Decomposition convolution model, lesion region identification method and related device

Abstract

This disclosure provides a decomposed convolution model, a lesion region identification method, and related equipment. The decomposed convolution model includes several decomposed convolutional layers and a normalization layer. Each decomposed convolutional layer corresponds to a single normalization layer, and the output of the decomposed convolutional layer is connected to the input of the normalization layer. Each decomposed convolutional layer includes two low-rank matrices, which are obtained by decomposing the standard convolutional layer according to its kernel parameters based on the weight matrix of the corresponding standard convolutional layer. The normalization layer normalizes the values ​​of the two low-rank matrices, ensuring they follow the same distribution. By replacing the standard convolutional layer of an existing convolutional model with the decomposed convolutional layer and the normalization layer, a decomposed convolution model is obtained, reducing model parameters. In application, the decomposed convolution model is deployed on a local device. Feature maps corresponding to case images are extracted using the decomposed convolution model, and these feature maps are input into a target detection model to identify lesion regions, ensuring high security of case data.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of data processing, in particular to a decomposed convolution model, a lesion region identification method and related equipment. BACKGROUND

[0002] AI medical image assisted diagnosis is to use the image processing ability of artificial intelligence to realize functions such as lesion detection, target organ segmentation, risk assessment, lesion type classification, organ and tissue marker positioning and segmentation, etc., to meet the clinical auxiliary needs of lesion identification and marking, disease type classification, image three-dimensional reconstruction, and even automatic delineation of radiotherapy target area, and to provide assistance and reference for doctors to read films and delineate. Deploying an AI medical image assisted diagnosis system on a medical display device will become a new trend in future products. However, due to the limited computing resources of medical display devices and the large computational overhead of the convolutional network model relied on by AI medical image assisted diagnosis, the combination of the two becomes a dilemma. SUMMARY

[0003] The present disclosure aims to provide a decomposed convolution model, a lesion region identification method and related equipment, which has small computational overhead of a decomposed convolution model constructed based on a decomposed convolution layer and a standardization layer, can be deployed on a medical display device with limited computing resources, and has ideal performance.

[0004] To achieve the above purpose, the present disclosure adopts the following technical solution: a decomposed convolution model, comprising a plurality of decomposed convolution layers and a plurality of standardization layers, a single decomposed convolution layer corresponding to a single standardization layer, and the output end of the decomposed convolution layer being connected to the input end of the standardization layer.

[0005] The decomposed convolution layer comprises two low-rank matrices, and the two low-rank matrices are obtained by decomposing the weight matrix of the corresponding standard convolution layer according to the convolution kernel parameters of the standard convolution layer.

[0006] The standardization layer is used to standardize the values of the two low-rank matrices, so that the two low-rank matrices conform to the same distribution.

[0007] The present disclosure also provides a lesion region identification method based on a case image, comprising:

[0008] Obtaining a case image;

[0009] Inputting the case image into a decomposed convolution model for feature extraction to obtain a feature map corresponding to the case image, wherein the decomposed convolution model is the decomposed convolution model as described above.

[0010] Inputting the feature map into a target detection model to identify and mark the lesion region on the case image.

[0011] The disclosure also provides a lesion area recognition device based on a case image, comprising:

[0012] An acquisition module is configured to acquire a case image;

[0013] An extraction module is configured to input the case image into a decomposed convolution model to extract features, so as to obtain a feature map corresponding to the case image, wherein the decomposed convolution model is the decomposed convolution model as described above.

[0014] A recognition module is configured to input the feature map into a target detection model to recognize and mark a lesion area on the case image.

[0015] The disclosure also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method according to any one of the above.

[0016] The disclosure also provides a computer-readable storage medium storing a computer program, wherein the computer program is executed by a processor to implement the steps of the method according to any one of the above.

[0017] The decomposed convolution model, the lesion area recognition method and the related device provided in the disclosure, the decomposed convolution model comprises a plurality of decomposed convolution layers and a plurality of normalization layers, a single decomposed convolution layer corresponds to a single normalization layer, and the output end of the decomposed convolution layer is connected to the input end of the normalization layer. The decomposed convolution layer comprises two low-rank matrices, which are obtained by decomposing the weight matrix of the corresponding standard convolution layer according to the convolution kernel parameters of the standard convolution layer; the calculation complexity of a single convolution is reduced, and the number of model parameters is reduced. The normalization layer is used to normalize the values of the two low-rank matrices, so that the two low-rank matrices conform to the same distribution and are no longer independent, and can be used to approximate the weight matrix before the decomposed convolution layer is decomposed, thereby ensuring that the decomposed convolution layer has ideal performance. All standard convolution layers in the existing convolution model are replaced by the decomposed convolution layer and the normalization layer, thereby obtaining the decomposed convolution model provided in the disclosure. In application, the decomposed convolution model can be deployed on a medical display device with limited computing resources. The medical display device acquires a case image, and then inputs the case image into the decomposed convolution model to extract features, so as to obtain a feature map corresponding to the case image. Finally, the feature map is input into a target detection model to recognize and mark a lesion area on the case image. Since the decomposed convolution model is deployed locally on the medical display device, the security of the case data can be ensured, and the instability of network transmission can be avoided, and the lesion area can be quickly recognized. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a structural schematic diagram of the decomposed convolution model in an embodiment of the disclosure.

[0019] Figure 2 This is a schematic diagram of the steps of a lesion area identification method based on a case image in one embodiment of the present disclosure;

[0020] Figure 3 This is an overall structural block diagram of a lesion area identification device based on a case image according to an embodiment of the present disclosure;

[0021] Figure 4 This is a schematic block diagram of the structure of a computer device according to an embodiment of the present disclosure.

[0022] The realization of the purpose, functional features and advantages of this disclosure will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this disclosure clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this disclosure.

[0024] Reference Figure 1 In one embodiment of this disclosure, a decomposed convolution model is provided, which includes a plurality of decomposed convolutional layers A1 and a plurality of normalized layers A2, wherein a single decomposed convolutional layer A1 corresponds to a single normalized layer A2, and the output of the decomposed convolutional layer A1 is connected to the input of the normalized layer A2.

[0025] The decomposed convolutional layer A1 includes two low-rank matrices, which are obtained by decomposing the standard convolutional layer A1 according to the kernel parameters of the standard convolutional layer based on the weight matrix of the standard convolutional layer.

[0026] The normalization layer A2 is used to normalize the values ​​of the two low-rank matrices so that the two low-rank matrices follow the same distribution.

[0027] In this embodiment, the decomposed convolution model includes several decomposed convolutional layers A1 and several normalized layers A2. Each decomposed convolutional layer A1 corresponds to a single normalized layer A2, and the output of the decomposed convolutional layer A1 is connected to the input of the normalized layer A2. The decomposed convolutional layer A1 includes two low-rank matrices. These two low-rank matrices are obtained by decomposing the weight matrix of the standard convolutional layer corresponding to the decomposed convolutional layer A1 according to the kernel parameters of the standard convolutional layer (i.e., decomposing the existing weight matrix of the standard convolutional layer into two low-rank matrices to obtain the decomposed convolutional layer A1 in this embodiment). The model parameters of the decomposed convolutional layer A1 are much smaller than those of the standard convolutional layer. During forward inference, the two low-rank matrices of the decomposed convolutional layer A1 perform the convolution operation twice, thereby reducing computational overhead (computational overhead includes the memory and CPU / GPU of the hardware device where the decomposed convolutional model is deployed). Since model performance (i.e., expressive power) is positively correlated with model parameters, generally, reducing the number of model parameters will lead to a decrease in model performance. To ensure the ideal performance of the decomposed convolution model, the decomposed convolutional layer A1 is followed by the normalization layer A2 (i.e., the output of the decomposed convolutional layer A1 is connected to the input of the normalization layer A2). The normalization layer A2 is used to normalize the values ​​of the two low-rank matrices of the decomposed convolutional layer A1, so that the two low-rank matrices follow the same distribution, thereby establishing a connection between the two low-rank matrices and satisfying the constraint that their product approximates the weight matrix of the standard convolutional layer before decomposition. This ensures enhanced model expressiveness while reducing model parameters.

[0028] In this embodiment, the decomposed convolutional layer A1 is followed by the normalized layer A2, which is used to replace the standard convolutional layer in the existing model. This significantly reduces the model parameters while effectively ensuring the ideal performance of the model, allowing the decomposed convolutional model of this embodiment to be deployed on hardware devices with limited computing resources.

[0029] Furthermore, the convolution kernel parameters include the number of input channels c of the convolution kernel. i Number of output channels c o And the kernel size k×k;

[0030] The weight matrix has a size of m×n, where m=c i ×k, n=c o ×k;

[0031] The two low-rank matrices are low-rank matrix P and low-rank matrix Q, where the size of low-rank matrix P is m×r and the size of low-rank matrix Q is r×n, where r represents the rank of low-rank matrix P and low-rank matrix Q, and r < 0. <min{m,n}。

[0032] In this embodiment, it is assumed that the weight matrix W of the 2D standard convolutional layer is of size m×n, where m=ci ×k, n = c o ×k, c i Represents the number of input channels of the convolutional kernel, c o Represents the number of output channels of the convolutional layer, k×k represents the size of the convolutional kernel. Assume the input feature map F i , with size c i ×h i ×w i ; and its output feature map F o , with size c o ×h o ×w o ; The forward calculation process of standard convolution can be expressed as: Where Represents the cyclic convolution process. Considering a single convolution in the cyclic convolution process, its computational complexity is o(c i c o k 2 ). Decompose the weight matrix W into two low-rank matrices P and low-rank matrix Q, that is, W≈P×Q, where the size of the low-rank matrix P is m×r, and the size of the low-rank matrix Q is r×n. r represents the rank of the low-rank matrix P and the low-rank matrix Q, and r << min{m,n}, that is, the rank r of the low-rank matrix P and the low-rank matrix Q is much smaller than the minimum of m and n, which plays a role in significantly reducing the model parameters. After decomposing the weight matrix of the 2D standard convolutional layer into two low-rank matrices, the decomposed convolutional layer A1 of this embodiment is obtained; in the forward inference process, the two low-rank matrices of the decomposed convolutional layer A1 are used to complete the convolution operation in two times, so as to reduce the computational overhead. Specifically, the decomposed convolutional layer A1 decomposes the sequential operation of the k×k convolutional kernel of the standard convolutional layer into two operations of 1×k convolutional kernels. Among them, the low-rank matrix Q is recombined into a convolutional kernel r×c i ×1×k, and the matrix P is recombined into c o ×1×k×r, and then a convolution operation is performed separately. The forward calculation process of the decomposed convolution is expressed as: The computational complexity of a single convolution of the decomposed convolutional layer A1 is: o(c i kr + c o kr). It can be seen from this that the computational complexity of the decomposed convolutional layer A1 is greatly reduced compared with that of the standard convolutional layer.

[0033] Furthermore, the specific steps for the normalization layer A2 to normalize the values of the two low-rank matrices so that the two low-rank matrices follow the same distribution are:

[0034] Call the mean formula to calculate the means of the low-rank matrix P and the low-rank matrix Q, and call the variance formula to calculate the variances of the low-rank matrix P and the low-rank matrix Q. Among them, the mean formula is: u is the mean; the variance formula is: N = (m × r + r × n), v 2 Let x be the variance. i The values ​​in the low-rank matrix P and the low-rank matrix Q are represented;

[0035] The values ​​of the low-rank matrix P and the low-rank matrix Q are standardized using the mean, the variance, and the standardization formula, wherein the standardization formula is: ε is the standardized value in the low-rank matrix, and ε is a near-zero floating-point number.

[0036] In this embodiment, after introducing the decomposed convolutional layer A1 into the model's network structure, both the forward inference and backward propagation gradient updates of the convolution are based on two low-rank matrices P and Q. These two low-rank matrices are not necessarily related, which actually contradicts the low-rank approximation of the initial matrix decomposition. To establish the connection between the two low-rank matrices and enhance the model's expressive power while reducing model parameters, a normalization layer A2 is added after the decomposed convolutional layer A1 to normalize the values ​​of the two low-rank matrices P and Q, making them follow the same distribution. Specifically, the mean formula is first used to calculate the mean of the low-rank matrices P and Q, and the variance formula is used to calculate the variance of the low-rank matrices P and Q. The mean formula is: u is the mean; the variance formula is: N = (m × r + r × n), v 2 Let x be the variance. i The values ​​in low-rank matrices P and Q are represented. Having the mean and variance, the values ​​in low-rank matrices P and Q can be standardized. The standardization formula is used to standardize the values ​​of low-rank matrices P and Q using the calculated mean, variance, and standardization formula; the standardization formula is: Let ε be the standardized value of the low-rank matrix, and let ε be a near-zero floating-point number. The calculation process involves standardizing the low-rank matrices P and Q with the same mean and variance. The two standardized low-rank matrices conform to the same submatrix and are no longer independent, thus approximating the undecomposed weight matrix W of a standard convolutional layer. The standardization layer A2 unifies the weights of the decomposed convolution after each iteration, acting as a constraint to ensure enhanced model expressiveness while reducing model parameters.

[0037] Reference Figure 2 This disclosure also provides an embodiment of a lesion region identification method based on medical case images, comprising:

[0038] S1: Acquire case images;

[0039] S2: Input the case image into the decomposition convolution model for feature extraction to obtain the feature map corresponding to the case image, wherein the decomposition convolution model is any of the decomposition convolution models described above;

[0040] S3: Input the feature map into the target detection model to identify and mark the lesion area on the case image.

[0041] In this embodiment, the recognition system acquires a case image (e.g., a chest X-ray) input by the user, and then inputs the case image into any of the decomposed convolutional models described above for feature extraction to obtain a feature map corresponding to the case image. The recognition system inputs the feature map into a target detection model for processing, thereby identifying the lesion region on the case image and marking the lesion region on the case image. The target detection model uses the case image as training data and is trained through deep learning. The trained target detection model can identify the lesion region in the input image. Preferably, the decomposed convolutional model and the target detection model are deployed on local hardware devices, such as medical display devices. Because the decomposed convolutional model has fewer model parameters, it can significantly reduce computational complexity, thus making it suitable for local hardware devices with limited computing resources. Furthermore, deploying the model on local hardware devices can ensure the security of case data and avoid the instability of network transmission, thereby improving the speed of identifying lesion regions in case images.

[0042] Furthermore, the step of inputting the feature map into the target detection model to identify and label the lesion region on the case image includes:

[0043] S301: Input the feature map into the target detection model to identify several lesion regions on the case image, as well as the disease information corresponding to each lesion region. The disease information includes the disease type and the probability of disease, and each lesion region corresponds to at least one disease information.

[0044] S302: Circle each of the lesion areas on the case image and associate each of the disease information with its corresponding lesion area.

[0045] In this embodiment, the recognition system inputs the feature map of the case image into the target detection model for recognition, obtaining several lesion regions contained in the case image, and the disease information corresponding to each lesion region. The disease information includes the disease type and the probability of disease occurrence; a single lesion region corresponds to at least one disease information (i.e., a single lesion region may correspond to multiple disease types, and a single disease type corresponds to a single probability of disease occurrence). The recognition system circles each lesion region on the case image and links each disease information with its corresponding lesion region, providing assistance and reference for doctors interpreting the images.

[0046] Furthermore, after the step of inputting the feature map into the target detection model to identify and label the lesion region on the case image, the method includes:

[0047] S4: Output the case image marked with the lesion area to the display interface of the medical display device, and monitor whether an expansion command is received, the expansion command including several specified lesion areas;

[0048] S5: If an expand command is received, expand and display the disease information corresponding to each of the specified lesion areas;

[0049] S6: If no unfold command is received, hide the disease information corresponding to each lesion area.

[0050] In this embodiment, the recognition system outputs the recognition results of the target detection model, i.e., the case image marked with lesion areas, to the display interface of the medical display device, and monitors in real time whether an expansion command is received; wherein, the expansion command includes several specified lesion areas. If an expansion command is received, the recognition system expands and displays the disease information corresponding to each specified lesion area contained in the expansion command (for example, when a doctor clicks on a lesion area, this click action is regarded as an expansion command, and the recognition system expands and displays the disease information of that lesion area), providing assistance and reference for the doctor to read the image. If no expansion command is received, the recognition system hides the disease information corresponding to each lesion area (i.e., in the initial stage of outputting the case image marked with lesion areas to the display interface, all disease information is hidden), to avoid displaying too much disease information and affecting the doctor's viewing of the case image.

[0051] Reference Figure 3 This disclosure also provides an embodiment of a lesion region identification device based on a medical record image, comprising:

[0052] Module 1 is used to acquire case images;

[0053] Extraction module 2 is used to input the case image into a decomposition convolution model for feature extraction to obtain the feature map corresponding to the case image, wherein the decomposition convolution model is any of the decomposition convolution models described above;

[0054] The recognition module 3 is used to input the feature map into the target detection model to identify and mark the lesion area on the case image.

[0055] Furthermore, the extraction module 2 includes:

[0056] The identification unit is used to input the feature map into the target detection model to identify several lesion regions on the case image, as well as the disease information corresponding to each lesion region. The disease information includes the disease type and the probability of disease, and each lesion region corresponds to at least one disease information.

[0057] The labeling unit is used to circle each of the lesion areas on the case image and associate each of the disease information with its corresponding lesion area.

[0058] Furthermore, the lesion area identification device also includes:

[0059] The monitoring module 4 is used to output the case image marked with the lesion area to the display interface of the medical display device, and monitor whether an expansion command is received, the expansion command including several specified lesion areas;

[0060] Expand module 5 is used to expand and display the disease information corresponding to each of the specified lesion areas if an expand command is received;

[0061] The hiding module 6 is used to hide the disease information corresponding to each lesion area if no unfolding instruction is received.

[0062] In this embodiment, each module and unit in the lesion area identification device is used to perform the corresponding steps in the above-mentioned lesion area identification method based on case images. The specific implementation process is not described in detail here.

[0063] This embodiment provides a lesion region identification method based on medical case images. A decomposed convolutional model is deployed on a local medical display device. The medical display device acquires medical case images and then inputs these images into the decomposed convolutional model for feature extraction, obtaining a feature map corresponding to the medical case image. Finally, the feature map is input into a target detection model to identify and label lesion regions on the medical case image. Because the decomposed convolutional model is deployed locally on the medical display device, the security of medical case data is ensured, and the instability of network transmission is avoided, enabling rapid identification of lesion regions.

[0064] Reference Figure 4 This disclosure also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 4As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data such as medical records images. The network interface allows communication with external terminals via a network connection. When executed by the processor, the computer program implements a lesion area identification method based on medical records images.

[0065] The processor described above performs the following steps in the lesion region identification method based on case images:

[0066] S1: Acquire case images;

[0067] S2: Input the case image into the decomposition convolution model for feature extraction to obtain the feature map corresponding to the case image, wherein the decomposition convolution model is any of the decomposition convolution models described above;

[0068] S3: Input the feature map into the target detection model to identify and mark the lesion area on the case image.

[0069] Furthermore, the step of inputting the feature map into the target detection model to identify and label the lesion region on the case image includes:

[0070] S301: Input the feature map into the target detection model to identify several lesion regions on the case image, as well as the disease information corresponding to each lesion region. The disease information includes the disease type and the probability of disease, and each lesion region corresponds to at least one disease information.

[0071] S302: Circle each of the lesion areas on the case image and associate each of the disease information with its corresponding lesion area.

[0072] Furthermore, after the step of inputting the feature map into the target detection model to identify and label the lesion region on the case image, the method includes:

[0073] S4: Output the case image marked with the lesion area to the display interface of the medical display device, and monitor whether an expansion command is received, the expansion command including several specified lesion areas;

[0074] S5: If an expand command is received, expand and display the disease information corresponding to each of the specified lesion areas;

[0075] S6: If no unfold command is received, hide the disease information corresponding to each lesion area.

[0076] One embodiment of this disclosure also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a lesion region identification method based on medical case images. Specifically, the lesion region identification method based on medical case images includes:

[0077] S1: Acquire case images;

[0078] S2: Input the case image into the decomposition convolution model for feature extraction to obtain the feature map corresponding to the case image, wherein the decomposition convolution model is any of the decomposition convolution models described above;

[0079] S3: Input the feature map into the target detection model to identify and mark the lesion area on the case image.

[0080] Furthermore, the step of inputting the feature map into the target detection model to identify and label the lesion region on the case image includes:

[0081] S301: Input the feature map into the target detection model to identify several lesion regions on the case image, as well as the disease information corresponding to each lesion region. The disease information includes the disease type and the probability of disease, and each lesion region corresponds to at least one disease information.

[0082] S302: Circle each of the lesion areas on the case image and associate each of the disease information with its corresponding lesion area.

[0083] Furthermore, after the step of inputting the feature map into the target detection model to identify and label the lesion region on the case image, the method includes:

[0084] S4: Output the case image marked with the lesion area to the display interface of the medical display device, and monitor whether an expansion command is received, the expansion command including several specified lesion areas;

[0085] S5: If an expand command is received, expand and display the disease information corresponding to each of the specified lesion areas;

[0086] S6: If no unfold command is received, hide the disease information corresponding to each lesion area.

[0087] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media provided in this disclosure and in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0088] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, first object, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, first object, or method. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, first object, or method that includes that element.

[0089] The above description is only a preferred embodiment of this disclosure and does not limit the patent scope of this disclosure. Any equivalent structural or procedural changes made based on the content of this disclosure and its drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this disclosure.

Claims

1. A method for lesion region identification based on medical case images, wherein, include: Acquire case images; The case image is input into a decomposition convolution model for feature extraction to obtain a feature map corresponding to the case image. The decomposition convolution model includes several decomposition convolutional layers and several normalization layers, with each decomposition convolutional layer corresponding to a single normalization layer. The output of each decomposition convolutional layer is connected to the input of the normalization layer. Each decomposition convolutional layer includes two low-rank matrices, which are obtained by decomposing the two low-rank matrices according to the weight matrix of the corresponding standard convolutional layer based on the kernel parameters of the standard convolutional layer. The normalization layer is used to normalize the values ​​of the two low-rank matrices so that they follow the same distribution. The feature map is input into the target detection model to identify and label the lesion areas on the case image.

2. The lesion region identification method based on case images according to claim 1, wherein, The convolution kernel parameters include the number of input channels c of the convolution kernel. i Number of output channels c o and kernel size ; The size of the weight matrix is ,in, , ; The two low-rank matrices are low-rank matrix P and low-rank matrix Q, wherein the size of low-rank matrix P is... The size of the low-rank matrix Q is r represents the rank of the low-rank matrix P and the low-rank matrix Q. .

3. The lesion region identification method based on case images according to claim 2, wherein, The standardization layer standardizes the values ​​of the two low-rank matrices to make them conform to the same distribution. The specific steps are as follows: The mean of the low-rank matrix P and the low-rank matrix Q are calculated using the mean formula, and the variance of the low-rank matrix P and the low-rank matrix Q is calculated using the variance formula. The mean formula is: u is the mean; the variance formula is: , , Let x be the variance. i The values ​​in the low-rank matrix P and the low-rank matrix Q are represented; The values ​​of the low-rank matrix P and the low-rank matrix Q are standardized using the mean, the variance, and the standardization formula, wherein the standardization formula is: , These are the standardized values ​​in a low-rank matrix. It is a floating-point number close to zero.

4. The lesion region identification method based on case images according to claim 1, wherein, The step of inputting the feature map into the target detection model to identify and label the lesion region on the case image includes: The feature map is input into the target detection model to identify several lesion regions on the case image, as well as the disease information corresponding to each lesion region. The disease information includes the disease type and the probability of disease, and each lesion region corresponds to at least one disease information. Each of the lesion areas is circled on the case image, and each of the disease information is associated with its corresponding lesion area.

5. The lesion region identification method based on case images according to claim 4, wherein, After the step of inputting the feature map into the target detection model to identify and label the lesion areas on the case image, the following steps are included: The case image marked with the lesion area is output to the display interface of the medical display device, and the device monitors whether an expansion command is received, the expansion command including several specified lesion areas; If an expand command is received, the disease information corresponding to each of the specified lesion areas will be expanded and displayed. If no unfold command is received, the disease information corresponding to each lesion area will be hidden.

6. A lesion region identification device based on medical case images, used to implement the lesion region identification method based on medical case images as described in any one of claims 1-5, wherein, include: The acquisition module is used to acquire case images; The extraction module is used to input the case image into a decomposition convolution model for feature extraction to obtain a feature map corresponding to the case image. The decomposition convolution model includes several decomposition convolution layers and several normalization layers, with each decomposition convolution layer corresponding to a single normalization layer. The output of each decomposition convolution layer is connected to the input of the normalization layer. Each decomposition convolution layer includes two low-rank matrices, which are obtained by decomposing the two low-rank matrices according to the weight matrix of the corresponding standard convolution layer based on the kernel parameters of the standard convolution layer. The normalization layer is used to normalize the values ​​of the two low-rank matrices so that they follow the same distribution. The recognition module is used to input the feature map into the target detection model to identify and mark the lesion areas on the case image.

7. The lesion region identification device based on medical record images according to claim 6, wherein, The extraction module includes: The identification unit is used to input the feature map into the target detection model to identify several lesion regions on the case image, as well as the disease information corresponding to each lesion region. The disease information includes the disease type and the probability of disease, and each lesion region corresponds to at least one disease information. The labeling unit is used to circle each of the lesion areas on the case image and associate each of the disease information with its corresponding lesion area.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein... When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

9. A computer-readable storage medium having a computer program stored thereon, wherein, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images