Analysis method for breast cancer metabolic heterogeneity based on label-free hyperspectral raman imaging

Abstract

The application provides a breast cancer metabolic heterogeneity analysis method based on a label-free hyperspectral Raman imaging, and belongs to the technical field of medical diagnosis and spectral imaging. The method comprises the following steps: obtaining a fresh or low-temperature stored breast cancer tissue sample; using a hyperspectral Raman imaging system to perform large-area scanning on the sample, and obtaining an original spectral data cube containing complete biochemical fingerprint information; pre-processing the original data to remove noise and background; using a non-negative least squares regression algorithm, based on a pre-established standard spectrum library of lipids and extracellular matrix, demixing the pre-processed spectral data, and quantitatively obtaining spatial distribution abundance maps of at least 15 kinds of lipid molecules and various ECM components. The application realizes in-situ, label-free, multi-component and quantitative metabolic phenotype imaging of breast cancer tissue, and provides a new technical means for accurate pathological diagnosis and individualized treatment.

Description

Technical Field

[0001] This invention relates to the fields of medical detection and biophotonics technology, specifically to a method, system, and application for quantitative analysis of metabolic heterogeneity in breast cancer tissue based on label-free hyperspectral Raman imaging. Background Technology

[0002] Breast cancer is the most common malignant tumor among women worldwide, and its high heterogeneity is a major cause of treatment resistance and recurrence. Currently, clinical diagnosis mainly relies on histopathological examination and immunohistochemistry (IHC) to detect markers such as estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), and Ki-67. However, these methods have significant limitations: 1. They rely on staining and labeling, a cumbersome process that can introduce human error; 2. They can only detect a limited number of targets, failing to comprehensively reflect the metabolic state of the tumor; 3. They provide semi-quantitative information based on color intensity, lacking precise quantitative capabilities; 4. Most importantly, they cannot visually present the in-situ distribution and interactions of various metabolic molecules (such as different lipid types) while maintaining the integrity of the tissue's spatial structure.

[0003] Metabolic reprogramming, particularly lipid metabolism abnormalities, has been proven to be a key driver of breast cancer development and progression. Different lipid molecules play different roles in energy supply, membrane structure formation, and signal transduction. Therefore, elucidating the spatial heterogeneity of lipid metabolism within tumors is crucial for understanding tumor biological behavior, differentiating subtypes, predicting prognosis, and identifying therapeutic targets. Existing techniques such as mass spectrometry can provide molecular spatial information, but they typically require complex sample pretreatment, a vacuum environment, and are destructive to tissues. Fluorescence imaging offers high throughput but requires labeling, and the types of labelers are limited.

[0004] Raman spectroscopy is a label-free optical technique based on molecular vibrational fingerprinting, capable of directly acquiring biochemical composition information of tissues. Hyperspectral Raman imaging can acquire the full spectrum of any point in the sample space, and combined with chemometric methods, it can theoretically achieve in-situ visualization of multiple components. However, applying it to the precise lipid metabolism subtyping of clinical breast cancer samples faces the following technical challenges: 1. How to stably and accurately demix and extract quantitative information of up to dozens of lipids and ECM components from complex spectral signals; 2. How to design an effective analytical framework to transform massive spatial-spectral data into biologically significant features that can distinguish subtypes; 3. How to determine a sufficient sampling area to overcome intratumoral heterogeneity and ensure the representativeness of the results. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method, system, and medium for analyzing metabolic heterogeneity in breast cancer based on label-free hyperspectral Raman imaging. This method aims to achieve in situ, multi-component, quantitative, and spatially resolved lipid metabolism phenotypic analysis of breast cancer tissue and to accurately correlate it with clinicopathological subtypes.

[0006] The objective of this invention can be achieved through the following technical solution: a method for analyzing metabolic heterogeneity in breast cancer based on label-free hyperspectral Raman imaging, comprising:

[0007] S1: Obtain fresh frozen sections of breast cancer tissue, with a section thickness of 10-60μm, and place them on a glass slide without staining or labeling.

[0008] S2: A confocal Raman imaging system is used to image and scan tissue sections (on a glass slide), acquiring the Raman spectrum of each pixel to form a hyperspectral data cube;

[0009] S3: Perform spectral preprocessing on the hyperspectral data cube, including cosmic ray removal, fluorescence background subtraction, noise filtering, and intensity normalization;

[0010] S4: Using a non-negative least squares regression algorithm, the spectrum of each preprocessed pixel is decomposed into a linear combination of reference spectra of various biochemical components. The reference spectra include at least the pure substance standard spectra of triglycerides, diacylglycerol, phosphatidylcholine, phosphatidylethanolamine, lysophosphatidic acid, cholesterol, cholesterol esters, cardiolipin, ceramide, sphingosine, docosahexaenoic acid, type I collagen, and laminin, thereby obtaining a quantitative abundance distribution map of each biochemical component in the space of the tissue section.

[0011] S5: Based on the quantitative abundance distribution map of each biochemical component, calculate the spatial Pearson correlation coefficient between all pairs of components and construct the correlation matrix;

[0012] S6: Based on the features extracted from the quantitative abundance distribution map and the correlation matrix, extract and comprehensively evaluate at least two of the following features: a: analyze the spatial enrichment degree and distribution heterogeneity of specific lipid types in the tumor region; b: analyze the co-localization relationship between lipids and extracellular matrix components in spatial distribution; c: analyze the network topology of the correlation matrix, including the degree of modularity, key hub nodes and differences in correlation patterns between different subtypes, to distinguish between healthy tissues and cancerous tissues and to identify molecular subtypes of breast cancer.

[0013] Preferably, in step S2, the confocal Raman imaging system is equipped with a 532nm or 785nm excitation laser, the imaging scan is a point-by-point scan or a line scan, and the scanned field of view area is not less than [amount missing]. And the spatial resolution is better than 10 The Raman spectral range covers 50 Up to 3800 .

[0014] Preferably, in step S4, the reference spectra of the multiple biochemical components are obtained from a pre-established reference spectral library, which contains standard spectra obtained by measuring pure substance standards. The reference spectral library contains standard spectra of at least 20 different lipid molecules and at least 4 extracellular matrix proteins. The multiple known biochemical components include multiple lipid molecules and multiple extracellular matrix proteins.

[0015] Preferably, the identification criteria for the molecular subtype are:

[0016] Triple-negative breast cancer: Supercluster structures with lipid enrichment, spatial heterogeneity exceeding a preset first threshold, and lipid-lipid correlation based on the correlation matrix exceeding a preset second threshold indicate the supercluster structure with the highest spatial heterogeneity and the strongest lipid-lipid correlation.

[0017] HER2 positive: Phospholipids and sterols showed focal enrichment in the tumor nest region, and the lipid-sterol network based on the correlation matrix was within the preset intermediate range;

[0018] Lumen type A: The increased ordered spatial distribution of phospholipids and the spatial correlation between metabolic components and interstitial components below the preset third threshold indicate a weak metabolic-interstitial association.

[0019] The first, second, and third preset thresholds, as well as the medium range, are determined through statistical analysis or machine learning model optimization of training sample datasets of known subtypes.

[0020] The beneficial effects of this invention are:

[0021] 1. This invention requires no staining, labeling or destructive treatment, and directly obtains native metabolic state information from fresh or frozen tissue sections, maintaining the integrity of the tissue structure.

[0022] 2. This invention uses hyperspectral imaging combined with advanced spectral unmixing algorithms to quantitatively visualize the spatial distribution of more than 20 lipids and ECM components in one go, providing a molecular information dimension far exceeding that of traditional IHC.

[0023] 3. This invention not only focuses on the abundance of a single component, but also uses the ventilation space correlation matrix as the core of the analysis to reveal the synergistic or antagonistic relationships between different metabolic pathways and between metabolism and the microenvironment, which has extremely high subtype identification value.

[0024] 4. The analytical workflow established in this invention is compatible with standard pathology workflows, and the analytical results can be directly compared with H&E and IHC images. By defining a clearly defined lower limit for the imaging area, the representativeness of the analytical results for the entire tumor is ensured, avoiding the bias of small sample biopsies.

[0025] 5. The method of this invention can not only distinguish subtypes, but also reveal the specific metabolic vulnerabilities of each subtype, such as the extensive lipid storage and mitochondrial remodeling in TNBC and the tight coupling of the phospholipid-sterol network in HER2+ type, providing a direct basis for developing precision therapies that target metabolism. Attached Figure Description

[0026] The present invention will now be further described with reference to the accompanying drawings.

[0027] Figure 1 This is a schematic diagram of the overall process of the method of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0029] Please see Figure 1 This invention provides a method for analyzing metabolic heterogeneity in breast cancer based on label-free hyperspectral Raman imaging, comprising:

[0030] S1: Obtain fresh frozen sections of breast cancer tissue, with a section thickness of 10-60μm, and place them on a glass slide without staining or labeling.

[0031] S2: A confocal Raman imaging system is used to scan tissue sections (on a glass slide), and the Raman spectrum of each pixel is acquired to form a hyperspectral data cube.

[0032] S3: Perform spectral preprocessing on the hyperspectral data cube, including cosmic ray removal, fluorescence background subtraction, noise filtering, and intensity normalization;

[0033] S4: Using a non-negative least squares regression algorithm, the spectrum of each preprocessed pixel is decomposed into a linear combination of reference spectra of multiple biochemical components. The reference spectra include at least the pure substance standard spectra of triglycerides, diacylglycerol, phosphatidylcholine, phosphatidylethanolamine, lysophosphatidic acid, cholesterol, cholesterol esters, cardiolipin, ceramide, sphingosine, docosahexaenoic acid, type I collagen, and laminin, thereby obtaining a quantitative abundance distribution map of each biochemical component in the space of the tissue section.

[0034] S5: Based on the quantitative abundance distribution map of each biochemical component, calculate the spatial Pearson correlation coefficient between all pairs of components and construct the correlation matrix;

[0035] S6: Based on the features extracted from the quantitative abundance distribution map and the correlation matrix, extract and comprehensively evaluate at least two of the following features: a: analyze the spatial enrichment degree and distribution heterogeneity of specific lipid types in the tumor region; b: analyze the co-localization relationship between lipids and extracellular matrix components in spatial distribution; c: analyze the network topology of the correlation matrix, including the degree of modularity, key hub nodes and differences in correlation patterns between different subtypes, to distinguish between healthy tissues and cancerous tissues and to identify molecular subtypes of breast cancer.

[0036] In step S2, the confocal Raman imaging system is equipped with a 532nm or 785nm excitation laser, and the imaging scan is either a point-by-point scan or a line scan, with a scanned field of view area not less than [amount missing]. And the spatial resolution is better than 10 The Raman spectral range covers 50 Up to 3800 .

[0037] In step S4, the reference spectra of the various biochemical components are obtained from a pre-established reference spectral library, which contains standard spectra obtained by measuring pure substance standards. The reference spectral library contains standard spectra of at least 20 different lipid molecules and at least 4 extracellular matrix proteins. The various known biochemical components include various lipid molecules and various extracellular matrix proteins.

[0038] The identification criteria for the molecular subtype are as follows:

[0039] Triple-negative breast cancer: Supercluster structures with lipid enrichment, spatial heterogeneity exceeding a preset first threshold, and lipid-lipid correlation based on the correlation matrix exceeding a preset second threshold indicate the supercluster structure with the highest spatial heterogeneity and the strongest lipid-lipid correlation.

[0040] HER2 positive: Phospholipids and sterols showed focal enrichment in the tumor nest region, and the lipid-sterol network based on the correlation matrix was within the preset intermediate range;

[0041] Lumen type A: The increased ordered spatial distribution of phospholipids and the spatial correlation between metabolic components and interstitial components below the preset third threshold indicate a weak metabolic-interstitial association.

[0042] The first, second, and third preset thresholds, as well as the medium range, are determined through statistical analysis or machine learning model optimization of training sample datasets of known subtypes.

[0043] Example 1: Raman imaging and metabolic map generation of breast cancer tissue samples.

[0044] Sample Preparation: Fresh breast cancer surgical specimens were collected with ethical approval and informed consent from the patients. The specimens were rapidly harvested on ice, embedded in OCT embedding medium, and serially sectioned on a cryostat to a thickness of 20 μm. The sections were then mounted on CaF2 slides and stored at -80°C until analysis. Simultaneously, adjacent 5 μm sections were prepared for routine H&E and IHC (ER, PR, HER2, Ki-67) staining to complete pathological diagnosis and molecular typing.

[0045] Data Acquisition: A commercial confocal Raman microscope system was used. Sample sections were equilibrated at room temperature and placed on the stage. A 532nm laser was used as the excitation source, focused through a 100x oil immersion objective, with a laser power of approximately 30mW at the sample. The spectrometer integration time was set to 0.5 seconds per pixel. For each tissue sample, a region containing the tumor mass and some adjacent normal tissue was selected for step scanning. The scanning range was approximately 6000×6000μm², with a step size (spatial resolution) of 1μm, ultimately obtaining a hyperspectral data cube containing millions of spectra.

[0046] Data preprocessing: Data processing was performed using a custom Python script. First, a Raman spectroscopy processing library (such as RamPy) was applied for cosmic ray removal and fluorescence background correction (using iterative polynomial fitting). Subsequently, all spectra were vector normalized to eliminate overall intensity differences caused by minor fluctuations in slice thickness or laser power.

[0047] Spectral unmixing and quantitative imaging:

[0048] a. Establish a reference spectral library: Purchase lipid and ECM component standards with a purity >99% (e.g., Sigma-Aldrich), including: palmitic acid, oleic acid, trioleic acid glycerides, dioleic acid glycerides, phosphatidylcholine, phosphatidylethanolamine, sphingomyelin, cholesterol, cholesterol oleate, cardiolipin, ceramide, type I collagen, etc., totaling 25 substances. Measure the spectrum of each pure substance under the same Raman instrument conditions.

[0049] b. Unmixing: For the spectrum of each pixel in the tissue sample data cube, the scipy.optimize.nnls function (non-negative least squares) is used for fitting. The preprocessed tissue spectrum is represented as a linear combination of the reference spectral library matrix, and the non-negative coefficient vector that minimizes the residual is solved. This coefficient represents the relative abundance of each component at that pixel.

[0050] c. Generating a distribution map: The coefficient values ​​of each component at all pixels are rearranged into a two-dimensional matrix, thus obtaining a quantitative distribution map of the component in tissue space. Through pseudo-color encoding, the differential distribution of different lipids in tumor areas, stroma areas, and normal areas can be visually observed.

[0051] Correlation matrix construction: The quantitative distribution maps of the 25 components obtained above (each map is a two-dimensional matrix, expanded into a one-dimensional vector) are input for calculation. The Pearson correlation coefficients r between each pair of these 25 vectors are calculated using the numpy.corrcoef function, resulting in a 25x25 symmetric matrix. The seaborn.heatmap function is used to draw a heatmap, and the rows and columns are sorted using hierarchical clustering to highlight relevant modules.

[0052] Metabolic phenotypic analysis and subtype identification: Taking triple-negative breast cancer (TNBC) samples as an example, their metabolic profiles show the following characteristics, which can distinguish them from other subtypes:

[0053] Lipid Accumulation Pattern: On the quantitative distribution map, the vast majority of lipids in the TNBC samples (especially triglycerides, lysophospholipids, and cardiolipin) showed high-intensity, highly spatially heterogeneous "hotspot" regions. Specifically, the process for identifying "hotspot" regions is as follows: The quantitative abundance distribution map of the target lipid component X is viewed as a matrix V (with dimensions m×n). The quantitative abundance distribution map is a grayscale image, where the brightness value of each pixel represents the abundance of lipid component X at that location. Specifically, a matrix V is constructed, with m rows and n columns. The elements in the i-th row and j-th column of the matrix are... Its value is directly read from the corresponding spatial coordinates in the image. Quantitative abundance. For pixels in matrix V. Calculate its standardization strength and surrounding standard deviation Specifically, standardization intensity The normalized intensity is the result of normalizing the original abundance value by dividing it by the abundance values ​​of all pixels in the entire image. This represents the relative intensity of the pixel; a value closer to 1 indicates a higher relative lipid content in that pixel. (Surrounding standard deviation) Based on pixels Construct a k×k range centered on a given point, calculate the standard deviation of all abundances within this range, and use this value as the standard deviation of the surrounding area. Surrounding standard deviation The larger the value, the more uneven and drastic the lipid distribution in the area surrounding that point. The normalized intensity of each pixel. and surrounding standard deviation The feature vectors of all pixels constitute a dataset. Principal component analysis is then used to project this dataset onto a two-dimensional plane formed by the first two principal components. On this projection plane, most points will densely cluster together (the main cluster), representing a background region with uniform content and gradual changes. The center point c of the main cluster is found, and the projection values ​​for each point are calculated. Euclidean geometric distance to center point c Furthermore, by analyzing known normal and homogeneous tissue sample data, a distance threshold R was obtained, and Euclidean geometric distance was applied. Compare with the distance threshold R to obtain the Euclidean geometric distance. Points with a distance greater than a threshold R are marked as candidate hotspot pixels, meaning they have high relative intensity and high variability. A black-and-white image is generated based on the positions of these candidate hotspot pixels, with candidate points as black and other points as white. The generated black-and-white image is then dilated to connect closely spaced black points into smaller, independent black blocks. The area of ​​each black point contained within each small black block is calculated to obtain the area value. Where m is a positive integer representing the number of black dots, the total area At is calculated, and finally, the result is obtained using the formula... Calculate the structural entropy Hs of the area, where, This represents the area of ​​the u-th black dot. If all black dots are mainly concentrated in a few very large black patches, the Hs value is small; if black dots are scattered in many black patches of roughly the same size, the Hs value is large, indicating that the spatial distribution of candidate points is fragmented and patchy, i.e., high spatial heterogeneity. A high-intensity, high-spatial-heterogeneity hotspot region can only be confirmed in the lipid image when the proportion of candidate hotspot pixels to the total number of pixels is greater than a preset reference proportion and the structural entropy is greater than the reference structural entropy. The reference proportion and reference structural entropy are obtained through statistical analysis of data from a batch of pathologically confirmed, evenly distributed control samples.

[0054] Correlation matrix characteristics: Its correlation matrix heatmap presents a highly integrated supercluster, meaning that the correlation coefficients among various lipids (such as glycerides, phospholipids, and unsaturated fatty acids) are generally high (r>0.7), indicating that their metabolism is highly coordinated. At the same time, the sterol metabolism module and the ECM module show a unique connection pattern.

[0055] Comparative analysis: The chromatogram of this TNBC sample was compared with a known luminal type A chromatogram database. Luminal type A typically shows moderate phospholipid enrichment and a clearly modular correlation matrix, lacking strongly integrated superclusters. Based on these differential characteristic combinations, the sample can be identified as belonging to the TNBC subtype. This result is consistent with traditional IHC-based typing results (ER-, PR-, HER2-), but this invention provides additional information on metabolic function.

[0056] Example 2: In this example, the above method is integrated into an automated analysis system, specifically including:

[0057] System Architecture: The system software platform adopts a client-server architecture. The server-side deploys algorithm packages that perform data preprocessing, demixing, and correlation matrix calculations. The client serves as the user interface for uploading data, setting parameters, and visualizing results.

[0058] Model Training: A Raman imaging dataset of breast cancer containing at least 100 known subtypes (healthy, luminal A, HER2+, TNBC) was collected as the training set. For each case, in addition to extracting the features described in Example 1 (such as the average intensity of various lipids, spatial heterogeneity index, global network properties of the correlation matrix, etc.), deep features were also extracted directly from the lipid distribution map using a deep learning convolutional neural network (CNN). These features, along with the pathological subtype labels, were fed into a support vector machine (SVM) or random forest classifier for training, and the model parameters were optimized through cross-validation.

[0059] Application: When a new, unknown sample is obtained, the system automatically executes steps S1-S5 of this method to generate a feature vector. Then, it calls the trained classification model to directly output the probability prediction of the sample belonging to each subtype. The system can also generate a visualization report, highlighting the most discriminative lipid distribution map and correlation matrix module to assist pathologists in final diagnosis and biological interpretation.

[0060] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. A method for analyzing metabolic heterogeneity in breast cancer based on label-free hyperspectral Raman imaging, characterized in that, include: S1: Obtain fresh frozen sections of breast cancer tissue; S2: A confocal Raman imaging system is used to scan tissue sections, and the Raman spectrum of each pixel is acquired to form a hyperspectral data cube; S3: Perform spectral preprocessing on the hyperspectral data cube; S4: Using a non-negative least squares regression algorithm, the spectrum of each preprocessed pixel is decomposed into a linear combination of reference spectra of multiple biochemical components to obtain a quantitative abundance distribution map of each biochemical component in the tissue slice space. S5: Based on the quantitative abundance distribution map of each biochemical component, calculate the spatial Pearson correlation coefficient between all pairs of components and construct the correlation matrix; S6: Based on the features extracted from the quantitative abundance distribution map and the correlation matrix, distinguish healthy tissue from cancerous tissue and identify the molecular subtype of breast cancer.

2. The method for analyzing metabolic heterogeneity in breast cancer based on label-free hyperspectral Raman imaging according to claim 1, characterized in that, In step S2, the confocal Raman imaging system is equipped with a 532nm or 785nm excitation laser, and the imaging scan is either a point-by-point scan or a line scan, with a scanned field of view area not less than [amount missing]. And the spatial resolution is better than 10 The Raman spectral range covers 50 Up to 3800 .

3. The method for analyzing metabolic heterogeneity in breast cancer based on label-free hyperspectral Raman imaging according to claim 2, characterized in that, In step S4, the reference spectra of the various biochemical components are obtained from a pre-established reference spectral library, which contains standard spectra obtained by measuring pure substance standards. The various known biochemical components include various lipid molecules and various extracellular matrix proteins.

4. The method for analyzing metabolic heterogeneity in breast cancer based on label-free hyperspectral Raman imaging according to claim 3, characterized in that, The identification criteria for the molecular subtype are as follows: Triple-negative breast cancer: lipid enrichment, spatial heterogeneity exceeding a preset first threshold, and lipid-lipid correlation based on the correlation matrix exceeding a preset second threshold supercluster structure; HER2 positive: Phospholipids and sterols showed focal enrichment in the tumor nest region, and the lipid-sterol network based on the correlation matrix was within the preset intermediate range; Type A lumen: Phospholipids are distributed in an ordered spatial manner, and the spatial correlation between metabolic components and interstitial components is lower than the preset third threshold.

Images