A noise-resistant, lightweight cryo-electron microscopy particle selection method and system

Abstract

This invention relates to the field of cryo-electron microscopy (cryo-EM) technology and discloses a noise-resistant, lightweight cryo-EM particle selection method and system, comprising the following steps: S1. Obtaining a serialized file of a cryo-EM image; S2. Preprocessing the serialized file of the cryo-EM image; S3. Constructing and training a lightweight cryo-EM particle selection model based on the YOLOX model using the preprocessed data; S4. Performing cryo-EM particle selection using the trained lightweight cryo-EM particle selection model. This method solves the problem of weak ability to distinguish noise from particles to be selected in existing technologies, and features convenient training, simple structure, and high accuracy.

Description

Technical Field

[0001] This invention relates to the field of cryo-electron microscopy, and more specifically, to a noise-resistant, lightweight cryo-electron microscopy particle selection method and system. Background Technology

[0002] Cryo-electron microscopy (cryo-EM) is a commonly used tool for determining protein structure. A single cryo-EM image can contain hundreds of protein particles with random orientations. By selecting particles from the cryo-EM image and then digitally aligning, classifying, averaging, and back-projecting them, the three-dimensional structure of the protein can be obtained. Therefore, the efficiency of protein particle selection and the completeness and accuracy of particle identification play a crucial role in three-dimensional reconstruction.

[0003] Initially, with limited knowledge of protein particle shapes and the distribution of projected views, manual particle selection was the only option. However, manual selection allows for focusing on more familiar and easily identifiable particles. Currently, methods based on traditional computer vision have been developed, including algorithms based on edge segmentation and image segmentation, Gaussian difference, and simulated annealing. With the development of deep learning, deep networks can extract more abstract features from images and have brought significant convenience to cryo-electron microscopy image research. These include methods for classifying image regions using sliding windows; automatic particle picking methods based on Faster-RCNN networks; and target selection methods based on YOLO. However, for noisy cryo-electron microscopy images, traditional computer vision methods struggle to accurately select the correct protein particles for 3D reconstruction. Traditional deep learning methods require complex pre-training steps and cannot directly achieve end-to-end network output, while their accuracy is significantly reduced due to noise.

[0004] Existing technology includes a method, apparatus, and electronic device for selecting particles in cryo-electron microscopy based on deep learning, relating to the field of image recognition technology. The method includes: acquiring a target cryo-electron microscopy image; inputting the target cryo-electron microscopy image into a particle segmentation model, outputting a segmented image representing a category, including background category or particle category; and performing average pooling on the segmented image. Non-maximum suppression operation is performed to determine the operation result. If the operation result includes at least one local maximum, the center coordinates of the particles in the segmented image are determined based on the position coordinates of at least one local maximum.

[0005] However, existing technologies have the problem of weak ability to distinguish between noise and particles to be selected. Therefore, how to invent a noise-resistant, lightweight cryo-electron microscopy particle selection method is an urgent problem to be solved in this technical field. Summary of the Invention

[0006] To address the problem of the weak ability of existing technologies to distinguish between noise and particles to be selected, this invention provides a noise-resistant, lightweight cryo-electron microscopy particle selection method, which is characterized by convenient training, simple structure, and high accuracy.

[0007] To achieve the above-mentioned objectives of this invention, the technical solution adopted is as follows:

[0008] A noise-resistant, lightweight cryo-electron microscopy particle selection method includes the following steps:

[0009] S1. Obtain the serialized file of the cryo-electron microscopy image;

[0010] S2. Preprocess the serialized files of cryo-electron microscopy images;

[0011] S3. Construct and train a lightweight cryo-electron microscopy particle selection model based on the YOLOX model using preprocessed data; the feature extraction module of the lightweight cryo-electron microscopy particle selection model based on the YOLOX model adopts the residual module B-ResBlockX module; the B-ResBlockX module includes a first image processing branch, a second image processing branch, a Concat layer, and a 3×3 convolutional layer; the number of channels in the first image processing branch and the second image processing branch are half the number of input channels respectively; the first image processing branch is configured with a 1×1 convolutional layer, a 3×3 convolutional layer, and a ResXBlock module connected in sequence to obtain the first image features; the second image processing branch is configured with a 1×1 convolutional layer to extract the second image features; B-ResBlockX concatenates the first image features and the second image features through the Concat layer, and finally extracts the features through the 3×3 convolutional layer to output the final extracted features;

[0012] S4. Perform cryo-electron microscopy particle selection using a trained lightweight cryo-electron microscopy particle selection model.

[0013] Preferably, in step S2, the images used for training are preprocessed, specifically as follows:

[0014] S201. Read the serialized file of each cryo-electron microscopy image and convert it into several visualization images;

[0015] S202. Several visualization images converted from each cryo-electron microscopy image are overlaid, and then drift correction is performed;

[0016] S203. Perform image denoising on the corrected image;

[0017] S204. The denoised image is segmented into fragments;

[0018] S205. Remove the bounding box information of the cut segments and organize the final segment information into training set, test set, and validation set.

[0019] Furthermore, the 1×1 convolutional layer and 3×3 convolutional layer in the first image processing branch of the B-ResBlockX module are used to extract features sequentially based on the downsampling mode, refining semantic information; the ResXBlock module is used to supplement the information lost after the convolutional layers extract features.

[0020] Furthermore, the specific process of the ResXBlock module is as follows:

[0021] After convolution by the 1×1 convolutional layer in the first image processing branch, the feature map is uniformly divided into s feature subsets, x i Let K represent the i-th feature subset. i () represents the 3×3 convolutional layer of the first image processing branch, denoted by y. i K represents i The output of () is used; when passing through the 3×3 convolutional layer, it receives the feature information from the previous feature subset, thus finally outputting a feature map containing information from different receptive fields:

[0022]

[0023] The ResXBlock module includes several filters, each with the same channels, and the filters are connected in a residual style. The output features of the previous group are concatenated with the input features of another group and sent to the next group of filters. This process is repeated several times until the convolutional layers have finished extracting features, thereby obtaining information from different dimensions of loss.

[0024] Furthermore, the feature fusion module and output of the lightweight cryo-electron microscopy particle selection model based on the YOLOX model are combined into a FastHead module; the FastHead module includes a projection layer, three dilated convolutional modules, and an output.

[0025] The projection layer includes a 1×1 convolutional layer and a 3×3 convolutional layer connected in sequence. The 1×1 convolutional layer is used to reduce the channel dimension, and the 3×3 convolutional layer is used to refine the semantic context of the image after the channel dimension is reduced.

[0026] The three dilated convolutional modules each include a first 1×1 convolutional layer, a 3×3 convolutional layer, and a second 1×1 convolutional layer connected in sequence.

[0027] The output terminal is a 32x downsampling channel.

[0028] Furthermore, the three dilated convolutional modules are stacked with three consecutive dilated residual blocks with different dilation rates of 2, 4, and 8, respectively. As the dilation rate increases, the size of the receptive field also increases to generate output features with multiple receptive fields, covering the scale of all objects, thereby achieving a multi-scale fusion effect.

[0029] Furthermore, when training the lightweight cryo-electron microscopy particle selection model based on the YOLOX model, the position loss function used is Loss. diou :

[0030]

[0031] in, and These represent the predicted bounding box and the ground truth bounding box, respectively. ( , () represents the Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box; It represents the diagonal distance of the smallest closure region that can simultaneously contain both the predicted bounding box and the ground truth bounding box.

[0032] A noise-resistant, lightweight cryo-electron microscopy particle selection system includes an image acquisition module, an image preprocessing module, a model building module, and a particle selection module.

[0033] The image acquisition module is used to obtain serialized files of cryo-electron microscopy images;

[0034] The image preprocessing module is used to preprocess the serialized files of cryo-electron microscopy images;

[0035] The model building module is used to construct and train a lightweight cryo-electron microscopy particle selection model based on the YOLOX model using preprocessed data. The feature extraction module of the lightweight cryo-electron microscopy particle selection model based on the YOLOX model adopts the residual module B-ResBlockX module. The B-ResBlockX module includes a first image processing branch, a second image processing branch, a Concat layer, and a 3×3 convolutional layer. The number of channels in the first and second image processing branches are half the number of input channels. The first image processing branch is configured with a 1×1 convolutional layer, a 3×3 convolutional layer, and a ResXBlock module connected in sequence to obtain the first image features. The second image processing branch is configured with a 1×1 convolutional layer to extract the second image features.

[0036] B-ResBlockX concatenates the features of the first and second images through a Concat layer, and finally extracts the features through a 3×3 convolutional layer to output the final extracted features.

[0037] The particle selection module is used to perform cryo-electron microscopy particle selection using a trained lightweight cryo-electron microscopy particle selection model.

[0038] The beneficial effects of this invention are as follows:

[0039] This invention discloses a noise-resistant, lightweight cryo-electron microscopy particle selection method. By obtaining a serialized file of a cryo-electron microscopy image, a lightweight cryo-electron microscopy particle selection model based on the YOLOX model is constructed and trained using preprocessed data. This invention solves the problem of the weak ability of existing technologies to distinguish between noise and particles to be selected, and provides a noise-resistant, lightweight cryo-electron microscopy particle selection method with the characteristics of convenient training, simple structure, and high accuracy. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the noise-resistant, lightweight cryo-electron microscopy particle selection method of the present invention.

[0041] Figure 2 This is a schematic diagram of the complete process flow of the noise-resistant lightweight cryo-electron microscopy particle selection method of the present invention.

[0042] Figure 3 This is a schematic diagram of the preprocessing of the noise-resistant, lightweight cryo-electron microscopy particle selection method of the present invention.

[0043] Figure 4 This is a schematic diagram illustrating the specific implementation of the B-ResBlockX module in the noise-resistant, lightweight cryo-electron microscopy particle selection method of the present invention.

[0044] Figure 5 This is a schematic diagram illustrating the specific implementation of the FastHead module of the noise-resistant lightweight cryo-electron microscopy particle selection method of the present invention.

[0045] Figure 6 This is a schematic diagram comparing the noise-resistant, lightweight cryo-electron microscopy particle selection method of this invention with mainstream target detection methods.

[0046] Figure 7 This is a schematic diagram comparing the noise-resistant, lightweight cryo-electron microscopy particle selection method of this invention with mainstream cryo-electron microscopy particle selection methods. Detailed Implementation

[0047] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0048] Example 1

[0049] like Figure 1 As shown, a noise-resistant, lightweight cryo-electron microscopy particle selection method includes the following steps:

[0050] S1. Obtain the serialized file of the cryo-electron microscopy image;

[0051] S2. Preprocess the serialized files of cryo-electron microscopy images;

[0052] S3. Construct and train a lightweight cryo-electron microscopy particle selection model based on the YOLOX model using preprocessed data;

[0053] S4. Perform cryo-electron microscopy particle selection using a trained lightweight cryo-electron microscopy particle selection model.

[0054] Example 2

[0055] like Figure 2 As shown, a noise-resistant, lightweight cryo-electron microscopy particle selection method includes the following steps:

[0056] S1. Obtain the serialized file of the cryo-electron microscopy image;

[0057] S2. Preprocess the serialized files of cryo-electron microscopy images;

[0058] S3. Construct and train a lightweight cryo-electron microscopy particle selection model based on the YOLOX model using preprocessed data;

[0059] S4. Perform cryo-electron microscopy particle selection using a trained lightweight cryo-electron microscopy particle selection model.

[0060] In this embodiment, the lightweight cryo-electron microscopy particle selection model is named AF-YOLOX.

[0061] In one specific embodiment, step S2 involves preprocessing the images used for training, specifically as follows:

[0062] S201. Read the serialized file of each cryo-electron microscopy image and convert it into several visualization images;

[0063] S202. Several visualization images converted from each cryo-electron microscopy image are overlaid, and then drift correction is performed;

[0064] S203. Perform image denoising on the corrected image;

[0065] S204. The denoised image is segmented into fragments;

[0066] S205. Remove the bounding box information of the cut segments and organize the final segment information into training set, test set, and validation set.

[0067] In this embodiment, as Figure 3As shown, firstly, the relevant dataset is downloaded from the publicly available cryo-electron microscopy dataset EMPIAR. The downloaded dataset is a serialized file, which needs to be converted into images that can be used for training through frame addition, drift correction, and denoising steps. Then, the images are randomly cropped and rotated, and divided into training set, validation set, and test set. The training set and validation set are then put into the network proposed in this invention for training.

[0068] In this embodiment, the feature extraction module is equipped with three downsampling modules, and each downsampling is 8 times, 16 times, and 32 times downsampling respectively.

[0069] In one specific embodiment, such as Figure 4 As shown, the feature extraction module of the lightweight cryo-electron microscopy particle selection model based on the YOLOX model adopts the novel residual module B-ResBlockX module.

[0070] The B-ResBlockX module includes a first image processing branch, a second image processing branch, a Concat layer, and a 3×3 convolutional layer; the number of channels in the first and second image processing branches are each half the number of input channels;

[0071] The first image processing branch is configured with a 1×1 convolutional layer, a 3×3 convolutional layer, and a ResXBlock module connected in sequence to obtain the first image features;

[0072] The second image processing branch has a 1×1 convolutional layer for extracting features from the second image;

[0073] B-ResBlockX concatenates the features of the first and second images using a Concat layer, and finally extracts the features using a 3×3 convolutional layer to output the final extracted features.

[0074] In one specific embodiment, the 1×1 convolutional layer and the 3×3 convolutional layer of the first image processing branch of the B-ResBlockX module are used to extract features sequentially based on the downsampling mode to refine semantic information; the ResXBlock module is used to supplement the information lost after the convolutional layers extract features.

[0075] In one specific embodiment, in the first image processing branch of the ResXBlock module, the specific process of using a 1×1 convolutional layer and a 3×3 convolutional layer to extract features based on the downsampling mode is as follows:

[0076] After convolution by the 1×1 convolutional layer in the first image processing branch, the feature map is uniformly divided into s feature subsets, x i Let K represent the i-th feature subset. i() represents the 3×3 convolutional layer of the first image processing branch, denoted by y. i K represents i The output of () is used; when passing through the 3×3 convolutional layer, it receives the feature information from the previous feature subset, thus finally outputting a feature map containing information from different receptive fields:

[0077] .

[0078] In one specific embodiment, the ResXBlock module includes several filters, each with the same channels, and the filters are connected in a residual style. The output features of the previous group are concatenated with the input features of another group and sent to the next group of filters. This process is repeated several times until the convolutional layer has finished processing the extracted features, thereby obtaining information of different dimensions of loss.

[0079] In this embodiment, the B-ResBlockX module is a novel residual module. It uses grouped filters to generate multiple feature channels, rather than simply superimposing large-scale dimensional information through skip connections. This improves the fusion and extraction capabilities of multi-scale features at a finer granular level, capturing more detailed features and thus enabling the network to have greater discriminative power when extracting noise and target particle features. Ultimately, this enhances the backbone network's extraction capabilities and improves accuracy.

[0080] In this embodiment, as Figure 5 As shown, the feature fusion module and output of the lightweight cryo-electron microscopy particle selection model based on the YOLOX model are combined into the FastHead module; the FastHead module includes a projection layer, three dilated convolution modules, and an output.

[0081] In one specific embodiment, the projection layer includes a 1×1 convolutional layer and a 3×3 convolutional layer connected in sequence, wherein the 1×1 convolutional layer is used to reduce the channel dimension, and the 3×3 convolutional layer is used to refine the semantic context of the image after the channel dimension is reduced.

[0082] The three dilated convolutional modules each include a first 1×1 convolutional layer, a 3×3 convolutional layer, and a second 1×1 convolutional layer connected in sequence.

[0083] The output terminal is a 32x downsampling channel.

[0084] In one specific embodiment, three dilated convolutional modules are stacked with three consecutive dilated residual blocks with different dilation rates of 2, 4, and 8, respectively. As the dilation rate increases, the size of the receptive field also increases to generate output features with multiple receptive fields, covering the scale of all objects, thereby achieving a multi-scale fusion effect.

[0085] In this embodiment, the FastHead module is a novel feature fusion module. Previous feature fusion modules mostly used feature pyramids, but their network structure is complex, with many parameters, leading to long training times. The feature maps extracted by the backbone network already contain rich contextual information, allowing for the fusion of information from different dimensions by combining these rich feature maps with a more lightweight feature fusion module. Therefore, our method uses multi-level dilated convolutions instead of a feature pyramid structure. Multi-level dilated convolutions have multi-level receptive field output features, achieving scale fusion while being more lightweight. Finally, the output is simplified to a single channel, resulting in a more lightweight network structure without sacrificing accuracy.

[0086] In this embodiment, the output typically uses IoU and GIoU, but these constraints are too simple, resulting in slow convergence. Furthermore, when the predicted bounding box is inside the ground truth bounding box, gradient backpropagation fails, and the degree of overlap is not reflected, significantly impacting the network's convergence speed and accuracy. To further improve accuracy and convergence speed, this method incorporates a Euclidean distance constraint between the target bounding box and the predicted bounding box into the position loss function. This directly minimizes the distance between the center points, resolving the issue of the predicted bounding box being inside the ground truth bounding box, and improving both convergence speed and accuracy.

[0087] In one specific embodiment, when training the lightweight cryo-electron microscopy particle selection model based on the YOLOX model, the position loss function used is Loss. diou :

[0088]

[0089] in, and These represent the predicted bounding box and the ground truth bounding box, respectively. ( , () represents the Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box; It represents the diagonal distance of the smallest closure region that can simultaneously contain both the predicted bounding box and the ground truth bounding box.

[0090] In this embodiment, as Figure 6 As shown, cryo-electron microscopy images with high noise levels were selected and compared with the YOLOX version. The first image from the left is the YOLOX result, the second image is the result of the method of this invention, and the third image is the label image. It can be seen that the method of this invention is more sensitive to the difference between background noise and target, and has a lower false positive rate. YOLOX, on the other hand, will incorrectly identify background noise as target particles.

[0091] In this embodiment, several mainstream target detection algorithms were selected and compared with the method used in this invention: YOLOv3, YOLOv3-spp, YOLOv3-tiny, YOLOv5s, YOLOv5m, and YOLOX. The comparison between this invention and other mainstream algorithms is shown in the table below:

[0092]

[0093] It can be seen that the method of this invention outperforms other algorithms in AP(0.5) and has fewer parameters. Compared with YOLOv3-spp, although mAP(0.5) is slightly lower by 0.2%, the method of this invention has only 8.4% of its network parameters.

[0094] In this embodiment, as Figure 7 As shown, the first image from the left is the RELION image, the second is the TOPAZ image, the third is the method of this invention, and the fourth is the annotation image; the boxed area indicates the sensitive region, which has a lot of background noise and is very similar to the target to be selected. Compared with mainstream cryo-electron microscopy particle selection algorithms, the method of this invention has a lower false positive rate, better detail feature extraction capability, and can more accurately identify the target and noise.

[0095] The following table compares the methods of Topaz and Relion for mainstream particle selection using cryo-electron microscopy with the method of this invention:

[0096]

[0097] In this embodiment, to address the challenge of cryo-electron microscopy images, where noise is highly similar to the target particle, a lightweight cryo-electron microscopy particle selection method based on YOLOX is proposed, offering strong noise resistance. This method utilizes the B-ResBlockX module to enhance the network's fine-grainedness, improving information representation and detailed feature capture capabilities. Test results demonstrate that B-ResBlockX achieves high accuracy in distinguishing between noise and target particles. The feature fusion module uses FastHead instead of the commonly used feature pyramid to fuse multi-scale features, simplifying the output to a single channel. This significantly reduces network complexity and improves training speed without compromising accuracy. Finally, using DIoU as the position loss function further enhances convergence speed and accuracy. Furthermore, compared to current mainstream target detection algorithms and cryo-electron microscopy target detection software, the proposed algorithm demonstrates improved performance and is better suited for target detection in cryo-electron microscopy images.

[0098] This invention discloses a noise-resistant lightweight cryo-electron microscopy particle selection method, comprising the following parts: (1) First, the publicly available cryo-electron microscopy dataset is preprocessed, converting the binary source file into trainable images and performing relevant transformations. (2) The dataset and validation set are put into the method of this invention to train the lightweight cryo-electron microscopy particle selection model based on the YOLOX model, wherein the lightweight cryo-electron microscopy particle selection model based on the YOLOX model comprises three parts: First, the abstract features of the image are extracted through the feature extraction module, which can better extract features and distinguish noise and particles to be selected; then, the high- and low-dimensional information is fused through the feature fusion module to supplement feature information; finally, the network parameters are updated in reverse iteration according to the loss function through the output end. This invention can better resist noise, better distinguish noise and particles, has a better AP (0.5) index, and the training speed of the model is improved, while the complexity of the model and the parameters of the network are significantly reduced. It can perform cryo-electron microscopy particle selection work with higher efficiency and better accuracy. This invention solves the problem of the weak ability of existing technologies to distinguish between noise and particles to be selected, and provides a noise-resistant, lightweight cryo-electron microscopy particle selection method, which is easy to train, simple in structure, and has high accuracy.

[0099] Example 3

[0100] A noise-resistant, lightweight cryo-electron microscopy particle selection system includes an image acquisition module, an image preprocessing module, a model building module, and a particle selection module.

[0101] The image acquisition module is used to obtain serialized files of cryo-electron microscopy images;

[0102] The image preprocessing module is used to preprocess the serialized files of cryo-electron microscopy images;

[0103] The model building module is used to build and train a lightweight cryo-electron microscopy particle selection model based on the YOLOX model using preprocessed data;

[0104] The particle selection module is used to perform cryo-electron microscopy particle selection using a trained lightweight cryo-electron microscopy particle selection model.

[0105] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A noise-resistant, lightweight cryo-electron microscopy particle selection method, characterized in that: Includes the following steps: S1. Obtain the serialized file of the cryo-electron microscopy image; S2. Preprocess the serialized files of cryo-electron microscopy images; S3. Construct and train a lightweight cryo-electron microscopy particle selection model based on the YOLOX model using preprocessed data; the feature extraction module of the lightweight cryo-electron microscopy particle selection model based on the YOLOX model adopts the residual module B-ResBlockX module; the B-ResBlockX module includes a first image processing branch, a second image processing branch, a Concat layer, and a 3×3 convolutional layer; the number of channels in the first image processing branch and the second image processing branch are half the number of input channels respectively; the first image processing branch is configured with a 1×1 convolutional layer, a 3×3 convolutional layer, and a ResXBlock module connected in sequence to obtain the first image features; the second image processing branch is configured with a 1×1 convolutional layer to extract the second image features; B-ResBlockX concatenates the first image features and the second image features through the Concat layer, and finally extracts the features through the 3×3 convolutional layer to output the final extracted features; S4. Perform cryo-electron microscopy particle selection using a trained lightweight cryo-electron microscopy particle selection model.

2. The noise-resistant, lightweight cryo-electron microscopy particle selection method according to claim 1, characterized in that: In step S2, the images used for training are preprocessed, specifically as follows: S201. Read the serialized file of each cryo-electron microscopy image and convert it into several visualization images; S202. Several visualization images converted from each cryo-electron microscopy image are overlaid, and then drift correction is performed; S203. Perform image denoising on the corrected image; S204. The denoised image is segmented into fragments; S205. Remove the bounding box information of the cut segments and organize the final segment information into training set, test set, and validation set.

3. The noise-resistant, lightweight cryo-electron microscopy particle selection method according to claim 1, characterized in that: The 1×1 and 3×3 convolutional layers in the first image processing branch of the B-ResBlockX module are used to extract features based on the downsampling mode in sequence, refining semantic information; the ResXBlock module is used to supplement the information lost after the convolutional layers extract features.

4. The noise-resistant, lightweight cryo-electron microscopy particle selection method according to claim 3, characterized in that: The specific process of the ResXBlock module is as follows: After convolution by the 1×1 convolutional layer in the first image processing branch, the feature map is uniformly divided into s feature subsets, x i Let K represent the i-th feature subset. i () represents the 3×3 convolutional layer of the first image processing branch, denoted by y. i K represents i The output of () is used; when passing through the 3×3 convolutional layer, it receives the feature information from the previous feature subset, thus finally outputting a feature map containing information from different receptive fields: The ResXBlock module includes several filters, each with the same channels, and the filters are connected in a residual style. The output features of the previous group are concatenated with the input features of another group and sent to the next group of filters. This process is repeated several times until the convolutional layers have finished extracting features, thereby obtaining information from different dimensions of loss.

5. The noise-resistant, lightweight cryo-electron microscopy particle selection method according to claim 1, characterized in that: The feature fusion module and output of the lightweight cryo-electron microscopy particle selection model based on the YOLOX model are combined into the FastHead module; the FastHead module includes a projection layer, three dilated convolution modules, and an output. The projection layer includes a 1×1 convolutional layer and a 3×3 convolutional layer connected in sequence. The 1×1 convolutional layer is used to reduce the channel dimension, and the 3×3 convolutional layer is used to refine the semantic context of the image after the channel dimension is reduced. The three dilated convolutional modules each include a first 1×1 convolutional layer, a 3×3 convolutional layer, and a second 1×1 convolutional layer connected in sequence. The output terminal is a 32x downsampling channel.

6. The noise-resistant, lightweight cryo-electron microscopy particle selection method according to claim 5, characterized in that: Three dilated convolutional modules are stacked with three consecutive dilated residual blocks with different dilation rates of 2, 4, and 8, respectively. As the dilation rate increases, the size of the receptive field also increases to generate output features with multiple receptive fields, covering the scale of all objects, thereby achieving a multi-scale fusion effect.

7. The noise-resistant, lightweight cryo-electron microscopy particle selection method according to claim 1, characterized in that: When training the lightweight cryo-electron microscopy particle selection model based on the YOLOX model, the position loss function used is Loss. diou : in, and These represent the predicted bounding box and the ground truth bounding box, respectively. ( , () represents the Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box; It represents the diagonal distance of the smallest closure region that can simultaneously contain both the predicted bounding box and the ground truth bounding box.

8. A noise-resistant, lightweight cryo-electron microscopy particle sorting system, characterized in that: It includes an image acquisition module, an image preprocessing module, a model building module, and a particle selection module; The image acquisition module is used to obtain serialized files of cryo-electron microscopy images; The image preprocessing module is used to preprocess the serialized files of cryo-electron microscopy images; The model building module is used to construct and train a lightweight cryo-electron microscopy particle selection model based on the YOLOX model using preprocessed data. The feature extraction module of the lightweight cryo-electron microscopy particle selection model based on the YOLOX model adopts the residual module B-ResBlockX module. The B-ResBlockX module includes a first image processing branch, a second image processing branch, a Concat layer, and a 3×3 convolutional layer. The number of channels in the first and second image processing branches are half the number of input channels. The first image processing branch is configured with a 1×1 convolutional layer, a 3×3 convolutional layer, and a ResXBlock module connected in sequence to obtain the first image features. The second image processing branch is configured with a 1×1 convolutional layer to extract the second image features. B-ResBlockX concatenates the features of the first and second images through a Concat layer, and finally extracts the features through a 3×3 convolutional layer to output the final extracted features. The particle selection module is used to perform cryo-electron microscopy particle selection using a trained lightweight cryo-electron microscopy particle selection model.

Images