WO2022204868A1

WO2022204868A1 - Method for correcting image artifacts on basis of multi-constraint convolutional neural network

Info

Publication number: WO2022204868A1
Application number: PCT/CN2021/083561
Authority: WO
Inventors: 郑海荣; 李彦明; 万丽雯; 胡战利; 邓富权
Original assignee: 深圳高性能医疗器械国家研究院有限公司
Priority date: 2021-03-29
Filing date: 2021-03-29
Publication date: 2022-10-06

Abstract

Disclosed is a method for correcting image artifacts on the basis of a multi-constraint convolutional neural network. The method comprises: constructing a convolutional neural network having a trunk structure layer and a plurality of branch structure layers, wherein the trunk structure layer uses an original image as input, the plurality of branch structure layers respectively use images having different levels of resolutions obtained by down-sampling the original image as input, and the feature image matrix of the last convolutional layer of each branch structure layer is restored to a matrix image of the same size as a target image by means of up-sampling; and training the convolutional neural network with the goal of convergence of a set total loss function, learning a mapping relationship from a low-dose original image to a standard-dose target image, and setting a loss value for each of the trunk structure layer and the plurality of branch structure layers in the training process. The present invention enhances image detail information while improving the peak signal to noise ratio and the structural similarity of images.

Description

A Method for Correcting Image Artifacts Based on Convolutional Neural Networks with Multiple Constraints

technical field

The invention relates to the technical field of medical image processing, and more particularly, to a method for correcting image artifacts based on a convolutional neural network with multiple constraints.

Background technique

Research and development to remove motion artifacts in images that produce artifacts has important scientific significance for the current field of medical diagnosis. Taking Coronary Computed Tomography Angiography (CCTA) as an example, it refers to a non-invasive imaging method that uses computers and X-rays to obtain a patient's cardiac tomographic images after intravenous injection of an appropriate contrast agent. The detection method, which has the advantages of short scanning time, extensive component information and non-invasive visualization of the vessel wall, is suitable for the diagnosis of suspected coronary heart disease, the follow-up of coronary artery bypass surgery, the assessment of valvular heart disease and the assessment of heart quality. However, CCTA-acquired images may exhibit motion artifacts and require re-examination. A large amount of X-ray exposure will cause the cumulative effect of radiation dose, which will greatly increase the possibility of various diseases, thereby affecting the physiological functions of the human body, destroying human tissues and organs, and even endangering the life safety of patients. Therefore, the artifact correction of medical images has broad application prospects in the field of medical diagnosis.

The formation of motion artifacts during coronary CT imaging is due to the displacement of image pixels when the CT acquires projection data from different angles. The degree of motion artifacts depends on the unique rate and correction results of the image reconstruction algorithm. Generally, the elimination of motion artifacts starts from two aspects: the first is to control the heart rate, reduce the heart rate of the subject, prolong the cardiac cycle, slow down the movement of the coronary arteries and prolong the time of the low-speed movement of the coronary arteries, which can reduce the time resolution of imaging. The second is to improve the temporal resolution. Based on the basic principles of MDCT coronary imaging and the causes of motion artifacts, to achieve CCTA imaging and avoid motion artifacts, it is necessary to compare the highest temporal resolution obtained under a specific scanning method with the coronary artifact. Phases with minimal arterial motion are matched.

The improvement of time resolution starts from two aspects, one is to improve the time resolution from the hardware method, and the other is to improve the time resolution from the software aspect. Specifically, in terms of hardware, the time resolution is improved by increasing the rotation speed of the tube ball, using a wide-body detector and adopting a dual-detector technology. However, increasing the rotational speed of the tube is limited by physical properties, using multi-detector technology is limited by space, and using wide-body detector technology is limited by economic costs. In terms of software, the use of multi-sector reconstruction technology, image reconstruction technology based on compressed sensing (Prior Image Constrained Compressed Sensing, PICCS), motion estimation and compensation algorithms and motion correction technology (Snap Shot Freeze, SSF) can effectively improve the temporal resolution . However, the use of multi-sector reconstruction techniques requires maintaining a stable patient's heart rate and is limited by the ball rotation time and scan pitch. Image reconstruction techniques based on compressed sensing have not yet been validated. Motion estimation and compensation algorithms rely on a lot of computation and evaluation. The motion correction technology needs to collect the phase with relatively small relative motion artifacts and good image quality in the image, and eliminate the motion artifacts through complex calculations.

In addition, for the existing technical solutions of using deep learning to remove image artifacts, the two methods proposed by Philips Research Center to remove CCTA motion artifacts are to use deep learning models to identify and estimate the level of motion artifacts, and assist traditional methods to perform motion. Compensation and Estimation. However, the effect of this scheme in removing motion artifacts needs to be improved.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the above-mentioned defects of the prior art, and to provide a method for correcting image artifacts based on a convolutional neural network with multiple constraints.

According to a first aspect of the present invention, a method for correcting image artifacts based on a convolutional neural network with multiple constraints is provided. The method includes the following steps:

Constructing a convolutional neural network with a backbone structure layer and a plurality of branch structure layers, wherein the backbone structure layer takes the original image as input, and the multiple branch structure layers are respectively distinguished by different levels obtained by downsampling the original image. The image of the rate is used as input, and the feature image matrix of the last convolutional layer of each branch structure layer is restored to a matrix image of the same size as the target image through upsampling;

The convolutional neural network is trained with the goal of convergence of the set overall loss function, and the mapping relationship between the low-dose original image and the standard-dose target image is learned, and during the training process, the backbone structure layer and the multiple The loss value is set for each layer of the branch structure layer.

According to a second aspect of the present invention, a medical image processing method is provided. The method includes: performing down-sampling on the to-be-processed image multiple times to obtain images of different levels of resolution; inputting the images of different levels of resolution and the to-be-processed image into a trained convolutional neural network obtained by using the above method of the present invention , to get the output image.

Compared with the prior art, the advantage of the present invention is that the motion artifact of the image is eliminated by using the convolutional neural network, in terms of hardware, it is not limited by physical characteristics, space and economic costs; in terms of software, it does not require complex calculation. And it is not affected by the change of the patient's heart rate, the image can be obtained to eliminate the artifacts; in terms of deep learning, it can be effectively used as a tool to directly remove the motion artifacts of medical images, and no traditional methods are needed for denoising. The invention realizes image noise reduction based on multi-constrained multi-level convolutional neural network, improves image peak signal-to-noise ratio and structural similarity, and enhances image detail information, thereby obtaining medical images that better meet diagnostic requirements.

Other features and advantages of the present invention will become apparent from the following detailed description of exemplary embodiments of the present invention with reference to the accompanying drawings.

Description of drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention.

1 is a flowchart of a method for correcting image artifacts based on a convolutional neural network with multiple constraints according to an embodiment of the present invention;

FIG. 2 is a schematic structural diagram of a densely connected residual block according to an embodiment of the present invention;

3 is a schematic structural diagram of a stacked dense residual block according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of an attention mechanism according to an embodiment of the present invention;

5 is a schematic diagram of a residual block according to the present invention;

6 is a schematic diagram of a multi-constrained multi-level convolutional neural network according to an embodiment of the present invention;

7 is a schematic diagram of an experimental result according to an embodiment of the present invention;

In the attached figure, conv-convolutional layer; res-residual block.

Detailed ways

Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that the relative arrangement of components and steps, numerical expressions and numerical values set forth in these embodiments do not limit the scope of the invention unless specifically stated otherwise.

The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the invention, its application, or uses.

Techniques, methods, and apparatus known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, such techniques, methods, and apparatus should be considered part of the specification.

In all examples shown and discussed herein, any specific values should be construed as illustrative only and not limiting. Accordingly, other instances of the exemplary embodiment may have different values.

It should be noted that like numerals and letters refer to like items in the following figures, so once an item is defined in one figure, it does not require further discussion in subsequent figures.

The present invention can be applied to motion artifact correction, noise reduction, etc. of CCTA images or other types of medical images, or can be applied to the field of image super-resolution after proper modification. For the sake of clarity, the artifact correction of CT images is taken as an example for description below.

In order to solve the problems of poor CT imaging quality and many noise artifacts under low-dose conditions, the present invention designs a multi-level convolutional neural network network based on multiple constraints to improve the low-dose CT imaging quality. In addition, the reuse of low-dimensional and high-dimensional information and the fusion of local and non-local information can be greatly improved through the attention mechanism, thereby enhancing the performance of traditional convolution operations, which can largely eliminate the noise and artifacts of low-dose CT. film. On the other hand, a joint loss function is specially designed to improve the quality of CT images, and the generated CT images are further guaranteed to meet the requirements of medical diagnosis by combining multiple loss functions.

Specifically, as shown in FIG. 1 , the provided method for correcting image artifacts based on a convolutional neural network with multiple constraints includes the following steps.

Step S110, design a stacked dense residual block structure.

The traditional convolutional neural network stacks multiple convolutional layers, and realizes the feature extraction of the image through these convolutional layers, thereby enhancing the features and details of the image. In one embodiment of the present invention, the DCR (Densely connected and residual block) module proposed in "Densely Connected Hierarchical Network for Image Denoising" published by Bumjun Park on CVPR in 2019 is used to extract CCTA image features. The DCR structure shown in Figure 2 has the advantage that this module has a better convergence speed than the same number of stacked convolutional layers.

Further, the present invention regards DCR as an ordinary convolution module, and performs densely stacked, and builds a stacked dense residual block structure (DDCR), thereby obtaining a large number of features from low-dimensional to high-dimensional, and compared with the same number of stacked For the convolutional layer of , its convergence is better, and the DDCR structure is shown in Figure 3. By using DCR module and DDCR module instead of ordinary convolutional layer stacking, the extracted features are more comprehensive and can provide computational efficiency.

Step S120, designing an attention mechanism.

In the method of coronary motion artifact correction in CCTA images, it is more difficult to correct local features into images without artifacts. In order to enable the network to perform artifact correction on its local features, especially coronary artery features, in one embodiment, an attention mechanism is designed to assist the main branch of the network to learn the local features of CCTA, as shown in Figure 4.

The use of attention mechanism can improve the utilization of image feature information. When it is applied to the field of CCTA coronary motion artifact removal, the attention mechanism can fuse the local and global features of the image, thereby improving the artifact removal effect. Furthermore, the attention mechanism can be thought of as a plug-and-play module that can be added to any traditional convolutional neural network workflow to improve the performance of the network.

Step S130, designing the residual block of the convolutional neural network.

In convolutional neural networks, residual blocks are a common method to extract low-dimensional to high-dimensional information from images. Compared with general convolutional layers, it is not easy to lose low-dimensional features. In image de-artifacting, this method can maintain the overall characteristics of the image, and at the same time can learn the clear features of coronary arteries, so as to de-artifact the CCTA image. For example, the residual block structure is shown in Figure 5.

Step S140, designing a multi-constrained multi-level convolutional neural network.

In traditional convolutional neural networks, the usual calculation method of the loss function is to calculate the loss value between the input image and the output image. However, in an embodiment of the present invention, the calculation method adopted is to set a value of calculation loss for each processing layer, which is used to constrain the optimization and convergence of each processing layer, so as to prevent the processing layer from being ignored by the backbone network. When calculating the loss function, the feature image matrix of the last convolutional layer of each layer can be restored to a matrix image of the same size as the target image through the pixel shuffle function in the PyTorch library.

As shown in Figure 6, the constructed convolutional neural network includes a backbone structure layer and a plurality of branch structure layers, wherein the backbone structure layer takes the original image as input, and the multiple branch structure layers are respectively obtained by down-sampling the original image (such as using The images of different levels of resolution obtained by pixel un-shuffle function) are used as input, and the feature image matrix of the last convolutional layer of each branch structure layer is restored to the same size as the target image through upsampling (such as pixel shuffle). Matrix image, and the image features extracted from each branch structure layer are fused to the backbone structure layer from top to bottom (ie, from low-resolution images to higher-resolution images), and are processed by the attention mechanism in the backbone structure layer. Thereby learning the mapping relationship between the input image and the output target image.

For example, there are three loss functions: mean square error (MSE), perceptual loss, and kernel loss. Among them, the mean square error is to improve the details of the matrix pixels; the perceptual loss is to improve the similarity of the overall features of the image; and the kernel loss is to improve the convergence speed of the network.

Still combined with Figure 6, the input and output backbone network is used as the first layer, and the image is downsampled once by pixel un-shuffle as the second layer, and so on. The higher the number of layers, the more the learned features tend to be holistic. Also, in order to improve the convergence of each processing layer. Each additional processing layer will decrease the loss weight of each layer. Therefore, MSE is calculated as follows:

Loss _MSE = loss _{mse_1} +w _{mse_2} ×loss _{mse_2} +w _{mse_3} ×loss _{mse_3} +w _{mse_4} ×loss _{mse_4} where w _{mse_2} , w _{mse_3} and w _{mse_4} are the weights of MSE loss for each processing layer, and loss _{mse_1} , loss _{mse_2} , loss _{mse_3} and loss _{mse_4} are the MSE losses for each processing layer.

Likewise, the perceptual loss is calculated as follows:

Loss _per =loss _{per_1} +w _{per_2} ×loss _{pre_2} +w _{per_3} ×loss _{per_3} +w _{per_4} ×loss _{per_4}

where w _{per_2} , w _{per_3} and w _{per_4} are the weights of the perceptual loss of each processing layer, and loss _{per_1} , loss _{per_2} , loss _{per_3} and loss _{per_4} are the perceptual losses of each processing layer.

Similarly, the kernel loss is calculated as follows:

Loss _ker = loss _{ker_1} +w _{ker_2} ×loss _{ker_2} +w _{ker_3} ×loss _{ker_3} +w _{ker_4} ×loss _{ker_4}

Among them, w _{ker_2} , w _{ker_3} and w _{ker_4} are the weights of the kernel loss of each processing layer, and loss _{ker_1} , loss _{ker_2} , loss _{ker_3} and loss _{ker_4} are the kernel losses of each processing layer.

The overall loss function of the convolutional neural network is expressed as:

Loss=w _mse ×Loss _MSE +w _per ×Loss _per +w _ker ×Loss _ker

Among them, w _mse , w _per and w _ker are the weights of MSE, perceptual loss and kernel loss, respectively, and the weights can be adjusted according to needs or simulation results.

It should be noted that in practical applications, more branch structure layers can be set according to the requirements for image reconstruction effect or processing speed. In this case, for example, the MSE loss can be generally expressed as:

Loss _MSE = loss _{mse_1} +w _{mse_2} ×loss _{mse_2} +w _{mse_3} ×loss _{mse_3} +w _{mse_4} ×loss _{mse_4} +,…,+w _{mse_k} ×loss _{mse_k} , where k represents the branch structure layer index. Correspondingly, other loss items are set, which will not be repeated here.

By setting up multiple processing layers such as backbone structure layer and branch structure layer, image features from different perspectives can be fused, and the quality of CCTA images can be improved by using a multi-level joint loss function with multiple constraints.

Step S150, train the convolutional neural network with the set loss function convergence as the goal.

For example, the entire network is optimized using the Adam optimizer. The training optimization process takes the artifact data in the CCTA image data as input, and takes the coronary artery motion artifact-free data in the CCTA image data of the same patient as the reference data, until the network reaches a convergence state, such as the loss function value is less than the set value. set threshold.

Step S160, using the trained convolutional neural network to perform image processing.

Using the trained convolutional neural network, the collected actual images can be de-artifacted and denoised in real time to obtain reconstructed images, that is, the images to be processed are down-sampled multiple times to obtain images with different levels of resolution; The images of different levels of resolution and the images to be processed are input into the above-mentioned trained convolutional neural network to obtain an output image.

In order to further verify the effect of the present invention, experiments are carried out, and the results are shown in FIG. 7 , the rows from top to bottom show the input image, the output image and the target image respectively. It can be seen that the method of the present invention can effectively correct the motion artifacts of the coronary arteries on the CCTA image. At the same time, image detail information can be recovered to a certain extent.

To sum up, compared with the prior art, the beneficial effects of the present invention are as follows: 1) Based on the multi-constraint and multi-level joint loss function enhancement network, the convergence and optimization of the multi-layer processing layer of the network are enhanced, and the local coronary artery is enhanced. details, and improve the effect of artifact correction; 2) Use DCR and DDCR modules to replace ordinary convolutional layer stacking, reduce network parameters, and improve network convergence; 3) Directly move CCTA images based on convolutional neural networks Artifact correction, input the image of coronary artery with motion artifact, and directly obtain the image after coronary motion artifact correction.

The present invention may be a system, method and/or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the present invention.

A computer-readable storage medium may be a tangible device that can hold and store instructions for use by the instruction execution device. The computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (non-exhaustive list) of computer readable storage media include: portable computer disks, hard disks, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM) or flash memory), static random access memory (SRAM), portable compact disk read only memory (CD-ROM), digital versatile disk (DVD), memory sticks, floppy disks, mechanically coded devices, such as printers with instructions stored thereon Hole cards or raised structures in grooves, and any suitable combination of the above. Computer-readable storage media, as used herein, are not to be construed as transient signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (eg, light pulses through fiber optic cables), or through electrical wires transmitted electrical signals.

The computer readable program instructions described herein may be downloaded to various computing/processing devices from a computer readable storage medium, or to an external computer or external storage device over a network such as the Internet, a local area network, a wide area network, and/or a wireless network. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and/or edge servers. A network adapter card or network interface in each computing/processing device receives computer-readable program instructions from a network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in each computing/processing device .

The computer program instructions for carrying out the operations of the present invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state setting data, or instructions in one or more programming languages. Source or object code written in any combination, including object-oriented programming languages, such as Smalltalk, C++, Python, etc., and conventional procedural programming languages, such as the "C" language or similar programming languages. The computer readable program instructions may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer, or entirely on the remote computer or server implement. In the case of a remote computer, the remote computer may be connected to the user's computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (eg, using an Internet service provider through the Internet connect). In some embodiments, custom electronic circuits, such as programmable logic circuits, field programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), can be personalized by utilizing state information of computer readable program instructions. Computer readable program instructions are executed to implement various aspects of the present invention.

Aspects of the present invention are described herein with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer readable program instructions.

These computer readable program instructions may be provided to a processor of a general purpose computer, special purpose computer or other programmable data processing apparatus to produce a machine that causes the instructions when executed by the processor of the computer or other programmable data processing apparatus , resulting in means for implementing the functions/acts specified in one or more blocks of the flowchart and/or block diagrams. These computer readable program instructions can also be stored in a computer readable storage medium, these instructions cause a computer, programmable data processing apparatus and/or other equipment to operate in a specific manner, so that the computer readable medium on which the instructions are stored includes An article of manufacture comprising instructions for implementing various aspects of the functions/acts specified in one or more blocks of the flowchart and/or block diagrams.

Computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other equipment to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other equipment to produce a computer-implemented process , thereby causing instructions executing on a computer, other programmable data processing apparatus, or other device to implement the functions/acts specified in one or more blocks of the flowcharts and/or block diagrams.

The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more functions for implementing the specified logical function(s) executable instructions. In some alternative implementations, the functions noted in the blocks may occur out of the order noted in the figures. For example, two blocks in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It is also noted that each block of the block diagrams and/or flowchart illustrations, and combinations of blocks in the block diagrams and/or flowchart illustrations, can be implemented in dedicated hardware-based systems that perform the specified functions or actions , or can be implemented in a combination of dedicated hardware and computer instructions. It is well known to those skilled in the art that implementation in hardware, implementation in software, and implementation in a combination of software and hardware are all equivalent.

Various embodiments of the present invention have been described above, and the foregoing descriptions are exemplary, not exhaustive, and not limiting of the disclosed embodiments. Numerous modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The terminology used herein was chosen to best explain the principles of the embodiments, the practical application or technical improvement in the marketplace, or to enable others of ordinary skill in the art to understand the embodiments disclosed herein. The scope of the invention is defined by the appended claims.

Claims

A method for correcting image artifacts based on a convolutional neural network with multiple constraints, comprising the following steps:

Constructing a convolutional neural network with a backbone structure layer and a plurality of branch structure layers, wherein the backbone structure layer takes the original image as input, and the multiple branch structure layers are respectively distinguished by different levels obtained by downsampling the original image. The image of the rate is used as input, and the feature image matrix of the last convolutional layer of each branch structure layer is restored to a matrix image of the same size as the target image through upsampling;

The convolutional neural network is trained with the goal of convergence of the set overall loss function, and the mapping relationship between the low-dose original image and the standard-dose target image is learned, and during the training process, the backbone structure layer and the multiple The loss value is set for each layer of the branch structure layer.
The method according to claim 1, wherein the overall loss function is expressed as:

Loss=w mse ×Loss MSE +w per ×Loss per +w ker ×Loss ker

Among them, Loss MSE is the mean square error loss, Loss per is the perceptual loss, Loss ker is the kernel loss, and w mse , w per and w ker are the weights of the corresponding items, respectively.
The method according to claim 2, wherein the mean square error loss is expressed as:

Loss MSE = loss mse_1 +w mse_2 ×loss mse_2 +w mse_3 ×loss mse_3 +w mse_4 ×loss mse_4 +,…,+w mse_k ×loss mse_k ;

Among them, w mse_2 , w mse_3 , w mse_4 and w mse_k are the weights of the mean square error loss of the corresponding branch structure layer, loss mse_2 , loss mse_3 , loss mse_4 and loss mse_k are the mean square error loss of the corresponding branch structure layer, and loss mse_1 is the backbone The loss of the structure layer, k is the index of the branch structure layer.
The method of claim 2, wherein the perceptual loss is expressed as:

Loss per =loss per_1 +w per_2 ×loss pre_2 +w per_3 ×loss per_3 +w per_4 ×loss per_4 +,…,+w per_k ×loss per_k ;

Among them, w per_2 , w per_3 , w per_4 and w per_k are the weights of the perceptual loss of the corresponding branch structure layer, loss per_2 , loss per_3 , loss per_4 and loss per_k are the perceptual losses of the corresponding branch structure layer, loss per_1 is the backbone structure layer The perceptual loss of , k is the index of the branch structure layer.
The method according to claim 2, wherein the core loss is expressed as:

Loss ker = loss ker_1 +w ker_2 ×loss ker_2 +w ker_3 ×loss ker_3 +w ker_4 ×loss ker_4 +,…,+w ker_k ×loss ker_k ;

Among them, w ker_2 , w ker_3 , w ker_4 and w ker_k are the weights of the kernel loss corresponding to the branch structure layer, loss ker_2 , loss ker_3 , loss ker_4 and loss ker_k are the kernel losses of the corresponding branch structure layer, and loss ker_1 is the backbone structure layer The kernel loss of , k is the index of the branch structure layer.
The method according to claim 1, wherein the backbone structure layer comprises a multi-layer convolution layer and a stacked densely connected residual block structure, and adopts an attention mechanism to fuse features extracted from each branch structure layer to obtain an output image.
The method according to claim 6, wherein, in the training process, images of different levels of resolution are respectively input into corresponding branch structure layers, each branch structure layer comprises a multi-layer convolution layer and a stacked densely connected residual block structure , and the output features of the branch structure layer corresponding to the lower resolution image are sequentially cascaded with the output of the first convolution layer in the branch structure layer corresponding to the higher resolution image, and finally fused to the backbone network layer.
A medical image processing method, comprising:

Downsampling the to-be-processed image multiple times to obtain images with different levels of resolution;

The images with different levels of resolution and the images to be processed are input into the trained convolutional neural network obtained by the method according to any one of claims 1 to 7 to obtain an output image.
A computer-readable storage medium having stored thereon a computer program, wherein the program, when executed by a processor, implements the steps of the method according to any one of claims 1 to 8.
A computer device, comprising a memory and a processor, a computer program that can be run on the processor is stored in the memory, and characterized in that, when the processor executes the program, any one of claims 1 to 8 is implemented The steps of the method described in item.