] However, when processing endometrial bleeding samples, the curvature of the hemoglobin fluorescence background does not change uniformly, and its waveform width overlaps with the wide Raman peak of biological macromolecules such as amide band. The airPLS relies on a single global smoothing parameter, which makes it difficult to consider the local fitting needs of the entire waveband: if the parameter focuses on fitting the rapidly changing background, the wide Raman peak overlapping with it will be misjudged as background and will be reduced (overfitting); if the parameter focuses on protecting the wide Raman peak, it will be difficult to fit the non-uniformly changing background baseline, resulting in residual fluctuating background noise in the spectrum after subtraction (underfitting). This local wave non-linear distortion caused by the mismatch between the global parameter and the local waveform characteristics makes the feature signal-to-noise ratio of the processed spectrum differ greatly between different wavebands, making it difficult for the subsequent classification model to distinguish whether the signal in some waveband is a real pathological feature or a residual background artifact, and reducing the accuracy of subsequent class judgment based on spectrum features.

[0005] In addition, the existing Raman spectrum classification methods mostly use general convolutional neural networks, often ignoring the local feature correlation of Raman spectrum as a molecular fingerprint. After baseline correction, the weak local features remaining in the spectrum are crucial for distinguishing highly similar tissue categories, but conventional global feature extraction networks are difficult to capture these subtle local band differences, limiting further improvement of detection accuracy. SUMMARY

[0006] The present application provides a Raman spectrum-based endometrial tissue spectrum data processing method and device, aiming to solve the problems of difficult feature extraction and low classification accuracy caused by low signal-to-noise ratio and non-uniform background of bleeding samples in related technologies.

[0007] In a first aspect, the present application provides a Raman spectrum-based endometrial tissue spectrum data processing method, applied to a spectrum data processing device, the method comprising: obtaining original Raman spectrum data of a to-be-detected endometrial tissue, the original Raman spectrum data containing background interference signals; performing denoising and background subtraction processing on the original Raman spectrum data using a preset bleeding interference adaptive suppression strategy to obtain purified spectrum data; the bleeding interference adaptive suppression strategy includes sequentially executed smoothing filtering processing, iteration baseline fitting processing based on a Raman peak protection mechanism, and non-uniform spline optimization processing; according to the waveform morphological features of the purified spectrum data, the purified spectrum data is adaptively segmented to obtain a plurality of spectral segment data; the spectral segment data is input into a preset segmented fusion feature extraction network to extract a target feature vector of the to-be-detected endometrial tissue; the target feature vector is input into a preset integrated classification model to output a classification result label corresponding to the original Raman spectrum data, the classification result label representing the Raman spectrum physicochemical feature category of the to-be-detected endometrial tissue, used to assist medical personnel in clinical evaluation; the integrated classification model performs feature filtering on the target feature vector through a waveband screening layer, and then analyzes and fuses the filtered features through a multi-classifier layer and an attention weighting mechanism to output the classification result label.

[0008] By employing the above technical solution, the spectral data processing equipment first utilizes an adaptive suppression strategy for bleeding interference to smooth and baseline-correct the raw Raman spectral data in stages, gradually removing high-frequency noise and non-uniformly distributed hemoglobin fluorescence background to restore the true Raman scattering intensity of the tissue. Next, the equipment performs adaptive segmentation based on waveform morphology characteristics, conforming to the fingerprint spectral band distribution logic of biomolecules and avoiding fixed segmentation that severs key chemical bond information. Subsequently, the equipment uses a segmented fusion feature extraction network to deeply mine the intrinsic correlations of the spectral data from local to global perspectives. Finally, the equipment outputs classification results through a multi-level mechanism integrating a classification model. In summary, this solution systematically overcomes the problems of background masking and baseline distortion caused by endometrial bleeding, constructs a complete processing chain from high-quality data cleaning to high-dimensional feature analysis, and improves the accuracy of extracting lesion features from endometrial tissue and the reliability of classification.

[0009] In conjunction with some embodiments of the first aspect, in some embodiments, before the iterative baseline fitting process based on the Raman peak protection mechanism, the method further includes: calculating the gradient change sequence of the smoothed and filtered spectral data in the wavelength dimension; detecting the zero-crossing positions in the gradient change sequence and the numerical sign state between adjacent zero-crossing positions to determine several main morphological change intervals; performing continuity verification on the wavelength width of the main morphological change intervals, marking the intervals that meet the preset width conditions as Raman characteristic peak intervals, and determining the unmarked remaining band range as background baseline intervals.

[0010] By employing the above technical solution, the spectral data processing equipment first calculates the gradient change sequence of the smoothed and filtered spectral data in the wavelength dimension, thereby transforming the microscopic fluctuations of spectral intensity into a quantifiable slope trend. Next, the equipment detects the zero-crossing points of the gradient and the state of adjacent signs, thereby identifying all potential peak morphological boundaries in the spectrum. Furthermore, the equipment verifies the continuity of the wavelength width within the main morphological change intervals, using the inherent half-width at half-maximum (HWHM) of Raman peaks as a screening criterion to filter narrowband noise and broadband background fluctuations. In summary, this solution achieves accurate division of the Raman characteristic peak interval and the background baseline interval without human intervention, providing a reliable spatial constraint for subsequent baseline fitting targeting only the background region and preventing misjudgment of valid signals.

[0011] In conjunction with some embodiments of the first aspect, in some embodiments, the iterative baseline fitting process based on the Raman peak protection mechanism specifically includes: initializing and setting the fitting weight of each data point in the original Raman spectral data, wherein the fitting weight of the data points within the Raman characteristic peak interval is set to zero or much smaller than the fitting weight of the data points within the background baseline interval; performing least-squares fitting calculation on the original Raman spectral data according to the current fitting weight to generate temporary baseline data; calculating the residual values ​​between the original Raman spectral data and each temporary baseline data; updating the fitting weight within the background baseline interval according to the residual values ​​within the background baseline interval, while keeping the fitting weight within the Raman characteristic peak interval from increasing, and performing the next round of iterative fitting calculation until the change amplitude of the temporary baseline data meets the convergence condition to obtain intermediate baseline data.

[0012] By adopting the above technical solution, the spectral data processing equipment first initializes and sets the fitting weights. The equipment then forces the weights within the Raman characteristic peak region to zero or a minimum value, thereby blocking the pulling effect of strong signal regions on baseline generation. During the iteration process, the equipment performs least-squares calculations based on the current fitting weights to generate a temporary baseline and calculates the residuals. Subsequently, the equipment updates the weights of that region based on the residuals of the background baseline region, while strictly maintaining that the weights within the Raman characteristic peak region do not increase. In summary, this solution solves the "overfitting" problem in traditional methods, where Raman peaks are easily mistakenly fitted as background, leading to peak intensity reduction. It effectively removes the background while fully preserving the intensity information and original shape of the characteristic peaks.

[0013] In conjunction with some embodiments of the first aspect, in some embodiments, the non-uniform spline optimization process includes: extracting data points located within the background baseline interval from the intermediate baseline data as interpolation anchor points; generating a smooth curve based on the interpolation anchor points as the final fluorescence background curve; the selection density of the interpolation anchor points is positively correlated with the spectral curvature within the background baseline interval, and no interpolation anchor points are set within the Raman characteristic peak interval.

[0014] By employing the above technical solution, the spectral data processing equipment first extracts intermediate baseline data points only from the background baseline interval as interpolation anchor points, thereby physically isolating the interference of Raman characteristic peaks on curve construction. Then, the equipment adjusts the anchor point density according to the spectral curvature within the background baseline interval, increasing anchor points in high-curvature regions with drastic changes in fluorescence background and decreasing anchor points in flat regions. Finally, the equipment generates a smooth fluorescence background curve based on these anchor points. In summary, this solution specifically addresses the challenge of non-uniform and nonlinear changes in the background baseline in endometrial bleeding samples, eliminates potential localized wavy artifacts after a single fitting, and ensures the baseline flatness of the final purified spectral data.

[0015] In conjunction with some embodiments of the first aspect, in some embodiments, the segmented fusion feature extraction network includes a feature embedding layer, a local analysis layer, and a pyramid-level aggregation layer connected in sequence. The network inputs segmented spectral data into a preset segmented fusion feature extraction network to extract the target feature vector of the endometrial tissue to be detected. Specifically, this includes: mapping each segmented spectral data into high-dimensional independent feature units through the feature embedding layer; extracting local detail information within each independent feature unit using a multi-head self-attention mechanism in the local analysis layer; and performing layer-by-layer dimensionality compression and feature splicing on each independent feature unit containing local detail information through the pyramid-level aggregation layer, gradually fusing the dispersed local features into a target feature vector representing the context information of the entire spectral band.

[0016] By adopting the above technical solution, the spectral data processing device first maps discrete spectral segment data into high-dimensional independent feature units through a feature embedding layer, thereby enhancing the feature representation capability of the data. Next, the device utilizes a multi-head self-attention mechanism in the local analysis layer to extract local detail information within each unit, capturing subtle patterns of chemical bond vibrations within specific spectral bands. Finally, the device performs layer-by-layer dimensional compression and stitching through a pyramid-level aggregation layer. In summary, the target feature vector constructed by this scheme not only contains the fine texture of local spectral bands but also integrates the biochemical distribution patterns across the entire spectrum, improving the feature vector's ability to identify subtle spectral differences between different pathological tissues.

[0017] In conjunction with some embodiments of the first aspect, in some embodiments, the integrated classification model performs feature filtering on the target feature vector through a band filtering layer, specifically including: calculating the mutual information values ​​between each dimension of the target feature vector and the preset classification label, wherein the mutual information values ​​represent the statistical dependence between the features and the category; sorting the mutual information values ​​in descending order, and extracting the feature dimensions whose cumulative information content meets the preset cumulative threshold, thereby generating the dimensionality-reduced key feature vector, which is used as input data for subsequent analysis.

[0018] By employing the above technical solution, the spectral data processing device first calculates the mutual information values ​​between each dimension of the target feature vector and the classification label. This value quantifies the statistical dependence of the feature on the category. Subsequently, the device sorts the mutual information values ​​and extracts feature dimensions whose cumulative information content meets a preset cumulative threshold. Finally, the device generates a dimensionality-reduced key feature vector. In summary, this solution generates a highly condensed key feature vector, significantly reducing the computational complexity of subsequent models while eliminating interference from irrelevant features on the decision boundary, allowing the input data to focus more on essential features highly correlated with pathological changes.

[0019] In conjunction with some embodiments of the first aspect, in some embodiments, the integrated classification model adopts a stacked generalization architecture, which includes at least a base classifier layer and a meta-classifier layer; the filtered features are analyzed and fused through a multi-classifier layer and an attention weighting mechanism, specifically including: using multiple base classifiers constructed differentially in the base classifier layer to predict key feature vectors respectively, and outputting multiple independent probability prediction vectors; inputting each probability prediction vector into the multi-classifier attention weighting layer, and generating a fused feature matrix according to the performance weights of each base classifier; inputting the fused feature matrix into the meta-classifier layer for final decision, and obtaining the classification result label corresponding to the original Raman spectral data.

[0020] By adopting the above technical solution, the spectral data processing device first utilizes multiple differentially constructed base classifiers in the base classifier layer to process key feature vectors in parallel. The device then obtains diversified probability prediction vectors through different algorithmic logics. Subsequently, the device employs a multi-classifier attention weighting mechanism to weight and fuse the prediction results according to the performance weights of each base classifier. This strengthens the decision-making influence of the model that performs better under specific data distributions. Finally, the device outputs the final label through the meta-classifier layer. In summary, this solution utilizes a stacked generalization architecture to compensate for the inductive bias of a single model on specific samples, and improves the robustness and accuracy of the classification result labels when facing complex samples through complementary advantages.

[0021] In a second aspect, this application provides a spectral data processing device, which includes: one or more processors and a memory; the memory is coupled to one or more processors, and the memory is used to store computer program code, which includes computer instructions, and the one or more processors invoke the computer instructions to cause the spectral data processing device to perform the method described in the first aspect and any possible implementation thereof.

[0022] Thirdly, this application provides a computer-readable storage medium storing computer instructions that, when executed on a spectral data processing device, cause the spectral data processing device to perform the method described in the first aspect and any possible implementation thereof.

[0023] Fourthly, this application provides a computer program product, including a computer program / instructions that, when run on a spectral data processing device, cause the spectral data processing device to perform the method described in the first aspect and any possible implementation thereof.

[0024] It is understood that the spectral data processing apparatus provided in the second aspect, the computer-readable storage medium provided in the third aspect, and the computer program product provided in the fourth aspect are all used to execute the methods provided in the embodiments of this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0025] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0026] 1. By employing an adaptive suppression strategy for bleeding interference combined with an adaptive segmentation and segmented fusion feature extraction network, the spectral data processing equipment can accurately remove non-uniformly distributed hemoglobin fluorescence background in stages and deeply mine the biochemical fingerprint features of local and global fusion from the purified data. This effectively solves the problems of background subtraction artifacts caused by a single global smoothing parameter and the neglect of local band correlation in feature extraction in related technologies, thereby achieving high-precision automated assessment of endometrial tissue lesion categories under complex bleeding interference environments.

[0027] 2. By employing a technique that first calculates the gradient change sequence and then verifies the continuity of the wavelength width in the main morphological change interval, the spectral data processing equipment can abandon simple intensity threshold judgment and instead use mathematical morphology and physical bandwidth characteristics to identify effective signals. This effectively solves the problem that related technologies easily misjudge narrowband noise or broadband background fluctuations as characteristic peaks when processing low signal-to-noise ratio spectra. As a result, it achieves accurate definition of the Raman characteristic peak interval and the background baseline interval, providing a pollution-free "protected area" and "fitting area" for subsequent baseline fitting.

[0028] 3. Because iterative baseline fitting processing based on Raman peak protection mechanism is adopted and the fitting weight within the Raman characteristic peak range is strictly kept unchanged during the iteration process, the spectral data processing equipment can physically isolate the "pulling" effect of strong signal peaks on baseline generation in the least squares fitting calculation. This effectively solves the problem that related technologies (such as conventional airPLS) are prone to misidentifying wide Raman peaks as background during iteration, resulting in the reduction of characteristic peak intensity (overfitting). Thus, it achieves complete preservation of the tissue's true Raman scattering intensity information while closely adhering to the bottom of the background. Attached Figure Description

[0029] Figure 1 This is a schematic flowchart of a method for processing spectral data of endometrial tissue based on Raman spectroscopy in an embodiment of this application;

[0030] Figure 2 This is another flowchart illustrating a method for processing spectral data of endometrial tissue based on Raman spectroscopy, as described in this application.

[0031] Figure 3 This is a schematic diagram of the physical device structure of a spectral data processing device in the embodiments of this application. Detailed Implementation

[0032] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0033] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0034] To facilitate understanding, the application scenarios of the embodiments of this application are described below.

[0035] In related technologies, preliminary denoising and feature classification of Raman spectral data of biological tissues can be achieved by using general spectral preprocessing algorithms (such as polynomial fitting and wavelet transform) in conjunction with traditional convolutional neural network models. The following describes a scenario using a Raman spectroscopy-based method for processing spectral data of endometrial tissue. In actual spectral data acquisition, due to limitations in the acquisition environment or sample preparation, non-target interference, such as residual blood adhering to the tissue surface, often gets mixed into the calibration optical path. When processing this high-fluorescence background data using existing conventional spectral analysis systems, the fluorescence interference generated by hemoglobin in the blood has a wide frequency band coverage and intensity much higher than the Raman scattering signal. Related technologies typically employ globally uniform smoothing parameters or baseline correction algorithms of fixed order. This coarse processing method easily introduces "data distortion": if the smoothing is too strong, it will erase weak characteristic peaks that can characterize the biochemical components of the tissue (overfitting); if the denoising is insufficient, the residual broadband fluorescence will be mistaken for a valid signal (underfitting). The feature vectors input to the final model are masked by noise or distorted by the signal, which greatly reduces the accuracy of feature extraction and makes it difficult for the computer model to effectively distinguish the spectral data patterns of different tissue categories.

[0036] The data processing method based on adaptive suppression of bleeding interference and segmented fusion feature extraction in this application embodiment achieves background iterative subtraction under feature peak protection through gradient change sequence verification, and realizes accurate restoration of low signal-to-noise ratio spectral data and robust extraction of high-dimensional features through an adaptive segmented fusion network based on spectral waveform features. The following describes a scenario using a Raman spectroscopy-based endometrial tissue spectral data processing method from this application. Faced with the same spectral data mixed with a strong fluorescence background, the data processing device of this application first performs physical-level signal analysis, using continuity verification to intelligently identify the "effective signal interval" and "background interference interval," accurately stripping non-uniform baseline noise while ensuring the true Raman peak shape is not destroyed. The processed clean data then enters the segmented fusion network. The system not only captures fine-grained features of local bands (such as fingerprint regions corresponding to specific biochemical bonds), but also fuses contextual information across the entire spectrum. Even with limited raw data quality, this processing logic can still keenly capture specific differences hidden in complex backgrounds and output high-confidence category labels.

[0037] As can be seen, by adopting the hemorrhage interference adaptive suppression strategy and segmented fusion feature extraction technology in the embodiments of this application, while achieving high-fidelity restoration of tissue spectral data under complex backgrounds, it can also effectively solve the problem that the classification model has poor generalization ability and low recognition accuracy due to the sensitivity of the feature extraction algorithm to strong background interference. In this way, it realizes the automated and high-precision identification of specific tissue sample categories and improves the application efficiency of spectral analysis technology in the field of intelligent sample data processing.

[0038] To facilitate understanding, the method provided in this implementation will be described in detail below, using the above scenario as an example. Please refer to [link / reference]. Figure 1 This is a schematic flowchart of a Raman spectroscopy-based method for processing spectral data of endometrial tissue.

[0039] S101. Obtain the raw Raman spectral data of the endometrial tissue to be tested. The raw Raman spectral data contains background interference signals.

[0040] Raw Raman spectral data refers to the spectral signal records acquired by spectral data processing equipment through acquisition units such as confocal Raman spectrometers after laser excitation of the endometrial tissue to be tested. This data is presented as a correlation between wavenumber and light intensity and contains the fingerprint information of biomolecules within the tissue. The endometrial tissue to be tested refers to the biological sample used for testing, including but not limited to healthy tissue, benign lesion tissue, or cancerous tissue, and this tissue may be in different bleeding states. Background interference signals refer to non-target tissue characteristic signals present in the raw Raman spectral data, mainly including hemoglobin fluorescence background caused by uterine bleeding, instrument noise, baseline drift, and environmental stray light, etc. These signals can mask the true Raman characteristic peaks.

[0041] Specifically, when initiating the detection process, the spectral data processing equipment first controls the Raman spectroscopy acquisition component via a hardware interface. The equipment sends acquisition commands to the spectrometer, sets preset acquisition parameters (such as integration time and laser power), and scans the endometrial tissue to be tested on the detection stage. Due to the specific nature of clinical samples, the acquired signals are often superimposed with strong background interference from blood components. The spectral data processing equipment receives the analog signals transmitted from the sensor and converts them into digital signals, generating raw Raman spectral data containing full-band information. This data is unprocessed at this stage, and the waveform contains significant noise spikes and a wide background envelope, requiring temporary storage in memory for subsequent purification processing.

[0042] S102. The original Raman spectral data is denoised and background subtracted using a preset adaptive suppression strategy for bleeding interference to obtain purified spectral data. The adaptive suppression strategy for bleeding interference includes sequential smoothing filtering, iterative baseline fitting based on Raman peak protection mechanism, and non-uniform spline optimization.

[0043] The preset adaptive suppression strategy for bleeding interference refers to a set of serial data processing logic configured within the spectral data processing equipment, used to specifically eliminate blood background and noise. This strategy includes three processing stages: smoothing filtering, baseline fitting, and spline optimization. Denoising and background subtraction processing refers to the process of removing random high-frequency fluctuations and low-frequency background trend terms superimposed on the real signal through algorithmic calculations. Purified spectral data refers to Raman spectral data with improved signal-to-noise ratio, flattened baseline correction, and preserved effective characteristic peaks after processing using the above strategies. Smoothing filtering represents the calculation process for suppressing random noise; iterative baseline fitting based on the Raman peak protection mechanism refers to the algorithmic steps of identifying and protecting Raman peak regions from participating in or reducing their weight in baseline generation when fitting the spectral baseline, preventing the real signal from being mistakenly used as background subtraction; non-uniform spline optimization refers to the process of using spline functions to refine and correct complex backgrounds based on the initial fitting. The preset peak identification threshold (implicit parameter) is used to determine which regions belong to Raman peaks that need protection during baseline fitting. This threshold is set according to the statistical characteristics of the spectrum to ensure that peak regions are correctly preserved.

[0044] Specifically, the spectral data processing device reads the raw Raman spectral data acquired in step S101 and calls the hemorrhage interference adaptive suppression strategy module. First, the device performs smoothing filtering, using a preset filtering algorithm to remove high-frequency random noise and cosmic ray interference from the spectrum, smoothing the spectral curve. Next, the device performs iterative baseline fitting based on a Raman peak protection mechanism. The device automatically identifies the positions of characteristic peaks in the spectrum and reduces the weight of these peak regions or truncates them when constructing the baseline model. Through multiple iterative calculations, it gradually approximates the non-Raman scattering background (such as fluorescence background) of the spectrum. Subsequently, to further improve the accuracy of background subtraction, especially for irregular background morphology caused by hemorrhage, the device performs non-uniform spline optimization. Based on the residuals or morphological characteristics of the initial fitting results, it adaptively adjusts the distribution of spline nodes and performs secondary optimization fitting to generate the final background curve. Finally, the device subtracts the generated background curve from the raw spectral data to obtain cleaned spectral data containing only tissue Raman features, providing standardized input for subsequent feature extraction.

[0045] S103. Based on the waveform characteristics of the purification spectral data, adaptive segmentation is performed on the purification spectral data to obtain several spectral segment data.

[0046] Waveform morphology characteristics refer to the geometric shape properties of purified spectral data in different wavenumber ranges, including the number, position, width, density of peaks, and distribution of troughs. Adaptive segmentation refers to the operation of spectral data processing equipment dynamically dividing the entire spectrum into multiple sub-ranges based on the drastic local changes in the spectral curve. The purpose is to group related characteristic peaks within the same range, avoiding fixed segmentation that cuts off effective features. Segmented spectral data represents multiple sub-sequences obtained by cutting the complete purified spectral data according to the defined boundaries. Each segment represents tissue biochemical information within a specific wavenumber range.

[0047] Specifically, after acquiring the purified spectral data, the spectral data processing device activates the segmentation algorithm module. The device first scans the entire spectral curve, calculating its first or second derivative and other indices to quantify waveform morphology, identifying the start and end positions of the main Raman peaks and flat troughs in the spectrum. Based on these morphological features, the device uses an adaptive algorithm to dynamically determine segmentation points. These segmentation points are typically located in troughs or flat regions between two characteristic peak groups, ensuring the integrity of each characteristic unit. According to the determined segmentation points, the device logically divides the purified spectral data into several continuous spectral segments and adds an index label to each segment so that subsequent networks can process these local characteristic intervals in parallel or serially, thereby preserving the local fine structure information of the spectrum.

[0048] S104. Input the spectral segmentation data into the preset segmented fusion feature extraction network to extract the target feature vector of the endometrial tissue to be detected.

[0049] The pre-defined segmented fusion feature extraction network refers to a deep learning model structure deployed in spectral data processing equipment. It is specifically designed to simultaneously extract local spectral features and integrate global information, enabling it to handle variable-length spectral segment inputs. The target feature vector is a highly compressed and abstracted numerical vector output by this network. This vector integrates the local features of each segment as well as the correlation information between segments, and is the core basis for the final classification decision.

[0050] Specifically, the spectral data processing device takes the several spectral segment data obtained in step S103 as input and feeds them into a preset segmented fusion feature extraction network. Within the network, the device first performs feature encoding on each spectral segment data, extracting local texture and peak shape features (such as vibrational modes of specific chemical bonds) within each band. Subsequently, the fusion layer mechanism within the network (such as feature splicing or weighted fusion structure) begins to work, interacting and integrating the scattered local features in the channel or spatial dimensions, establishing long-distance dependencies between different spectral segments, thereby capturing the overall biochemical fingerprint features of endometrial tissue across the entire spectral range. Through forward propagation calculations using this network, the spectral data processing device ultimately outputs a fixed-dimensional target feature vector. This vector eliminates redundant information and highly condenses the key discriminative features that distinguish normal tissue from diseased tissue.

[0051] Optionally, in some embodiments, the training process of the segmented fusion feature extraction network includes:

[0052] The spectral data processing equipment collects Raman spectral samples of endometrial tissue labeled with lesion categories to construct a training dataset. Each sample, after preprocessing and adaptive segmentation, forms several spectral segments and corresponds to a classification label. The spectral data processing equipment divides the dataset into training, validation, and test sets in an 8:1:1 ratio.

[0053] During the network construction phase, the spectral data processing device initializes the parameters of each layer: the embedding dimension of the feature embedding layer is set to 256 dimensions, and the weights are initialized using Xavier; the multi-head self-attention mechanism of the local analysis layer is set with 8 attention heads, each with 32 dimensions, and the Query, Key, and Value projection matrices are initialized; the pyramid-level aggregation layer is designed with multiple levels of convolutional kernels, the first level uses a 1×3 convolutional kernel for dimensionality compression, the second level uses a 1×5 convolutional kernel for feature fusion, and finally, a fixed-dimensional target feature vector is generated through global average pooling.

[0054] In the training process, the spectral data processing device employs end-to-end supervised learning. The device reads spectral segment data in batches (32 samples per batch), inputs it into the network for forward propagation, connects a temporary classification head (fully connected layer plus Softmax) at the end of the network to output predicted probabilities, calculates the cross-entropy loss between the predicted results and the true labels, and then backpropagates to update the parameters. The device uses the Adam optimizer with an initial learning rate of 0.001, which decays by a factor of 0.5 every 30 epochs. To prevent overfitting, a Dropout rate of 0.3 and an L2 regularization coefficient of 0.0001 are set.

[0055] After each training round, the spectral data processing equipment evaluates the model performance on the validation set, calculating accuracy, precision, recall, and F1 score. An early stopping mechanism is employed: training terminates and optimal parameters are saved if the validation set loss does not decrease for 10 consecutive rounds. After training is complete, generalization ability is evaluated on the test set.

[0056] During deployment, the spectral data processing equipment solidifies the network parameters, removes temporary classification heads, and exports the model in ONNX or TensorFlow SavedModel format before loading it into the spectral analysis system. Dropout layers are disabled during inference to ensure deterministic output.

[0057] S105. Input the target feature vector into the preset integrated classification model and output the classification result label corresponding to the original Raman spectral data. The classification result label represents the Raman spectral physicochemical feature category of the endometrial tissue to be detected, which is used to assist medical staff in clinical evaluation.

[0058] The ensemble classification model filters the target feature vector through a band filtering layer, and then analyzes and fuses the filtered features through a multi-classifier layer and an attention weighting mechanism to output the classification result label.

[0059] The pre-defined ensemble classification model refers to the decision model loaded onto the spectral data processing equipment. This model improves the accuracy and robustness of diagnosis by combining the predictive capabilities of multiple classifiers. The band filtering layer is a pre-processing layer in the ensemble classification model, used to perform secondary filtering of input features, removing noisy features that do not contribute to classification. The multi-classifier layer represents multiple base classifiers with different structures (such as machine learning or deep learning classifiers based on different principles) contained in parallel within the model, used to analyze features from different dimensions. The attention weighting mechanism is an adaptive decision fusion strategy used to dynamically allocate the weights of each base classifier in the final decision based on their performance. The Raman spectral physicochemical feature categories represent the biophysical meaning of the raw Raman spectral data, reflecting changes in the molecular composition within tissues.

[0060] Specifically, the spectral data processing device inputs the target feature vector into the integrated classification model. First, the data passes through a band filtering layer. Based on pre-trained feature importance parameters, the device retains highly discriminative feature dimensions, further refining the input features. The processed features are then fed into a multi-classifier layer. The device runs multiple base classifiers concurrently, each inferring the same feature vector based on its unique algorithm logic, outputting preliminary predicted probabilities or classification scores. To obtain more reliable results, the device uses an attention-weighted mechanism to weight and fuse these preliminary predictions, assigning higher decision weights to base classifiers that are more robust in training and validation or more sensitive to the features of the current sample. After weighted calculation, the device outputs the final classification result label, which visually represents the Raman spectral physicochemical characteristics category of the endometrial tissue being tested. The device displays or transmits this result to medical personnel as an important auxiliary reference for clinical assessment.

[0061] The classification result labels output in this embodiment only reflect the similarity of the physicochemical properties of tissue samples in the spectral dimension, and are not the final pathological diagnosis conclusion. Medical personnel need to make a comprehensive judgment by combining clinical symptoms, imaging examinations, and other multi-dimensional information.

[0062] Optionally, in some embodiments, the training process of the ensemble classification model includes:

[0063] The spectral data processing equipment uses a trained segmented fusion feature extraction network to extract features from all samples in the training dataset, generating a set of target feature vectors. These target feature vectors are then used as input data for an ensemble classification model, retaining the true classification labels and dividing the dataset into training, validation, and test sets in an 8:1:1 ratio.

[0064] During the feature selection phase, the spectral data processing equipment performs mutual information analysis on the target feature vectors in the training set. For each dimension, its value distribution across different class samples is statistically analyzed, and the mutual information value with the classification label is calculated. The spectral data processing equipment sorts the mutual information values ​​and tests the classification performance corresponding to different preset cumulative thresholds (80% to 98%, step size 2%) on the validation set using grid search, selecting the threshold that yields the highest F1 score. Based on the determined threshold, high mutual information feature dimensions are extracted to generate a set of key feature vectors.

[0065] During the base classifier layer construction phase, the spectral data processing device employs a differentiated strategy to train multiple base classifiers: the first base classifier uses a Support Vector Machine (SVM), optimizing the radial basis kernel parameters and penalty parameter C (ranging from 0.1 to 100) through grid search; the second base classifier uses a Gradient Boosting Decision Tree (GBDT), fine-tuning the number of trees (100 to 500), maximum depth (3 to 10 layers), and learning rate (0.01 to 0.3); the third base classifier uses a Lightweight Convolutional Neural Network (CNN), designing an architecture of two one-dimensional convolutional layers plus one fully connected layer, trained using the Adam optimizer. After training, the spectral data processing device evaluates the performance of each base classifier on a validation set, recording accuracy, precision, recall, and F1 score.

[0066] During the construction phase of the multi-classifier attention weighting layer, the spectral data processing device calculates performance weights based on the F1 scores of each base classifier on the validation set, and normalizes the weights so that the sum of the weights is 1. The spectral data processing device uses training set samples to generate probability prediction vectors through each base classifier, and then sums them according to the performance weights to form a fusion feature matrix set.

[0067] During the training phase of the meta-classifier layer, the spectral data processing equipment uses the fused feature matrix as input and the ground truth labels as supervisory signals to train the meta-classifier. Logistic regression is used as the meta-classifier, with an L1 regularization coefficient of 0.01, and the L-BFGS optimization algorithm is used to solve for the parameters. The spectral data processing equipment evaluates the overall performance of the complete ensemble classification model on the validation set, iteratively optimizing until performance convergence. After training, accuracy, sensitivity, specificity, and other metrics are calculated on the test set.

[0068] During deployment, the spectral data processing equipment packages and saves the mutual information screening threshold, parameters of each base classifier, performance weight coefficients, and meta-classifier parameters in a unified package, integrating them into the spectral analysis system. In practical applications, it sequentially executes the complete inference process of feature extraction, mutual information screening, parallel prediction by base classifiers, attention-weighted fusion, and meta-classifier decision, outputting classification result labels and confidence information. The spectral data processing equipment periodically collects new clinical samples and updates parameters through incremental learning or model retraining to continuously optimize diagnostic performance.

[0069] In this embodiment, an adaptive suppression strategy for bleeding interference is employed to perform staged smoothing and denoising of the original Raman spectral data and remove non-uniform background. Furthermore, an adaptive segmentation and fusion feature extraction network based on waveform morphology features is used to extract target feature vectors. Finally, an integrated classification model is used for decision-making. Therefore, the spectral data processing device can accurately remove background interference and deeply mine biochemical features that fuse local details and global context from complex signals heavily masked by blood fluorescence. This effectively solves the problems of low signal-to-noise ratio due to the lack of targeted blood background suppression methods in related technologies, and insufficient classification accuracy due to the difficulty of traditional feature extraction methods in simultaneously considering local band correlation and global feature integration. This improves the accuracy and reliability of extracting endometrial tissue lesion features, and enhances the practical value of Raman spectroscopy in gynecological tumor screening.

[0070] Based on the above embodiments, the method provided in this embodiment will be described in further detail below. Please refer to... Figure 2 This is another flowchart illustrating a Raman spectroscopy-based method for processing spectral data of endometrial tissue.

[0071] S201. Obtain the raw Raman spectral data of the endometrial tissue to be tested. The raw Raman spectral data contains background interference signals.

[0072] This step is similar to step S101 in the above embodiment, and will not be repeated here.

[0073] S202. Perform smoothing filtering on the original Raman spectral data to obtain smoothed spectral data.

[0074] Smoothing filtering refers to the data purification operation performed by the spectral data processing equipment before baseline fitting to reduce the impact of random noise on subsequent signal feature extraction. It is the first stage of the adaptive suppression strategy for bleeding interference. The weighted median filtering algorithm is a nonlinear filtering method that preserves signal edge features, used to remove impulse noise (such as cosmic rays) and high-frequency thermal noise from the spectrum. The preset filter window size refers to the width of the local data interval that slides during the filtering process (e.g., set to 5-11 wavenumber points). This parameter needs to be adaptively adjusted according to the overall signal-to-noise ratio of the original Raman spectral data, balancing the denoising effect with spectral resolution.

[0075] Specifically, after the spectral data processing equipment acquires the raw Raman spectral data containing background interference, it immediately initiates a smoothing filtering process to avoid high-frequency noise interfering with subsequent gradient calculations and peak identification. The equipment first assesses the noise intensity of the raw spectrum and determines an appropriate preset filtering window size *w*. Then, it uses an improved weighted median filtering algorithm to iterate through each wavenumber point of the raw spectrum. For the current wavenumber point, the equipment extracts the data sequence within a window centered on that point, assigns weighting coefficients to each data point within the window according to a preset linear decreasing rule, copies the data points to the corresponding weighted number of copies, sorts them, and selects the median value as the filtered intensity value for that wavenumber point. Through this process, the equipment effectively removes cosmic ray interference and randomly fluctuating thermal noise in the form of isolated spikes, while preserving the original height and half-width of the Raman characteristic peaks to the greatest extent possible, resulting in smoothed and filtered spectral data.

[0076] S203. Before the iterative baseline fitting process based on the Raman peak protection mechanism, calculate the gradient change sequence of the spectral data in the wavelength dimension after smoothing and filtering.

[0077] In this context, the wavelength dimension represents the horizontal axis direction of the Raman spectral data (i.e., the Raman shift or wavenumber direction), reflecting the distribution of scattered light at different energies. The gradient variation sequence is used to represent the set of first derivatives or difference values ​​of spectral intensity as a function of wavelength. This sequence intuitively reflects the slope and trend of the spectral curve at each point (such as rising, falling, or flattening), and is a key mathematical feature for identifying the morphology of the spectral waveform.

[0078] Specifically, after the spectral data processing equipment completes the smoothing and filtering process, in order to accurately locate the Raman characteristic peak regions that need to be protected, it is necessary to first analyze the morphological variation patterns of the spectrum. The spectral data processing equipment reads the smoothed and filtered spectral data and calculates the intensity difference between adjacent data points point by point along the wavelength dimension, or uses a differential operator to calculate the first derivative value at each wavenumber point. The spectral data processing equipment arranges the calculated differences or derivative values ​​in wavelength order to construct a complete gradient variation sequence. In this sequence, positive values ​​represent the rising edge of the spectral curve, negative values ​​represent the falling edge, and values ​​close to zero represent peaks, troughs, or flat baseline regions.

[0079] S204. Detect the zero-crossing positions in the gradient change sequence and the numerical sign state between adjacent zero-crossing positions to determine several main morphological change intervals;

[0080] The zero-crossing point refers to the critical wavenumber point in the gradient change sequence where the value changes from positive to negative (corresponding to the peak) or from negative to positive (corresponding to the trough), or the position where the gradient value is directly zero. The main morphological change interval is used to represent the range of candidate wavebands that may contain Raman peaks or background fluctuations, defined by the zero-crossing point and sign state. Each interval usually contains a complete "rising-falling" morphological unit.

[0081] Specifically, after obtaining the gradient change sequence, the spectral data processing equipment scans and analyzes it to segment the spectral structure. The equipment detects all zero-crossing points where the numerical signs in the sequence flip; these points mathematically correspond to the zeros of the first derivative of the original spectrum, i.e., extreme points. Using these zero-crossing points as boundaries, the equipment analyzes the numerical sign state of the sequence between adjacent zero-crossing points, identifying continuous "positive value sequence-zero-crossing point-negative value sequence" structures. This structure corresponds to a convex peak shape in the original spectrum. The equipment defines the range from the starting trough to the ending trough covered by each such structure as a major morphological change interval. The resulting series of intervals contains genuine Raman peaks, but may also contain spurious peaks caused by incomplete baseline fluctuations, requiring further verification.

[0082] S205. Perform continuity verification on the wavelength width of the main morphological change range, mark the range that meets the preset width condition as the Raman characteristic peak range, and determine the remaining unmarked band range as the background baseline range.

[0083] The preset width condition refers to the threshold standard for determining whether a range is a valid Raman peak (e.g., setting a minimum width threshold and a maximum width threshold). This parameter is set based on the statistical law of the full width at half maximum (FWHM) of typical biomolecular characteristic peaks in Raman spectra, and its function is to filter out invalid signals. The Raman characteristic peak range is used to represent the range of bands that have been verified and confirmed to contain real biomolecular fingerprint information. This region will be protected in subsequent baseline fitting. The background baseline range is used to represent the remaining bands that have not been marked as characteristic peaks, usually representing a pure background region, and is the main basis for baseline fitting.

[0084] Specifically, the spectral data processing device reviews each of the major morphological change intervals determined in step S204. The device calculates the wavelength width of each interval (i.e., the end wavenumber minus the start wavenumber) and compares this width with preset width conditions. If the width of an interval is within a preset reasonable range (i.e., neither an extremely narrow noise pulse nor an extremely wide background diffuse), the device classifies it as a valid signal region and tags it with an electronic label indicating a Raman characteristic peak interval. Conversely, if the interval width does not meet the conditions, or if there are flat areas between peak intervals, the device classifies it as a background baseline interval.

[0085] S206. Initialize the fitting weight of each data point in the original Raman spectrum data. The fitting weight of the data points in the Raman characteristic peak interval is set to zero or much smaller than the fitting weight of the data points in the background baseline interval.

[0086] Among them, the fitting weight is used to represent the degree of influence or contribution of each data point to the final fitting result in the least squares fitting calculation. The larger the weight value, the stronger the constraint of the data point in the fitting process. The smaller the weight value or zero, the more the data point basically does not participate or is excluded from the fitting calculation.

[0087] Specifically, after the spectral data processing device completes the division and marking of the Raman characteristic peak interval and the background baseline interval, it needs to start the iterative baseline fitting processing module based on the Raman peak protection mechanism. Before this module officially starts iterative calculation, the spectral data processing device first performs an initialization setting operation to assign an initial fitting weight to each wavenumber point in the original Raman spectral data. The spectral data processing device reads the interval marking information generated in step S205 and assigns differentiated weights according to the interval type to which each data point belongs: for data points located in the background baseline interval, the spectral data processing device initializes their fitting weights uniformly to 1 or other standard positive values, indicating that these points fully participate in baseline fitting; while for data points located in the Raman characteristic peak interval, the spectral data processing device sets their fitting weights to zero or a very small value (such as 0.001) much smaller than the weight of the background region, thereby mathematically achieving soft truncation or weak constraint on these peak regions, ensuring that the subsequently fitted baseline curve will not be pulled upward by the high intensity value of the Raman peak and produce an upward bulge distortion.

[0088] S207. Perform least-squares fitting calculation on the original Raman spectral data according to the current fitting weights to generate temporary baseline data; calculate the residual values ​​between the original Raman spectral data and each temporary baseline data.

[0089] The current fitting weights refer to the set of weight values ​​used in a certain round of the iterative algorithm. The first round uses the initial weights set in step S206, and subsequent rounds use the weights updated in the previous round. Least squares fitting calculation is a mathematical optimization method that solves for the parameters of the fitting curve by minimizing the weighted sum of squared residuals between the observed data and the fitting curve. In the baseline fitting scenario, it is used to generate a smooth background trend line. The temporary baseline data is used to represent the numerical sequence of the fitting curve calculated in a certain round of iteration. This sequence is an estimate of the original spectral background components by the spectral data processing equipment, but it has not yet reached the optimal level and needs to be gradually approximated to the true background through multiple rounds of iteration. The residual value refers to the difference between the measured intensity value of the original Raman spectral data at each wavenumber point and the fitted value of the temporary baseline data at the corresponding wavenumber point. A positive residual indicates that the original spectrum is higher than the baseline (possibly due to Raman peak contribution), and a negative residual indicates that the original spectrum is lower than the baseline (possibly due to fitting overshoot).

[0090] Specifically, after the spectral data processing equipment completes the initialization of the fitting weights or the weight update of a certain iteration, it enters the baseline fitting calculation stage for that round. The spectral data processing equipment calls the weighted penalized least squares algorithm module, reads the current fitting weight vector and the original Raman spectral data, and constructs an objective function (this function includes a data fitting term and a curve smoothing penalty term, with the smoothness controlled by a preset penalty coefficient λ). The spectral data processing equipment solves for the baseline curve parameters that minimize the objective function using numerical optimization methods (such as the conjugate gradient method or direct matrix solving), thereby generating a temporary baseline data under the current weight constraints. Since the weights of data points in the Raman characteristic peak regions are set to extremely low values, the influence of these regions on the fitting results is significantly weakened, and the temporary baseline is mainly determined by the shape of the data points in the background baseline regions. Immediately afterwards, the spectral data processing equipment calculates the difference between the original Raman spectral data and the newly generated temporary baseline data point by point, forming a set of residual numerical sequences.

[0091] S208. Update the fitting weights in the background baseline interval based on the residual values ​​in the background baseline interval, and keep the fitting weights in the Raman characteristic peak interval from increasing. Perform the next round of iterative fitting calculation until the change magnitude of the temporary baseline data meets the convergence condition, and obtain the intermediate baseline data.

[0092] Weight update refers to the dynamic adjustment process of adjusting the fitting weights of each data point in the next iteration based on the fitting effect evaluation results of the current round. It is the core mechanism of adaptive optimization of the iterative algorithm. Convergence condition refers to the preset criteria for determining whether an iteration should terminate (e.g., the maximum difference between the temporary baseline data generated by two adjacent iterations is less than a preset convergence threshold, or the residual change rate is lower than a preset threshold). This condition is used to prevent excessive iteration from wasting computational resources while ensuring that the fitting result reaches the required accuracy. The preset convergence threshold is usually set based on the statistical characteristics of the average spectral intensity, for example, set to 0.1% to 1% of the average spectral intensity. Its function is to quantify the mathematical criterion that "the baseline has stabilized." Intermediate baseline data is used to represent the final baseline fitting result output after iteration convergence. It is called "intermediate" because this baseline data will be used as the input for the next stage of non-uniform spline optimization processing, rather than the final fluorescence background curve.

[0093] Specifically, after the spectral data processing device obtains the temporary baseline data and residual values ​​for the current round in step S207, it immediately initiates the weight adaptive update mechanism. The spectral data processing device first analyzes the residual values ​​of each data point within the background baseline interval: for data points with negative residuals (indicating that the original spectrum is lower than the current fitted baseline, meaning the fitted baseline is too high), the spectral data processing device maintains or slightly increases the weight of that point to enhance its constraint on the next round of fitting, forcing the baseline to move lower towards that point; for data points with positive residuals (indicating that the original spectrum is higher than the current baseline, possibly due to residual peaks or noise), the spectral data processing device reduces the weight of that point to weaken its impact on fitting and prevent the baseline from being raised by non-background components. For data points within the Raman characteristic peak interval, regardless of their residual values, the spectral data processing device strictly maintains that their fitting weights do not increase, ensuring that the peak region does not participate in baseline construction in any iteration round. After completing the weight update, the spectral data processing device uses the updated weights as the new current fitting weights and executes the fitting and residual calculation in step S207 again, thus forming an iterative loop.

[0094] At the end of each iteration, the spectral data processing device checks the convergence condition: it calculates the absolute value of the difference between the temporary baseline data generated in this round and the temporary baseline data in the previous round at all wavenumber points. If the maximum difference is less than the preset convergence threshold, or the rate of change of the sum of squared residuals for several consecutive rounds is lower than the set value, then the iteration is considered to have converged. At this time, the spectral data processing device outputs the temporary baseline data of the last round as intermediate baseline data. This data has approximated the background trend of the original spectrum quite well. However, due to the limitation of the global smoothing parameter λ, there may still be fitting bias in regions where the background curvature changes rapidly (such as non-uniform fluorescence background caused by hemorrhage interference). Therefore, it is necessary to proceed to the next stage of spline optimization processing.

[0095] S209. Extract data points located within the background baseline interval from the intermediate baseline data as interpolation anchor points;

[0096] Interpolation anchor points are used to represent data coordinate pairs (wavenumber position and corresponding baseline intensity value) that serve as control points or nodes in the spline interpolation algorithm. These anchor points determine the basic direction and local shape of the spline curve and form the skeleton for constructing a smooth curve.

[0097] Specifically, after the spectral data processing equipment obtains intermediate baseline data through iterative fitting, in order to further improve the baseline's ability to fit non-uniform backgrounds (especially rapidly changing fluorescence backgrounds caused by hemorrhage), a non-uniform spline optimization processing stage needs to be initiated. The primary task in this stage is to determine the set of anchor points required for spline interpolation. The spectral data processing equipment reads the intermediate baseline data and the interval marking information generated in step S205, extracting data points only from the wavenumber range marked as background baseline intervals. The spectral data processing equipment traverses the background baseline intervals, extracting the wavenumber coordinates and corresponding fitting intensity values ​​of the intermediate baseline data within these intervals in pairs to construct the interpolation anchor point set. Since the Raman characteristic peak intervals have been explicitly excluded, data points in these areas will not be selected as anchor points, ensuring that the subsequently generated spline curves are constructed only based on the morphological characteristics of the background region and will not be affected by Raman peaks, resulting in local bulges.

[0098] S210. Generate a smooth curve based on the interpolation anchor points as the final fluorescence background curve; the selection density of the interpolation anchor points is positively correlated with the spectral curvature within the background baseline interval, and no interpolation anchor points are set within the Raman characteristic peak interval.

[0099] The smooth curve refers to a continuous and well-smooth fitting curve constructed between discrete anchor points using mathematical interpolation methods (such as B-spline functions) (usually requiring continuity of the first or second derivative). The final fluorescence background curve represents the curve output by the spectral data processing equipment after complete hemorrhage interference adaptive suppression, which best approximates the non-Raman scattering background components (mainly hemoglobin fluorescence) in the original Raman spectrum. This curve is subtracted from the original spectrum to obtain purified spectral data. Spectral curvature represents the degree of bending of the spectral curve at a certain wavenumber position (mathematically, the absolute value of the second derivative). Regions with high curvature correspond to locations where the background changes rapidly (such as steep rises or falls in the fluorescence background), while regions with low curvature correspond to locations where the background changes gently.

[0100] Specifically, after the spectral data processing equipment completes the extraction of interpolation anchor points, it enters the spline curve generation stage. The equipment first analyzes the curvature distribution characteristics of the intermediate baseline data corresponding to the set of interpolation anchor points, calculating the local curvature within each background baseline interval (estimated through second-order difference or derivative). Based on the curvature analysis results, the equipment adaptively adjusts the selection density of interpolation anchor points: in regions with high curvature (such as abrupt changes in fluorescence background caused by hemorrhage), the equipment increases the anchor point sampling density, possibly selecting an anchor point every few wavenumber points; while in flat regions with low curvature, the equipment decreases the sampling density, possibly selecting an anchor point only every tens of wavenumber points. Simultaneously, the equipment strictly adheres to the principle of not setting interpolation anchor points within Raman characteristic peak intervals, ensuring that the spline curve crosses these regions naturally and smoothly without being stretched by peaks. Based on the final set of anchor points after screening and density adjustment, the spectral data processing equipment calls a non-uniform B-spline interpolation algorithm (typically using a cubic B-spline function to ensure the continuity of the curve's second derivative). Using the wavenumber coordinates of the anchor points as node vectors and the intensity values ​​of the anchor points as control parameters, it calculates and generates a smooth curve that runs through all background regions. This curve, while maintaining overall smoothness, can flexibly adapt to differences in background curvature in different regions, accurately approximating the complex non-uniform fluorescence background morphology caused by hemorrhage. The spectral data processing equipment defines this generated smooth curve as the final fluorescence background curve and uses its output as the basis for background subtraction.

[0101] Subsequently, the spectral data processing equipment subtracts the corresponding value of the fluorescence background curve from the original Raman spectral data point by point to obtain purified spectral data containing only Raman characteristic signals, thus completing the preprocessing process of the entire adaptive suppression strategy for hemorrhage interference.

[0102] S211. Based on the waveform characteristics of the purification spectral data, adaptive segmentation is performed on the purification spectral data to obtain several spectral segment data.

[0103] This step is similar to step S103 in the above embodiment, and will not be repeated here.

[0104] S212. Input the spectral segmentation data into the preset segmented fusion feature extraction network to extract the target feature vector of the endometrial tissue to be detected; the segmented fusion feature extraction network includes a feature embedding layer, a local analysis layer and a pyramid-level aggregation layer connected in sequence.

[0105] This step specifically includes:

[0106] Each spectral segment data is mapped into a high-dimensional independent feature unit through a feature embedding layer;

[0107] The multi-head self-attention mechanism in the local analysis layer is used to extract local detail information within each independent feature unit;

[0108] By using a pyramid-level aggregation layer, the independent feature units containing local details are compressed and spliced ​​layer by layer, gradually merging the scattered local features into a target feature vector that represents the context information of the entire band.

[0109] The feature embedding layer is the first layer of the segmented fusion feature extraction network. It transforms the input spectral segment data from the original one-dimensional intensity sequence into a representation in a high-dimensional vector space, laying the mathematical foundation for subsequent deep feature extraction. High-dimensional independent feature units represent feature vectors obtained after the embedding layer mapping, with dimensions much higher than the original spectral data. Each spectral segment corresponds to an independent feature unit. The local analysis layer is a processing layer specifically responsible for extracting detailed features within individual feature units. It captures local texture information such as peak position relationships and intensity distribution patterns within spectral segments through deep learning mechanisms. The multi-head self-attention mechanism is a neural network structure that computes multiple sets of attention weights in parallel, enabling simultaneous attention to the dependencies between elements within feature units from different representation subspaces. Local detail information represents fine features within spectral segments, such as the relative intensity ratio of adjacent peaks within a specific wavenumber range, peak width variation trends, and weak shoulder peaks—local spectral patterns that are significant for lesion diagnosis. The pyramid-level aggregation layer refers to a network structure that employs a multi-scale, layer-by-layer fusion strategy, similar to a feature pyramid in image processing. Through hierarchical feature compression and stitching operations, it progressively integrates local features from different spectral segments into global features. Full-band contextual information is used to represent a global characterization that integrates features from all segments of the entire Raman spectrum, from low to high wavenumbers. This information includes not only the local features of each segment but also macroscopic patterns such as the relative positional relationships between segments and intensity gradient changes.

[0110] Specifically, after the spectral data processing device completes the adaptive segmentation of the purified spectral data and obtains several spectral segment data, these segment data need to be input into a preset segmented fusion feature extraction network for deep feature mining. In the first stage of network computation, the spectral data processing device starts the feature embedding layer module and reads the original intensity sequence of each spectral segment data (for example, a segment contains intensity values ​​of 50 wavenumber points). The spectral data processing device performs mapping calculations on each segment data through the linear transformation matrix (the weight parameters have been optimized and determined during the model training stage) or convolution kernel inside the embedding layer, projecting it from the low-dimensional original space to the high-dimensional feature space, generating corresponding high-dimensional independent feature units (for example, mapping a 50-dimensional intensity sequence to a 256-dimensional feature vector). After completing the embedding mapping, the spectral data processing device inputs each high-dimensional independent feature unit into the local analysis layer in sequence. In the local analysis layer, the spectral data processing device calls the multi-head self-attention mechanism module to calculate multiple sets of attention weight matrices in parallel for each dimension of feature elements within each independent feature unit. Specifically, the spectral data processing device first linearly transforms the feature units into three sets of representations: query, key, and value. Then, within each attention head, it calculates the similarity score between the query and key, normalizes it, and uses this score as attention weights. These weights are then used to weighted summations of the values, thereby achieving adaptive information interaction between feature elements. Through multi-head parallel computation, the spectral data processing device can simultaneously capture local details within feature units from multiple subspace perspectives. For example, the first attention head might focus on the spacing between peaks, the second on the relative intensity ratio, and the third on the symmetry of peak shapes. After processing by the local analysis layer, each independent feature unit is infused with rich internal structural information, but the connections between segments are not yet established.

[0111] Next, the spectral data processing device initiates a pyramid-level aggregation layer, fusing independent feature units containing local details across segments. The device employs a hierarchical processing strategy: at the first level, two or more adjacent feature units are combined into a larger feature block through feature concatenation, followed by dimensionality compression via convolutional or fully connected layers to generate intermediate feature representations; at the second level, the device continues to concatenate and compress the intermediate features from the previous level, further expanding the receptive field of the features; this process is repeated until the top layer completely fuses all segment features into a single global feature vector. During dimensionality compression at each level, the device enhances the expressive power of features using nonlinear activation functions (such as ReLU or GELU) and preserves detailed information from lower-level features through residual connections or skip connections, preventing excessive information loss during multi-layer transmission. After layer-by-layer processing by the pyramid-level aggregation layer, the local features originally scattered across various spectral segments are organically integrated into a fixed-dimensional target feature vector (e.g., 512-dimensional or 1024-dimensional).

[0112] S213. Input the target feature vector into the preset integrated classification model and output the classification result label corresponding to the original Raman spectral data. The classification result label represents the Raman spectral physicochemical feature category of the endometrial tissue to be detected, which is used to assist medical staff in clinical assessment.

[0113] The ensemble classification model employs a stacked generalization architecture, containing at least a base classifier layer and a meta-classifier layer.

[0114] This step specifically includes:

[0115] Calculate the mutual information values ​​between each dimension of the target feature vector and the preset classification labels. The mutual information values ​​represent the statistical dependence between the features and the categories.

[0116] The mutual information values ​​are sorted in descending order, and the feature dimensions whose cumulative information content meets the preset cumulative threshold are extracted to generate the dimensionality-reduced key feature vectors. The key feature vectors are used as input data for subsequent analysis.

[0117] Multiple base classifiers, constructed using differentiation in the base classifier layer, are used to predict key feature vectors, outputting multiple independent probability prediction vectors.

[0118] Each probability prediction vector is input into the attention weighting layer of the multi-classifier, and a fusion feature matrix is ​​generated based on the performance weights of each base classifier.

[0119] The fused feature matrix is ​​input into the meta-classifier layer for final decision-making, resulting in the classification label corresponding to the original Raman spectral data.

[0120] Stacked generalization architecture refers to an ensemble learning strategy that improves classification performance and robustness by using the predictions of multiple base classifiers as new features to input into a higher-level meta-classifier for secondary learning. The base classifier layer is the first layer composed of multiple classifiers using different algorithms in parallel, such as Support Vector Machines (SVM), Gradient Boosting Decision Trees (GBDT), and Lightweight Convolutional Neural Networks (CNN). The meta-classifier layer is the second layer used to integrate the predictions of the base classifiers and output the final classification conclusion; it typically uses logistic regression or a neural network. Mutual information is a quantitative indicator representing the strength of the statistical dependency between feature dimensions and classification labels; a higher value indicates a more significant contribution of the feature to classification. The preset cumulative threshold is a standard for the proportion of cumulative information (e.g., 85%, 90%, or 95%) set in feature selection to determine how many high-mutual-information features to retain. This threshold is determined through cross-validation experiments and serves to balance feature dimensionality and classification performance. The key feature vector represents the feature vector obtained after dimensionality reduction through mutual information filtering, retaining only the feature dimensions that contribute the most to classification. The probability prediction vector represents the predicted probability distribution of each category output by the base classifiers. The multi-classifier attention weighting layer is a fusion mechanism that dynamically calculates weights based on the historical performance of each base classifier on the validation set. Performance weights represent weight coefficients that quantify the reliability of each base classifier, typically calculated based on the validation set accuracy. The fusion feature matrix represents the comprehensive feature representation formed by weighting and fusing the probability prediction vectors of each base classifier.

[0121] Specifically, after the spectral data processing device obtains the target feature vector from the segmented fusion feature extraction network, it needs to be input into the integrated classification model to determine the lesion category.

[0122] In the first stage, the spectral data processing equipment activates a band filtering layer to calculate the mutual information value between each dimension of the target feature vector and the preset classification label, quantifying the statistical dependence of that dimension on distinguishing different lesion categories. The spectral data processing equipment sorts the mutual information values ​​of all dimensions from largest to smallest and accumulates them starting from the top. When the accumulated information content reaches a preset accumulation threshold (e.g., 90%), the feature dimension before the corresponding position is truncated to generate a dimensionality-reduced key feature vector, effectively removing redundant features and noise interference.

[0123] In the second stage, the spectral data processing device inputs the key feature vectors in parallel into multiple differentially constructed base classifiers in the base classifier layer. Each base classifier independently performs classification prediction based on its trained model parameters and outputs the corresponding probability prediction vector (e.g., [0.15, 0.25, 0.60] in a three-class classification task).

[0124] In the third stage, the spectral data processing device inputs each probability prediction vector into the attention weighting layer of the multi-classifier, reads the performance weights of each base classifier that have been statistically analyzed on the validation set, and performs a weighted summation on each probability prediction vector to generate a fusion feature matrix.

[0125] In the final stage, the spectral data processing equipment inputs the fused feature matrix into the meta-classifier layer. The meta-classifier uses the weight parameters learned during the training phase to calculate the final classification probability distribution and selects the category with the highest probability value as the classification result label. This classification result label represents the Raman spectral physicochemical characteristics category of the endometrial tissue to be tested. The spectral data processing equipment transmits this label, along with confidence information, to clinical medical staff as a reference for auxiliary diagnosis.

[0126] In this embodiment, the Raman characteristic peak interval and background baseline interval are accurately divided by first calculating the gradient change sequence and wavelength width for continuity verification, and then iterative baseline fitting processing based on the Raman peak protection mechanism is adopted. Therefore, the spectral data processing device can construct reliable fitting constraints based on the physical slope and bandwidth characteristics of the spectral waveform, and physically shield the weight influence of the strong signal region when iteratively approximating the background. This effectively solves the problems of difficulty in identifying characteristic peaks due to low signal-to-noise ratio and non-uniform background when processing hemorrhage samples, and the problem of signal distortion caused by overfitting of characteristic peaks by traditional baseline fitting algorithms. Thus, it can ensure the integrity and authenticity of Raman spectral fingerprint information while removing various forms of hemoglobin fluorescence interference.

[0127] The spectral data processing device in the embodiments of this invention is described below from the perspective of hardware processing. Please refer to [link / reference needed]. Figure 3 This is a schematic diagram of the physical device structure of a spectral data processing device in the embodiments of this application.

[0128] It should be noted that, Figure 3 The structure of the spectral data processing device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of the present invention.

[0129] like Figure 3 As shown, the spectral data processing device includes a CPU 301, which can perform various appropriate actions and processes according to a program stored in the read-only memory ROM 302 or a program loaded from the storage section 308 into the random access memory RAM 303, such as performing the methods described in the above embodiments. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An I / O interface 305 is also connected to the bus 304.

[0130] The following components are connected to I / O interface 305: input section 306 including audio input devices, push-button switches, etc.; output section 307 including a liquid crystal display (LCD) and audio output devices, indicator lights, etc.; storage section 308 including a hard disk, etc.; and communication section 309 including a network interface card such as a LAN (Local Area Network) card, modem, etc. Communication section 309 performs communication processing via a network such as the Internet. Drive 310 is also connected to I / O interface 305 as needed. Removable media 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 310 as needed so that computer programs read from them can be installed into storage section 308 as needed.

[0131] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program / instructions carried on a computer-readable medium, the computer program / instructions containing computer program / instructions for performing the methods shown in the flowcharts. In such embodiments, the computer program / instructions can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by CPU 301, it performs the various functions defined in the present invention.

[0132] It should be noted that specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0133] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, program segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.

[0134] Specifically, the spectral data processing device in this embodiment includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it implements the Raman spectral data processing method for endometrial tissue based on Raman spectroscopy provided in the above embodiment.

[0135] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the spectral data processing device described in the above embodiments; or it may exist independently and not assembled into the spectral data processing device. The storage medium carries one or more computer programs that, when executed by a processor of the spectral data processing device, cause the spectral data processing device to implement the Raman spectroscopy-based endometrial tissue spectral data processing method provided in the above embodiments.

[0136] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0137] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as meaning "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0138] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for processing spectral data of endometrial tissue based on Raman spectroscopy, applied to a spectral data processing device, characterized in that, The method includes: Acquire raw Raman spectral data of the endometrial tissue to be tested, wherein the raw Raman spectral data contains background interference signals; The original Raman spectral data is denoised and background subtracted using a preset adaptive suppression strategy for bleeding interference to obtain purified spectral data. The adaptive suppression strategy for bleeding interference includes sequential smoothing filtering, iterative baseline fitting based on Raman peak protection mechanism, and non-uniform spline optimization. Based on the waveform characteristics of the purification spectral data, the purification spectral data is adaptively segmented to obtain several spectral segment data. The spectral segmentation data is input into a preset segmented fusion feature extraction network to extract the target feature vector of the endometrial tissue to be detected; The target feature vector is input into a preset integrated classification model, and the classification result label corresponding to the original Raman spectral data is output. The classification result label represents the Raman spectral physicochemical feature category of the endometrial tissue to be detected, which is used to assist medical staff in clinical assessment. The integrated classification model filters the target feature vector through a band filtering layer, and then analyzes and fuses the filtered features through a multi-classifier layer and an attention weighting mechanism to output the classification result label.

2. The method according to claim 1, characterized in that, Prior to the iterative baseline fitting process based on the Raman peak protection mechanism, the method further includes: Calculate the gradient change sequence of the smoothed spectral data in the wavelength dimension; Detect the zero-crossing positions in the gradient change sequence and the numerical sign state between adjacent zero-crossing positions to determine several main morphological change intervals; The wavelength width of the main morphological change range is continuously verified, and the range that meets the preset width condition is marked as the Raman characteristic peak range, and the remaining unmarked range is determined as the background baseline range.

3. The method according to claim 2, characterized in that, The iterative baseline fitting process based on the Raman peak protection mechanism specifically includes: Initialize the fitting weight of each data point in the original Raman spectrum data, and set the fitting weight of the data points in the Raman characteristic peak interval to zero or much smaller than the fitting weight of the data points in the background baseline interval. Based on the current fitting weights, the original Raman spectral data are subjected to least squares fitting calculation to generate temporary baseline data; Calculate the residual values ​​between the original Raman spectral data and each of the provisional baseline data; The fitting weights within the background baseline interval are updated based on the residual values ​​within the background baseline interval, while keeping the fitting weights within the Raman characteristic peak interval unchanged. The next round of iterative fitting calculation is then performed until the change magnitude of the temporary baseline data meets the convergence condition, thus obtaining the intermediate baseline data.

4. The method according to claim 3, characterized in that, The non-uniform spline optimization process includes: Data points located within the background baseline interval are extracted from the intermediate baseline data and used as interpolation anchor points; A smooth curve is generated based on the interpolation anchor points, which serves as the final fluorescence background curve. The selection density of the interpolation anchor points is positively correlated with the spectral curvature within the background baseline interval, and no interpolation anchor points are set within the Raman characteristic peak interval.

5. The method according to claim 1, characterized in that, The segmented fusion feature extraction network comprises a feature embedding layer, a local analysis layer, and a pyramid-level aggregation layer connected in sequence. The step of inputting the spectral segmentation data into a preset segmented fusion feature extraction network to extract the target feature vector of the endometrial tissue to be detected specifically includes: The feature embedding layer maps each spectral segment data into a high-dimensional independent feature unit; The multi-head self-attention mechanism in the local analysis layer is used to extract local detail information within each independent feature unit; The pyramid-level aggregation layer performs dimensional compression and feature splicing on each independent feature unit containing local details, gradually merging the scattered local features into the target feature vector that represents the context information of the entire band.

6. The method according to claim 1, characterized in that, The ensemble classification model performs feature filtering on the target feature vector through a band filtering layer, specifically including: Calculate the mutual information value between each dimension of the target feature vector and the preset classification label, wherein the mutual information value represents the statistical dependence between the feature and the category; The mutual information values ​​are sorted in descending order, and the feature dimensions whose cumulative information content meets the preset cumulative threshold are extracted to generate the dimensionality-reduced key feature vectors. The key feature vectors are used as input data for subsequent analysis.

7. The method according to claim 6, characterized in that, The ensemble classification model adopts a stacked generalization architecture, which includes at least a base classifier layer and a meta classifier layer. The process of analyzing and fusing the filtered features through a multi-classifier layer and an attention weighting mechanism specifically includes: Multiple base classifiers, constructed differentially in the base classifier layer, are used to predict the key feature vectors respectively, outputting multiple independent probability prediction vectors; Each of the probability prediction vectors is input into the attention weighting layer of the multi-classifier, and a fusion feature matrix is ​​generated based on the performance weights of each base classifier. The fused feature matrix is ​​input into the meta-classifier layer for final decision-making to obtain the classification result label corresponding to the original Raman spectral data.

8. A spectral data processing device, characterized in that, The spectral data processing device includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the spectral data processing device to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium storing computer instructions, characterized in that, When the computer instructions are executed on the spectral data processing device, the spectral data processing device causes the spectral data processing device to perform the method as described in any one of claims 1-7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are run on the spectral data processing device, the spectral data processing device performs the method as described in any one of claims 1-7.

Images