A method for segmenting MR images of the fetal cerebellar vermis

Abstract

The present invention provides a fetal cerebellar vermis MR image segmentation method, belonging to the field of image processing technology, the method comprising acquiring a fetal brain MR image, and preparing a whole brain target detection data set and a cerebellar vermis image segmentation data set; locating the midsagittal plane of the fetal brain MR image to be segmented according to the whole brain target detection data set and the cerebellar vermis image segmentation data set; constructing a fetal cerebellar vermis MR image segmentation network based on global multi-scale feature aggregation; using the constructed fetal cerebellar vermis MR image segmentation network, segmenting the fetal cerebellar vermis MR image on the screened midsagittal plane. The present invention effectively achieves the segmentation of the fetal cerebellar vermis MR image.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular relates to an MR image segmentation method of a fetal cerebellar vermis. Background Art

[0002] Fetal cerebellar vermis MRI segmentation is an important research point in medical image segmentation, and it is of key significance in the assessment of fetal growth and development. MRI is magnetic resonance imaging, which is highly safe, does not cause harm to the body, and performs well in soft tissue imaging. It can clearly show the subtle differences in brain tissue structure. When observing the gray matter and white matter structure of the brain, MRI can provide more detailed images than ultrasound. Therefore, when it comes to tasks such as fetal brain tissue segmentation and developmental assessment, the use of MRI can provide experts with clearer guidance. By accurately segmenting the fetal cerebellar vermis image, it can effectively provide assistance to clinicians and improve the accuracy of fetal cerebellar vermis assessment.

[0003] At present, deep learning-based methods have become the mainstream medical image segmentation method and have been applied to cerebellar vermis MRI. However, the existing technology uses deep learning models to automatically locate the pons and vermis on fetal brain MRI to assist radiologists in selecting high-quality fetal brain images and reducing the time spent on solution measurement. There are two problems with this method. First, the MRI image itself is a 3D data, which contains multiple slices. In clinical practice, the median sagittal plane needs to be screened out and then segmented. A single 2D segmentation task cannot meet the needs well. Secondly, the original UNet does not fully utilize multi-scale information, which will lead to reduced segmentation accuracy.

[0004] In the prior art, a real-time segmentation network for fetal brain MRI is proposed. The network uses two branches. At the beginning of the network, a downsampling is used to modify the feature map size, and then UNet and a separate branch are used to perform image segmentation operations. Although the network can improve the reasoning speed and has a high accuracy for fetal brain segmentation, the fetal brain tissue itself is very small. If the downsampling method is used at the beginning, a large amount of information will be lost, affecting the image accuracy. In the prior art, a multi-scale FCN network is used to segment fetal brain images, but because the FCN network only uses downsampling and upsampling processes, it is easy to lose detailed information in the image. Compared with the UNet network, it is relatively weak in capturing details in the image. A fully automatic fetal brain reconstruction framework is proposed in the prior art, which uses a convolutional network to achieve slice positioning, and uses a segmentation network implemented by deep supervision plus dilated convolution to extract brain regions. In this framework, the positioning method is not efficient, and it is not possible to accurately find the slices that need to be segmented, such as the midsagittal plane. In addition, the convolutional network does not fully utilize the information at each level, and the deep supervision method only downsamples the final result to obtain the loss, which affects the accuracy to a certain extent. A multi-task network is proposed in the prior art. The network proposes a separate molecule for image classification at the highest level. The main structure adds more levels on the basis of the UNet network, and adds a convolution for feature extraction in the jump connection part. The multi-level network may cause overfitting and increase the risk of gradient explosion.

[0005] In the general segmentation method, considering that the features of each layer in the decoder should be lower-level features, directly merging them with high-level features may have semantic gaps, and the fusion of incompatible features may cause some differences in the entire learning process, which will have an adverse effect on the prediction process. Therefore, some convolutional layers are merged in the jump connection, and residual connections are introduced to make learning easier. Although this method effectively improves the impact of semantic differences, it does not fully utilize multi-scale information, which affects the segmentation accuracy to a certain extent. Summary of the invention

[0006] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method for segmenting the MR image of the fetal cerebellar vermis, which solves the problem of finding the midsagittal plane where the cerebellar vermis is located from 3D fetal brain MR images and the problem that the current image segmentation network has low parameter quantity and cannot take into account high precision at the same time.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is: a MR image segmentation method of the fetal cerebellar vermis, comprising the following steps:

[0008] S1. Obtain fetal brain MR images and create a whole-brain target detection dataset and a cerebellar vermis image segmentation dataset;

[0009] S2, locating the midsagittal plane of the fetal brain MR image to be segmented based on the whole-brain target detection dataset and the cerebellar vermis image segmentation dataset;

[0010] S3, constructing a fetal cerebellar vermis MR image segmentation network based on global multi-scale feature aggregation;

[0011] S4. Using the constructed fetal cerebellar vermis MR image segmentation network, segment the MR image of the fetal cerebellar vermis on the screened midsagittal plane.

[0012] The beneficial effects of the present invention are as follows: the present invention accurately locates the median sagittal plane of a 3D fetal brain MR image, and proposes a low-parameter and high-precision fetal cerebellar vermis MR image segmentation network to effectively segment the fetal cerebellar vermis MR image.

[0013] Furthermore, the S1 is specifically:

[0014] Obtain fetal brain MR images;

[0015] The fetal brain MR images are annotated to obtain a whole-brain target detection dataset, where the annotation form is a rectangular box and the annotation categories are divided into three categories: category one is the median sagittal plane, category two is the layer beside the median sagittal plane, and category three is the layer other than category one and category two;

[0016] The fetal brain MR images are annotated to obtain a cerebellar vermis image segmentation dataset, wherein the annotation is only for the cerebellar vermis region that needs to be segmented.

[0017] The beneficial effect of the above further scheme is: by constructing a high-precision target detection and image segmentation dataset, it provides data support and guarantee for the efficient training of subsequent neural networks.

[0018] Furthermore, the S2 is specifically:

[0019] According to the whole-brain target detection dataset and the cerebellar vermis image segmentation dataset, the required slices were screened out using the Yolov8 target detection network;

[0020] According to the slices obtained by screening, the midsagittal plane of the fetal brain MR image to be segmented is located using the gray-level co-occurrence matrix.

[0021] The beneficial effect of the above further scheme is: locating the midsagittal plane from the 3D fetal brain MR image and filtering out the data required for the subsequent image segmentation task.

[0022] Furthermore, the fetal cerebellar vermis MR image segmentation network includes:

[0023] An encoder, used for extracting MR image features of the fetal cerebellar vermis on the screened midsagittal plane through convolution operations and pooling operations, and compressing the spatial dimension of the MR image of the fetal cerebellar vermis;

[0024] The decoder is used to restore the feature maps partially compressed by the encoder through deconvolution and convolution operations, generate segmentation results, and output the mask information of each level for the loss function to optimize the learning process of the network;

[0025] The global multi-scale feature aggregation block is used to modify the output of each layer of skip connection, fuse the feature map output by each layer encoder with the mask information output by the decoder, and use the channel visual Mamba fusion module to extract global features from the feature map containing information of each scale based on the fusion processing results. Among them, the channel visual Mamba fusion module uses the 2D selective scanning mechanism of the original Mamba network to learn global features, and adds a channel attention mechanism to realize the ability to extract channel information.

[0026] Furthermore, the loss function of the fetal cerebellar vermis MR image segmentation network is expressed as follows:

[0027]

[0028]

[0029]

[0030]

[0031] in, represents the loss function of the fetal cerebellar vermis MR image segmentation network, Represents the weights of different levels, Indicates level, Indicates The loss of levels, represents the dice loss function, represents the total number of samples, represents the predicted probability of the i-th sample, represents the true label of the i-th sample, represents the binary cross entropy loss function, Indicates The predicted probability that a sample is a positive class.

[0032] The beneficial effect of the above further scheme is: through the designed fetal cerebellar vermis MR image segmentation network, the U-shaped segmentation network can effectively improve the full utilization of mask information, global information, and multi-level information, improve the learning ability of the neural network, optimize the learning process, and thus improve the segmentation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The figure is a flow chart of the method of the present invention.

[0034] Figure 2 This is an example of a fetal brain MR image used in this embodiment.

[0035] Figure 3 This is an example diagram of target detection annotation in this embodiment.

[0036] Figure 4 This is an example diagram of image segmentation and annotation in this embodiment.

[0037] Figure 5 Schematic diagram of the midsagittal plane positioning process in this embodiment.

[0038] Figure 6 Schematic diagram of the YOLOv8 network structure in this embodiment.

[0039] Figure 7 This is a schematic diagram of the YOLOv8 network implementation results in this embodiment.

[0040] Figure 8 Schematic diagram of the gray-level co-occurrence matrix screening results in this embodiment.

[0041] Fig. 9 This is a diagram of the global multi-scale feature aggregation U-shaped segmentation network structure in this embodiment.

[0042] Fig.10 This is a diagram of the global multi-level feature aggregation block structure in this embodiment.

[0043] Fig.11 This is the SS2D structure diagram in this embodiment.

[0044] Fig.12 Schematic diagram of the structure of the visual mamba block and the channel visual mamba block in this embodiment. DETAILED DESCRIPTION

[0045] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0046] Example

[0047] The basic idea of ​​the present invention is to produce a whole-brain target detection dataset and a cerebellar vermis image segmentation dataset based on the fetal brain MRI images provided by the hospital, and use the efficient yolov8 target detection network to locate the midsagittal plane of the fetal brain MR data to be segmented, preliminarily screen out the required slices, and then extract the texture information through the grayscale co-occurrence moment, compare the slices after the initial screening, obtain data with more complex textures, and achieve midsagittal plane positioning. A fetal cerebellar vermis MR image segmentation network based on global multi-scale feature aggregation is proposed, and the cerebellar vermis is segmented using a segmentation network. The proposed network can better utilize mask and multi-level information, significantly improve the segmentation accuracy, and use a deep supervision method to supervise the output of each level in the decoder, which significantly improves the network training process.

[0048] like Figure 1 As shown, the present invention provides a method for segmenting MR images of the fetal cerebellar vermis, and the implementation method thereof is as follows:

[0049] S1. Obtain fetal brain MR images and create a whole-brain target detection dataset and a cerebellar vermis image segmentation dataset. The implementation method is as follows:

[0050] Obtain fetal brain MR images, such as Figure 2 As shown;

[0051] The fetal brain MR images are annotated to obtain a whole-brain target detection dataset, where the annotation form is a rectangular box and the annotation categories are divided into three categories: category one is the median sagittal plane, category two is the layer beside the median sagittal plane, and category three is the layer other than category one and category two;

[0052] The fetal brain MR images are annotated to obtain a cerebellar vermis image segmentation dataset, wherein the annotation is only for the cerebellar vermis region that needs to be segmented.

[0053] In this embodiment, the target detection data is annotated using Labelme, the annotation form is a rectangular frame, and the annotation is divided into three categories: the first category is the median sagittal plane, the second category is the layer beside the median sagittal plane, and the third category is other layers. The target detection annotation example is shown in FIG. Figure 3 shown.

[0054] In this embodiment, the image segmentation data is annotated using ITK-SNAP software, and the annotation is only for the cerebellar vermis area to be segmented. The image segmentation annotation example is shown in FIG. Figure 4 shown.

[0055] S2. Based on the whole brain target detection dataset and the cerebellar vermis image segmentation dataset, the midsagittal plane of the fetal brain MR image to be segmented is located. The implementation method is as follows: Figure 5 As shown:

[0056] According to the whole-brain target detection dataset and the cerebellar vermis image segmentation dataset, the required slices were screened out using the Yolov8 target detection network;

[0057] According to the slices obtained by screening, the midsagittal plane of the fetal brain MR image to be segmented is located using the gray-level co-occurrence matrix.

[0058] In this embodiment, the YOLOv8 network is used for training and reasoning, and the grayscale co-occurrence matrix is ​​used to achieve midsagittal plane positioning. The network is improved from YOLOv5. It consists of three parts: the backbone network (Backbone), the neck network (Neck), and the head network (head). The backbone network part is used for feature extraction, which is composed of convolution, channel pixel (C2f), and multi-scale feature extraction modules. The channel pixel module extracts basic features through the initial convolution block, and uses multiple bottleneck blocks to further capture detail information; the multi-scale feature extraction module uses a fast spatial pyramid pooling method. This method uses three maximum pooling to achieve downsampling of the feature map to reduce the computational complexity, and then uses different convolution kernels to extract features, and finally splices it in the channel dimension, so that the network can process information of different scales, thereby achieving the purpose of enhancing the feature extraction ability of the network. The neck network part is used for feature fusion, and the multi-scale feature fusion technology is used to fuse the feature maps of different scales from the backbone network. The head network part is a decoupled head structure, which is used to achieve the final target detection and classification tasks. The YOLOv8 structure diagram is shown in Figure 6 shown.

[0059] In this embodiment, after the initial screening by the YOLOv8 network, the slice selection can be effectively realized. However, since some brain slices have two or more slices that are extremely similar to the median sagittal plane, these slices can even be considered as the median sagittal plane. Figure 7 However, clinically, it is believed that the midsagittal plane shows clearer grooves. When there are multiple good slices, this method can be used for judgment. Therefore, the gray level co-occurrence matrix is ​​used to judge the groove texture. The gray level co-occurrence matrix describes the texture method through the spatial correlation characteristics of gray. The calculation formula of the energy parameter is as follows:

[0060]

[0061] in, represents the energy parameter, Represents the number of gray levels after quantization, is the gray level co-occurrence matrix Line The elements of the column represent the grayscale values ​​in the image. and The normalized frequency with which a pair of pixels occurs at a selected direction and distance.

[0062] The energy parameter can measure the uniformity and smoothness of the image texture, which reflects the distribution of elements in the gray-level co-occurrence matrix. The higher the value, the more regular the texture, and the lower the value, the more complex the texture. The results after screening by the gray-level co-occurrence matrix are as follows: Figure 8 As shown, it can be seen that this method can effectively determine the slices with more complex texture information, thereby achieving the effect of selecting the required slices.

[0063] S3, constructing a fetal cerebellar vermis MR image segmentation network based on global multi-scale feature aggregation;

[0064] The fetal cerebellar vermis MR image segmentation network includes:

[0065] An encoder, used for extracting MR image features of the fetal cerebellar vermis on the screened midsagittal plane through convolution operations and pooling operations, and compressing the spatial dimension of the MR image of the fetal cerebellar vermis;

[0066] The decoder is used to restore the feature maps partially compressed by the encoder through deconvolution and convolution operations, generate segmentation results, and output the mask information of each level for the loss function to optimize the learning process of the network;

[0067] The global multi-scale feature aggregation block is used to modify the output of each layer of skip connection, fuse the feature map output by each layer encoder with the mask information output by the decoder, and use the channel visual Mamba fusion module to extract global features from the feature map containing information of each scale based on the fusion processing results. Among them, the channel visual Mamba fusion module uses the 2D selective scanning mechanism of the original Mamba network to learn global features, and adds a channel attention mechanism to realize the ability to extract channel information.

[0068] In this embodiment, the encoder gradually extracts the features of the image through convolution and pooling operations, compresses the spatial dimensions of the image, and extracts high-level feature information. In each layer, the first convolution operation is used to extract the low-level features of the feature map of the previous layer and increase the number of channels; the second convolution operation is used to further extract deeper features; finally, the maximum pooling layer is used to reduce the dimensionality of the feature map space to reduce the width and height of the feature map, retain key features, enable the model to capture a larger receptive field, and reduce computational overhead.

[0069] In this embodiment, the decoder gradually restores the partially compressed feature map of the encoder through deconvolution and convolution operations to generate the segmentation result. The deconvolution in each layer is used to restore the spatial resolution and modify the high-dimensional feature map to a low-dimensional one so that it has the same number of channels as the feature map transmitted by the jump connection in each layer; the first convolution is used to fuse and modify the number of channels of the feature map spliced ​​by the jump connection feature map and the feature map after each upsampling, extract useful feature information, and eliminate redundant information; the second convolution further refines the features obtained by the previous convolution, refines and enhances the features; the convolution branch in each layer is used to output the mask information of each level for loss function calculation, optimize the model training process, and improve the utilization of mask information.

[0070] In this embodiment, the skip connection avoids the loss of low-level detail information due to the gradual decrease in the spatial resolution of the feature map and the gradual enhancement of the semantic information of the feature as the encoder network goes deeper. The global multi-scale feature aggregation module is used to modify the output of each layer of the skip connection, fuse the feature map output by each layer of the encoder with the mask information output by the decoder, and use the 2D selective scanning mechanism to learn the global features to optimize the network feature extraction capability.

[0071] In this embodiment, the present invention proposes a global multi-scale feature aggregation network, such as Fig. 9 As shown in the figure, the global multi-scale feature aggregation network consists of four parts, namely encoder, decoder, jump connection, and loss module. In both the encoder and decoder, the original UNet structure is used. In each layer of the encoder, the first ordinary convolution operation is performed and the number of channels is modified. The second convolution operation is performed to extract features, and then the maximum pooling is used to achieve spatial dimension reduction of the feature map. In the decoder, transposed convolution is used after each layer of convolution operation is performed to achieve upsampling and channel number modification. A global multi-scale feature aggregation module is proposed to modify the jump connection, and the feature map calculated in the encoder is fused with the mask output by each layer of the decoder to achieve multi-scale information feature aggregation operation. In the loss module, the information of each level of the decoder is output separately using 1*1 convolution to obtain the result of each layer. The real mask is modified to match the size of the feature map of each level using maximum pooling. Finally, the loss function is calculated using deep supervision.

[0072] In this embodiment, the global multi-scale feature aggregation block is a novel block proposed by the present invention. At present, many networks have taken into account the impact of multi-scale information on network accuracy, but they only consider the use of information of different receptive fields and different levels, but ignore the impact of mask information on model accuracy. In addition, some networks that do not use Transformer only use dilated convolution to achieve receptive field expansion, which is insufficient for global feature learning and affects the segmentation accuracy to a certain extent. The present invention absorbs the new methods proposed in recent years and applies them to the network to realize feature fusion of information at each level and mask information, and extract global features, so as to make fuller use of information and achieve the purpose of optimizing the accuracy of the network.

[0073] In this embodiment, a global multi-level feature aggregation block is proposed such as Fig.10 As shown, taking the third-level feature map as an example, the width of the third-level feature map is Gao Wei , the number of channels is First, the width and height of the low-dimensional feature maps of the first and second levels in the network are reduced to the size of the third-level feature maps by bilinear interpolation, and the width and height of the high-dimensional feature maps of the fourth and fifth levels are expanded to the size of the third-level feature maps, so that they can be spliced ​​with the third-level feature maps; then all feature maps are divided into blocks using functions, and each feature map is evenly divided into 4 blocks according to the channel dimension, and their labels from front to back are {0, 1, 2, 3}. In order to ensure that the mask information can be fully utilized, the blocked feature maps and masks are grouped and spliced, channel 1 is the mask, and the rest are spliced ​​to the back in sequence according to the label and feature map order, and four new sub-feature maps are obtained, each feature map size is , these four sub-feature maps all contain mask information and information of each level; then all feature maps are modified to the size of block feature maps using point-by-point convolution, and merged to restore the feature maps to the size of the third-level feature maps. To ensure the effectiveness of the merging of multi-scale features, normalization is required after merging; then the channel visual mamba CVMamba fusion module is used to perform global feature extraction on the feature maps containing information of various scales, and in order to avoid the disappearance of gradients, the residual connection is used to fuse the original feature map of the third layer with the new feature map to obtain the final result.

[0074] In this embodiment, Channel Visual Mamba CVMamba is formed by supplementing the channel attention mechanism on the basis of Visual Mamba VMamba. Visual Mamba VMamba integrates the advantages of convolutional neural networks and visual Transformer, and has stronger feature extraction capabilities and better computing efficiency. Visual Mamba VMamba is based on the state-space model. It adapts the continuous-time model to the deep learning framework through discretization, so that the model can process both time series data and non-time series data such as images and graphics. It also innovates the 2D Selective Scan Mechanism (SS2D). In the cross-scanning part, the scanning results are expanded into one-dimensional sequences from four directions (upper left to lower right, lower right to upper left, upper right to lower left, and lower left to upper right), and global information is aggregated from different directions to achieve the overall feature capture of the image; in the S6 module part, the one-dimensional sequence of each direction obtained by scanning is independently input into the S6 module, and important information is dynamically selected through the gating mechanism. The state-space model (SSM) is used to introduce a parameterized selective mechanism, and the gating weights are generated through learnable parameters to control the intensity of information transmission. Finally, the S6 module remaps the output sequence of each direction back to the image space; in the cross-fusion part, the sequences processed by the S6 module are fused and reorganized to generate a two-dimensional feature map as the final output, such as Fig.11 As shown. However, since the 2D selective scanning mechanism mainly captures spatial features by scanning the four directions of the image, it does not pay special attention to the information between channels, resulting in weak learning of channel features, which affects the segmentation accuracy to a certain extent. The channel attention mechanism can adaptively adjust the weight of each channel according to the importance of each channel in the input feature map, which can more effectively extract and integrate information from different channels, complementing the spatial information processing in the 2D selective scanning mechanism, thereby improving the overall feature expression ability. Therefore, SE (Squeeze-and-Excitation) channel attention is added to the VSS block of Visual Mamba VMamba to modify the original module. SE channel attention processes the input feature map using a residual method. In the squeezing module of the feature map processing branch, the input feature map is first compressed into a scalar for each channel through a global average pooling operation to extract global information and reduce the redundancy of the spatial dimension. In the excitation module, a fully connected layer is first used to reduce the feature dimension, reduce the amount of calculation and retain important features, and then the ReLU activation function is applied to introduce nonlinear features. Then, the generated weights are normalized using the Sigmoid activation function, and these weights are applied to the input feature map to weight the channels, thereby highlighting important features and suppressing unimportant features. The block structures of Visual Mamba VMamba and Channel Visual Mamba CVMamba are as follows: Fig.12 shown.

[0075] In this embodiment, the deep supervision technology is used to optimize the real loss function. The deep supervision outputs the feature map of each layer in the decoder using a 1*1 convolution to obtain the feature map of each layer, and then downsamples the real segmentation result using the maximum pooling to match the segmentation result output by each layer, and then calculates the loss. This method can obtain feedback information in advance, thereby accelerating convergence, optimizing the training process, and achieving the purpose of improving accuracy.

[0076] The loss function used in the present invention is composed of Dice loss and Binary Cross Entropy (BCE) loss.

[0077] The Dice loss function formula is as follows:

[0078]

[0079] The BCE loss function formula is as follows:

[0080]

[0081] The loss function calculation at each layer is the Dice loss function value and the BCE loss function value multiplied by a weight of 0.5 as shown below:

[0082]

[0083] The total loss function is the loss of each layer multiplied by a weight, and the calculation formula from high dimension to low dimension is {1, 0.5, 0.4, 0.3, 0.2} as shown below:

[0084]

[0085] in, represents the loss function of the fetal cerebellar vermis MR image segmentation network, Represents the weights of different levels, Indicates level, Indicates The loss of levels, represents the dice loss function, represents the total number of samples, represents the predicted probability of the i-th sample, represents the true label of the i-th sample, represents the binary cross entropy loss function, Indicates The predicted probabilities of samples being positive are {1, 0.5, 0.4, 0.3, 0.2} from high dimension to low dimension.

[0086] S4. Using the constructed fetal cerebellar vermis MR image segmentation network, segment the MR image of the fetal cerebellar vermis on the screened midsagittal plane.

[0087] The present invention can select the required slices of 3D MRI data by using the midsagittal plane positioning method. The present invention effectively utilizes multi-scale information to achieve more accurate fetal cerebellar vermis image segmentation.

[0088] To demonstrate the effectiveness of the proposed segmentation network, experiments were conducted on a dataset consisting of 300 fetal cerebellar vermis images. The dataset was split into training, validation, and test sets in a 6:2:2 ratio. Due to the small size of the dataset, three datasets were generated using a 5-fold cross-validation approach for the experiment, and the experimental results were averaged.

[0089] This experimental platform uses the Ubantu22.04 system, Python 3.8, PyTorch deep learning framework version 2.0.0, CUDA version 11.3, CPU: Intel 12600kf, system memory size: 64GB, GPU: NVIDIA 2080ti graphics processor (video memory 22GB).

[0090] The accuracy evaluation indicators are Dice coefficient and intersection-over-union ratio. Dice coefficient is an image segmentation indicator used to quantify the segmentation quality and is used to compare the degree of overlap between the segmentation result and the true value, as shown in the following formula:

[0091]

[0092] in, represents the region predicted by the model, Indicates the true area. The value range of the Dice coefficient is [0,1]. The larger the value, the more consistent the prediction is with the truth.

[0093] The intersection-over-union ratio mainly evaluates the accuracy of the model by calculating the degree of overlap between the segmentation result and the true value. The calculation formula is shown as follows:

[0094]

[0095] Among them, the value range of the IoU coefficient is [0,1]. The larger the value, the more consistent the prediction is with the truth.

[0096] Select the number of parameters (Params) and floating-point operations (Flops) as the model scale indicators. The number of parameters refers to the total number of trainable parameters in the model, measured in millions (M). The number of parameters reflects the complexity and capacity of the model. The more parameters there are, the greater the storage requirements and memory usage during training, and the slower the model training and reasoning speed. The number of floating-point operations refers to the number of floating-point operations required by the model when performing a forward propagation, measured in billions (G). It is an indicator of the computational complexity of the model, indicating the amount of computation required during the reasoning process. The higher the FLOPs, the greater the computational requirements of the model, and the slower the reasoning speed may be.

[0097] To prove the effectiveness of the proposed method, it is compared with 10 existing mainstream segmentation networks. The experimental results are shown in Table 1. The results show that the segmentation network proposed in this invention is superior to other segmentation networks in terms of Dice coefficient and intersection over union accuracy, and is only higher than EGEUNet in terms of parameter quantity and floating-point operation quantity. The Fetal Vermis Cerebelli dataset used in the experiment in Table 1.

[0098] Table 1

[0099] Model Parameter quantity (M) Floating point operations (G) Dice coefficient (%) Intersection ratio (%) UNet 31.04 85.53 91.02 84.70 MultiResUNet 7.25 29.31 85.85 78.57 SwinUNet 27.15 12.21 70.41 57.68 UCTransnet 66.24 67.24 90.99 84.87 PAttUNet 11.17 26.31 89.90 83.69 EGEUNet 0.05 0.11 86.28 77.91 RFBSNET 1.94 4.57 90.94 84.26 VMUNet 34.62 11.82 88.04 79.95 VMUNetv2 17.91 6.88 74.06 63.55 HVMUNet 6.44 1.16 86.97 78.65 GMFA_UNet 0.54 1.71 91.74 85.33

[0100] In order to reflect the effectiveness of the various structural improvements of this network, the present invention conducted ablation experiments on the jump connection and the loss function, as shown in Table 2, which is an ablation experiment table. The present invention selects the UNet network with 8 channels as the baseline network (Baseline), and explores the influence of the multi-level feature group aggregation module and the deep supervision loss function on the model segmentation accuracy and parameter quantity through two ablation experiments. In the first ablation experiment, the global multi-scale feature aggregation block was added to the jump connection, which enabled the network to make rational use of masks, multi-level information and global information, and effectively improved the network accuracy. In the second ablation experiment, the deep supervision method was added to optimize the network training process. By calculating the loss of the feature map of each layer of the decoder, the training process of the model was improved, and the accuracy was improved. The experimental results show that the addition of both parts can effectively improve the accuracy of the network.

[0101] Table 2

[0102] Model Parameter quantity (M) Floating point operations (G) Dice coefficient (%) Intersection ratio (%) Baseline(UNet8) 0.49 1.38 89.02 82.59 Baseline+SkipC 0.54 1.71 91.25 85.23 Baseline+Loss 0.49 1.38 90.23 83.90 GMFA_UNet 0.54 1.71 91.74 85.33

[0103] In summary, through the above design, the present invention can accurately locate the midsagittal plane of the 3D fetal brain MR image, and propose a low-parameter and high-precision fetal cerebellar vermis MR image segmentation network, which effectively realizes the segmentation of the fetal cerebellar vermis MR image.

Claims

1. A method for segmenting MR images of the fetal cerebellar vermis, characterized in that: The following steps are involved: S1. Obtain fetal brain MR images and create a whole-brain target detection dataset and a cerebellar vermis image segmentation dataset; S2, locating the midsagittal plane of the fetal brain MR image to be segmented based on the whole-brain target detection dataset and the cerebellar vermis image segmentation dataset; S3, constructing a fetal cerebellar vermis MR image segmentation network based on global multi-scale feature aggregation; S4, using the constructed fetal cerebellar vermis MR image segmentation network, segmenting the fetal cerebellar vermis MR image on the screened midsagittal plane; The fetal cerebellar vermis MR image segmentation network comprises: An encoder, used for extracting MR image features of the fetal cerebellar vermis on the screened midsagittal plane through convolution operations and pooling operations, and compressing the spatial dimension of the MR image of the fetal cerebellar vermis; The decoder is used to restore the feature maps partially compressed by the encoder through deconvolution and convolution operations, generate segmentation results, and output the mask information of each level for the loss function to optimize the learning process of the network; The global multi-scale feature aggregation block is used to modify the output of each layer of skip connection, fuse the feature map output by each layer encoder with the mask information output by the decoder, and use the channel visual Mamba fusion module to extract global features from the feature map containing information of each scale based on the fusion processing results; the channel visual Mamba fusion module uses the 2D selective scanning mechanism of the original Mamba network to learn global features, and adds a channel attention mechanism to realize the ability to extract channel information; The global multi-level feature aggregation block first reduces the width and height of the low-dimensional feature maps of the first and second levels in the network to the size of the third-level feature maps by bilinear interpolation, and expands the width and height of the high-dimensional feature maps of the fourth and fifth levels to the size of the third-level feature maps, so that they can be spliced ​​with the third-level feature maps; then all feature maps are divided into blocks using a function, and each feature map is evenly divided into 4 blocks according to the channel dimension, which are numbered {0, 1, 2, 3} from front to back, and the block feature maps and masks are grouped and spliced, channel 1 is used as the mask, and the remaining channels are sequentially spliced ​​to the back according to the number and feature map order, to obtain four new sub-feature maps, which contain mask information and information of each level; then all feature maps are modified to the size of the block feature maps using point-by-point convolution, and merged so that the feature maps are restored to the size of the third-level feature maps, and normalized after merging; then the original feature map of the third layer is fused with the new feature map using residual connection to obtain the final result.

2. The MR image segmentation method of the fetal cerebellar vermis according to claim 1, characterized in that: The S1 is specifically: Obtain fetal brain MR images; The fetal brain MR images are annotated to obtain a whole-brain target detection dataset, where the annotation form is a rectangular box and the annotation categories are divided into three categories: category one is the median sagittal plane, category two is the layer beside the median sagittal plane, and category three is the layer other than category one and category two; The fetal brain MR images are annotated to obtain a cerebellar vermis image segmentation dataset, wherein the annotation is only for the cerebellar vermis region that needs to be segmented.

3. The MR image segmentation method of the fetal cerebellar vermis according to claim 1, characterized in that: The S2 is specifically: According to the whole-brain target detection dataset and the cerebellar vermis image segmentation dataset, the required slices were screened out using the Yolov8 target detection network; According to the slices obtained by screening, the midsagittal plane of the fetal brain MR image to be segmented is located using the gray-level co-occurrence matrix.

4. The MR image segmentation method of the fetal cerebellar vermis according to claim 1, characterized in that: The expression of the loss function of the fetal cerebellar vermis MR image segmentation network is as follows: ； ； ； ； in, represents the loss function of the fetal cerebellar vermis MR image segmentation network, Represents the weights of different levels, Indicates level, Indicates The loss of levels, represents the dice loss function, represents the total number of samples, represents the predicted probability of the i-th sample, represents the true label of the i-th sample, represents the binary cross entropy loss function, represents the true label of the i-th sample, Indicates The predicted probability that a sample is a positive class.

Images