Mouse tumor detection method based on contrast agent enhanced nuclear magnetic resonance image

Abstract

The invention provides a mouse tumor detection method based on a contrast agent enhanced nuclear magnetic resonance image. The method comprises the following steps: firstly, inputting a nuclear magnetic resonance image of a mouse enhanced by using a contrast agent into a data preprocessing enhancement module to obtain an enhanced image feature; then inputting the original mouse nuclear magnetic resonance image without using the contrast agent into an image embedding module to obtain a series of decoding query tokens; and inputting the enhanced image features into an image encoder module to obtain image encoding features. And subsequently, the image coding features are input into a Transform encoder module to obtain coding features. And finally, inputting the decoding query token and the coding features into a cross attention filtering decoder module to obtain a target detection bounding box. According to the method, multiple image enhancement technologies and an improved deep learning model are combined, and the accuracy and robustness of mouse tumor detection are remarkably improved by fully utilizing the original and enhanced image features.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image processing, and particularly relates to an image processing technology for detecting mouse tumors by processing contrast agent enhanced magnetic resonance images. BACKGROUND

[0002] Magnetic resonance imaging (MRI) is an important medical imaging technique widely used in diagnosis and research. Through MRI, high-resolution soft tissue images can be obtained to help doctors accurately locate and evaluate tumors. However, conventional MRI has lower sensitivity in detecting small tumors or early tumors in some cases. In order to improve detection accuracy, contrast agents are often used to enhance the contrast of tumor areas, so that the lesion site can be more clearly displayed. Detecting changes in mouse tumors is a direct means of evaluating the efficacy of anticancer drugs. By observing changes in tumor volume, number and growth rate in mice before and after drug treatment, researchers can judge the anti-tumor effect of the drug. This is of great significance for new drug screening and optimization.

[0003] Currently, tumor detection methods in MRI images have been widely applied, but there are still some technical challenges in their application to small animal models such as mice. Further, conventional tumor detection in MRI images rarely applies to the assistance of contrast agent images. Due to the small size of mice, higher resolution and more sensitive detection algorithms are needed for tumor detection and identification, so the assistance of contrast agent images is particularly important. In addition, existing methods still have a lot of room for improvement in image processing and feature extraction, especially in how to effectively use enhanced image features to improve detection accuracy. For example, the imaging contrast agent prepared by the amphiphilic polymer nanomicelles containing polydopamine amino acid chelated ferric ion in patent CN110496231A can enhance the image quality under the premise of non-toxicity. Making full use of the auxiliary images enhanced by contrast agents can further improve the efficiency of detection and guide subsequent treatment.

[0004] To overcome the above problems, the present application proposes a mouse tumor detection method based on contrast agent enhanced magnetic resonance images. This method combines various image enhancement techniques and deep learning models, and makes full use of the features of original and enhanced images to significantly improve the accuracy and robustness of tumor detection. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a mouse tumor detection method based on contrast agent enhanced magnetic resonance images, to overcome the problem that the prior art cannot fully utilize contrast agent images, thereby improving the recognition accuracy.

[0006] To solve the above technical problems, the present application adopts the following technical solutions:

[0007] In a first aspect, the present application provides a mouse tumor detection method based on contrast agent enhanced magnetic resonance imaging, comprising the following steps:

[0008] S1, collecting the magnetic resonance image Img of the mouse after using the contrast agent for enhancement enhance , inputting the data preprocessing enhancement module M p , and obtaining the enhanced image feature F aug after processing;

[0009] S2, collecting the original magnetic resonance image Img of the mouse without using the contrast agent origin , inputting the image embedding module M e , and obtaining the series decoding query token Query after processing de ;

[0010] S3, inputting the enhanced image feature F aug obtained in step S1 into the image encoder module M cen for encoding processing to obtain the image encoding feature F cen ;

[0011] S4, inputting the image encoding feature F cen obtained in step S3 into the Transformer encoder module M ten for encoding processing to obtain the encoding feature F ten ;

[0012] S5, inputting the decoding query token Query obtained in step S2 and the encoding feature F de obtained in step S4 into the cross-attention filtering decoder module M de to obtain the target detection bounding box Bbox after processing.

[0013] Further, the step S1 specifically comprises the following steps:

[0014] S1.1, performing intensity normalization calculation on the input image Img enhance with a resolution size of HxWxC in length, width and channel number to obtain the image feature F n ;

[0015] S1.2, applying geometric enhancement Augment geometry to the image feature F n after intensity normalization calculation to obtain the image feature F g with a feature size of HxWxC;

[0016] S1.3, applying contrast enhancement Augment g to the image feature F g after geometric enhancement to obtain the image feature F contrast .contrast the processed image feature F c with feature size HxWxC.

[0017] S1.4, for the image feature F c applying pseudo-color augmentation Augment pseudo the processed image feature F aug with feature size HxWxC1.

[0018] Further, for the image feature F c the intensity feature value at each image position

[0019] Further, applying pseudo-color augmentation Augment pseudo the processed image feature F p includes the following steps:

[0020] S2.1, for the intensity feature value at each point, calculating the three-channel intensity values after JET pseudo-color transformation where i, j are the positions corresponding in F c .

[0021] S2.2, for the intensity feature value at each point, calculating the three-channel intensity values after rainbow pseudo-color transformation where i, j are the positions corresponding in F c .

[0022] S2.3, for the intensity feature value at each point, calculating the three-channel intensity values after heat map pseudo-color transformation where i, j are the positions corresponding in F c .

[0023] S2.4, combining along the channel dimension into a new feature map F aug with feature size HxWxC1.

[0024] Further, the working process of the image embedding module M e includes the following steps:

[0025] S3.1, performing intensity normalization processing on the original mouse magnetic resonance image Img origin before using contrast agent, to obtain a normalized image feature F n2 with feature size HxWxC.

[0026] S3.2, for the normalized image feature F n2Input into the pre-trained convolutional neural network model ConvNet, extract multi-level image features F e2 , whose feature size is H2 x W2 x C2.

[0027] S3.3, the multi-level image features F e2 Dimension reduction processing is performed through the full connection layer, generating a series of decoding query tokens Query de , whose feature size is H3 x W3 x C3.

[0028] Further, the image encoder module M cen includes enhanced image features F aug Input the pre-trained convolutional neural network model ConvNet, extract multi-level encoding features F cen , whose feature size is H4 x W4 x C4.

[0029] Further, the Transformer encoder module M ten includes image encoding features F cen Input the multi-layer Transformer encoder, each layer includes a multi-head self-attention layer and a feedforward neural network layer, to obtain the encoding features F ten , whose feature size is H4 x W4 x C4.

[0030] Further, the cross-attention filtering decoding module M de The working process includes the following steps:

[0031] S4.1, the decoding query token Query de According to the confidence filtering layer L filter , the first layer of query tokens Query de1 , whose feature dimension size is H3 x W3 x f1C3;

[0032] S4.2, the decoding query token Query de1 And the encoding features F ten Calculate the first layer cross-attention Attention1, input the first layer feedforward network to obtain the decoder features F de1 , whose dimension size is H4 x W4 x C4;

[0033] S4.3, the decoding query token Query de According to the confidence filtering layer L filter , the second layer of query tokens Query de2 , whose feature dimension size is H3 x W3 x f2C3;

[0034] S4.4, the decoding query token Query de2 with the first layer decoder feature F de1 calculate the second layer cross attention Attention2, input the second layer feedforward network to obtain the decoder feature F de2 , the dimension size is H4×W4×C4;

[0035] S4.5, the decoding query token Query de According to the confidence filtering layer L filter with the filtering factor f3 to calculate the third layer query token Query de3 , the feature dimension size is H3×W3×f3C3;

[0036] S4.6, the decoding query token Query de3 with the second layer decoder feature F de2 calculate the third layer cross attention Attention3, input the third layer feedforward network to obtain the decoder feature F de3 , the dimension size is H4×W4×C4;

[0037] S4.7, the decoding query token Query de According to the confidence filtering layer L filter with the filtering factor f4 to calculate the fourth layer query token Query de4 , the feature dimension size is H3×W3×f4C3;

[0038] S4.8, the decoding query token Query de4 with the third layer decoder feature F de3 calculate the fourth layer cross attention Attention4, input the fourth layer feedforward network to obtain the decoder feature F de4 , the dimension size is H4×W4×C4;

[0039] S4.9, the updated query feature Query de Through the feedforward network calculation, generate the candidate target detection bounding box BBox Candidate and its corresponding classification Cls;

[0040] S4.10, the corresponding BBox Candidate and the real label are matched to the final prediction BBox using the Hungarian matching algorithm, and the comprehensive loss including classification loss, overlap loss and bounding box loss is calculated to update the weight.

[0041] Further, the working process of the confidence filtering layer includes the following steps:

[0042] S5.1, calculate the query token foreground confidence s;

[0043] S5.2, according to the foreground confidence, selecting a filter factor to update the query token with high confidence;

[0044] The confidence calculation process is:

[0045] s l-1 = MLP(F l-1 (1+ UP(alpha l x s l ))

[0046] Wherein, s l represents the confidence of the query token at the lth layer, s l-1 represents the confidence of the query token at the (l-1)th layer, F l represents the query token feature map at the lth layer, F l-1 represents the query token feature map at the (l-1)th layer, alpha l represents a learnable parameter, and UP represents a bilinear interpolation up-sampling function.

[0047] In a second aspect, the present application provides a mouse tumor detection system based on contrast agent enhanced nuclear magnetic resonance images, which comprises a data preprocessing enhancement module M p , an image embedding module M e , an image encoder module M cen , a Transformer encoder module M ten , and a cross-attention filtering decoder module M de , and the system is configured to implement the mouse tumor detection method based on contrast agent enhanced nuclear magnetic resonance images according to the first aspect of the present application.

[0048] In a third aspect, the present application provides a computer device, which comprises a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements the steps of the mouse tumor detection method based on contrast agent enhanced nuclear magnetic resonance images according to the first aspect of the present application when executing the computer program.

[0049] In a fourth aspect, the present application provides a computer readable storage medium, which stores a computer program, and the program implements the steps of the mouse tumor detection method based on contrast agent enhanced nuclear magnetic resonance images according to the first aspect of the present application when executed by a processor.

[0050] The present application has the following beneficial effects:

[0051] The application provides a mouse tumor detection method based on contrast agent enhanced nuclear magnetic resonance imaging. First, the nuclear magnetic resonance imaging of the mouse after being enhanced by the contrast agent is input into a data preprocessing enhancement module to obtain enhanced image features. Then, the original nuclear magnetic resonance imaging of the mouse before being enhanced by the contrast agent is input into an image embedding module to obtain a series of decoding query tokens. Further, the enhanced image features are input into an image encoder module to obtain image coding features. Then, the image coding features are input into a Transformer encoder module to obtain coding features. Finally, the decoding query tokens and the coding features are input into a cross-attention filtering decoder module to obtain a target detection bounding box. The mouse tumor detection method based on the contrast agent enhanced nuclear magnetic resonance imaging has the advantages of high data utilization rate and high recognition accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 It is a mouse tumor detection method based on contrast agent enhanced nuclear magnetic resonance imaging.

[0053] Figure 2 It is a pseudo-color enhancement data enhancement method schematic diagram of a mouse tumor detection method based on contrast agent enhanced nuclear magnetic resonance imaging.

[0054] Figure 3 It is a cross-attention filtering decoding module structure diagram based on contrast agent enhanced nuclear magnetic resonance imaging. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and not to limit the application.

[0056] In one embodiment, as shown in Figure 1 , a mouse tumor detection method based on contrast agent enhanced nuclear magnetic resonance imaging is provided, comprising:

[0057] S1, inputting the nuclear magnetic resonance imaging Img enhance of the mouse after being enhanced by the contrast agent into a data preprocessing enhancement module M p to obtain enhanced image features F aug .

[0058] Specifically, the contrast agent herein is an imaging contrast agent prepared from amphiphilic polymer nanomicelles containing polydopamine amino acid chelated ferric ions.

[0059] S2, inputting the original nuclear magnetic resonance imaging Img origin of the mouse before being enhanced by the contrast agent into an image embedding module M e to obtain a series of decoding query tokens Queryde .

[0060] S3, the enhanced image feature F aug input image encoder module M cen obtained image encoding feature F cen .

[0061] S4, the image encoding feature F cen input Transformer encoder module M ten obtained encoding feature F ten .

[0062] S5, the decoding query token Query de and the encoding feature F ten input cross-attention filtering decoder module M de obtained target detection bounding box Bbox.

[0063] Specifically, the Transformer encoder module M ten includes, the image encoding feature F cen input multi-layer Transformer encoder, each layer includes multi-head self-attention layer and feedforward neural network layer, to obtain the encoding feature F ten , the feature size is H4xH4xC4.

[0064] Specifically, the contrast-enhanced nuclear magnetic resonance image Img enhance input pre-processing enhancement module M p , obtain the enhanced image feature F aug , including:

[0065] S1.1, for the input image Img enhance , the length and width and the number of channels resolution size is HxWxC, after intensity normalization calculation to obtain the image feature F n .

[0066] S1.2, for the normalized image feature F n application geometry enhancement Augment geometry obtained image feature F g , the feature size is HxWxC.

[0067] S1.3, for the geometry enhanced image feature F g application contrast enhancement Augment contrast obtained image feature F c , the feature size is HxWxC.

[0068] S1.4, for the contrast enhanced image feature F cApplying pseudo-color augmentation Augment pseudo The image feature F is obtained aug , and the feature size is HxWxC2.

[0069] Specifically, the pseudo-color augmentation Augment pseudo The image feature F is obtained p , and the feature size is HxWxC2. c The intensity feature value of each image position

[0070] S2.1, for the intensity feature value of each point, calculate the three-channel intensity value after JET pseudo-color transformation Where i, j are the positions corresponding in F c .

[0071] Specifically, the RGB three channels of the JET pseudo-color transformation calculation formula can be represented as:

[0072]

[0073] S2.2, for the intensity feature value of each point, calculate the three-channel intensity value after rainbow pseudo-color transformation Where i, j are the positions corresponding in F c .

[0074] Specifically, the RGB three channels of the rainbow pseudo-color transformation calculation formula can be represented as:

[0075] H=240(1-I)

[0076]

[0077] S2.3, for the intensity feature value of each point, calculate the three-channel intensity value after heatmap pseudo-color transformation Where i, j are the positions corresponding in F c .

[0078] Specifically, the RGB three channels of the heatmap pseudo-color transformation calculation formula can be represented as:

[0079]

[0080] S2.4, combine F along the channel dimension to a new feature map F aug , and the feature size is HxWxC1.

[0081] Specifically, the image encoder module M cen includes, the enhanced image feature F augInput the pre-trained convolutional neural network model ConvNet to extract multi-level encoding features F cen with feature size H4×H4×C4.

[0082] Specifically, the image embedding module M e comprises the following steps:

[0083] S3.1, input the original mouse nuclear magnetic resonance image Img origin before using the contrast agent to perform intensity normalization processing to obtain normalized image features F n2 with feature size H×W×C.

[0084] S3.2, input the normalized image features F n2 into the pre-trained convolutional neural network model ConvNet to extract multi-level image features F e2 with feature size H2×W2×C2.

[0085] S3.3, input the multi-level image features F e2 into the fully connected layer for dimensionality reduction processing to generate a series of decoding query tokens Query de with feature size H3×W3×C3.

[0086] Specifically, the cross-attention filtering decoding module M de comprises the following steps:

[0087] S4.1, input the decoding query token Query de according to the confidence filtering layer L filter and the filtering factor f1 to calculate the query token Query de1 that needs to be updated in the first layer, with feature dimension size H3×H3×f1C3.

[0088] S4.2, input the decoding query token Query de1 and the encoding features F ten to calculate the first layer cross-attention Attention1, and input the first layer feedforward network to obtain the decoder features F de1 with dimension size H4×H4×C4.

[0089] S4.3, input the decoding query token Query de according to the confidence filtering layer L filter and the filtering factor f2 to calculate the query token Query de2 that needs to be updated in the second layer, with feature dimension size H3×H3×f2C3.

[0090] S4.4, input the decoding query token Query de2With the first layer decoder feature F de1 Calculate the second layer of cross attention Attention2, input the second layer of feedforward network to obtain the decoder feature F de2 , the dimension size is H4×H4×C4.

[0091] S4.5, decode the query token Query de Filter layer L according to confidence filter The query token Query that needs to be updated in the third layer is calculated with the filter factor f3 de3 , and its feature dimension size is H3×H3×f3C3.

[0092] S4.6, decode the query token Query de3 With the second layer decoder feature F de2 Calculate the third layer cross attention Attention3, input the third layer feedforward network to obtain the decoder feature F de3 , the dimension size is H4×H4×C4.

[0093] S4.7, decode the query token Query de Filter layer L according to confidence filter The query token Query that needs to be updated in the fourth layer is calculated with the filter factor f4 de4 , and its feature dimension size is H3×H3×f4C3.

[0094] S4.8, decode the query token Query de4 and the third layer decoder feature F de3 Calculate the fourth layer cross attention Attention4, input the fourth layer feedforward network to obtain the decoder feature F de4 , the dimension size is H4×H4×C4.

[0095] S4.9, update the query feature Query de Generate the target detection bounding box BBox to be selected through feedforward network calculation Candidate Its corresponding category is Cls.

[0096] S4.10, corresponding to BBox Candidate The final predicted BBox is matched with the true label using the Hungarian matching algorithm, and the comprehensive loss including classification loss, overlap loss, and bounding box loss is calculated and then the weights are updated by backpropagation.

[0097] The confidence filtering layer includes the following steps:

[0098] S5.1, calculate the query token foreground confidence s.

[0099] S5.2, according to the foreground confidence, select the filter factor to update the part of query tokens with high confidence.

[0100] The confidence calculation process is:

[0101] s l-1 = MLP(F l-1 (1+UP(a l *s l )))

[0102] where s l represents the confidence of the l-th layer query token, F l indicates the feature map of the l-th layer query token, a l indicates a learnable parameter, and UP indicates a bilinear interpolation up-sampling.

[0103] In one embodiment, a mouse tumor detection system based on contrast-enhanced nuclear magnetic resonance images is provided, which includes a data preprocessing enhancement module M p , an image embedding module M e , an image encoder module M cen , a Transformer encoder module M ten , and a cross-attention filtering decoder module M de , and the system is configured to implement the mouse tumor detection method based on contrast-enhanced nuclear magnetic resonance images as described in the above embodiments.

[0104] In one embodiment, a computer device is provided, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements the steps of the mouse tumor detection method based on contrast-enhanced nuclear magnetic resonance images as described in the above embodiments when executing the computer program.

[0105] In one embodiment, a computer readable storage medium is provided, which stores a computer program executable by a processor to implement the steps of the mouse tumor detection method based on contrast-enhanced nuclear magnetic resonance images as described in the embodiments of the present application.

[0106] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments of each method. In the embodiments of the present application, any reference to memory, storage, database or other medium can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0107] The application has been described in several embodiments, which are more specific and detailed, but should not be understood as limiting the scope of the patent. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A method for detecting mouse tumors based on contrast agent-enhanced magnetic resonance imaging, characterized in that: The mouse tumor detection method comprises the following steps: S1, collecting MRI images of mice after contrast agent enhancement enhance , input data preprocessing enhancement module M p , after processing, the enhanced image feature F is obtained aug ; S2, acquisition of original MRI images of mice without contrast agent (Img) origin , input image embedding module M e , after processing, we get a series of decoded query tokens Query de ; S3, the enhanced image feature F obtained in step S1 aug Input image encoder module M cen After encoding processing, the image encoding feature F is obtained cen ; S4, the image encoding feature F obtained in step S3 cen Input Transformer encoder module M ten After encoding processing, the encoding feature F is obtained ten ; S5, decode the query token Query obtained in step S2 de and the encoding feature F obtained in step S4 ten Input cross attention filter decoder module M de After processing, the target detection bounding box Bbox is obtained.

2. The method for detecting mouse tumors based on contrast agent-enhanced magnetic resonance imaging according to claim 1, wherein: The step S1 specifically includes the following steps: S1.1, for the input image Img enhance , its length, width and channel number resolution is H×W×C, and the image feature F is obtained after intensity normalization calculation n ; S1.2, image feature F calculated after intensity normalization n Applying geometric augmentation geometry After processing, the image feature F is obtained g , its characteristic size is H×W×C, where C is the number of channels, H is the height of a single layer, and W is the width of a single layer; S1.3, the geometrically enhanced image feature F g Apply contrast enhancement contrast After processing, the image feature F is obtained c , whose characteristic dimensions are H×W×C; S1.4, for the contrast-enhanced image feature F c Apply pseudo-color augmentation pseudo Then we get the image feature F aug , whose feature size is H×W×C1, where C1 is the number of channels after pseudo color enhancement.

3. The method for detecting mouse tumors based on contrast agent-enhanced magnetic resonance imaging according to claim 2, wherein: For the image feature F calculated by intensity normalization and contrast image enhancement method c , the intensity feature value of each image position Here, R represents the image pixel space with a resolution of H×W×C.

4. The method for detecting mouse tumors based on contrast agent-enhanced magnetic resonance imaging according to claim 3, wherein: Apply pseudo-color augmentation pseudo Get image feature F p The specific steps include: S2.1, for each point intensity characteristic value, calculate the three-channel intensity value after JET pseudo-color transformation Where i, j are the corresponding c Position in S2.2, for each point’s intensity feature value, calculate the three-channel intensity value after rainbow pseudo-color transformation Where i, j are the corresponding c Position in S2.3, for each point’s intensity feature value, calculate the three-channel intensity value after pseudo-color transformation of the heat map. Where i, j are the corresponding c Position in S2.4, Combine along the channel dimension into a new feature map F aug , whose characteristic dimensions are H×W×C1.

5. The method for detecting mouse tumors based on contrast agent-enhanced magnetic resonance imaging according to claim 1, wherein: The image embedding module M e The working process includes the following steps: S3.1, the original MRI image of the mouse before the use of contrast agent Img origin Perform intensity normalization to obtain the normalized image feature F n2 , whose characteristic dimensions are H×W×C; S3.2, the normalized image feature F n2 Input into the pre-trained convolutional neural network model ConvNet to extract multi-level image features F e2 , its feature size is H2×W2×C2, where C2 is the number of channels after normalization and multi-level extraction, H2 is the height of a single layer after normalization and multi-level extraction, and W2 is the width of a single layer after normalization and multi-level extraction; S3.3, multi-level image features F e2 Perform dimensionality reduction processing through the fully connected layer to generate a series of decoded query tokens Query de , its characteristic size is H3×W3×C3, where C3 is the number of channels after dimensionality reduction, H3 is the height of the single layer after dimensionality reduction, and W3 is the width of the single layer after dimensionality reduction.

6. The method for detecting mouse tumors based on contrast agent-enhanced magnetic resonance imaging according to claim 1, wherein: The image encoder module M cen Including, the enhanced image feature F aug Input the pre-trained convolutional neural network model ConvNet to extract multi-level encoding features F cen , its characteristic size is H4×W4×C4, where C4 is the number of channels after enhancement and multi-level extraction, H4 is the height of a single layer after enhancement and multi-level extraction, and W4 is the width of a single layer after enhancement and multi-level extraction.

7. The method for detecting mouse tumors based on contrast agent-enhanced magnetic resonance imaging according to claim 1, wherein: The Transformer encoder module M ten Including, encoding the image feature F cen Input a multi-layer Transformer encoder, each layer includes a multi-head self-attention layer and a feedforward neural network layer, and obtain the encoded feature F ten , its characteristic dimensions are H4×W4×C4.

8. The method for detecting mouse tumors based on contrast agent-enhanced magnetic resonance imaging according to claim 1, wherein: The cross-attention filter decoding module M de The working process includes the following steps: S4.1, decode the query token Query de Filter layer L according to confidence filter The query token Query that needs to be updated in the first layer is calculated with the filter factor f1 de1 , its feature dimension size is H3×H3×f1C3; S4.2, decode the query token Query de1 and encoding feature F ten Calculate the first layer of cross attention Attention1, input the first layer of feedforward network to obtain the decoder feature F de1 , the dimension size is H4×H4×C4; S4.3, decode the query token Query de Filter layer L according to confidence filter The query token Query that needs to be updated in the second layer is calculated with the filter factor f2 de2 , its feature dimension size is H3×H3×f2C3; S4.4, decode the query token Query de2 With the first layer decoder feature F de1 Calculate the second layer of cross attention Attention2, input the second layer of feedforward network to obtain the decoder feature F de2 , the dimension size is H4×H4×C4; S4.5, decode the query token Query de Filter layer L according to confidence filter The query token Query that needs to be updated in the third layer is calculated with the filter factor f3 de3 , its feature dimension size is H3×H3×f3C3; S4.6, decode the query token Query de3 With the second layer decoder feature F de2 Calculate the third layer cross attention Attention3, input the third layer feedforward network to obtain the decoder feature F de3 , the dimension size is H4×H4×C4; S4.7, decode the query token Query de Filter layer L according to confidence filter The query token Query that needs to be updated in the fourth layer is calculated with the filter factor f4 de4 , its feature dimension size is H3×H3×f4C3; S4.8, decode the query token Query de4 and the third layer decoder feature F de3 Calculate the fourth layer cross attention Attention4, input the fourth layer feedforward network to obtain the decoder feature F de4 , the dimension size is H4×H4×C4; S4.9, update the query feature Query de Generate the target detection bounding box BBox to be selected through feedforward network calculation Candidate Its corresponding classification Cls; S4.10, corresponding to BBox Candidate The final predicted BBox is matched with the true label using the Hungarian matching algorithm, and the comprehensive loss including classification loss, overlap loss, and bounding box loss is calculated and then the weights are updated by backpropagation.

9. The method for detecting mouse tumors based on contrast agent-enhanced magnetic resonance imaging according to claim 8, wherein: The working process of the confidence filtering layer includes the following steps: S5.1, calculate the query token foreground confidence s; S5.2, selecting a filter factor according to the foreground confidence and updating the query tokens corresponding to the high confidence; The confidence calculation process is: s l-1 =MLP(F l-1 (1+UP(α l ×s l ))) Among them, s l represents the confidence of the query token at level l, s l-1 represents the confidence of the query token at level l-1, F l represents the query token feature map at layer l, F l-1 represents the query token feature map at layer l-1, α l Represents a learnable parameter, and UP represents a bilinear interpolation upsampling function.

10. A mouse tumor detection system based on contrast agent-enhanced magnetic resonance imaging, characterized in that: The system includes a data preprocessing enhancement module M p , image embedding module M e , image encoder module M cen , Transformer encoder module M ten and the cross-attention filter decoder module M de The system is configured to implement the mouse tumor detection method based on contrast agent enhanced magnetic resonance imaging according to any one of claims 1 to 9.

Images