A breast tissue detection system and method for cross-scale spectral analysis

Abstract

The application discloses a breast tissue detection system and method for cross-scale spectral analysis, and relates to the technical field of spectral analysis. Firstly, microscopic hyperspectral images of ex vivo breast tissue sections are collected, and are stored in association with pathological annotation information, patient clinical information and tissue state labels to form a standardized benchmark sample library. Then, near-infrared diffuse reflection spectrum signals of in vivo breast tissue are collected, and in vivo mixed spectrum vectors are generated through normalization correction. A cross-scale spectral analysis model is constructed and trained based on the standardized benchmark sample library, the contribution weights of different breast tissue components are separated from the in vivo mixed spectrum vectors, and the probabilities of different breast tissue states are output, so that an evaluation report containing the proportion of tissue components and the contribution of characteristic wave bands is generated. When an abnormal outlier signal is identified in the detection process, the current sample is determined to be an abnormal spectrum sample, and microscopic hyperspectral scanning is performed again to obtain new data, and the standardized benchmark sample library is updated, so that the system closed-loop iterative optimization is realized.

Description

Technical Field

[0001] This invention relates to the field of spectral analysis technology, specifically a breast tissue detection system and method based on cross-scale spectral analysis. Background Technology

[0002] Currently, breast tissue testing mainly relies on two methods: pathological section analysis and in vivo optical detection. Pathological examination can obtain microscopic tissue characteristics as a diagnostic reference, but it is an ex vivo test and difficult to use for real-time in vivo analysis. In vivo near-infrared optical detection has the advantages of being non-invasive and rapid, and can obtain deep tissue optical response signals, making it suitable for clinical screening. However, in existing detection methods, there is no unified correspondence between ex vivo pathological information and in vivo detection signals. The two differ significantly in sampling scale and detection mechanism, and in vivo spectra are difficult to directly map to microscopic tissue characteristics. Breast tissue has a strong light scattering effect, and the near-infrared echo signal is composed of multiple components such as fat, water, blood, and glands, resulting in significant signal mixing, which easily masks subtle features and makes it difficult to accurately distinguish tissue states. Traditional analytical methods can only process data at a single scale and cannot combine microscopic pathological features with macroscopic in vivo signals for comprehensive judgment. Spectral data has a large amount of information and high dimensionality, lacks efficient screening and purification methods, is easily interfered with, and results are unstable. Most detection systems can only complete a single detection and result output, and have not established new sample return and model optimization mechanisms, so the system cannot continuously improve the accuracy of analysis in long-term use. Summary of the Invention

[0003] The purpose of this invention is to provide a breast tissue detection system and method for cross-scale spectral analysis to solve the problems raised in the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a breast tissue detection system for cross-scale spectral analysis, comprising a benchmark library construction module, a signal acquisition module, a feature analysis module, a report management module, and a closed-loop update module; The benchmark library construction module acquires hyperspectral images of isolated breast tissue sections using microscopic hyperspectral imaging to obtain hyperspectral image data containing multiple continuous spectral channels; it then associates and stores the hyperspectral image data with pathological annotation information of the breast tissue region, patient clinical information, and breast tissue status labels to form a standardized benchmark sample library. The signal acquisition module irradiates breast tissue with a near-infrared light source, acquires the diffuse reflectance spectral signal returned after being scattered by the breast tissue, and normalizes and corrects the diffuse reflectance spectral signal to generate an in vivo mixed spectral vector. The feature analysis module constructs and trains a cross-scale spectral analysis model based on a standardized benchmark sample library. The cross-scale spectral analysis model is used to analyze the in vivo mixed spectral vector, calculate the matching degree between the in vivo mixed spectrum and the reference spectrum, separate the contribution weights of different breast tissue components from the in vivo mixed spectrum, and output the probability that the target detection area belongs to different breast tissue states. The report management module is used to store in vivo mixed spectral vectors, contribution weights of different breast tissue components, and probabilities of different breast tissue states, and to generate an assessment report containing information on the proportion of breast tissue components and contribution information of characteristic bands. When the closed-loop update module identifies an abnormal outlier signal during the detection process, it marks the corresponding spectral sample as an abnormal spectral sample; it then performs a new microscopic hyperspectral scan on the ex vivo breast tissue corresponding to the abnormal spectral sample to obtain new hyperspectral image features and corresponding breast tissue state annotation information, and writes the new data into the standardized benchmark sample library.

[0005] Furthermore, the benchmark library construction module includes a hyperspectral imaging unit, a labeling and association unit, and a structured storage unit; The hyperspectral imaging unit scans ex vivo breast tissue sections to acquire 256 continuous channels of hyperspectral image cube X-rays in the 400-1000 nm band. path ∈R H×W×256 Where H and W represent the spatial height and width of the pathological slide image, respectively; 256 represents the number of continuous spectral channels acquired in the 400-1000 nm band. The annotation association unit binds pathological annotation information, patient clinical information, and breast tissue status tags; Structured storage units perform registration, alignment, and encoding on hyperspectral image data, pathological annotation information, patient clinical information, and breast tissue status labels to form a standardized benchmark sample library. ;in, M represents the cube containing the pathological hyperspectral image of the nth ex vivo breast tissue sample. (n) C represents the annotation information of the lesion region corresponding to the nth isolated breast tissue sample. (n) y represents the clinicopathological attribute information of the patient associated with the nth ex vivo breast tissue sample. (n) This indicates the tissue status label of the nth isolated breast tissue sample, determined based on clinical test results.

[0006] Furthermore, the signal acquisition module includes a signal acquisition unit, a signal processing unit, and a signal quality assessment unit; The signal acquisition unit irradiates breast tissue in vivo in the 700nm to 1700nm wavelength range using a near-infrared light source and detector, and acquires the diffuse reflection signal of the breast tissue in vivo to obtain the original light intensity signal of the breast tissue under test. The signal processing unit performs normalization calculation on the diffuse reflection signal based on the light intensity of the tissue under test, the dark current signal, and the standard reference light intensity to obtain the normalized reflectance signal. The signal quality assessment unit calculates the signal-to-noise ratio (SNR) of the normalized signal, determines the signal that meets the preset SNR threshold as the effective signal, and arranges the effective signals according to the wavelength dimension to form a one-dimensional bulk mixed spectral vector.

[0007] Furthermore, the feature analysis module includes a domain-aligned attention unit, a spectral demixing unit, a classification output unit, and a pre-training unit; The domain-aligned attention unit performs local feature extraction and embedding transformation on the in vivo mixed spectral vector to obtain a spectral feature tensor; it extracts reference spectral prototypes of various breast tissue components from hyperspectral image data in a standardized benchmark sample library to form a reference spectral prototype set R; then, domain-aligned attention calculation is performed. ; Wherein, DAAttn() represents the breast spectral features output after domain-aligned attention-weighted aggregation of the in vivo mixed spectrum and the reference spectral prototype; Q is the query matrix, representing the in vivo mixed spectral features of the currently input in vivo breast tissue to be tested; K is the key matrix, representing the spectral features of various components of breast tissue extracted from the standardized benchmark sample library; V is the value matrix, representing the original spectral information of breast tissue; d k The feature dimension of the key vector is used to scale the dot product matching results of spectral features; G is the standard dot product attention term, representing the similarity between different spectral channels; G is the reference spectral guidance matrix, representing the degree of cross-scale matching between the spectral features of the in vivo breast tissue to be tested and the reference spectral prototype of breast tissue in the in vitro benchmark library; B is the spectral channel contribution bias matrix, used to highlight the feature bands that have a higher discriminative contribution to the differences in breast tissue composition and tissue state; β and γ are learnable weight coefficients, used to adjust the strength of the role of the reference spectral guidance term and the channel contribution bias term in the overall attention weight allocation process of breast spectral data; softmax() represents the normalization function, used to convert the association matching scores of each spectral channel into corresponding weights.

[0008] The reference spectral guiding matrix G is calculated as follows: ; Where Norm() represents the normalization operation; W q r With W rAs a learnable projection matrix, the in vivo mixed spectral features and the reference spectral prototype are mapped to the same alignment space, respectively; RW r This represents the mapping result of the reference spectral prototype in the alignment space; Construct the channel contribution gating vector g: ; Where F represents the input 256-channel spectral feature tensor; Pool() represents the global pooling operation, which compresses the high-dimensional spectral features into a global statistical description; z represents the global feature vector after pooling; W g Represents the gating mapping matrix; b g The bias term is represented by σ; σ() represents the Sigmoid activation function; for the j-th spectral channel, its contribution weight is denoted as g. j ∈(0,1), representing the relative contribution of the j-th band to the determination of breast tissue status; the contribution weights of all spectral channels constitute the gating vector g=[g1,g2,…,g 256 Subsequently, based on the channel contribution gating vector g, the dimensionality is expanded to construct the spectral channel contribution bias matrix B, which participates in the domain alignment attention calculation of the domain alignment attention unit.

[0009] The spectral demixing unit separates the tissue components of the breast spectral features output by the domain-aligned attention unit, using the formula: ; Among them, S in (λ) represents the response intensity of the mixed spectrum acquired in vivo at wavelength λ; M represents the number of categories of reference breast tissue components; w i w represents the contribution weight of the i-th type of breast tissue component in the current mixed spectrum. i ≥0 and ∑ i=1 M w i =1;S ref,i (λ) represents the reference spectrum corresponding to the i-th type of breast tissue component; ε(λ) represents the unmixing residual term; by minimizing the unmixing residual term ε(λ), the contribution weights of various types of breast tissue components are obtained, and the proportions of different breast tissue components are separated from the mixed spectrum in vivo. The classification output unit fuses the spectral features output by the domain-aligned attention unit with the contribution weights of breast tissue components output by the spectral demixing unit to obtain a fused feature vector h, which is then input into the classification layer to determine the probability of breast tissue category. ; Among them, W c b is the classification mapping matrix; c For classification bias terms; p = [p1, p2, ..., p c[p] represents the probability distribution of breast tissue categories corresponding to the region to be tested. c This represents the probability that the tested area belongs to the c-th type of breast tissue state.

[0010] The pre-training unit first uses in vitro hyperspectral data from a standardized benchmark sample library as training input and breast tissue state labels obtained from pathological annotations as training targets to pre-train the cross-scale spectral analysis model constructed by the domain alignment attention unit, spectral unmixing unit, and classification output unit. The parameters of the cross-scale spectral analysis model are iteratively optimized using a multi-objective joint loss function. The specific calculation formula is as follows: ; Where L represents the overall training loss of the cross-scale spectral analysis model; L comp This represents the unmixing loss of breast tissue components, used to constrain the tissue component contribution weights in the model output to maintain consistency with the reference spectral weights in the standardized benchmark sample library; L cls L represents the classification loss for breast tissue state, used to constrain the class probability distribution of the model output to remain consistent with the true pathological labels; reg denoted by regularization loss; α is the balance weight coefficient for spectral component unmixing loss, η is the balance weight coefficient for tissue state classification loss, and τ is the balance weight coefficient for regularization loss of cross-scale spectral analysis model parameters. The expression for the tissue component unmixing loss is as follows: ; Among them, w i This indicates the contribution weight of breast tissue components in the output of the spectral demixing unit. This represents the reference spectra of various breast tissue components in a standardized benchmark sample library and the pre-determined weights of the labeled components based on tissue state labels. 2 2 represents the square norm operation; The expression for the organizational state classification loss is: ; Among them, y c p represents the true label indicator corresponding to the c-th type of breast tissue state. c This represents the probability value corresponding to the c-th type of breast tissue state output by the classification output unit; under the constraint of the multi-objective joint loss function, the model parameters are iteratively converged to obtain the trained cross-scale spectral analysis model.

[0011] Furthermore, the report management module includes a storage unit, a contribution analysis unit, and a report generation unit; The storage unit receives and saves the in vivo mixed spectral vector, the contribution weight of breast tissue components, and the probability distribution of breast tissue state output by the feature analysis module. The contribution analysis unit calculates the overall contribution score for each spectral channel. The specific calculation process is as follows: ; Among them, zc i A represents the overall contribution score of the i-th spectral channel; j,i The j-th multi-head attention head represents the attention weight assigned to the i-th spectral channel; HD represents the total number of multi-head attention heads; g i The global gating weight corresponding to the i-th spectral channel is represented; the channels are sorted in descending order according to their comprehensive contribution scores, and the top u spectral bands are selected as the feature frequency bands that make key contributions to the identification of breast tissue status. Based on the comprehensive contribution scores of each spectral channel, and combined with multi-round random perturbation sampling analysis, a band marginal contribution score is constructed. The specific calculation process is as follows: ; Where, Φ i The average marginal contribution score of the i-th spectral channel is represented by T; T represents the number of random perturbation samplings; x i (t) represents the input vector after removing or replacing the i-th spectral channel under the t-th perturbation; f() represents the breast tissue heterogeneity determination function; x represents the in vivo mixed spectral vector; The report generation unit generates the proportion information of various breast tissue components based on the contribution weight of breast tissue components. It also generates the contribution ranking information of each spectral band to the judgment of breast tissue status based on key characteristic frequency bands, comprehensive contribution scores of each spectral channel, and marginal contribution scores of the bands. Finally, it integrates and generates a visual evaluation report that includes a breast tissue heterogeneity distribution map, a component proportion table, and a characteristic band contribution map.

[0012] Furthermore, the closed-loop update module includes an anomaly identification unit, an incremental analysis unit, and an update execution unit; When the anomaly identification unit detects an abnormal outlier signal during the detection process, it determines that the current sample is an abnormal spectral sample and marks it. The incremental analysis unit sends a hyperspectral reacquisition command to the benchmark library construction module to perform supplementary scanning and annotation of the ex vivo breast tissue corresponding to the abnormal spectral samples, and obtains the newly added hyperspectral image data and breast tissue status annotation information. The update execution unit incrementally writes the newly added hyperspectral data, pathological annotation information, and tissue status labels after registering and aligning them according to the format of the standardized benchmark sample library, thereby completing the dynamic expansion of the standardized benchmark sample library. Simultaneously, it triggers the pre-training unit of the feature analysis module to perform incremental learning and iterative optimization of model parameters based on the updated standardized benchmark sample library.

[0013] A method for detecting breast tissue using multi-scale spectral analysis, comprising: S1: By acquiring microscopic hyperspectral images of isolated breast tissue sections, hyperspectral image data containing multiple continuous spectral channels is obtained; the hyperspectral image data is associated and stored with pathological annotation information of breast tissue regions, patient clinical information, and breast tissue status labels to form a standardized benchmark sample library; S2: Irradiate breast tissue with a near-infrared light source, collect the diffuse reflectance spectral signal returned after being scattered by the breast tissue, and normalize and correct the diffuse reflectance spectral signal to generate an in vivo mixed spectral vector. S3: Construct and train a cross-scale spectral analysis model based on a standardized benchmark sample library. Analyze the in vivo mixed spectral vector through the cross-scale spectral analysis model, calculate the matching degree between the in vivo mixed spectrum and the reference spectrum, separate the contribution weights of different breast tissue components from the in vivo mixed spectrum, and output the probability that the target detection area belongs to different breast tissue states. S4: Store the in vivo mixed spectral vector, the contribution weights of different breast tissue components, and the probabilities of different breast tissue states, and generate an evaluation report containing information on the proportion of breast tissue components and the contribution information of characteristic bands. S5: When an abnormal outlier signal is identified during the detection process, the corresponding spectral sample is marked as an abnormal spectral sample; the ex vivo breast tissue corresponding to the abnormal spectral sample is re-scanned with microscopic hyperspectral to obtain new hyperspectral image features and corresponding breast tissue state annotation information, and the new data is written into the standardized benchmark sample library.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention constructs a standardized reference sample library of microscopic hyperspectral images of isolated breast tissue sections, and associates and stores hyperspectral image data with pathological annotation information, patient information and tissue state labels. It establishes a unified correspondence between isolated spectra and in vivo detection signals, solves the problem of large differences in sampling scale between isolated and in vivo samples and the inability to directly map features, and improves the accuracy of breast tissue state discrimination. 2. This invention employs a cross-scale spectral analysis model that combines domain-aligned attention units and spectral unmixing units. By calculating domain-aligned attention, the mixed spectral features in vivo are matched with reference spectral prototypes in the benchmark sample library. The linear unmixing model separates the contribution weights of different breast tissue components from the near-infrared diffuse reflectance signal. This can suppress the interference of background signals such as fat and water, highlight the feature bands that make key contributions to tissue state discrimination, and improve the problems of signal aliasing and weak features being easily submerged. 3. This invention introduces a channel contribution gating vector and a reference spectrum guidance mechanism into the feature analysis module. It adaptively enhances the weight of high contribution bands through global pooling and gating mapping, and optimizes the model by combining a multi-objective joint loss function composed of tissue component unmixing loss and tissue state classification loss. This improves the utilization efficiency of high-dimensional spectral data, enhances the robustness and generalization ability of the model, and improves the accuracy of detection results. 4. When an abnormal outlier signal is identified during the detection process through the closed-loop update module, the ex vivo tissue corresponding to the abnormal sample is re-scanned with microscopic hyperspectral imaging. The new data is written into the benchmark sample library and the incremental learning of the model is triggered. This forms a closed-loop iterative mechanism of detection, identification, resampling, updating and optimization, realizing the continuous evolution of the sample library and analysis model. It can continuously improve the detection stability and accuracy in long-term applications. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the process of a breast tissue detection system based on cross-scale spectral analysis according to the present invention. Detailed Implementation

[0016] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example: Figure 1 As shown, the present invention provides a technical solution, a breast tissue detection system for cross-scale spectral analysis, including a benchmark library construction module, a signal acquisition module, a feature analysis module, a report management module, and a closed-loop update module; The benchmark library construction module acquires hyperspectral images of isolated breast tissue sections using microscopic hyperspectral imaging to obtain hyperspectral image data containing multiple continuous spectral channels; it then associates and stores the hyperspectral image data with pathological annotation information of the breast tissue region, patient clinical information, and breast tissue status labels to form a standardized benchmark sample library. The signal acquisition module irradiates breast tissue with a near-infrared light source, acquires the diffuse reflectance spectral signal returned after being scattered by the breast tissue, and normalizes and corrects the diffuse reflectance spectral signal to generate an in vivo mixed spectral vector. The feature analysis module constructs and trains a cross-scale spectral analysis model based on a standardized benchmark sample library. The cross-scale spectral analysis model is used to analyze the in vivo mixed spectral vector, calculate the matching degree between the in vivo mixed spectrum and the reference spectrum, separate the contribution weights of different breast tissue components from the in vivo mixed spectrum, and output the probability that the target detection area belongs to different breast tissue states. The report management module is used to store in vivo mixed spectral vectors, contribution weights of different breast tissue components, and probabilities of different breast tissue states, and to generate an assessment report containing information on the proportion of breast tissue components and contribution information of characteristic bands. When the closed-loop update module identifies an abnormal outlier signal during the detection process, it marks the corresponding spectral sample as an abnormal spectral sample; it then performs a new microscopic hyperspectral scan on the ex vivo breast tissue corresponding to the abnormal spectral sample to obtain new hyperspectral image features and corresponding breast tissue state annotation information, and writes the new data into the standardized benchmark sample library.

[0018] The benchmark library construction module includes a hyperspectral imaging unit, a labeling and association unit, and a structured storage unit; The hyperspectral imaging unit scans ex vivo breast tissue sections to acquire 256 continuous channels of hyperspectral image cube X-rays in the 400-1000 nm band. path ∈R H×W×256 Where H and W represent the spatial height and width of the pathological slide image, respectively; 256 represents the number of continuous spectral channels acquired in the 400-1000 nm band. The annotation and association unit is first established by pathologists based on stained sections and relevant examination results, who perform morphological annotation of the lesion area and combine it with immunohistochemistry, molecular typing, or pathological diagnosis results to complete the biochemical characteristics and tissue grade labeling. Then, a one-to-one correspondence between pathological images, hyperspectral data, and test results is established through section number, patient number, and lesion area coordinates. The pathological area annotation is mapped to the hyperspectral image space, patient information is uniquely bound to the sample, and the test conclusion is converted into a tissue state label. For example, based on the tissue heterogeneity grade determined by pathological analysis, the sample is labeled as one of four categories: normal tissue, fat-rich area, glandular hyperplasia area, or abnormal heterogeneous area. Structured storage units perform registration, alignment, and encoding on hyperspectral image data, pathological annotation information, patient clinical information, and breast tissue status labels to form a standardized benchmark sample library. ;in, M represents the cube containing the pathological hyperspectral image of the nth ex vivo breast tissue sample. (n) C represents the annotation information of the lesion region corresponding to the nth isolated breast tissue sample. (n) y represents the clinicopathological attribute information of the patient associated with the nth ex vivo breast tissue sample. (n) This indicates the tissue status label of the nth isolated breast tissue sample, determined based on clinical test results.

[0019] The signal acquisition module includes a signal acquisition unit, a signal processing unit, and a signal quality assessment unit; The signal acquisition unit irradiates in-vivo breast tissue in the 700nm to 1700nm waveband through a near-infrared light source and a detector, collects the diffuse reflection signal of the in-vivo breast tissue, and obtains the original light intensity signal of the breast tissue to be measured; signal acquisition is mainly based on diffuse reflection theory. Since biological tissue is a highly scattering medium, incident light returns to the detection end after experiencing multiple scattering and absorption inside the tissue, and its diffuse reflectance R d and the absorption coefficient μ a and the reduced scattering coefficient μ s ' satisfy a nonlinear relationship, which is expressed by the diffusion approximation model as: ; wherein λ represents the acquisition wavelength, R d (λ) represents the diffuse reflectance of breast tissue at wavelength λ, A0 represents a dimensionless boundary correction parameter related to tissue boundary conditions and refractive index mismatch, and its value is positive. In this embodiment, A0 is determined according to the equivalent refractive index of the contact interface between the measurement cover, air, coupling medium and breast tissue, with a value range of 0.5 to 5.0; A0 is specifically expressed as: A0=(1+Reff) / (1-Reff), wherein Reff is the equivalent internal reflection coefficient, 0≤Reff<1, which is calculated by the Fresnel reflection relationship from the refractive index difference between the breast tissue and the external medium, or calibrated by a standard diffuse reflector / tissue phantom; a'(λ) represents the single-scattering albedo corresponding to wavelength λ, which is used to characterize the relative proportion of scattering in the process of light transmission in tissue, with a value range of 0<a'(λ)<1; in near-infrared detection of breast tissue, since breast tissue is usually a highly scattering medium, the value range of a'(λ) is 0.7 to 0.99, and a'(λ) is expressed as: ; μ a (λ) represents the absorption coefficient of breast tissue corresponding to wavelength λ, which is used to characterize the absorption capacity of breast tissue for near-infrared light energy; μ s '(λ) represents the reduced scattering coefficient of breast tissue corresponding to wavelength λ, which is used to characterize the effective scattering intensity of internal scattering of breast tissue after anisotropy correction; μ a (λ) and μ s '(λ) are both non-negative parameters varying with wavelength, and in near-infrared detection of breast tissue, μ s '(λ)>μ a (λ) is satisfied; Specifically, μ a (λ) is determined by fitting according to the reference absorption spectra and corresponding concentration weights of main absorbing components such as water, fat, hemoglobin and collagen in breast tissue, and is expressed as: μ a(λ)=Σ k m k e k (λ); Where k represents the absorber component number; m k Indicates the concentration or relative content weight of the k-th absorbent component; e k (λ) represents the reference absorption coefficient of the k-th absorbing component at wavelength λ; μ s '(λ) can be determined through tissue simulation calibration, standard reference measurement, or power-law scattering model fitting, and its expression is: μ s '(λ)=A s (λ / λ0) -b ; Where λ0 represents the preset reference wavelength, A s The parameter represents the scattering amplitude at the reference wavelength λ0; b represents the scattering spectrum index; A s b is obtained through tissue simulation experiments or by regression analysis based on historical spectral data of the same tissue type in a standardized benchmark sample library; In actual testing, the system first uses dark current signal, standard reference light intensity, and tissue simulator to complete instrument calibration, obtaining A0 and A2. s The initial values ​​of b and the weights of each absorber component are determined; then, based on the measured diffuse reflectance spectral signal of the breast tissue to be tested, the diffuse reflectance R calculated by the diffusion approximation model is obtained through least squares fitting or iterative inversion. d (λ) and the measured normalized reflectance R norm The error between (λ) is minimized, thus determining the μ corresponding to the current sample to be tested. a (λ), μ s '(λ) and a'(λ).

[0020] The signal processing unit normalizes the diffuse reflectance signal based on the light intensity of the tissue under test, the dark current signal, and the standard reference light intensity to obtain the normalized reflectance signal. The calculation formula is: μ a (λ)=Σ k m k e k (λ); Where k represents the absorber component number; m k Indicates the concentration or relative content weight of the k-th absorbent component; e k (λ) represents the reference absorption coefficient of the k-th absorbing component at wavelength λ; μ s '(λ) can be determined through tissue simulation calibration, standard reference measurement, or power-law scattering model fitting, and its expression is: μ s'(λ)=A s (λ / λ0) -b ; Where λ0 represents the preset reference wavelength, A s The parameter represents the scattering amplitude at the reference wavelength λ0; b represents the scattering spectrum index; A s b is obtained through tissue simulation experiments or by regression analysis based on historical spectral data of the same tissue type in a standardized benchmark sample library; In actual testing, the system first uses dark current signal, standard reference light intensity, and tissue simulator to complete instrument calibration, obtaining A0 and A2. s The initial values ​​of b and the weights of each absorber component are determined; then, based on the measured diffuse reflectance spectral signal of the breast tissue to be tested, the diffuse reflectance R calculated by the diffusion approximation model is obtained through least squares fitting or iterative inversion. d (λ) and the measured normalized reflectance R norm The error between (λ) is minimized, thus determining the μ corresponding to the current sample to be tested. a (λ), μ s '(λ) and a'(λ).

[0021] The signal processing unit normalizes the diffuse reflectance signal based on the light intensity of the tissue under test, the dark current signal, and the standard reference light intensity to obtain the normalized reflectance signal. The calculation formula is as follows: ; Where λ represents the acquisition wavelength; R norm (λ) represents the normalized reflectance at wavelength λ; I sample (λ) represents the original measured light intensity of the breast tissue sample at wavelength λ; I dark (λ) represents the dark current background signal measured by the system under conditions of no incident light; I white (λ) represents the reference light intensity measured at the same wavelength by a standard white board or a reference highly reflective object; through the above normalization process, the original signal is converted into a comparable standard reflection response, thereby improving the consistency of data from different devices, at different times and under different environments. The signal quality assessment unit calculates the signal-to-noise ratio (SNR) of the normalized signal, identifies signals that meet a preset SNR threshold as valid signals, and arranges the valid signals according to the wavelength dimension to form a one-dimensional bulk mixed spectral vector. The formula for calculating the SNR is: ; Among them, SNR dB P represents the signal-to-noise ratio, expressed in decibels. signal Indicates effective signal power; P noise Indicates noise power.

[0022] The feature analysis module includes a domain-aligned attention unit, a spectral demixing unit, a classification output unit, and a pre-training unit; The domain-aligned attention unit performs local feature extraction and embedding transformation on the in vivo mixed spectral vector to obtain a spectral feature tensor; it extracts reference spectral prototypes of various breast tissue components from hyperspectral image data in a standardized benchmark sample library to form a reference spectral prototype set R; then, domain-aligned attention calculation is performed. ; Wherein, DAAttn() represents the breast spectral features output after domain-aligned attention-weighted aggregation of the in vivo mixed spectrum and the reference spectral prototype; Q is the query matrix, representing the in vivo mixed spectral features of the currently input in vivo breast tissue to be tested; K is the key matrix, representing the spectral features of various components of breast tissue extracted from the standardized benchmark sample library; V is the value matrix, representing the original spectral information of breast tissue; d k The feature dimension of the key vector is used to scale the dot product matching results of spectral features; denoted as the standard dot product attention term, representing the similarity between different spectral channels; G is the reference spectral guidance matrix, representing the degree of cross-scale matching between the spectral features of the in vivo breast tissue to be tested and the reference spectral prototype of breast tissue in the in vitro benchmark library; B is the spectral channel contribution bias matrix, used to highlight the feature bands that have a higher discriminative contribution to the differences in breast tissue composition and tissue state; β and γ are learnable weight coefficients, used to adjust the strength of the role of the reference spectral guidance term and the channel contribution bias term in the overall attention weight allocation process of breast spectral data; softmax() represents the normalization function, used to convert the association matching scores of each spectral channel into corresponding weights.

[0023] The reference spectral guiding matrix G is calculated as follows: ; Where Norm() represents the normalization operation; W q r With W r As a learnable projection matrix, the in vivo mixed spectral features and the reference spectral prototype are mapped to the same alignment space, respectively; RW r This represents the mapping result of the reference spectral prototype in the alignment space; Construct the channel contribution gating vector g: ; Where F represents the input 256-channel spectral feature tensor; Pool() represents the global pooling operation, which compresses the high-dimensional spectral features into a global statistical description; z represents the global feature vector after pooling; W g Represents the gating mapping matrix; b gThe bias term is represented by σ; σ() represents the Sigmoid activation function; for the j-th spectral channel, its contribution weight is denoted as g. j ∈(0,1), representing the relative contribution of the j-th band to the determination of breast tissue status; the contribution weights of all spectral channels constitute the gating vector g=[g1,g2,…,g 256 Subsequently, based on the channel contribution gating vector g, the dimensionality is expanded to construct the spectral channel contribution bias matrix B, which participates in the domain alignment attention calculation of the domain alignment attention unit.

[0024] The spectral demixing unit separates the tissue components of the breast spectral features output by the domain-aligned attention unit, using the formula: ; Among them, S in (λ) represents the response intensity of the mixed spectrum acquired in vivo at wavelength λ; M represents the number of categories of reference breast tissue components; w i w represents the contribution weight of the i-th type of breast tissue component in the current mixed spectrum. i ≥0 and ∑ i=1 M w i =1;S ref,i (λ) represents the reference spectrum corresponding to the i-th type of breast tissue component; ε(λ) represents the unmixing residual term; by minimizing the unmixing residual term ε(λ), the contribution weights of various types of breast tissue components are obtained, and the proportions of different breast tissue components are separated from the mixed spectrum in vivo. The classification output unit fuses the spectral features output by the domain-aligned attention unit with the contribution weights of breast tissue components output by the spectral demixing unit to obtain a fused feature vector h, which is then input into the classification layer to determine the probability of breast tissue category. ; Among them, W c b is the classification mapping matrix; c For classification bias terms; p = [p1, p2, ..., p c [ ] represents the probability distribution of breast tissue categories corresponding to the region to be tested, c is the total number of breast tissue categories, and p c This represents the probability that the tested area belongs to the c-th type of breast tissue state.

[0025] The pre-training unit first uses in vitro hyperspectral data from a standardized benchmark sample library as training input and breast tissue state labels obtained from pathological annotations as training targets to pre-train the cross-scale spectral analysis model constructed by the domain alignment attention unit, spectral unmixing unit, and classification output unit. The parameters of the cross-scale spectral analysis model are iteratively optimized using a multi-objective joint loss function. The specific calculation formula is as follows: ; Where L represents the overall training loss of the cross-scale spectral analysis model; L comp This represents the unmixing loss of breast tissue components, used to constrain the tissue component contribution weights in the model output to maintain consistency with the reference spectral weights in the standardized benchmark sample library; L cls L represents the classification loss for breast tissue state, used to constrain the class probability distribution of the model output to remain consistent with the true pathological labels; reg denoted by regularization loss; α is the balance weight coefficient for spectral component unmixing loss, η is the balance weight coefficient for tissue state classification loss, and τ is the balance weight coefficient for regularization loss of cross-scale spectral analysis model parameters. The expression for the tissue component unmixing loss is as follows: ; Among them, w i This indicates the contribution weight of breast tissue components in the output of the spectral demixing unit. This represents the reference spectra of various breast tissue components in a standardized benchmark sample library and the pre-determined weights of the labeled components based on tissue state labels. 2 2 represents the square norm operation; The expression for the organizational state classification loss is: ; Among them, y c p represents the true label indicator corresponding to the c-th type of breast tissue state. c This represents the probability value corresponding to the c-th type of breast tissue state output by the classification output unit; under the constraint of the multi-objective joint loss function, the model parameters are iteratively converged to obtain the trained cross-scale spectral analysis model.

[0026] The report management module includes a storage unit, a contribution analysis unit, and a report generation unit; The storage unit receives and saves the in vivo mixed spectral vector, the contribution weight of breast tissue components, and the probability distribution of breast tissue state output by the feature analysis module. The contribution analysis unit calculates the overall contribution score for each spectral channel. The specific calculation process is as follows: ; Among them, zc i A represents the overall contribution score of the i-th spectral channel; j,i The j-th multi-head attention head represents the attention weight assigned to the i-th spectral channel; HD represents the total number of multi-head attention heads; g iThe global gating weight corresponding to the i-th spectral channel is represented; the channels are sorted in descending order according to their comprehensive contribution scores, and the top u spectral bands are selected as the feature frequency bands that make key contributions to the identification of breast tissue status. Based on the comprehensive contribution scores of each spectral channel, and combined with multi-round random perturbation sampling analysis, a band marginal contribution score is constructed. The specific calculation process is as follows: ; Where, Φ i The average marginal contribution score of the i-th spectral channel is represented by T; T represents the number of random perturbation samplings; x i (t) denoted as the input vector after removing or replacing the i-th spectral channel under the t-th perturbation; x represents the in vivo mixed spectral vector; f() represents the breast tissue heterogeneity determination function, with the in vivo mixed spectral vector x as the input and the breast tissue heterogeneity score Sh(x) as the output. The larger the value of Sh(x), the more mixed the tissue components in the current detection area, the more uncertain the tissue state discrimination, and the larger the unmixing residual between the in vivo mixed spectrum and the reference spectrum, the higher the possibility of abnormal heterogeneity in the corresponding detection area. Specifically, the in vivo mixed spectral vector x is input into the trained cross-scale spectral analysis model to obtain the breast tissue component contribution weight vector, the breast tissue state probability distribution, and the spectral unmixing residual. The breast tissue heterogeneity determination function f(x) is expressed as: f(x)=Sh(x)=θ1Hc(x)+θ2Hs(x)+θ3En(x); Wherein, θ1, θ2, and θ3 represent the weighting coefficients of the breast tissue component confounding term, the breast tissue state uncertainty term, and the normalized spectral unmixing residual term, respectively, and θ1≥0, θ2≥0, and θ3≥0, and θ1+θ2+θ3=1; Hc(x) represents the degree of breast tissue component confounding, which is quantified using normalized information entropy based on the contribution weight of breast tissue components. The larger Hc(x) is, the higher the degree of mixing of different breast tissue components in the current detection area; based on the probability distribution of breast tissue state, normalized information entropy is used. The calculation is performed using a quantitative method. The larger the value of Hs(x), the more uncertain the model is in judging the tissue state of the current detection area. En(x) represents the normalized spectral unmixing residual. First, the mean square error of the spectral unmixing residual corresponding to each wavelength position is calculated to obtain the overall aggregated residual value. Then, the aggregated residual value is normalized by dividing it by the sum of the aggregated residual value and the regularization parameter, so that the result is mapped to the [0,1] interval. The regularization parameter is determined based on the average unmixing residual of normal breast tissue samples in the standardized benchmark sample library.

[0027] The report generation unit generates the proportion information of various breast tissue components based on the contribution weight of breast tissue components. It also generates the contribution ranking information of each spectral band to the judgment of breast tissue status based on key characteristic frequency bands, comprehensive contribution scores of each spectral channel, and marginal contribution scores of the bands. Finally, it integrates and generates a visual evaluation report that includes a breast tissue heterogeneity distribution map, a component proportion table, and a characteristic band contribution map.

[0028] The closed-loop update module includes an anomaly identification unit, an incremental analysis unit, and an update execution unit; When the anomaly identification unit detects an abnormal outlier signal during the detection process, it determines that the current sample is an abnormal spectral sample and marks it. The incremental analysis unit sends a hyperspectral reacquisition command to the benchmark library construction module to perform supplementary scanning and annotation of the ex vivo breast tissue corresponding to the abnormal spectral samples, and obtains the newly added hyperspectral image data and breast tissue status annotation information. The update execution unit incrementally writes the newly added hyperspectral data, pathological annotation information, and tissue status labels after registering and aligning them according to the format of the standardized benchmark sample library, thereby completing the dynamic expansion of the standardized benchmark sample library. Simultaneously, it triggers the pre-training unit of the feature analysis module to perform incremental learning and iterative optimization of model parameters based on the updated standardized benchmark sample library.

[0029] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A breast tissue detection system based on cross-scale spectral analysis, characterized in that: The system includes a benchmark library construction module, a signal acquisition module, a feature analysis module, a report management module, and a closed-loop update module; The benchmark library construction module acquires hyperspectral images of isolated breast tissue sections using microscopic hyperspectral imaging to obtain hyperspectral image data containing multiple continuous spectral channels; the hyperspectral image data is then associated and stored with pathological annotation information of the breast tissue region, patient clinical information, and breast tissue status labels to form a standardized benchmark sample library. The signal acquisition module irradiates breast tissue with a near-infrared light source, acquires the diffuse reflectance spectral signal returned after being scattered by the breast tissue, and normalizes and corrects the diffuse reflectance spectral signal to generate an in vivo mixed spectral vector. The feature analysis module constructs and trains a cross-scale spectral analysis model based on a standardized benchmark sample library. The cross-scale spectral analysis model is used to analyze the in vivo mixed spectral vector, calculate the degree of matching between the in vivo mixed spectrum and the reference spectrum, separate the contribution weights of different breast tissue components from the in vivo mixed spectrum, and output the probability that the target detection area belongs to different breast tissue states. The report management module is used to store in vivo mixed spectral vectors, contribution weights of different breast tissue components, and probabilities of different breast tissue states, and to generate an assessment report containing information on the proportion of breast tissue components and contribution information of characteristic bands. When the closed-loop update module identifies an abnormal outlier signal during the detection process, it marks the corresponding spectral sample as an abnormal spectral sample. The ex vivo breast tissue corresponding to the abnormal spectral sample was re-scanned using microscopic hyperspectral scanning to obtain new hyperspectral image features and corresponding breast tissue state annotation information, and the new data was written into the standardized benchmark sample library.

2. The breast tissue detection system for cross-scale spectral analysis according to claim 1, characterized in that: The benchmark library construction module includes a hyperspectral imaging unit, an annotation and association unit, and a structured storage unit; The hyperspectral imaging unit scans isolated breast tissue slices to acquire a cube X-ray hyperspectral image with 256 continuous channels in the 400-1000nm band. path ∈R H×W×256 Where H and W represent the spatial height and width of the pathological slide image, respectively; 256 represents the number of continuous spectral channels acquired in the 400-1000 nm band. The annotation association unit binds pathological annotation information, patient clinical information, and breast tissue status tags; The structured storage unit performs registration, alignment, and encoding on hyperspectral image data, pathological annotation information, patient clinical information, and breast tissue status labels to form a standardized benchmark sample library. ;in, M represents the cube containing the pathological hyperspectral image of the nth ex vivo breast tissue sample. (n) C represents the annotation information of the lesion region corresponding to the nth isolated breast tissue sample. (n) y represents the clinicopathological attribute information of the patient associated with the nth ex vivo breast tissue sample. (n) This indicates the tissue status label of the nth isolated breast tissue sample, determined based on clinical test results.

3. The breast tissue detection system for cross-scale spectral analysis according to claim 2, characterized in that: The signal acquisition module includes a signal acquisition unit, a signal processing unit, and a signal quality evaluation unit; The signal acquisition unit irradiates the breast tissue in vivo with a near-infrared light source and detector in the 700nm to 1700nm wavelength range, and acquires the diffuse reflection signal of the breast tissue in vivo to obtain the original light intensity signal of the breast tissue to be tested. The signal processing unit performs normalization calculation on the diffuse reflection signal based on the light intensity of the tissue under test, the dark current signal, and the standard reference light intensity to obtain the normalized reflectance signal. The signal quality assessment unit calculates the signal-to-noise ratio of the normalized signal, determines the signal that meets the preset signal-to-noise ratio threshold as a valid signal, and arranges the valid signals according to the wavelength dimension to form a one-dimensional bulk mixed spectral vector.

4. The breast tissue detection system for cross-scale spectral analysis according to claim 3, characterized in that: The feature analysis module includes a domain-aligned attention unit, a spectral demixing unit, a classification output unit, and a pre-training unit; The domain-aligned attention unit performs local feature extraction and embedding transformation on the in vivo mixed spectral vector to obtain a spectral feature tensor; it extracts reference spectral prototypes of various breast tissue components from hyperspectral image data in a standardized benchmark sample library to form a reference spectral prototype set R; then, domain-aligned attention calculation is performed. ； Wherein, DAAttn() represents the breast spectral features output after domain-aligned attention-weighted aggregation of the in vivo mixed spectrum and the reference spectral prototype; Q is the query matrix, representing the in vivo mixed spectral features of the currently input in vivo breast tissue to be tested; K is the key matrix, representing the spectral features of various components of breast tissue extracted from the standardized benchmark sample library; V is the value matrix, representing the original spectral information of breast tissue; d k The feature dimension of the key vector is used to scale the dot product matching results of spectral features; denoted as the standard dot product attention term, representing the similarity between different spectral channels; G is the reference spectral guidance matrix, representing the degree of cross-scale matching between the spectral features of the in vivo breast tissue to be tested and the reference spectral prototype of breast tissue in the in vitro benchmark library; B is the spectral channel contribution bias matrix, used to highlight the feature bands that have a higher discriminative contribution to the differences in breast tissue composition and tissue state; β and γ are learnable weight coefficients, used to adjust the strength of the role of the reference spectral guidance term and the channel contribution bias term in the overall attention weight allocation process of breast spectral data; softmax() represents the normalization function, used to convert the association matching scores of each spectral channel into corresponding weights.

5. The breast tissue detection system for cross-scale spectral analysis according to claim 4, characterized in that: The reference spectral guiding matrix G is calculated as follows: ； Where F represents the input 256-channel spectral feature tensor; Pool() represents the global pooling operation, which compresses the high-dimensional spectral features into a global statistical description; z represents the global feature vector after pooling; W g Represents the gating mapping matrix; b g The bias term is represented by σ; σ() represents the Sigmoid activation function; for the j-th spectral channel, its contribution weight is denoted as g. j ∈(0,1), representing the relative contribution of the j-th band to the determination of breast tissue status; the contribution weights of all spectral channels constitute the gating vector g=[g1,g2,…,g 256 Subsequently, based on the channel contribution gating vector g, the dimensionality is expanded to construct the spectral channel contribution bias matrix B, which participates in the domain alignment attention calculation of the domain alignment attention unit.

6. The breast tissue detection system for cross-scale spectral analysis according to claim 5, characterized in that: The spectral demixing unit performs tissue component separation on the breast spectral features output by the domain-aligned attention unit, using the formula: ； Among them, S in (λ) represents the response intensity of the mixed spectrum acquired in vivo at wavelength λ; M represents the number of categories of reference breast tissue components; w i w represents the contribution weight of the i-th type of breast tissue component in the current mixed spectrum. i ≥0 and ∑ i= 1 M w i =1;S ref,i (λ) represents the reference spectrum corresponding to the i-th type of breast tissue component; ε(λ) represents the unmixing residual term; by minimizing the unmixing residual term ε(λ), the contribution weights of various types of breast tissue components are obtained, and the proportions of different breast tissue components are separated from the mixed spectrum in vivo. The classification output unit fuses the spectral features output by the domain-aligned attention unit with the contribution weights of breast tissue components output by the spectral demixing unit to obtain a fused feature vector h, which is then input into the classification layer to determine the probability of breast tissue category. ； Among them, W c b is the classification mapping matrix; c For classification bias terms; p = [p1, p2, ..., p c [p] represents the probability distribution of breast tissue categories corresponding to the region to be tested. c This represents the probability that the tested area belongs to the c-th type of breast tissue state.

7. The breast tissue detection system for cross-scale spectral analysis according to claim 6, characterized in that: The pre-training unit first uses in vitro hyperspectral data from a standardized benchmark sample library as training input and breast tissue state labels obtained from pathological annotation as training target to pre-train the cross-scale spectral analysis model constructed by the domain alignment attention unit, spectral unmixing unit, and classification output unit. The parameters of the cross-scale spectral analysis model are iteratively optimized using a multi-objective joint loss function. The specific calculation formula is as follows: ； Where L represents the overall training loss of the cross-scale spectral analysis model; L comp This represents the unmixing loss of breast tissue components, used to constrain the tissue component contribution weights in the model output to maintain consistency with the reference spectral weights in the standardized benchmark sample library; L cls L represents the classification loss for breast tissue state, used to constrain the class probability distribution of the model output to remain consistent with the true pathological labels; reg denoted by regularization loss; α is the balance weight coefficient for spectral component unmixing loss, η is the balance weight coefficient for tissue state classification loss, and τ is the balance weight coefficient for regularization loss of cross-scale spectral analysis model parameters. The expression for the tissue component unmixing loss is: ； Among them, w i This indicates the contribution weight of breast tissue components in the output of the spectral demixing unit. This represents the reference spectra of various breast tissue components in a standardized benchmark sample library and the pre-determined weights of the labeled components based on tissue state labels. 2 2 represents the square 2 norm operation; The expression for the organizational state classification loss is: ； y c p represents the true label indicator corresponding to the c-th type of breast tissue state. c This represents the probability value corresponding to the c-th type of breast tissue state output by the classification output unit; under the constraint of the multi-objective joint loss function, the model parameters are iteratively converged to obtain the trained cross-scale spectral analysis model.

8. The breast tissue detection system for cross-scale spectral analysis according to claim 7, characterized in that: The report management module includes a storage unit, a contribution analysis unit, and a report generation unit; The storage unit receives and saves the in vivo mixed spectral vector, the contribution weight of breast tissue components, and the probability distribution of breast tissue state output by the feature analysis module. The contribution analysis unit calculates the overall contribution score for each spectral channel. The specific calculation process is as follows: ； Among them, zc i A represents the overall contribution score of the i-th spectral channel; j,i The j-th multi-head attention head represents the attention weight assigned to the i-th spectral channel; HD represents the total number of multi-head attention heads; g i The global gating weight corresponding to the i-th spectral channel is represented; the channels are sorted in descending order according to their comprehensive contribution scores, and the top u spectral bands are selected as the feature frequency bands that make key contributions to the identification of breast tissue status. Based on the comprehensive contribution scores of each spectral channel, and combined with multi-round random perturbation sampling analysis, a band marginal contribution score is constructed. The specific calculation process is as follows: ； Where, Φ i The average marginal contribution score of the i-th spectral channel is represented by T; T represents the number of random perturbation samplings; x i (t) represents the input vector after removing or replacing the i-th spectral channel under the t-th perturbation; f() represents the breast tissue heterogeneity determination function; x represents the in vivo mixed spectral vector; The report generation unit generates the proportion information of various breast tissue components based on the contribution weight of breast tissue components, generates the contribution ranking information of each spectral band to the judgment of breast tissue status based on key feature frequency bands, comprehensive contribution scores of each spectral channel and marginal contribution scores of bands, and integrates them to generate a visual evaluation report containing a breast tissue heterogeneity distribution map, a component proportion table and a feature band contribution map.

9. The breast tissue detection system for cross-scale spectral analysis according to claim 8, characterized in that: The closed-loop update module includes an anomaly identification unit, an incremental analysis unit, and an update execution unit. When the anomaly identification unit identifies an abnormal outlier signal during the detection process, it determines that the current sample is an abnormal spectral sample and marks it. The incremental analysis unit sends a hyperspectral reacquisition command to the benchmark library construction module to perform supplementary scanning and annotation of the ex vivo breast tissue corresponding to the abnormal spectral samples, and obtains newly added hyperspectral image data and breast tissue status annotation information. The update execution unit incrementally writes the newly added hyperspectral data, pathological annotation information, and tissue status labels after registering and aligning them according to the format of the standardized benchmark sample library, thereby completing the dynamic expansion of the standardized benchmark sample library. Simultaneously, it triggers the pre-training unit of the feature analysis module to perform incremental learning and iterative optimization of model parameters based on the updated standardized benchmark sample library.

10. A method for detecting breast tissue using cross-scale spectral analysis, applied to the breast tissue detection system using cross-scale spectral analysis as described in any one of claims 1-9, characterized in that: The method includes: S1: By acquiring microscopic hyperspectral images of isolated breast tissue sections, hyperspectral image data containing multiple continuous spectral channels is obtained; the hyperspectral image data is associated and stored with pathological annotation information of breast tissue regions, patient clinical information, and breast tissue status labels to form a standardized benchmark sample library; S2: Irradiate breast tissue with a near-infrared light source, collect the diffuse reflectance spectral signal returned after being scattered by the breast tissue, and normalize and correct the diffuse reflectance spectral signal to generate an in vivo mixed spectral vector. S3: Construct and train a cross-scale spectral analysis model based on a standardized benchmark sample library. Analyze the in vivo mixed spectral vector through the cross-scale spectral analysis model, calculate the matching degree between the in vivo mixed spectrum and the reference spectrum, separate the contribution weights of different breast tissue components from the in vivo mixed spectrum, and output the probability that the target detection area belongs to different breast tissue states. S4: Store the in vivo mixed spectral vector, the contribution weights of different breast tissue components, and the probabilities of different breast tissue states, and generate an evaluation report containing information on the proportion of breast tissue components and the contribution information of characteristic bands. S5: When an abnormal outlier signal is identified during the detection process, the corresponding spectral sample is marked as an abnormal spectral sample; the ex vivo breast tissue corresponding to the abnormal spectral sample is re-scanned with microscopic hyperspectral to obtain new hyperspectral image features and corresponding breast tissue state annotation information, and the new data is written into the standardized benchmark sample library.

Images