Automatic segmentation method for horizontal sinus sinus sigmoideus junction by flat scanning CT

Abstract

The invention discloses an automatic segmentation method for a horizontal sinus sinus sigmoideus junction through a plain-scan CT, and the method mainly comprises the steps: obtaining a pseudo plain-scan CT through the processing of an enhanced CT, carrying out the brain tissue segmentation of the enhanced CT through a deep learning pre-training model, correcting a segmentation result, unifying the gray value of the brain tissue as a background value, and carrying out the segmentation of the pseudo plain-scan CT. Removing blood vessel gray scale information to generate a pseudo-plain-scan CT image; then, preprocessing the pseudo-plain-scan CT data, and constructing an annotation data set; a 3D-UNet network model is adopted, a context coding module is introduced, feature enhancement is performed by learning a group of scaling factors, and the segmentation precision and robustness are improved; deep learning training is carried out on the pseudo-plain-scan CT data, a Dice loss function is adopted, and the network performance is optimized through a deep supervision method. Finally, the trained network model can be used for automatic segmentation of the transverse sinus sinus sigmoideus junction of the plain-scan CT image of an actual patient. According to the scheme of the invention, the pseudo-plain-scan CT is obtained by processing the enhanced CT image, and the automatic segmentation of the transverse sinus sigmoideal sinus junction in the plain-scan CT image is realized by using the deep learning network in which the context coding module is introduced.

Description

Technical Field

[0001] The invention relates to the fields of computer vision and medicine, and more particularly to an automatic segmentation method for the junction of the transverse sinus and sigmoid sinus of plain scan CT. Background Art

[0002] Cranial nerve diseases such as trigeminal neuralgia, hemifacial spasm, masseter spasm, and glossopharyngeal neuralgia are common neurological diseases with a high incidence rate. Microvascular Decompression (MVD) has the characteristics of less trauma, high cure rate, and low complication rate. It has become the preferred treatment option and has received widespread attention and application. MVD surgery is usually performed by craniotomy through the retrosigmoid approach. When performing craniotomy, the location of the bone window needs to be designed to avoid important blood vessels and nerve structures and be close to the transverse sinus and sigmoid sinus. Therefore, it is crucial to determine the exact location of the junction of the transverse sinus and sigmoid sinus during craniotomy.

[0003] However, due to clinical needs and equipment limitations, many hospitals only use plain CT scans without angiography information for preoperative evaluation. Plain CT images lack the visualization of blood vessels, which makes it more difficult for doctors to locate the transverse sinus-sigmoid junction relying solely on images. Traditional image segmentation methods, such as threshold-based segmentation and edge detection, usually rely on grayscale differences and morphological features of CT images. These methods are difficult to effectively segment the transverse sinus-sigmoid junction when processing plain CT images, and often require preprocessing of the original image and subsequent manual optimization. The accuracy of the segmentation results is difficult to guarantee and cannot meet the high-precision requirements in clinical surgery.

[0004] With the emergence of surgical robots and the development of technologies such as artificial intelligence algorithms, deep learning methods can process large-scale data to achieve higher-precision image segmentation and positioning. Deep learning models based on convolutional neural networks (CNN) and UNet classic networks can further improve the segmentation effect, so that the accuracy of the results meets clinical requirements. Existing methods have not yet conducted research on the segmentation of the transverse sinus and sigmoid junction of the patient's brain. The present invention proposes an automatic segmentation method for the transverse sinus and sigmoid junction of plain scan CT. According to actual clinical application needs, the segmentation mask of the transverse sinus and sigmoid junction is predicted using plain scan CT images, aiming to provide doctors with a more accurate and reliable reference and improve the safety and success rate of surgery. Summary of the invention

[0005] In order to solve the limitations of existing methods, an automatic segmentation method for the transverse sinus and sigmoid junction of plain scan CT is proposed, which includes the following steps:

[0006] Step 1: Use enhanced CT to perform label segmentation of the transverse sinus-sigmoid junction to obtain the patient's enhanced CT data and the segmentation label of the transverse sinus-sigmoid junction.

[0007] Step 2: Use the deep learning pre-trained model for automatic head bone removal to segment the patient's brain tissue in the enhanced CT, manually correct the segmentation result, and obtain the patient's brain tissue segmentation map. Process the enhanced CT, set the grayscale of the brain tissue to a uniform background value, and obtain a pseudo plain scan CT.

[0008] Step 3: Preprocess the pseudo-plain scan CT data, extract the region of interest (ROI) at the junction of the transverse sinus and sigmoid sinus of each patient, unify the image size and resolution through resampling operations, and obtain a labeled data set.

[0009] Step 4: Use pseudo plain scan CT and the label of the transverse sinus and sigmoid junction to train the deep learning network. Use the 3D-UNet network model to input the pseudo plain scan CT image into the network, introduce a context encoding module, and obtain the predicted segmentation label map through the encoder-decoder structure. Calculate the Dice loss function between the label of the transverse sinus and sigmoid junction, and use the deep supervision method to train the network, iterating until the model converges.

[0010] Step 5: After uniform preprocessing of the patient's actual cranial CT image, the trained deep learning network model is input to output the corresponding transverse sinus-sigmoid sinus junction segmentation map.

[0011] Furthermore, in step 2, the deep learning pre-trained model for automatic head bone removal is used to segment the patient's brain tissue in the enhanced CT, including:

[0012] A 3D fully convolutional deep neural network was used to segment the patient's cranial CTA images. A pre-trained model based on the UNet architecture was used. The structure included: 3 groups of bottleneck residual convolution blocks with 3×3×3 voxel kernels, each group had 3, 3, and 4 convolution blocks, and each group was connected to each other by a downsampling convolution block with a step size of 2; the decoder structure was the same, using a transposed convolution upsampling method. After training, the brain tissue mask was obtained and manually corrected again to improve the segmentation accuracy of the transverse sinus and sigmoid sinus junction.

[0013] The enhanced CT is processed to obtain a pseudo plain scan CT. The specific operation is as follows:

[0014] In the patient's enhanced CT, the corrected brain tissue segmentation is used to set the pixel value at the position where the segmentation label is 1 to the uniform background gray value of 0 in the plain scan CT, remove the gray information of the blood vessels, retain the skull information, and output the patient's pseudo plain scan CT.

[0015] Furthermore, in step 3, the pseudo plain scan CT data is preprocessed, including:

[0016] The ROI of each patient's image was manually framed using a tool, and the physical coordinates of the origin and pixel size of the three-dimensional block were recorded. The actual physical size of the ROI was calculated using the resolution of each pseudo-plain scan CT image, and the maximum value was selected as the physical size for uniform cutting. The coordinates of the center point of the ROI manually framed by each patient were fixed, and the ROI was re-cut with a uniformly determined size to obtain the ROI data set.

[0017] Furthermore, in step 4, a 3D-UNet network model is used for deep learning training. The model has an encoding path and a decoding path, and a context encoding module is introduced. In the encoder, each layer contains two 3×3×3 convolutional layers with a stride of 1. Each convolution is followed by an instance normalization layer (IN) and a nonlinear activation function (ReLU), followed by a 2×2×2 maximum pooling layer with a stride of 2. In the decoder, the structure of each layer is the same as that of the encoder, and the upsampling method of the transposed convolution is used to restore it to the same size as the input image. Jump connections are connected in series between the encoder and decoder of each layer. The last layer uses a 1×1×1 convolutional layer to reduce the number of output channels to the number of labels, followed by an activation function to output the prediction result.

[0018] The context encoding module operates on a latent feature map of shape C×H×W×E, where C is the number of channels and H×W×E is the height, width and depth of the feature map, respectively, and generates the encoding context through a series of convolution processes. The encoding context learns a set of scaling factors that increase the weights of specific channel features related to the transverse sinus-sigmoid junction. The scaling factors are activated by a sigmoid function and used as a channel attention mechanism. The residual of the original latent space is combined with the scaled features to avoid zero weights suppressing a feature and promote network back propagation. The output of this attention module can be expressed as:

[0019]

[0020] Where P is the latent feature map, k is the scaling factor, is the channel multiplication, is element-wise addition.

[0021] The network model calculates the Dice loss between the predicted segmentation result and the true segmentation label:

[0022]

[0023] Among them, TP is True Positive, that is, the number of voxels predicted as positive samples are actually positive samples; FP is False Positive, that is, the number of voxels predicted as positive samples are actually negative samples; FN is False Negative, that is, the number of voxels predicted as negative samples are actually positive samples.

[0024] The deep supervision method is used for network training, and the model is iterated until convergence. The patient's CT image is preprocessed and input into the trained network, and the corresponding transverse sinus-sigmoid sinus junction segmentation map can be output.

[0025] The beneficial effects of the present invention are as follows:

[0026] The present invention utilizes the patient's enhanced CT images and transverse sinus and sigmoid sinus junction segmentation labels, and based on a deep learning network model, removes vascular grayscale information from the enhanced CT to obtain a pseudo plain scan CT, learns the shape and position relationship between the skull and the transverse sinus and sigmoid sinus in the CT image, and realizes the segmentation of the transverse sinus and sigmoid sinus junction in the patient's CT image in conjunction with context information, obtains predicted images and evaluation results, which can further guide the application of clinical surgery and improve the accuracy and stability of MVD surgery.

[0027] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the specific implementation of the present invention or the technical solution in the prior art, the following briefly introduces the drawings required for use in the specific implementation or the prior art description. Obviously, the drawings in the following description are Figure 1 This is an implementation mode of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0029] Figure 1 It is a flow chart of the automatic segmentation method of the transverse sinus and sigmoid sinus junction of plain scan CT according to an embodiment of the present invention;

[0030] Figure 2 Schematic diagram of the deep learning network model structure of the present invention. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the drawings. The described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments.

[0032] Figure 11 is a flow chart of a method for automatic segmentation of the transverse sinus and sigmoid sinus junction of plain scan CT according to an embodiment of the present invention. The method includes the following five steps.

[0033] Step 1: Use enhanced CT to perform label segmentation of the transverse sinus-sigmoid junction to obtain the patient's enhanced CT data and the segmentation label of the transverse sinus-sigmoid junction.

[0034] Step 2: Use the deep learning pre-trained model for automatic head bone removal to segment the patient's brain tissue in the enhanced CT, manually correct the segmentation result, and obtain the patient's brain tissue segmentation map. Process the enhanced CT, set the grayscale of the brain tissue to a uniform background value, and obtain a pseudo plain scan CT.

[0035] The deep learning pre-trained model with automatic head bone removal is used to segment the patient's brain tissue in enhanced CT, including: using a 3D fully convolutional deep neural network to segment the patient's cranial CTA images, using a pre-trained model based on the UNet architecture, the structure includes: 3 groups of bottleneck residual convolution blocks with 3×3×3 voxel kernels, each group has 3, 3, and 4 convolution blocks, each group is connected to each other by a downsampling convolution block with a step size of 2; the decoder structure is the same, using a transposed convolution upsampling method. After training, the brain tissue mask is obtained, and then manually corrected again to accurately divide the boundary between the skull and the transverse sinus and sigmoid sinus vessels, and improve the segmentation accuracy of the transverse sinus and sigmoid sinus junction.

[0036] In the patient's enhanced CT, the corrected brain tissue segmentation is used to set the pixel value at the position where the segmentation label is 1 to the uniform background gray value of 0 in the plain scan CT, remove the gray information of the blood vessels, retain the skull information, and output the patient's pseudo plain scan CT.

[0037] Step 3: Preprocess the pseudo-plain scan CT data, extract the region of interest (ROI) at the junction of the transverse sinus and sigmoid sinus of each patient, unify the image size and resolution through resampling operations, and obtain a labeled data set.

[0038] The specific operations are as follows: use the tool to manually frame the ROI of each patient's image, record the physical coordinates of the origin and pixel size of the three-dimensional block; calculate the actual physical size of the ROI through the resolution of each pseudo-plain scan CT image, and select the maximum value as the physical size of the unified cutting; fix the coordinates of the center point of the ROI manually framed by each patient, and re-cut it with a uniformly determined size to obtain the ROI data set.

[0039] Step 4: Use pseudo-unenhanced CT and transverse sinus-sigmoid junction labels to train a deep learning network. The network model is as follows: Figure 2As shown, a 3D-UNet network model is used to input pseudo plain scan CT images into the network, and a context encoding module is introduced. The predicted segmentation label map is obtained through the encoder-decoder structure, and the Dice loss function between the transverse sinus and sigmoid sinus junction label is calculated. The network is trained using a deep supervision method and iterated until the model converges.

[0040] The 3D-UNet network model is used for deep learning training. The model has an encoding path and a decoding path, and introduces a context encoding module. In the encoder, each layer contains two 3×3×3 convolutional layers with a stride of 1. Each convolution is followed by an instance normalization layer (IN) and a nonlinear activation function (ReLU), followed by a 2×2×2 maximum pooling layer with a stride of 2. In the decoder, each layer has the same structure as the encoder, and uses transposed convolution upsampling to restore it to the same size as the input image. Jump connections are connected in series between the encoder and decoder of each layer. The last layer uses a 1×1×1 convolutional layer to reduce the number of output channels to the number of labels, followed by an activation function to output the prediction result.

[0041] The context encoding module is Figure 2 As shown in Figure 1, the module operates on a latent feature map of shape C×H×W×E, where C is the number of channels, and H×W×E is the height, width, and depth of the feature map, respectively, to generate an encoding context through a series of convolution processes. The encoding context learns a set of scaling factors that increase the weights of specific channel features related to the transverse sinus-sigmoid junction. The scaling factors are activated by a sigmoid function and used as a channel attention mechanism. The residual of the original latent space is combined with the scaled features to avoid zero weights suppressing a feature and promote network back propagation. The output of this attention module can be expressed as:

[0042]

[0043] Where P is the latent feature map, k is the scaling factor, is the channel multiplication, is element-wise addition.

[0044] The network model calculates the Dice loss between the predicted segmentation result and the true segmentation label, as shown in formula (2):

[0045]

[0046] Among them, TP is True Positive, that is, the number of voxels predicted as positive samples are actually positive samples; FP is False Positive, that is, the number of voxels predicted as positive samples are actually negative samples; FN is False Negative, that is, the number of voxels predicted as negative samples are actually positive samples.

[0047] Step 5: After uniform preprocessing of the patient's actual cranial CT image, the trained deep learning network model is input to output the corresponding transverse sinus-sigmoid sinus junction segmentation map.

[0048] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for automatic segmentation of the transverse sinus and sigmoid sinus junction in plain scan CT, characterized in that: The following steps are involved: Step 1: Use enhanced CT to perform label segmentation of the transverse sinus-sigmoid junction to obtain the patient's enhanced CT data and the segmentation label of the transverse sinus-sigmoid junction. Step 2: Use the deep learning pre-trained model for automatic head bone removal to segment the patient's brain tissue in the enhanced CT, manually correct the segmentation result, and obtain the patient's brain tissue segmentation map. Process the enhanced CT, set the grayscale of the brain tissue to a uniform background value, and obtain a pseudo plain scan CT. Step 3: Preprocess the pseudo-plain scan CT data, extract the region of interest (ROI) at the junction of the transverse sinus and sigmoid sinus of each patient, unify the image size and resolution through resampling operations, and obtain a labeled data set. Step 4: Use pseudo plain scan CT and the label of the transverse sinus and sigmoid junction to train the deep learning network. Use the 3D-UNet network model to input the pseudo plain scan CT image into the network, introduce a context encoding module, and obtain the predicted segmentation label map through the encoder-decoder structure. Calculate the Dice loss function between the label of the transverse sinus and sigmoid junction, and use the deep supervision method to train the network, iterating until the model converges. Step 5: After uniform preprocessing of the patient's actual cranial CT image, the trained deep learning network model is input to output the corresponding transverse sinus-sigmoid sinus junction segmentation map.

2. The method according to claim 1, characterized in that: The deep learning pre-trained model for automatic head bone removal is used to segment the patient's brain tissue in enhanced CT, including: A 3D fully convolutional deep neural network was used to segment the patient's cranial CTA images. A pre-trained model based on the UNet architecture was used. The structure included: 3 groups of bottleneck residual convolution blocks with 3×3×3 voxel kernels, each group had 3, 3, and 4 convolution blocks, and each group was connected to each other by a downsampling convolution block with a step size of 2; the decoder structure was the same, using a transposed convolution upsampling method. After training, the brain tissue mask was obtained and manually corrected again to improve the segmentation accuracy of the transverse sinus and sigmoid sinus junction.

3. The method according to claim 1, characterized in that The enhanced CT is processed to set the grayscale of the brain tissue to a uniform background value to obtain a pseudo plain scan CT, including: In the patient's enhanced CT, the corrected brain tissue segmentation is used to set the pixel value at the position where the segmentation label is 1 to the uniform background gray value of 0 in the plain scan CT, remove the gray information of the blood vessels, retain the skull information, and output the patient's pseudo plain scan CT.

4. The method according to claim 1, characterized in that: Preprocessing operations are performed on pseudo plain scan CT data, including: The ROI of each patient's image was manually framed using a tool, and the physical coordinates of the origin and pixel size of the three-dimensional block were recorded. The actual physical size of the ROI was calculated using the resolution of each pseudo-plain scan CT image, and the maximum value was selected as the physical size for uniform cutting. The coordinates of the center point of the ROI manually framed by each patient were fixed, and the ROI was re-cut with a uniformly determined size to obtain the ROI data set. The ROI image was resampled and linear interpolation was used to obtain a data set with consistent pixel size and resolution. Finally, an image with a uniform resolution (1.0 mm × 1.0 mm × 0.8 mm) was obtained, and its size was 96 × 112 × 80.

5. The method according to claim 1, characterized in that Using the 3D-UNet network model, a context encoding module is introduced for deep learning training, including: The model has an encoding path and a decoding path. In the encoder, each layer contains two 3×3×3 convolutional layers with a stride of 1. Each convolution is followed by an instance normalization layer (IN) and a nonlinear activation function (ReLU), followed by a 2×2×2 max pooling layer with a stride of 2. In the decoder, each layer has the same structure as the encoder, and uses transposed convolution to upsample to restore to the same size as the input image. Jump connections are connected in series between the encoder and decoder of each layer. The last layer uses a 1×1×1 convolutional layer to reduce the number of output channels to the number of labels, followed by an activation function to output the prediction result. The context encoding module operates on a latent feature map of shape C×H×W×E, where C is the number of channels and H×W×E is the height, width and depth of the feature map, respectively, and generates the encoding context through a series of convolution processes. The encoding context learns a set of scaling factors that increase the weights of specific channel features related to the transverse sinus-sigmoid junction. The scaling factors are activated by a sigmoid function and used as a channel attention mechanism. The residual of the original latent space is combined with the scaled features to avoid zero weights suppressing a feature and promote network back propagation. The output of this attention module can be expressed as: Where P is the latent feature map, k is the scaling factor, is the channel multiplication, is element-wise addition.

6. The method according to claim 1, characterized in that Calculate the Dice loss between the predicted segmentation result and the true segmentation label: Among them, TP is True Positive, that is, the number of voxels predicted as positive samples are actually positive samples; FP is False Positive, that is, the number of voxels predicted as positive samples are actually negative samples; FN is False Negative, that is, the number of voxels predicted as negative samples are actually positive samples.

Images