A method, device, equipment, and medium for determining similarity in cell-dense biomedical images.
By using a deep convolutional neural network model based on contrastive learning, target features of cell-dense biomedical images are detected and clustered, and group features of images are calculated. This solves the accuracy problem of image similarity judgment in traditional methods and achieves accurate similarity judgment for cell-dense biomedical images.
Patent Information
- Application Number
- CN202410253103.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-06
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-03-06
AI Technical Summary
Traditional methods struggle to accurately determine the similarity of cell-dense biomedical images, especially when the number of target objects is uncertain and their spatial location varies.
A deep convolutional neural network model based on contrastive learning is adopted to calculate the group features of images through object detection, feature extraction and unsupervised clustering, including class ratio, mean of intra-class features and spatial distribution features. Similarity judgment is performed by combining Pearson similarity, cosine similarity and Frobenius norm.
It achieves accurate similarity judgment for cell-dense biomedical images, avoiding calculation errors caused by differences in the number and angle of target objects, and has strong robustness and interpretability.
Smart Images

Figure CN118097664B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to a method, apparatus, device, and medium for determining the similarity of cell-dense biomedical images. Background Technology
[0002] Images commonly used in biomedical research, such as pathological images stained with hematoxylin and eosin (HE), IHC immunohistochemical images, and antinuclear antibody (ANA) fluorescent images, share the common characteristic of containing a large number of densely packed cells within a single image (e.g., Figure 1 (As shown). For this type of image, common image feature descriptions include global features and local features. Global features include the number of target objects (cells) in the image, their density, and whether there are clustered areas, etc. Local features include whether there are special target objects in the image (such as whether there are tumor cells or broken other objects), the feature accumulation of special target objects, etc.
[0003] Different image similarity metrics include distance methods, histogram methods, and Pearson correlation coefficients. These image similarity metrics are not suitable for cell-dense biomedical images that rely on multiple dense target features. For example... Figure 2 The two images shown, a and c, both contain several fluorescent cells. The classification features of the images depend on certain key cellular features, such as the presence of aggregated chromatin and some coarse-grained cells. This indicates that the two images belong to the same category of antinuclear antibody. Figure 2 In the image, a and c are two fluorescent images of the same type of ANA antinuclear antibody, while b is a structurally similar image to a but without any cellular imaging features. However, according to traditional similarity index calculation methods, the resulting index cannot link the two ANA images to similarity. For example, calculating separately... Figure 2 The similarity indices between each pair of images a, b, and c were calculated using Pearson correlation coefficients. The similarity between a and b was 0.7250, a and c was 0.5893, and b and c was 0.5845, indicating that a and b were the most similar images. Alternatively, the Euclidean distance between the images was calculated: a and b were 50105.70, a and c were 42484.48, and b and c were 39196.68, indicating that b and c were the most similar.
[0004] like Figure 3The three images shown are simulated dense cell images. Images a and b (where the cells are simulated as squares) are clusters of several square objects, while image c is a cluster of circular objects (simulating cell morphology). The differences between a and b are slight color variations, and some individuals in b have changed angles, but overall, it remains a dense cluster of square objects. However, when we use traditional image similarity calculation methods: the histogram chi-square distance results are 42.4126 for a and b, 1.8352 for a and c, and 558.5504 for b and c, indicating that a and c are the most similar images. The Euclidean distance results are 98912.32 for a and b, 87709.19 for a and c, and 137340.03 for b and c, again indicating that a and c are the most similar.
[0005] Therefore, in densely packed images, changes in the angle or position of certain target objects, or variations in the number of objects of a particular category, will affect pixel similarity calculations using traditional methods. For specific image detection / recognition / segmentation tasks, traditional methods are no longer suitable for determining whether two images filled with an uncertain number of targets (objects) in different spatial locations are similar. Summary of the Invention
[0006] The purpose of this invention is to provide a method, apparatus, device, and medium for judging the similarity of cell-dense biomedical images, so as to achieve accurate similarity judgment between cell-dense biomedical images filled with an uncertain number of target objects and with different spatial positions of the targets.
[0007] The objective of this invention can be achieved through the following technical solutions:
[0008] A method for determining the similarity of cell-dense biomedical images includes the following steps:
[0009] S1, Obtain the biomedical image dataset;
[0010] S2, detecting densely packed single targets in biomedical images, wherein the target is a cell or cell nucleus;
[0011] S3. Based on the single target detected in step S2, the image features of the target are obtained using a deep convolutional neural network model based on contrastive learning.
[0012] S4, based on the clustering results of image features of a large number of targets, calculate the group characteristics of dense targets;
[0013] S5, performs image similarity judgment based on group characteristics.
[0014] The biomedical image dataset in step S1 includes label information for the location detection of individual target objects.
[0015] In step S2, a YOLO-based target detection model is established to detect dense individual targets in biomedical images.
[0016] Step S3 specifically involves: constructing a deep convolutional neural network model based on contrastive learning; using the individual target objects detected in step S2 as the training set to train the model; obtaining image features of the target objects through contrastive learning; wherein, based on the characteristic that the image features of individual target objects are not affected by angular rotation, during the training process, each individual target object is randomly rotated by 1 to 360 degrees, and a preset logarithmic original image and the corresponding rotated augmented image are randomly selected to calculate image features through the encoder module of the deep convolutional neural network model; calculating the loss function using the image features of a pair of images; and minimizing the loss function to achieve model training.
[0017] Step S4 specifically involves: performing cluster analysis on the image features of all target objects using an unsupervised clustering method to obtain the optimal number of clusters; and statistically analyzing the group characteristics based on the clustering results, including the proportion of each category, the mean of image features within each category, and spatial distribution characteristics.
[0018] Let F1 be the first representation of the group characteristic, representing the proportion of each category, where F1 = (K1, K2, K3, ..., K). n ), where n is the total number of categories, i.e., the optimal number of clusters, and K i This represents the proportion of the number of target objects in category i to the total number of objects.
[0019] Let F2 be the mean of image features within a category, that is, the mean of image features of all objects in each category. Assuming that the dimension of image features in each category is m, then F2 is represented as a set containing n m-dimensional vectors, F2={V1,V2,V3,…,V n}, V i The mean vector of image features representing the target object of the i-th category, where i∈{1,2,3,…,n};
[0020] Let F3, the third representation of the group characteristics, be the spatial distribution characteristic, represented by an n×n distance matrix, where k is the element of the matrix. ij The median of the nearest Euclidean distances between all target objects of class i and class j, i, j∈{1,2,3,…,n};
[0021] Combining the three representations, we obtain the group feature representation of an image as {F1, F2, F3}.
[0022] Step S5 specifically involves: Based on the group feature representations of the two images, calculating similarity indices using these group feature representations. Specifically, for the first group feature representation F1, which is an n-dimensional vector, calculating the Pearson similarity value between the two n-dimensional vectors corresponding to the two images, serving as the first similarity index; for the second group feature representation F2, which is an m×n-dimensional matrix, expanding the matrix column-wise into a one-dimensional vector, calculating the cosine similarity between the two one-dimensional vectors corresponding to the two images, serving as the second similarity index; and for the third group feature representation F3, which is an n×n symmetric matrix, performing feature decomposition, and calculating the Frobenius norm using the difference between the eigenvalues of the two images to construct a diagonal matrix, serving as the third similarity index; and determining the similarity between the two images based on the first, second, and / or third similarity indices.
[0023] A cell-dense biomedical image similarity assessment device includes:
[0024] The data acquisition module is used to acquire biomedical image datasets;
[0025] The target detection module is used to detect densely packed single targets in biomedical images, wherein the target is a cell or cell nucleus;
[0026] The image feature extraction module is used to obtain the image features of a single target detected by the target detection module using a deep convolutional neural network model based on contrastive learning.
[0027] The group feature calculation module is used to calculate the group features of dense targets based on the clustering results of the image features of the targets.
[0028] The similarity judgment module is used to judge image similarity based on group characteristics.
[0029] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described above.
[0030] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.
[0031] Compared with the prior art, the present invention has the following beneficial effects:
[0032] (1) This invention maps the overall features of an image to the features of a group, avoiding inaccurate results caused by differences in the number of target objects or differences in the display angle of a single target object. It can effectively improve the image similarity calculation effect and has robustness.
[0033] (2) This invention utilizes a target object feature pre-training learning method based on contrastive learning. By using image augmentation methods such as random rotation of the target object in the training set, the image features of the same target object remain unchanged after rotating at different angles.
[0034] (3) The similarity calculation method proposed in this invention describes the similarity between two images from three perspectives, which has strong interpretability. The three perspectives include category ratio similarity, intra-class image feature mean similarity, and group spatial distribution density similarity. Different weights are assigned to the three group features for different points of interest to obtain the overall image similarity. Attached Figure Description
[0035] Figure 1 These are schematic diagrams of common cell-dense biomedical images, where (1a) is a yeast image, (1b) is a pathological image stained with HE, (1c) is an ANA antinuclear antibody fluorescent image, and (1d) is an IHC immunohistochemical image.
[0036] Figure 2 Examples of antinuclear antibody images and similar images;
[0037] Figure 3 Example of a simulated dense cell image;
[0038] Figure 4 This is a flowchart of the method of the present invention;
[0039] Figure 5 This is a schematic diagram of a deep convolutional neural network model structure based on contrastive learning in one embodiment;
[0040] Figure 6 This is a schematic diagram of the target detection and group feature calculation process of the present invention;
[0041] Figure 7 This is a schematic diagram of the cell-dense image similarity judgment process of the present invention. Detailed Implementation
[0042] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.
[0043] This embodiment provides a method for judging the similarity of cell-dense biomedical images, such as... Figure 4 As shown, it includes the following steps:
[0044] S1. Obtain a biomedical image dataset, which includes label information for the location detection of individual target objects, for training and testing of target detection in step S2.
[0045] S2 detects densely packed single targets in biomedical images.
[0046] Specifically, a YOLO-based target detection model is established to detect densely packed individual targets in biomedical images, wherein the targets are cells or cell nuclei.
[0047] S3. Based on the single target detected in step S2, the image features of the target are obtained using a deep convolutional neural network model based on contrastive learning.
[0048] Specifically, constructing deep convolutional neural network models based on contrastive learning, such as... Figure 5 As shown, the patch image of a single target object detected in step S2 is used as the training set to train the model, and the image features of the target object are obtained through contrastive learning. Based on the characteristic that the image features of a single target object are not affected by angular rotation, during the training process, each single target object is randomly rotated by 1 to 360 degrees, and 100 pairs of original images and corresponding rotated augmented images are randomly selected. The image features are calculated by the encoder module of the deep convolutional neural network model, and the loss function is calculated using the image features of a pair of images. The model training is achieved by minimizing the loss function.
[0049] After training for 100–300 epochs until the loss function curve is nearly stationary, a computational model for the image features of the target object is obtained.
[0050] S4, based on the clustering results of the image features of the targets, calculates the group characteristics of dense targets.
[0051] like Figure 6 As shown, firstly, step S2 is used to detect the location of the target in the dense target image and obtain the small image of the target object. Then, step S3 is used to calculate the image features of the target object. Next, unsupervised clustering is used to perform cluster analysis on the image features of all target objects to obtain the optimal number of clusters (set as n). Finally, the group features are statistically analyzed based on the clustering results, and F1, F2, and F3 are used to represent the three representations of the group features.
[0052] Let F1 be the first representation of the group characteristic, representing the proportion of each category, where F1 = (K1, K2, K3, ..., K). n ), where n is the total number of categories, i.e., the optimal number of clusters, and K i This represents the proportion of the number of target objects in category i to the total number of objects.
[0053] Let F2 be the mean of image features within a category, that is, the mean of image features of all objects in each category. Assuming that the dimension of image features in each category is m, then F2 is represented as a set containing n m-dimensional vectors, F2={V1,V2,V3,…,V n}, V i The mean vector of image features represents the target object of the i-th category, where i∈{1,2,3,…,n}.
[0054] Let F3, the third representation of the group characteristics, be the spatial distribution characteristic, represented by an n×n distance matrix, where k is the element of the matrix. ij Let represent the median of the nearest Euclidean distances between all target objects of class i and class j, where i, j ∈ {1, 2, 3, ..., n}.
[0055] Combining the three representations, we obtain the group feature representation of an image as {F1, F2, F3}.
[0056] S5, performs image similarity judgment based on group characteristics.
[0057] Based on the group feature representations of two images, a similarity index is calculated separately for each group feature representation. Since the group features consist of three representations, the similarity index is calculated separately for each of the three representations, with each part having a different focus. Specifically, for example... Figure 7 As shown, for the first representation F1 of the group features, which is an n-dimensional vector, the Pearson similarity value of the two n-dimensional vectors corresponding to the two images is calculated as the first similarity index. For the second representation F2 of the group features, which is an m×n-dimensional matrix, the matrix is expanded column-wise into a one-dimensional vector, and the cosine similarity of the two one-dimensional vectors corresponding to the two images is calculated as the second similarity index. For the third representation F3 of the group features, which is an n×n symmetric matrix, eigenvalue decomposition is performed, and the Frobenius norm is calculated using the difference between the eigenvalues of the two images to construct a diagonal matrix, which is used as the third similarity index. The three similarity indices are weighted and summed or used individually in different application scenarios to obtain the similarity result between the two images.
[0058] To verify the performance of the present invention, this embodiment was tested on a simulated dataset.
[0059] Step 1: Construct a simulated dataset for object detection. Using Python's PIL package, construct 10,000 black images, each 1000*1000 pixels in size. Randomly fill each image with 10-100 triangles and rectangles of random size (maximum circumscribed circle radius varies from 10-50 pixels), random grayscale values (ranges from 100-255), and random positions (objects do not overlap). Calculate the maximum and minimum X and Y coordinates of each object's (triangle or rectangle) region to obtain the bounding box of each object (the top-left corner coordinates of the bounding box are...). min ,Y min ), bottom right corner coordinates (X max ,Y max The constructed graphs and their corresponding bounding boxes serve as the training dataset for object detection.
[0060] Step 2: Train the object detection neural network based on the YOLO V5 framework. Train the network using the training set constructed in Step 5, setting the number of iterations (epochs) to 200 and the batch size to 20. The trained network is then used to identify objects (triangles / rectangles) in new images.
[0061] Step 3: Construct a simulated dataset for target feature calculation. Using Python's PIL package, construct a black 56x56 pixel (300 dpi) square image. Randomly select three points within the image (these three points cannot lie on the same side of the square), and connect them to form a white triangle. This is a random triangle image, with black as the background. Then, reconstruct another black 56x56 pixel square image. Randomly select two points within this image, and using these two points as vertices, draw the largest possible white rectangle within the square's area. This is a randomly constructed rectangle image, with black as the background. Finally, randomly construct 1000 triangle images and 1000 rectangle images, each 56x56 pixels in size.
[0062] Step 4: Construct the dataset for the deep convolutional neural network model. The deep convolutional neural network model takes out a batch (batch_size is customizable, 2 is recommended) of random images each time. The images are randomly rotated from 0 to 360 degrees, and the grayscale values are randomly changed (grayscale value changes from 100 to 255; the original white image has a grayscale value of 255). Ten new images are randomly added to each image (each batch obtains 10 * batch_size number of added images).
[0063] Step 5: Construct a deep convolutional neural network model framework based on contrastive learning. For each input image, the feature tensors obtained by the feature extraction module from the two augmented images are made as similar as possible. The loss is estimated using negative cosine similarity, and the network parameters of the feature extraction module are trained by minimizing the loss. The trained contrastive learning framework serves as the feature extractor for the target image.
[0064] Step 6: Calculate the clustering features of the target image. Based on the features calculated by the model in Step 5, perform unsupervised clustering on all training sets to obtain the optimal number of clusters (2 in this example).
[0065] Step 7: Construct test images. Construct a 1000*1000 pixel black image 'a', and randomly fill it with rectangles of random side length (10-50 pixels), angle, and position. The grayscale value of the rectangles varies randomly (range 100-255). Construct a 1000*1000 pixel black image 'b', and fill it with equilateral triangles. The center of the triangles should be the same as the center of the rectangle in image 'a', the side length of the triangles should be the same as the shortest side of the rectangle at that position in image 'a', and the grayscale value of the triangles should be the same as the grayscale value of the rectangle at that position in image 'a'. Construct another 1000*1000 pixel black image 'c', and randomly fill it with rectangles of random side length (10-50 pixels), angle, and position. The grayscale value of the rectangles varies randomly (range 100-255).
[0066] Step 8: Calculate the traditional similarity indices for the test graphs. Calculate the traditional similarity indices (Pearson correlation coefficient, histogram distance, and Euclidean distance) for graphs a and b, a and c, and b and c respectively.
[0067] Step 9: Calculate the similarity index proposed in this invention for the test images. For images a, b, and c, use the target detection model trained in Step 2 to detect the target locations, and crop the target regions to obtain smaller images (images a, b, and c correspond to set1, set2, and set3 respectively); combine set1, set2, and set3 together for clustering into two classes (based on the optimal number of clusters obtained in Step 6); calculate the group features of set1, set2, and set3, denoted as: {F 11 ,F 12 ,F 13}、{F 21 ,F 22 ,F 23} and {F 31 ,F 32 ,F 33 The process of calculating the similarity index for Figures a, b, and c is as follows:
[0068] 1) The process of calculating the similarity index between Figure a and Figure b: Calculate F 11 With F 21 Let V1 be the Pearson similarity value; calculate F. 12 With F 22 The cosine similarity value is denoted as V2; for F respectively 13 With F 23 Perform matrix eigenvalue decomposition, and construct a new diagonal matrix by the difference of the eigenvalues of the two diagonal matrices. Calculate the Frobenius norm, denoted as V3. In this embodiment, V1, V2 and V3 are added together as the similarity index between Figure a and Figure b.
[0069] 2) The process of calculating the similarity index between Figure a and Figure c: Calculate F 11 With F 31 Let V1 be the Pearson similarity value; calculate F. 12 With F 32 The cosine similarity value is denoted as V2; for F respectively 13 With F 33 Perform matrix eigenvalue decomposition, and construct a new diagonal matrix by the difference of the eigenvalues of the two diagonal matrices. Calculate the Frobenius norm, denoted as V3. In this embodiment, V1, V2 and V3 are added together as the similarity index between Figure a and Figure c.
[0070] 3) The process of calculating the similarity index between Figure b and Figure c: Calculate F 21 With F 31 Let V1 be the Pearson similarity value; calculate F. 22 With F 32 The cosine similarity value is denoted as V2; for F respectively 23 With F 33 Perform matrix eigenvalue decomposition, and construct a new diagonal matrix by the difference of the eigenvalues of the two diagonal matrices. Calculate the Frobenius norm, denoted as V3. In this embodiment, V1, V2 and V3 are added together as the similarity index between Figure b and Figure c.
[0071] Step 10: Repeat steps 7 and 9 100 times to calculate the mean and standard deviation of the similarity index between Figure a and Figure b, Figure a and Figure c, and Figure b and Figure c, and compare the traditional method with the method of this invention.
[0072] The above is an introduction to the method embodiments. The following describes the solution of the present invention further through device embodiments.
[0073] This embodiment provides a cell-dense biomedical image similarity judgment device, including:
[0074] The data acquisition module is used to acquire biomedical image datasets;
[0075] The target detection module is used to detect densely packed single targets in biomedical images, wherein the target is a cell or cell nucleus;
[0076] The image feature extraction module is used to obtain the image features of a single target detected by the target detection module using a deep convolutional neural network model based on contrastive learning.
[0077] The group feature calculation module is used to calculate the group features of dense targets based on the clustering results of the image features of the targets.
[0078] The similarity judgment module is used to judge image similarity based on group characteristics.
[0079] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.
[0080] In one embodiment, the electronic device includes a computing unit that can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) or a computer program loaded from a storage unit into a random access memory (RAM). The RAM may also store various programs and data required for device operation. The computing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.
[0081] Multiple components in an electronic device are connected to an I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.
[0082] The computing unit can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of computing units include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit performs the various methods and processes described above, such as image processing methods and / or model training methods. For example, in some embodiments, the image processing methods and / or model training methods can be implemented as computer software programs tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the computing unit, one or more steps of the image processing methods and / or model training methods described above can be performed. Alternatively, in other embodiments, the computing unit can be configured to perform image processing methods and / or model training methods by any other suitable means (e.g., by means of firmware).
[0083] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.
[0084] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.
[0085] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.
[0086] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A method for judging the similarity of cell-dense biomedical images, characterized in that, Includes the following steps: S1, Obtain the biomedical image dataset; S2, detecting densely packed single targets in biomedical images, wherein the target is a cell or cell nucleus; S3. Based on the single target detected in step S2, the image features of the target are obtained using a deep convolutional neural network model based on contrastive learning. Specifically, a deep convolutional neural network model based on contrastive learning is constructed. The single target object detected in step S2 is used as the training set to train the model. The image features of the target object are obtained through contrastive learning. The image features of a single target object are not affected by the angle rotation. During the training process, each single target object is randomly rotated from 1 to 360 degrees. The original image and the corresponding rotated augmented image of a preset logarithm are randomly selected. The image features are calculated by the encoder module of the deep convolutional neural network model. The loss function is calculated using the image features of a pair of images. The model training is achieved by minimizing the loss function. S4, based on the clustering results of the image features of the targets, calculate the group characteristics of dense targets; S5, performs image similarity judgment based on group characteristics.
2. The method for determining the similarity of cell-dense biomedical images according to claim 1, characterized in that, The biomedical image dataset in step S1 includes label information for the location detection of individual target objects.
3. The method for determining the similarity of cell-dense biomedical images according to claim 1, characterized in that, In step S2, a YOLO-based target detection model is established to detect dense individual targets in biomedical images.
4. The method for determining the similarity of cell-dense biomedical images according to claim 1, characterized in that, Step S4 specifically involves: performing cluster analysis on the image features of all target objects using an unsupervised clustering method to obtain the optimal number of clusters; and statistically analyzing the group characteristics based on the clustering results, including the proportion of each category, the mean of image features within each category, and spatial distribution characteristics.
5. The method for determining the similarity of cell-dense biomedical images according to claim 4, characterized in that, The first representation of the group feature F1 is the proportion of the number of each class, F1 = (K1, K2, K3, …, K n ), where n is the total number of classes, i.e. the optimal clustering number, K i represents the proportion of the number of target objects of the i-th class to the total number of objects; Let F2 be the mean of image features within a category, that is, the mean of image features of all objects in each category. Assuming that the dimension of image features in each category is m, then F2 is represented as a set containing n m-dimensional vectors, F2={V1,V2,V3,…,V...} n }, V i The mean vector of image features for the target object of category i, where i∈{1, 2, 3,…,n}; Let F3, the third representation of the group characteristics, be the spatial distribution characteristic, represented by an n×n distance matrix, where k is the element of the matrix. ij The median of the nearest Euclidean distances between all target objects of class i and class j, i, j∈{1,2,3,…,n}; Combining the three representations, we obtain the group feature representation of an image as {F1, F2, F3}.
6. The method for determining the similarity of cell-dense biomedical images according to claim 5, characterized in that, Step S5 specifically involves: Based on the group feature representations of the two images, calculating similarity indices using these group feature representations. Specifically, for the first group feature representation F1, which is an n-dimensional vector, calculating the Pearson similarity value between the two n-dimensional vectors corresponding to the two images, serving as the first similarity index; for the second group feature representation F2, which is an m×n-dimensional matrix, expanding the matrix column-wise into a one-dimensional vector, calculating the cosine similarity between the two one-dimensional vectors corresponding to the two images, serving as the second similarity index; and for the third group feature representation F3, which is an n×n symmetric matrix, performing feature decomposition, and calculating the Frobenius norm using the difference between the eigenvalues of the two images to construct a diagonal matrix, serving as the third similarity index; and determining the similarity between the two images based on the first, second, and / or third similarity indices.
7. A cell-dense biomedical image similarity judgment device, characterized in that, The method of claim 1 includes: The data acquisition module is used to acquire biomedical image datasets; The target detection module is used to detect densely packed single targets in biomedical images, wherein the target is a cell or cell nucleus; The image feature extraction module is used to obtain the image features of a single target detected by the target detection module using a deep convolutional neural network model based on contrastive learning. The group feature calculation module is used to calculate the group features of dense targets based on the clustering results of the image features of the targets. The similarity judgment module is used to judge image similarity based on group characteristics.
8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 6.
Citation Information
Patent Citations
Blood disease leukocyte scatter diagram similarity analysis method based on structural analysis
CN111274949A
