A cerebrovascular image segmentation method based on multi-scale attention network

Abstract

A cerebrovascular medical image segmentation method based on a multi-scale attention network, first pre-processing the original cerebrovascular MRA image, and then training a multi-scale attention UNet network; the encoding part of the network model extracts the features of the cerebrovascular image through a multi-scale attention module to improve the learning ability of effective features; the decoding part fuses multi-scale features through skip connections to improve the accuracy of model segmentation; finally, the test data to be segmented is input into the trained model to obtain the segmentation result and perform three-dimensional reconstruction. The present invention takes into account the complex characteristics of cerebrovascular images in brain images, and specifically proposes a network model for the segmentation and three-dimensional reconstruction of cerebrovascular images, which solves the segmentation problems caused by the high complexity of cerebral blood vessels and different positions and morphologies, and improves the segmentation accuracy of MRI cerebrovascular images.

Description

Technical Field

[0001] The invention relates to a cerebral blood vessel image segmentation method. Background Art

[0002] Vascular diseases have become a major threat to human health due to their high morbidity and mortality. The cerebrovascular system is a complex network of arteries and veins that provide vital nutrients and oxygen to brain cells. In addition, individual differences between brain vessels at a more refined level are still not fully understood. Automatic and accurate cerebrovascular segmentation is a key step in cerebrovascular structure visualization, vascular diagnosis, and computer-assisted intervention.

[0003] Current medical angiography techniques, such as computed tomography angiography (CT), digital subtraction angiography, and magnetic resonance angiography (MRA), provide a large amount of data in different ways and have greatly helped the study of cerebral vascular disease. Time-of-Flight magnetic resonance angiography (TOF-MRA) is based on blood flow or inflow angiography, which uses short echo time and flow compensation to make flowing blood much brighter than stationary tissue. Head TOF-MRA imaging is used for arterial angiography (while some venous imaging results are usually included in data acquisition) to evaluate vascular abnormalities such as stenosis, aneurysm, and arteriovenous anomalies. However, due to the complexity of cerebral vascular branches, there are not only many tiny capillaries in cerebral vascular MRA, but also a thick internal carotid artery (ICA). This requires that segmentation targets of different scales be considered in the task of cerebral vascular segmentation. At the same time, during the digital subtraction angiography imaging process, the concentration of the contrast agent will continue to change and diffuse unevenly, which may cause artifacts and noise in the cerebral vascular digital subtraction angiography images. Due to the above characteristics, how to automatically and accurately segment cerebral vessels from MRA images becomes a challenging problem.

[0004] However, it is time-consuming and impractical for medical experts to manually segment and label cerebrovascular images in large-scale related studies. Although many different automatic vascular segmentation methods have been developed for different medical imaging modalities in the past two decades, cerebrovascular segmentation is still a task that only a few researchers have explored. Summary of the invention

[0005] The present invention aims to overcome the above-mentioned shortcomings of the prior art. In order to improve the performance of cerebrovascular segmentation in brain medical images, an automatic cerebrovascular image segmentation method based on a multi-scale attention mechanism network model is proposed.

[0006] The present invention provides a method for automatic segmentation of cerebrovascular medical images based on a multi-scale attention network, comprising the following four steps: cerebrovascular medical image preprocessing, multi-scale attention UNet (MSAUNet) network model design, network model training, segmentation and three-dimensional reconstruction.

[0007] 1) Preprocess the cerebrovascular medical images. First, correct and normalize the image intensity to [0,255] through the n4 bias field, then resample the image to make the layer thickness uniform, and then fill the surrounding 0 operation to make the image voxel size H×W×D satisfy H, W, D are all integer multiples of 32, and finally cut the window position according to the image intensity characteristics.

[0008] 2) Design a multi-scale attention UNet network model. In order to improve the performance of cerebrovascular segmentation, the present invention proposes a MSAUNet model, which is similar to the topological structure of UNet. This model has an encoder and a decoder, and uses skip connections to combine the features of the encoder stage with the features corresponding to the decoder, where MSAB is a multi-scale attention module and MSFB is a multi-scale feature fusion module. In the encoding stage, MSAB is used to extract multi-scale features, and four pooling operations are used to increase the receptive field of the features and extract high-level features with strong semantic representation capabilities. In the decoding stage, while the upsampling operation is used to restore the size of the high-level features, MSFB is used to extract high-resolution features with stronger semantic representation capabilities in combination with the low-level features of the encoding stage. Finally, a 1×1 convolution operation and Sigmoid normalization are used to achieve the mapping from features to pixel categories, and a segmentation mask consistent with the size of the input image is obtained.

[0009] The implementation of multi-scale and attention is as follows:

[0010] (2.1) Extract multi-scale information features of the image. This paper adopts a serial extraction method to reduce memory consumption. Two 3×3 convolutions can extract features with receptive fields of 3×3 and 5×5. Similarly, three 3×3 convolutions can replace a 7×7 convolution block. Finally, the module can combine the contextual information of 3×3, 5×5 and 7×7 regions to obtain feature representations of multiple receptive fields.

[0011] (2.2) Use the attention mechanism to add different weights to multi-scale information. In the simple linear aggregation method of features of different scales, for different targets, the features of each receptive field are treated equally. However, for the complex and variable cerebrovascular segmentation target, different input images have different sensitivities to the features of different receptive fields, which limits the generalization ability of the model to a certain extent. In MSAB, we extract two low-scale features and two high-scale features through convolution and concatenate them on the channel to obtain two preliminary fused feature information of different scales, and then pass them through a series of channel attention modules (CAM) and a spatial attention module (SAM), and finally add the two high- and low-scale features at the pixel level to obtain the final output. We use two independent attention modules in series to reduce the number of network parameters, and perform channel and spatial attention weighting on the cerebrovascular image features. The calculation of the channel attention module is shown in formula (1), and the calculation of the spatial attention is shown in formula (2).

[0012] M C (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) (1)

[0013] M s (F) = σ(f 5×5 ([AvgPool(F);MaxPool(F)])) (2)

[0014] Among them, F represents the input feature, AvgPool is the average pooling, MaxPool is the maximum pooling, the two MLPs are a weight-shared multi-layer perceptron, σ represents sigmoid, and [f1; f2] represents the concat operation of f1 and f2 features.

[0015] 3) Train the MSAUNet network model proposed in step 2. The loss function uses the general Dice loss function:

[0016]

[0017] 4) Finally, the test data to be segmented is input into the trained network model to obtain the segmentation result and perform three-dimensional reconstruction.

[0018] Taking into account the characteristics of cerebral blood vessels in brain images, such as high complexity and different positions and morphologies, the present invention specifically proposes an MSAUNet network model based on multi-scale feature extraction and attention mechanism to perform segmentation and three-dimensional reconstruction of cerebral vascular images.

[0019] The advantages of the present invention are: solving the segmentation problem caused by the high complexity and different positions and shapes of brain blood vessels, and improving the segmentation accuracy of MRI cerebral blood vessel images. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is the overall process of the method of the present invention.

[0021] Figure 2 It is a schematic diagram of the overall framework of the model of the present invention.

[0022] Figure 3 It is a structural diagram of the multi-scale attention module MSAB of the present invention. DETAILED DESCRIPTION

[0023] The present invention will be further described below in conjunction with the accompanying drawings.

[0024] Figure 1 The overall process of the cerebrovascular image segmentation method based on the multi-scale attention network proposed in the present invention is as follows. The cerebrovascular medical image segmentation method based on the multi-scale attention network proposed in the present invention mainly includes the following steps: preprocessing, designing the network model, MSAUNet training and testing segmentation and 3D reconstruction. The overall structure of the model is as follows Figure 2 The present invention provides a cerebrovascular image segmentation method based on a multi-scale attention network, and the specific steps are as follows:

[0025] 1) Preprocess the original image;

[0026] This step first uses n4 bias field calibration processing, and the obtained image pixel intensity is normalized to [0,255]. In view of the problem of different image layer thicknesses and discontinuity in existing label data, the present invention resamples the original MRI image and label image with a size of 448x448x128 voxels and a layer thickness of 0.51mm x 0.51mm x 0.80mm, so that the layer thickness is unified to 1mm x 1mm x 1mm. After resampling, the image and its label size we obtained are 460x460x205 voxels. The resampled image is padding-filled with 0s around to obtain data with a size of 480x480x224 voxels. Finally, the data is cut with a window position of [25,75] to obtain the final input image.

[0027] 2) We designed a multi-scale attention U-shaped network model. The network design is as follows: (1) Based on the morphological characteristics of cerebrovascular vessels in the entire image, we applied the multi-scale idea in the encoding part and extracted the vascular features at different scales to effectively deal with the complex vascular problems. (2) We applied the attention mechanism, which enables the network to pay more attention to the small cerebrovascular part that occupies the entire image by connecting channel attention and spatial attention in series. Finally, we constructed the multi-scale attention module MSAB, whose structure is shown in the figure below. Figure 3 shown.

[0028] 3) Send the preprocessed input image to MSAUNet for training.

[0029] During the training phase, the learning rate was set to 0.0001 and the Adam optimizer was selected. A total of 80 epochs were trained, and the learning rate was reduced by 10 times every 20 epochs. During this period, a batch size of 8 samples was used for gradient update, and the loss function used the general Dice loss function.

[0030] 4) Finally, the test data to be segmented is input into the trained model to obtain the segmentation result and the slice data is stacked for 3D reconstruction.

[0031] The contents described in the embodiments of this specification are merely an enumeration of implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for automatic segmentation of cerebrovascular medical images based on a multi-scale attention network, comprising the following steps: 1) Preprocessing of cerebrovascular medical images: First, the image intensity is normalized to [0, 255] by using the n4 bias field correction, then the image is resampled to make the layer thickness uniform, and then the image voxel size H×W×D is filled with zeros to satisfy the requirement that H, W, and D are all integer multiples of 32, and finally the window position is cut according to the image intensity characteristics; 2) Design a multi-scale attention UNet network model; In order to improve the performance of cerebrovascular segmentation, a MSAUNet model is proposed. Similar to the topological structure of UNet, the MSAUNet model has an encoder and a decoder, and uses skip connections to combine the features of the encoder stage with the corresponding features of the decoder. MSAB is a multi-scale attention module, and MSFB is a multi-scale feature fusion module. In the encoding stage, MSAB is used to extract multi-scale features, and four pooling operations are used to increase the receptive field of the features and extract high-level features with strong semantic representation capabilities; In the decoding stage, the upsampling operation is used to restore the size of high-level features, and at the same time, the low-level features in the encoding stage are combined to extract high-resolution features with stronger semantic representation capabilities using MSFB; Finally, a 1×1 convolution operation and Sigmoid normalization are used to map features to pixel categories, and a segmentation mask with the same size as the input image is obtained. The implementation of multi-scale and attention is as follows: (2.1) Extract multi-scale information features of the image; using serial extraction can reduce memory consumption. Two 3×3 convolutions can be used to extract features with receptive fields of 3×3 and 5×5. Similarly, three 3×3 convolutions can replace a 7×7 convolution block. Finally, the module can combine the contextual information of 3×3, 5×5 and 7×7 regions to obtain feature representations of multiple receptive fields. (2.2) The attention mechanism is used to add different weights to multi-scale information. In the simple linear aggregation method of features of different scales, the features of each receptive field are treated equally for different targets. However, for the complex and variable cerebrovascular segmentation target, different input images have different sensitivities to the features of different receptive fields, which limits the generalization ability of the model. In MSAB, the two low-scale features and the two high-scale features extracted by convolution are concat on the channel to obtain two preliminary fused feature information of different scales, and then a series of channel attention modules CAM and a spatial attention module SAM are respectively passed through. Finally, the two high- and low-scale features are added at the pixel level to obtain the final output. By using two independent attention modules in series, the number of network parameters is reduced, and the cerebrovascular image features are weighted in terms of channel and space. The calculation of the channel attention module is shown in formula (1), and the calculation of the spatial attention is shown in formula (2). M C (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) (1) M s (F)=σ(f 5×5 ([AvgPool(F);MaxPool(F)])) (2) Among them, F represents the input feature, AvgPool is the average pooling, MaxPool is the maximum pooling, the two MLPs are a weight-sharing multi-layer perceptron, σ represents sigmoid, [f1; f2] represents the concat operation of f1 and f2 features; 3) Train the MSAUNet network model proposed in step 2); the loss function adopts the general Dice loss function: 4) Finally, the test data to be segmented is input into the trained network model to obtain the segmentation result and perform three-dimensional reconstruction.

2. The method for automatic segmentation of cerebrovascular medical images based on a multi-scale attention network as claimed in claim 1, characterized in that: Step 1) To address the problems of different image layer thicknesses and discontinuity in existing label data, the original MRI image and label image with a size of 448x 448x 128 voxels and a layer thickness of 0.51mm x 0.51mm x 0.80mm were resampled to make the layer thickness uniform to 1mm x 1mm x 1mm. After resampling, the size of the image and its label was 460x 460x 205 voxels. The resampled image was padded with 0s to obtain data with a size of 480x480x 224 voxels. Finally, the data was windowed with a window of [25,75] to obtain the final input image.

3. The method for automatic segmentation of cerebrovascular medical images based on a multi-scale attention network as claimed in claim 1, characterized in that: In step 3), the learning rate is set to 0.0001, and the Adam optimizer is selected; a total of 80 epochs are trained, and the learning rate is reduced by 10 times every 20 epochs. During this period, a batch size of 8 samples is used for gradient update, and the loss function adopts the Dice loss function.

Images