A 3D medical image segmentation method based on context information fusion

Abstract

The present invention discloses a three-dimensional medical image segmentation method based on context information fusion. In the present invention, the three-dimensional medical image network mainly adopts a sub-block-based method for both the training stage and the prediction stage. Through the dilated convolution branch and the self-attention branch in parallel of the three-dimensional medical image network, features containing context information are extracted, and after fusion, a predicted segmentation result is generated. Then, the obtained predicted segmentation result is used to fuse the segmentation results of adjacent predicted sub-blocks by Gaussian weighting, which alleviates the problem of inaccurate prediction of sub-block boundaries and further improves the segmentation accuracy. In addition, during the training of the three-dimensional medical image network, a sample balancing sampling strategy is used to sample training sub-blocks, avoiding the possible class imbalance problem in the data and improving the segmentation accuracy of the foreground class. The method of the present invention can effectively improve the accuracy of the three-dimensional image segmentation algorithm, is applicable to the three-dimensional medical image segmentation tasks of various organs, and has good universality and generality.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image segmentation, and particularly relates to a three-dimensional medical image segmentation method based on context information fusion. Background Art

[0002] Three-dimensional medical image data is usually represented by voxels, such as the common CT images now. A voxel is a normalized representation method, conceptually similar to the smallest unit in two-dimensional space - a pixel, and can be regarded as the smallest unit in the three-dimensional space partition of data. Semantic segmentation has a wide range of applications in biomedical image analysis, such as X-rays, MRI scans, digital pathology, and endoscopes. Therefore, semantic segmentation is a very important task in medical image analysis. By quickly and automatically segmenting organs from three-dimensional medical images and then determining the spatial geometry and volume of the organs, it can help doctors formulate accurate medical plans. Therefore, organ segmentation based on three-dimensional medical images has important research significance and clinical value.

[0003] According to the different dimensions of the network input data, there are mainly two types of methods for three-dimensional medical image segmentation. The first type of method is to divide the data into 2D slices and use a single slice or multiple adjacent slices as the input, so that a two-dimensional segmentation network can be used for segmentation, and finally the segmentation results are summarized into a three-dimensional form according to the slices. The second type of method is to directly use the three-dimensional data as the input and process it with a 3D CNN. 3D CNN refers to replacing operations such as convolution and pooling in the common 2D CNN with corresponding three-dimensional operations. The computational cost and memory cost of 3D CNN are very high, so it is usually necessary to cut the original three-dimensional data in the form of a sliding window, perform segmentation separately, and finally summarize the style results. Compared with the slice-based method, using 3D CNN can more effectively utilize the information in all directions of the data and theoretically has better expressive ability.

[0004] When performing image segmentation, the prediction of a pixel category not only needs to consider the information of the pixel itself, but also needs to consider the information of its surrounding pixels and the overall information of the picture. The information that combines the information of other pixels around the pixel and the overall characteristics of the picture is called local context information and global context information respectively, and is collectively called context information. Existing research shows that making full use of context information can effectively improve the accuracy of image segmentation algorithms.

[0005] However, many existing 3D CNN-based segmentation methods use networks with an encoder-decoder structure. For medical image segmentation tasks, the network with a three-dimensional encoder-decoder structure has two limitations. On the one hand, downsampling and upsampling are divided into multiple stages, and each stage has a considerable number of parameters, which results in a large number of parameters in the overall network that need to be trained, and there is a risk of overfitting. On the other hand, the spatial size of the input three-dimensional image sub-blocks is relatively small, and further downsampling by the encoder will cause the loss of fine-grained information at the boundaries of the input images, thus affecting the performance of dense prediction tasks such as semantic segmentation. Summary of the Invention

[0006] This part of the content of the present application is used to briefly introduce concepts, which will be described in detail in the following detailed implementation part. This part of the content of the present application is not intended to identify the key features or essential features of the claimed technical solution, nor is it intended to limit the scope of the claimed technical solution.

[0007] Aiming at the problems and deficiencies in the prior art, the purpose of the present invention is to provide a three-dimensional medical image segmentation method based on context information fusion. The entire three-dimensional image segmentation network is trained and predicted in a sub-block-based manner, combining the information of other pixels around the pixel and the information of the overall characteristics of the picture, extracting features containing context information through parallel dilated convolution branches and self-attention branches, and then fusing them to generate a segmentation result. During training, a sample balancing sampling strategy is used to sample training sub-blocks, avoiding the possible class imbalance problem in the data and improving the segmentation accuracy of the foreground class; during prediction, the three-dimensional image data is cropped by a sliding window to obtain prediction sub-blocks as the network input, and the segmentation results of adjacent prediction sub-blocks are fused using Gaussian weighting, thereby alleviating the problem of inaccurate prediction at the sub-block boundaries and further improving the segmentation accuracy, with good universality and generality, to solve the problems proposed in the above background technology.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] The present invention discloses a three-dimensional medical image segmentation method based on context information fusion. The three-dimensional medical image segmentation prediction includes the following steps:

[0010] Step 1, preprocess the obtained three-dimensional medical image data to obtain optimized three-dimensional medical image data after data resampling and data normalization;

[0011] Step 2, randomly crop at different positions of the optimized three-dimensional image data to obtain a plurality of training sub-blocks;

[0012] Step 3: Taking half of the size of the training sub-block as the step size, using a sliding window strategy to crop prediction sub-blocks with the same size as the training sub-block from the optimized 3D image data;

[0013] Step 4: Inputting the prediction sub-blocks into the trained 3D medical image segmentation network based on context information fusion, and obtaining the prediction segmentation results of each prediction sub-block through forward calculation;

[0014] Step 5: Using a Gaussian weighting strategy to fuse the prediction segmentation results of adjacent prediction sub-blocks to obtain the final predicted 3D medical image segmentation result.

[0015] Furthermore, the specific operation of data resampling in Step 1 is as follows:

[0016] Using the median of the dataset spacing of the 3D medical image data as the target spacing, and performing resampling on its image data and label data distribution using third-order Spline interpolation and nearest-neighbor interpolation.

[0017] Furthermore, the specific operation of data normalization in Step 1 is as follows:

[0018] Calculating the mean and standard deviation of the voxel intensities of all foreground classes in the 3D medical image data, as well as their 0.5 and 99.5 percentiles, then clipping all images to the 0.5 and 99.5 percentiles, and finally subtracting the mean and dividing by the standard deviation.

[0019] Furthermore, the training steps of the 3D medical image segmentation network in Step 4 are specifically as follows:

[0020] Step 4.1: Sampling the training sub-blocks using a sample balancing sampling strategy, and using the sampled training sub-blocks and their corresponding segmentation annotations as training samples;

[0021] Step 4.2: Defining a 3D medical image segmentation network based on context information fusion, and inputting the training samples into the 3D medical image segmentation network to obtain training segmentation results through forward calculation;

[0022] Step 4.3: Calculating the target loss function according to the training segmentation results and the true segmentation labels, and updating the parameters of the 3D medical image segmentation network;

[0023] Step 4.4: Judging whether the preset number of training rounds is reached until the training ends.

[0024] Furthermore, the 3D medical image segmentation network extracts features containing multi-scale local context information and global context using parallel dilated convolution branches and self-attention branches, and fuses the features of different branches to generate the results. The dilated convolution branch is stacked by G groups of dilated convolutions, and the self-attention branch is stacked by a convolutional layer, L layers of 3D axial self-attention layers, and a transposed convolutional layer.

[0025] Furthermore, the sample balancing sampling strategy in step 4.1 is specifically as follows:

[0026] Among the sampled training sub-blocks, two-thirds of the sub-blocks come from random positions, and the other one-third of the sub-blocks come from ensuring the inclusion of foreground categories.

[0027] Furthermore, the objective loss function of the 3D medical image segmentation network in step 4.3 is expressed as:

[0028] L total = λ 1 ·L CE + λ 2 ·L Dice ;

[0029] where L CE and L Dice are the cross-entropy loss and Dice loss between the segmentation result and the true segmentation label respectively, and λ 1 and λ 2 are the weights of the cross-entropy loss and Dice loss.

[0030] Furthermore, the steps of the Gaussian weighting strategy in step 5 are specifically as follows:

[0031] Use the Gaussian distribution as the weight for the segmentation result of each prediction sub-block, then normalize the weights between the segmentation results of adjacent prediction sub-blocks, and take the weighted sum of the prediction segmentation results as the final prediction segmentation result.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention provides a three-dimensional medical image segmentation method based on context information fusion. In the present invention, the three-dimensional medical image network mainly adopts a sub-block-based method for the training stage and the prediction stage. Through the dilated convolution branch and the self-attention branch in parallel of the three-dimensional medical image network, features containing context information are extracted, and after fusion, a predicted segmentation result is generated. Then, the obtained predicted segmentation result is used to fuse the predicted segmentation results of adjacent prediction sub-blocks by Gaussian weighting, alleviating the problem of inaccurate prediction at the sub-block boundaries and further improving the segmentation accuracy. In addition, during the training of the three-dimensional medical image network, a sample balancing sampling strategy is used to sample training sub-blocks, avoiding the possible class imbalance problem in the data and improving the segmentation accuracy of the foreground classes. The method of the present invention can effectively improve the accuracy of the three-dimensional image segmentation algorithm, be applicable to the three-dimensional medical image segmentation tasks of various organs, and has good universality and generality. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings that form a part of this application are used to provide a further understanding of this application, making other features, objectives, and advantages of this application more obvious. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation of this application.

[0034] In the drawings:

[0035] Figure 1 : is the structural diagram of the three-dimensional medical image segmentation network of a three-dimensional medical image segmentation method based on context information fusion implemented in the present invention;

[0036] Figure 2 : is the flowchart of the three-dimensional medical image network training of a three-dimensional medical image segmentation method based on context information fusion implemented in the present invention;

[0037] Figure 3 : is the flowchart of the image segmentation prediction of a three-dimensional medical image segmentation method based on context information fusion implemented in the present invention;

[0038] Figure 4 : is the visualization result of the output of a three-dimensional medical image segmentation method based on context information fusion implemented in the present invention on the NIH pancreas dataset: (a) is the true segmentation result, and (b) is the prediction result using the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although some embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. On the contrary, these embodiments are provided to more thoroughly and completely understand the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not used to limit the protection scope of the present disclosure.

[0040] In addition, it should be noted that for ease of description, only parts related to the relevant invention are shown in the drawings. Without conflict, the embodiments in the present disclosure and the features in the embodiments can be combined with each other.

[0041] The present invention discloses a three-dimensional medical image segmentation method for fusing human context information. The present disclosure will be described in detail below with reference to the accompanying drawings and in combination with embodiments.

[0042] Referring to Figure 3 as shown, the three-dimensional medical image segmentation prediction mainly includes the following steps:

[0043] Step 1: Preprocess the obtained three-dimensional medical image data to obtain optimized three-dimensional medical image data after data resampling and data normalization;

[0044] Step 2: Randomly crop multiple training sub-blocks at different positions of the optimized three-dimensional image data;

[0045] Step 3: Using half of the training sub-block size as the step size, adopt a sliding window strategy to crop prediction sub-blocks with the same size as the training sub-blocks from the optimized three-dimensional image data;

[0046] Step 4: Input the prediction sub-blocks into the trained three-dimensional medical image segmentation network based on context information fusion, and obtain the prediction segmentation results of each prediction sub-block through the forward calculation method;

[0047] Step 5: Use the Gaussian weighting strategy to fuse the prediction segmentation results of adjacent prediction sub-blocks to obtain the final predicted three-dimensional medical image segmentation result.

[0048] Specifically, the specific operation of data resampling is as follows: The median value of the dataset spacing of three-dimensional medical image data is used as the target spacing, and third-order Spline interpolation and nearest neighbor interpolation are used to resample the distribution of its image data and label data. Resampling refers to the process of interpolating the information of one type of pixel to obtain the information of another type of pixel. Resampling is mainly divided into upsampling and downsampling. When it is less than the original signal, it is downsampling, and the signal needs to be decimated; when it is greater than the original signal, it is upsampling, and the signal needs to be interpolated. The specific operation of data normalization is as follows: Calculate the mean and standard deviation of the voxel intensities of all foreground classes in the three-dimensional medical image data, as well as their 0.5 and 99.5 percentiles, then clip all images to the 0.5 and 99.5 percentiles, and finally subtract the mean and divide by the standard deviation to convert the dimensional quantity into a dimensionless quantity. Data normalization is a commonly used data preprocessing operation, aiming to convert data of different specifications to a unified specification or convert data of different distributions to a specific range to reduce the impact of scale, features, distribution differences, etc. on the model. Therefore, the optimized three-dimensional medical image data obtained after data resampling and data normalization of three-dimensional image data can ensure the consistency of the spacing of different three-dimensional image data, and thus better maintain the spacing between samples.

[0049] Random cropping is performed at different positions of the optimized three-dimensional image data to obtain multiple training sub-blocks, and with half of the size of the training sub-block as the step size, a sliding window strategy is used to crop prediction sub-blocks of the same size as the training sub-block from the optimized three-dimensional image data. The sliding window strategy operates on a string or array of a specific window size instead of the entire string and array, which reduces the complexity of the problem and thus also reduces the nested depth of the loop. Here, prediction sub-blocks are cropped through the sliding window strategy, and the number of prediction sub-blocks is controlled for transmission to accelerate data transmission, improve network throughput, and avoid data congestion.

[0050] Refer to Figure 2 As shown, the training steps of the three-dimensional medical image segmentation network include:

[0051] Step 4.1, sample the training sub-blocks using the sample balancing sampling strategy, and use the sampled training sub-blocks and their corresponding segmentation annotations as training samples;

[0052] Step 4.2, define a three-dimensional medical image segmentation network based on context information fusion, and input the training samples into the three-dimensional medical image segmentation network to obtain the training segmentation result through forward calculation;

[0053] Step 4.3, calculate the target loss function according to the training segmentation result and the true segmentation label, and update the parameters of the three-dimensional medical image segmentation network;

[0054] Step 4.4, determine whether the preset number of training rounds is reached until the training ends.

[0055] Specifically, first, preprocess the 3D medical image data using the same method as in the 3D medical image segmentation prediction stage to obtain the data after data resampling and data normalization. Then, adopt a sample balancing strategy to sample sub-blocks from the training sub-blocks, and use the sampled training sub-blocks and their corresponding segmentation annotations as training samples. The sample balancing strategy is specifically that two-thirds of the sub-blocks come from random positions, and the other one-third of the sub-blocks come from ensuring the inclusion of foreground categories. Then define a 3D medical image segmentation network based on context information fusion to calculate the training segmentation result according to the input of the sampled training samples, and calculate the objective loss function L total , and use stochastic gradient descent with Nesterov momentum as the optimization algorithm to update the parameters of the 3D medical image segmentation network. Among them, the initial learning rate is 0.01, the momentum is 0.9, and the weight decay rate is 10-4. The learning rate uses polynomial decay with an exponent of 0.9. Generally speaking, 250 batches are sampled per round, and the batch size is 2. Determine whether the preset total number of training rounds of 200 rounds is reached according to the current number of training rounds. If the training termination condition is reached, that is, the number of training rounds reaches the total number of rounds, then end the training and output the trained 3D medical image segmentation network; otherwise, return to Step 4.1 and continue to execute.

[0056] Furthermore, the objective loss function L total is expressed as:

[0057] L total = λ 1 ·L CE + λ 2 ·L Dice ;

[0058] Among them, L CE and L Dice are the cross-entropy loss and Dice loss between the training segmentation result and the true segmentation label respectively, and λ 1 and λ 2 are the weights of the cross-entropy loss and Dice loss.

[0059] Referring to Figure 1 the 3D medical image segmentation network structure shown, the 3D medical image segmentation network uses parallel dilated convolution branches and self-attention branches to extract features containing multi-scale local context information and global context, and fuse the features of different branches to generate the segmentation result; the dilated convolution branch is stacked by G groups of dilated convolution layers and convolution layers, and the self-attention branch is stacked by convolution layers, L layers of 3D axial self-attention layers and transposed convolution layers.

[0060] Specifically, the 3D medical image segmentation network includes a parallel dilated convolution branch, a self-attention branch, and a segmentation head. The dilated convolution branch is composed of four groups of dilated convolutional layers and convolutional layers stacked together, where the g-th group contains two layers of dilated convolutional layers, and the dilation rate is 2g - 1 for both. The dilated convolutional layer inserts dilations at equal intervals in the convolutional kernel of the convolutional layer, so that the receptive field of the feature is increased without changing the computational amount. The gradually increasing dilation rate makes the size of the network receptive field increase with the network depth. The self-attention branch is composed of a convolutional layer, six 3D axial self-attention layers, and a transposed convolutional layer stacked together. The 3D axial self-attention calculates self-attention successively according to the width, height, and depth directions to model the global relationship. In addition, the 3D axial self-attention uses shared relative position encoding to introduce position information.

[0061] Input the prediction sub-blocks into the trained 3D medical image segmentation network, and obtain the predicted segmentation results of each prediction sub-block through the forward calculation method. The forward calculation is the process of the forward propagation of the signal. The specific propagation process is as follows: First, use the 3×3×3 convolutional layer in the dilated convolution branch to process the output of each group of dilated convolutional layers, then fuse the output of each dilated convolutional layer after convolution with the output of the last transposed convolutional layer of the self-attention branch by element-wise addition, and finally generate the predicted segmentation result through the output of the 1×1×1 convolutional layer.

[0062] In step 5, use the Gaussian weighted strategy to fuse the predicted segmentation results of adjacent prediction sub-blocks to obtain the final 3D medical image segmentation result. Specifically, use the Gaussian distribution as the weight for the segmentation result of each prediction sub-block, then normalize the weights between the segmentation results of adjacent prediction sub-blocks, and take the predicted segmentation result after weighted summation as the final prediction result. This can effectively alleviate the problem of inaccurate prediction at the sub-block boundary, enabling it to be applied to the 3D medical image segmentation tasks of various organs, and having good universality and generality.

[0063] Embodiment The present invention conducts experiments on the NIH pancreas dataset

[0064] The NIH pancreas dataset includes 82 abdominal CT enhanced scans. Each CT scan is a 3D data of size 512×512×C1, where C1 ranges from 181 to 466. The labels of the pancreas in the data are provided by senior radiologists. The experiments conducted here only use pancreas segmentation as an example to verify the effectiveness of the present invention. However, the application scope of the present invention is not limited to pancreas segmentation and can also be applied to the segmentation of other organ regions.

[0065] The experiment compares the method of the present invention with the representative methods in 3D medical image segmentation, namely "3DU-Net" (a 3D image segmentation method based on convolutional networks) proposed by [authors] and "V-Net" (a volumetric fully convolutional 3D image segmentation method) proposed by Fausto Milletari et al. To verify the effectiveness of each structure in the network model (3D-CANet) proposed in the present invention, as well as the effectiveness of the sample balanced sampling in the training stage and the Gaussian weighted strategy in the prediction stage, 3D-CANet-wo-GC and 3D-CANet-single respectively represent removing the self-attention branch and only using the output of the last layer of the dilated convolution branch on the basis of the original 3D-CANet network. 3D-CANet-wo-sample and 3D-CANet-wo-weight respectively represent not adopting the sample balanced sampling strategy during training and not using the Gaussian weighted strategy during prediction. The Dice similarity coefficient (DSC) of the foreground class is used as the evaluation index in the experiment.

[0066] The experimental results are shown in the following table:

[0067] Method DSC 3D U-Net 0.8332 V-Net 0.8343 3D-CANet-wo-GC 0.8291 3D-CANet-single 0.8143 3D-CANet-wo-sample 0.8317 3D-CANet-wo-weight 0.8521 3D-CANet 0.8557

[0068] It can be seen from the table that the method of the present invention is superior to the 3D U-Net and V-Net methods because the method of the present invention avoids the resolution reduction and loss of detailed information caused by multi-stage downsampling in the aforementioned methods, and thus has better segmentation accuracy in the 3D pancreas segmentation task here. Secondly, the segmentation performance of 3D-CANet-wo-GC and 3D-CANet-single is lower than that of the original 3D-CANet network, which shows the effectiveness of using the multi-scale dilated convolution branch and the self-attention branch output in the method proposed in the present invention. Finally, the segmentation performance of 3D-CANet-wo-sample and 3D-CANet-wo-weight is lower than that of the original 3D-CANet network, which shows the effectiveness of the sample balanced sampling strategy and the Gaussian weighted strategy in the present invention.

[0069] Refer to Figure 4 As shown, the segmentation results of the present invention on the NIH pancreas dataset are presented. Columns (a) and (b) in the figure respectively represent the true segmentation results and the results predicted using the method of the present invention. It can be seen that the method proposed in the present invention can segment the pancreas well, and the segmentation results are very close to the true labels.

[0070] The above description is only some preferred embodiments of the present disclosure and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by the specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with the technical features (but not limited to) having similar functions disclosed in the embodiments of the present disclosure.

Claims

1. A 3D medical image segmentation method based on context information fusion, characterized in that, the 3D medical image segmentation prediction includes the following steps: Step 1, preprocess the obtained 3D medical image data to obtain optimized 3D medical image data after data resampling and data standardization; Step 2, randomly crop at different positions of the optimized 3D medical image data to obtain multiple training sub-blocks; Step 3, with half of the size of the training sub-block as the step size, use a sliding window strategy to crop prediction sub-blocks with the same size as the training sub-block from the optimized 3D medical image data; Step 4, input the prediction sub-blocks into the trained 3D medical image segmentation network based on context information fusion, and obtain the prediction segmentation results of each of the prediction sub-blocks through the forward calculation method; Step 5, use the Gaussian weighted strategy to fuse the prediction segmentation results of adjacent prediction sub-blocks to obtain the final predicted 3D medical image segmentation result; The training steps of the 3D medical image segmentation network in Step 4 are specifically as follows: Step 4.1, sample the training sub-blocks using the sample equalization sampling strategy, and use the sampled training sub-blocks and their corresponding segmentation annotations as training samples. Two-thirds of the sub-blocks in the training sub-blocks come from random positions, and the other one-third of the sub-blocks come from ensuring the inclusion of foreground categories; Step 4.2, define a 3D medical image segmentation network based on context information fusion, and input the training samples into the 3D medical image segmentation network to obtain the training segmentation result through forward calculation; Step 4.3, calculate the target loss function according to the training segmentation result and the true segmentation label, and update the parameters of the 3D medical image segmentation network; Step 4.4, judge whether the preset number of training epochs is reached until the training ends; The target loss function is expressed as, L total = λ 1 ·L CE + λ 2 ·L Dice ; Among them, L CE and L Dice are the cross-entropy loss and Dice loss between the segmentation result and the ground-truth segmentation label respectively, and λ 1 and λ 2 are the weights of the cross-entropy loss and Dice loss; The 3D medical image segmentation network uses parallel dilated convolution branches and self-attention branches to extract features containing multi-scale local context information and global context, and fuses the features of different branches to generate results; The dilated convolution branch is stacked by G groups of dilated convolutions, and the self-attention branch is stacked by a convolutional layer, L layers of 3D axial self-attention layers, and a transposed convolutional layer.

2. A 3D medical image segmentation method based on context information fusion according to claim 1, characterized in that, the specific operation of the data resampling in Step 1 is: Use the median of the dataset spacing of the 3D medical image data as the target spacing, and perform data resampling on its image data and label data distribution using third-order Spline interpolation and nearest neighbor interpolation.

3. A 3D medical image segmentation method based on context information fusion according to claim 2, characterized in that, the specific operation of the data standardization in Step 1 is: Calculate the average value and standard deviation of the voxel intensities of all foreground classes in the 3D medical image data, as well as their 0.5 and 99.5 percentiles, then clip all images to the 0.5 and 99.5 percentiles, and finally subtract the average value and divide by the standard deviation.

4. A 3D image segmentation method based on context information fusion according to claim 3, characterized in that, the steps of the Gaussian weighting strategy described in step 5 are specifically as follows: Taking the segmentation result of each of the prediction sub-blocks as weights using a Gaussian distribution, then normalizing the weights between the segmentation results of adjacent prediction sub-blocks, and using the weighted sum of the predicted segmentation results as the final predicted segmentation result.

Images