Colorectal cancer space transcriptome prediction method and device based on deep learning and storage medium

Abstract

The invention relates to a colorectal cancer space transcriptome prediction method and device based on deep learning and a storage medium. The prediction method comprises the following steps: collecting paired digital pathological images and spatial transcriptome data to form a data set; preprocessing the spatial transcriptome data; the method comprises the following steps of: constructing an STFormer model; the Style-Aug module is used for performing feature space conversion on an input digital pathological image; the dense connection convolutional network DenseNet module is used for constructing a deep mapping relation between a pathological image and a spatial transcriptome; the inter-batch full-image crossing Transform module is used for effectively avoiding overfitting of the model due to data volume limitation through an inter-batch attention mechanism; and performing spatial transcriptome prediction on the input digital pathological image by using the trained STFormer model to obtain a prediction result. Under the condition of a small sample, the performance and the general stability of a pathological image prediction space transcriptome are remarkably improved, and an efficient and reliable deep learning tool is provided for analyzing a tumor microenvironment.

Description

Technical Field

[0001] The present invention relates to digital pathology image AI, and in particular to a spatial transcriptome prediction method for colorectal cancer based on deep learning. Background Art

[0002] Spatial transcriptomics technology is of great value in analyzing gene expression patterns in the tumor microenvironment, but its high cost limits its widespread application. Using histological images to predict spatial transcriptome data has become a cost-effective alternative. However, existing methods such as STNet, HistoGene, and Hist2ST have shortcomings in dealing with staining variations and relationships between spots, especially when data is limited. These methods either ignore the staining differences between different tissue sections, resulting in poor model generalization ability, or fail to effectively explore the complex relationships between spots, affecting prediction accuracy.

[0003] STNet is based on the DenseNet architecture. Although DenseNet performs well in feature utilization, STNet may focus too much on the features of a single spot and ignore the relationships between spots. This neglect may cause the model to perform poorly in capturing the spatial heterogeneity of the tumor microenvironment, which is crucial for understanding gene expression patterns.

[0004] HistoGene uses the VisionTransformers architecture, and such models typically require a large amount of data to be effectively trained. In the field of medical imaging, this becomes a significant limitation due to the scarcity of data and the high cost of acquisition. In addition, differences in preparation techniques and equipment for pathological sections can lead to significant differences in image style. HistoGene may be sensitive to these differences, resulting in unstable performance on new datasets.

[0005] Hist2ST combines convmixer, global Transformer encoder and graph neural network. Although graph neural network can capture the relationship between spots, this complexity may lead to overfitting, especially in the case of limited data. In addition, due to the inclusion of graph neural network and Transformer layers, the high memory consumption of Hist2ST may make it difficult to apply in environments with limited computing resources.

[0006] It can be seen that the existing spatial transcriptome prediction methods have the problem of difficulty in ensuring prediction accuracy and high computational cost. Summary of the invention

[0007] To solve the above problems, the present invention provides a spatial transcriptome prediction model component based on digital pathology images, and the specific technical solution is as follows: A method for spatial transcriptome prediction of colorectal cancer based on deep learning, comprising the following steps: Paired digital pathology images and spatial transcriptome data were collected to form a dataset; Preprocessing of spatial transcriptome data, including dataset partitioning, gene selection, and image processing; Constructing an STFormer model, the STFormer model includes a Style-Aug module for sequentially processing images for performing feature space conversion on input digital pathology images, a densely connected convolutional network DenseNet module for constructing a deep mapping relationship between pathology images and spatial transcriptomes, and an inter-batch cross-full-slice image Transformer module for effectively avoiding overfitting of the model due to data volume limitation through an inter-batch attention mechanism; The trained STFormer model is used to perform spatial transcriptome prediction on the input digital pathology images to obtain the prediction results.

[0008] Preferably, the data set division includes: a first data set, which includes whole-slice images of several colorectal cancer patients and their corresponding spatial transcriptome data, each whole slice has multiple first spots, each first spot corresponds to a gene expression vector, and each first spot has a diameter of 65 microns, corresponding to 280 pixels on the whole slice, and is used to analyze the expression of 18,000 to 22,000 genes; and a second data set, which includes a whole slice data, with a total of 2,660 second spots, each second spot has a diameter of 61.4 to 62.4 microns, corresponding to 94 pixels, and is used to analyze the expression of 17,943 genes.

[0009] Furthermore, the gene selection includes the following steps: performing variance analysis on the gene expression values ​​in the first data set, and selecting the top N genes with the largest variance for subsequent prediction; dividing the first data set using a leave-one-out cross-validation method so that part of the data is used for training and the rest of the data is used for testing.

[0010] Furthermore, the image processing includes: adjusting the histological image to a uniform size by an interpolation method; and matching the uniform size with the physical diameter of the spatial transcriptome sites in the training set.

[0011] Furthermore, the training steps of the STFormer model are as follows: The digital pathological images corresponding to the spots of the spatial transcriptome are input into the Style-Aug module for real-time image enhancement. The image after Style-Aug style enhancement is represented as { x i Aug}(i =1,2,… M ), where M is the number of spots in the training batch, x i Aug It is i An enhanced spot image; The enhanced image is input into the densely connected convolutional network DenseNet module, and the corresponding image features are output. F Aug ={ f i Aug}( i =1,2,… M ); The enhanced image features are input into the batch-to-batch cross-slice image Transformer module to interact between different full slices across batches and capture the association between images at different sites. The features processed by the Transformer module across the entire image slice between batches are input into the shared prediction head. The true value of the spatial transcriptome is used as supervision to construct the regression task loss function and iteratively update the parameters.

[0012] Furthermore, the batch-to-batch cross-image Transformer module includes a multi-head self-attention module: In the formula, X = F Aug ; W Q , W K , W V are all learnable weight matrices; d k is the dimension of the key vector; i Indicates i Head, h is the number of heads, W O is the output weight matrix.

[0013] Furthermore, the batch-by-batch Transformer module builds multi-head self-attention for images in each batch, which avoids model overfitting due to data size limitation while learning the association between images: In the formula, MSA stands for multi-head self-attention module, MLP stands for multi-layer perceptron, The result output of DenseNet is: l = 1,2,3,4,5,6, LN is the layer normalization operation, for Output after MSA, MLP and LN, and so on.

[0014] Furthermore, the prediction head of the Transformer module across the entire image slice between batches is a shared predictor, which learns the global correlation of different spots in the training phase; in the inference phase, the shared predictor makes the predicted results stable and generalizable.

[0015] A device for a method for predicting spatial transcriptome of colorectal cancer based on deep learning, comprising: a processor; a memory on which a computer program that can be run on the processor is stored; wherein the computer program, when executed by the processor, implements the steps of the method for predicting spatial transcriptome of colorectal cancer based on deep learning.

[0016] A computer-readable storage medium having a data processing program stored thereon, wherein the data processing program, when executed by a processor, implements the steps of a method for spatial transcriptome prediction of colorectal cancer based on deep learning.

[0017] Compared with the prior art, the present invention has the following beneficial effects: Under the condition of small samples, the present invention significantly improves the performance and pan-stability of spatial transcriptome prediction in pathological images, and provides an efficient and reliable deep learning tool for analyzing the tumor microenvironment. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of a method provided by an embodiment of the present invention; Figure 2 It is a structural diagram of the STFormer model provided by an embodiment of the present invention; Figure 3 It is a Style-Aug style enhancement network structure diagram provided by an embodiment of the present invention; Figure 4 It is a multi-head self-attention architecture for cross-batch digital pathology images provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The present invention will be further described with reference to the accompanying drawings.

[0020] 1) Construct an image enhancement module (Style-Aug module) based on style transfer, and perform feature space transformation on the input digital pathology image through a pre-trained style transfer network, effectively eliminating the image style differences between different datasets and improving the model's cross-dataset generalization ability; 2) Using a densely connected convolutional network (DenseNet) as a feature extractor, the dense connection mechanism is used to capture the local detail features and multi-level semantic information of pathological images, and to construct a deep mapping relationship between pathological images and spatial transcriptomes; 3) Design a batch-by-batch cross-whole-slice image Transformer module (Cross-WSI Transformer), which captures the global relationship between spatial sites (spots) in different pathological images through the self-attention mechanism, establishes long-range dependencies between spots, and comprehensively explores the spatial interaction characteristics between spots. While enhancing the image-spatial transcriptome mapping capability, the batch-by-batch attention mechanism effectively avoids overfitting of the model due to data limitation.

[0021] The above modules 1-3 are organically integrated into a unified model component, named STFormer (SpatialTranscriptomics Former). When training the model, a training set containing digital pathology images and their corresponding spatial transcriptome data is used for end-to-end training; when applying the model, the digital pathology image to be predicted is input, and the corresponding spatial transcriptome prediction result is output through STFormer.

[0022] like Figures 1 to 4 As shown, a colorectal cancer spatial transcriptome prediction method based on deep learning includes the following steps: Step 1: Collect paired digital pathology images and spatial transcriptome data to form a dataset; Step 2: Preprocess the gene expression data, including gene selection and image processing; Step 3: Build the STFormer model, including: The Style-Aug module is an image enhancement module based on style transfer. It transforms the feature space of the input digital pathology image through a pre-trained style transfer network, effectively eliminating the image style differences between different datasets and improving the model's cross-dataset generalization ability. The densely connected convolutional network (DenseNet) module is an image feature extractor that captures local detail features and multi-level semantic information of pathological images through a dense connection mechanism, and constructs a deep mapping relationship between pathological images and spatial transcriptomes. The batch-to-batch cross-whole-slice image Transformer module (Cross-WSI Transformer) captures the global relationship between spatial sites (spots) between different pathological images through the self-attention mechanism, establishes long-range dependencies between spots, and comprehensively explores the spatial interaction characteristics between spots. While enhancing the image-spatial transcriptome mapping capability, the batch-to-batch attention mechanism effectively avoids overfitting of the model due to data limitation.

[0023] Step 4: Use the trained model to perform spatial transcriptome prediction on new digital pathology images to obtain prediction results.

[0024] The specific steps are as follows: Step 1: Collect histological images of colorectal cancer and corresponding spatial transcriptome data to form a data set.

[0025] A first dataset is used, which can be named the "Charting" dataset. The dataset contains 14 whole-slice images from 7 patients with colorectal cancer and their corresponding spatial transcriptome data. There are multiple first spots on each whole slice, and each first spot corresponds to a gene expression vector; wherein, each first spot has a diameter of 65 microns, corresponding to 280 pixels on the whole slice, and is used to analyze the expression of approximately 20,000 genes. In order to increase the generalization ability of the model, another second dataset is used. The second dataset can be named the "34_Intestine_CRC" dataset, which contains a whole slice data, a total of 2,660 second spots, each second spot has a diameter of approximately 61.9 microns, corresponding to 94 pixels, and is used to analyze the expression of 17,943 genes. 34_Intestine_CRC dataset.

[0026] Step 2: Data preprocessing, mainly including data set division, gene selection and image processing.

[0027] The dataset was divided using leave-one-out cross-validation. The training was conducted using the data of 6 patients in the Charting dataset (12 full slides and their corresponding spatial transcriptome data), and the data of the remaining 1 patient (2 full slides and their corresponding spatial transcriptome data) was used for testing. Another 4 patients' paired data (including digital pathology images and spatial transcriptomes) were used as an independent external test set to evaluate the model performance.

[0028] In the gene selection stage, variance analysis is first performed on the gene expression values ​​in the training data. Specifically, the variance of the expression values ​​of each gene in all samples is calculated, and the top N genes with the largest variance are selected for subsequent prediction. This allows us to focus on genes that have large variations in different samples, which tend to have higher biological significance and information content. When selecting the number of genes N, a trade-off is made based on experimental data and computing resources, and usually hundreds to thousands of genes are selected. In addition, other preprocessing steps are performed on the gene expression data, such as standardization and normalization, to eliminate the impact of differences in the expression magnitude of different genes.

[0029] A variance analysis was performed on the gene expression values ​​in the Charting dataset, and the 1000 genes with the largest variance were selected. Finally, 418 genes were determined as prediction targets through cross-validation. The expression values ​​of these genes were normalized by minimum-maximum normalization to preserve their distribution pattern.

[0030] During the image processing stage, the histology images were resized to a uniform size. This step is usually achieved by interpolation methods, such as bilinear interpolation or bicubic interpolation, to ensure that image quality is maintained during image scaling. For the external dataset “34_Intestine_CRC”, the standard diameter of its spot images was adjusted to the same 65 μm as the Charting dataset, and the corresponding spot space size was adjusted to 128 × 128 pixels.

[0031] Step 3: Build the STFormer model, such as Figure 2 As shown; 3.1, such as Figure 3 As shown in the figure, we first perform real-time style transfer enhancement through the pre-trained Style-Aug. Style-Aug transfers the style of natural images to digital pathology images through the AdaIN function. This process aims to eliminate the staining differences between different tissue sections and improve the generalization ability of the model. It ensures that the semantic content of the image is not changed during the style transfer process, and only the style features are replaced to maintain the diagnostic information of the image.

[0032] AdaIN is the core function of Style-Aug: In the formula, the characteristic For a given content graph, features For a given style map, Represents the mean of the style map, used for shifting, Represents the variance of the style map, which is used to scale and normalize the content input. Through AdaIN, the content map Convert to style.

[0033] The Style-Aug module generates visual representations that are insensitive to specific medical styles, retains domain-independent content, and avoids interference from domain-specific styles, thereby enhancing the generalization ability of the model on different full-slice datasets. It is then input into the DenseNet module for processing.

[0034] 3.2. Input the enhanced spot image into the densely connected convolutional network DenseNet module and output image features F Aug ={ f i Aug}( i =1,2,… M ): The image after Style-Aug style enhancement is represented as { x i Aug}( i =1,2,… M ),in M is the number of spots in the training batch, x i Aug It is i The enhanced spot corresponds to the image. f(·) represents DenseNet, and the output image features are F Aug ={ f i Aug}( i =1,2,… M ).

[0035] 3.3. The output features are input into the batch-to-batch cross-slice image Transformer module, and the spot image features between different slices implement the self-attention mechanism in the same batch to capture the association between spots.

[0036] like Figure 4 As shown in Figure 2, the calculation formula of the single-head self-attention module is: In the formula, q , k , v is a linear projection, d k yes k Dimension.

[0037] make X = ; Then the formula of the multi-head self-attention (MSA) module is In the formula,W Q , W K , W V are all learnable weight matrices; d k is the dimension of the key vector; i Indicates i Head, h is the number of heads, W O is the output weight matrix.

[0038] The batch-to-batch Transformer module across the entire image is composed of six cascades of multi-head self-attention (MSA). l = 1,2,3,4,5,6. is the DenseNet output. The final output of this stage is

[0039] Where MSA stands for multi-head self-attention module and MLP stands for multi-layer perceptron: Where W l and b l are the learnable parameters of the model, f and g They are all activation functions, used to introduce nonlinearity, z l is the output of each layer, z L is the final output of the model; LN is layer normalization (LayerNorm): in, x is the input vector with dimension d , μ is the mean of the input vector, σ 2 is the variance of the input vector, γ is the scaling parameter dimension d , β is the offset parameter, with dimension , is a small constant used to prevent the denominator from being zero, and ⊙ represents element-by-element multiplication. for Output after MSA, MLP and LN, and so on.

[0040] 3.4. Training STFormer based on the prediction results of the shared prediction head go through l= The Cross-WSI Transformer output after 6 layers is , sent to the shared prediction head, and the mean square error between the prediction result and the real space transcription result is used as the loss function to update the model parameters.

[0041] In the model training phase, the Style-Aug module and the Cross-WSI Transformer module are combined to jointly learn the mapping relationship between histological images and spatial transcriptome data. An appropriate loss function is used to measure the difference between the predicted gene expression value and the true value. Adam is selected as the optimizer, and the learning rate is adjusted according to the performance of the validation set.

[0042] The experiment was conducted based on the PyTorch deep learning framework and optimized using the Adam optimizer with a learning rate set to 1e-4. When the training loss was stable for 10 epochs, the learning rate was decreased by 10%. The proposed model was trained on an NVIDIA A6000 GPU with a batch size of 64. A total of 50 epochs were trained, and the epoch with the smallest loss during training was finally selected as the trained model. For the training of other models, the batch size was set to 1 for training.

[0043] Step 4: Use the trained STFormer model to perform spatial transcriptome prediction on new histological images.

[0044] Specifically, in order to maintain feature consistency during the testing phase, a shared prediction head was introduced; its function is to migrate the previously learned correlations between images to the prediction stage when there is only one tissue slice and the correlations with other slices cannot be established.

[0045] Through the verification of internal and external datasets, the STFormer model performed well in the spatial transcriptome prediction task. The specific experimental results are as follows: In internal cross validation, the average median Pearson correlation coefficient of STFormer is 0.2493, significantly better than STNet (0.1238), HistoGene (0.1130), and Hist2ST (0.0136).

[0046] On the external dataset “34_Intestine_CRC”, the average median Pearson correlation coefficient of STFormer is 0.1461, which is better than STNet (0.0534), HistoGene (0.1228), and Hist2ST (0.0022), further demonstrating its robustness and applicability on different datasets.

[0047] The trained STFormer model can directly infer spatial transcriptome results based on digital pathology images, thereby improving the ability of digital pathology images to analyze tumor microenvironment.

[0048] This application significantly improves the accuracy and generalization of spatial transcriptome prediction by introducing the Style-Aug module and the Cross-WSI Transformer module. The STFormer model performs well under limited data conditions and can effectively capture the relationship between spots, solving the shortcomings of existing methods in modeling staining differences and spot relationships.

[0049] When training the model, a training set containing digital pathology images and their corresponding spatial transcriptome data is used for end-to-end training; when applying the model, the digital pathology image to be predicted is input, and the corresponding spatial transcriptome prediction result is output through STFormer. The technical advantages of this application are reflected in: 1) Eliminate data bias through the Style-Aug module and significantly improve the model's ability to migrate across data sets; 2) Combining the local feature extraction capability of DenseNet and the global relationship modeling of Transformer to achieve more accurate image-transcriptome mapping prediction; 3) Innovatively introduce the batch-to-batch cross-image Transformer mechanism to effectively mine the spatial relationship features between spots; 4) Even with limited data, it can still maintain high prediction performance and generalization ability.

[0050] The present invention provides an efficient and reliable deep learning tool for analyzing tumor microenvironment based on digital pathological images.

[0051] A device for a method for predicting spatial transcriptome of colorectal cancer based on deep learning, comprising: a processor; a memory on which a computer program that can be run on the processor is stored; wherein the computer program, when executed by the processor, implements the steps of the method for predicting spatial transcriptome of colorectal cancer based on deep learning.

[0052] The memory and the processor are electrically connected directly or indirectly to achieve data transmission or interaction. For example, these elements can be electrically connected to each other through one or more communication buses or signal lines, such as through a bus connection. The memory stores computer-executable instructions for implementing the data access control method, including at least one software function module that can be stored in the memory in the form of software or firmware. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory.

[0053] The memory may be, but is not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), etc. The memory is used to store programs, and the processor executes the programs after receiving the execution instruction.

[0054] The processor may be an integrated circuit chip having the ability to process signals. The above-mentioned processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc. The methods, steps, and logic block diagrams disclosed in the embodiments of the present application may be implemented or executed. The general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc.

[0055] A computer-readable storage medium having a data processing program stored thereon, wherein the data processing program, when executed by a processor, implements the steps of a method for spatial transcriptome prediction of colorectal cancer based on deep learning.

[0056] The embodiments of the present invention are described with reference to flowcharts of methods, terminal devices (systems), and computer program products according to the embodiments of the present invention. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate means for implementing the functions specified in the flowchart and / or.

[0057] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device that implements the functions specified in the flowchart.

[0058] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operation steps are executed on the computer or other programmable terminal device to produce computer-implemented processing, whereby the instructions executed on the computer or other programmable terminal device provide steps for implementing the functions specified in the flowchart.

[0059] The technical principle of the present invention is described above in conjunction with specific embodiments. These descriptions are only for explaining the principle of the present invention and cannot be interpreted as limiting the protection scope of the present invention in any way. Based on the explanations herein, those skilled in the art can associate other specific implementations of the present invention without paying creative labor, and these methods will fall within the protection scope of the claims of the present invention.

Claims

1. A method for spatial transcriptome prediction of colorectal cancer based on deep learning, characterized in that: The following steps are involved: Paired digital pathology images and spatial transcriptome data were collected to form a dataset; Preprocessing of spatial transcriptome data, including dataset partitioning, gene selection, and image processing; Constructing an STFormer model, the STFormer model includes a Style-Aug module for sequentially processing images for performing feature space conversion on input digital pathology images, a densely connected convolutional network DenseNet module for constructing a deep mapping relationship between pathology images and spatial transcriptomes, and an inter-batch cross-full-slice image Transformer module for effectively avoiding overfitting of the model due to data volume limitation through an inter-batch attention mechanism; The trained STFormer model is used to perform spatial transcriptome prediction on the input digital pathology images to obtain the prediction results.

2. According to claim 1, a method for predicting spatial transcriptome of colorectal cancer based on deep learning, characterized in that: The data set division includes: A first data set, comprising whole-slice images of a number of colorectal cancer patients and their corresponding spatial transcriptome data, each whole slice having a plurality of first spots, each first spot corresponding to a gene expression vector, each first spot having a diameter of 65 micrometers, corresponding to 280 pixels on the whole slice, and used for analyzing the expression of 18,000 to 22,000 genes; and The second data set includes a whole slice data, with a total of 2660 second spots, each with a diameter of 61.4-62.4 microns, corresponding to 94 pixels, and is used to analyze the expression of 17943 genes.

3. The method for predicting spatial transcriptome of colorectal cancer based on deep learning according to claim 2, characterized in that: The gene selection comprises the following steps: Perform variance analysis on the gene expression values ​​in the first data set and select the top N genes with the largest variance for subsequent prediction; The first data set is divided using a leave-one-out cross-validation approach so that part of the data is used for training and the rest of the data is used for testing.

4. The method for predicting spatial transcriptome of colorectal cancer based on deep learning according to claim 2, characterized in that: The image processing comprises: The histological images were adjusted to a uniform size using an interpolation method; The uniform size matches the physical diameter of the spatial transcriptome loci in the training set.

5. The method for predicting colorectal cancer spatial transcriptome based on deep learning according to claim 2, characterized in that: The training steps of the STFormer model are as follows: The digital pathological images corresponding to the spots of the spatial transcriptome are input into the Style-Aug module for real-time image enhancement. The image after Style-Aug style enhancement is represented as { x i Aug }( i =1,2,… M ), where M is the number of spots in the training batch, x i Aug It is i An enhanced spot image; The enhanced image is input into the densely connected convolutional network DenseNet module, and the corresponding image features are output. F Aug ={ f i Aug }( i =1,2,… M ); The enhanced image features are input into the batch-to-batch cross-slice image Transformer module to interact between different full slices across batches and capture the association between images at different sites. The features processed by the Transformer module across the entire image slice between batches are input into the shared prediction head. The true value of the spatial transcriptome is used as supervision to construct the regression task loss function and iteratively update the parameters.

6. A method for predicting spatial transcriptome of colorectal cancer based on deep learning according to claim 5, characterized in that: The batch-to-batch cross-image Transformer module includes a multi-head self-attention module: In the formula, X = F Aug ; W Q , W K , W V are all learnable weight matrices; d k is the dimension of the key vector; i Indicates i Head, h is the number of heads, W O is the output weight matrix.

7. The method for predicting spatial transcriptome of colorectal cancer based on deep learning according to claim 5, characterized in that: The batch-by-batch Transformer module builds multi-head self-attention for images in each batch, learning the associations between images while avoiding model overfitting due to data size limitations: In the formula, MSA stands for multi-head self-attention module, MLP stands for multi-layer perceptron, The result output of DenseNet is: l = 1,2,3,4,5,6, LN is the layer normalization operation, for Output after MSA, MLP and LN, and so on.

8. The method for predicting spatial transcriptome of colorectal cancer based on deep learning according to claim 5, characterized in that: The prediction head of the Transformer module across the entire image slice between batches is a shared predictor, which learns the global association of different spots in the training phase; in the inference phase, the shared predictor makes the predicted results stable and generalizable.

9. A device for a method for predicting spatial transcriptome of colorectal cancer based on deep learning, characterized in that: include: processor; a memory having stored thereon a computer program executable on the processor; Wherein, when the computer program is executed by the processor, the steps of a method for predicting spatial transcriptome of colorectal cancer based on deep learning are implemented as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a data processing program, and when the data processing program is executed by the processor, the steps of a colorectal cancer spatial transcriptome prediction method based on deep learning as described in any one of claims 1 to 8 are implemented.

Images