Neonatal brain magnetic resonance image segmentation method based on deep learning

Abstract

The present application relates to a method for segmenting a nuclear magnetic resonance image of a newborn brain based on deep learning, the method comprising: inputting a first slice of a magnetic resonance T1 modality image and a second slice of a magnetic resonance T2 modality image into a segmentation model of a nuclear magnetic resonance image of a newborn brain, and obtaining a final image segmentation result corresponding to the nuclear magnetic resonance image to be segmented; the segmentation model of a nuclear magnetic resonance image of a newborn brain comprises a parallel information encoding module and a feature decoding module. The present application performs tissue segmentation on a magnetic resonance image of a newborn brain, and makes full use of the multimodal information of the magnetic resonance image so that the model can achieve accurate segmentation for images of newborns with irregular tissue distribution and rapid tissue changes; compared with the prior art, the present application performs more accurate tissue segmentation on a magnetic resonance image of a newborn brain, and the error between the predicted tissue segmentation result and the doctor's manual labeling is smaller, which is more conducive to clinical application.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular, to a method for segmenting magnetic resonance images of the brain of a newborn based on deep learning. Background Art

[0002] In the medical field, in order to meet the needs of disease diagnosis and treatment plan formulation, it is often necessary to scan patients to determine the condition of internal organs. Nuclear magnetic resonance technology is widely used in the medical field. It makes the biological tissue structure such as brain and cardiovascular visible. It emits radio frequency pulses of a specific frequency to a stationary human body, resonates with hydrogen nuclei in the human body and induces MR signals to form images.

[0003] Compared with CT imaging, magnetic resonance imaging (MRI) has obvious advantages: first, MRI is non-destructive and is performed without any ionizing radiation; second, MRI has a flow effect and can identify heart and vascular abnormalities; finally, MRI uses three-dimensional imaging and can simultaneously obtain magnetic resonance images of the patient's brain in the coronal, sagittal and horizontal planes, which is conducive to the location of diseased tissue.

[0004] There are three modalities of MRI in clinical application. These three modalities have their own imaging characteristics and can reflect different types of biological information of the same tissue, so that doctors can choose different modalities to observe tissues according to specific needs. These three modalities correspond to three different imaging methods, namely T1 imaging, T2 imaging and Flair imaging, among which T1 weighted images and T2 weighted images are more commonly used in clinical practice. T1 weighted images are generated with shorter repetition time and echo time, while T2 weighted images are generated with longer repetition time and echo time. Echo time refers to the time between the emission of a radio pulse and the reception of an echo signal, and repetition time refers to the duration between two consecutive pulse sequences on the same image slice. In the observation of brain magnetic resonance images, the location and structure of white matter, gray matter and cerebrospinal fluid in human brain tissue are the key observations of doctors. In T1 modality brain MR images, there is a clear contrast between different cranial nerve soft tissue structures, so the anatomical structure of brain tissue can be better observed. In T2 mode images, cerebrospinal fluid will show a high signal, which is often used to display diseased tissue and is suitable for observing the lesion site and performing qualitative or quantitative analysis.

[0005] The neonatal period is a critical period for brain development. The segmentation study of neonatal brain MRI images is of great significance for the prevention and treatment of neonatal brain diseases: First, the neonatal brain is in a critical period of development, and the risk of brain diseases is high. The segmentation processing of neonatal brain MR images can provide valuable information for the diagnosis of brain diseases, which is helpful for the early diagnosis and treatment of brain diseases. Secondly, there are obvious differences between neonatal brain MR images and adult brain images. According to the T1-weighted MRI of adult and neonatal brains, it can be seen that the contrast of white matter-gray matter in adult and neonatal brain tissue is opposite; and the contrast of neonatal brain tissue is low, and the intra-tissue heterogeneity is large. Therefore, the adult brain MR segmentation algorithm is not suitable for neonatal brain MR image segmentation. Finally, the low resolution of neonatal brain MR images makes accurate segmentation more difficult. In clinical practice, neonates are mainly examined and diagnosed through multimodal MRI of the brain; due to factors such as the practical experience, personal knowledge accumulation and working hours of different doctors, the analysis results of the same image may vary greatly. Summary of the invention

[0006] In order to overcome at least one deficiency in the prior art, an embodiment of the present application provides a method for segmenting a neonatal brain magnetic resonance image based on deep learning.

[0007] In a first aspect, a newborn brain magnetic resonance image segmentation model is provided, comprising: a parallel information encoding module and a feature decoding module; the parallel information encoding module comprises a long sequence information encoder, a convolution encoder and a feature fusion module, and the feature decoding module comprises a reconstruction output module and a skull stripping module;

[0008] The magnetic resonance image to be segmented includes a magnetic resonance T1 modality image and a magnetic resonance T2 modality image, and for each first slice of the magnetic resonance T1 modality image and each second slice of the magnetic resonance T2 modality image, the convolution encoder is used to obtain a first hierarchical feature corresponding to the first slice and a second hierarchical feature corresponding to the second slice;

[0009] The long sequence information encoder is used to obtain the third-level features corresponding to the first slice and the fourth-level features corresponding to the second slice;

[0010] The feature fusion module is used to fuse the first-level features and the second-level features to obtain the features after convolution coding fusion; to fuse the third-level features and the fourth-level features to obtain the features after information coding fusion; and to concatenate the features after convolution coding fusion and the features after information coding fusion to obtain the final fused features;

[0011] The reconstruction output module is used to obtain the preliminary image segmentation results based on the final fused features;

[0012] The skull stripping module is used to obtain the image segmentation result without the skull based on the final fused features;

[0013] According to the preliminary image segmentation results and the image segmentation results without the skull, the final image segmentation results corresponding to the first slice and the second slice are obtained; for all the first slices and all the second slices, all the final image segmentation results are spliced ​​to obtain the final image segmentation results corresponding to the magnetic resonance image to be segmented.

[0014] In one embodiment, the convolution encoder includes two convolution encoder structures, respectively used to obtain a first hierarchical feature corresponding to the first slice and a second hierarchical feature corresponding to the second slice;

[0015] Each convolutional encoder structure includes four convolutional encoder units connected in sequence, and each convolutional encoder unit includes a dense connection module, a maximum pooling layer, and a CoT attention module connected in sequence;

[0016] Each convolutional encoder unit outputs a hierarchical feature, and the four hierarchical features output by four encoder units constitute the first hierarchical feature or the second hierarchical feature.

[0017] In one embodiment, the long sequence information encoder includes two information encoder structures, respectively used to obtain the third hierarchical features corresponding to the first slice and the fourth hierarchical features corresponding to the second slice;

[0018] Each information encoder structure includes four information encoder units connected in sequence, wherein the first information encoder unit includes a linear embedding layer and a Swin Transformer block, and each of the remaining three information encoder units includes an image block fusion layer and a Swin Transformer block;

[0019] Each information encoder unit outputs a hierarchical feature, and the four hierarchical features output by the four information encoder units constitute the third hierarchical feature or the fourth hierarchical feature.

[0020] In one embodiment, the reconstruction output module includes five decoders connected in sequence, each decoder including an unpooling layer, a concatenation layer, and a dense connection module;

[0021] The unpooling layer is used to upsample the input to obtain the upsampled features;

[0022] The concatenation layer is used to concatenate and output the upsampled features and the shallow features in the final fused features using skip connections;

[0023] The densely connected module is used to perform convolution on the concatenated output.

[0024] In one embodiment, the skull stripping module includes five decoders connected in sequence, the first four encoders of the five decoders are the first four encoders of the reconstruction output module; the parameters of the last encoder of the five encoders are different from the parameters of the last encoder of the reconstruction output module.

[0025] In a second aspect, a method for segmenting a neonatal brain magnetic resonance image based on deep learning is provided, comprising:

[0026] The magnetic resonance images to be segmented include magnetic resonance T1 modality images and magnetic resonance T2 modality images;

[0027] Inputting each first slice of the magnetic resonance T1 modality image and each second slice of the magnetic resonance T2 modality image into a newborn brain nuclear magnetic resonance image segmentation model to obtain a final image segmentation result corresponding to the first slice and the second slice;

[0028] For all first slices and all second slices, all final image segmentation results are spliced ​​to obtain a final image segmentation result corresponding to the magnetic resonance image to be segmented;

[0029] The neonatal brain MRI image segmentation model is a neonatal brain MRI image segmentation model according to any one of claims 1-5.

[0030] In one embodiment, it further includes:

[0031] The newborn brain MRI image segmentation model was trained. When training the reconstruction output module, the training data used included MRI T1 modality images, MRI T2 modality images, and brain tissue benchmark images without skull removal. When training the skull stripping module, the training data used included MRI T1 modality images, MRI T2 modality images, and brain tissue benchmark images without skull removal.

[0032] Compared with the prior art, the present application has the following beneficial effects: the present application performs tissue segmentation on the magnetic resonance images of the brains of newborns, and fully utilizes the multimodal information of the magnetic resonance images so that the model can achieve accurate segmentation for images of newborns with irregular tissue distribution and rapid tissue changes; compared with the prior art, the present application performs tissue segmentation on magnetic resonance images of the brains of newborns more accurately, and the error between the predicted tissue segmentation results and the manual labels of doctors is smaller, which is more conducive to clinical application. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The present application may be better understood by referring to the following description given in conjunction with the accompanying drawings, which together with the following detailed description are included in this specification and form a part of this specification. In the drawings:

[0034] Figure 1 A schematic diagram of a neonatal brain magnetic resonance image segmentation model according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0035] The exemplary embodiments of the present application will be described below in conjunction with the accompanying drawings. For the sake of clarity and conciseness, not all features of the actual embodiments are described in the specification. However, it should be understood that many implementation-specific decisions can be made in the process of developing any such actual embodiments in order to achieve the specific goals of the developer, and these decisions may vary from embodiment to embodiment.

[0036] It is also necessary to explain here that, in order to avoid obscuring the present application due to unnecessary details, only the device structure closely related to the scheme according to the present application is shown in the drawings, while other details that are not closely related to the present application are omitted.

[0037] It should be understood that the present application is not limited to the described implementation forms due to the following description with reference to the accompanying drawings. In this article, where feasible, the embodiments can be combined with each other, features between different embodiments can be replaced or borrowed, and one or more features can be omitted in one embodiment.

[0038] This application proposes a deep learning-based segmentation method for neonatal brain MRI images based on the characteristics of neonatal brain MRI images, and uses skull stripping tasks to assist segmentation; the main content is to use deep learning methods to build a new medical image segmentation model to perform tissue segmentation and skull stripping on neonatal brain MRI images; and to verify the effectiveness of this network structure through experiments. Among them, the types of neonatal brain tissue segmentation include: cerebrospinal fluid (CSF), cortical gray matter (CGM), white matter (WM), ventricles (VT), cerebellum (CB), deep gray matter (DGM), brainstem (BS), hippocampus and amygdala (HA).

[0039] The deep learning-based newborn brain MRI image segmentation method of the present application is a brain tissue segmentation method based on a two-stage network. The main body of the two-stage network model consists of two parts: an information encoding module and a feature decoding module; wherein, the information encoding consists of two parts, namely, a parallel information encoding part and a feature fusion part; the feature decoding module completes the skull stripping task and the segmentation task, and is mainly composed of two parts, namely, a skull stripping part and a reconstruction output part. The two-modal 3D magnetic resonance images are sliced ​​and input into the network in sequence to obtain the segmentation results of the corresponding slices, and the segmentation results are spliced ​​in sequence and saved as files with the suffix nii format to obtain a 3D brain tissue segmentation result.

[0040] The present application embodiment provides a newborn brain magnetic resonance image segmentation model. Figure 1 A schematic diagram of a newborn brain nuclear magnetic resonance image segmentation model according to an embodiment of the present application is shown, and the model is used to segment the nuclear magnetic resonance image to be segmented to obtain an image segmentation result, that is, a brain tissue segmentation result. Before using the model for image segmentation, the magnetic resonance T1 modality image and the magnetic resonance T2 modality image of the nuclear magnetic resonance image to be segmented are first sliced ​​and divided into a plurality of first slices and a plurality of second slices, respectively, and then the first slices and the second slices are input into the model to obtain the segmentation results corresponding to the slices.

[0041] Specifically, the neonatal brain magnetic resonance image segmentation model includes: a parallel information encoding module and a feature decoding module; the parallel information encoding module includes a long sequence information encoder, a convolution encoder and a feature fusion module, and the feature decoding module includes a reconstruction output module and a skull stripping module;

[0042] The magnetic resonance image to be segmented includes a magnetic resonance T1 modality image and a magnetic resonance T2 modality image, and for each first slice of the magnetic resonance T1 modality image and each second slice of the magnetic resonance T2 modality image, the convolution encoder is used to obtain a first hierarchical feature corresponding to the first slice and a second hierarchical feature corresponding to the second slice;

[0043] The long sequence information encoder is used to obtain the third-level features corresponding to the first slice and the fourth-level features corresponding to the second slice;

[0044] The feature fusion module is used to fuse the first hierarchical features and the second hierarchical features to obtain convolutional coding fused features; to fuse the third hierarchical features and the fourth hierarchical features to obtain information coding fused features; to concatenate the convolutional coding fused features and the information coding fused features to obtain the final fused features, where the final fused features are also hierarchical features; here, the first hierarchical features and the second hierarchical features are fused, specifically by selecting the larger pixel value of the corresponding position of the first hierarchical features and the second hierarchical features as the fused value of the corresponding position.

[0045] The reconstruction output module is used to obtain the preliminary image segmentation results based on the final fused features;

[0046] The skull stripping module is used to obtain the image segmentation result without the skull based on the final fused features;

[0047] According to the preliminary image segmentation result and the image segmentation result after removing the skull, the final image segmentation result corresponding to the first slice and the second slice is obtained; here, the preliminary image segmentation result and the image segmentation result after removing the skull can be multiplied to obtain the final image segmentation result corresponding to the first slice and the second slice; for all first slices and all second slices, all the final image segmentation results are spliced ​​to obtain the final image segmentation result corresponding to the magnetic resonance image to be segmented.

[0048] In the above embodiments of the present application, the magnetic resonance T1 modality image and the magnetic resonance T2 modality image are sliced ​​and input, and a convolution encoder and a long sequence information encoder are used to learn the local features and global features of the image respectively; at the same time, a dual-modality input is used to cleverly utilize the redundancy and complementarity between different modalities of MRI. The present application performs tissue segmentation on the magnetic resonance images of the brain of newborns, and makes full use of the multimodal information of the magnetic resonance images so that the model can achieve accurate segmentation for newborn images with irregular tissue distribution and rapid tissue changes.

[0049] In one embodiment, the convolution encoder includes two convolution encoder structures, respectively used to obtain a first hierarchical feature corresponding to the first slice and a second hierarchical feature corresponding to the second slice;

[0050] Each convolutional encoder structure includes four convolutional encoder units connected in sequence, and each convolutional encoder unit includes a dense connection module, a maximum pooling layer, and a CoT attention module connected in sequence;

[0051] Each convolutional encoder unit outputs a hierarchical feature, and the four hierarchical features output by four encoder units constitute the first hierarchical feature or the second hierarchical feature.

[0052] In this embodiment, the convolution encoder is used to learn the local features of the image. For the target slice with index i, the first slice or the second slice with index i-5 to i+5 is selected for channel-by-channel splicing, and the spliced ​​slice is used as the input of the first convolution encoder unit. The dense connection module of the first convolution encoder unit uses a convolution layer to repeatedly downsample the spliced ​​slice, and the convolution kernel size of the convolution layer is 5×5. Zero padding is used for each convolution to keep the size of the output feature unchanged; after convolution, the PReLU activation function is used to process the result, and then the maximum pooling layer is used for maximum pooling, the pooling window size is 2×2, and the downsampling step is 2; the output of the maximum pooling layer is input to the CoT attention module for feature enhancement, and the hierarchical features output by the first convolution encoder unit are obtained; the hierarchical features output by the first convolution encoder unit are input into the second convolution encoder unit, and then a hierarchical feature is output; and so on, four hierarchical features output by four encoder units are finally obtained, and the four hierarchical features output by the four encoder units constitute the first hierarchical features or the second hierarchical features.

[0053] In one embodiment, the long sequence information encoder includes two information encoder structures, respectively used to obtain the third hierarchical features corresponding to the first slice and the fourth hierarchical features corresponding to the second slice;

[0054] Each information encoder structure includes four information encoder units connected in sequence, wherein the first information encoder unit includes a linear embedding layer and a Swin Transformer block, and each of the remaining three information encoder units includes an image block fusion layer and a Swin Transformer block;

[0055] Each information encoder unit outputs a hierarchical feature, and the four hierarchical features output by the four information encoder units constitute the third hierarchical feature or the fourth hierarchical feature.

[0056] In this embodiment, a long sequence information encoder is used to learn the global features of an image. The target slice indexed as i is used as the input of the information encoder structure, and the first information encoder unit includes a linear embedding layer and a Swin Transformer block, where the linear embedding layer uses a convolution operation to serialize the two-dimensional image information of the input first slice or the second slice, and the Swin Transformer block includes a plurality of residual modules connected in sequence, each residual module includes a normalization layer and a multi-head attention module, and the first information encoder unit outputs a hierarchical feature. Each of the remaining three information encoder units includes an image block fusion layer and a Swin Transformer block. Here, the output of the first information encoder unit is input into the second information encoder unit. The image block fusion layer in the second information encoder unit downsamples the input in order to obtain hierarchical features similar to the convolutional coding part, so that the obtained features have multi-scale information. Specifically, a feature map with a dimension of H×W×C, where H is the height of the image, W is the width of the image, and C is the number of feature channels of the image, is extracted from the H×W two-dimensional matrix according to certain rules, and then the extracted image blocks are spliced ​​in the channel direction to obtain a feature map with a dimension of The concatenated feature map is normalized and linearly mapped. The second information encoder unit also outputs a hierarchical feature; and so on, four hierarchical features output by four information encoder units are finally obtained, and the four hierarchical features output by the four information encoder units constitute the third hierarchical feature or the fourth hierarchical feature.

[0057] In one embodiment, the reconstruction output module includes five decoders connected in sequence, each decoder includes a reverse pooling layer, a concatenation layer and a dense connection module; and a PReLU activation function is used in the decoder.

[0058] The unpooling layer is used to upsample the input to obtain the upsampled features; here, the pooling window size is 2×2;

[0059] The concatenation layer is used to concatenate and output the upsampled features and the shallow features in the final fused features using skip connections;

[0060] The densely connected module is used to perform convolution on the concatenated output.

[0061] In this embodiment, the input of the first decoder is the deep features of the features after the final fusion, where the deep features refer to the last hierarchical features of the features after the final fusion, and the output is the convolution result; the input of the second decoder is the output of the first decoder, and the output is input to the third decoder; and so on, the convolution result output by the dense connection module of the fifth decoder is the preliminary image segmentation result, and the preliminary image segmentation result is the probability that each pixel of the image belongs to different categories of brain tissue; here, the preliminary image segmentation result is the image segmentation result without removing the skull.

[0062] Further, in other embodiments, in order to guide the brain tissue segmentation network to focus on the brain tissue area rather than the non-brain tissue area, a skull stripping auxiliary task is used to extract brain tissue from the magnetic resonance image. The skull stripping module includes five decoders connected in sequence, the first four encoders of the five decoders are the first four encoders of the reconstruction output module; the parameters of the last encoder of the five encoders are different from the parameters of the last encoder of the reconstruction output module.

[0063] In this embodiment, the first four encoders of the skull stripping module and the reconstruction output module are shared, and the last encoder is trained separately during the training process, using different training data, so as to obtain encoders with different parameters. When the reconstruction output module is trained, the training data used include magnetic resonance T1 modal images, magnetic resonance T2 modal images, and reference images of brain tissue without removing the skull, and when the skull stripping module is trained, the training data used include magnetic resonance T1 modal images, magnetic resonance T2 modal images, and reference images of brain tissue without removing the skull.

[0064] For the neonatal brain tissue segmentation model, this application uses the Dice coefficient (DSC) as an evaluation indicator. DSC characterizes the degree of pixel overlap between the sample label and the segmentation result predicted by the model. The formula is as follows:

[0065]

[0066] Among them, |A| and |B| represent the number of elements in the sample label and the segmentation result predicted by the model, respectively, and |A∩B| represents the number of identical elements in the sample label and the segmentation result predicted by the model.

[0067] The above formula can be written in another form:

[0068]

[0069] Among them, TP means the image is correctly segmented when the model predicts; FP means the image is classified or segmented as a positive example when predicted but is actually a negative example; FN means the image is classified or segmented as a negative example when predicted but is actually a positive example.

[0070] The present application also provides a method for segmenting a newborn brain magnetic resonance image based on deep learning, the method comprising:

[0071] The magnetic resonance images to be segmented include magnetic resonance T1 modality images and magnetic resonance T2 modality images;

[0072] Inputting each first slice of the magnetic resonance T1 modality image and each second slice of the magnetic resonance T2 modality image into a newborn brain nuclear magnetic resonance image segmentation model to obtain a final image segmentation result corresponding to the first slice and the second slice;

[0073] For all first slices and all second slices, all final image segmentation results are spliced ​​to obtain a final image segmentation result corresponding to the magnetic resonance image to be segmented;

[0074] The neonatal brain MRI image segmentation model is the neonatal brain MRI image segmentation model according to the above embodiment.

[0075] In one embodiment, the method further comprises:

[0076] The newborn brain MRI image segmentation model was trained. When training the reconstruction output module, the training data used included MRI T1 modality images, MRI T2 modality images, and brain tissue benchmark images without skull removal. When training the skull stripping module, the training data used included MRI T1 modality images, MRI T2 modality images, and brain tissue benchmark images without skull removal.

[0077] During the model training process, the Dice loss function and optimizer are used to monitor and automatically update the training of the network model. The calculation formula of the Dice loss function d is as follows:

[0078]

[0079] The value range of d is [0,1]. The closer it is to 1, the closer the predicted value is to the true value. If the predicted result is exactly the same as the sample label, d takes the value of 1. If the predicted result and the sample label do not have the same elements, d takes the value of 0.

[0080] The above are only various implementations of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A newborn brain magnetic resonance image segmentation model, characterized in that: include: A parallel information encoding module and a feature decoding module; the parallel information encoding module includes a long sequence information encoder, a convolution encoder and a feature fusion module, and the feature decoding module includes a reconstruction output module and a skull stripping module; The magnetic resonance image to be segmented includes a magnetic resonance T1 modality image and a magnetic resonance T2 modality image, and for each first slice of the magnetic resonance T1 modality image and each second slice of the magnetic resonance T2 modality image, the convolution encoder is used to obtain a first hierarchical feature corresponding to the first slice and a second hierarchical feature corresponding to the second slice; The long sequence information encoder is used to obtain a third hierarchical feature corresponding to the first slice and a fourth hierarchical feature corresponding to the second slice; The feature fusion module is used to fuse the first hierarchical feature and the second hierarchical feature to obtain a convolutional coded fused feature; to fuse the third hierarchical feature and the fourth hierarchical feature to obtain an information coded fused feature; and to concatenate the convolutional coded fused feature and the information coded fused feature to obtain a final fused feature; The reconstruction output module is used to obtain a preliminary image segmentation result according to the final fused features; The skull stripping module is used to obtain the image segmentation result without the skull according to the final fused features; Obtaining final image segmentation results corresponding to the first slice and the second slice according to the preliminary image segmentation result and the image segmentation result after the skull is removed; For all the first slices and all the second slices, all the final image segmentation results are spliced ​​to obtain the final image segmentation result corresponding to the nuclear magnetic resonance image to be segmented; The long sequence information encoder includes two information encoder structures, which are respectively used to obtain the third hierarchical features corresponding to the first slice and the fourth hierarchical features corresponding to the second slice; Each of the information encoder structures comprises four information encoder units connected in sequence, wherein the first information encoder unit comprises a linear embedding layer and a Swin Transformer block, and each of the remaining three information encoder units comprises an image block fusion layer and a Swin Transformer block; Each of the information encoder units outputs a hierarchical feature, and the four hierarchical features output by the four information encoder units constitute the third hierarchical feature or the fourth hierarchical feature.

2. The model according to claim 1, characterized in that The convolution encoder includes two convolution encoder structures, respectively used to obtain a first hierarchical feature corresponding to the first slice and a second hierarchical feature corresponding to the second slice; Each of the convolutional encoder structures includes four convolutional encoder units connected in sequence, and each convolutional encoder unit includes a dense connection module, a maximum pooling layer, and a CoT attention module connected in sequence; Each of the convolutional encoder units outputs a hierarchical feature, and the four hierarchical features output by the four convolutional encoder units constitute the first hierarchical feature or the second hierarchical feature.

3. The model according to claim 1, characterized in that The reconstruction output module includes five decoders connected in sequence, each of which includes an anti-pooling layer, a splicing layer and a dense connection module; The depooling layer is used to upsample the input to obtain upsampled features; The concatenation layer is used to concatenate and output the upsampled features and the shallow features in the final fused features by using a skip connection method; The dense connection module is used to perform convolution on the result of the splicing output.

4. The model according to claim 3, characterized in that The skull stripping module includes five decoders connected in sequence, the first four decoders of the five decoders are the first four decoders of the reconstruction output module; the parameters of the last decoder of the five decoders are different from the parameters of the last decoder of the reconstruction output module.

5. A method for segmenting neonatal brain magnetic resonance images based on deep learning, characterized in that: include: The magnetic resonance images to be segmented include magnetic resonance T1 modality images and magnetic resonance T2 modality images; Inputting each first slice of the magnetic resonance T1 modality image and each second slice of the magnetic resonance T2 modality image into a newborn brain nuclear magnetic resonance image segmentation model to obtain final image segmentation results corresponding to the first slice and the second slice; For all the first slices and all the second slices, all the final image segmentation results are spliced ​​to obtain the final image segmentation result corresponding to the nuclear magnetic resonance image to be segmented; The neonatal brain MRI image segmentation model is the neonatal brain MRI image segmentation model according to any one of claims 1-4.

6. The method according to claim 5, characterized in that Also includes: The neonatal brain magnetic resonance image segmentation model is trained. When the reconstruction output module is trained, the training data used include magnetic resonance T1 modality images, magnetic resonance T2 modality images and brain tissue reference images without skull removal. When the skull stripping module is trained, the training data used include magnetic resonance T1 modality images, magnetic resonance T2 modality images and brain tissue reference images without skull removal.

Images