A multi-modal spatial omics data fusion method, device and system

Abstract

The present application relates to the technical field of data fusion processing, and especially relates to a multi-modal spatial omics data fusion method, device and system, the multi-modal spatial omics data fusion method, device and system provided by the present application, the method comprises: performing standardization format conversion on first platform protein imaging data to generate a standardized multi-channel image containing channel priority dimensions and pixel physical size information.The method provided by the present application solves the multi-platform data format compatibility problem through standardization conversion; the robustness and precision of cross-platform image spatial alignment are improved by using a hybrid registration strategy; efficient extraction of large-scale single-cell protein features is realized through parallel block calculation; finally, the protein features are automatically associated with transcriptome data at the single-cell level to generate structured fusion data, providing a reliable foundation for spatial multi-omics joint analysis.

Description

Technical Field

[0001] This invention relates to the field of data fusion processing technology, and in particular to a method, apparatus and system for fusion of multimodal spatial omics data. Background Technology

[0002] Spatial omics, by analyzing the molecular expression profiles of tissue samples while preserving their spatial location, has become an important tool in life science research. Currently, multiple technology platforms coexist; for example, 10xGenomics' Xenium platform provides high-resolution spatial transcriptome data, while AkoyaBiosciences' PhenoCycler-Fusion (PCF) platform enables multiplex protein imaging. To gain a more comprehensive understanding of cellular state and function, researchers increasingly need to integrate and correlate heterogeneous data from different platforms (such as transcriptomics and proteomics) at the single-cell level. This falls under the technical field of multimodal spatial omics data fusion.

[0003] However, existing technologies face a series of challenges in achieving such multi-platform data fusion. First, in the data preprocessing stage, there are compatibility issues between image formats output from different platforms (such as QPTIFF for PCF) and standard analysis workflows (such as OME-TIFF). Existing conversion tools often require extensive manual adjustments to handle multi-dimensional channels and maintain metadata consistency, easily leading to channel information mismatch. Second, in the core image registration stage, conventional methods (such as registration based on mutual information or single feature points) are insufficiently robust to cross-platform images with multiple channels, low feature contrast, or scale rotation differences, easily leading to registration failures or significant deviations, directly affecting the accuracy of subsequent cell-level correspondence. Finally, in the data extraction and fusion stage, performing pixel statistical calculations on tens of thousands of cell regions in high-resolution full-frame images presents bottlenecks in traditional serial or simple parallel methods due to high memory consumption and low computational efficiency, making it difficult to meet the demands of large-scale data processing.

[0004] Therefore, there is an urgent need for a solution that can automate, be highly robust, and be efficient in completing the process from format standardization and accurate spatial registration to single-cell statistical fusion, so as to achieve reliable integration of cross-platform spatial omics data at the single-cell level. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method, apparatus and system for fusing multimodal spatial omics data.

[0006] In a first aspect, embodiments of the present invention provide a method for fusing multimodal spatial omics data, the method comprising:

[0007] The protein imaging data from the first platform is standardized and converted to generate a standardized multi-channel image that includes channel priority dimension and pixel physical size information.

[0008] Selected channel images from the standardized multi-channel image are image registered with a spatial reference image from the second platform to obtain a spatially aligned multi-channel registered image.

[0009] Based on cell segmentation information from the second platform, the pixel statistical feature values ​​of each cell region in each image channel are extracted in parallel from the spatially aligned multi-channel registration image.

[0010] Pixel statistical feature values ​​are correlated with cell-identified transcriptome data from a second platform to generate fused spatial single-cell data.

[0011] In conjunction with the first aspect, the steps of standardizing and converting protein imaging data from the first platform to generate a standardized multichannel image containing channel priority dimension and pixel physical size information include:

[0012] Obtain the input path of the protein imaging data to be converted, the output path of the target normalized multi-channel image, and the optional parameters;

[0013] Based on the image sequence index and resolution level index in the optional parameters, the corresponding image data is read from the protein imaging data;

[0014] Parse the axis order information of the image data and rearrange the axis order of the image data to unify it into a channel-first standardized order;

[0015] Extract pixel resolution parameters from the tag information of protein imaging data and convert them into pixel size information in standard physical units;

[0016] Determine the channel name information corresponding to the number of channels in the image data;

[0017] Based on the rearranged image data, pixel size information, and channel name information, a standard format multichannel image is generated.

[0018] In conjunction with the first aspect, the steps of image registration between a selected channel image from the standardized multi-channel image and a spatial reference image from the second platform to obtain the geometric transformation model include:

[0019] Use the selected channel image as the moving image and the spatial reference image as the fixed image;

[0020] Downsampling and normalization preprocessing are performed on both moving and stationary images;

[0021] The scale-invariant feature transform algorithm is used to perform feature point detection, descriptor calculation and matching on the preprocessed image;

[0022] Based on the random sampling consensus algorithm, the initial geometric transformation model from the moving image to the fixed image is estimated from the matched feature points;

[0023] Based on the initial geometric transformation model, the final geometric transformation model for mapping the normalized multi-channel image to the spatial reference image coordinate space is obtained;

[0024] Using the final geometric transformation model, the images of each channel of the standardized multi-channel image are mapped to the coordinate space of the spatial reference image to obtain a spatially aligned multi-channel registration image.

[0025] In conjunction with the first aspect, the steps for estimating the initial geometric transformation model from the moving image to the fixed image from the matched feature points, based on the random sample consensus algorithm, include:

[0026] The confidence level of the matched feature point pairs is used to filter out mismatched point pairs.

[0027] Based on the downsampling scaling factor, scale recovery is performed on the coordinates of the matched feature point pairs, mapping them back from the downsampled image coordinate system to the original image coordinate system;

[0028] Based on the feature matching point pairs after scale recovery, iterative estimation is performed using the random sampling consensus algorithm, prioritizing the estimation of the similarity transformation model;

[0029] If the similarity transformation model meets the preset validity conditions, it will be used as the initial geometric transformation model.

[0030] If the similarity transformation model does not meet the validity condition, then we fall back to estimating the affine transformation model and use the successfully estimated affine transformation model as the initial geometric transformation model.

[0031] In conjunction with the first aspect, the steps for obtaining the final geometric transformation model for mapping a normalized multi-channel image to a spatial reference image coordinate space, based on the initial geometric transformation model, include:

[0032] If fine registration optimization is not enabled, the initial geometric transformation model will be used as the final geometric transformation model.

[0033] If fine registration optimization is enabled, the initial geometric transformation model is converted into an affine transformation matrix as the starting point for optimization. Based on this starting point, the enhanced correlation coefficient method is used for iterative optimization on the constructed low-resolution image pairs to determine the final geometric transformation model.

[0034] In conjunction with the first aspect, the steps of mapping each channel of the normalized multi-channel image to the coordinate space of the spatial reference image using the final geometric transformation model to obtain a spatially aligned multi-channel registered image include:

[0035] Obtain the number of channels and spatial dimensions of each channel of the standardized multichannel image, as well as the spatial dimensions of the spatial reference image;

[0036] Based on the spatial dimensions of the spatial reference image, initialize the registered image space to accommodate the mapping results;

[0037] Using the final geometric transformation model, the image data of each channel of the standardized multi-channel image is resampled and mapped from the coordinate space of the standardized multi-channel image to the registered image space.

[0038] The mapped image data of each channel are combined to form a spatially aligned multi-channel registered image that is consistent with the spatial reference image space.

[0039] Combining the first aspect, based on cell segmentation information from the second platform, the step of extracting pixel statistical feature values ​​of each cell region in each image channel in parallel from spatially aligned multi-channel registered images includes:

[0040] Load cell segmentation information from the second platform and obtain cell tag data;

[0041] Load spatially aligned multi-channel registered images and acquire image data for each channel;

[0042] Establish a spatial mapping relationship between cell label data and image data of each channel;

[0043] Based on spatial mapping relationships, a block-based parallel computing approach is adopted to process the image data of each channel and classify each pixel value into the corresponding cell label according to its spatial coordinates.

[0044] Based on the classification results, the pixel statistical feature values ​​of each cell label in each image channel are calculated. The pixel statistical feature values ​​include the mean and median.

[0045] In conjunction with the first aspect, the step of associating pixel statistical feature values ​​with cell-identified transcriptome data from the second platform includes:

[0046] The pixel statistical feature values ​​are correlated and matched with the transcriptome data indexed by cell tags provided by the second platform;

[0047] The matched pixel statistical feature values ​​are added to the transcriptome data as proteome data fields to generate spatial single-cell data containing fused transcriptome and proteome data.

[0048] Secondly, this application provides a fusion apparatus for multimodal spatial omics data, the apparatus being used to perform the above-described method, the apparatus comprising:

[0049] The standardization processing module is used to perform standardization format conversion on protein imaging data from the first platform to generate a standardized multi-channel image containing channel priority dimension and pixel physical size information.

[0050] The image registration module is used to register a selected channel image from a standardized multi-channel image with a spatial reference image from a second platform to obtain a spatially aligned multi-channel registered image.

[0051] The feature statistics module is used to extract the pixel statistical feature values ​​of each cell region in each image channel in parallel from the spatially aligned multi-channel registered images based on cell segmentation information from the second platform.

[0052] The data fusion module is used to associate pixel statistical feature values ​​with the transcriptome data of the corresponding cells to generate fused spatial single-cell data.

[0053] Thirdly, this application provides a fusion system for multimodal spatial omics data, including the apparatus described above.

[0054] The embodiments of the present invention bring the following beneficial effects: The present application provides a method, apparatus and system for fusing multimodal spatial omics data. The method includes: performing a standardized format conversion on protein imaging data from a first platform to generate a standardized multichannel image containing channel priority dimension and pixel physical size information; performing image registration on selected channel images from the standardized multichannel image with a spatial reference image from a second platform to obtain a spatially aligned multichannel registered image; extracting pixel statistical feature values ​​of each cell region in each image channel from the spatially aligned multichannel registered image in parallel based on cell segmentation information from the second platform; and associating the pixel statistical feature values ​​with cell-identified transcriptome data from the second platform to generate fused spatial single-cell data.

[0055] The method provided in this application solves the problem of data format compatibility across multiple platforms through standardized conversion; it improves the robustness and accuracy of cross-platform image spatial alignment by adopting a hybrid registration strategy; it achieves efficient extraction of large-scale single-cell protein features through parallel block computing; and finally, it automatically associates protein features with transcriptome data at the single-cell level to generate structured fusion data, providing a reliable foundation for spatial multi-omics joint analysis.

[0056] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0057] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0058] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0059] Figure 1 A flowchart illustrating a method for fusing multimodal spatial omics data provided in an embodiment of the present invention;

[0060] Figure 2 A schematic diagram of a multimodal spatial omics data fusion device provided in an embodiment of the present invention;

[0061] Figure 3 This is an example diagram before registration in an embodiment of the present invention;

[0062] Figure 4 To Figure 3 The registered image obtained by registering the provided example image;

[0063] Figure 5 for Figure 3 and Figure 4 A diagram illustrating the effect after overlay;

[0064] Figure 6 The provided example is a diagram showing the matching effect of the new chip point outline;

[0065] Figure 7 for Figure 6 A magnified schematic diagram of a local area.

[0066] Figure label:

[0067] 10 - Standardization processing module, 20 - Image registration module, 30 - Feature statistics module, 40 - Data fusion module. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions 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, 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.

[0069] To facilitate understanding of this embodiment, the technical terms used in this application will be briefly introduced below.

[0070] PCF (Pheno Cycler-Fusion): A spatial proteomics imaging platform developed by Akoya Biosciences, which outputs QPTIFF format image data containing multiple protein labeling channels.

[0071] Xenium: A spatial transcriptomics in situ analysis platform developed by 10xGenomics, whose output includes gene expression matrices, cell boundary (segmentation) information, and morphological reference images.

[0072] SIFT (Scale-Invariant Feature Transform): A scale-invariant feature transformation algorithm used to extract local feature points and their descriptors from an image that remain stable despite changes in scale, rotation, and brightness.

[0073] RANSAC (Random Sample Consensus): A robust model parameter estimation method that can fit the optimal model from data containing a lot of noise (mismatches).

[0074] ECC (Enhanced Correlation Coefficient): An enhanced correlation coefficient method, a registration optimization algorithm based on image pixel intensity similarity measurement, used to achieve fine alignment at the sub-pixel level.

[0075] Cell tag data: A type of raster data in which the value of each pixel represents a unique cell identifier, used to define the spatial region of each cell in an image.

[0076] After introducing the technical terms used in this application, the application scenarios and design concepts of the embodiments of this application will be briefly described below.

[0077] In existing space omics research, it is difficult to accurately integrate multi-platform data (such as transcriptomics and proteomics) at the single-cell level. Existing methods lack robustness in cross-platform image registration, are prone to errors in data format conversion, and are inefficient in single-cell statistical computation for large-scale images, thus hindering multimodal joint analysis.

[0078] Based on this, embodiments of this application provide a method, apparatus, and system for fusing multimodal spatial omics data.

[0079] Example 1

[0080] This application provides a method for fusing multimodal spatial omics data, combining... Figure 1 As shown, the method includes:

[0081] S110 performs a standardized format conversion on protein imaging data from the first platform to generate a standardized multichannel image that includes channel priority dimension and pixel physical size information.

[0082] S120, the selected channel image in the standardized multi-channel image is image registered with the spatial reference image from the second platform to obtain a spatially aligned multi-channel registered image.

[0083] S130, based on cell segmentation information from the second platform, extracts the pixel statistical feature values ​​of each cell region in each image channel in parallel from the spatially aligned multi-channel registered image.

[0084] S140 associates pixel statistical feature values ​​with cell-identified transcriptome data from a second platform to generate fused spatial single-cell data.

[0085] This application first automatically identifies and unifies image dimensions through data standardization processing in step S110, while extracting information such as pixel physical size, and outputs a standardized multi-channel image. This achieves automatic and reliable conversion from proprietary format to open standard format, solving the problems of inconsistent data input across multiple platforms and reliance on manual adjustments, and enhancing the repeatability and traceability of the process. Next, in step S120, a hybrid strategy combining feature matching and intensity optimization is employed to improve registration success rate and spatial alignment accuracy. Then, all channels of the standardized multi-channel image are uniformly mapped to the coordinate space of the reference image, generating a spatially aligned multi-channel registration image to ensure the accuracy of protein information. The cell number is accurately located. Then, in step S130, a block-parallel mechanism is introduced to efficiently traverse and statistically analyze the spatially aligned multi-channel registered images, extracting the pixel statistical features of each cell in each protein channel to reduce memory requirements and improve processing speed, making batch sample analysis feasible and enhancing the practicality and scalability of the method. Finally, in step S140, the extracted cell-level protein statistical features are associated and matched with the transcriptome data provided by the second platform according to cell identifiers to generate a unified spatial single-cell multimodal dataset, realizing the parallel integration of transcriptome and proteome information in the same cell, and providing a directly usable data structure for downstream joint analysis.

[0086] In conjunction with the first aspect, step S110 includes:

[0087] S111: Obtain the input path of the protein imaging data to be converted, the output path of the target normalized multi-channel image, and optional parameters.

[0088] S112, read the corresponding image data from the protein imaging data according to the image sequence index and resolution level index in the optional parameters.

[0089] S113, parse the axis order information of the image data and rearrange the axis order of the image data to unify it into a channel-first standardized order.

[0090] S114 extracts pixel resolution parameters from the tag information of protein imaging data and converts them into pixel size information in standard physical units.

[0091] S115, determine the channel name information corresponding to the number of channels in the image data.

[0092] S116 generates a standard format multichannel image based on the rearranged image data, pixel size information, and channel name information.

[0093] The purpose of step S110 is to convert proprietary format imaging data from the first platform (such as AkoyaPCF) into a standardized format widely supported by downstream analysis processes, and to ensure the integrity and consistency of key metadata (such as channel order and physical size) in order to solve compatibility problems caused by inconsistencies in data format, channel organization and spatial metadata in different image analysis processes.

[0094] Step S111 requires specifying three key pieces of information: 1) the source data location (i.e., the path to the QPTIFF file to be converted); 2) the result output location; and 3) optional parameters controlling the conversion behavior. Specifically, the optional parameters include: image sequence index, used to process QPTIFF files containing multiple image sequences, specifying the specific sequence to be converted; resolution level index, used to process images with a multi-level resolution pyramid structure, specifying which level of image data to use; channel name list, allowing the user to explicitly provide biomarker names corresponding to the image channel order (e.g., "CD8", "DAPI"); and photometric interpretation, defining the color interpretation method of the output image (e.g., "minisblack" for grayscale, "rgb" for color), ensuring consistent display across different software.

[0095] Step S112 uses the image sequence index and resolution level index provided in S111 as key parameters to accurately parse and read the input protein imaging data (usually in QPTIFF format). This step first uses the image sequence index to locate the target image sequence to correctly handle complex file structures that may contain multiple independent image sequences (e.g., different fields of view or imaging rounds). Then, based on the resolution level index, it reads the corresponding level of image data from the selected sequence. This is specifically designed to support files with multi-level resolution pyramid structures, allowing users to flexibly choose full-resolution data for precise analysis or low-resolution data for rapid preview based on downstream computational needs.

[0096] It is worth noting that step S112 also includes:

[0097] S1121, if reading image data according to the resolution level index fails, it will automatically fall back to reading the original resolution image data of the specified image sequence, and record the actual read image data and related parameters in the log or metadata.

[0098] Understandably, when the system detects that the source file does not contain a multi-resolution pyramid structure, it will automatically ignore the hierarchical index parameter and directly read the original resolution image data of the specified sequence, thus ensuring broad compatibility and processing reliability for heterogeneous data sources. This refined reading strategy fundamentally overcomes the problems that existing general-purpose conversion tools may cause when dealing with complex, multi-level QPTIFF files, such as sequence confusion, misuse of resolution levels, or incomplete data due to the use of simple or default reading modes. This provides an accurate and error-free raw data foundation for all subsequent processing steps.

[0099] Step S113 is the core operation of structural analysis and standardization of the loaded raw protein imaging data. The specific process is as follows: First, the dimensional arrangement (axis order) of the image data in memory is automatically analyzed, for example, identifying it as a common format such as "height (Y)-width (X)-channel (C)" or "channel (C)-height (Y)-width (X)"; this analysis is fundamental to understanding the data organization structure. Next, a crucial rearrangement operation is performed, uniformly converting the image data into a standardized three-dimensional array of "channel (C)-row (Y)-column (X)". It is worth noting that step S113 also includes:

[0100] S1131, if the protein imaging data is determined to be a two-dimensional matrix structure, then the protein imaging data is confirmed to be a single-channel image, and a channel axis is inserted before the first dimension and displayed as a three-dimensional array as input for the axis order rearrangement operation.

[0101] This standardization process involves an important technical detail: when the input data is detected as a single-channel image without explicitly assigned channel dimensions (e.g., stored as a two-dimensional matrix), the system proactively introduces a three-dimensional array by expanding the dimensions. This array serves as the input for the axis order rearrangement operation, thus achieving indiscriminate treatment at the data structure level compared to multi-channel images. This eliminates the axis order confusion caused by differences in data sources, providing a unique and reliable data structure foundation for all subsequent quantitative calculations, visualizations, and cross-platform comparisons based on channel indexes. It also avoids algorithmic errors or misinterpretations of results caused by format ambiguity in existing technologies.

[0102] Step S114 aims to assign a precise spatial scale to the image data, enabling the mapping of pixel coordinates to real physical space. Specifically, it involves extracting pixel resolution parameters and their physical units from the metadata tags of the original QPTIFF file, and then uniformly converting them to physical pixel dimensions in micrometers (µm). Regardless of whether the original unit is "pixels / micrometer," "micrometers / pixels," or "inches," it is calculated and converted to the actual micrometer length corresponding to each pixel. This conversion provides a unified spatial measurement benchmark for data from different platforms with different resolution definitions, thus providing a necessary prerequisite for high-precision spatial registration with data from platforms such as Xenium, and providing a reliable basis for spatial quantitative analysis involving actual distance measurements (such as intercellular distance and region area), solving the problem of easily lost or inconsistent spatial metadata in existing processes.

[0103] Step S115 is responsible for establishing accurate biological identifiers for each channel of the image. Specifically, it prioritizes using a user-provided list of custom channel names; if no such list is provided, default names are automatically generated based on the number of channels. When a user provides a list, the system rigorously verifies whether its length matches the total number of image channels: if they match, the list is used as the official channel identifier, bound to the channel index, and written into standardized metadata; if they do not match, the process is immediately terminated and an error is reported. This verification mechanism fundamentally eliminates the risk of mismatch between channel names and data, ensuring the accuracy of data annotation and the authenticity of its biological significance.

[0104] Step S116 is the integration and output stage of the data standardization process. Based on the rearranged image data, standardized pixel physical dimensions, and a verified list of channel names, embedded metadata conforming to the OME data model is constructed. Subsequently, the image data and this metadata are written together into an OME-TIFF file. This standard format ensures broad compatibility with downstream tools such as ImageJ, QuPath, and bioformats. Simultaneously, a separate JSON format metadata file is generated, fully recording the entire chain of traceability information, including input / output paths, parameters used, channel names, and pixel dimensions.

[0105] The system ultimately outputs an ".ome.tif" image file and a "_metadata.json" metadata file. This dual-file output mechanism not only completes the format conversion but also ensures full data disclosure and ultimate traceability. It overcomes the problems of deeply embedded, difficult-to-extract, or easily lost metadata in existing tools, laying a reliable data foundation for reproducible scientific research analysis and automated processes.

[0106] Step S110 implements the entire conversion process from raw QPTIFF to standardized OME-TIFF and JSON metadata. First, it receives the user-specified input and output paths and optional parameters (such as image sequence, resolution level, channel name, etc.). Then, the core processing unit executes the following sequentially: reading the image data of the specified sequence and level, standardizing the axis order to a "channel-row-column" format (including automatic adaptation to single-channel data), extracting the pixel resolution and converting it to micrometer-scale physical dimensions, and simultaneously determining and strictly verifying the channel name list against the channel count. Finally, the system integrates the rearranged image data, standardized physical dimensions, and verified channel names to generate a TIFF file containing complete OME metadata, and simultaneously outputs a separate JSON file recording the entire process traceability information.

[0107] In conjunction with the first aspect, step S120 includes:

[0108] S121, use the selected channel image as the moving image and the spatial reference image as the fixed image.

[0109] S122 performs downsampling and normalization preprocessing on both moving and stationary images.

[0110] S123 performs feature point detection, descriptor calculation, and matching on the preprocessed image based on the scale-invariant feature transformation algorithm.

[0111] S124, based on the random sampling consensus algorithm, estimates the initial geometric transformation model from the moving image to the fixed image from the matched feature points.

[0112] S125, based on the initial geometric transformation model, obtains the final geometric transformation model for mapping the normalized multi-channel image to the spatial reference image coordinate space.

[0113] S126. Using the final geometric transformation model, the images of each channel of the standardized multi-channel image are mapped to the coordinate space of the spatial reference image to obtain a spatially aligned multi-channel registration image.

[0114] Step S120 aims to accurately map the feature image from the first platform (such as AkoyaPCF) to the spatial coordinate system defined by the second platform (such as 10xXenium).

[0115] In step S121, the spatial reference for registration is first determined: a high-contrast feature channel (usually the 0th layer DAPI image of PCF) is selected from the normalized multi-channel image as the "moving image", and the morphological reference image (morphology.tiff) from the Xenium platform is defined as the "fixed image", thereby establishing the mathematical goal of aligning the moving image to the coordinate space of the fixed image.

[0116] Step S122 performs standardization preprocessing on both the fixed and moving images to optimize computational efficiency and algorithm stability. First, both images are downsampled proportionally. While preserving sufficient structural information, the image resolution is reduced, significantly decreasing the computational load for subsequent feature extraction and matching. The downsampling uses area interpolation with a preset scaling factor, thus reducing image resolution while preserving structural features as much as possible, thereby reducing the computational load for subsequent feature extraction and matching, improving running speed, and reducing resource requirements. Second, grayscale normalization and format conversion are performed, linearly mapping pixel values ​​to a preset range (e.g., 0-255) and converting them to a common format such as 8-bit unsigned integers. This processing eliminates potential contrast and brightness differences between the original images, ensuring consistency of input for the feature extraction algorithm and providing stable and efficient input data for subsequent steps.

[0117] Step S123 employs the Scale Invariant Feature Transform (SIFT) algorithm to detect stable feature points and calculate their feature descriptors on downsampled and normalized image pairs, obtaining descriptor sets for each of the two images. If no valid descriptor is detected in either image, the process terminates; otherwise, a fast approximate nearest neighbor search algorithm is used for feature matching, comparing descriptor similarity to find corresponding local structures between the two images.

[0118] Since the initial matching inevitably contains a large number of errors (outside points), the core task of step S124 is to robustly estimate a global geometric transformation model.

[0119] Subsequently, step S125 performs optional fine-tuning based on the initial estimate: the initial model can be directly used as the final geometric transformation model; alternatively, an image intensity-based optimization process can be initiated, which iteratively optimizes low-resolution image pairs using an enhanced correlation coefficient method. This fine-tunes parameters to maximize the overall pixel correlation between images, correcting subtle system biases and achieving sub-pixel-level alignment. Regardless of whether optimization is enabled, this step ultimately outputs a high-precision geometric transformation model, providing a foundation for subsequent multi-channel mapping. This hybrid strategy (feature matching provides robustness, while intensity optimization improves accuracy) ensures registration success rate and accuracy in complex cross-platform scenarios.

[0120] Finally, step S126 uses the final geometric transformation model to independently perform geometric transformation and resampling on each channel of the standardized multi-channel image, accurately mapping all channel data to the coordinate space of the reference image, and finally combining and outputting a standard-format spatially aligned multi-channel registered image, providing a precise spatial basis for single-cell-level data fusion.

[0121] Figures 3 to 5 The effectiveness and key intermediate results of the hybrid registration scheme described in step S120 were directly verified. Figure 3 This is a morphological reference image (fixed image) for the Xenium platform. Figure 4 The image is a PCF DAPI channel image (moving image) registered using this method. Figure 5 The three images directly demonstrate the high-precision spatial alignment achieved through feature point detection and matching (S123), initial model estimation based on RANSAC (S124), and optional ECC fine-tuning (S125). This shows that the hybrid strategy effectively overcomes platform differences and provides a reliable geometric transformation basis for generating spatially aligned multi-channel registered images.

[0122] In conjunction with the first aspect, step S124 includes:

[0123] S1241, perform confidence filtering on the matched feature point pairs to eliminate mismatched point pairs.

[0124] S1242, based on the downsampling scaling factor, scales the coordinates of the matched feature point pairs, mapping them back from the downsampled image coordinate system to the original image coordinate system.

[0125] S1243 uses a stochastic consensus algorithm for estimation based on feature matching point pairs after scale recovery, prioritizing the estimation of similarity transformation models that include rotation, uniform scale transformation, and translation parameters.

[0126] S1244: If the similarity transformation model estimation fails or the number of its interior points does not meet the preset conditions, then the process reverts to estimating the affine transformation model.

[0127] S1245, the successfully estimated similarity transformation model or affine transformation model is used as the initial geometric transformation model.

[0128] First, step S1241 filters the preliminary matching results (e.g., based on the nearest neighbor distance ratio): using a multidimensional search structure, it finds multiple nearest neighbor candidates for each feature descriptor in another image; for each candidate match, it calculates the distance ratio between its nearest neighbor and second nearest neighbor. If the ratio is less than a preset threshold, it is determined to be a valid matching point pair; otherwise, it is considered unreliable and is removed, thus obtaining a set of high-confidence matching points.

[0129] For computational efficiency, feature detection and matching operations are typically performed on downsampled, low-resolution images. However, the final estimated geometric transformation model must be applied to the original resolution image. Therefore, the coordinates of the selected matching points need to be restored from the low-resolution coordinate system to the original coordinate system. By applying a downsampling scaling factor (0.25 if the image is downsampled to 1 / 4 of its original size), an inverse scaling transformation (usually a division operation) is performed on each pair of matching point coordinates to eliminate the systematic bias introduced by the resolution difference and ensure that subsequent geometric calculations are based on a unified original image space.

[0130] Even after screening, matching points may still contain noise. Therefore, a random sampling consensus algorithm is used to robustly estimate the geometric transformation model: the algorithm generates candidate models by repeatedly randomly sampling the minimum point set and evaluates the number of "interior points" that match the model. After iteration, the model with the most interior points is selected as the optimal estimate. Priority is given to estimating similarity transformation models (containing only rotation, uniform scaling, and translation), which are more stable in rigid or near-rigid deformation scenarios. If similarity transformation estimation fails (e.g., insufficient interior points or unreasonable parameters), it automatically reverts to estimating affine transformation models. Affine transformations have six degrees of freedom and can describe complex deformations such as non-uniform scaling and shearing; their estimation is also based on the RANSAC framework. This adaptive strategy significantly improves the method's adaptability to different image deformations and registration success rate. If the similarity transformation meets the preset validity conditions, it is directly used as the initial geometric transformation model.

[0131] In conjunction with the first aspect, step S125 includes:

[0132] S1251, if fine registration optimization is not enabled, the initial geometric transformation model will be used as the final geometric transformation model.

[0133] S1252, if fine registration optimization is enabled, the initial geometric transformation model is converted into an affine transformation matrix as the optimization starting point. Based on the optimization starting point, the enhanced correlation coefficient method is used to perform iterative optimization on the constructed low-resolution image pairs to determine the final geometric transformation model.

[0134] In practical applications, based on judgments regarding data quality, time efficiency, or specific requirements, the computationally intensive fine-grained optimization may be omitted. When fine-grained registration optimization is not enabled by system configuration or user instructions, step S1251 is executed: the initial geometric transformation model estimated by robust feature matching (SIFT+RANSAC) in step S124 is directly adopted as the final geometric transformation model. This decision is understandably based on the premise that the initial model has passed rigorous interior point screening and validity checks, providing sufficiently reliable, pixel-level spatial alignment to meet the needs of many analysis scenarios. It saves computational resources while ensuring the reliability of the output while pursuing efficiency.

[0135] When fine optimization is enabled, step S1252 performs the following core process to achieve sub-pixel level high-precision registration:

[0136] First, the initial geometric transformation model is converted into an affine transformation matrix, providing a compatible and near-optimal iterative starting point for subsequent optimization based on enhanced correlation coefficients. Then, low-resolution image pairs are constructed to improve optimization efficiency and stability: both fixed and moving images are uniformly scaled, with the moving image pre-aligned using the affine matrix before scaling to ensure approximate alignment; simultaneously, the linear and translation parameters of the matrix are scaled according to the scaling ratio to precisely adapt to the low-resolution space.

[0137] Next, to improve optimization efficiency and stability, a low-resolution image pair for fine registration is constructed: the fixed image and the moving image are scaled to the same size. Before scaling, the moving image is pre-aligned using an initial affine transformation matrix to ensure that the image pair is roughly aligned. At the same time, based on the ratio between the low-resolution and original images, the linear and translation parameters of the matrix are scaled to adapt to the low-resolution space, providing accurate initial conditions for iterative optimization.

[0138] Subsequently, on the preprocessed low-resolution image pairs, the geometric transformation model is iteratively optimized using the enhanced correlation coefficient method. By fine-tuning the scale-adapted affine transformation matrix parameters, the global pixel correlation between images is maximized to achieve sub-pixel-level fine-tuning. This process does not rely on feature points and can correct subtle systematic biases that may remain in feature matching.

[0139] Finally, the final model and fallback mechanism are determined: The optimization process generates an optimized geometric transformation model. If optimization is successful, this model is adopted as the final geometric transformation model. However, if the ECC optimization process encounters an anomaly (e.g., numerical instability) or fails to converge (e.g., exceeding the iteration limit or failing to improve the correlation coefficient), the system will safely terminate the fine-tuning optimization and fall back to using the initial geometric transformation model estimated by RANSAC as the final model. This mechanism ensures the entire process is highly robust to challenging data.

[0140] In conjunction with the first aspect, step S126 includes:

[0141] S1261, Obtain the number of channels and spatial dimensions of each channel of the standardized multi-channel image, as well as the spatial dimensions of the spatial reference image.

[0142] S1262, Based on the spatial dimensions of the spatial reference image, initialize the registered image space to accommodate the mapping results.

[0143] S1263 utilizes the final geometric transformation model to resample and map the image data of each channel of the standardized multi-channel image from the coordinate space of the standardized multi-channel image to the registered image space.

[0144] S1264 combines the mapped image data of each channel to form a spatially aligned multi-channel registered image that is consistent with the spatial reference image space.

[0145] Step S126 is the final execution and data generation stage of the image registration process. Its core task is to apply the determined precise spatial transformation relationship to the complete image data of all protein channels and generate standardized data products that can be directly used for downstream single-cell analysis.

[0146] In step S1261, the normalized multi-channel image (usually in OME-TIFF format) is first parsed to obtain the total number of channels and the height (number of rows) and width (number of columns) of each channel. Simultaneously, the spatial dimensions are obtained from a reference image serving as the spatial benchmark (such as Xenium's morphology.tiff). This metadata forms the basis for all subsequent memory allocations, loop calculations, and coordinate mappings, ensuring a strict match between the processing procedure and the actual data structure, and avoiding array out-of-bounds errors or information pruning due to size misjudgments.

[0147] In step S1262, a three-dimensional or multi-dimensional array is created and initialized in memory based on the spatial dimensions (height, width) of the reference image. The spatial dimensions (Y, X) of this array are completely consistent with those of the reference image to ensure a one-to-one correspondence between the output image and the reference spatial coordinates; its channel dimension is equal to the number of channels in the multi-channel image to be processed. During initialization, all elements of the array are typically set to preset background values ​​(such as 0 or NaN). This operation predefines the standard structure and spatial framework of the output data, preparing for efficient and orderly filling of the registration results for each channel.

[0148] Subsequently, in step S1263, for each individual biomarker channel (such as CD8, PD-1, DAPI, etc.) in the normalized multi-channel image, the following operations are performed:

[0149] Apply the transformation model: Use the final geometric transformation model determined in step S125 (which may be a high-precision model optimized by ECC or a robust initial model) that mathematically defines a mapping function from PCF image coordinates to Xenium reference coordinates.

[0150] Resampling calculation is performed: For each pixel in the original channel image, or using a more efficient grid traversal method, its new coordinates in the target space (i.e., the image space initialized by S1262) are calculated according to the transformation model. Since the new coordinates are usually non-integer, the system uses an interpolation algorithm (such as bilinear interpolation or cubic spline interpolation) to calculate the intensity value of the pixel at that location, thereby generating high-quality, continuous resampling results.

[0151] Result filling: Write the calculated pixel intensity value, which has been transformed to the new coordinate space, into the corresponding position of the corresponding channel in the registered image space initialized by S1262.

[0152] This process is performed independently and sequentially (or in parallel) on each channel, ensuring that the geometric transformation of each protein signal is precise and consistent.

[0153] Finally, step S1264 is the final integration and encapsulation of the data. After all channels have undergone independent coordinate mapping and resampling, and their results are filled into the pre-allocated image space, the data stored in this image space constitutes a complete, spatially aligned, multi-channel registered image dataset. Subsequently, the system encapsulates this dataset and writes it into an image file conforming to the OME-TIFF standard. This output file not only contains the registered pixel data for all channels but also records key information such as channel names, pixel physical dimensions (micrometers / pixel), and spatial coordinate systems in the file header or accompanying metadata. This standardized output ensures the interchangeability and reproducibility of the data and serves as a direct input for protein data extraction and data fusion pairing, providing a rigorously spatially calibrated foundational image for achieving precise correlation between protein expression quantification and transcriptome data at the single-cell level.

[0154] In conjunction with the first aspect, step S130 includes:

[0155] S131, Load cell segmentation information from the second platform and obtain cell label data.

[0156] S132, load the spatially aligned multi-channel registration image and acquire image data for each channel.

[0157] S133, establish the spatial mapping relationship between cell label data and image data of each channel.

[0158] S134, based on spatial mapping relationships, uses a block-based parallel computing approach to process image data from each channel, classifying each pixel value to its corresponding cell label according to its spatial coordinates.

[0159] S135, based on the classification results, calculate the pixel statistical feature value of each cell label in each image channel. The pixel statistical feature value includes the mean and median.

[0160] The purpose of step S130 is to transform the spatially registered, image-level multichannel protein signals into quantitative feature vectors per cell through efficient and parallel statistical computation, and to complete the structured integration before fusion with transcriptome data.

[0161] First, based on the externally input operating parameters, the system initializes the operating environment, including resolving the spatial data output path, the second platform (Xenium) data source path, the registered first platform (PCF) image file path, and the parallel computing resource configuration, and creates the necessary directory structure to provide a unified resource management foundation for all subsequent data operations.

[0162] In step S131, the system loads a spatial data object from the second platform (Xenium) according to the parameters. This object can be constructed directly from the original output directory or loaded from an existing Zarr format file, which integrates cell spatial location, boundary, and transcriptomic phenotypic data. Key cell segmentation information, namely "cell tag data"—a two-dimensional raster image in which the integer value of each pixel represents a unique cell ID (0 indicates background)—is extracted.

[0163] During the data preparation phase, the system performs two key loading steps: First, it loads the Xenium spatial data object. If a corresponding morphological image exists, it reads, parses, and converts it into a standard (channel, row, column) format before injecting it into the object to construct a unified data structure containing cell location, boundaries, and transcriptome information. Second, it loads the spatially aligned PCF multi-channel image (such as OME-TIFF) output by the registration module, reads it, merges layers, normalizes dimensions, and parses channel names, and injects it into the same spatial data object as a whole. This ensures that cell segmentation information and protein image data coexist in the same object and have the same coordinate system.

[0164] Subsequently, the system establishes a precise spatial correspondence: cell tag raster data (each pixel value is a cell ID) is extracted from the spatial data object, and the registered PCF image is aligned with the cell tag data in the spatial coordinate system to construct a one-to-one mapping relationship between each pixel of the PCF image and the cell tag. This allows for rapid querying of the cell to which any pixel belongs and its intensity value in each protein channel.

[0165] Subsequently, a block-based parallel computing architecture is employed to process large-scale data: First, based on the number of channels and cell label range of the PCF image, accumulation, counting, and caching structures for pixel statistics are initialized. Next, each channel image is processed sequentially. Dask is used to divide each channel image into multiple spatial blocks, and joblib is used to distribute the tasks of each block to multiple cores for parallel execution. In each parallel task, each pixel of the assigned image block is traversed, and its corresponding cell label is queried based on the established spatial mapping. If the label is valid (not background), the pixel intensity value is accumulated into the statistical structure corresponding to that cell ID (e.g., summing, increasing the count, or caching the value). This method significantly reduces memory usage and greatly improves processing speed through parallelization.

[0166] Finally, in the result generation and fusion preparation stage, the cumulative results of each parallel task are summarized. For each cell, statistical features such as mean (based on the sum and number of pixel values) and median (based on the cached pixel value list) are calculated on each PCF channel. The results are then converted into a structured data table indexed by cell labels, ultimately forming a structured dataset with cells as rows and protein features as columns.

[0167] In conjunction with the first aspect, step S140 includes:

[0168] S141, the pixel statistical feature values ​​are associated and matched with the transcriptome data provided by the second platform, which is indexed by cell tags.

[0169] S142, the matched pixel statistical feature values ​​are added to the transcriptome data as a proteome data field to generate spatial single-cell data containing the fusion of transcriptome and proteome data.

[0170] Step S141 is a key operation for achieving accurate alignment of multimodal data, specifically including:

[0171] Transcriptome data from the second platform and pixel statistical feature values ​​calculated from images from the first platform (PCF) were both indexed using the same cell tag (CellID). This tag is a unique identifier assigned to each segmented cell in the Xenium data analysis pipeline.

[0172] The algorithm uses cell tags as "keys" to match and merge "pixel statistical feature value tables" (containing columns such as cell ID, CD8_mean, and CD8_median) representing protein expression levels with "transcriptome feature matrices" (usually containing columns such as cell ID, Gene1, and Gene2) representing gene expression levels. This operation is equivalent to "primary key joins" in a database, ensuring that all molecular information from the same physical cell is accurately merged together.

[0173] Step S142 completes the final data structure integration and generation, specifically including: based on successful matching, adding the protein statistical characteristics (mean, median, etc.) of each cell as new fields to the transcriptome data of the corresponding cell, so that each cell simultaneously contains gene expression profiles and protein expression profiles. After this operation, all cells form a structured fusion dataset with cells as rows and multi-omics features as columns. This dataset can be directly used for subsequent cross-modal association analyses such as gene-protein co-expression. Finally, the fused spatial single-cell data and its spatial location, boundaries, and other metadata are written to the storage system as a complete spatial data object in Zarr format. This format supports efficient block read / write and parallel access, and is suitable for processing large-scale data. This persistence step not only saves the fusion results but also ensures that the data can be reused by downstream tools, supports cross-platform interaction and reproducible analysis, and achieves data reproducibility, scalability, and interactivity.

[0174] Combination Figure 6 This paper demonstrates the fusion effect at a complete chip spot scale after registration using the proposed method. The base image is a PCF DAPI (nuclear staining) channel image, representing the spatial information of the protein imaging platform; the superimposed white semi-transparent outline is the cell segmentation boundary provided by the Xenium platform. As can be seen in the figure, the spatial location and morphology of cells defined by the two very different data sources are highly consistent, proving that the hybrid registration strategy of this invention can achieve accurate spatial alignment at the overall scale across platforms. Figure 7 for Figure 6 The magnified view of the selected area reveals the detailed precision at the cellular level. Under this magnified view, it is clearly observed that each Xenium cell segmentation boundary (white outline) closely encloses and corresponds one-to-one with the corresponding nuclear signal (bright area) in the PCF image. This demonstrates that the registration accuracy of this invention reaches the subcellular level, ensuring that in subsequent steps, each PCF image pixel can be accurately classified to its corresponding Xenium cell identifier, thus providing a reliable spatial basis for multimodal data fusion at the single-cell level. This indicates that this method achieves not only geometric spatial alignment but also accurate correlation of biological signals, thereby ensuring that the final generated fused spatial single-cell data has high biological reliability and can reliably support subsequent cross-modal joint analysis.

[0175] Secondly, embodiments of this application provide a fusion device for multimodal spatial omics data, combined with Figure 2As shown, the device includes: a normalization processing module 10, an image registration module 20, a feature statistics module 30, and a data fusion module 40. The normalization processing module 10 performs normalization format conversion on protein imaging data from a first platform to generate a normalized multi-channel image containing channel priority dimension and pixel physical size information. The image registration module 20 performs image registration of selected channels from the normalized multi-channel image with a spatial reference image from a second platform to obtain a spatially aligned multi-channel registered image. The feature statistics module 30 extracts pixel statistical feature values ​​for each cell region in each image channel from the spatially aligned multi-channel registered image in parallel, based on cell segmentation information from the second platform. The data fusion module 40 associates the pixel statistical feature values ​​with the transcriptome data of the corresponding cells to generate fused spatial single-cell data.

[0176] Thirdly, embodiments of this application provide a fusion system for multimodal spatial omics data, including the apparatus described above.

[0177] Fourthly, embodiments of this application also provide an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor runs the computer program to cause the electronic device to perform the above-described method. Furthermore, the electronic device also includes a bus and a communication interface, and the processor, communication interface, and memory are interconnected.

[0178] The memory includes both volatile and non-volatile storage media. The communication interface supports wired or wireless connections, enabling communication with other network elements through various networks. The bus handles data transmission and control between components. The processor is an integrated circuit chip capable of implementing the steps of the method described in this invention through its hardware logic or by executing software instructions stored in the memory. The processor can be a general-purpose processor, digital signal processor, application-specific integrated circuit, programmable logic device, or other hardware component with signal processing capabilities.

[0179] Fifthly, embodiments of this application provide a readable storage medium storing computer program instructions, which are read and executed by a processor to perform the above-described method.

[0180] Finally, it should be noted that the above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for fusing multimodal spatial omics data, characterized in that, The method includes: The protein imaging data from the first platform is standardized and converted to generate a standardized multi-channel image that includes channel priority dimension and pixel physical size information. The selected channel image in the standardized multi-channel image is image registered with the spatial reference image from the second platform to obtain a spatially aligned multi-channel registered image. Based on cell segmentation information from the second platform, pixel statistical feature values ​​of each cell region in each image channel are extracted in parallel from the spatially aligned multi-channel registration image. The pixel statistical feature values ​​are correlated with transcriptome data from the second platform, identified by the cells, to generate fused spatial single-cell data. The step of image registration between a selected channel image from the standardized multi-channel image and a spatial reference image from the second platform to obtain a spatially aligned multi-channel registered image includes: The selected channel image is used as the moving image, and the spatial reference image is used as the fixed image; The moving image and the fixed image are preprocessed by downsampling and normalization; The scale-invariant feature transform algorithm is used to perform feature point detection, descriptor calculation and matching on the preprocessed image; Based on the random sampling consensus algorithm, the initial geometric transformation model from the moving image to the fixed image is estimated from the matched feature points; Based on the initial geometric transformation model, a final geometric transformation model for mapping the standardized multi-channel image to the spatial reference image coordinate space is obtained; Using the final geometric transformation model, each channel of the standardized multi-channel image is mapped to the coordinate space of the spatial reference image to obtain the spatially aligned multi-channel registration image.

2. The method according to claim 1, characterized in that, The steps for standardizing and converting protein imaging data from the first platform to generate a standardized multichannel image containing channel priority dimension and pixel physical size information include: Obtain the input path of the protein imaging data to be converted, the output path of the target normalized multi-channel image, and the optional parameters; Based on the image sequence index and resolution level index in the optional parameters, the corresponding image data is read from the protein imaging data; The axis order information of the image data is parsed, and the axis order of the image data is rearranged to unify it into a channel-first standardized order; Pixel resolution parameters are extracted from the label information of the protein imaging data and converted into pixel size information in standard physical units; Determine the channel name information corresponding to the number of channels in the image data; Based on the rearranged image data, the pixel size information, and the channel name information, a standard format multi-channel image is generated.

3. The method according to claim 1, characterized in that, The step of estimating the initial geometric transformation model from the moving image to the fixed image from matched feature points based on the random sampling consensus algorithm includes: The confidence level of the matched feature point pairs is used to filter out mismatched point pairs. Based on the downsampling scaling factor, the coordinates of the matched feature point pairs are scaled back, and they are mapped from the downsampled image coordinate system back to the original image coordinate system. Based on the feature matching point pairs after scale recovery, iterative estimation is performed using the random sampling consensus algorithm, prioritizing the estimation of the similarity transformation model; If the similarity transformation model satisfies the preset validity conditions, it is used as the initial geometric transformation model; If the similarity transformation model does not meet the validity condition, then the process reverts to estimating the affine transformation model, and the successfully estimated affine transformation model is used as the initial geometric transformation model.

4. The method according to claim 1, characterized in that, The step of obtaining a final geometric transformation model for mapping the normalized multi-channel image to the spatial reference image coordinate space based on the initial geometric transformation model includes: If fine registration optimization is not enabled, the initial geometric transformation model will be used as the final geometric transformation model. If fine registration optimization is enabled, the initial geometric transformation model is converted into an affine transformation matrix as the starting point for optimization. Based on this starting point, the enhanced correlation coefficient method is used for iterative optimization on the constructed low-resolution image pairs to determine the final geometric transformation model.

5. The method according to claim 1, characterized in that, The step of mapping each channel of the normalized multi-channel image to the coordinate space of the spatial reference image using the final geometric transformation model to obtain the spatially aligned multi-channel registration image includes: Obtain the number of channels and spatial dimensions of each channel of the standardized multi-channel image, as well as the spatial dimensions of the spatial reference image; Based on the spatial dimensions of the spatial reference image, initialize the registered image space to accommodate the mapping results; Using the final geometric transformation model, the image data of each channel of the standardized multi-channel image is resampled and mapped from the coordinate space of the standardized multi-channel image to the registered image space. The image data mapped from each channel are combined to form a spatially aligned multi-channel registered image that is consistent with the spatial reference image space.

6. The method according to claim 1, characterized in that, Based on cell segmentation information from the second platform, the step of extracting pixel statistical feature values ​​of each cell region in each image channel in parallel from the spatially aligned multi-channel registered image includes: Load the cell segmentation information from the second platform and obtain cell tag data; Load the spatially aligned multi-channel registered image and acquire image data for each channel; Establish a spatial mapping relationship between the cell tag data and the image data of each channel; Based on the spatial mapping relationship, a block-parallel computing method is used to process the image data of each channel, and classify each pixel value into the corresponding cell label according to its spatial coordinates. Based on the classification results, the pixel statistical feature values ​​of each cell label in each image channel are calculated, and the pixel statistical feature values ​​include the mean and median.

7. The method according to claim 1, characterized in that, The step of associating the pixel statistical feature values ​​with transcriptome data identified by the cell from the second platform includes: The pixel statistical feature values ​​are then correlated and matched with the transcriptome data indexed by cell tags provided by the second platform. The matched pixel statistical feature values ​​are added to the transcriptome data as proteome data fields to generate spatial single-cell data containing the fusion of transcriptome and proteome data.

8. A device for fusing multimodal spatial omics data, characterized in that, The apparatus is configured to perform the method as described in any one of claims 1-7, the apparatus comprising: The standardization processing module is used to perform standardization format conversion on protein imaging data from the first platform to generate a standardized multi-channel image containing channel priority dimension and pixel physical size information. The image registration module is used to perform image registration between a selected channel image in the standardized multi-channel image and a spatial reference image from the second platform to obtain a spatially aligned multi-channel registered image. The feature statistics module is used to extract pixel statistical feature values ​​of each cell region in each image channel in parallel from the spatially aligned multi-channel registered image based on cell segmentation information from the second platform. The data fusion module is used to associate the pixel statistical feature values ​​with the transcriptome data of the corresponding cells to generate fused spatial single-cell data; The step of image registration between a selected channel image from the standardized multi-channel image and a spatial reference image from the second platform to obtain a spatially aligned multi-channel registered image includes: The selected channel image is used as the moving image, and the spatial reference image is used as the fixed image; The moving image and the fixed image are preprocessed by downsampling and normalization; The scale-invariant feature transform algorithm is used to perform feature point detection, descriptor calculation and matching on the preprocessed image; Based on the random sampling consensus algorithm, the initial geometric transformation model from the moving image to the fixed image is estimated from the matched feature points; Based on the initial geometric transformation model, a final geometric transformation model for mapping the standardized multi-channel image to the spatial reference image coordinate space is obtained; Using the final geometric transformation model, each channel of the standardized multi-channel image is mapped to the coordinate space of the spatial reference image to obtain the spatially aligned multi-channel registration image.

9. A fusion system for multimodal spatial omics data, characterized in that, The apparatus as described in claim 8.

Images