CN116205844A - Full-automatic heart magnetic resonance imaging segmentation method based on expansion residual error network - Google Patents

Abstract

The invention discloses a full-automatic heart magnetic resonance imaging segmentation method based on an expansion residual error network. The method comprises the following steps: acquiring a heart magnetic resonance image; and inputting the heart magnetic resonance image into a trained segmentation network to divide a right ventricle area, a myocardial area and a left ventricle area, wherein the segmentation network is constructed based on a residual network U-Net, and a bottleneck layer of the residual network adopts an expansion convolution block with a set expansion rate to combine an encoding path and a decoding path. The invention can accurately divide the areas such as the right ventricle, the left ventricle, the cardiac muscle and the like from the heart magnetic resonance image, realize the full-automatic division of the heart image and improve the performance of the heart area image division.

Description

Full-automatic heart magnetic resonance imaging segmentation method based on expansion residual error network

Technical Field

The invention relates to the technical field of biomedical engineering, in particular to a full-automatic cardiac magnetic resonance imaging segmentation method based on an expansion residual error network.

Background

The heart disease seriously threatens the life of people, and in order to effectively treat and prevent the disease, accurate calculation, modeling and analysis of the whole heart structure are important for the research and application in the medical field. However, during acquisition of CMRI (magnetic resonance cine imaging), the constant beating of the heart increases the difficulty in acquiring clear images, and especially for cardiovascular disease patients, there is a greater likelihood of experiencing arrhythmias, breath holding difficulties, and the like. This results in that the image of an MRI (magnetic resonance imaging) scanner may contain various image artifacts, making it difficult to evaluate the image quality. If the image data is not segmented correctly, the clinician may draw erroneous conclusions from the image data. The existing manual image segmentation is time-consuming and the precision is difficult to guarantee. Therefore, it is desirable to achieve automatic segmentation of cardiac regions for solving practical problems in the field of cardiac medical.

Cardiac image segmentation refers to the division of a cardiac image into a plurality of anatomically significant regions based on which quantitative measures, such as myocardial mass, wall thickness, left Ventricle (LV) and Right Ventricle (RV) volumes, can be extracted. Therefore, it is particularly important to design an accurate full-automatic heart segmentation algorithm. In recent years, deep Convolutional Neural Networks (DCNNs) have proven to segment the left and right ventricles and myocardium better than traditional computer vision methods. For example, the U-Net architecture is task independent and has been applied to a variety of biomedical segmentation tasks, with little or no substantial modification, U-Net being the most efficient underlying network model (backbone) for the ventricular segmentation algorithm.

In the prior art, patent application CN202210321078.8 provides a CT image heart segmentation method and system based on artificial intelligence semantic segmentation. The technology reduces the influence of noise through optimizing the class probability, and realizes accurate image segmentation. However, this approach does not involve specific left ventricular, right ventricular, and inter-myocardial segmentation problems in the heart, and does not allow independent cardiac tissue images to be obtained.

Patent application CN202110391121.3 describes a method and apparatus for training a heart segmentation model and a pathology classification model based on cardiac MRI, which can greatly suppress background interference and promote rapid convergence of neural network training, but no improvement method is proposed for image segmentation accuracy and robustness.

According to analysis, the prior art lacks research on a U-Net bottleneck layer, and the background area is far larger than a mask in an image, so that the problems of pixel degradation, time and space information loss caused by deepening the network layer number cause the insufficient extraction capability of the network to the image sparse features.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provides a full-automatic cardiac magnetic resonance imaging segmentation method based on an expansion residual error network. The method comprises the following steps:

acquiring a heart magnetic resonance image;

inputting the heart magnetic resonance image into a trained segmentation network to segment a right ventricle area, a myocardial area and a left ventricle area;

the split network is constructed based on a residual network U-Net, and a bottleneck layer of the residual network adopts an expansion convolution block with a set expansion rate to combine the coding path and the decoding path.

Compared with the prior art, the full-automatic cardiac MRI (magnetic resonance imaging) segmentation method based on the dilation residual error network has the advantages that the regions such as the right ventricle, the left ventricle and the cardiac muscle can be accurately segmented from the cardiac MRI images, the full-automatic segmentation of the cardiac images is realized, and the performance of the cardiac region image segmentation is improved.

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

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.

FIG. 1 is a flow chart of a fully automated cardiac magnetic resonance imaging segmentation method based on a dilated residual network in accordance with one embodiment of the present invention;

figure 2 is a schematic illustration of a process from raw magnetic resonance image data to image segmentation in accordance with one embodiment of the present invention;

FIG. 3 is a diagram of an automatic U-Net based image segmentation architecture in accordance with one embodiment of the present invention;

FIG. 4 is a schematic diagram of an architecture of a dilated residual block in accordance with one embodiment of the present invention;

FIG. 5 is a graph of image segmentation results for an ACDC test dataset according to one embodiment of the present invention;

in the figure, conv-convolution; norm-regularization; maxpool-max pooling; upConv-deconvolution; deconvolution-Deconvolution; skip-connection-hopping connection; pixel-wise addition-Pixel-by-Pixel addition; kernel; condition-Concatenation; stride-step length; end Systole-End Systole; end diaston-End Diastole.

Detailed Description

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 the components and steps, numerical expressions and numerical values set forth in these embodiments do not limit the scope of the present invention unless it is specifically stated otherwise.

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

Techniques, methods, and apparatus known to one of ordinary skill in the relevant art may not be discussed in detail, but are intended to be part of the specification where appropriate.

In all examples shown and discussed herein, any specific values should be construed as merely illustrative, and not a limitation. Thus, other examples of exemplary embodiments may have different values.

It should be noted that: like reference numerals and letters denote like items in the following figures, and thus once an item is defined in one figure, no further discussion thereof is necessary in subsequent figures.

The present invention develops a fully automatic segmentation method for segmenting the Right Ventricle (RV), myocardium (MYO) and Left Ventricle (LV) by combining short axis CMRI (magnetic resonance cine imaging) sequence images. The method captures multi-resolution features in the U-Net by expanding a convolutional residual network (DRN), thereby significantly increasing spatial and temporal information and maintaining positioning accuracy. And, the outputs of each extension path are added pixel by pixel to improve the training response.

As shown in fig. 1 and 2, the provided full-automatic cardiac magnetic resonance imaging segmentation method based on the dilation residual error network comprises the following steps:

in step S110, the data set is preprocessed to obtain training samples.

Taking magnetic resonance cine imaging as an example, the three-dimensional image has dimensions of l×w×h, where L is the image sequence length, W is the image width, and H is the image length. In the dataset, the tag values of the image are set to four types of tags by using a mapping mode, namely: black background=0, rv=1, myo=2, lv=3.

A significant difference in the display space size h×w of the magnetic resonance cine images and the range of the intensity distribution is considered. In one embodiment, training samples are obtained through a data preprocessing process, specifically exemplified by ACDC (Adverse Conditions Dataset with Correspondences) images. First, an input image is resampled. ACDC data sets have a voxel spacing problem in that the convolutional neural network cannot interpret the voxel spacing by resampling all images to the same voxel spacing of 1.52 x 6.35mm.

The data preprocessing process takes into account that the voxel spacing directly affects the overall voxel size of the image and also affects the amount of context information extracted from the image patch by the convolutional neural network. Furthermore, if the voxel spacing is greatly increased, the image size is reduced to the point that details are lost, so that a trade-off between the amount of context information contained in the network patch size and the amount of details retained in the image data needs to be ensured for optimal performance.

In one embodiment, for training data, all images are resampled to a median of 256×256 pixels. Magnetic resonance images of the ACDC dataset are then obtained using the multi-layer magnetic resonance cine images. For example, 2D-MRI (magnetic resonance imaging) slices and their associated annotations for each patient are extracted. And performs normalization on a slice-by-slice basis for each time frame.

Step S120, expanding training samples through data enhancement, and constructing a training set.

Due to the limited training data, the model cannot learn the desired invariance and robustness features, resulting in overfitting. Thus, a variety of data enhancement techniques may be applied to the training data to expand the number of samples. For example, basic image transformation techniques are employed, including random rotation, random elastic deformation, scaling, flipping, and gamma correction. Such data enhancement techniques, when applied to the original training image, can effectively generate multiple views of the same image. By expanding the training samples with a variety of data enhancement methods, the over-fitting and class imbalance problems can be solved.

Step S130, constructing a segmentation network based on the dilation residual network.

In this context, a heart segmentation network based on a U-net network is illustrated, see fig. 3 and 4, wherein fig. 4 corresponds to the dilated residual block architecture of fig. 3. Throughout the segmentation process, the segmentation network follows the overall architecture of the encoder-decoder, from the input image to the final output. For example, a shrink path is constructed using 5 coding blocks; each block consists of 2 convolutional layers with a 3 x 3 kernel and a 2 x 2 max pooling operation, step size of 2. Initially, 32 convolution kernels are selected. After each max pooling operation, the convolution kernels will increase, producing 320 convolution kernels in the bottleneck layer of the U-Net. Similarly, the spatial dimension of the feature map is reduced by a factor of 2 by the downsampling operation. The linear rectifying unit (ReLU) is replaced with leaky linear rectification and example regularization is used instead of normalization (BN).

The encoding and decoding paths are combined at the U-Net bottleneck layer by a extended residual network (DRN) that captures global context and recovers spatial and temporal information without affecting the resolution of the segmentation map. In addition, the extended residual network can effectively adjust the depth of the convolutional layer without degrading network performance. For example, the receptive field in the dilated residual network block is dilated by using dilation convolutions with different dilation rates (d=1, 3 and 5). The previously generated feature is then concatenated with the current feature through a residual connection. After every 3×3 convolution in a DRN (extended residual network) block, a pause (dropout) operation with a forgetting rate of 0.5 is performed to prevent overfitting. Thus, the extended residual network captures contextual image information, high spatial resolution, and multi-texture features. The process of decoding the path is similar to that of encoding the path, however the order of operations is reversed. The U-Net architecture provides the advantage of reusing the encoded signature from the encoded blocks to their corresponding levels, where the spatial dimensions match. This may be achieved by a channel-specific concatenation. A 1 x 1 kernel projection operation is used at the final stage of the decoding path to align the output channel dimensions with the classes (left ventricle, myocardium and right ventricle). Finally, all the extended path outputs are aggregated by upsampling and pixel-by-pixel addition to enhance the training response.

Typically, natural images contain many objects whose identity and relative position are important to understanding the scene. However, segmentation may become more difficult when the target object is spatially unobtrusive, e.g., when the target object is small compared to the background. If the features of the target object are lost during the downsampling process, they are not easily recovered during training. But if high (extensive) spatial and temporal information is maintained throughout the network and output features are provided that densely cover the input features, the back propagation can learn important features from smaller and less prominent objects. Therefore, the invention adopts an expanded convolution network, and predicts small and dense image features by increasing receptive fields and extracting more spatial information. The discrete dilation convolutions are as follows:

wherein,,

is an input and output discrete function, k is of size (2d+1) ² Discrete cores _l To expand convolution, in the summation process, s+lt=p, s denotes the expansion step, l denotes the scaling factor, p denotes the receptive field, and t denotes the integer sequence, i.e. t=1, 2,3.

An expanded residual network can better expand the receptive field to achieve a promising result and avoid image information loss at the bottleneck of the U-Net. The dilation convolution introduces a new parameter called the "dilation rate" into the convolution layer, which defines the spacing of the values as the convolution kernel processes the data, and enlarges the receptive field by adding holes. The dilated convolution layer is based on a conventional convolution with a dilation factor (d=1, 3 and 5). For example, a 1×1 kernel is selected for the normal convolution layer, and a 3×3 kernel is selected for the expanded convolution.

Wherein y is _ij Representing the input as x _ij Is a convolution kernel having a length M and a width N, M, N being the input variables of the expansion convolution. w (i, j) is the corresponding weight, i represents the image length index, j represents the image width index, and d represents the expansion ratio.

Step S140, training the segmentation network by using the set loss function.

The purpose of segmentation is to detect the target object and draw a contour around it. The automatic segmentation profile Cp (predicted) is compared with the corresponding annotation image to measure the accuracy of the proposed method. The pixels enclosed by the outline are referred to herein as A _p And A _g 。

In the segmentation network training, a variety of loss functions may be employed. Such as dice similarity coefficients or hausdorff distance or other loss function types.

For example, for Dice Similarity Coefficient (DSC), it is the ratio between the predicted profile and the ground truth profile that represents the DSC score, typically expressed as a percentage between 0 and 1. A high dice value indicates a good match.

Wherein A is _p Representing pixels surrounded by a predicted contour, A _g Representing the pixels enclosed by the true contour.

The Hausdorff Distance (HD) is a symmetric distance between the predicted and actual contours that is compared and provides the spatial resolution of the magnetic resonance cine images. The lower the HD value, the better the segmentation matching performance.

Wherein C is _p For predictive automatic segmentation of contours, C _g For the corresponding true mark contour, d (i, j) represents the distance between the ground true value and the predicted contour, i represents the pixel point of the predicted contour, and j represents the pixel point of the ground true contour. A significant class imbalance is considered in the image between the region of interest (ROI) and the background. To solve this problem, different loss functions were tested, including dice loss and weighted cross entropy loss.

In a preferred embodiment, the segmentation network is trained using a dual loss function comprising dice loss and cross entropy loss. Specifically, the cross entropy loss is defined as follows:

wherein C represents the total number of categories; c represents a category indication, w= (W ₁ ，w ₂ ，w ₃ ...w _n ) Is a series of weight capable of learning, w _n Is the weight matrix of the n layer; p (Y) _i |X _i W) represents a predicted pixel X _i Probability of error in classifying the label pixels relative to ground truth; y (c, x)Representing a target label corresponding to input x;

the activation function value for the prediction class c corresponding to the input x is represented. For example, for the category denoted by c, black background=0, rv=1, myo=2, lv=3.

The training of the model was performed for a total of 500 iterations, with 250 images randomly extracted from the dataset in each iteration of the training set until all image data was traversed. To enhance generalization, slices are randomly cropped from the training image and the network is evaluated after each iteration on the validation set. For example, the segmentation network is trained using a multi-class variant of dice loss of the following formula.

Wherein u and v are image segmentation tag values corresponding to the one-hot encoding vector and the class identifier output by the activation function softmax, i represents an image length index, and k represents an image width index; c e C is the class identifier, i.e. left ventricle, right ventricle, myocardium and background of the heart; epsilon is a small constant. After each traversal, the learning rate lr is recalculated according to the following equation. Finally, the best model was chosen to evaluate the test set to ensure that the verification of RV (right ventricle), MYO (myocardium) and LV (left ventricle) reached the highest DSC (dice similarity coefficient). The network provides consistent and stable performance across all folds.

Wherein initial_learning_rate is the initial learning rate, currentepoch is the current iteration number, and total iteration number is the total iteration number.

Step S150, for the acquired target magnetic resonance image, the areas such as the right ventricle, the cardiac muscle and the left ventricle are identified by using the trained segmentation network.

After the training of the segmentation network is completed, the optimized model parameters can be obtained, and further, the trained segmentation network can be utilized to accurately distinguish the regions of interest such as the right ventricle, the cardiac muscle and the left ventricle and the complete heart contour, and further, quantitative measures such as the quality of the cardiac muscle, the volumes of the left ventricle and the right ventricle and the like can be extracted based on the regions.

To further demonstrate the effects of the present invention, experimental tests have been performed on a plurality of patient cardiac magnetic resonance images. See the schematic of the different slice positions of the heart shown in fig. 5. Experimental results show that the invention obtains higher segmentation precision and segmentation speed of the left ventricle, the right ventricle and the cardiac muscle, and obtains the overall dice similarity coefficient of 0.92+/-0.02 and the average Hausdorff distance of 8.06+/-0.05 mm. Also, the invention increases the speed of image segmentation, e.g. processing 2D magnetic resonance images takes on average 0.28 seconds. Furthermore, the network of the present invention is designed to predict individual magnetic resonance images to segment ventricular areas, successfully enabling automatic segmentation of cardiac images.

In summary, compared with the prior art, the invention has the following technical effects:

1) The invention introduces an expansion convolution residual error network, enhances the performance of the U-Net bottleneck layer, realizes full-automatic accurate segmentation of a heart MRI (magnetic resonance imaging) image, solves the limitation of the U-Net bottleneck layer, obviously enhances the space and time information, and improves the precision while maintaining the space consistency.

2) The invention designs an extended residual network (DRN) block to replace the original bottleneck layer of U-Net. And a plurality of loss functions are used, so that the heart characteristic training model is better utilized in the process of segmenting the heart image, and the accuracy is improved.

3) The invention has higher calculation speed and robustness, and can be applied to diversified cardiac CMRI (magnetic resonance imaging) data sets. For example, the processed data are magnetic resonance images of the patient under two different magnetic intensities, and the processed data can simultaneously obtain images of the complete heart contour, the left ventricle, the right ventricle and the myocardial contour of the patient.

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 embodied thereon for causing a processor to implement aspects of the present invention.

The computer readable storage medium may be a tangible device that can hold and store instructions for use by an instruction execution device. The computer readable storage medium may be, for example, but not limited to, an electronic 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 (a non-exhaustive list) of the computer-readable storage medium would include the following: 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 Disks (DVD), memory sticks, floppy disks, mechanical coding devices, punch cards or in-groove structures such as punch cards or grooves having instructions stored thereon, and any suitable combination of the foregoing. Computer-readable storage media, as used herein, are not to be construed as transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., optical pulses through fiber optic cables), or electrical signals transmitted through wires.

The computer readable program instructions described herein may be downloaded from a computer readable storage medium to a respective computing/processing device 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 transmissions, wireless transmissions, routers, firewalls, switches, gateway computers and/or edge servers. The network interface card or network interface in each computing/processing device receives computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium in the respective computing/processing device.

Computer program instructions for carrying out operations of the present invention may be assembly instructions, instruction Set Architecture (ISA) instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, c++, python, and the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer readable program instructions may be executed 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. 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 (for example, through the Internet using an Internet service provider). In some embodiments, aspects of the present invention are implemented by personalizing electronic circuitry, such as programmable logic circuitry, field Programmable Gate Arrays (FPGAs), or Programmable Logic Arrays (PLAs), with state information for computer readable program instructions, which can execute the computer readable program instructions.

Various 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, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks. These computer readable program instructions may also be stored in a computer readable storage medium that can direct a computer, programmable data processing apparatus, and/or other devices to function in a particular manner, such that the computer readable medium having the instructions stored therein includes an article of manufacture including instructions which implement the function/act specified in the flowchart and/or block diagram block or blocks.

The computer readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other devices to produce a computer implemented process such that the instructions which execute on the computer, other programmable apparatus or other devices implement the functions/acts specified in the flowchart and/or block diagram block or blocks.

The flowcharts 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 executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown 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 will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions. It is well known to those skilled in the art that implementation by hardware, implementation by software, and implementation by a combination of software and hardware are all equivalent.

The foregoing description of embodiments of the invention has been presented for purposes of illustration and description, and is not intended to be exhaustive or limited to the embodiments disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the various embodiments described. The terminology used herein was chosen in order to best explain the principles of the embodiments, the practical application, or the technical improvements 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 (10)

1. A full-automatic heart magnetic resonance imaging segmentation method based on an expansion residual error network comprises the following steps:

acquiring a heart magnetic resonance image;

inputting the heart magnetic resonance image into a trained segmentation network to segment a right ventricle area, a myocardial area and a left ventricle area;

the split network is constructed based on a residual network U-Net, and a bottleneck layer of the residual network adopts an expansion convolution block with a set expansion rate to combine the coding path and the decoding path.

2. The method according to claim 1, wherein the segmentation network is trained according to the steps of:

constructing a training set, wherein the training set comprises a plurality of pieces of sample data, each piece of sample data is a magnetic resonance image with a labeling category, and the labeling category is used for distinguishing a right ventricle area, a myocardial area and a left ventricle area;

image enhancing the training set to generate multiple views for the same magnetic resonance image, the image enhancing comprising one or more of random rotation, random elastic deformation, scaling, flipping, and gamma correction;

and training the segmentation network by using the training set after image enhancement and using a set loss function to obtain optimized parameters.

3. The method of claim 1, wherein training the segmentation network with a set loss function comprises:

training a segmentation network by using a cross entropy loss function within a range set to iteration times, and extracting a set number of sample images from the training set each time of iteration;

randomly cropping slices from the training image and evaluating the segmentation network after each iteration on the validation set and training the segmentation network using a multi-class variant of dice loss;

after each traversal, recalculating the learning rate;

a segmentation network meeting set performance requirements is selected as the trained segmentation network.

4. A method according to claim 3, characterized in that the cross entropy loss is expressed as:

wherein C represents the total number of labeling categories; c represents a label category indication, w= (W ₁ ，w ₂ ，w ₃ ...w _n ) Is a series of weights to be learned, w _n Is the weight matrix of the n-th layer, p (Y _i |X _i W) represents a predicted pixel X _i Relative ground truth value label pixel Y _i Probability of classification error, Y (c, x) represents the target label corresponding to input x;

the activation function value for the prediction class c corresponding to the input x is represented.

5. A method according to claim 3, wherein the dice loss is expressed as:

wherein u and v are image segmentation label values corresponding to a single thermal coding vector and a labeling category indication C epsilon C output by an activation function softmax, i represents an image length index, k represents an image width index, epsilon is a set constant, and C represents the total number of labeling categories.

6. A method according to claim 3, wherein the learning rate is updated according to the following formula:

wherein initial_learning_rate is the initial learning rate, currentepoch is the current iteration number, and total iteration number is the total iteration number.

7. The method of claim 1, wherein a last stage of a decoding path of the split network uses a 1 x 1 kernel projection operation.

8. The method of claim 1, wherein the expansion ratio is set to d = 1, 3 or 5.

9. A computer readable storage medium having stored thereon a computer program, wherein the computer program when executed by a processor realizes the steps of the method according to any of claims 1 to 8.

10. A computer device comprising a memory and a processor, on which memory a computer program is stored which can be run on the processor, characterized in that the processor implements the steps of the method according to any of claims 1 to 8 when the computer program is executed.

Images