A method for automatic segmentation of deep gray matter structures in the brain based on T1W and QSM using deep learning combined with personalized atlas

Abstract

The present invention discloses a method for automatically segmenting deep gray matter structures (caudate nucleus, putamen, globus pallidus, substantia nigra, red nucleus, dentate nucleus) of the brain based on T1W and QSM by combining deep learning with personalized atlas, including the steps of: collecting T1W and QSM magnetic resonance data, and manually annotating deep gray matter structures of the brain by doctors; extracting feature vectors of T1W brain images; performing unsupervised clustering on the feature vectors of all T1W brain images, selecting representative images and doctor's annotations to form a personalized atlas; registering the T1W images and annotations in the personalized atlas with other subjects' T1W images to obtain preliminary segmentation results, forming training set data; constructing a deep learning network model based on the attention mechanism; and training the model to segment deep gray matter structures of the brain. The present invention uses the contrast advantages of T1W and QSM images to accurately and automatically segment deep gray matter structures of the brain, and is expected to be used in brain imaging research and auxiliary clinical diagnosis of movement disorders such as Parkinson's disease in the future.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep brain gray matter structure segmentation, and specifically relates to a method for automatically segmenting deep brain gray matter structures based on T1W and quantitative magnetic susceptibility imaging (QSM) by combining deep learning with a personalized atlas. Background Art

[0002] In the past decade, the rapid development of susceptibility-weighted imaging technology in the field of magnetic resonance imaging (MRI) has made it possible to visualize and quantify iron in the living human brain [1]. Specifically, the combination of gradient echo (GRE) sequences with post-processing techniques and modeling methods has effectively enhanced the image contrast related to magnetic susceptibility and facilitated the quantitative calculation of iron content. Conventionally acquired structural MRI images (such as T1WI) can clearly display the contours of the caudate nucleus (CN), but the display of deep gray matter structures such as the putamen (PUT), globus pallidus (GP), substantia nigra (SN), subthalamic nucleus (STN), red nucleus (RN) and dentate nucleus (DN) relies on the unique properties of quantitative susceptibility imaging (QSM) technology [2,3]. QSM reflects the distribution of magnetic susceptibility of local tissues, highlighting the contours and internal details of iron-containing tissue structures, and can clearly present structures that are difficult to distinguish on conventional T1W-weighted images. The application of QSM technology has led to increasing attention to the importance of iron in brain development, aging, and a variety of neurological diseases including Parkinson's disease, Alzheimer's disease, and multiple sclerosis. In addition, neurosurgery mainly uses stereotactic surgery to treat refractory Parkinson's disease, epilepsy, Alzheimer's disease and other neurological diseases. Deep gray matter structures such as CN, PUT, GP, STN, etc. are common target areas in surgery, and accurate positioning of the target area is the key to the success of the surgery. The brain maps embedded in existing neuroimaging analysis toolkits (such as FreeSurfer, FSL, and SPM, etc.) are limited by technological development, and most of them do not include gray matter structures such as SN and STN. At the same time, the segmentation effect of other deep gray matter structures in the brain needs to be further optimized. Comprehensively exploring image structure and contrast features to achieve accurate segmentation of deep gray matter structures in the brain has important clinical and scientific value.

[0003] At present, a series of studies related to the automatic segmentation of deep gray matter structures in the brain have been published, including semi-automatic methods [4], fully automated patch-based segmentation algorithms [5], level set methods [6], majority voting label fusion algorithms [7], Bayesian methods [8], and methods based on standard spatial template registration [9]. The main problem with these traditional segmentation methods is that the segmentation accuracy is insufficient and it is difficult to reliably guide clinical decision-making. Deep learning-related methods have shown great advantages

[10] . The convolutional neural network method for segmenting deep gray matter structures in the brain can automatically learn useful feature expressions from big data. However, due to its nested nonlinear structure, it is difficult to explain the working principle of the network.

[0004] References:

[0005] [1]Ropele S,Langkammer C.Iron quantification with susceptibility.NMRBiomed 2017;30.

[0006] [2] Bilgic B, Pfefferbaum A, Rohlfing T, Sullivan E V., Adalsteinsson E. MRI estimates of brain iron concentration in normal aging using quantitative susceptibility mapping. Neuroimage 2012;59:2625–35.

[0007] [3]Liu C, Li W, Tong KA, Yeom KW, Kuzminski S. Susceptibility-weightedimaging and quantitative susceptibility mapping in the brain. J Magn ResonImaging 2015;42:23–41.

[0008] [4] Kim J, Lenglet C, Duchin Y, Sapiro G, Harel N. Semiautomatic Segmentation of Brain Subcortical Structures from High-Field MRI. IEEE JBiomed Heal Informatics 2014;18:1678–95.

[0009] [5]Haegelen C,CoupéP,Fonov V,Guizard N,Jannin P,Morandi X,etal.Automated segmentation of basal ganglia and deep brain structures in MRIof Parkinson’s disease.Int J Comput Assist Radiol Surg 2013;8:99–110.

[0010] [6]Basukala D,Mukundan R,Melzer T,Keenan R.Segmentation of SubstantiaNigra for the Automated Characterization of Parkinson’s Disease.IEEE 3rd IntConf Image Process Appl Syst IPAS 2018 2018:85–90.

[0011] [7]Xiao Y,Jannin P,D’Albis T,Guizard N,Haegelen C,Lalys F,etal.Investigation of morphometric variability of subthalamic nucleus,rednucleus,and substantia nigra in advanced Parkinson’s disease patients usingautomatic segmentation and PCA-based analysis.Hum Brain Mapp 2014;35:4330–44.

[0012] [8]Visser E,Keuken MC,Forstmann BU,Jenkinson M.Automated segmentationof the substantia nigra,subthalamic nucleus and red nucleus in 7T data atyoung and old age.Neuroimage 2016;139:324–36.

[0013] [9] Lim IAL, Faria A V., Li

[0014]

[10] Lundervold AS, Lundervold A. An overview of deep learning in medical imaging focusing on MRI. Z Med Phys 2019;29:102–27. Summary of the invention

[0015] In view of the above-mentioned deficiencies in the prior art, the purpose of the present invention is to provide a method for automatically segmenting deep gray matter structures of the brain based on T1W and QSM by combining deep learning with a personalized atlas, aiming to solve the problem that the existing deep learning segmentation network model ignores the anatomical structure information of the tissue and improve the segmentation accuracy.

[0016] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0017] A method for automatically segmenting deep gray matter structures of the brain based on T1W and QSM by combining deep learning with a personalized atlas, the method comprising the following specific steps:

[0018] Step 1: Collect original T1W and QSM magnetic resonance imaging data, and doctors manually mark the deep gray matter structure areas of the brain based on T1W and QSM images to form a complete data set;

[0019] Step 2: Using the anatomical structure of the T1W brain image in the complete data set, the structural features of the entire brain are extracted to form the structural feature vector of the T1W image;

[0020] Step 3: Perform unsupervised clustering on the structural feature vectors of all T1W images, and iteratively divide the structural feature vectors into several categories so that the loss function corresponding to the clustering result is minimized; select the image corresponding to the structural feature vector closest to the cluster center point in each group of clusters as the representative T1W image of that class, and all representative T1W images and the corresponding deep brain gray matter structure areas manually marked by doctors form a personalized atlas;

[0021] Step 4: Align the representative T1W images and doctor-marked regions of each cluster in the personalized atlas with other T1W images in the group to obtain preliminary segmentation results, which together with the original T1W and QSM magnetic resonance data form a training dataset for deep learning;

[0022] Step 5: Build a deep learning segmentation network model based on multiple attention mechanisms;

[0023] Step 6: Input the training data obtained in step 4 into the constructed segmentation network model for learning to obtain the final segmentation result.

[0024] The acquisition of original T1W and QSM magnetic resonance imaging data described in step 1 includes: performing a brain scan on the subject using a magnetic resonance device, acquiring brain magnetic resonance data including T1W and QSM modalities, and performing image resolution normalization processing on the T1W and QSM modality data; the doctor manually marks the deep brain gray matter structure area, specifically: the doctor manually outlines the caudate nucleus, putamen, globus pallidus, substantia nigra, red nucleus and dentate nucleus areas in the deep brain gray matter on the magnetic resonance image; the magnetic resonance data and the doctor-marked areas constitute a complete data set.

[0025] The step 2 of extracting the structural features of the entire brain specifically includes: using Freesurfer software to extract the skull contour of the T1W brain magnetic resonance image, extracting the surface area to volume ratio, sphericity and spherical proportion shape features of the skull contour, and forming a structural feature vector for each T1W magnetic resonance image.

[0026] The unsupervised clustering described in step 3 is performed using the K-means algorithm; the loss function is defined as the sum of squared errors of each structural feature vector from the center point of the category to which it belongs.

[0027] The other images in the group described in step 4 refer to images in the complete data set excluding the images in the personalized atlas.

[0028] The step 5 is specifically as follows: constructing a deep learning segmentation network model based on multiple attention mechanisms, the segmentation network model includes a segmentation branch and an anatomical branch; the segmentation branch uses U-net as the backbone network model, and its input is two modal data, T1W and QSM, including encoding and decoding stages, and uses jump connections to splice the features of the encoding stage with the features of the decoding stage, and the segmentation branch uses both the spatial attention mechanism and the channel attention mechanism to quantify the features of the deep gray matter structure area of ​​the brain and suppress irrelevant background areas; the input of the anatomical branch is the preliminary segmentation result after alignment, and the anatomical attention module guides the segmentation branch to perform accurate multi-scale deep gray matter structure segmentation based on the prior knowledge of the deep brain gray matter structure learned from the preliminary segmentation; in the output stage of the network model, the scale attention mechanism will integrate features of different scales to obtain the final segmentation probability map.

[0029] The step 6 specifically includes: firstly unifying the resolution of the training set data and standardizing the grayscale value, and using the contrast-limited adaptive histogram equalization (CLAHE) algorithm to enhance the image contrast; inputting the training data into the constructed segmentation network model, and in the process of training the network model, in order to prevent the network model from overfitting, performing online data augmentation on the input data, using random rotation, stretching, translation, flipping, gamma transformation, affine and elastic transformation; after the training is completed, the final segmentation result is obtained.

[0030] Beneficial effects of the present invention:

[0031] The present invention establishes a method for automatically segmenting deep brain gray matter structures by combining deep learning with personalized atlases based on T1W and QSM modal magnetic resonance images, and constructs a personalized atlas from a complete data set by using an unsupervised clustering method of image structure features; a preliminary segmentation result of the training data set image is obtained by aligning with the data in the personalized atlas; an anatomical attention mechanism is used to extract anatomical information of the deep brain gray matter structure from the preliminary segmentation result, thereby improving the performance of the network model; at the same time, the spatial attention mechanism, the channel attention mechanism and the scale attention mechanism are used to locate the features of the region of interest while suppressing irrelevant background areas; and the purpose of accurately segmenting multiple deep brain gray matter structures such as the caudate nucleus, putamen, globus pallidus, substantia nigra, red nucleus and dentate nucleus in the left and right hemispheres of the brain is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a flow chart of the present invention;

[0033] Figure 2 It is a schematic diagram of the deep gray matter structural area of ​​the brain on QSM and T1W images;

[0034] Figure 3It is a structural diagram of the segmentation network model of the present invention;

[0035] Figure 4 It is a module structure diagram of the anatomical and spatial attention mechanism of the present invention;

[0036] Figure 5 It is a structural diagram of the channel and scale attention mechanism module of the present invention. DETAILED DESCRIPTION

[0037] The present invention is further described in detail in conjunction with the following specific embodiments and drawings. The process, conditions, experimental methods, etc. for implementing the present invention, except for the contents specifically mentioned below, are all common knowledge and common common sense in the art, and the present invention has no special restrictions. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the art without making creative work are within the scope of protection of the present invention.

[0038] See also Figure 1 , the present invention comprises the following specific steps:

[0039] S101, collect T1W and QSM magnetic resonance imaging data and doctors manually mark the deep gray matter structure areas of the brain to form a complete data set;

[0040] S102, using the anatomical structure of the T1W brain image in the complete data set, extracting the structural features of the entire brain to form a structural feature vector of the T1W image;

[0041] S103, performing unsupervised clustering on the structural feature vectors of all T1W brain images, and dividing the structural feature vectors into several categories through iteration so that the loss function corresponding to the clustering result is minimized; selecting the image corresponding to the structural feature vector closest to the cluster center point in each group of clusters as the representative T1W image of the class, and all representative T1W images and the corresponding deep gray matter structural areas of the brain manually marked by doctors form a personalized atlas;

[0042] S104, registering the representative T1W images and doctor-marked areas of each cluster in the personalized atlas with other T1W images in the group to obtain preliminary segmentation results, which together with the T1W and QSM magnetic resonance data constitute a training data set for deep learning;

[0043] S105. Construct a deep learning segmentation network model based on multiple attention mechanisms;

[0044] S106: Input the training data obtained in S104 into the network model constructed in S105 for learning to obtain the final segmentation result.

[0045] In order to optimize the above technical solutions, the specific measures taken also include:

[0046] Furthermore, the acquisition of T1W and QSM magnetic resonance imaging data in step S101 specifically includes: performing a brain scan on the subject using a magnetic resonance device, acquiring brain magnetic resonance data including T1W and QSM modalities, and performing image resolution normalization processing on the T1W and QSM modality data;

[0047] Furthermore, the doctor manually marks the deep brain gray matter structure area in step S101, specifically including: the doctor manually outlines the caudate nucleus, putamen, globus pallidus, substantia nigra, red nucleus and dentate nucleus areas in the deep brain gray matter on the magnetic resonance image; the image data and the doctor-marked area constitute a complete data set.

[0048] Furthermore, in step S102, the structural features of the entire brain are extracted, specifically including: using Freesurfer software to extract the skull contour of the T1W brain magnetic resonance image, extracting the surface area to volume ratio, sphericity and spherical proportion shape features of the skull contour, and forming a structural feature vector for each T1W magnetic resonance image.

[0049] Furthermore, in step S103, a personalized atlas is constructed, which specifically includes: using the K-means algorithm to perform unsupervised clustering on the structural feature vectors of all T1W images, and dividing the structural feature vectors into several categories (such as twenty groups) through iteration, so that the loss function corresponding to the clustering result is minimized, that is, the sum of squared errors of each structural feature vector from the center point of the category to which it belongs is minimized; the image corresponding to the structural feature vector closest to the center point of each cluster is selected as the representative image of the class, and all representative T1W images and the corresponding deep brain gray matter structure areas manually marked by doctors constitute a personalized atlas.

[0050] Furthermore, in step S104, the representative T1W images and doctor-marked regions of each cluster in the personalized atlas are registered with other T1W images in the group to obtain preliminary segmentation results, which together with the T1W and QSM magnetic resonance data constitute a training data set for deep learning;

[0051] Furthermore, in step S105, the QSM and T1W data are spliced ​​in the channel dimension, and are simultaneously input into the deep learning segmentation network model together with the preliminary annotation results after registration.

[0052] Furthermore, in step S105, the constructed deep learning segmentation network model is an end-to-end trainable network model, including a segmentation branch and an anatomical branch.

[0053] Further, in step S105, the segmentation branch of the deep learning segmentation network model uses U-net as the backbone network model, and its input is two modal data, T1W and QSM, including encoding and decoding stages, and uses jump connections to splice the features of the encoding stage with the features of the decoding stage. The encoding stage compresses the information of the image in the feature map, extracts the representative feature expression on the image, includes three downsampling, and the number of feature channels gradually increases, and the resolution of the feature map gradually decreases. In the decoding stage, the resolution of the features compressed in the encoding stage is gradually restored to the same resolution as the original image size through three deconvolution operations, and the number of feature channels gradually decreases during the process.

[0054] Furthermore, in step S105, the anatomical branch in the deep learning segmentation network model, whose input is the preliminary segmentation result of the training set data, can automatically learn and acquire the prior knowledge of the deep brain gray matter structure through the anatomical attention module, and guide the segmentation branch to perform accurate multi-scale deep brain gray matter structure segmentation.

[0055] Furthermore, in step S105, the channel and spatial attention mechanisms are introduced in the decoding stage of the constructed deep learning segmentation network model. Most of the feature channels in the encoding stage contain low-level information, and their corresponding decoders contain more semantic information. Therefore, the importance of features on different channels in the encoding and decoding stages of the segmentation task is different. In order to better utilize the most useful feature channels, the channel attention mechanism can automatically highlight relevant feature channels while suppressing irrelevant channels. The spatial attention mechanism can strengthen the region of interest on the feature map while suppressing the background or irrelevant parts.

[0056] Furthermore, in step S105, a scale attention mechanism is introduced in the output stage of the constructed deep learning segmentation network model. In order to segment the deep brain gray matter structures of different scales, the features of different scales in the decoding process are upsampled and combined on the channel to jointly predict multiple deep brain gray matter structure areas. The scale attention module can automatically learn the image-level weights for each scale of the deep brain gray matter structure and correct the features of different scales at the end of the network model.

[0057] Furthermore, in the process of training the network model in step S106, the training set data needs to be unified in resolution and standardized in grayscale value before being input into the network model, and the image contrast is enhanced using the Contrast-Limited Adaptive Histogram Equalization (CLAHE) algorithm. In order to prevent the network model from overfitting, the input data is augmented in real time, including random rotation, stretching, translation, flipping, gamma transformation, radiation and elastic transformation. The focalloss function is used as the loss function, and the Adam algorithm is used as the optimization algorithm.

[0058] Example

[0059] See also Figure 1 , this embodiment specifically includes:

[0060] Step S101, collecting T1W and QSM magnetic resonance imaging data and manually marking the deep gray matter structure area of ​​the brain by a doctor, specifically including using a magnetic resonance device to scan the brain of the subject, collecting brain magnetic resonance data including T1W and QSM modalities, and normalizing the image resolution of the T1W and QSM modal data; the doctor manually marks the caudate nucleus, putamen, globus pallidus, substantia nigra, red nucleus and dentate nucleus in the deep gray matter of the brain on the magnetic resonance image Figure 2 The image data and the doctor-marked area constitute a complete data set.

[0061] Step S102, extracting the structural features of the T1W brain image specifically includes extracting the skull contour of the T1W brain magnetic resonance image in the complete data set using Freesurfer software, extracting the surface area to volume ratio, sphericity and spherical proportion shape features of the skull contour, and forming a structural feature vector for each T1W magnetic resonance image.

[0062] Step S103, constructing a personalized atlas specifically includes: using the K-means algorithm to perform unsupervised clustering on the structural feature vectors of all T1W images, and dividing the structural feature vectors into several categories (such as twenty groups) through iteration, so that the loss function corresponding to the clustering result is minimized, that is, the sum of squared errors of each structural feature vector from the center point of the category to which it belongs is minimized; selecting the image corresponding to the structural feature vector closest to the center point of each cluster in each group of clusters as the representative image of the class, and all representative T1W images and the corresponding deep brain gray matter structure areas manually marked by doctors constitute a personalized atlas.

[0063] Step S104, obtaining a training data set for deep learning specifically includes aligning the representative T1W images in each cluster and the deep brain gray matter structure areas annotated by the doctor with other T1W images in the group to obtain preliminary segmentation results, which together with the T1W and QSM magnetic resonance data constitute a training data set for deep learning.

[0064] Furthermore, in step S104, the training data set includes T1W and QSM MRI data, as well as the preliminary segmentation results after data registration and the deep gray matter structure areas of the brain manually annotated by doctors. Before the data is input into the deep learning network model, the resolution of the data needs to be unified and the grayscale value needs to be standardized: first, the grayscale value of the Z-score of the image is standardized, that is, the image mean is subtracted and the variance is divided; then, all data are cut into 160×160×64 pixels in the center. 3A three-dimensional data block of this size can contain all the deep gray matter structural areas of the brain and discard irrelevant background areas. In the process of training the network model, in order to prevent the network model from overfitting, the input data is augmented in real time, including random rotation, stretching, translation, flipping, gamma transformation, affine and elastic transformation.

[0065] Step S105: Construct a deep learning segmentation network model based on multiple attention mechanisms. Figure 3 The network model includes a segmentation branch and an anatomical branch. The magnetic resonance data of T1W and QSM modalities are input into the segmentation branch of the network model, and the corresponding preliminary segmentation areas are input into the anatomical branch of the network model.

[0066] Further, in step S105, CBR represents the convolution module, including the convolution layer (the convolution kernel size is 3×3×3, the step size is 1), the excitation layer (ReLU function) and the batch normalization layer (Batch Normalization). MP represents the maximum pooling layer (the pooling area size is 2×2×2, the step size is 2) and DC represents the deconvolution layer (the convolution kernel size is 3×3×3, the step size is 2). AG represents the anatomical attention module, SA represents the spatial attention mechanism module, CA represents the channel attention mechanism module and LA represents the scale attention mechanism module.

[0067] Furthermore, in step S105, the two modal data of T1W and QSM are merged on the channel and simultaneously input into the segmentation branch of the network model. The segmentation branch uses U-net as the backbone network model, includes encoding and decoding stages, and uses jump connections to splice the features of the encoding stage with the features of the decoding stage. In the encoding stage, the information of the image is compressed in the feature map, and the representative feature expression on the image is extracted through three down-sampling modules (two layers of CBR and one maximum pooling layer), and the number of feature channels gradually increases, and the resolution of the feature map gradually decreases; in the decoding stage, the resolution of the features compressed in the encoding stage is gradually restored to the same resolution as the original image size through three up-sampling modules (two layers of CBR and one deconvolution layer), and the number of feature channels gradually decreases in the process.

[0068] Furthermore, in step S105, the preliminary segmentation results in the training data set are input into the anatomical branch of the network model. The anatomical branch only contains the encoding stage, and the prior knowledge of the deep brain gray matter structure is automatically learned through the anatomical attention module, and guides the segmentation branch to perform accurate multi-scale deep brain gray matter structure segmentation.

[0069] Further, in step S105, the structure of the anatomical attention module is shown in Figure 4Its input is the features of the segmentation branch encoding stage and the features extracted from the preliminary segmentation results. They are spliced ​​on the channel and passed through two parallel convolution modules and sigmoid functions with a convolution kernel size of 1×1×1 to generate a weight tensor for each input feature map. The weight tensor is then multiplied with the input features to obtain two weighted feature maps. Finally, the two feature maps are added point by point and fused to output a weighted and feature map guided by anatomical information.

[0070] Furthermore, in step S105, the features of the decoding stage of the segmentation branch will be input into the spatial attention mechanism module, whose structure is shown in Figure 4 The input features enter the global maximum pooling layer and the global average pooling layer in parallel, and the obtained features are spliced ​​in the channel dimension; after the spliced ​​features are input into the CBR layer and the sigmoid function, the number of channels is reduced to 1, and the weight value of each pixel in the space is obtained; finally, the spatial weight is multiplied by the original input feature to obtain the final spatial weighted feature map. The formula of its sigmoid function is:

[0071]

[0072] Among them, x represents the eigenvalue on the feature map.

[0073] Further, in step S105, the features of the spatial attention mechanism module are input into the channel attention mechanism module after passing through a CBR module. Figure 5 The input features are passed through the CBR module and the global maximum pooling layer to compress the features of each channel in spatial dimensions. The compressed features have a global field of view, and the number of channels of the features remains unchanged. It is followed by two fully connected layers and ReLU and sigmoid functions to reduce the dimension of the features and then increase the dimension to obtain the weight corresponding to each channel. The channel weight is multiplied by the original input feature to obtain the final channel-weighted feature map.

[0074] Furthermore, in step S105, the three scales of the channel in the decoding stage (40×40×16 pixels) are 3 , 80×80×32 pixels 3 , 160×160×64 pixels 3 ) feature map with the same size as the original input image (160×160×64 pixels) 3 ) feature map and concatenate them on the channel as the input of the scale attention mechanism module. Its structure is shown in Figure 5The spliced ​​feature maps are input into the global maximum pooling and global average pooling layers in parallel, and then pass through the fully connected layer and the sigmoid function to obtain the weight corresponding to each scale; the input feature is multiplied by the scale weight to obtain the scale-weighted feature map; the scale-weighted feature map passes through two CBR layers and the sigmoid function to generate a weight map based on each pixel, the feature map generated by the first layer of CBR is multiplied by the generated weight map, and added to the scale-weighted feature map to obtain a feature map with both spatial weighting and scale weighting; finally, the weighted feature map is added to the input feature to obtain the output feature map based on the residual.

[0075] Furthermore, in step S105, the feature map output by the scale attention mechanism module is output through the CBR module to output the feature map of six channels, and then the final segmentation probability map is obtained through the sigmoid function. The probability maps of the seven channels represent the segmentation results of the caudate nucleus, putamen, globus pallidus, substantia nigra, red nucleus, dentate nucleus and background respectively.

[0076] In step S106, the training data is input into the constructed segmentation network model for learning to obtain the final segmentation result. The optimizer used in the training process is the Adam algorithm, and its initial rate is set to 10 -4 , the input batch size is 16, the loss function is focal loss, and its formula is:

[0077] Focal Loss(p,p t )=-(α(1-p t ) γ plog(p t )+(1-α)p t γ (1-p)log(1-p t ))

[0078] Among them, p and p t are the category prediction probability and the labeled true category respectively. α and γ are hyperparameters, which are set to 0.25 and 2 in this embodiment. This loss function can be trained for imbalanced category samples.

[0079] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, it should be understood by those skilled in the art that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features therein may be replaced by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.

Claims

1. A method for automatic segmentation of deep gray matter structures of the brain based on T1W and QSM by combining deep learning with personalized atlas, characterized in that: The method comprises the following specific steps: Step 1: Collect original T1W and QSM magnetic resonance imaging data, and doctors manually mark the deep gray matter structure areas of the brain based on T1W and QSM images to form a complete data set; Step 2: Using the anatomical structure of the T1W brain image in the complete data set, the structural features of the entire brain are extracted to form the structural feature vector of the T1W image; Step 3: Perform unsupervised clustering on the structural feature vectors of all T1W images, and iteratively divide the structural feature vectors into several categories so that the loss function corresponding to the clustering result is minimized; select the image corresponding to the structural feature vector closest to the cluster center point in each group of clusters as the representative T1W image of that class, and all representative T1W images and the corresponding deep brain gray matter structure areas manually marked by doctors form a personalized atlas; Step 4: Align the representative T1W images and doctor-marked regions of each cluster in the personalized atlas with other T1W images in the group to obtain preliminary segmentation results, which together with the original T1W and QSM magnetic resonance data form a training dataset for deep learning; Step 5: Build a deep learning segmentation network model based on multiple attention mechanisms; Step 6: Input the training data obtained in step 4 into the constructed segmentation network model for learning to obtain the final segmentation result; where: The step 5 is specifically as follows: constructing a deep learning segmentation network model based on multiple attention mechanisms, wherein the segmentation network model includes a segmentation branch and an anatomical branch; the segmentation branch uses U-net as a backbone network model, and its input is two modal data, T1W and QSM, including encoding and decoding stages, and uses jump connections to splice the features of the encoding stage with the features of the decoding stage. The segmentation branch uses both the spatial attention mechanism and the channel attention mechanism to increase the features of the deep gray matter structure area of ​​the brain and suppress irrelevant background areas; the input of the anatomical branch is the preliminary segmentation result after registration, and the anatomical attention module guides the segmentation branch to perform accurate multi-scale deep gray matter structure segmentation by using the prior knowledge of the deep gray matter structure of the brain obtained from the preliminary segmentation learning; In the output stage of the network model, the scale attention mechanism will integrate features of different scales to obtain the final segmentation probability map.

2. According to claim 1, a method for automatic segmentation of deep gray matter structures of the brain based on T1W and QSM by combining deep learning with personalized atlas, characterized in that: The acquisition of original T1W and QSM magnetic resonance imaging data described in step 1 includes: performing a brain scan on the subject using a magnetic resonance device, acquiring brain magnetic resonance data including T1W and QSM modalities, and performing image resolution normalization processing on the T1W and QSM modality data; the doctor manually marks the deep brain gray matter structure area, specifically: the doctor manually outlines the caudate nucleus, putamen, globus pallidus, substantia nigra, red nucleus and dentate nucleus areas in the deep brain gray matter on the magnetic resonance image; the magnetic resonance data and the doctor-marked areas constitute a complete data set.

3. According to claim 1, a method for automatic segmentation of deep gray matter structures of the brain based on T1W and QSM by combining deep learning with personalized atlas, characterized in that: The step 2 of extracting the structural features of the entire brain specifically includes: using Freesurfer software to extract the skull contour of the T1W brain magnetic resonance image, extracting the surface area to volume ratio, sphericity and spherical proportion shape features of the skull contour, and forming a structural feature vector for each T1W magnetic resonance image.

4. According to claim 1, a method for automatic segmentation of deep gray matter structures of the brain based on T1W and QSM by combining deep learning with personalized atlas, characterized in that: The unsupervised clustering described in step 3 is performed using the K-means algorithm; the loss function is defined as the sum of squared errors of each structural feature vector from the center point of the category to which it belongs.

5. According to claim 1, a method for automatic segmentation of deep gray matter structures of the brain based on T1W and QSM by combining deep learning with personalized atlas, characterized in that: The other images in the group described in step 4 refer to images in the complete data set excluding the images in the personalized atlas.

6. According to claim 1, a method for automatic segmentation of deep gray matter structures of the brain based on T1W and QSM by combining deep learning with personalized atlas, characterized in that: The step 6 specifically includes: firstly unifying the resolution of the training set data and standardizing the grayscale value, and using the contrast-limited adaptive histogram equalization algorithm to enhance the image contrast; inputting the training data into the constructed segmentation network model, and in the process of training the network model, in order to prevent the network model from overfitting, performing online data augmentation on the input data, using random rotation, stretching, translation, flipping, gamma transformation, affine and elastic transformation methods; after the training is completed, the final segmentation result is obtained.

Images