A method for automatic identification and segmentation of intracranial aneurysms in 3.0t high-resolution mri t1 sequences

Abstract

The application relates to the technical field of medical image processing, in particular to a method for automatically identifying and segmenting intracranial aneurysms in 3.0T high-resolution MRI T1 sequences, which extracts high-dimensional features of intracranial aneurysms in the T1 sequence through convolution operation in the encoding path in a full convolutional neural network; further extracts higher-dimensional features and reconstructs the features through decoding path convolution and deconvolution operation in the network, removes irrelevant background voxels in the T1 image, retains voxels related to the intracranial aneurysm, and realizes accurate identification and segmentation of the intracranial aneurysm.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method for automatically identifying and segmenting intracranial aneurysms in a 3.0T high-resolution MRI T1 sequence. Background Technology

[0002] Intracranial aneurysms are abnormal bulges in cerebral arteries. Rupture of an intracranial aneurysm is the leading cause of non-traumatic subarachnoid hemorrhage, a catastrophic event with extremely high rates of disability and death. Therefore, early detection and intervention of intracranial aneurysms are essential, and the identification and segmentation of intracranial aneurysms has become a research hotspot in the field of medical image processing.

[0003] Current methods for identifying and segmenting intracranial aneurysms are based on invasive, radiation-exposed, costly, or time-consuming enhanced imaging techniques such as DSA, CTA, and MRA. Commonly used methods primarily employ convolutional neural networks (CNNs) for image block-level intracranial aneurysm identification and fully convolutional neural networks (WCNNs) for voxel-level segmentation. However, even with advanced enhanced imaging techniques, the results of voxel-level segmentation of intracranial aneurysms still fall short of clinical requirements, necessitating further optimization and refinement of identification and segmentation methods.

[0004] Using DSA, CTA, and MRA imaging techniques for intracranial aneurysm segmentation is not optimal. This is because DSA is an invasive procedure with significant radiation exposure; CTA also involves significant radiation exposure; and MRA is time-consuming and expensive. Furthermore, patient complaints, misleading information, and physician experience can lead to patients not undergoing these examinations, resulting in missed diagnoses of aneurysms. Therefore, using DSA, CTA, and MRA for intracranial aneurysm segmentation and identification has limitations. Thus, establishing a relatively universal imaging technique for intracranial aneurysm identification and segmentation is the most meaningful approach. Summary of the Invention

[0005] To address the aforementioned technical problems, the present invention aims to provide a method for automatically identifying and segmenting intracranial aneurysms in a 3.0T high-resolution MRI T1 sequence. This method addresses the shortcomings of current intracranial aneurysm identification and segmentation methods, which require further optimization and improvement, and the lack of a universally applicable identification and segmentation method. By introducing the more universal MRI T1 sequence, a method for automatically identifying and segmenting intracranial aneurysms in a 3.0T high-resolution MRI T1 sequence is proposed.

[0006] To achieve the above-mentioned technical effects, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for automatically identifying and segmenting intracranial aneurysms in a 3.0T high-resolution MRI T1 sequence, specifically comprising the following steps:

[0008] S1: Acquire high-resolution data for T1;

[0009] Specifically, the acquisition parameters for the T1 high-resolution data are as follows: head scanning using a head coil; layer thickness of 0.6 mm; interlayer spacing of 0.3 mm; pixel spacing of 0.2976 mm; flip angle of 90°; repetition time of 800 ms; echo time of 18.722 ms; acquisition time of 1.7 min; and matrix of 672×672. The head is scanned using these sequence of scanning parameters to obtain the T1 high-resolution data.

[0010] S2: Select the training sample set and the test sample set from the high-resolution data of T1;

[0011] Specifically, 4 / 5 of the high-resolution data T1 obtained in step S1 is selected as the training sample set for model training, and the remaining 1 / 5 is used as the test sample set to test the model.

[0012] S3: Perform training set preprocessing and test set preprocessing on the selected training sample set and test sample set respectively;

[0013] Furthermore, the training set preprocessing includes the following steps:

[0014] S3a1: Perform intensity value clipping on the original training sample set data to obtain the intensity value-clipped training sample set data. The clipping formula is as follows:

[0015]

[0016] Where f'(x,y,z) and f(x,y,z) represent the intensity values ​​of the cropped and original images, respectively.

[0017] When the voxel intensity value in the matrix is ​​less than 0, the voxel intensity value is 0; when the voxel intensity value is between 0 and 2000, the voxel intensity value remains unchanged; when the voxel intensity value in the matrix is ​​greater than 2000, the voxel intensity value is 2000.

[0018] S3a2: An adaptive region of interest (ROI) selection method is used to extract 3D image blocks of intracranial aneurysms from the training sample set data to reduce the image size and thus reduce the GPU memory resources consumed by the 3D data. Furthermore, since patients may have multiple aneurysms, traditional extraction methods can lead to the severing of surrounding aneurysms, affecting model iteration. The flowchart of the adaptive ROI selection method used in this invention is attached. Figure 2 As shown;

[0019] S3a3: Standardize the extracted region of interest. The standardization process can accelerate model convergence. The standardization process is as follows:

[0020]

[0021] Where, f(x,y,z) norm f(x,y,z) represents the intensity values ​​of the region of interest before and after standardization. μ and f σ These are the mean and standard deviation of the intensity values ​​in the region of interest, respectively.

[0022] S3a4: Data augmentation is performed on the standardized image to obtain the T1 region of interest image. This data augmentation is a method to prevent overfitting in deep learning models and also has a positive effect on convolutional neural networks. Therefore, this invention utilizes random flipping, rotation, and elastic transformation to process the region of interest image, alleviating overfitting of the convolutional neural network through image augmentation.

[0023] Furthermore, the test set preprocessing includes the following steps:

[0024] S3b1: The intensity values ​​of the original test sample set data are clipped to obtain the intensity-clipped test sample set data. The clipping formula is as follows:

[0025]

[0026] Where f'(x,y,z) and f(x,y,z) represent the intensity values ​​of the cropped and original images, respectively.

[0027] S3b2: The aneurysm images of the test sample set data after intensity value cropping are cropped to reduce the image size. Specifically, the obtained test sample set data is high-resolution data, so the three-dimensional size of the images is very large. Intracranial aneurysms grow in arteries, and there are no aneurysms in the image background and around the skull. Therefore, the size of the aneurysm images can be cropped to remove irrelevant background voxels in the T1 images and reduce the image size.

[0028] S3b3: Transform the image layer count so that the transformed layer count is divisible by 16. Specifically, the layer count of the test sample data after cropping often does not meet this condition. This can be achieved by adding black images to both ends of the test sample data to increase the layer count, or by removing the images from both ends to decrease the layer count. The transformed layer count is calculated as follows:

[0029]

[0030] Among them, s t and s o These represent the number of layers in the data before and after the transformation, respectively. This is for rounding down. If s t Less than s o Then, delete images at both ends of the image to reduce the layer number s. t equals s o If s t Greater than s o Then, black images are added to both ends of the image to make the number of layers s. t equals s o The T1 data transformed using this layer transformation method will not have any aneurysm cut off or lost, ensuring that complete data containing the aneurysm is input into the model for testing.

[0031] S4: Establish a fully convolutional neural network model;

[0032] The fully convolutional neural network model comprises three parts: an encoder, a decoder, and a voxel-level classifier. The encoder is primarily used for extracting intracranial aneurysm features. The decoder further extracts high-level features and reconstructs the aneurysm. The voxel-level classifier fuses the reconstructed aneurysm feature maps and assigns the voxel category using the sigmoid function, thus achieving aneurysm segmentation. More specifically:

[0033] (4a) The encoder consists of 5 basic units, with the following structure: Unit 1 -> Pooling Layer -> Unit 2 -> Pooling Layer -> Unit 3 -> Pooling Layer -> Unit 4 -> Pooling Layer -> Unit 5. Each unit is composed of: Input Features -> Convolutional Layer -> Residual Concatenation (Input Features, Convolutional Layer) -> Channel Attention Layer. Each unit has feature extraction capabilities; residual concatenation can handle gradient vanishing issues; and the channel attention layer can select effective features. The 5 units are concatenated to form the encoder of this invention, enabling the extraction of intracranial aneurysm features.

[0034] The encoder layers are configured as follows:

[0035] Unit n (n = 1, 2, 3, 4, 5) contains two identical convolutional layers. The number of convolutional windows in each layer is set to 32 * 2n - 1, the kernel size is 3 × 3, and the stride is 1. Group normalization and ELU activation units are used. Residual concatenation aggregates and concatenates the input features with the convolutional output features. The number of input features is the sum of the input features and the convolutional output features. The specific implementation of the channel attention layer is as follows:

[0036] f a =l2(l1(Avg(f) o)))

[0037] f b =l2(l1(Max(f) o )))

[0038] f out =f o *(sigmoid(add(f a ,f b )))

[0039] Where Avg and Max represent global average pooling and global max pooling, respectively; add and sigmoid represent voxel-by-voxel addition and sigmoid operation of the feature maps, respectively; * represents voxel-by-voxel multiplication of the feature maps. l1 and l2 are two fully connected layers with different numbers of neurons, f a and f b For the features mapped by the fully connected layer, f o and f out The input and output features are kept constant, and the number of feature maps remains unchanged. Through the analysis of the attention layer, the weights of the feature maps can be redistributed, thereby giving higher weights to important features and suppressing irrelevant information such as noise, accelerating model training and improving model performance; the pooling layer has a pooling kernel size of 2, which downsamples the feature maps to half their original size.

[0040] (4b) The decoder consists of four basic units, with the following structure: encoder -> transposed convolution -> unit 1 -> transposed convolution -> unit 2 -> transposed convolution -> unit 3 -> transposed convolution -> unit 4. Each unit is composed of: input features -> residual concatenation 1 (input features, features of the corresponding unit in the encoder) -> channel attention layer -> convolutional layer -> residual concatenation 2 (features of residual concatenation 1, convolutional layer) -> channel attention layer. Each unit has a feature extraction function, further extracting high-level features and combining them with transposed convolution to reconstruct the aneurysm. The four units are concatenated together to form the decoder of this invention, realizing the reconstruction of intracranial aneurysms.

[0041] The layers of the decoder are configured as follows:

[0042] In unit n (n = 1, 2, 3, 4, 5), the convolutional layer has two identical convolutional operations. The number of convolutional windows in the convolutional operations is set to 32 * 2n - 1, the kernel size is 3 × 3, the stride is 1, and group normalization and ELU activation units are used. Residual concatenation aggregates and concatenates the input features and the convolutional output features. The number of input features is the sum of the input features and the convolutional output features. The channel attention layer is the same as the encoder. The number of convolutional windows in the transposed convolution is the same as the number of convolutional windows in the convolutional operations of the same unit. The kernel size is 2 × 2, the stride is 2, and group normalization and ELU activation units are used.

[0043] (4c) The voxel-level classifier is implemented using a 1×1 convolution kernel and a sigmoid activation function. The kernel size is 1×1, and the kernel window size is 1. The convolution operation in the voxel-level classifier fuses the reconstructed aneurysm feature map. The sigmoid function maps the voxel values ​​of the fused feature map to the range [0,1]. The mapped value represents the probability that the voxel is an aneurysm voxel. Based on the probability values ​​of the voxels, the actual aneurysm region and the background region can be obtained, thus achieving the segmentation of intracranial aneurysms.

[0044] (4d) Setting the loss function is also an important step in building the model. The loss function of the fully convolutional neural network set in this invention is as follows:

[0045]

[0046] Among them, y p and y t These represent the results of model segmentation and manually labeled results, respectively, where N represents the number of input samples. This represents the loss function value. The segmentation results obtained from model training are compared with the labels to obtain the loss function. Based on the loss function, the model is trained using backpropagation.

[0047] S5: Training a fully convolutional neural network model, specifically including the following steps:

[0048] S5a: The Xavier initialization method is applied to initialize the weights in the fully convolutional network;

[0049] S5b: Input the enhanced T1 region of interest image into a fully convolutional neural network;

[0050] S5c: The input image first enters the encoder, the encoder features are input into the decoder, and the decoder features are finally passed through the voxel classifier to achieve the segmentation of intracranial aneurysms;

[0051] S5d: Compare the model segmentation results with the corresponding manually labeled results and calculate the loss function;

[0052] S5e: Based on the loss function, an adaptive moment estimation optimizer is applied to update the weight parameters of the encoder, decoder, and voxel classifier; through forward and backward propagation, the parameters are continuously optimized and updated to obtain the optimal model of the finally trained fully convolutional neural network.

[0053] S6: Obtain the segmentation results from the model, which includes the following steps:

[0054] S6a: Input the T1 test set sample image set into the optimal model of the trained fully convolutional neural network, and extract the feature map of the intracranial aneurysm input data through the encoder;

[0055] S6b: Input the feature map of the test sample into the decoder, further extract high-level features and reconstruct the feature map to obtain the reconstructed feature map;

[0056] S6c: The reconstructed feature map is input into the voxel-level classifier. First, the high-dimensional features are fused by a convolution with a kernel of 1. Then, the sigmoid function is applied to map the feature map to the range [0,1], where the value represents the probability of belonging to an aneurysm.

[0057] S6d: By setting a probability threshold, the segmentation result of intracranial aneurysm is obtained, thus realizing the segmentation of intracranial aneurysm.

[0058] S7: Post-processing yields the optimal segmentation result. Based on the application of a fully convolutional model to segment aneurysms, this invention performs post-processing on the segmentation result, removing false positive regions by setting a volume threshold to obtain the final optimal segmentation result.

[0059] First, connected component analysis is performed on the model segmentation results. The volume of each connected component is calculated, and a volume threshold is set. Connected components with volumes smaller than the threshold are designated as background, and their corresponding voxels are set to 0. This volume thresholding method effectively removes false positives, further improving the segmentation results.

[0060] Secondly, the present invention also provides a system for automatically identifying and segmenting intracranial aneurysms in a 3.0T high-resolution MRI T1 sequence. The system includes a data preprocessing module, an image storage module, and an image recognition module, which are communicatively connected.

[0061] The data preprocessing module is used to preprocess the acquired intracranial aneurysm MRI images;

[0062] The image storage module is used to store the image processed by the preprocessing module;

[0063] The image recognition module includes a fully convolutional neural network model and is used to recognize intracranial aneurysm MRI images obtained after data preprocessing.

[0064] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0065] First, the method of the present invention extracts high-dimensional features of intracranial aneurysms in T1 sequences through convolution operations in the encoding path of a fully convolutional neural network; through convolution and deconvolution operations in the decoding path of the network, it further extracts higher-dimensional features and reconstructs the features, removes irrelevant background voxels in the T1 image, retains voxels related to intracranial aneurysms, and achieves accurate identification and segmentation of intracranial aneurysms.

[0066] Secondly, existing technologies for intracranial aneurysm identification and segmentation are all based on imaging techniques such as DSA, CTA, and MRA, with almost no technology or method designed based on T1 sequences. Compared to these other three methods, T1 sequence examination has the advantages of being non-invasive, radiation-free, and having the lowest detection time and cost. Furthermore, T1 sequence is the most widely used examination method; almost all brain disease examinations include T1 sequences, and it is almost impossible to omit a patient's T1 examination due to patient complaints or physician experience. Therefore, T1 sequence-based intracranial aneurysm detection has the widest applicability but is also the most challenging, and currently, no technology is designed specifically for T1 sequences. This invention addresses the current lack of technologies and methods for detecting intracranial aneurysms based on T1 sequences by proposing a method for automatically identifying and segmenting intracranial aneurysms in 3.0T high-resolution MRI T1 sequences, achieving accurate segmentation.

[0067] Third, existing technologies only segment intracranial aneurysms using sequence data such as DSA, CTA, and MRA. This invention, even based on non-contrast-enhanced T1 sequences, achieves segmentation results comparable to DSA, CTA, and MRA. A review of the literature on intracranial aneurysms reveals that the segmentation index, the Dice coefficient, is even superior to the segmentation results of existing technologies.

[0068] Fourth, this invention uses a simple volume thresholding method to remove false positives from the segmentation results of the fully convolutional network, achieving more accurate segmentation with almost no increase in segmentation time. Compared with existing technologies, the number of false positives in this invention is relatively low. Attached Figure Description

[0069] Figure 1 This is a flowchart of the present invention;

[0070] Figure 2 This is a flowchart of the adaptive region of interest selection method of the present invention;

[0071] Figure 3 This is a visualization of the test sample set segmentation results provided in the embodiments of the present invention;

[0072] Figure 4 This is a scatter plot showing the correlation between the volume of manually labeled regions and the volume of model segmented regions on the test sample set provided in this embodiment of the invention. Detailed Implementation

[0073] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0074] like Figure 1 As shown in this embodiment, a method for automatically identifying and segmenting intracranial aneurysms in a 3.0T high-resolution MRI T1 sequence is as follows: High-dimensional features of intracranial aneurysms in the T1 sequence are extracted through convolution operations in the encoding path of a fully convolutional neural network; higher-dimensional features are further extracted and reconstructed through convolution and deconvolution operations in the decoding path of the network; irrelevant background voxels in the T1 image are removed, and voxels related to the intracranial aneurysm are retained. This achieves the identification and segmentation of intracranial aneurysms. Specifically, the method includes the following steps:

[0075] S1: Obtain high-resolution data from T1:

[0076] The method described in this invention involves analyzing high-resolution MRI T1 sequences to achieve intracranial aneurysm segmentation.

[0077] The T1 sequence was acquired using MRI with the following parameters: head scan using a head coil; slice thickness of 0.6 mm; slice spacing of 0.3 mm; pixel spacing of 0.2976 mm; flip angle of 90°; repetition time of 800 ms; echo time of 18.722 ms; acquisition time of 1.7 min; and matrix of 672×672. High-resolution T1 data was obtained using these parameters.

[0078] S2: Select training and testing sample sets

[0079] Four-fifths of the T1 high-resolution dataset is selected as the test set for model training, and the remaining one-fifth is used as the test set to test the model.

[0080] S3: Different preprocessing methods are applied to the training and test sets, specifically as follows:

[0081] (3a) Preprocessing of the training set

[0082] (3a1) The original T1 data is trimmed by intensity values ​​to obtain the trimmed T1 data. The trimming formula is as follows:

[0083]

[0084] Where f'(x,y,z) and f(x,y,z) represent the intensity values ​​of the cropped and original images, respectively.

[0085] (3a2) Adaptive Region of Interest (ROI) Selection Method: Due to the large amount of video memory and other resources consumed by 3D data, it is necessary to reduce the image size during training. A common method is to extract a 3D image block sufficient to contain the intracranial aneurysm and its boundaries, based on artificially marked intracranial aneurysms. However, since patients may have multiple aneurysms, traditional extraction methods can lead to the severing of surrounding aneurysms, affecting model iteration. This invention proposes an adaptive RIO selection method to extract the region of interest (ROI) of the aneurysm that has not been severed. The algorithm flowchart is attached. Figure 2 As shown.

[0086] (3a3) Standardize the extracted regions of interest. Standardization can accelerate model convergence. The standardization process is as follows:

[0087]

[0088] Where f(x,y,z) norm f(x,y,z) represents the intensity values ​​of the region of interest before and after standardization. μ and f σ These are the mean and standard deviation of the intensity values ​​in the region of interest, respectively.

[0089] (3a4) Data augmentation is a method to prevent overfitting in deep learning models and also has a positive effect on convolutional neural networks. Therefore, this invention uses random flipping, rotation, and elastic transformation to process the region of interest image, thereby alleviating the overfitting of the convolutional neural network through image augmentation.

[0090] (3b) Preprocessing of the test set

[0091] (3b1) Trim the intensity values ​​of the original T1 data.

[0092] (3b2) The T1 data used in this invention is high-resolution data, resulting in a very large three-dimensional image size. Intracranial aneurysms grow within arteries, and there are no aneurysms in the image background or around the skull. Therefore, the aneurysm size is cropped to reduce the image size. The test data and training data of the fully convolutional network used in this invention do not need to be the same size; their three-dimensional dimensions only need to be divisible by 16 (4 pooling layers, downsampling by 16 times). The number of layers in the data often does not meet this condition. This invention increases the number of layers by adding black images at both ends, or reduces the number of layers by removing the images at both ends, ensuring that the transformed number of layers meets the requirements.

[0093] (3b3) Standardize the image after it has been cropped to the correct size.

[0094] The T1 data used in this invention is high-resolution data, resulting in a very large three-dimensional image size. Intracranial aneurysms grow within arteries, and are not present in the image background or around the skull. Therefore, the aneurysm size is cropped to reduce the image size. The test and training data sizes for the fully convolutional network used in this invention do not need to be the same; their three-dimensional dimensions only need to be divisible by 16 (4 pooling layers, downsampling by 16x). The number of layers in the data often does not meet this condition. This invention increases the number of layers by adding black images at both ends, or reduces the number of layers by removing the images at both ends, ensuring the transformed number of layers meets the requirements. The calculation method for the transformed number of layers is as follows:

[0095]

[0096] Among them, s t and s o These represent the number of layers in the data before and after the transformation, respectively. This is for rounding down. If s t Less than s o Then, delete images at both ends of the image to reduce the layer number s. t equals s o If s t Greater than s o Then, black images are added to both ends of the image to make the number of layers s. t equals s o The T1 data transformed using this layer transformation method will not have any aneurysm cut off or lost, ensuring that complete data containing the aneurysm is input into the model for testing.

[0097] S4: Establish a fully convolutional neural network model:

[0098] The network consists of three parts: an encoder, a decoder, and a voxel-level classifier. The encoder is mainly used to extract intracranial aneurysm features, the decoder is used to further extract high-level features and reconstruct the aneurysm, and the voxel-level classifier fuses the reconstructed aneurysm feature maps and gives the voxel category through the Sigmoid function to achieve aneurysm segmentation.

[0099] (4a) The encoder consists of 5 basic units, with the following structure: Unit 1 -> Pooling Layer -> Unit 2 -> Pooling Layer -> Unit 3 -> Pooling Layer -> Unit 4 -> Pooling Layer -> Unit 5. Each unit is composed of: Input Features -> Convolutional Layer -> Residual Concatenation (Input Features, Convolutional Layer) -> Channel Attention Layer. Each unit has feature extraction capabilities; residual concatenation can handle gradient vanishing issues; and the channel attention layer can select effective features. The 5 units are concatenated together to form the encoder of this invention, enabling the extraction of intracranial aneurysm features.

[0100] The encoder layers are configured as follows:

[0101] Unit n (n = 1, 2, 3, 4, 5) contains two identical convolutional layers. The number of convolutional windows in each layer is set to 32 * 2^n - 1, the kernel size is 3 × 3, and the stride is 1. Group normalization and ELU activation units are used. Residual concatenation aggregates and concatenates the input features with the convolutional output features. The number of input features is the sum of the input features and the convolutional output features. The specific implementation of the channel attention layer is as follows:

[0102] f a =l2(l1(Avg(f) o )))

[0103] f b =l2(l1(Max(f) o )))

[0104] f out =f o *(sigmoid(add(f a ,f b )))

[0105] Where Avg and Max represent global average pooling and global max pooling, respectively; add and sigmoid represent voxel-by-voxel addition and sigmoid operation of the feature maps, respectively; * represents voxel-by-voxel multiplication of the feature maps. l1 and l2 are two fully connected layers with different numbers of neurons, f a and f b For the features mapped by the fully connected layer, f o and f out The input and output features are kept constant, and the number of feature maps remains unchanged. Through the analysis of the attention layer, the weights of the feature maps can be redistributed, thereby giving higher weights to important features and suppressing irrelevant information such as noise, accelerating model training and improving model performance; the pooling layer has a pooling kernel size of 2, which downsamples the feature maps to half their original size.

[0106] (4b) The decoder consists of four basic units, with the following structure: Encoder -> Transposed Convolution -> Unit 1 -> Transposed Convolution -> Unit 2 -> Transposed Convolution -> Unit 3 -> Transposed Convolution -> Unit 4. Each unit is composed of: Input Features -> Residual Concatenation 1 (Input Features, features of the corresponding unit in the encoder) -> Channel Attention Layer -> Convolutional Layer -> Residual Concatenation 2 (Features of Residual Concatenation 1, Convolutional Layer) -> Channel Attention Layer. Each unit has a feature extraction function, further extracting high-level features and combining them with the transposed convolution to reconstruct the aneurysm. The four units are concatenated together to form the decoder of this invention, enabling the reconstruction of intracranial aneurysms.

[0107] The layers of the decoder are configured as follows:

[0108] In unit n (n = 1, 2, 3, 4, 5), the convolutional layer has two identical convolutional operations. The number of convolutional windows in the convolutional operations is set to 32 * 2n - 1, the kernel size is 3 × 3, the stride is 1, and group normalization and ELU activation units are used. Residual concatenation aggregates and concatenates the input features and the convolutional output features. The number of input features is the sum of the input features and the convolutional output features. The channel attention layer is the same as the encoder. The number of convolutional windows in the transposed convolution is the same as the number of convolutional windows in the convolutional operations of the same unit. The kernel size is 2 × 2, the stride is 2, and group normalization and ELU activation units are used.

[0109] (4c) The voxel-level classifier is implemented using a 1×1 convolution kernel and a sigmoid activation function. The kernel size is 1×1, and the kernel window size is 1. The convolution operation in the voxel-level classifier fuses the reconstructed aneurysm feature map. The sigmoid function maps the voxel values ​​of the fused feature map to the range [0,1]. The mapped value represents the probability that the voxel is an aneurysm voxel. Based on the probability values ​​of the voxels, the actual aneurysm region and the background region can be obtained, thus achieving the segmentation of intracranial aneurysms.

[0110] (4d) Setting the loss function is also an important step in building the model. The loss function of the fully convolutional neural network set in this invention is as follows:

[0111]

[0112] Among them, y p and y t These represent the results of model segmentation and manually labeled results, respectively, where N represents the number of input samples. This represents the loss function value. The segmentation results obtained from model training are compared with the labels to obtain the loss function. Based on the loss function, the model is trained using backpropagation.

[0113] S5: Training a fully convolutional neural network model:

[0114] (5a) The weights in the fully convolutional network are initialized using the Xavier initialization method.

[0115] (5b) Input the enhanced T1 region of interest image into a fully convolutional neural network.

[0116] (5c) The input image first enters the encoder, the encoder features are input into the decoder, and the decoder features are finally passed through the voxel classifier to achieve the segmentation of intracranial aneurysms.

[0117] (5d) Compare the model segmentation results with the corresponding manually labeled results and calculate the loss function. Based on the loss function, apply the adaptive moment estimation optimizer to update the weight parameters of the encoder, decoder, and voxel classifier.

[0118] (5e) The parameters are continuously optimized and updated through forward and backward propagation to obtain the optimal model of the finally trained fully convolutional neural network.

[0119] S6: Obtain the segmentation results from the model:

[0120] (6a) Input the standardized test sample set into the optimal model of the trained fully convolutional neural network, and extract the feature map of the intracranial aneurysm input data through the encoder.

[0121] (6b) Input the feature map of the test sample into the decoder, further extract high-level features and reconstruct the feature map to obtain the reconstructed feature map.

[0122] (6c) The reconstructed feature map is input into the voxel-level classifier. First, the high-dimensional features are fused by a convolution with a kernel of 1. Then, the sigmoid function is applied to map the feature map to the range [0,1], where the value represents the probability of belonging to an aneurysm.

[0123] (6d) By setting a probability threshold, the segmentation result of intracranial aneurysm is obtained, thus realizing the segmentation of intracranial aneurysm.

[0124] S7: Post-processing yields the optimal segmentation result:

[0125] This invention, based on the application of a fully convolutional model for aneurysm segmentation, performs post-processing on the segmentation results. False positive regions are removed by setting a volume threshold, resulting in the final optimal segmentation result. First, connected component analysis is performed on the model segmentation results, calculating the volume of each connected region. A volume threshold is set; connected regions smaller than the threshold are designated as background, and their corresponding voxels are set to 0. This volume threshold method effectively removes false positives, further improving the segmentation results.

[0126] The present invention will be further described below with reference to specific experiments, and the results of the invention will be compared with those of methods in the prior art.

[0127] The high-resolution T1 sequence acquisition method used in the experiment of this invention is the same as described in the aforementioned patent method, containing a total of 141 data points. 113 cases were randomly selected as the training sample set, and 28 cases were selected as the test sample set. The specific information of the 141 patients is as follows:

[0128]

[0129] The experimental environment for the algorithm of this invention was as follows: the operating system was Ubuntu 20.04; the processor was an AMD Ryzen Threadripper PRO 3995WX 64-Cores with a main frequency of 2.7GHz; the graphics card was an RTX 8000; the Python version was 3.6; and the TensorFlow version was 2.4.0.

[0130] To verify the model's recognition capability, this invention quantifies the segmentation results, referencing existing literature on intracranial aneurysm segmentation. Two methods are primarily used in the quantification process: first, an intracranial aneurysm is considered identified when the model's segmentation result overlaps with the manually labeled data; second, an intracranial aneurysm is considered identified when the overlap between the model's segmentation result and the manually labeled data exceeds 30%. This invention employs the second method: when the number of voxels in the aneurysm segmented by this invention that overlap with the manually labeled data exceeds 30%, the aneurysm is considered successfully identified. The test results are as follows: the sensitivity for aneurysm recognition is 90.3%, the false positive rate is 0.714 FPs / case, and the average test time is 6 seconds / case.

[0131] The identification results are compared with those of six existing methods, as shown in Table 1.

[0132] Table 1 Comparison of the recognition capabilities of the present invention's method and existing methods

[0133]

[0134]

[0135] As shown in Table 1, currently no other method for identifying intracranial aneurysms has achieved identification in T1 sequences like this invention. This invention is the first to achieve identification in T1 sequences. Furthermore, the identification results of this invention are comparable to, and even superior to, most other methods, achieving accurate identification in challenging sequences.

[0136] This invention verifies the segmentation results of the model by comparing them with manually labeled results. A visualization of the segmentation results versus manually labeled results is attached. Figure 3 As shown.

[0137] The segmentation results of the method of this invention are as follows: the average Dice coefficient is 0.769, and the median Dice coefficient is 0.867. In the test data, when the intracranial aneurysm volume is greater than 100 mm... 3 At that time, the average Dice coefficient was 0.84, and the median Dice coefficient was 0.912. To further verify the segmentation results of the method of the present invention, the volume of the segmented region by the method of the present invention was compared with the volume of the expert-marked region, and a scatter plot was drawn, as shown in the attached figure. Figure 4 As shown.

[0138] The scatter plot results of the method of the present invention and the manual method are as follows: R2 is 0.9432, p<0.0001; the consistency correlation coefficient is 0.9454; and the Pearson coefficient is 0.9712, which proves that the method of the present invention and the manual method have a high degree of consistency.

[0139] Table 2 shows a comparison of the segmentation results with those of the three existing methods.

[0140] Table 2 Comparison of segmentation accuracy between the method of the present invention and existing methods

[0141]

[0142]

[0143] As shown in Table 2, no current method for intracranial aneurysm identification achieves segmentation in T1 sequences like this invention. This invention is the first to achieve segmentation in T1 sequences. Furthermore, the method of this invention outperforms other current aneurysm segmentation methods in other sequences, achieving accurate segmentation even in challenging sequences.

[0144] In summary, the T1 sequence used in this invention has better universality than commonly used sequences and achieves excellent results with T1 sequences. Verification experiments of the method of this invention confirm its correctness and effectiveness.

[0145] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications and substitutions should be covered within the scope of the claims of the present invention. Technical aspects, shapes, and structures not described in detail in this invention are all well-known technologies.

Claims

1. A method for automatically identifying and segmenting intracranial aneurysms in a 3.0T high-resolution MRI T1 sequence, characterized in that, Includes the following steps: S1: Acquire high-resolution data for T1; S2: Select training and test sample sets from the T1 high-resolution data; S3: Perform training set preprocessing and test set preprocessing on the training sample set and test sample set respectively; S4: Establish a fully convolutional neural network model; S5: Train a fully convolutional neural network model; S6: Obtain the segmentation results from the model; S7: Post-processing yields the optimal segmentation result; The training set preprocessing includes the following steps: S3a1: Perform intensity value clipping on the original training sample set data to obtain the intensity value clipped training sample set data; S3a2: An adaptive region of interest selection method is used to extract three-dimensional image blocks of intracranial aneurysms from the training sample set data; S3a3: Standardize the extracted region of interest; S3a4: Perform data augmentation on the standardized image to obtain the T1 region of interest image; The test set preprocessing includes the following steps: S3b1: The intensity values ​​of the original test sample set data are clipped to obtain the intensity value-clipped test sample set data; S3b2: Crops the aneurysm images of the test sample set data after intensity value cropping to reduce the image size; S3b3: Standardize the aneurysm images after size cropping to obtain T1 test set sample images; The cutting formula for strength value cutting is as follows: in, and These represent the intensity values ​​of the cropped and original images, respectively. The convolutional neural network model in S4 includes an encoder, a decoder, and a voxel-level classifier, wherein... The encoder is used for the extraction of intracranial aneurysm features; The decoder is used to further extract high-level features and reconstruct the aneurysm; The voxel-level classifier fuses the reconstructed aneurysm feature maps and uses the Sigmoid function to give the voxel category, thereby achieving aneurysm segmentation. The encoder consists of 5 basic units, with the following structure: Unit 1 -> Pooling Layer -> Unit 2 -> Pooling Layer -> Unit 3 -> Pooling Layer -> Unit 4 -> Pooling Layer -> Unit 5. Each unit is composed of: Input Features -> Convolutional Layer -> Residual Concatenation -> Channel Attention Layer. Each unit has the function of feature extraction. Residual concatenation can deal with the vanishing gradient of the model, and the channel attention layer can select effective features. The 5 units are concatenated together to form the encoder, which realizes the extraction of intracranial aneurysm features. The encoder layers are configured as follows: Unit n (n=1,2,3,4,5) contains two identical convolutional layers. The number of convolutional windows in each layer is set to 32*(2n-1), the kernel size is 3×3, and the stride is 1. Group normalization and ELU activation units are used. Residual concatenation aggregates and concatenates the input features with the convolutional output features. The number of input features is the sum of the input features and the convolutional output features. The specific implementation of the channel attention layer is as follows: Where Avg and Max represent global average pooling and global max pooling, respectively; add and sigmoid represent voxel-by-voxel addition and sigmoid operation of the feature maps, respectively; and * represents voxel-by-voxel multiplication of the feature maps. and These are two fully connected layers with different numbers of neurons. and The features are those mapped after the fully connected layer. and The input and output features are kept constant, and the number of feature maps remains unchanged. Through the analysis of the attention layer, the weights of the feature maps can be redistributed, thereby giving higher weights to important features and suppressing noise, accelerating model training and improving model performance. The pooling kernel size of the pooling layer is 2, which downsamples the feature maps to half of their original size. The decoder consists of four basic units, with the following structure: encoder -> transposed convolution -> unit 1 -> transposed convolution -> unit 2 -> transposed convolution -> unit 3 -> transposed convolution -> unit 4. Each unit is composed of the following: input features -> residual concatenation 1 -> channel attention layer -> convolutional layer -> residual concatenation 2 -> channel attention layer. Each unit has a feature extraction function, further extracting high-level features and combining them with transposed convolution to reconstruct the aneurysm; the four units are spliced ​​together to form a decoder, realizing the reconstruction of intracranial aneurysms; The layers of the decoder are configured as follows: In unit n (n=1,2,3,4,5), the convolutional layer has two identical convolutional operations. The number of convolutional windows in the convolutional operations is set to 32*(2n-1), the kernel size is 3×3, the stride is 1, and group normalization and ELU activation units are used. Residual concatenation aggregates and concatenates the input features and the convolutional output features. The number of input features is the sum of the input features and the convolutional output features. The channel attention layer is the same as the encoder. The number of convolutional windows in the transposed convolution is the same as the number of convolutional windows in the convolutional operations of the same unit. The kernel size is 2×2, the stride is 2, and group normalization and ELU activation units are used. The voxel-level classifier is implemented by convolution with a kernel of 1 and a sigmoid activation function. The size of the convolution kernel is 1×1 and the number of convolution kernel windows is 1. The convolution operation in the voxel-level classifier fuses the reconstructed aneurysm feature map. The sigmoid function maps the voxel values ​​of the fused feature map to the range [0,1]. The mapped value represents the probability that the voxel is an aneurysm voxel. Based on the probability value of the voxel, the actual aneurysm region and the background region are obtained, thus realizing the segmentation of intracranial aneurysms. The loss function for a fully convolutional neural network is as follows: in, and These represent the results of model segmentation and manually labeled results, respectively, where N represents the number of input samples. The loss function value is obtained by comparing the segmentation results obtained from model training with the labels. Based on the loss function, the model is trained using backpropagation. S5 includes the following steps: S5a: The Xavier initialization method is applied to initialize the weights in the fully convolutional network; S5b: Input the enhanced T1 region of interest image into a fully convolutional neural network; S5c: The input image first enters the encoder, the encoder features are input into the decoder, and the decoder features are finally passed through the voxel classifier to achieve the segmentation of intracranial aneurysms; S5d: Compare the model segmentation results with the corresponding manually labeled results and calculate the loss function; S5e: Based on the loss function, an adaptive moment estimation optimizer is applied to update the weight parameters of the encoder, decoder, and voxel classifier; the parameters are continuously optimized and updated through forward and backward propagation to obtain the optimal model of the finally trained fully convolutional neural network. S6 includes the following steps: S6a: Input the T1 test set sample image set into the optimal model of the trained fully convolutional neural network, and extract the feature map of the intracranial aneurysm input data through the encoder; S6b: Input the feature map of the test sample into the decoder, further extract high-level features and reconstruct the feature map to obtain the reconstructed feature map; S6c: The reconstructed feature map is input into the voxel-level classifier. First, the high-dimensional features are fused by a convolution with a kernel of 1. Then, the sigmoid function is applied to map the feature map to the range [0,1], where the value represents the probability of belonging to an aneurysm. S6d: By setting a probability threshold, the segmentation result of intracranial aneurysm is obtained, thus realizing the segmentation of intracranial aneurysm; S1 specifically refers to an MRI scan, which uses a head coil to scan the head; The ratio of the number of training samples to the number of test samples in the T1 high-resolution data is 7:

3.

2. A system for automatically identifying and segmenting intracranial aneurysms in a 3.0T high-resolution MRI T1 sequence according to claim 1, comprising a data preprocessing module, an image storage module, and an image recognition module, wherein, The data preprocessing module is used to preprocess the acquired intracranial aneurysm MRI images; The image storage module is used to store the image processed by the preprocessing module; The image recognition module includes a fully convolutional neural network model and is used to recognize intracranial aneurysm MRI images obtained after data preprocessing.

Images