Unmarked fluorescence imaging method based on random matrix optics and deep learning
Through the label-free fluorescence imaging method of random matrix optics and deep learning, white light and coherent light illumination are used to acquire image sequences, and MTT images are generated in combination with mixed attention-intensive U-shaped neural networks, which solves the problem of contrast agent dependence in ICG fluorescence imaging. It has achieved markless vascular information recovery and is suitable for clinical surgical operations.
Patent Information
- Application Number
- CN202411319916.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-23
- Publication Date
- 2025-01-03
AI Technical Summary
In clinical surgery, ICG fluorescence imaging requires injection of contrast agents. Some patients are allergic to contrast agents, and ICG fluorescence contrast agents cannot be used after surgical cutting, so the generation of label-free fluorescence average conversion time (MTT) maps cannot be achieved.
Using label-free fluorescence imaging method based on random matrix optics and deep learning, image sequences are obtained using white light and coherent light illumination, deep tissue contrast images are extracted through random matrix optics statistical gating method, combined with mixed attention-intensive U-shaped nerves The network generates MTT maps to achieve recovery of vascular information without the need for contrast agent injection.
Label-free imaging is achieved, and MTT images similar to ICG fluorescence imaging are generated. It is suitable for clinical applications, reduces dependence on contrast agents, and provides a more flexible vascular information recovery solution.
Smart Images
Figure CN119251336A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of imaging technology, and in particular to a label-free fluorescence imaging method based on random matrix optics and deep learning. Background Art
[0002] Indocyanine green (ICG) fluorescence imaging is often used during clinical surgical procedures to obtain vascular information in the surgical area, especially to distinguish between arteries and veins, confirm the direction of blood flow, and develop deep blood vessels. ICG fluorescence imaging requires the injection of contrast agent ICG, which some patients cannot use due to allergies to contrast agents. In clinical use, two consecutive injections require waiting for the contrast agent of the previous injection to be metabolized out of the surgical field (at least 20 minutes). Once surgical cutting and other operations are started, ICG fluorescence imaging cannot be used because the ICG fluorescent contrast agent will leak along the cutting wound to the entire surgical field.
[0003] Therefore, how to use label-free imaging without the need to inject contrast agents to achieve the generation of fluorescence-like mean transition time (MTT) maps is a technical problem that needs to be solved urgently. Summary of the invention
[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a label-free quasi-fluorescence imaging method based on random matrix optics and deep learning. The label-free imaging method without the injection of contrast agent is used to realize the generation of a quasi-fluorescence mean transition time (MTT) diagram.
[0005] The purpose of the present invention can be achieved by the following technical solutions:
[0006] A label-free fluorescence imaging method based on random matrix optics and deep learning, comprising the following steps:
[0007] Using white light for illumination and obtaining a sequence of white light images of the imaging area through a camera; Using coherent light for illumination and obtaining a sequence of coherent light images of the imaging area through a camera;
[0008] The deep tissue contrast image sequence is extracted from the coherent light image sequence by using random matrix optical statistical gating method.
[0009] Selecting a white light image and a deep tissue contrast image at the same time from the white light image sequence and the deep tissue contrast image sequence, and generating an MTT image similar to ICG fluorescence imaging by deep learning for developing a vascular network;
[0010] The deep learning specifically adopts a hybrid attention-intensive U-type neural network to generate an MTT map based on a white light map and a deep tissue contrast map.
[0011] Furthermore, the method further comprises:
[0012] Select blood vessels according to the generated MTT diagram, and draw a curve from point A to point B of each blood vessel along the center line of the blood vessel;
[0013] The MTT data values were extracted along the serial numbers of each point in the drawn curve and fitted into a straight line;
[0014] If the slope of the straight line corresponding to the blood vessel is greater than zero, the blood flow direction is from point A to point B, and an arrow is added to the corresponding curve; if the slope of the straight line corresponding to the blood vessel is less than zero, the blood flow direction is from point B to point A, and an arrow is added to the corresponding curve; if the slope of the straight line corresponding to the blood vessel is equal to zero or the straight line fitting fails, the corresponding curve is marked as a dotted line.
[0015] Furthermore, the main structure of the hybrid attention dense U-type neural network is a U-type neural network, including an encoder path and a decoder path, each layer of the encoder path and the decoder path includes a double-layer dense module and a hybrid attention module, and each layer of the decoder path is also provided with an upsampling layer for upsampling the feature map to the feature map of the corresponding layer of the encoder path. Figure 1 of the same size and merged;
[0016] The double-layer dense module is a two-layer structure connected to each other, each layer includes a local activation batch normalization layer, a transition layer and a bottleneck layer connected in sequence;
[0017] The hybrid attention module includes a channel attention module and a spatial attention module. The channel attention module includes a pooling layer, a multi-layer perceptron and a hidden layer connected in sequence, and outputs a channel attention feature map through an activation function; the spatial attention module includes a pooling layer and a convolution layer connected to each other. The pooling layer outputs spatial feature maps of different channels, and is cascaded together through the convolution layer to generate a spatial attention feature map.
[0018] Furthermore, the local activation batch normalization layer includes a batch normalization layer and a nonlinear activation layer connected in sequence; the transition layer includes a local activation batch normalization layer and a convolution layer connected in sequence; and the bottleneck layer includes a convolution layer with a kernel size of 1×1.
[0019] Furthermore, the number of feature channels on the encoder path doubles from 16 per step, while the image block size decreases from 256 per step by 2; the feature image block size channels and the decoder path adopt a change process opposite to that of the encoder path.
[0020] Furthermore, the cost function calculation process in the hybrid attention dense U-type neural network training process is specifically as follows:
[0021] The hidden layer outputs of each layer of the decoder path are processed through the transition layer and the convolution layer to obtain the corresponding prediction results. Each prediction result is compared with the corresponding true value pixel by pixel to obtain the final cost value.
[0022] Furthermore, the calculation expression of the cost function is:
[0023]
[0024] Where L is the cost, M is the number of layers in the decoder path, and α i is a constant coefficient in the range of 0-1 and satisfies N is the total number of pixels, β j is the pixel coefficient, y j is the true value of pixel j, is the prediction result of pixel j at the i-th layer of the decoder path.
[0025] Further, the hybrid attention intensive U-shaped neural network includes a first processing layer, a first encoding processing layer, a second encoding processing layer, a third encoding processing layer, a fourth encoding processing layer, a first decoding processing layer, a second decoding processing layer, a third decoding processing layer, a fourth decoding processing layer and a second processing layer connected in sequence;
[0026] The first processing layer includes a convolution layer and a local activation batch normalization layer connected in sequence, the first encoding processing layer, the second encoding processing layer, the third encoding processing layer and the fourth encoding processing layer each include a double-layer dense module and a hybrid attention module connected to each other; the first decoding processing layer, the second decoding processing layer, the third decoding processing layer and the fourth decoding processing layer each include a double-layer dense module, an upsampling layer and a hybrid attention module; the second processing layer includes a convolution layer;
[0027] The output of the first processing layer is jump-connected with the output of the fourth decoding processing layer as the input of the second processing layer, the output of the first encoding processing layer is jump-connected with the output of the third decoding processing layer as the input of the fourth encoding processing layer, the output of the second encoding processing layer is jump-connected with the output of the second decoding processing layer as the input of the third decoding processing layer, and the output of the third encoding processing layer is jump-connected with the output of the first decoding processing layer as the input of the second decoding processing layer;
[0028] The output of the first decoding processing layer is processed by three third processing layers and one fourth processing layer and used to calculate the cost function, the output of the second decoding processing layer is processed by two third processing layers and one fourth processing layer and used to calculate the cost function, the output of the third decoding processing layer is processed by one third processing layer and one fourth processing layer and used to calculate the cost function, and the output of the second processing layer is used to calculate the cost function; the third processing layer includes interconnected transition layers and convolution layers, and the fourth processing layer includes convolution layers.
[0029] Furthermore, the method of extracting a deep tissue contrast image sequence from a coherent light image sequence is specifically as follows: using a random matrix optical statistical gating method, from the coherent light image sequence, using a time sliding window strategy, a deep tissue contrast image containing vascular information is extracted for every N images to form a deep tissue contrast image sequence.
[0030] Furthermore, the coherent light is a near-infrared laser, and the central wavelength of the near-infrared laser is between 780nm and 850nm.
[0031] Compared with the prior art, the present invention has the following advantages:
[0032] (1) The present invention uses coherent light illumination, combined with a random matrix optical statistical gating method, to extract a deep tissue contrast image sequence, which can effectively restore vascular information similar to ICG fluorescence imaging, and is combined with a white light image. Both are label-free imaging, that is, no contrast agent is required, and sufficient information is provided for reconstructing / generating a fluorescent MTT image. Finally, by introducing a hybrid attention dense U-type neural network with a dense connection mechanism and a hybrid attention mechanism, the generation of an MTT image similar to ICG fluorescence imaging is achieved, which is suitable for training with small data volumes and provides convenience for clinical applications.
[0033] (2) The hybrid attention dense U-shaped neural network of the present invention is based on a densely connected convolutional network, which can achieve feature reuse and reduce overfitting, thereby obtaining better learning performance with fewer parameters; the decoder path adopts the opposite process. In addition, a hybrid attention module is introduced on the decoder path to perform deep supervision on the hidden layer to optimize the intermediate feature representation and accelerate convergence. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 A flowchart of a label-free fluorescence imaging method based on random matrix optics and deep learning provided in an embodiment of the present invention;
[0035] Figure 2 A schematic flow chart of a label-free fluorescence imaging method based on random matrix optics and deep learning provided in an embodiment of the present invention;
[0036] Figure 3 Schematic diagram of a hybrid attention-intensive U-shaped neural network structure for a label-free fluorescence imaging method based on random matrix optics and deep learning provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0037] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.
[0038] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0039] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.
[0040] Example 1
[0041] like Figure 1 and Figure 2 As shown, this embodiment provides a label-free fluorescence imaging method based on random matrix optics and deep learning, comprising the following steps:
[0042] S1: using white light illumination and obtaining a sequence of white light images of the imaging area through a camera; using coherent light illumination and obtaining a sequence of coherent light images of the imaging area through a camera;
[0043] Coherent light refers to two beams of light that maintain the same phase difference during propagation and have the same frequency or completely consistent waveform. Such light waves can produce stable interference phenomena, including constructive interference and destructive interference. In reality, perfect coherent light does not exist, but a light source close to coherence can be obtained through various methods.
[0044] Optionally, the coherent light is a near-infrared laser, and the central wavelength of the near-infrared laser is between 780nm and 850nm.
[0045] S2: Using random matrix optical statistical gating method, deep tissue contrast image sequence is extracted from coherent light image sequence;
[0046] Specifically, the random matrix optical statistical gating method is used to extract a deep tissue contrast map containing vascular information from every N images using a time sliding window strategy from the coherent light image sequence to form a deep tissue contrast image sequence, which can effectively restore vascular information similar to that in ICG fluorescence imaging.
[0047] The random matrix optical statistical gating method utilizes the statistical properties of coherent light to achieve statistical separation of multiple scattered signals in wide-field illumination reflective imaging.
[0048] S3: Select the white light image and deep tissue contrast image at the same time from the white light image sequence and the deep tissue contrast image sequence (both are label-free imaging, that is, no contrast agent is required), and use deep learning to generate MTT images similar to ICG fluorescence imaging, which are used to develop vascular networks and assist clinical surgical operations;
[0049] Deep learning specifically uses a hybrid attention-intensive U-shaped neural network to generate an MTT map based on the white light image and the deep tissue contrast image.
[0050] Preferably, the method further comprises:
[0051] S4: Select blood vessels according to the generated MTT map, and draw a curve from point A to point B of each blood vessel along the center line of the blood vessel;
[0052] S5: extracting MTT data values along the sequence number of each point in the drawn curve and fitting them into a straight line;
[0053] S6: If the slope of the straight line corresponding to the blood vessel is greater than zero, the blood flow direction is from point A to point B, and an arrow is added to the corresponding curve; if the slope of the straight line corresponding to the blood vessel is less than zero, the blood flow direction is from point B to point A, and an arrow is added to the corresponding curve; if the slope of the straight line corresponding to the blood vessel is equal to zero or the straight line fitting fails, the corresponding curve is marked as a dotted line.
[0054] The hybrid attention dense U-shaped neural network is based on a densely connected convolutional network (DenseNet), which enables feature reuse and reduces overfitting, thereby achieving better learning performance with fewer parameters. The decoder path adopts the opposite process. In addition, a hybrid attention module is introduced on the decoder path to perform deep supervision on the hidden layer to optimize the intermediate feature representation and accelerate convergence.
[0055] Specifically, Figure 3 As shown, the main structure of the hybrid attention dense U-type neural network is a U-type neural network, including an encoder path and a decoder path. Each layer of the encoder path and the decoder path includes a sequentially connected double-layer dense module and a hybrid attention module. Each layer of the decoder path is also provided with an upsampling layer for upsampling the feature map to the feature map of the corresponding layer of the encoder path. Figure 1 of the same size and merged;
[0056] The two-layer dense module is a two-layer structure connected to each other, each layer includes a local activation batch normalization layer, a transition layer and a bottleneck layer connected in sequence;
[0057] The local activation batch normalization layer includes a batch normalization layer and a nonlinear activation layer connected in sequence; the transition layer includes a local activation batch normalization layer and a convolution layer connected in sequence; the bottleneck layer includes a convolution layer with a kernel size of 1×1;
[0058] The hybrid attention module includes a channel attention module and a spatial attention module. The channel attention module includes a pooling layer, a multi-layer perceptron and a hidden layer connected in sequence, and outputs a channel attention feature map through an activation function; the spatial attention module includes a pooling layer and a convolution layer connected to each other. The pooling layer outputs spatial feature maps of different channels, which are cascaded together through the convolution layer to generate a spatial attention feature map.
[0059] The cost function calculation process in the hybrid attention dense U-shaped neural network training process is as follows:
[0060] The hidden layer outputs of each layer of the decoder path are processed through the transition layer and the convolution layer to obtain the corresponding prediction results. Each prediction result is compared with the corresponding true value pixel by pixel to obtain the final cost value.
[0061] This embodiment specifically includes the following steps:
[0062] (1) using white light illumination and an RGB camera to obtain a white light image sequence of the imaging area;
[0063] (2) using near-infrared laser illumination (the central wavelength of the laser is between 780 nm and 850 nm) and a near-infrared camera to obtain a near-infrared light image sequence of the imaging area;
[0064] (3) Using the random matrix optical statistical gating method, a deep tissue contrast image (including vascular information) is extracted from every N images from the near-infrared image sequence using a temporal sliding window strategy to form a deep tissue contrast image sequence;
[0065] (4) taking out one image from each of the white light image sequence and the deep tissue contrast image sequence according to the time correspondence;
[0066] (5) The two images are input into a hybrid attention dense U-shaped neural network, which outputs a fluorescent MTT-like image for visualizing the vascular network and distinguishing arteries from veins;
[0067] (6) From a blood vessel in the MTT diagram, manually draw a curve from point A to point B along the center line of the blood vessel, marked in white;
[0068] (7) Extract the MTT data value along the point number of the curve and fit it into a straight line y=ax+b; where x is the number of the point where the curve is manually drawn, y is the MTT data value, and a and b are the fitted parameters;
[0069] (8) When a>0, the blood flow direction is from point A to point B, and an arrow is added to the curve; when a<0, the blood flow direction is from point B to point A, and an arrow is added to the curve. When a=0 or the fitting fails, the curve is marked as a dotted line.
[0070] Preferably, in this embodiment, on the encoder path, the number of feature channels is doubled from 16, and the image block size is reduced by 2 steps from 256. The decoder path adopts the opposite process. In addition, the hidden layer is deeply supervised on the decoder path to optimize the intermediate feature representation and accelerate convergence.
[0071] In the two-layer dense module, the number of channels in each layer increases at a rate of k=N / 2, where N represents the number of input channels. Each layer contains a local activation batch normalization (ABN) layer and a transition layer. ABN is an integration of batch normalization (BN) layers and nonlinear activation layers to reduce computational memory consumption. The transition layer includes an ABN layer and a convolutional layer to achieve downsampling of feature map channels. Each layer also contains a bottleneck layer, which is a convolutional layer with a kernel size of 1×1, thereby reducing the number of channels of the input feature map to 4k, which greatly reduces the computational complexity.
[0072] The hybrid attention module includes a channel attention module and a spatial attention module. In the decoder part, the hybrid attention module upsamples the feature map of the intermediate layer and outputs a channel attention feature map and a spatial attention feature map of uniform size through a convolutional layer. The channel attention module contains a pooling layer, a multi-layer perceptron and a hidden layer, which uses a sigmoid activation function to output a channel attention feature map. The spatial attention module contains a pooling layer and a convolutional layer. The pooling layer outputs spatial feature maps of different channels. These spatial feature maps are cascaded together and passed through the convolutional layer to finally generate a spatial attention feature map.
[0073] The calculation expression of the cost function is:
[0074]
[0075] Where L is the cost, M is the number of layers in the decoder path, and α i is a constant coefficient in the range of 0-1 and satisfies N is the total number of pixels, βj is the pixel coefficient, y j is the true value of pixel j, is the prediction result of pixel j at the i-th layer of the decoder path.
[0076] Specifically, the hybrid attention dense U-shaped neural network includes a first processing layer, a first encoding processing layer, a second encoding processing layer, a third encoding processing layer, a fourth encoding processing layer, a first decoding processing layer, a second decoding processing layer, a third decoding processing layer, a fourth decoding processing layer and a second processing layer connected in sequence;
[0077] The first processing layer includes a convolution layer and a local activation batch normalization layer connected in sequence, the first encoding processing layer, the second encoding processing layer, the third encoding processing layer and the fourth encoding processing layer each include a double-layer dense module and a hybrid attention module connected to each other; the first decoding processing layer, the second decoding processing layer, the third decoding processing layer and the fourth decoding processing layer each include a double-layer dense module, an upsampling layer and a hybrid attention module; the second processing layer includes a convolution layer;
[0078] The output of the first processing layer is jump-connected with the output of the fourth decoding processing layer as the input of the second processing layer, the output of the first encoding processing layer is jump-connected with the output of the third decoding processing layer as the input of the fourth encoding processing layer, the output of the second encoding processing layer is jump-connected with the output of the second decoding processing layer as the input of the third decoding processing layer, and the output of the third encoding processing layer is jump-connected with the output of the first decoding processing layer as the input of the second decoding processing layer;
[0079] The output of the first decoding processing layer is processed by three third processing layers and one fourth processing layer and used to calculate the cost function, the output of the second decoding processing layer is processed by two third processing layers and one fourth processing layer and used to calculate the cost function, the output of the third decoding processing layer is processed by one third processing layer and one fourth processing layer and used to calculate the cost function, and the output of the second processing layer is used to calculate the cost function; the third processing layer includes interconnected transition layers and convolution layers, and the fourth processing layer includes convolution layers.
[0080] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.
Claims
1. A label-free fluorescence imaging method based on random matrix optics and deep learning, characterized in that: The following steps are involved: Using white light for illumination and obtaining a sequence of white light images of the imaging area through a camera; Using coherent light for illumination and obtaining a sequence of coherent light images of the imaging area through a camera; The deep tissue contrast image sequence is extracted from the coherent light image sequence by using random matrix optical statistical gating method. Selecting a white light image and a deep tissue contrast image at the same time from the white light image sequence and the deep tissue contrast image sequence, and generating an MTT image similar to ICG fluorescence imaging by deep learning for developing a vascular network; The deep learning specifically adopts a hybrid attention-intensive U-type neural network to generate an MTT map based on a white light map and a deep tissue contrast map.
2. A label-free fluorescence imaging method based on random matrix optics and deep learning according to claim 1, characterized in that: The method further comprises: Select blood vessels according to the generated MTT diagram, and draw a curve from point A to point B of each blood vessel along the center line of the blood vessel; The MTT data values were extracted along the serial numbers of each point in the drawn curve and fitted into a straight line; If the slope of the straight line corresponding to the blood vessel is greater than zero, the blood flow direction is from point A to point B, and an arrow is added to the corresponding curve; if the slope of the straight line corresponding to the blood vessel is less than zero, the blood flow direction is from point B to point A, and an arrow is added to the corresponding curve; if the slope of the straight line corresponding to the blood vessel is equal to zero or the straight line fitting fails, the corresponding curve is marked as a dotted line.
3. A label-free fluorescence imaging method based on random matrix optics and deep learning according to claim 1, characterized in that: The main structure of the hybrid attention dense U-type neural network is a U-type neural network, including an encoder path and a decoder path, each layer of the encoder path and the decoder path includes a double-layer dense module and a hybrid attention module, and each layer of the decoder path is also provided with an upsampling layer for upsampling the feature map to the same size as the feature map of the corresponding layer of the encoder path, and fusing them; The double-layer dense module is a two-layer structure connected to each other, each layer includes a local activation batch normalization layer, a transition layer and a bottleneck layer connected in sequence; The hybrid attention module includes a channel attention module and a spatial attention module. The channel attention module includes a pooling layer, a multi-layer perceptron and a hidden layer connected in sequence, and outputs a channel attention feature map through an activation function; the spatial attention module includes a pooling layer and a convolution layer connected to each other. The pooling layer outputs spatial feature maps of different channels, and is cascaded together through the convolution layer to generate a spatial attention feature map.
4. A label-free fluorescence imaging method based on random matrix optics and deep learning according to claim 3, characterized in that: The local activation batch normalization layer includes a batch normalization layer and a nonlinear activation layer connected in sequence; the transition layer includes a local activation batch normalization layer and a convolution layer connected in sequence; the bottleneck layer includes a convolution layer with a kernel size of 1×1.
5. The label-free fluorescence imaging method based on random matrix optics and deep learning according to claim 3, characterized in that: The number of feature channels on the encoder path doubles from 16 per step, while the image block size decreases by 2 per step from 256; the feature image block size channels and the decoder path adopt a change process opposite to that of the encoder path.
6. The label-free fluorescence imaging method based on random matrix optics and deep learning according to claim 3, characterized in that: The cost function calculation process in the hybrid attention dense U-type neural network training process is specifically as follows: The hidden layer outputs of each layer of the decoder path are processed through the transition layer and the convolution layer to obtain the corresponding prediction results. Each prediction result is compared with the corresponding true value pixel by pixel to obtain the final cost value.
7. A label-free fluorescence imaging method based on random matrix optics and deep learning according to claim 6, characterized in that: The calculation expression of the cost function is: Where L is the cost, M is the number of layers in the decoder path, and α i is a constant coefficient in the range of 0-1 and satisfies N is the total number of pixels, β j is the pixel coefficient, y j is the true value of pixel j, is the prediction result of pixel j at the i-th layer of the decoder path.
8. The label-free fluorescence imaging method based on random matrix optics and deep learning according to claim 3, characterized in that: The hybrid attention intensive U-shaped neural network comprises a first processing layer, a first encoding processing layer, a second encoding processing layer, a third encoding processing layer, a fourth encoding processing layer, a first decoding processing layer, a second decoding processing layer, a third decoding processing layer, a fourth decoding processing layer and a second processing layer connected in sequence; The first processing layer includes a convolution layer and a local activation batch normalization layer connected in sequence, the first encoding processing layer, the second encoding processing layer, the third encoding processing layer and the fourth encoding processing layer each include a double-layer dense module and a hybrid attention module connected to each other; the first decoding processing layer, the second decoding processing layer, the third decoding processing layer and the fourth decoding processing layer each include a double-layer dense module, an upsampling layer and a hybrid attention module; the second processing layer includes a convolution layer; The output of the first processing layer is jump-connected with the output of the fourth decoding processing layer as the input of the second processing layer, the output of the first encoding processing layer is jump-connected with the output of the third decoding processing layer as the input of the fourth encoding processing layer, the output of the second encoding processing layer is jump-connected with the output of the second decoding processing layer as the input of the third decoding processing layer, and the output of the third encoding processing layer is jump-connected with the output of the first decoding processing layer as the input of the second decoding processing layer; The output of the first decoding processing layer is processed by three third processing layers and one fourth processing layer and used to calculate the cost function, the output of the second decoding processing layer is processed by two third processing layers and one fourth processing layer and used to calculate the cost function, the output of the third decoding processing layer is processed by one third processing layer and one fourth processing layer and used to calculate the cost function, and the output of the second processing layer is used to calculate the cost function; the third processing layer includes interconnected transition layers and convolution layers, and the fourth processing layer includes convolution layers.
9. The label-free fluorescence imaging method based on random matrix optics and deep learning according to claim 1, characterized in that: The method of extracting a deep tissue contrast image sequence from a coherent light image sequence specifically comprises: utilizing a random matrix optical statistical gating method, using a time sliding window strategy, extracting a deep tissue contrast image containing vascular information from every N images from the coherent light image sequence, and forming a deep tissue contrast image sequence.
10. The label-free fluorescence imaging method based on random matrix optics and deep learning according to claim 1, characterized in that: The coherent light is near-infrared laser, and the central wavelength of the near-infrared laser is between 780nm and 850nm.