Method for training brain tissue segmentation model using single atlas image and application thereof

Abstract

This invention discloses a method for training a brain tissue segmentation model using a single atlas image and its application, belonging to the field of medical image segmentation. The method includes: when training a segmentation network based on a registration-segmentation dual-model iterative learning, using the registration network to predict the atlas image I. a and unlabeled image I u The offset field φ between them, and the spectral image I according to φ. a and its label S a Perform deformation operations separately to obtain deformation atlas images and deformation atlas labels. Calculate the registration confidence map C, which is used to represent and I. u Alignment at each pixel; according to C-alignment and I-alignment. u Regions with varying degrees of alignment were individually subjected to IST style transfer and then stitched together to obtain a style-transferred image. Regions with higher alignment levels exhibited higher style transfer coefficients. This was used as supervisory information, with I... u Using the sum as input, the segmentation network is trained, with the training loss for the sum being a confidence-guided loss. This invention can improve the performance of brain tissue segmentation when only a single atlas image exists.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of medical image segmentation, and more particularly relates to a method for training a brain tissue segmentation model using a single atlas image and applications thereof. BACKGROUND

[0002] Brain tissue segmentation refers to dividing a brain image into different functional regions, which has important guiding value in clinical surgery. Using the labels of brain tissue segmentation, doctors can more intuitively understand the brain structure, so as to more accurately achieve surgical planning, disease diagnosis and treatment.

[0003] Traditional segmentation algorithms realize the aggregation of image elements of the same category through morphological operations or similarity measures. Such methods are mostly based on iterative optimization, which is usually time-consuming, and some also need to provide good artificial initialization (such as region growing algorithm usually needs to provide good initialization seed points) to obtain better segmentation results. These defects greatly limit their application in clinical practice.

[0004] In recent years, deep learning algorithms have become the mainstream method in medical image segmentation due to their good performance and fast inference time. For segmentation networks, full-supervised learning requires labels for all images, which can usually achieve good segmentation accuracy. However, due to the fine and complex structure of the human brain, manually outlining brain tissue is usually time-consuming and prone to error, and lacks quantitative standards. Therefore, it has important practical value to achieve accurate brain tissue segmentation under the condition of a small number of labels, especially a single label.

[0005] In order to realize brain tissue segmentation under a single label, some studies have proposed methods based on atlas registration. Such methods first define or synthesize a labeled template image (atlas), then register the labeled atlas to the unlabeled image, and deform the atlas label accordingly using the pixel-level mapping relationship (offset field) generated by registration, thereby predicting the label of the image to be segmented. However, due to the lack of label guidance, such registration methods are difficult to align the fine and complex structure of brain tissue, and the segmentation accuracy is usually low, which is difficult to meet the clinical needs.

[0006] A more effective method is to use a registration-segmentation dual model iterative learning. In the iterative process, registration acts as a data augmenter, deforming the atlas image and label to generate data for segmentation training; the segmentation network predicts a more refined pseudo-label, further optimizing the registration. However, the error generated by the registration network will affect the subsequent segmentation learning, which to some extent limits the performance of segmentation. SUMMARY

[0007] To address the shortcomings and improvement needs of existing technologies, this invention provides a method and its application for training a brain tissue segmentation model using a single atlas image. The purpose is to optimize the training of the segmentation network based on the registration error of the perceived registration network during the registration-segmentation dual-model iterative learning process, thereby improving the performance of brain tissue segmentation when only a single atlas image exists.

[0008] To achieve the above objectives, according to one aspect of the present invention, a method for training a brain tissue segmentation model using a single atlas image is provided, comprising:

[0009] The semi-supervised training steps for a segmentation network include:

[0010] (S1) Predicting spectral images using a registration network I a and unlabeled images I u offset fields between and according to the offset field For spectral images I a and its labels S a Perform deformation operations separately to obtain deformation spectrum images. Deformation map tags The atlas images are brain tissue images with segmentation labels, while the unlabeled images are brain tissue images without segmentation labels.

[0011] (S2) Calculate the registration confidence plot C Used to represent deformation spectrum images and unlabeled images I u Alignment level at each pixel;

[0012] (S3) According to the registration confidence map C, the deformation spectrum image is... and unlabeled images I u The image is divided into multiple regions with different degrees of alignment for the deformation spectrum. Each region undergoes a style transfer based on the aligned image, and then the images are stitched together to obtain a style-transferred image. When performing style transfer, the higher the alignment of the area, the higher the transfer coefficient.

[0013] (S4) Label the deformation map As supervisory information, unlabeled images I u and style transfer images The segmentation network is trained using the input.

[0014] The steps for weakly supervised training of a registered network include:

[0015] (W1) Predicting unlabeled images using a segmentation network I u The segmentation results are used as pseudo-labels ;

[0016] (W2) Predicting spectral images using a registration network I a and unlabeled images I u offset fields between , with spectral images I a tags S a According to the offset field Deformation map labels obtained after performing deformation operations Compared to pseudo tags The error is used as the training loss to train the registration network;

[0017] The dual-model iterative learning process includes repeatedly and alternately performing semi-supervised training steps for the segmentation network and weakly supervised training steps for the registration network until the segmentation network and the registration network converge.

[0018] Further, in step (S3), the registration confidence map is calculated. C ,include:

[0019] Image I a and unlabeled images I u Mirror the images respectively and ;

[0020] Predicting images using registration networks and offset field between ;

[0021] according to Calculate the composite migration field ;

[0022] according to Calculate the registration error E ;

[0023] according to Calculate the registration confidence plot C ;

[0024] in, Indicates deformation operation. This represents the equivalent offset field for the mirror operation. Represents the L2 norm, This represents the standard deviation of the registration error.

[0025] Furthermore, in step (S4), when training the segmentation network, for style-transfer images... Input it into the segmentation network to obtain the prediction result. Afterwards, the segmentation loss The calculation method is as follows:

[0026]

[0027] For unlabeled images I u Input it into the segmentation network to obtain the segmentation result. Afterwards, the segmentation loss The calculation method is as follows:

[0028]

[0029] in, Dice This represents the Dice coefficient.

[0030] Furthermore, in step (W2), when training the registration network, the training loss also includes: similarity loss. and regularization terms The expressions for the two are as follows:

[0031]

[0032]

[0033] in, NLCC This represents the locally normalized correlation coefficient; Representing spectral images I a According to the offset field The deformed spectral image; This represents the gradient operator.

[0034] Furthermore, in the dual-model iterative learning step, before repeatedly and alternately executing the semi-supervised training step of the segmentation network and the weakly supervised training step of the registration network, the following step is also included: performing an unsupervised registration initialization step; the unsupervised registration initialization step includes:

[0035] Predicting spectral images using registration networks I a and unlabeled images I u offset fields between loss based on similarity and regularization terms The registration network is trained using this training loss.

[0036] Furthermore, in the dual-model iterative learning step, before performing the unsupervised registration initialization step, the process further includes: performing a preprocessing operation; the preprocessing operation includes at least one of the following operations:

[0037] Perform bias field correction on all images;

[0038] Resample all images to the same size;

[0039] Perform an affine transformation on all images to the atlas image.

[0040] According to another aspect of the present invention, a method for segmenting brain tissue is provided, comprising:

[0041] The brain tissue image to be segmented is input into the segmentation network trained by the method of training a brain tissue segmentation model using a single atlas image provided by the present invention, and the brain tissue segmentation result is obtained.

[0042] According to another aspect of the present invention, a method for segmenting brain tissue is provided, comprising:

[0043] The atlas image and the brain tissue image to be segmented are input into the registration network trained by the method described above for training a brain tissue segmentation model using a single atlas image, as provided by this invention, to obtain the offset field between the atlas image and the brain tissue image to be segmented. The atlas images are brain tissue images with segmentation labels.

[0044] According to the offset field The labels of the atlas image are deformed to obtain deformed labels, which are then used as the brain tissue segmentation results of the brain tissue image to be segmented.

[0045] According to another aspect of the present invention, a computer-readable storage medium is provided, including a stored computer program; when the computer program is executed by a processor, it controls the device where the computer-readable storage medium is located to execute the method for training a brain tissue segmentation model using a single atlas image provided by the present invention, and / or the brain tissue segmentation method provided by the present invention.

[0046] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects:

[0047] (1) When training a brain tissue segmentation network based on a registration-segmentation dual-model iterative learning framework, this invention estimates a registration confidence map to reflect the alignment degree at each pixel location, thereby achieving registration error perception; in the training scenario of a single atlas image, the deformed atlas image All are derived from the same atlas image. Ia Deformation generation, with relatively uniform grayscale style, this invention is based on the registration error perception result, and transforms the deformation spectrum image. By dividing the image into multiple regions with varying degrees of alignment and setting appropriate style transfer coefficients for each region, a weighted style transfer is achieved. Based on this style transfer method, not only is the spatial correspondence between the atlas image and label pairs after registration deformation preserved, but also diverse and high-fidelity training data is generated. This effectively reduces the impact of errors generated by the registration network on segmentation learning and effectively improves the accuracy of brain tissue segmentation even with only a single atlas image.

[0048] (2) In the preferred embodiment of the present invention, the mirror error is calculated by calculating the equivalent offset field and the mirror error is used as the registration error. It is converted into the registration confidence by the Gaussian function. This calculation method can accurately perceive the registration error and accurately calculate the registration confidence map.

[0049] (3) Unlabeled images There is a certain spatial misalignment between the unlabeled image and the deformation map label. In a preferred embodiment of the present invention, the unlabeled image-deformation map label is used to... When training the segmentation network, according to By calculating the segmentation loss, the segmentation network can focus more on learning the image-label alignment region when there is spatial misalignment between the labels in unlabeled images and deformed atlases, thereby further improving the segmentation accuracy.

[0050] (4) In a preferred embodiment of the present invention, the segmentation result of the segmentation network for the unlabeled image is used as a pseudo-label, and a corresponding training loss term is designed to constrain the registration deformation map label. With pseudo-tags The consistency of the model enables weakly supervised training of the registration network. In addition, its training loss includes similarity loss and regularization term. The former constrains the similarity between the registered and deformed atlas image and the unlabeled image, while the latter constrains the smoothness of the registration offset field. This enables the registration network to have a stronger brain tissue alignment ability.

[0051] (5) In a preferred embodiment of the present invention, before performing dual-model iterative learning, the registration network is trained based on similarity loss and regularization terms used to constrain the smoothness of the registration offset field, and iterative initialization is achieved using an unsupervised registration network.

[0052] (6) In a preferred embodiment of the present invention, at the initial time, preprocessing operations such as bias field correction, resampling, and affine coarse registration are performed, thereby reducing the learning burden of the subsequent registration network. Attached Figure Description

[0053] Figure 1 A flowchart illustrating a method for training a brain tissue segmentation model using a single atlas image, as provided in an embodiment of the present invention.

[0054] Figure 2 This is a schematic diagram of the registration-segmentation dual-model iterative learning framework provided in an embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram of the registration error calculation process provided in an embodiment of the present invention;

[0056] Figure 4 This is a schematic diagram of weighted style transfer provided in an embodiment of the present invention;

[0057] Figure 5 This is a comparison diagram of the conversion results of the weighted style transfer method and the image-aligned conversion method provided in the embodiments of the present invention;

[0058] Figure 6 Segmentation model prediction results for different registration and segmentation iteration methods;

[0059] Figure 7 The curves showing the changes in registration accuracy and segmentation accuracy of the method of the present invention during the iteration process under different experimental conditions set for weighted style transfer and confidence-guided loss are shown. Among them, (a) compares the registration accuracy under different conditions, and (b) compares the segmentation accuracy under different conditions. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0061] In this invention, the terms "first," "second," etc. (if present) in the invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0062] To address the problem that existing methods for training brain tissue segmentation models based on a registration-segmentation dual-model iterative learning framework suffer from low segmentation accuracy due to registration network errors, this invention provides a method and its application for training a brain tissue segmentation model using a single atlas image. The overall idea is to achieve registration error awareness and integrate the perceived registration error into the dual-model learning framework, effectively eliminating the negative impact of registration error on segmentation learning and significantly improving segmentation accuracy.

[0063] The following is an example.

[0064] Example 1:

[0065] A method for training a brain tissue segmentation model using a single atlas image, the process of which is as follows: Figure 1 As shown in the figure. The method for training a brain tissue segmentation model using a single atlas image provided in this embodiment employs a registration-segmentation dual-model iterative learning framework. Optionally, in this embodiment, the registration network uses an existing advanced model, PC-Net, which has strong brain tissue registration capabilities; the segmentation network uses Unet, and a deep supervision strategy is introduced to improve network performance. The framework based on the registration-segmentation dual-model iterative learning is as follows: Figure 2 As shown. When training the segmentation network, the registration network is fixed; when training the registration network, the segmentation network is fixed.

[0066] This embodiment requires only a single labeled atlas image and multiple unlabeled images to complete training. After training, the trained segmentation network or registration network can be used to achieve high-precision segmentation of brain tissue images. It is easy to understand that in this embodiment, the atlas image is a brain tissue image with segmentation labels, and the unlabeled images are brain tissue images without segmentation labels.

[0067] In this embodiment, preprocessing operations were performed on the atlas images and all unlabeled images at the initial moment. Specifically, considering that intensity differences in the same tissue images due to imaging variations may affect subsequent segmentation, N4 bias field correction was first performed on all brain images. After correction, all image data were resampled to 1mm×1mm×1mm, and then the ANTS toolkit was used to perform affine transformations on all image data to the atlas to achieve coarse registration, ensuring that all images are in the same spatial coordinate system, and also reducing the learning burden of the subsequent registration network to some extent.

[0068] In this embodiment, before iterative training of the registration-segmentation network, an unsupervised registration initialization step is first performed, including:

[0069] Predicting spectral images using registration networks I a and unlabeled images I u offset fields between loss based on similarity and regularization terms The registration network is trained using the training loss.

[0070] Among them, similarity loss Constraint map image I a According to the offset field Deformation spectral images and unlabeled images Similarity, regularization terms Constrained registration migration field The smoothness of the surface. The calculation formula is:

[0071]

[0072]

[0073] Let || denote the gradient operator, and || denote the L2 norm calculation; This is the locally normalized correlation coefficient, calculated using the following formula:

[0074]

[0075] in It is a 9×9×9 image block. and Represent and Mean gray level within an image patch.

[0076] After the above preprocessing operations and unsupervised registration initialization, this embodiment will execute a dual-model iterative learning step, including: repeatedly and alternately executing the semi-supervised training step of the segmentation network and the weakly supervised training step of the registration network until the segmentation network and the registration network converge. Figure 2 As shown, the semi-supervised training steps for the segmentation network include:

[0077] (S1) Predicting spectral images using a registration network I a and unlabeled images I u offset fields between and according to the offset field For spectral images I a and its labels S a Perform deformation operations separately to obtain deformation spectrum images. Deformation map tags The atlas images are brain tissue images with segmentation labels, while the unlabeled images are brain tissue images without segmentation labels.

[0078] (S2) Calculate the registration confidence plot C Used to represent deformation spectrum images and unlabeled images I u Alignment level at each pixel;

[0079] (S3) According to the registration confidence plot CDeformation map image and unlabeled images I u The image is divided into multiple regions with different degrees of alignment for the deformation spectrum. Each region undergoes a style transfer based on the aligned image, and then the images are stitched together to obtain a style-transferred image. When performing style transfer, the higher the alignment of the area, the higher the transfer coefficient.

[0080] (S4) Label the map As supervisory information, unlabeled images I u and style transfer images The segmentation network is trained using the input.

[0081] like Figure 2 As shown, the weakly supervised training steps for the registration network include:

[0082] (W1) Predicting unlabeled images using a segmentation network I u The segmentation results are used as pseudo-labels ;

[0083] (W2) Predicting spectral images using a registration network I a and unlabeled images I u offset fields between , with spectral images I a tags S a According to the offset field Deformation map labels obtained after performing deformation operations Compared to pseudo tags The error is used as the training loss to train the registration network.

[0084] In this embodiment, when training the brain tissue segmentation network based on the registration-segmentation dual-model iterative learning framework, a registration confidence map is estimated to reflect the alignment degree at each pixel position, thereby realizing the perception of registration error. In order to accurately realize the perception of registration error, this embodiment proposes a method for calculating registration error based on mirror error, that is, it is assumed that if the input is a mirror image pair to be registered, its output is also a mirror image when there is no error, and the mirror difference between the outputs can reflect the registration error. Figure 3 The calculation process for mirror error is shown below:

[0085] Input mirrored image pairs to be registered and ; and These are the spectral images. I a and unlabeled images I u The image obtained after mirroring.

[0086] The registration network can predict the offset field of each image pair. and Deformation map image The offset field can be used For spectral images The deformation is achieved as follows:

[0087]

[0088] in, Indicates deformation operation. This indicates that the arrow's starting point (left side) tends towards its ending point (right side). express x The offset value at that location.

[0089] Similarly, for deformed images output in mirror space Similarly, the following conditions are met:

[0090]

[0091] Will Mirroring creates two target images, both of which are unlabeled. I u The registration paths are: the path via mirror flip-deformation-mirror flip, and the path via direct deformation.

[0092]

[0093] in, This indicates that the arrow's starting point (right side) tends towards the arrow's ending point (left side), and the superscript "*" indicates a mirror operation.

[0094] To measure the registration error, this invention first expresses the mirroring operation equivalently as an offset field. The calculation formula is as follows:

[0095]

[0096] in, The axis is perpendicular to the plane of symmetry, and the plane of symmetry is chosen to be the sagittal plane; N x This indicates the length of the horizontal axis. The starting images for the two registration paths mentioned above can also be obtained via... Unified as The formula is as follows:

[0097]

[0098] in, This indicates a two-way approximation.

[0099] Using the offset field superposition formula, The three deformations combine into a single offset field The calculation formula is:

[0100]

[0101]

[0102] The absolute value of the offset field difference between the two registration paths, flip-deformation-flip and direct deformation, is calculated as the registration error:

[0103]

[0104] The registration error is converted into registration confidence using a Gaussian function:

[0105]

[0106] in, Represents the L2 norm, This represents the standard deviation of the registration error.

[0107] This embodiment uses mirror error to measure registration error and calculates registration confidence from it. It is a pixel-level registration confidence map that can accurately reflect the alignment degree at each pixel position.

[0108] To enrich the training data, the semi-supervised training step of the segmentation network in this embodiment simultaneously uses registered and deformed atlas image-label pairs. and unlabeled image-deformation map label pairs Train the segmentation network. During training, As labeling information, the deformed spectral image is registered. and unlabeled images As input to the segmentation network.

[0109] Considering the deformed images paired with each unlabeled image All are composed of single atlas images The deformation-generated grayscale style directly inherits from... The image is relatively simple. In this embodiment, before inputting it into the segmentation network, style transfer is performed to add deformation atlas images. The diversity of styles enhances the segmentation performance of the trained segmentation network. Research shows that the Fourier amplitude spectrum of an image primarily contains grayscale style information, while the phase spectrum primarily contains structural information. Based on this principle, the bispectral reconstruction method preserves the phase spectrum information of the original image and linearly combines the amplitude spectrum information of the starting and target images, thereby altering the style of the original image. Some researchers have transferred this method to coarsely aligned image pairs. and Above, we propose an image-aligned style transformation (IST) method:

[0110]

[0111] in, and These are the amplitude spectrum and phase spectrum of the Fourier transform, respectively. It is the inverse Fourier transform. It is the converted deformation spectrum image. yes A random number between [a certain number of points].

[0112] The aforementioned style transfer method (IST) based on aligned images has certain drawbacks, namely, the image... and Low alignment and conversion coefficient When the image size is large, this method will produce more artifacts, affecting the segmentation accuracy. To address this issue and achieve style transfer with higher fidelity, this embodiment proposes a Weighted Image-aligned Style Transformation (WIST) method, which follows the registration confidence map... C Deformation map image and unlabeled images I u The image is divided into multiple regions with different alignment levels. Each region undergoes a style transfer based on the aligned image, and then the images are stitched together to obtain a style-transferred image. When performing style transfer, the higher the alignment of the area, the higher the transfer coefficient. Figure 4 The implementation process is demonstrated. Specifically, using the obtained registration confidence map C, through... Each gate function divides the confidence score into different regions, resulting in... A binary mask:

[0113]

[0114] These masks Compared with deformation spectrum images respectively and unlabeled images Multiply, we get Pairs of regions with varying degrees of alignment are grouped. Style transfer is then performed on these regions using the style transfer method based on aligned images described above, with the transfer coefficients... directly from Mid-sampling. Finally, all the transformed regions are superimposed to obtain the style-transferred image. . Figure 5 This demonstrates style-transferred images produced by the two style transfer methods described above when registration accuracy is low. For ease of description, the weighted style transfer method proposed in this embodiment is abbreviated as WIST. Figure 5 As can be seen, images converted by WIST have fewer artifacts and higher image quality.

[0115] Therefore, this embodiment calculates the registration confidence map based on the registration error perception and performs weighted style transfer on the image based on the registration confidence map. This not only preserves the spatial correspondence between the registration deformation map image and label pair, but also generates training data with diverse styles and high fidelity. This can effectively reduce the impact of errors generated by the registration network on segmentation learning and effectively improve the accuracy of brain tissue segmentation when only a single map image is available.

[0116] Optionally, in step (S4) of this embodiment, when training the segmentation network, for style-transferred images... Input it into the segmentation network to obtain the prediction result. Afterwards, the segmentation loss The calculation method is as follows:

[0117]

[0118] in, Dice This represents the Dice coefficient, used to measure the degree of overlap between regions. The calculation formula is:

[0119]

[0120] in, A and B This indicates two regions.

[0121] Considering unlabeled images Deformation map tags There is a certain spatial misalignment. To address this issue, in this embodiment, for unlabeled images... I u Input it into the segmentation network to obtain the segmentation result. Afterwards, the segmentation loss The calculation method is as follows:

[0122]

[0123] The segmentation loss described above is a confidence-guided Dice loss. Calculating the loss in the above manner allows the segmentation network to focus more on learning the aligned regions in the image-label, further improving the segmentation accuracy of the trained segmentation network.

[0124] In the weakly supervised training step of the registration network in this embodiment, the unlabeled image is... The input is fed into the segmentation network to obtain a more accurate pseudo-label. The registration network was trained under weak supervision using this label, resulting in a stronger ability to align brain tissues. Compared to initial unsupervised registration, the weakly supervised training step introduced additional loss. Spectral labels used to constrain registration deformation Optimized pseudo-labels predicted by segmentation networks Consistency. The calculation formula is as follows:

[0125]

[0126] Repeatedly and alternately perform the semi-supervised training steps of the segmentation network and the weakly supervised training steps of the registration network until the segmentation network and the registration network converge. The trained segmentation network and registration network can then be used to segment brain tissue.

[0127] In summary, this embodiment achieves registration error awareness and integrates registration errors into the registration-segmentation dual-model iterative learning, greatly eliminating the negative impact of registration errors on segmentation learning within the registration-segmentation iterative framework. Specifically, this embodiment proposes weighted style transfer and confidence-guided loss, enabling the segmentation network to learn more diverse and accurate data. Weighted style transfer not only preserves the spatial correspondence between the registered and deformed atlas image-label pairs but also generates style-diverse and high-fidelity data; the confidence-guided loss allows the segmentation network to focus more intently on learning the alignment regions of unlabeled images and deformed atlas label pairs. Both significantly improve the performance of the segmentation network.

[0128] The training method provided in this embodiment can accurately predict on a single-label dataset and has higher accuracy compared to other existing single-label segmentation methods. The segmentation performance of this embodiment is verified and explained below by comparing the segmentation results.

[0129] This embodiment was tested on two publicly available brain datasets, OASIS and CANDIShare, using only one label, and compared with five existing registration-segmentation iterative learning algorithms (DeepAtlas, Brainstorm, PC-Reg-RT, BRBS, StyleSeg) and four graph registration-based methods (Ants, VoxelMorph, 10*VTN, PC-Net). The performance of the segmentation and registration models was measured using the Dice coefficient and Hausdorff Distance (HD). The former measures the overlap between regions, while the latter calculates the maximum distance between surfaces. The HD calculation formula is as follows:

[0130]

[0131] Here, A and B represent all the pixels contained in the two labels.

[0132] Table 1 presents the quantitative results of the segmentation model of this embodiment and similar methods on the OASIS and CANDIShare datasets, used to quantitatively compare the performance of the segmentation network in different registration-segmentation iteration methods. According to the quantitative results shown in Table 1, the segmentation Dice of this embodiment on these two datasets are 0.868±0.004 and 0.859±0.007, respectively, representing improvements of 2.0% and 0.9% compared to the best algorithms of the same type. Furthermore, the segmentation model provided in this embodiment has the minimum or second smallest (second) Hausdorff distance. To intuitively demonstrate the superior performance of the segmentation model of this invention, Figure 6 The segmentation prediction results of this embodiment and several other algorithms are shown. It can be seen that, in Figure 6 The area marked by the box in this embodiment is closer to the gold standard (Ground-truth), demonstrating that the segmentation network trained in this embodiment has higher segmentation accuracy than other mainstream algorithms in the case of single label.

[0133] Table 1 Performance of segmentation networks in different registration-segmentation iterative methods

[0134]

[0135] Consider extending this embodiment to other application scenarios, such as surgical navigation or arbitrary brain tissue segmentation. In such scenarios, the registration model can achieve preoperative and intraoperative image alignment, and can also map new atlas labels to the image to be segmented when a single label is provided, thereby achieving tissue segmentation. Therefore, the performance of the registration model is equally important. Table 2 shows the quantitative registration results of this embodiment, similar methods, and mainstream registration methods on two datasets, OASIS and CANDIShare. It can be seen that the registration Dice of the method of this embodiment on these two datasets are 0.847±0.004 and 0.844±0.005, respectively, which are 0.7% and 1.6% higher than the best algorithms of the same type. The results shown in Table 2 also show that this embodiment has the lowest Hausdorff distance compared to existing similar methods in terms of registration performance. These results all demonstrate that the present invention also has excellent registration performance and therefore has broad application prospects.

[0136] Table 2. Registration performance of different registration methods

[0137]

[0138] Furthermore, to verify the role of the weighted style transfer and confidence-guided loss proposed in this invention in improving performance, the existing registration-segmentation iterative learning algorithm StyleSeg is used as a comparison. StyleSeg uses the aforementioned registration model PC-Net and segmentation model U-Net for iteration, and uses image alignment style transfer IST during the training of the segmentation model, only displaying the style-transferred images. Deformation map tags The input network is used for learning. The method after introducing weighted style transfer alone is denoted as StyleSeg+WIST, and the method after introducing confidence-guided loss alone is denoted as Style+L. cgd The method that incorporates weighted style transfer and confidence-guided loss is denoted as StyleSeg V2. The registration accuracy of the registration model and the segmentation accuracy of the segmentation model trained by the four methods are as follows: Figure 7 As shown in (a) and (b), the horizontal axis represents the number of iterations. According to... Figure 7 The results show that the weighted style transfer and confidence-guided loss proposed in this invention can effectively improve the registration and segmentation effects of the registration-segmentation iterative learning method. Furthermore, when the weighted style transfer and confidence-guided loss are introduced simultaneously, the optimal registration and segmentation effects are obtained.

[0139] Example 2:

[0140] A method for segmenting brain tissue, comprising:

[0141] The brain tissue image to be segmented is input into the segmentation network trained by the method of training a brain tissue segmentation model using a single atlas image provided in Example 1 above, and the brain tissue segmentation result is obtained.

[0142] Example 3:

[0143] A method for segmenting brain tissue, comprising:

[0144] The atlas image and the brain tissue image to be segmented are input into the registration network trained by the method for training a brain tissue segmentation model using a single atlas image provided in Example 1 above, to obtain the offset field between the atlas image and the brain tissue image to be segmented. The atlas images are brain tissue images with segmentation labels.

[0145] According to the offset field The labels of the atlas image are deformed to obtain deformed labels, which are then used as the brain tissue segmentation results of the brain tissue image to be segmented.

[0146] Example 4:

[0147] A computer-readable storage medium includes a stored computer program; when the computer program is executed by a processor, it controls the device where the computer-readable storage medium is located to execute the method for training a brain tissue segmentation model using a single atlas image provided in Embodiment 1 above, and / or the brain tissue segmentation method provided in Embodiment 2 or 3 above.

[0148] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for training a brain tissue segmentation model using a single atlas image, characterized in that, include: The semi-supervised training steps for a segmentation network include: (S1) Predicting spectral images using a registration network I a and unlabeled images I u offset fields between and according to the offset field For spectral images I a and its labels S a Perform deformation operations separately to obtain deformation spectrum images. Deformation map tags The atlas images are brain tissue images with segmentation labels, while the unlabeled images are brain tissue images without segmentation labels. (S2) Calculate the registration confidence plot C Used to represent deformation spectrum images and unlabeled images I u Alignment level at each pixel; (S3) The deformation spectrum image is processed according to the registration confidence map C. and unlabeled images I u The image is divided into multiple regions with different degrees of alignment for the deformation spectrum. Each region undergoes a style transfer based on the aligned image, and then the images are stitched together to obtain a style-transferred image. When performing style transfer, the higher the alignment of the area, the higher the transfer coefficient. (S4) Label the deformation map As supervisory information, unlabeled images I u and style transfer images The segmentation network is trained using the input. The steps for weakly supervised training of a registered network include: (W1) Predict the unlabeled image using a segmentation network. I u The segmentation results are used as pseudo-labels ; (W2) Predicting spectral images using a registration network I a and unlabeled images I u offset fields between , with spectral images I a tags S a According to the offset field Deformation map labels obtained after performing deformation operations Compared to pseudo tags The error is used as the training loss to train the registration network; The dual-model iterative learning process includes repeatedly and alternately performing semi-supervised training steps for the segmentation network and weakly supervised training steps for the registration network until the segmentation network and the registration network converge.

2. The method for training a brain tissue segmentation model using a single atlas image as described in claim 1, characterized in that, In step (S2), the registration confidence map is calculated. C ,include: The atlas image I a and the unlabeled image I u Mirror the images respectively and ; Predicting images using the registration network and offset field between ; according to Calculate the composite migration field ; according to Calculate the registration error E ; according to Calculate the registration confidence map C ; in, Indicates deformation operation. This represents the equivalent offset field for the mirror operation. Represents the L2 norm, This represents the standard deviation of the registration error.

3. The method for training a brain tissue segmentation model using a single atlas image as described in claim 2, characterized in that, In step (S4), when training the segmentation network, for style-transferred images... Input it into the segmentation network to obtain the prediction result. Afterwards, the segmentation loss The calculation method is as follows: For unlabeled images I u Input it into the segmentation network to obtain the segmentation result. Afterwards, the segmentation loss The calculation method is as follows: in, Dice This represents the Dice coefficient.

4. The method for training a brain tissue segmentation model using a single atlas image as described in claim 1 or 2, characterized in that, In step (W2), when training the registration network, the training loss also includes: similarity loss. and regularization terms The expressions for the two are as follows: in, NLCC This represents the locally normalized correlation coefficient; Representing spectral images I a According to the offset field The deformed spectral image; Represents the gradient operator, This represents the L2 norm.

5. The method for training a brain tissue segmentation model using a single atlas image as described in claim 4, characterized in that, In the dual-model iterative learning step, before repeatedly and alternately executing the semi-supervised training step of the segmentation network and the weakly supervised training step of the registration network, the method further includes: performing an unsupervised registration initialization step; the unsupervised registration initialization step includes: Predicting spectral images using registration networks I a and unlabeled images I u offset fields between loss based on similarity and regularization terms The registration network is trained using this training loss.

6. The method for training a brain tissue segmentation model using a single atlas image as described in claim 5, characterized in that, In the dual-model iterative learning step, before performing the unsupervised registration initialization step, the method further includes: performing a preprocessing operation; the preprocessing operation includes at least one of the following operations: Perform bias field correction on all images; Resample all images to the same size; Perform an affine transformation on all images to the atlas image.

7. A method for segmenting brain tissue, characterized in that, include: The brain tissue image to be segmented is input into the segmentation network trained by the method of training a brain tissue segmentation model using a single atlas image as described in any one of claims 1 to 6, and the brain tissue segmentation result is obtained.

8. A method for segmenting brain tissue, characterized in that, include: The atlas image and the brain tissue image to be segmented are input into the registration network trained by the method described in any one of claims 1 to 6 for training a brain tissue segmentation model using a single atlas image, to obtain the offset field between the atlas image and the brain tissue image to be segmented. The atlas images are brain tissue images with segmentation labels. According to the offset field The labels of the atlas image are deformed to obtain deformed labels, which are then used as the brain tissue segmentation results of the brain tissue image to be segmented.

9. A computer-readable storage medium, characterized in that, The computer program includes a stored computer program; when executed by a processor, the computer program controls the device containing the computer-readable storage medium to perform the method for training a brain tissue segmentation model using a single atlas image as described in any one of claims 1 to 6, or the brain tissue segmentation method as described in claim 7 or 8.

Images