Fast X-ray dynamic real-time imaging denoising method based on deep learning

Abstract

The present invention discloses a fast X-ray dynamic real-time imaging denoising method based on deep learning, comprising the following steps: S1, preprocessing clinically collected X-ray dynamic real-time video data to obtain a data set; S2, constructing a deep learning model including at least one denoising module, wherein the at least one denoising module forms a one-layer network or a cascaded two-layer network; S3, using the data set to train the deep learning model; S4, inputting several continuous frames of low-radiation-dose X-ray dynamic real-time video into the deep learning model trained in step S3, and outputting the denoising results of the intermediate frames of several continuous frames. The fast X-ray dynamic real-time imaging video denoising method based on deep learning proposed by the present invention greatly shortens the denoising time while ensuring the denoising effect, and can be applied to real-time denoising of surgical videos that cannot be achieved by previous methods.

Description

Technical Field

[0001] The present invention relates to the technical field of video quality enhancement, and in particular to a method for denoising fast X-ray dynamic real-time imaging videos during surgery based on deep learning. Background Art

[0002] X-ray dynamic real-time imaging guided fluoroscopy systems are used to guide the surgeon in the use of surgical instruments during the dissection process. The real-time feedback provided by these fluoroscopy systems can capture the surgeon's precise movements and display them on the monitor in the operating room. Therefore, this dynamic real-time X-ray imaging has become the standard for many surgical procedures because it can reduce invasiveness while giving the surgeon more control over the operation, thereby reducing trauma and interference with tissues.

[0003] However, these fluoroscopic systems require sufficient radiation dose to obtain high-quality videos with sufficient tissue contrast. This radiation exposure is harmful to both patients and surgeons: it can cause DNA damage, accelerated aging, hair loss, burns and tissue necrosis in patients; surgeons exposed to radiation will have a higher risk of cancer. As for the countermeasures, in addition to using protective equipment, the radiation dose should be as low as possible while ensuring the function of the X-ray dynamic real-time imaging system. However, the reality is that the fluoroscopic video produced at a lower dose of radiation intensity has extremely high noise and is not suitable for clinical use. Therefore, a suitable video denoising method is particularly critical.

[0004] Currently, in the field of video denoising, the processing methods are mainly divided into two categories: one is patch-based algorithms, such as VBM4D and VNLB; the other is deep learning-based algorithms, such as DVDnet. However, both types of denoising algorithms have the problem that they cannot be applied to real-time video processing due to their long processing time. On the other hand, for deep learning-based algorithms, it is difficult to produce data sets for training in dynamic real-time X-ray videos because safe radiation doses cannot produce real data that is close to noise-free. Summary of the invention

[0005] The main purpose of the present invention is to overcome the shortcomings of the prior art and propose a fast X-ray dynamic real-time imaging denoising method based on deep learning to solve the problem that the X-ray fluoroscopy video produced under low-dose radiation intensity has extremely high noise and cannot be used for clinical guidance of surgery.

[0006] The technical solution proposed by the present invention to achieve the above-mentioned purpose is as follows:

[0007] A fast X-ray dynamic real-time imaging denoising method based on deep learning comprises the following steps:

[0008] S1. Preprocess the clinically collected X-ray dynamic real-time video data to obtain a data set; S2. Construct a deep learning model including at least one denoising module, wherein the at least one denoising module forms a one-layer network or a cascaded two-layer network; S3. Use the data set to train the deep learning model; S4. Input several consecutive frames of low-radiation-dose X-ray dynamic real-time video into the deep learning model trained in step S3, and output the denoising results of the intermediate frames of the several consecutive frames.

[0009] Furthermore, step S1 includes: S11, obtaining dynamic real-time video data of X-rays under normal radiation dose, recorded as B1; S12, performing fine denoising processing on B1 to obtain simulated noise-free data, recorded as B2; S13, subtracting B1 from B2 to obtain simulated normal radiation dose noise, recorded as C1; S14, determining a constant factor c used to characterize the multiple relationship between low radiation dose noise and normal radiation dose noise, multiplying C1 by the constant factor c to obtain simulated low radiation dose noise, recorded as C2; wherein c>1; S15, adding C2 to B2 to obtain simulated low radiation dose video data, recorded as B3; S16, using several consecutive video frames in B3 as input of the network during training, and the intermediate frames in the corresponding several consecutive video frames in B2 as the expected output of the network during training, to form data pairs, and multiple data pairs constitute the data set.

[0010] Furthermore, the fine denoising process described in step S12 includes: firstly, performing time domain averaging processing on B1, that is, normalizing adjacent frames in a neighborhood at the pixel scale, wherein for video intervals containing motion, the neighborhood range is reduced; then, performing denoising processing through the VBM4D algorithm to obtain B2.

[0011] Furthermore, the denoising module is a U-Net multi-layer network using residual learning; and the denoising module can receive a number of consecutive frame inputs and estimate the denoising results of the intermediate frames of the consecutive frames.

[0012] Furthermore, the network structure of the denoising module is, from the input end to the output end, a first encoding module, a second encoding module, a bottleneck module, a first decoding module and a second decoding module in sequence: the first encoding module serves as an input port and can receive a plurality of consecutive frame inputs. At the same time, the intermediate frames of the plurality of consecutive frames are also directly sent to the output of the second decoding module to perform a pixel-level addition operation with the output of the second decoding module, and the addition result is used as the final output frame of the denoising module; the output of the first encoding module serves as the input of the second encoding module, and at the same time, the output of the first encoding module is also directly connected to the output of the first decoding module to perform a pixel-level addition operation with the output of the first decoding module; the output of the second encoding module serves as the input of the bottleneck module, and at the same time, the output of the second encoding module is also directly connected to the output of the bottleneck module to perform a pixel-level addition operation with the output of the bottleneck module; the result obtained by performing a pixel-level addition operation on the output of the second encoding module and the output of the bottleneck module serves as the input of the first decoding module, and the result obtained by performing a pixel-level addition operation on the output of the first encoding module and the output of the first decoding module serves as the input of the second decoding module.

[0013] Furthermore, the network structure of the denoising module includes four convolutional networks, namely: a first convolutional network formed by connecting a convolutional layer, a BN normalization operation and a ReLU excitation operation in series; a second convolutional network formed by connecting a convolutional layer, a BN normalization operation and a ReLU excitation operation in series and having a step size of 2; a third convolutional network with an upsampling operation in series after the convolutional layer; a fourth convolutional network with only a convolutional layer and an output channel number of 1; wherein: the first encoding module is formed by connecting two of the first convolutional networks in series, wherein the first convolutional network in front can receive several continuous frame inputs and has an output channel number of 3; the second encoding module is formed by connecting one of the second convolutional networks and two of the first convolutional networks in series; the bottleneck module is formed by connecting one of the second convolutional networks, four of the first convolutional networks and one of the third convolutional networks in series; the first decoding module is formed by connecting two of the first convolutional networks and one of the third convolutional networks in series; the second decoding module is formed by connecting one of the first convolutional networks and one of the fourth convolutional networks in series.

[0014] Furthermore, in the first encoding module, the number of output channels of the first convolutional network is 32; in the second encoding module, the number of output channels of each first convolutional network and the second convolutional network is 64; in the bottleneck module, the number of output channels of each first convolutional network and the second convolutional network is 128, and the number of output channels of the third convolutional network is 256; in the first decoding module, the number of output channels of each first convolutional network is 64, and the number of output channels of the third convolutional network is 128; in the second decoding module, the number of output channels of the first convolutional network is 32.

[0015] Furthermore, the deep learning model is a single-layer network including a denoising module, and the denoising module receives continuous multi-frame inputs and outputs denoising results of intermediate frames in the continuous multi-frame inputs as the final output of the deep learning model.

[0016] Furthermore, the deep learning model is a two-layer network formed by cascading multiple denoising modules, wherein the second layer of the network is composed of a denoising module; each denoising module of the first layer of the network respectively receives several consecutive frames in the continuous multi-frame input, and the denoising results of each denoising module of the first layer of the network are simultaneously input into the denoising module of the second layer of the network, and the output of the denoising module of the second layer of the network is used as the final output of the deep learning model.

[0017] Furthermore, when training the deep learning model in step S3, the loss function adopts L1 loss, and is optimized using the Adam optimizer, with the learning rate set to 0.001 and the cycle to 40.

[0018] Compared with the prior art, the beneficial effects of the present invention are: on the one hand, the data set preparation method proposed in the present invention makes the training data noise distribution of the deep learning model closer to the actual low-radiation dose X-ray dynamic real-time imaging video noise distribution, which is beneficial to the application in real scenes; on the other hand, the existing methods (including deep learning methods) use explicit motion estimation and motion compensation processes when processing the temporal correlation of continuous video frames. These steps require a lot of calculations and are very time-consuming. The method for fast X-ray dynamic real-time imaging video denoising based on deep learning proposed in the present invention does not have the explicit process of motion estimation and motion compensation, and directly allows the neural network to learn the correlation of the time scale of video frames through continuous multi-frame input. On the basis of ensuring the denoising effect, the denoising time is greatly shortened, and can be applied to real-time denoising of surgical videos that cannot be achieved by previous methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of a fast X-ray dynamic real-time imaging denoising method based on deep learning proposed in an embodiment of the present invention;

[0020] Figure 2 is a flow chart of a method for preparing a data set for training a deep learning model according to an embodiment of the present invention;

[0021] Figure 3 is a deep learning model architecture diagram for a denoising method proposed in an embodiment of the present invention;

[0022] Figure 4 is an exemplary network architecture diagram of a denoising module proposed in an embodiment of the present invention;

[0023] Figure 5 This is a comparison diagram of the effects of denoising dynamic real-time X-ray imaging using the method of the present invention. DETAILED DESCRIPTION

[0024] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. It should be emphasized that the following description is only exemplary and is not intended to limit the scope of the present invention and its application.

[0025] Non-limiting and non-exclusive embodiments will be described with reference to the following figures, wherein like reference numerals refer to like parts unless otherwise specifically stated.

[0026] The present invention proposes a fast X-ray dynamic real-time imaging denoising method based on deep learning, such as Figure 1 As shown, the denoising method mainly includes the following steps S1 to S4:

[0027] Step S1: preprocess the clinically collected X-ray dynamic real-time video data to obtain a data set, which will be used to train the deep learning model proposed in the present invention.

[0028] Step S2, construct a deep learning model including at least one denoising module, wherein the at least one denoising module forms a one-layer network or a cascaded two-layer network. In some embodiments, the deep learning model is a single-layer network consisting of one denoising module; in other embodiments, the deep learning model is a two-layer network formed by cascading multiple denoising modules. The denoising module can be a U-Net multi-layer network using residual learning; it can also be a residual network of other structures.

[0029] Step S3, use the data set of step S1 to train the deep learning model constructed in step S2. The loss function during training can adopt L1 loss, that is, the absolute value difference between the network output and the noise-free data. It should be understood that L2 loss can also be used. In addition, since the deep learning model of the present invention is built under the Pytorch framework, the Adam optimizer is used for optimization during training, and the learning rate is set to 0.001 and the cycle is 40.

[0030] Step S4: input several consecutive frames of low-radiation-dose X-ray dynamic real-time video into the deep learning model trained in step S3, and output the denoising results of the intermediate frames of the several consecutive frames.

[0031] In some embodiments, reference Figure 2 The process of preprocessing the clinically collected X-ray dynamic real-time video data to obtain a data set in step S1 specifically includes:

[0032] S11. Acquire dynamic real-time X-ray video data under normal radiation dose, denoted as B1. This step requires acquiring dynamic real-time X-ray video data under normal radiation dose in a large range. The so-called large range refers to the range of variation of indicators such as video resolution, video frame number, and video frame rate. In one embodiment, the number of frames of the collected video data is between 20 and 400, the frame resolution includes three resolutions of 768*768, 1024*1024, and 1536*1536, and the video frame rate is between 4 and 30fps.

[0033] S12, perform fine denoising on B1 to obtain simulated noise-free data, recorded as B2. Fine denoising here means that since this is a step to make a data set for training, the denoising step does not require real-time performance and allows a longer running time. Specifically, first perform time-domain averaging on B1, that is, the adjacent frames of a neighborhood are normalized on a pixel scale, and the neighborhood range is appropriately reduced for video intervals containing motion; then perform denoising through the VBM4D algorithm, which takes a longer time, and finally obtain approximately noise-free data, that is, simulated noise-free data B2.

[0034] S13. Subtract B1 from B2 to obtain the simulated normal radiation dose noise, recorded as C1.

[0035] S14. Determine a constant factor c used to characterize the multiple relationship between low radiation dose noise and normal radiation dose noise, multiply C1 by the constant factor c, and obtain the simulated low radiation dose noise, denoted as C2; where c>1. Existing experimental studies have shown that noise is inversely proportional to the square root of radiation dose. Therefore, when the low radiation dose noise is set to be c times the normal radiation dose noise, the normal radiation dose is correspondingly c times the low radiation dose. 2 In some embodiments, c=2 or 3.

[0036] S15, adding C2 and B2 to obtain simulated low radiation dose video data, recorded as B3;

[0037] S16, using several continuous video frames in B3 as the input of the network during training, and the intermediate frames in the corresponding several continuous video frames in B2 as the expected output of the network during training, to form a data pair, and a plurality of the data pairs constitute the data set. When training a deep learning model, the data set used is composed of many such data pairs, and each data pair is composed of a simulated low radiation dose video data and a simulated noise-free data.

[0038] like Figure 3 The figure shows an exemplary deep learning model of the present invention, which is a two-layer network structure composed of multiple denoising modules in cascade, wherein the first layer of the network includes three denoising modules, namely denoising modules 1, 2 and 3, and the second layer includes a denoising module 4. The outputs of the three denoising modules in the first layer are simultaneously input to the denoising module 4 in the second layer, and the output of the denoising module 4 in the second layer is used as the final output of the deep learning model. In a preferred embodiment, it can be considered that the role of each denoising module is to take three consecutive frames as input and estimate the denoising result of the middle frame.

[0039] For example, Figure 3 In the example, three denoising modules in the first layer are cascaded with one denoising module in the second layer, and five consecutive video frames can be input, of which three consecutive input frames t-2, t-1 and t are input to denoising module 1, and denoising module 1 estimates the denoising result of the middle frame t-1; similarly, three consecutive input frames t-1, t and t+1 are input to denoising module 2, and denoising module 2 estimates the denoising result of the middle frame t; three consecutive input frames t, t+1 and t+2 are input to denoising module 3, and denoising module 3 estimates the denoising result of the middle frame t+1. Then, the denoising results of the three denoising modules 1, 2, and 3 are simultaneously sent to denoising module 4, and denoising module 4 estimates the denoised frame of frame t as the final output frame. In other words, the cascade network can utilize more temporal information to improve the denoising effect.

[0040] In other embodiments, the deep learning model may also have only one layer of network, that is, it is composed of only one denoising module. For example, 5 consecutive frames can be input, and then the denoising result of the middle frame, that is, the third frame, is estimated as the output. Therefore, it should be understood that the deep learning model of the present invention, when denoising the original video, the input of each denoising process is a continuous odd frame, and the output is the denoising result of the middle frame. In this way, continuous odd frames are continuously input, and denoising results are continuously output, so that real-time denoised frames can be continuously output without waiting for the entire video to be denoised before output, that is, the effect of real-time denoising of real-time imaging video during surgery is achieved.

[0041] like Figure 4As shown, the denoising module takes the U-Net multi-level network as an example, and may include a first encoding module 100, a second encoding module 200, a bottleneck module 300, a first decoding module 400 and a second decoding module 500 from the input end to the output end. The first encoding module 100 serves as an input port and can receive continuous multi-frame inputs, such as Figure 3 In accordance with the example shown, three consecutive frames can be input; at the same time, the intermediate frames of the consecutive multi-frame input are also directly sent to the output of the second decoding module 500 to perform pixel-level addition operation with the output of the second decoding module 500, and the result of the addition is used as the final output frame of the denoising module, thereby realizing residual learning. Figure 4 The denoising module 2 of the structure shown has the principle of residual learning as follows: after the second decoding module 500 outputs the denoised frame of frame t, the original frame t to be estimated is directly sent to the output end of the second decoding module 500 to perform a pixel-level addition operation with the denoised frame of frame t as the denoising result of the denoising module, thereby realizing residual learning. The output of the first encoding module 100 is used as the input of the second encoding module 200, and at the same time, the output of the first encoding module 100 is also directly connected to the output of the first decoding module 400 to perform a pixel-level addition operation with the output of the first decoding module 400. The output of the second encoding module 200 is used as the input of the bottleneck module 300, and at the same time, the output of the second encoding module 200 is also directly connected to the output of the bottleneck module 300 to perform a pixel-level addition operation with the output of the bottleneck module 300. The result of pixel-level addition operation on the output of the second encoding module 200 and the output of the bottleneck module 300 is used as the input of the first decoding module 400, and the result of pixel-level addition operation on the output of the first encoding module 100 and the output of the first decoding module 400 is used as the input of the second decoding module 500.

[0042] The encoding part of the denoising module is the process of encoding the original input image into a smaller-sized feature map, and the decoding part of the denoising module is to restore the small-sized feature map to the original-sized image. Therefore, for general X-ray dynamic real-time video data, the number of output channels of the first layer of the denoising module can be 3, and the number of output channels of the last layer of the network can be 1. Figure 4 ,In the network structure of the denoising module, four convolutional networks are included, namely:

[0043] The first convolutional network is formed by connecting the convolutional layer, batch normalization (BN) and ReLU excitation operation in series, that is, Figure 4 The second convolutional network with a step size of 2 is formed by connecting the convolutional layer, batch normalization operation and ReLU excitation operation in series, that is, Figure 4The rounded rectangle in the figure represents the third convolutional network with PixelShuffle upsampling operation in series after the convolutional layer, that is, Figure 4 The last one is the last network layer of the denoising module: the fourth convolutional network consisting of only a single convolutional layer, that is, Figure 4 Indicated by the oval.

[0044] On the basis of the above division, the first encoding module 100 is formed by two of the first convolutional networks connected in series, wherein the first convolutional network in front can receive continuous multi-frame inputs and has an output channel number of 3, while the output channel number of the first convolutional network in the back can be 32. The second encoding module 200 is formed by one of the second convolutional networks and two of the first convolutional networks connected in series, wherein the number of output channels of each of the first convolutional network and the second convolutional network is 64. The bottleneck module 300 is formed by one of the second convolutional networks, four of the first convolutional networks and one of the third convolutional networks connected in series, wherein the number of output channels of each of the first convolutional network and the second convolutional network is 128, and the number of output channels of the third convolutional network is 256. The first decoding module 400 is formed by two of the first convolutional networks and one of the third convolutional networks connected in series, wherein the number of output channels of each of the first convolutional networks is 64, and the number of output channels of the third convolutional network is 128. The second decoding module 500 is composed of one of the first convolutional networks and one of the fourth convolutional networks connected in series, wherein the number of output channels of the first convolutional network is 32.

[0045] based on Figure 3 and Figure 4 The denoising method of the network structure shown in the figure uses residual learning for each denoising module of the deep learning model: the intermediate noisy frames are directly connected to the output, making it easier for the network to learn the residual information.

[0046] like Figure 5 As shown, it is a comparison diagram of the effect of denoising dynamic real-time X-ray imaging using the denoising method of the present invention, where (a), (b), (c), and (d) are respectively: simulated noise-free data (i.e., the expected output during training), normal radiation dose data (with normal noise), low radiation dose data (with a lot of noise), and the denoising result of the method of the present invention. Among them:

[0047] TSNM (time-series noise metric): An evaluation metric used to estimate temporal and spatial noise when there is no control video. In this experiment, it is normalized so that the TSNM metric for low radiation dose data is 1 (the lower the value of this metric, the less noise). Figure 5It can be seen that the noise evaluation index of the normal radiation dose data (b) is 0.5; the noise evaluation index of the denoising result (d) of the method of the present invention is 0.42, that is, after denoising the input low radiation dose data (c), an image with less noise than the normal radiation dose data can be obtained, thereby making it possible to image at low radiation doses and denoise in real time to obtain clearer imaging.

[0048] SSIM (structural similarity index measure): Structural similarity index is an index used to measure the similarity between two digital images. It is more consistent with the human eye's judgment of image quality in measuring image quality (the closer the index is to 1, the more similar they are). Figure 5 It can be seen that the index of the denoising result (d) of the method of the present invention is 0.8, which is better than the image quality of the normal radiation dose data (b), and the index of the normal radiation dose data (b) is only 0.74.

[0049] PSNR (Peak signal-to-noise ratio): Peak signal-to-noise ratio is an engineering term that represents the ratio of the maximum possible signal power to the destructive noise power that affects its representation accuracy. This indicator focuses more on the difference in pixel level with the reference data (the larger the value of this indicator, the more similar it is). Figure 5 It can be seen that the PSNR index of the denoising result (d) of the method of the present invention is 28.45, which is also greater than 25.91 of the normal radiation dose data (b).

[0050] In summary, the denoising method of the present invention is used to denoise X-ray dynamic real-time imaging under low radiation dose, which can not only denoise quickly and in real time, but also obtain better image effects than X-ray dynamic real-time imaging under normal radiation dose.

[0051] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art of the present invention, several equivalent substitutions or obvious variations can be made without departing from the concept of the present invention, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A fast X-ray dynamic real-time imaging denoising method based on deep learning, characterized in that: The steps include: S1, preprocessing the clinically collected X-ray dynamic real-time video data to obtain a data set; S2. Constructing a deep learning model including at least one denoising module, wherein the at least one denoising module forms a one-layer network or a cascaded two-layer network; S3. Using the data set to train the deep learning model; S4, inputting a plurality of consecutive frames of low-radiation-dose X-ray dynamic real-time video into the deep learning model trained in step S3, and outputting denoising results of intermediate frames of the plurality of consecutive frames; Step S1 includes: S11, obtaining dynamic real-time video data of X-rays under normal radiation dose, recorded as B1; S12, perform fine denoising on B1 to obtain simulated noise-free data, recorded as B2; S13, subtract B1 from B2 to obtain the simulated normal radiation dose noise, recorded as C1; S14, determining a constant factor c used to characterize the multiple relationship between low radiation dose noise and normal radiation dose noise, multiplying C1 by the constant factor c to obtain simulated low radiation dose noise, recorded as C2; wherein c>1; S15, adding C2 and B2 to obtain simulated low radiation dose video data, recorded as B3; S16. Using several continuous video frames in B3 as input of the network during training and the corresponding intermediate frames in several continuous video frames in B2 as expected output of the network during training to form data pairs, and a plurality of the data pairs constitute the data set.

2. The fast X-ray dynamic real-time imaging denoising method based on deep learning as claimed in claim 1, characterized in that: The fine denoising process described in step S12 includes: first, performing time domain averaging processing on B1, that is, normalizing adjacent frames in a neighborhood at the pixel scale, wherein for video intervals containing motion, the neighborhood range is reduced; then, performing denoising processing through the VBM4D algorithm to obtain B2.

3. The fast X-ray dynamic real-time imaging denoising method based on deep learning as claimed in claim 1, characterized in that: The denoising module is a U-Net multi-level network that uses residual learning; and the denoising module can receive a number of consecutive frame inputs and estimate the denoising results of the intermediate frames of the consecutive frames.

4. The fast X-ray dynamic real-time imaging denoising method based on deep learning as claimed in claim 3, characterized in that: The network structure of the denoising module is, from the input end to the output end, a first encoding module, a second encoding module, a bottleneck module, a first decoding module, and a second decoding module: The first encoding module serves as an input port and can receive a plurality of consecutive frame inputs. Meanwhile, the intermediate frames of the plurality of consecutive frames are directly sent to the output of the second decoding module to perform a pixel-level addition operation with the output of the second decoding module, and the addition result is used as the final output frame of the denoising module; The output of the first encoding module is used as the input of the second encoding module. Meanwhile, the output of the first encoding module is also directly connected to the output of the first decoding module to perform a pixel-level addition operation with the output of the first decoding module; The output of the second encoding module is used as the input of the bottleneck module. Meanwhile, the output of the second encoding module is also directly connected to the output of the bottleneck module to perform a pixel-level addition operation with the output of the bottleneck module. The result obtained by performing a pixel-level addition operation on the output of the second encoding module and the output of the bottleneck module is used as the input of the first decoding module, and the result obtained by performing a pixel-level addition operation on the output of the first encoding module and the output of the first decoding module is used as the input of the second decoding module.

5. The fast X-ray dynamic real-time imaging denoising method based on deep learning as claimed in claim 4, characterized in that: The network structure of the denoising module includes four convolutional networks, namely: a first convolutional network formed by sequentially connecting a convolutional layer, a BN normalization operation, and a ReLU excitation operation; a second convolutional network formed by sequentially connecting a convolutional layer, a BN normalization operation, and a ReLU excitation operation with a step size of 2; a third convolutional network with an upsampling operation connected in series after the convolutional layer; and a fourth convolutional network with only a convolutional layer and an output channel number of 1. in: The first encoding module is formed by connecting two of the first convolutional networks in series, wherein the first convolutional network in front can receive a number of continuous frame inputs and has 3 output channels; The second encoding module is composed of one of the second convolutional networks and two of the first convolutional networks connected in series in sequence; The bottleneck module is composed of one of the second convolutional networks, four of the first convolutional networks, and one of the third convolutional networks connected in series in sequence; The first decoding module is formed by connecting two of the first convolutional networks and one of the third convolutional networks in series; The second decoding module is composed of one of the first convolutional networks and one of the fourth convolutional networks connected in series.

6. The fast X-ray dynamic real-time imaging denoising method based on deep learning as claimed in claim 5, characterized in that: In the first encoding module, the number of output channels of the first convolutional network is 32; In the second encoding module, the number of output channels of each of the first convolutional network and the second convolutional network is 64; In the bottleneck module, the number of output channels of each of the first convolutional network and the second convolutional network is 128, and the number of output channels of the third convolutional network is 256; In the first decoding module, the number of output channels of each first convolutional network is 64, and the number of output channels of the third convolutional network is 128; In the second decoding module, the number of output channels of the first convolutional network is 32.

7. The fast X-ray dynamic real-time imaging denoising method based on deep learning according to any one of claims 3 to 6, characterized in that: The deep learning model is a single-layer network including a denoising module, and the denoising module receives continuous multi-frame inputs and outputs denoising results of intermediate frames in the continuous multi-frame inputs as the final output of the deep learning model.

8. The fast X-ray dynamic real-time imaging denoising method based on deep learning according to any one of claims 3 to 6, characterized in that: The deep learning model is a two-layer network formed by cascading multiple denoising modules, wherein the second layer of the network is composed of a denoising module; each denoising module of the first layer of the network respectively receives several consecutive frames in the continuous multi-frame input, and the denoising results of each denoising module of the first layer of the network are simultaneously input into the denoising module of the second layer of the network, and the output of the denoising module of the second layer of the network is used as the final output of the deep learning model.

9. The fast X-ray dynamic real-time imaging denoising method based on deep learning as claimed in claim 1, characterized in that: When training the deep learning model in step S3, the loss function adopts L1 loss, and is optimized using the Adam optimizer, with the learning rate set to 0.001 and the cycle set to 40.

Images