CN-115829914-B - Method for automatically identifying and segmenting intracranial aneurysm in 3.0T high-resolution MRI T1 sequence

Abstract

The invention relates to the technical field of medical image processing, in particular to a method for automatically identifying and segmenting an intracranial aneurysm in a 3.0T high-resolution MRI T1 sequence, which extracts the high-dimensional characteristic of the intracranial aneurysm in the T1 sequence through convolution operation in a coding path in a full convolution neural network; through decoding path convolution and deconvolution operation in the network, further extracting higher-dimensional features and reconstructing the features, removing irrelevant background voxels in the T1 image, and reserving voxels related to intracranial aneurysms to realize accurate identification and segmentation of the intracranial aneurysms.

Inventors

• ZHANG XU
• LIU AIHUA
• QU JUNDA
• NIU HAO
• PENG FEI
• LI CHUNLIN
• XIA JIAXIANG
• Xu boya

Assignees

• 首都医科大学
• 北京市神经外科研究所

Dates

Publication Date

20260512

Application Date

20220815

Claims (2)

1. 1. A method for automatically identifying and segmenting an intracranial aneurysm in a 3.0T high resolution MRI T1 sequence, comprising the steps of: S1, acquiring high-resolution data of T1; S2, selecting a training sample set and a test sample set from the T1 high-resolution data; s3, respectively carrying out training set pretreatment and test set pretreatment on the training sample set and the test sample set; S4, establishing a full convolution neural network model; S5, training a full convolution neural network model; s6, obtaining a segmentation result of the model; S7, post-processing to obtain an optimal segmentation result; the training set preprocessing comprises the following steps: S3a1, performing intensity value cutting on original training sample set data to obtain training sample set data subjected to intensity value cutting; S3a2, extracting the intracranial aneurysm three-dimensional image block in the training sample set data by adopting a self-adaptive region-of-interest selection method; s3a3, carrying out standardization treatment on the extracted region of interest; S3a4, carrying out data enhancement on the standardized image to obtain a T1 region of interest image; the test set preprocessing comprises the following steps: S3b1, performing intensity value cutting on the original test sample set data to obtain test sample set data subjected to intensity value cutting; S3b2, performing size cutting on the aneurysm image of the test sample set data after the intensity value cutting, and reducing the image size; S3b3, normalizing the aneurysm image after size cutting to obtain a T1 test set sample image; The clipping formula for clipping the intensity values is as follows: Wherein, the And Representing the intensity values of the cut and original images respectively; The convolutional neural network model in S4 comprises an encoder, a decoder, and a voxel level classifier, wherein, The encoder is used for extracting intracranial aneurysm characteristics; The decoder is used for further extracting high-level characteristics and reconstructing the aneurysm; The voxel level classifier fuses the reconstructed aneurysm feature images, and gives the category of voxels through a Sigmoid function so as to realize the segmentation of aneurysms; The encoder consists of 5 basic units, wherein the structure of the encoder sequentially comprises a unit 1- > a pooling layer- > a unit 2- > a pooling layer- > a unit 3- > a pooling layer- > a unit 4- > a pooling layer- > a unit 5, the composition of each unit comprises an input characteristic- > a convolution layer- > residual splicing- > a channel attention layer, each unit has a characteristic extraction function, residual splicing can cope with the condition of model gradient dispersion, and the channel attention layer can select effective characteristics; The layers of the encoder are set as follows: The convolution layer in the unit n (n=1, 2,3,4, 5) has two identical convolution operations, the number of convolution windows of the convolution operations is set to be 32 x (2 n-1), the convolution kernel size is 3 x 3, the convolution step size is 1, a group standardization and elu activation unit is adopted, the input features and the convolution output features are subjected to aggregation and splicing by residual splicing, the number of the input features is the combination of the input features and the convolution output features, and the channel attention layer is specifically realized as follows: Wherein Avg and Max respectively represent global average pooling and global maximum pooling, add and sigmoid respectively represent voxel-by-voxel corresponding addition and sigmoid operation of the feature map, the feature map is multiplied voxel-by-voxel, And For two fully connected layers with different numbers of neurons, And For features that are mapped by the fully connected layer, And The feature images can be subjected to weight redistribution through analysis of the attention layer, so that higher weight is given to important features, noise is suppressed, training of a model is accelerated, and performance of the model is improved; The decoder consists of 4 basic units, and the structure of the decoder sequentially comprises an encoder- > transposition convolution- > unit 1- > transposition convolution- > unit 2- > transposition convolution- > unit 3- > transposition convolution- > unit 4, wherein the composition of each unit is as follows, input characteristics- > residual splicing 1- > channel attention layer- > convolution layer- > residual splicing 2- > channel attention layer; Each unit has a feature extraction function, further extracts high-level features, and reconstructs an aneurysm in combination with transpose convolution, 4 units are spliced together to form a decoder, so that reconstruction of the intracranial aneurysm is realized; the layers of the decoder are set as follows: The convolution layer in the unit n (n=1, 2,3,4, 5) has two identical convolution operations, the number of convolution windows of the convolution operations is set to be 32 x (2 n-1), the convolution kernel size is 3 x 3, the convolution step size is 1, a group standardization and elu activation unit is adopted, the residual error splicing carries out aggregation splicing on the input features and the convolution output features, the number of the input features is the combination of the input features and the convolution output features, the channel attention layer is consistent with the encoder, the number of convolution windows of the transposed convolution is the same as the number of convolution windows of the convolution operations of the same unit, the convolution kernel size is 2 x 2, the convolution step size is 2, and the group standardization and elu activation unit is adopted; The voxel level classifier is realized by convolution with a convolution kernel of 1 and a sigmoid activation function, wherein the size of the convolution kernel is 1 multiplied by 1, the number of convolution kernel windows is 1, the convolution operation in the voxel level classifier fuses the reconstructed aneurysm feature images, the values of the voxels of the fused feature images are mapped between [0,1] through the sigmoid function, the mapped values represent the probability that the voxels are aneurysm voxels, the actual area and the background area of the aneurysm are obtained according to the probability values of the voxels, and the segmentation of intracranial aneurysms is realized; The loss function of the fully convolutional neural network is as follows: Wherein, the And The result of the model segmentation and the result of the artificial labeling, respectively, N representing the number of input samples, Comparing the segmentation result obtained by model training with the mark to obtain a loss function, and training the model by a back propagation method according to the loss function; The step S5 comprises the following steps: s5a, initializing weights in the full convolution network by using xavier initialization method; S5b, inputting the T1 region of interest image after image enhancement into a full convolution neural network; s5c, firstly, inputting an input image into an encoder, inputting the characteristics of the encoder into a decoder, and finally, dividing intracranial aneurysms by the characteristics of the decoder through a voxel classifier; S5d, comparing the model segmentation result with a corresponding artificial marking result, and calculating a loss function; S5e, updating weight parameters of the encoder, the decoder and the voxel classifier by using a self-adaptive moment estimation optimizer according to the loss function, and continuously optimizing and updating the parameters through forward and backward propagation to obtain an optimal model of the finally trained full convolutional neural network; the step S6 comprises the following steps: S6a, inputting a T1 test set sample image set into an optimal model of a trained full convolution neural network, and extracting a characteristic diagram of intracranial aneurysm input data through an encoder; S6b, inputting the feature map of the test sample into a decoder, further extracting high-level features and reconstructing the feature map to obtain a reconstructed feature map; S6c, inputting the reconstructed feature map into a voxel level classifier, firstly fusing high-dimensional features by convolution with a convolution kernel of 1, then mapping the feature map between [0,1] by applying a sigmoid function, and representing the probability of belonging to aneurysm by a numerical value; S6d, obtaining a segmentation result of the intracranial aneurysm by setting a probability threshold value, and realizing segmentation of the intracranial aneurysm; The method comprises the following steps that S1 specifically comprises MRI scanning, and a head coil is used for scanning the head; The number ratio of the training sample set to the test sample set of the T1 high-resolution data is 7:3.
2. 2. A system for performing the method for automatically identifying and segmenting an intracranial aneurysm in a 3.0T high resolution MRI T1 sequence as in claim 1, comprising a data preprocessing module, an image storage module and an image identification module, wherein, The data preprocessing module is used for preprocessing the obtained MRI image of the intracranial aneurysm; the image storage module is used for storing the image processed by the preprocessing module; The image recognition module comprises a full convolution neural network model and is used for recognizing the MRI image of the intracranial aneurysm obtained through processing by the data preprocessing module.

Description

Method for automatically identifying and segmenting intracranial aneurysm in 3.0T high-resolution MRI T1 sequence Technical Field The invention relates to the technical field of medical image processing, in particular to a method for automatically identifying and segmenting intracranial aneurysms in a 3.0T high-resolution MRI T1 sequence. Background Intracranial aneurysms are abnormal distensions on cerebral arteries, and rupture of the intracranial aneurysm is the leading cause of non-traumatic subarachnoid hemorrhage, a catastrophic event with very high residual and mortality rates from subarachnoid hemorrhage. Therefore, early discovery and intervention of intracranial aneurysms is highly necessary, and identification and segmentation of intracranial aneurysms has become a research hotspot in the field of medical image processing. The existing intracranial aneurysm identification and segmentation methods are based on invasive, radiation, high-cost or time-consuming enhanced image examination means such as DSA, CTA and MRA. The common method mainly realizes the intracranial aneurysm identification at the image block level based on the convolutional neural network and the voxel level segmentation based on the full convolutional neural network. Even if the result of dividing the intracranial aneurysm at the voxel level based on the advanced enhanced image inspection means is still not satisfactory, the method cannot be applied to clinic, and further optimization and improvement of the identification and division method are required. The imaging means based on DSA, CTA and MRA are not optimal ways to achieve intracranial aneurysm segmentation. The reasons are that DSA is an invasive inspection means and has strong radiation, CTA has strong radiation, MRA inspection time and cost are high. Meanwhile, due to the factors of patient complaints, misleading information, doctor experience and the like, the patient does not perform the examination, and the aneurysm is missed. Therefore, the segmentation and identification of intracranial aneurysms based on the DSA, CTA and MRA detection means have certain limitations, so that the detection of images based on relative universality and the realization of the identification and segmentation of intracranial aneurysms based on the images are the most significant. Disclosure of Invention In order to solve the technical problems, the invention aims to provide a method for automatically identifying and segmenting an intracranial aneurysm in a 3.0T high-resolution MRI T1 sequence, which aims at solving the problems that the existing method for identifying and segmenting the intracranial aneurysm is to be optimized and perfected, and the existing method for identifying and segmenting the intracranial aneurysm is not developed for a universal image inspection means. The introduction of more general MRI T1 sequences, a method for automatically identifying and segmenting intracranial aneurysms in 3.0T high resolution MRI T1 sequences is presented. In order to achieve the technical effects, the invention adopts the following technical scheme: 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, comprising the steps of: S1, acquiring high-resolution data of T1; the acquisition parameters of the T1 high-resolution data comprise head scanning of a scanning part, head coil use, scanning layer thickness of 0.6mm, layer spacing of 0.3mm, pixel spacing of 0.2976mm, turning angle of 90 degrees, repetition time of 800ms, echo time of 18.722ms, acquisition time of 1.7min and matrix 672×672. The head is scanned by the scanning parameters of the sequence, thereby obtaining T1 high resolution data. S2, selecting a training sample set and a test sample set from the T1 high-resolution data; specifically, 4/5 of the data from the high-resolution data of T1 obtained in the step S1 is selected as a training sample set for training the model, and the remaining 1/5 is used as a test sample set for testing the model. S3, respectively carrying out training set pretreatment and test set pretreatment on the selected training sample set and test sample set; further, the training set preprocessing includes the steps of: s3a1, performing intensity value cutting on the original training sample set data to obtain training sample set data subjected to intensity value cutting, wherein a cutting formula is as follows: where f' (x, y, z) and f (x, y, z) represent intensity values of the cropped and original images, respectively. When the voxel intensity value in the matrix is smaller than 0, the voxel intensity value is 0, when the voxel intensity value is between 0 and 2000, the voxel intensity value is unchanged, and when the voxel intensity value in the matrix is larger than 2000, the voxel intensity value is 2000. S3a2, extracting the intracranial aneurysm three-dimensional image block in the