Intelligent screening method for slow-release drug carriers based on material microscopic image recognition

Abstract

The application discloses an intelligent sustained-release drug carrier screening method based on material microscopic image recognition, integrates nanometer hyperspectral, cryo-EM and Raman spectral data through a multi-modal data synchronous acquisition system, realizes nanoscale spatial alignment of microscopic morphology and component distribution, adopts a non-local mean denoising algorithm to enhance pore structure identification, combines adaptive baseline correction and multi-peak fitting technology, and quantifies chemical bond state changes, constructs a characteristic spectral library, generates a chemical distribution map through spectral angle mapping, combines spatial mutual information entropy to analyze the co-localization of carriers and drugs, innovatively develops a space-spectrum joint encoder, fuses multi-modal features by using a cross attention mechanism, and outputs quantitative indexes such as pore filling rate, crystallinity change and interaction stability score. The method breaks through the static and single-modal limitation of traditional methods, and establishes a complete screening system from microscopic structure analysis to release kinetics prediction.

Description

Technical Field

[0001] This invention relates to a method for screening sustained-release drug carriers, and more particularly to an intelligent method for screening sustained-release drug carriers based on microscopic image recognition of materials. Background Technology

[0002] The purpose of screening sustained-release drug carriers is to achieve targeted drug delivery by controlling the drug release rate, prolonging the duration of drug action and reducing the frequency of administration. Traditional carrier materials often suffer from problems such as uneven pore structure and weak interfacial bonding, leading to uncontrollable burst release or release cycles, directly impacting therapeutic efficacy and patient compliance. In clinical applications, there is an urgent need to establish quantitative evaluation indicators to achieve a high-precision match between carrier performance and drug release kinetics.

[0003] Existing screening methods mainly rely on offline characterization techniques, including scanning electron microscopy to observe carrier morphology, laser particle size analyzer to analyze particle size distribution, and ultraviolet spectrophotometer to detect in vitro release curves. These methods have the following problems: single-modal data cannot fully reflect the correlation between the carrier's microstructure and chemical composition; their static characterization cannot capture the dynamic changes during drug release; and manual feature extraction relies on empirical parameters, resulting in poor reproducibility of screening results. Summary of the Invention

[0004] This invention overcomes the shortcomings of the prior art and provides an intelligent method for screening sustained-release drug carriers based on material microscopic image recognition.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: an intelligent sustained-release drug carrier screening method based on material microscopic image recognition, comprising the following steps:

[0006] S1. Construct a multimodal data synchronous acquisition system, which integrates a nano-hyperspectral microscopy system, a cryo-electron microscope, and a confocal Raman spectrometer; drug carrier samples are prepared by liquid nitrogen quenching at -196℃ to -80℃, and nanoscale spatial alignment of multimodal data is achieved through a time synchronization trigger;

[0007] S2. Apply nonlocal mean denoising algorithm to enhance boundary recognition of drug carrier samples, extract structural features of pores and interfaces through threshold segmentation, and generate a binary structure mask with spatial constraints.

[0008] S3. Baseline correction and multi-peak fitting are used to extract characteristic peaks and carrier component peaks, calculate the redshift of peak position and the change of half peak width, and quantify the chemical bond state.

[0009] S4. Construct a characteristic spectral library, generate a chemical distribution map through spectral angle mapping, and calculate the spatial mutual information entropy between the carrier and the drug molecule to assess their colocalization in spatial distribution.

[0010] S5. Develop a multimodal joint encoder consisting of a spatial attention module and a spectral attention module, and take the data obtained in steps S2 to S4 as input; the output includes a correlation matrix of results containing pore filling rate, crystallinity change index and interaction stability score.

[0011] In a preferred embodiment of the present invention, in step S1, the multimodal data synchronous acquisition system is constructed based on the principle of three-dimensional coordinate mapping, and nanometer-level spatial coordinate calibration is achieved through feedback from the grating ruler of the displacement stage. The spatial alignment error satisfies: ;in For coordinate readings of a nano-hyperspectral microscopy system; The coordinate readings are from the cryo-electron microscope; and the electron beam energy of the cryo-electron microscope and the laser wavelength of the confocal Raman spectrometer satisfy an energy complementarity relationship: Where E is the electron beam energy; λ is the laser wavelength; L is Planck's constant; c is the speed of light; k is the energy matching coefficient of 0.8~1.2.

[0012] In a preferred embodiment of the present invention, in step S1, the step temperature gradient of the liquid nitrogen quenching method satisfies: ;in, The cooling rate; The thermal conductivity coefficient; T represents the liquid nitrogen temperature; T represents the real-time sample temperature; by controlling... The value maintains the cooling rate in the range of -196℃ to -80℃ at .

[0013] In a preferred embodiment of the present invention, in step S2, the weight coefficients of the nonlocal means denoising algorithm are calculated as follows:

[0014] ;

[0015] In the formula, These are the weight coefficients for the nonlocal mean denoising algorithm; These are the grayscale values ​​of pixels i and j, respectively; is the gradient vector of the corresponding pixel; h is the gray-level similarity parameter; g is the structural similarity parameter.

[0016] In a preferred embodiment of the present invention, in step S3, the baseline correction employs an adaptive iterative reweighted penalized least squares method, and the iterative update formula for the baseline is:

[0017] ;

[0018] In the formula, The baseline after iteration; The original spectrum; The baseline before iteration; D is the difference matrix; is the regularization parameter; p is the penalty factor, with a value range of 1≤p≤2;

[0019] The multi-peak fitting uses a Gaussian-Lorentz mixture function:

[0020] ;

[0021] In the formula, , These are the amplitudes of the Lorentz and Gaussian components, respectively; Characteristic peak position; It is half the peak width.

[0022] In a preferred embodiment of the present invention, in step S4, the angle of the spectral angle mapping is calculated as follows:

[0023] ;

[0024] In the formula, The angle mapped from the spectral angle; The sample spectral vector; The standard vector is from the spectral library; Indicates inner product operation; when When the value is ≤0.2 radians, it is considered a feature match, and the confidence level for generating a chemical distribution map is determined. ;in This is the maximum threshold.

[0025] In a preferred embodiment of the present invention, in step S4, the formula for calculating the spatial mutual information entropy is:

[0026] ;

[0027] In the formula, H(C) is the information entropy of the spatial distribution of the carrier; H(D) is the information entropy of the spatial distribution; H(C,D) is the joint information entropy; and MI is the spatial mutual information entropy. The larger the spatial mutual information entropy value, the stronger the co-localization of the drug carrier and the drug molecule in spatial distribution, that is, the higher the degree of spatial overlap between the two.

[0028] In a preferred embodiment of the present invention, in step S5, the weight vector of the spatial attention module is calculated as follows:

[0029] ;

[0030] In the formula, For the spatial attention module, the weight vector is used. Use the Sigmoid activation function; This is a 3×3 convolution operation; It is a spatial feature map vector;

[0031] The weight vector of the spectral attention module is calculated as follows:

[0032] ;

[0033] In the formula, This represents the weight vector of the spectral attention module; This is a 1D convolution operation; It is a spectral eigenvector;

[0034] Multimodal feature fusion employs a cross-attention mechanism: ;

[0035] In the formula, This represents the total weight vector after multimodal feature fusion.

[0036] In a preferred embodiment of the present invention, in step S5, the interaction stability score is calculated as follows:

[0037] ;

[0038] In the formula, Assess the stability of the interaction; Pore ​​filling rate; As an indicator of changes in crystallinity; This represents the fluctuation value of spatial mutual information entropy; , , The weighting coefficients and .

[0039] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0040] (1) The algorithm automatically filters image noise while preserving the tiny pores and boundary features on the surface of the drug carrier, thereby improving the accuracy of pore identification and avoiding misjudgment caused by noise interference.

[0041] (2) By combining hyperspectral, electron microscopy and Raman spectroscopy, the spatial positions of images captured by different devices are completely matched, thus protecting the sample from damage and ensuring the clarity of chemical signals, thereby significantly improving the screening efficiency.

[0042] (3) By rapidly cooling down, the growth of crystals and phase separation are suppressed, the transient structure is frozen, the porosity retention rate is improved, the structural collapse caused by the traditional drying method is avoided, the component segregation is reduced, and the spectral analysis is ensured to reflect the true drug loading state. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart of a preferred embodiment of the present invention. Detailed Implementation

[0045] 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.

[0046] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0047] It should be noted in advance that the core reason for using material microscopic image recognition technology for intelligent screening of sustained-release drug carriers is that it can systematically solve the inherent defects of traditional methods and construct a complete analytical system from microscopic mechanisms to macroscopic performance. Traditional screening methods suffer from problems such as insufficient correlation of single-modal data, inability of static characterization to capture dynamic changes, and poor reproducibility due to dependence on artificial parameters. These limitations keep the carrier performance evaluation at the level of surface characteristics, making it difficult to establish a quantitative relationship between structure, composition, and release behavior.

[0048] Material microscopic image recognition technology, by integrating a nanoscale hyperspectral microscopy system, cryo-electron microscopy, and confocal Raman spectroscopy, achieves nanoscale spatial alignment of microscopic morphology and chemical composition. This multimodal synchronous acquisition capability correlates pore structure parameters (such as porosity and particle size distribution) with changes in chemical bond states (such as peak redshift and half-maximum width adjustment), overcoming the dilemma of separating morphology analysis and composition detection in traditional methods. Combined with a nonlocal mean denoising algorithm and adaptive baseline correction technology, it can accurately extract pore features in complex backgrounds, providing a highly reliable data foundation for structural analysis.

[0049] The innovation of this technology lies in its dynamic quantification capabilities and the construction of an intelligent prediction system. By real-time monitoring of key indicators such as carrier pore filling rate and crystallinity changes during drug release, and combining spectral angle mapping and spatial mutual information entropy, three-dimensional visualization and quantitative scoring of drug-carrier interactions are achieved. The spatial-spectral joint encoder uses a cross-attention mechanism to fuse multimodal features, outputting an interaction stability score and a release curve prediction over several hours, compressing the screening cycle from weeks in traditional methods to hours.

[0050] This technology not only establishes a direct link between the microstructure of the carrier and its release kinetics, but also provides a data-driven optimization scheme for personalized carrier design through a quantitative index system (covering morphological, chemical, and kinetic parameters). Its core value lies in transforming the experience-dependent screening process into a repeatable and predictable intelligent decision-making process, laying the technological foundation for the precise development of sustained-release drug carriers.

[0051] Specifically, such as Figure 1 As shown, the intelligent sustained-release drug carrier screening method based on material microscopic image recognition includes the following steps:

[0052] S1. Construct a multimodal data synchronous acquisition system, which integrates a nano-hyperspectral microscopy system, a cryo-electron microscope, and a confocal Raman spectrometer; drug carrier samples are prepared in an extremely cold environment by liquid nitrogen quenching, and nanoscale spatial alignment of multimodal data is achieved through a time synchronization trigger.

[0053] In step S1, the multimodal data synchronous acquisition system is built based on the principle of three-dimensional coordinate mapping. Nanoscale spatial coordinate calibration is achieved through feedback from the grating ruler of the displacement stage, and the spatial alignment error satisfies: ;in For coordinate readings of a nano-hyperspectral microscopy system; The coordinate readings are from the cryo-electron microscope; and the electron beam energy of the cryo-electron microscope and the laser wavelength of the confocal Raman spectrometer satisfy an energy complementarity relationship: Where E is the electron beam energy; Where λ is the laser wavelength; L is Planck's constant; c is the speed of light; and k is the energy matching coefficient ranging from 0.8 to 1.2. Based on the photon energy formula in quantum mechanics and considering the requirements of multimodal imaging, a dynamic balance between electron beam and laser energy is achieved by introducing a matching coefficient k. This ensures that cryo-electron microscopy (electron beam imaging) and confocal Raman spectroscopy (laser imaging) are energy complementary when spatially aligned, thus improving the accuracy of multimodal data fusion.

[0054] Furthermore, the S1 multimodal data synchronous acquisition system employs a timing alignment scheme combining hard and soft synchronization. Hard synchronization generates trigger pulses via a square wave signal generator, which are transmitted to the GPIO interfaces of each device through a differential signal circuit, ensuring that the coordinate calibration error between the nanoscale hyperspectral microscopy system and the cryo-electron microscope is ≤5nm. Soft synchronization uses a timestamp alignment algorithm to record the data acquisition time of each device and map it to a unified time axis, controlling the timing deviation within 0.1μs. Device extrinsic parameter calibration uses a checkerboard target, and the three-dimensional coordinate transformation matrix is ​​obtained through the Zhang Zhengyou calibration method, achieving spatial fusion of multimodal data.

[0055] The step temperature gradient of liquid nitrogen quenching satisfies: ;in, The cooling rate; The thermal conductivity coefficient; T represents the liquid nitrogen temperature; T represents the real-time sample temperature; by controlling... The value maintains the cooling rate in the range of -196℃ to -80℃ at .

[0056] To further clarify, the liquid nitrogen quenching method employs a stepped cooling strategy: initially, the temperature is lowered from room temperature to -20°C at a rate of 10°C / min, maintained for 30 seconds, then switched to a rate of 0.5°C / min to slowly pass through the glass transition zone, ultimately maintaining the temperature at -150°C. During the cooling process, a PID controller adjusts the liquid nitrogen injection rate in real time, achieving closed-loop control through thermocouple feedback to ensure sample temperature fluctuations are ≤1°C. For special carrier materials, ampoules can be pre-cooled by immersing them in an -80°C ethanol bath before being transferred to the liquid nitrogen environment, increasing the cooling rate to 20°C / min.

[0057] S2. Apply nonlocal mean denoising algorithm to enhance boundary recognition of drug carrier samples. Extract structural features of pores and interfaces through threshold segmentation to generate a binary structure mask with spatial constraints.

[0058] In step S2, the weight coefficients of the nonlocal means denoising algorithm are calculated as follows:

[0059] ;

[0060] In the formula, These are the weight coefficients for the nonlocal mean denoising algorithm; These are the grayscale values ​​of pixels i and j, respectively; is the gradient vector of the corresponding pixel; h is the gray-level similarity parameter; g is the structural similarity parameter. Combining gray-level similarity (h parameter) and structural similarity (g parameter), the correlation strength between pixels is quantified through an exponential decay model. This effectively suppresses noise while preserving edge features, improving the signal-to-noise ratio for pore identification and providing a clear structural mask for subsequent spectral analysis.

[0061] Baseline correction employs an adaptive iterative reweighted penalized least squares method, and the iterative update formula for the baseline is:

[0062] ;

[0063] In the formula, The baseline after iteration; The original spectrum; The baseline before iteration; D is the difference matrix; is the regularization parameter; p is the penalty factor, with a value range of 1≤p≤2;

[0064] Multi-peak fitting uses a Gaussian-Lorentz mixture function:

[0065] ;

[0066] In the formula, , These are the amplitudes of the Lorentz and Gaussian components, respectively; Characteristic peak position; It is half the peak width.

[0067] Furthermore, the nonlocal means denoising algorithm employs an adaptive parameter adjustment strategy. Initially, h is set to 3. g=2 ( The structural similarity parameters are automatically adjusted by calculating the pixel gradient distribution (where the noise standard deviation is used). For periodic pore structures, a Fast Fourier Transform is introduced to accelerate the similarity search, reducing the computational complexity from O(n²) to O(nlog n). In actual processing, a sliding window mechanism is used, with the window size set to twice the average diameter of the pores to ensure feature integrity.

[0068] S3. Baseline correction and multi-peak fitting are used to extract characteristic peaks and carrier component peaks, calculate the redshift of peak position and the change of half peak width, and quantify the chemical bond state.

[0069] Furthermore, baseline correction employs an improved airPLS algorithm. Through a dynamic weight update mechanism, the weights are automatically reset to zero when a peak region is detected, effectively suppressing peak tailing. For complex spectra, wavelet transform is introduced for multi-scale decomposition. Baseline estimation is performed at the coarse-scale level, while characteristic peak information is preserved at the fine-scale level. Finally, the corrected spectrum is obtained through wavelet reconstruction.

[0070] S4. Construct a feature spectral library, generate a chemical distribution map through spectral angle mapping, and calculate the spatial mutual information entropy between the carrier and drug molecules to assess their co-localization in spatial distribution. During spectral angle mapping, a spectral library containing 5000 standard substances is established, and a KD tree structure is used to accelerate nearest neighbor search. After the chemical distribution map is generated, a spatial constraint is introduced: when the spectral angle difference between adjacent pixels exceeds 15°, a region growing algorithm is activated for correction. The spatial mutual information entropy calculation is implemented in parallel, dividing the carrier region into a 4×4 grid. Information entropy is calculated independently for each grid, and then weighted and fused. The weight coefficients are determined by the pore density.

[0071] In step S4, the angle of the spectral angle mapping is calculated as follows:

[0072] ;

[0073] In the formula, The angle mapped from the spectral angle; The sample spectral vector; The standard vector is from the spectral library; Indicates inner product operation; when When the value is ≤0.2 radians, it is considered a feature match, and the confidence level for generating a chemical distribution map is determined. ;in This is the maximum threshold.

[0074] The formula for calculating spatial mutual information entropy is:

[0075] ;

[0076] In the formula, H(C) is the information entropy of the spatial distribution of the carrier; H(D) is the information entropy of the spatial distribution; H(C,D) is the joint information entropy; the larger the MI value, the stronger the spatial colocation of the carrier and the information entropy.

[0077] S5. Develop a multimodal joint encoder consisting of a spatial attention module and a spectral attention module, and take the data obtained in steps S2 to S4 as input; the output includes a correlation matrix of results containing pore filling rate, crystallinity change index and interaction stability score.

[0078] In step S5, the weight vector of the spatial attention module is calculated as follows:

[0079] ;

[0080] In the formula, For the spatial attention module, the weight vector is used. Use the Sigmoid activation function; This is a 3×3 convolution operation; It is a spatial feature map vector;

[0081] The weight vector of the spectral attention module is calculated as follows:

[0082] ;

[0083] In the formula, This represents the weight vector of the spectral attention module; This is a 1D convolution operation; It is a spectral eigenvector;

[0084] Multimodal feature fusion employs a cross-attention mechanism: ;

[0085] In the formula, This represents the total weight vector after multimodal feature fusion.

[0086] In step S5, the interaction stability score is calculated as follows:

[0087] ;

[0088] In the formula, Assess the stability of the interaction; Pore ​​filling rate; As an indicator of changes in crystallinity; This represents the fluctuation value of spatial mutual information entropy; , , The weighting coefficients and .

[0089] During the training of the multimodal co-encoder, a progressive learning strategy is adopted: the first 20 epochs train only the spatial attention module, the middle 30 epochs train the spectral module with fixed spatial module parameters, and the last 50 epochs perform joint fine-tuning. A dynamic weight adjustment mechanism is introduced to automatically adjust the weights based on the current epoch number. , , Coefficient: Initial stage =0.7、 =0.2、 =0.1, later gradually adjusted to =0.5、 =0.3、 =0.2. The training data employs an online augmentation strategy, randomly adding Gaussian noise to the spectral data and applying elastic deformation to the image data, effectively improving the model's robustness.

[0090] It should be further explained that in the multimodal data acquisition stage, the system of this invention achieves high-speed data transmission between the nano-hyperspectral microscopy system, cryo-electron microscope, and confocal Raman spectrometer via a USB 3.0 interface. The synchronous triggering mechanism adopts a dual-modal design combining hardware and software: at the hardware level, a 1kHz square wave signal generator produces trigger pulses, which are transmitted to the GPIO interfaces of each device via a differential signal circuit, ensuring that the coordinate calibration error between the nano-hyperspectrology and cryo-electron microscope is controlled within 5nm; at the software level, time calibration is performed via the NTP protocol, mapping the acquisition time of each device to a unified time axis, with timing deviation strictly controlled within 0.1μs. The device extrinsic parameter calibration uses a checkerboard target, and the three-dimensional coordinate transformation matrix is ​​obtained through the Zhang Zhengyou calibration method, achieving spatial fusion of multimodal data.

[0091] In the image preprocessing stage, an improved nonlocal mean denoising algorithm is used. The values ​​of the gray-level similarity parameter h and the structural similarity parameter g in the weighting coefficient formula are dynamically adjusted according to the carrier material: for porous materials, h=3. g=2 To preserve detailed features, the polymer carrier uses h=2. g=1 To avoid boundary ambiguity, the parameter adaptive mechanism dynamically adjusts the parameters through sample variance calculation. The baseline correction process employs an adaptive iterative reweighted penalized least squares method with dual convergence criteria: automatic termination when the residual change rate is below 0.1% for five consecutive iterations, or forced termination when the maximum number of iterations (50) is reached. Simultaneously, a region growth algorithm is initiated through peak region detection to correct the weights.

[0092] The spectral analysis module constructs a spectral library containing 5000 standard substances. The selection of standard substances follows the principles of purity ≥99% and spectral resolution ≤1nm, covering common carrier materials such as PLGA and MSN, and drug components such as DOX and PTX. After generating chemical distribution maps through spectral angle mapping, the confidence level c is verified using TEM-EDS pixel-level comparison and NIST standard spectral library comparison. The threshold is determined through ROC curve analysis. The correlation between spatial mutual information entropy calculation results and carrier performance is experimentally defined with thresholds: MI > 0.8 indicates high colocalization, while MI < 0.5 requires adjustment of the carrier structure.

[0093] The joint encoder training employs a progressive strategy. The spatial attention module generates feature map weights through 3×3 convolution followed by a ReLU activation function, while the spectral attention module extracts deep features using two layers of 1D convolution. Data augmentation employs the CutMix method, randomly selecting 30% of the region to mix two samples. The mixing ratio λ is generated using a Beta distribution, and the boundaries are Gaussian blurred. A step-by-step cooling strategy based on material properties is implemented using liquid nitrogen quenching. A PID controller adjusts the liquid nitrogen injection volume in real time via thermocouple feedback, ensuring that temperature fluctuations in the metal-organic framework, polymer carrier, and biological samples are controlled within ±1℃.

[0094] In summary, the multimodal data synchronous acquisition system integrates nanoscale hyperspectral imaging and cryo-electron microscopy to achieve nanoscale spatial alignment of microstructure and composition distribution. Nonlocal mean denoising and adaptive baseline correction algorithms accurately extract chemical bond state changes from dynamic spectra. The spatial-spectral joint encoder establishes a quantitative correlation model between pore filling rate, crystallinity changes, and interaction stability through a cross-attention mechanism. This approach overcomes the static, single-modal limitations of traditional methods, forming a complete screening system from microstructure analysis to release kinetic prediction.

[0095] Based on the preferred embodiments of the present invention described above, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. An intelligent screening method for a sustained-release drug carrier based on material micro-image recognition, characterized in that, Includes the following steps: S1. Construct a multimodal data synchronous acquisition system, which integrates a nano-hyperspectral microscopy system, a cryo-electron microscope, and a confocal Raman spectrometer; drug carrier samples are prepared by liquid nitrogen quenching at -196℃ to -80℃, and nanoscale spatial alignment of multimodal data is achieved through a time synchronization trigger; S2. Apply nonlocal mean denoising algorithm to enhance boundary recognition of drug carrier samples, extract structural features of pores and interfaces through threshold segmentation, and generate a binary structure mask with spatial constraints. S3. Baseline correction and multi-peak fitting are used to extract characteristic peaks and carrier component peaks, calculate the redshift of peak position and the change of half peak width, and quantify the chemical bond state. S4. Construct a characteristic spectral library, generate a chemical distribution map through spectral angle mapping, and calculate the spatial mutual information entropy between the carrier and the drug molecule to assess their colocalization in spatial distribution. S5. Develop a multimodal joint encoder consisting of a spatial attention module and a spectral attention module, and take the data obtained in steps S2 to S4 as input; output a correlation matrix containing pore filling rate, crystallinity change index and interaction stability score. 2.The method of claim 1, wherein the method is characterized by: In step S1, the multimodal data synchronous acquisition system is built based on the principle of three-dimensional coordinate mapping. Nanoscale spatial coordinate calibration is achieved through feedback from the grating ruler of the displacement stage, and the spatial alignment error satisfies: ;in For coordinate readings of a nano-hyperspectral microscopy system; The coordinate readings are from the cryo-electron microscope; and the electron beam energy of the cryo-electron microscope and the laser wavelength of the confocal Raman spectrometer satisfy an energy complementarity relationship: Where E is the electron beam energy; λ is the laser wavelength; L is Planck's constant; c is the speed of light; k is the energy matching coefficient of 0.8~1.

2. 3.The method of claim 1, wherein the method is characterized by: In step S1, the step temperature gradient of the liquid nitrogen quenching method satisfies: ;in, The cooling rate; The thermal conductivity coefficient; T represents the liquid nitrogen temperature; T represents the real-time sample temperature; by controlling... The value maintains the cooling rate in the range of -196℃ to -80℃ at . 4.The method of claim 1, wherein the method is characterized by: In step S2, the weight coefficients of the nonlocal means denoising algorithm are calculated as follows: ； In the formula, These are the weight coefficients for the nonlocal mean denoising algorithm; These are the grayscale values ​​of pixels i and j, respectively; is the gradient vector of the corresponding pixel; h is the gray-level similarity parameter; g is the structural similarity parameter.

5. The intelligent screening method of the material micro-image-based intelligent sustained-release drug carrier according to claim 1, characterized in that: In step S3, the baseline correction employs an adaptive iterative reweighted penalized least squares method, and the iterative update formula for the baseline is: ； wherein is the baseline after iteration; is the original spectrum; is the baseline before iteration; D is a difference matrix; is a regularization parameter; p is a penalty factor, with a value range of 1≤p≤2; The multi-peak fitting uses a Gaussian-Lorentz mixture function: ； In the formula, , These are the amplitudes of the Lorentz and Gaussian components, respectively; Characteristic peak position; It is half the peak width.

6. The intelligent screening method of the material micro-image-based intelligent sustained-release drug carrier according to claim 1, characterized in that: In step S4, the angle of the spectral angle mapping is calculated as follows: ； In the formula, The angle mapped from the spectral angle; The sample spectral vector; The standard vector is from the spectral library; Indicates inner product operation; when When the value is ≤0.2 radians, it is considered a feature match, and the confidence level for generating a chemical distribution map is determined. ;in This is the maximum threshold.

7. The intelligent screening method of the material micro-image-based intelligent sustained-release drug carrier according to claim 1, characterized in that: In step S4, the formula for calculating the spatial mutual information entropy is: ； In the formula, H(C) is the information entropy of the spatial distribution of the carrier; H(D) is the information entropy of the spatial distribution; H(C,D) is the joint information entropy; and MI is the spatial mutual information entropy. 8.The method of claim 1, wherein the method is characterized by: In step S5, the weight vector of the spatial attention module is calculated as follows: ； In the formula, For the spatial attention module, the weight vector is used. Use the Sigmoid activation function; This is a 3×3 convolution operation; It is a spatial feature map vector; The weight vector of the spectral attention module is calculated as follows: ； wherein is a weight vector of the spectral attention module; is a 1D convolution operation; is a spectral feature vector; The multi-modal feature fusion adopts a cross attention mechanism: ; In the formula, is the total value of the weight vector after multi-modal feature fusion. 9.The intelligent screening method of the material micro-image-based intelligent sustained-release drug carrier according to claim 1, characterized in that: In step S5, the interaction stability score is calculated as follows: ； In the formula, Assess the stability of the interaction; Pore ​​filling rate; As an indicator of changes in crystallinity; This represents the fluctuation value of spatial mutual information entropy; , , The weighting coefficients and .

Images