Cell electron microscope image multi-level segmentation and organelle interaction quantification method

Abstract

The invention discloses a cell electron microscope image multi-level segmentation and organelle interaction quantification method, which comprises the following steps: S1, collecting an electron microscope image of an organelle, and preprocessing the electron microscope image; s2, training a Mask2Former model by using the preprocessed electron microscope image; s3, according to the trained Mask2Former model and cell boundary segmentation, carrying out multilevel segmentation on the organelle electron microscope image; s4, organelle interaction quantitative analysis is carried out according to the organelle multi-level segmentation result, multi-level organelle automatic interaction analysis is carried out on the large-range two-dimensional / three-dimensional electron microscope image on the basis of the multi-level segmentation result, efficiency is improved, and human intervention is reduced.

Description

Technical Field

[0001] The present application relates to the fields of image processing and deep learning, and in particular to a method for multi-level segmentation of cell electron microscope images and quantification of organelle interactions. Background Art

[0002] In two-dimensional electron microscope images, organelles are divided into two types according to whether they can be instantiated. One type is instantiable organelles, such as mitochondria, which have regular shapes and obvious instance boundaries and can be distinguished as individual instances. The other type is non-instantiable organelles, such as the endoplasmic reticulum, which do not have regular shapes and cannot be distinguished as individual instances and can only be regarded as a whole.

[0003] Existing organelle segmentation methods mainly use two deep learning models:

[0004] Semantic segmentation models, such as UNet. Semantic segmentation models can separate pixels in an image that belong to the same category, but cannot distinguish individual instances. For example, semantic segmentation can distinguish between organelle type A and organelle type B, but within organelle type A, it cannot further distinguish instances A1, A2, A3, ...

[0005] Instance segmentation models, such as MaskRCNN. Instance segmentation models can identify instantiable organelles and distinguish independent instance individuals. However, instance segmentation models cannot identify non-instantiable organelles.

[0006] Combining instance segmentation models with semantic segmentation models, such as Deepcontact, combines the advantages of both, using semantic segmentation models to identify non-instantiable organelles, while instantiable organelles are handled by instance segmentation models. However, this method requires training two different models for the two organelles, which increases the difficulty of parameter adjustment, and the segmentation results of the two models for the same image will inevitably overlap, increasing the difficulty of post-processing.

[0007] Currently, in the field of cell electron microscopy image processing, there is still a lack of segmentation and organelle interaction analysis methods for large-scale electron microscopy images (i.e., resolution above 100 million pixels).

[0008] The main method for 3D electron microscope image segmentation is 3D semantic segmentation based on 3D CNN, such as 3DUNet. This method can utilize the depth information in space and has high segmentation accuracy when the annotations are sufficient. However, 3D CNN has a high computational complexity, resulting in a low inference speed of the model. On the other hand, complete 3D annotations in existing electron microscope images are relatively scarce, which restricts the performance of 3D CNN models to a certain extent. Summary of the invention

[0009] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method for multi-level segmentation of cell electron microscope images and quantification of organelle interactions, which solves the problem that the existing algorithm can only analyze electron microscope images with a small local field of view, and performs small-scale local analysis by manually screening local areas of interest in biological tissues, which easily introduces personal bias and leads to incomplete analysis.

[0010] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a method for multi-level segmentation of cell electron microscope images and quantification of organelle interactions, comprising:

[0011] S1. Collect electron microscope images of organelles and perform preprocessing;

[0012] S2, use the preprocessed electron microscope images to train the Mask2Former model;

[0013] S3, based on the trained Mask2Former model and cell boundary segmentation, perform multi-level segmentation of organelle electron microscopy images;

[0014] S4. Based on the results of multi-level organelle segmentation, quantitative analysis of organelle interactions is performed.

[0015] Further: S101 includes:

[0016] S11, pixel-level annotation of organelles and cell boundaries in electron microscope images;

[0017] S12. Convert the annotation information into COCO format.

[0018] Further: in S12, the Mask2Former model includes: an image encoder, a pixel decoder, and an object decoder;

[0019] The image encoder is used to extract features of the input image in stages;

[0020] The pixel decoder is used to receive the features of the input image in stages and generate a feature pyramid to extract features from the pixels to obtain a high-resolution feature map.

[0021] The object decoder is used to receive high-resolution feature maps and perform learnable embedding, and replaces the standard cross attention mechanism with a mask attention mechanism to obtain segmentation masks and classification categories.

[0022] Further: In the pixel decoder, each layer includes a deformable attention module and a feedforward network, and the deformable attention extracted by the deformable attention module is represented as:

[0023]

[0024] Among them, MSAttn(.) represents deformable attention, z q To query features, represents a two-dimensional reference point, l=1,...,L is the pixel decoder layer index, L is the number of pixel decoder layers, m represents the attention head, k represents the sampled point, K represents the total number of sampled points, Δp mlqk and A mlqk denote the offset and attention weight of the kth sampling point of the mth attention head in the lth level feature map, Φ l (.) means rescaling the input resolution to the size of the l-th level feature map, f l represents the l-th level feature map of the input feature map, W m and W m ' are all weight matrices.

[0025] Further: In the pixel decoder, the calculation formula of the mask attention mechanism is:

[0026]

[0027] Among them, X l is the learnable embedding of the input of the lth layer of the object decoder, softmax(.) represents the activation function, K l and V l are all image features from the pixel decoder after linear transformation. The superscript T represents matrix transposition. l is the learnable embedding X of the input to the l-1 layer of the object decoder l-1 The query embedding obtained after linear transformation is l-1 is the attention mask.

[0028] Further: For two-dimensional organelle electron microscopy images, S3 includes:

[0029] S31, extracting the region of interest in the two-dimensional electron microscope image of the organelle and encapsulating it as an instance object;

[0030] The instance objects include cell bounding boxes, segmentation masks, and cell types;

[0031] S32, inputting the two-dimensional organelle electron microscope image into the trained Mask2Former model to perform two-dimensional organelle panoramic segmentation;

[0032] S33, cropping the two-dimensional organelle panorama segmentation result according to the cell bounding box in the instance object to obtain the bounding box range of the two-dimensional organelle;

[0033] S34, bitwise multiplying the segmentation mask in the instance object and the bounding box range of the organelle to obtain a two-dimensional organelle multi-level segmentation result.

[0034] Further: For two-dimensional organelle electron microscopy images, S4 includes:

[0035] S401, selecting organelle A and organelle B from the organelle multi-level segmentation results, and obtaining segmentation masks of organelle A and organelle B;

[0036] S402, extracting edge information of organelle A and organelle B respectively according to the segmentation masks of organelle A and organelle B;

[0037] S403, according to the edge information of organelle A and organelle B, using distance transformation or pixel traversal, calculate the shortest distance between organelle A and organelle B, and generate a distance map and a contact map between organelle A and organelle B;

[0038] Among them, the distance map is used to record the distance from each pixel in organelle A to the nearest organelle B to form a global distance distribution;

[0039] The contact map is used to mark the contact areas between organelle A and organelle B and to quantify the distribution characteristics between contact points;

[0040] S404, performing distance characteristic calculation, contact characteristic calculation and organelle geometric property calculation according to the distance map and contact map between organelle A and organelle B;

[0041] S405 , structuring the distance characteristic calculation results, the contact characteristic calculation results, and the organelle geometric property calculation results into a data format as a quantitative analysis result of the interaction between organelle A and organelle B.

[0042] Further: For 3D organelle electron microscopy images, S3 includes:

[0043] S311, extracting the region of interest in the electron microscope image section of the three-dimensional organelle layer by layer;

[0044] S312, using a tracking algorithm to perform z-axis matching on the region of interest in each section layer to establish a three-dimensional instance object;

[0045] The 3D instance object includes 3D cell bounding box, 3D segmentation mask and cell type;

[0046] S312, inputting the cross-section of the 3D organelle electron microscope image layer by layer into the trained Mask2Former model to perform 3D organelle panoramic segmentation;

[0047] S313, using the 3D cell bounding box and 3D segmentation mask in the 3D instance object to cut out the region of interest from the 3D organelle panoramic segmentation result, to obtain a multi-level segmentation result of the 3D organelle electron microscope image.

[0048] Further: For 3D organelle electron microscopy images, S4 includes:

[0049] S411, performing block processing on the multi-level segmentation results of the three-dimensional organelle electron microscope image;

[0050] S412, extracting the surface voxels of each organelle in each block;

[0051] S413, determining the contact area and contact points between the surface voxels of the organelle;

[0052] S414. Summarize the contact points, contact areas and organelle areas of all blocks, calculate the contact ratio, and complete the quantitative analysis of the interaction of three-dimensional organelle electron microscopy images.

[0053] The beneficial effects of the present invention are:

[0054] 1. This protocol can simultaneously perform quantitative analysis of organelle interactions at the single cell level / whole tissue level in a large range of two-dimensional / three-dimensional electron microscopy images;

[0055] 2. Based on the multi-level segmentation results, all regions of interest are automatically screened out in the image to improve efficiency and reduce human intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flow chart of the multi-level segmentation of cell electron microscopy images and the quantification of organelle interactions.

[0057] Figure 2 Schematic diagram of the multi-level segmentation and interaction quantification process of large-scale two-dimensional electron microscopy images.

[0058] Figure 3 Schematic diagram of the multi-level segmentation and interaction quantification process of three-dimensional electron microscopy images. DETAILED DESCRIPTION

[0059] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0060] like Figure 1 As shown, in one embodiment of the present invention, a method for multi-level segmentation of cell electron microscope images and quantification of organelle interactions is provided, comprising:

[0061] S1. Collect electron microscope images of organelles and perform preprocessing;

[0062] S2, use the preprocessed electron microscope images to train the Mask2Former model;

[0063] S3, based on the trained Mask2Former model and cell boundary segmentation, perform multi-level segmentation of organelle electron microscopy images;

[0064] S4. Based on the results of multi-level organelle segmentation, quantitative analysis of organelle interactions is performed.

[0065] Specifically, S1 includes:

[0066] S11, pixel-level annotation of organelles and cell boundaries in electron microscope images;

[0067] S12. Convert the annotation information into COCO format.

[0068] In S12, the Mask2Former model is a typical "encoder-decoder" structure, including: image encoder, pixel decoder and object decoder;

[0069] The image encoder is used to extract features of the input image in stages;

[0070] The pixel decoder is used to receive the features of the input image in stages and generate a feature pyramid to extract features from the pixels to obtain a high-resolution feature map.

[0071] The object decoder is used to receive high-resolution feature maps and perform learnable embedding, and replaces the standard cross attention mechanism with a mask attention mechanism to obtain segmentation masks and classification categories.

[0072] Specifically: For an image of size H0×W0, it is planned to use an image encoder such as ResNet and swin-transformer to extract image features.

[0073] The feature maps generated by the last L stages of the image encoder are input into the pixel decoder to generate a feature pyramid. Each layer of the pixel decoder consists of a deformable attention module and a feedforward network. The advantage of deformable attention is that it can focus on the sampling points around a specific pixel. Compared with the standard attention mechanism, it has lower computational cost and is easier to converge.

[0074] For a given input feature map f, query feature z q and 2D reference points The deformable attention is calculated by the following formula:

[0075]

[0076] Where MSAttn(.) represents deformable attention, l=1,...,L is the pixel decoder layer index, L is the number of pixel decoder layers, m represents the attention head, k represents the sampled point, K represents the total number of sampled points, Δp mlqk and A mlqk denote the offset and attention weight of the kth sampling point of the mth attention head in the lth level feature map, Φ l (.) means rescaling the input resolution to the size of the l-th level feature map, f l represents the l-th level feature map of the input feature map, W m and W′ m are weight matrices, W′ m Used to perform linear transformation on the input feature map, W m Used to linearly transform the output features of deformable attention.

[0077] The obtained high-resolution feature map is input into the object decoder to train the learnable embedding to obtain the segmentation mask and classification category. The mask attention mechanism is used in the object decoder to replace the standard cross attention mechanism, so that the attention is focused on the predicted mask area and the model convergence speed is accelerated. The calculation formula of the mask attention mechanism is as follows:

[0078]

[0079] Among them, X l is the learnable embedding of the input of the lth layer of the object decoder, softmax(.) represents the activation function, K l and V l are all image features from the pixel decoder after linear transformation. The superscript T represents matrix transposition. l is the learnable embedding X of the input to the l-1 layer of the object decoder l-1 The query embedding obtained after linear transformation is l-1 is the attention mask, and its corresponding value at the (x, y) position on the feature map is calculated as follows:

[0080]

[0081] Among them, F l-1 (x,y)=1 is the binarized output of the mask prediction of the lth layer in the object decoder.

[0082] like Figure 2 As shown, for two-dimensional organelle electron microscopy images, S3 includes:

[0083] S31, extracting the region of interest in the two-dimensional electron microscope image of the organelle and encapsulating it as an instance object;

[0084] For the segmentation of regions of interest (i.e., cell boundary segmentation), since it is an instance segmentation task (i.e., only instantiable individual cells are segmented), the open source sahi framework is used to identify and extract cell regions in the image window by window in the form of sliding windows, and then post-processed by nms (non-maximum suppression algorithm) to remove duplicate and redundant instances. Then, for instances across windows, the IoU threshold is used for merging. If the IoU of adjacent instances is greater than the threshold of 0.4, the two are merged, otherwise they are not merged. All extracted regions of interest are saved in a set of encapsulated instance objects. The individual objects in this set contain the bounding box, segmentation mask, confidence, and cell type of a single region of interest. By setting the confidence and bounding box size thresholds, low confidence and areas with too small an area are filtered out to improve the efficiency of subsequent processing.

[0085] In particular, instance objects include cell bounding boxes, segmentation masks, and cell types;

[0086] S32, inputting the two-dimensional organelle electron microscope image into the trained Mask2Former model to perform two-dimensional organelle panoramic segmentation;

[0087] S33, cropping the two-dimensional organelle panorama segmentation result according to the cell bounding box in the instance object to obtain the bounding box range of the two-dimensional organelle;

[0088] S34, bitwise multiplying the segmentation mask in the instance object and the bounding box range of the organelle to obtain a two-dimensional organelle multi-level segmentation result.

[0089] For 2D organelle electron microscopy images, S4 includes:

[0090] S401, selecting organelle A and organelle B from the organelle multi-level segmentation results, and obtaining segmentation masks of organelle A and organelle B;

[0091] S402, extracting edge information of organelle A and organelle B respectively according to the segmentation masks of organelle A and organelle B;

[0092] S403, according to the edge information of organelle A and organelle B, using distance transformation or pixel traversal, calculate the shortest distance between organelle A and organelle B, and generate a distance map and a contact map between organelle A and organelle B;

[0093] Among them, the distance map is used to record the distance from each pixel in organelle A to the nearest organelle B to form a global distance distribution;

[0094] The contact map is used to mark the contact areas between organelle A and organelle B and to quantify the distribution characteristics between contact points;

[0095] S404, performing distance characteristic calculation, contact characteristic calculation and organelle geometric property calculation according to the distance map and contact map between organelle A and organelle B;

[0096] S405 , structuring the distance characteristic calculation results, the contact characteristic calculation results, and the organelle geometric property calculation results into a data format as a quantitative analysis result of the interaction between organelle A and organelle B.

[0097] Taking the instantiable organelle mitochondria and the non-instantiable organelle endoplasmic reticulum as examples, the distance map: first traverse each pixel in the endoplasmic reticulum area and check whether its neighborhood within the specified range contains mitochondrial pixels. If there are mitochondrial pixels in the neighborhood, calculate the Euclidean distance and record the minimum distance as the nearest distance value of the point. The minimum distance values ​​of all points are summarized to form a distance map, which is used to quantify the spatial relationship between the endoplasmic reticulum and mitochondria.

[0098] Contact Map: First, traverse each pixel in the mitochondrial region and check whether there is an endoplasmic reticulum pixel in its neighborhood. If an endoplasmic reticulum pixel is found and the distance is less than or equal to the threshold, the mitochondrial pixel is marked as a contact point and its distance value is recorded in the contact map. If a mitochondrial pixel completely overlaps with an endoplasmic reticulum pixel (the distance is 0), it is directly marked as a contact point.

[0099] like Figure 3 As shown, for 3D organelle electron microscopy images, S3 includes:

[0100] S311, extracting the region of interest in the electron microscope image section of the three-dimensional organelle layer by layer;

[0101] S312, using a tracking algorithm to perform z-axis matching on the region of interest in each section layer to establish a three-dimensional instance object;

[0102] To solve the 3D ROI extraction problem, this application relies on an intuitive idea: in adjacent sections, if the 2D instance in the current section does not match the instance in the previous section, it is considered to be a new 3D instance. This problem can be expressed as a weighted bipartite graph matching problem. Instances on different sections are regarded as nodes in a bipartite graph, and the intersection over union (IoU) between instance masks is used as the weight of the edge. The Hungarian algorithm is selected as the bipartite graph matching algorithm. First, the IoU between instances is calculated, and an IoU matrix is ​​constructed as the cost matrix of the Hungarian matching algorithm. If the IoU between a pair of instances is lower than the set threshold, it is forced to be set to 0. Then, the Hungarian matching algorithm is applied to select the optimal matching solution. Since the goal is to select the solution with the maximum cost or the highest IoU, which is contrary to the original goal of the algorithm, we take the negative of the IoU matrix for calculation.

[0103] The 3D instance object includes 3D cell bounding box, 3D segmentation mask and cell type;

[0104] S312, inputting the cross-section of the 3D organelle electron microscope image layer by layer into the trained Mask2Former model to perform 3D organelle panoramic segmentation;

[0105] S313, using the 3D cell bounding box and 3D segmentation mask in the 3D instance object to cut out the region of interest from the 3D organelle panoramic segmentation result, to obtain a multi-level segmentation result of the 3D organelle electron microscope image.

[0106] For 3D organelle electron microscopy images, S4 includes:

[0107] S411, performing block processing on the multi-level segmentation results of the three-dimensional organelle electron microscope image;

[0108] S412, extracting the surface voxels of each organelle in each block;

[0109] The specific method is: perform an erosion operation (binary_erosion) on the voxels of each instance, and calculate the difference before and after the erosion to extract the surface voxels;

[0110] S413, determining the contact area and contact points between the surface voxels of the organelle;

[0111] The specific method is: Use cKDTree in the Python scientific computing library scipy to build a KD tree and calculate the distance between the surface voxels of the two organelles. For each distance value, determine whether the contact condition is met (i.e., whether it is less than or equal to the given maximum distance max_distance), and record the contact point;

[0112] S414. Summarize the contact points, contact areas and organelle areas of all blocks, calculate the contact ratio, and complete the quantitative analysis of the interaction of three-dimensional organelle electron microscopy images.

[0113] In one embodiment of the present invention, a comparison of the quantitative experimental results of the method of the present application and UNet in organelle segmentation is provided (as shown in Table 1) and a comparison of the quantitative experimental results of the present application in MaskRCNN in region of interest segmentation is provided (as shown in Table 2);

[0114] Table 1 Quantitative results of comparative experiments between this application and UNet in organelle segmentation

[0115]

[0116] Table 2 Quantitative results of comparative experiments between this application and MaskRCNN in region of interest segmentation

[0117]

[0118] The above comparison shows that the system has excellent segmentation effect on organelles, leading in multiple indicators on private datasets, and can simultaneously perform quantitative analysis of organelle interactions at the single-cell level / whole tissue level in a wide range of two-dimensional / three-dimensional electron microscopy images.

[0119] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for multi-level segmentation of cell electron microscope images and quantification of organelle interactions, characterized in that: include: S1. Collect electron microscope images of organelles and perform preprocessing; S2, use the preprocessed electron microscope images to train the Mask2Former model; S3, based on the trained Mask2Former model and cell boundary segmentation, perform multi-level segmentation of organelle electron microscopy images; S4. Based on the results of multi-level organelle segmentation, quantitative analysis of organelle interactions is performed.

2. The method for multi-level segmentation of cell electron microscope images and quantification of organelle interactions according to claim 1, characterized in that: S1 includes: S11, pixel-level annotation of organelles and cell boundaries in electron microscope images; S12. Convert the annotation information into COCO format.

3. The method for multi-level segmentation of cell electron microscope images and quantification of organelle interactions according to claim 1, characterized in that: In S12, the Mask2Former model includes: image encoder, pixel decoder and object decoder; The image encoder is used to extract features of the input image in stages; The pixel decoder is used to receive the features of the input image in stages and generate a feature pyramid to extract features from the pixels to obtain a high-resolution feature map. The object decoder is used to receive high-resolution feature maps and perform learnable embedding, and replaces the standard cross attention mechanism with a mask attention mechanism to obtain segmentation masks and classification categories.

4. The method for multi-level segmentation of cell electron microscope images and quantification of organelle interactions according to claim 3, characterized in that: In the pixel decoder, each layer includes a deformable attention module and a feedforward network. The deformable attention extracted by the deformable attention module is represented as: Among them, MSAttn(.) represents deformable attention, z q To query features, represents a two-dimensional reference point, l=1,...,L is the pixel decoder layer index, L is the number of pixel decoder layers, m represents the attention head, k represents the sampled point, K represents the total number of sampled points, Δp mlqk and A mlqk denote the offset and attention weight of the kth sampling point of the mth attention head in the lth level feature map, Φ l (.) means rescaling the input resolution to the size of the l-th level feature map, f l represents the l-th level feature map of the input feature map, W m and W m ' are all weight matrices.

5. The method for multi-level segmentation of cell electron microscope images and quantification of organelle interactions according to claim 3, characterized in that: In the pixel decoder, the calculation formula of the mask attention mechanism is: Among them, X l is the learnable embedding of the input of the lth layer of the object decoder, softmax(.) represents the activation function, K l and V l are all image features from the pixel decoder after linear transformation. The superscript T represents matrix transposition. l is the learnable embedding X of the input to the l-1 layer of the object decoder l-1 The query embedding obtained after linear transformation is l-1 is the attention mask.

6. The method for multi-level segmentation of cell electron microscope images and quantification of organelle interactions according to claim 1, characterized in that: For 2D organelle electron microscopy images, S3 includes: S31, extracting the region of interest in the two-dimensional electron microscope image of the organelle and encapsulating it as an instance object; The instance objects include cell bounding boxes, segmentation masks, and cell types; S32, inputting the two-dimensional organelle electron microscope image into the trained Mask2Former model to perform two-dimensional organelle panoramic segmentation; S33, cropping the two-dimensional organelle panorama segmentation result according to the cell bounding box in the instance object to obtain the bounding box range of the two-dimensional organelle; S34, bitwise multiplying the segmentation mask in the instance object and the bounding box range of the organelle to obtain a two-dimensional organelle multi-level segmentation result.

7. The method for multi-level segmentation of cell electron microscope images and quantification of organelle interactions according to claim 6, characterized in that: For 2D organelle electron microscopy images, S4 includes: S401, selecting organelle A and organelle B from the organelle multi-level segmentation results, and obtaining segmentation masks of organelle A and organelle B; S402, extracting edge information of organelle A and organelle B respectively according to the segmentation masks of organelle A and organelle B; S403, according to the edge information of organelle A and organelle B, using distance transformation or pixel traversal, calculate the shortest distance between organelle A and organelle B, and generate a distance map and a contact map between organelle A and organelle B; Among them, the distance map is used to record the distance from each pixel in organelle A to the nearest organelle B to form a global distance distribution; The contact map is used to mark the contact areas between organelle A and organelle B and to quantify the distribution characteristics between contact points; S404, performing distance characteristic calculation, contact characteristic calculation and organelle geometric property calculation according to the distance map and contact map between organelle A and organelle B; S405 , structuring the distance characteristic calculation results, the contact characteristic calculation results, and the organelle geometric property calculation results into a data format as a quantitative analysis result of the interaction between organelle A and organelle B.

8. The method for multi-level segmentation of cell electron microscope images and quantification of organelle interactions according to claim 1, characterized in that: For 3D organelle electron microscopy images, S3 includes: S311, extracting the region of interest in the electron microscope image section of the three-dimensional organelle layer by layer; S312, using a tracking algorithm to perform z-axis matching on the region of interest in each section layer to establish a three-dimensional instance object; The 3D instance object includes 3D cell bounding box, 3D segmentation mask and cell type; S312, inputting the cross-section of the 3D organelle electron microscope image layer by layer into the trained Mask2Former model to perform 3D organelle panoramic segmentation; S313, using the 3D cell bounding box and 3D segmentation mask in the 3D instance object to cut out the region of interest from the 3D organelle panoramic segmentation result, to obtain a multi-level segmentation result of the 3D organelle electron microscope image.

9. The method for multi-level segmentation of cell electron microscope images and quantification of organelle interactions according to claim 8, characterized in that: For 3D organelle electron microscopy images, S4 includes: S411, performing block processing on the multi-level segmentation results of the three-dimensional organelle electron microscope image; S412, extracting the surface voxels of each organelle in each block; S413, determining the contact area and contact points between the surface voxels of the organelle; S414. Summarize the contact points, contact areas and organelle areas of all blocks, calculate the contact ratio, and complete the quantitative analysis of the interaction of three-dimensional organelle electron microscopy images.

Images