CN112634285A

CN112634285A - Method for automatically segmenting abdominal CT visceral fat area

Info

Publication number: CN112634285A
Application number: CN202011542684.XA
Authority: CN
Inventors: 彭博; 左昊; 贾维; 张傲
Original assignee: Southwest Petroleum University
Current assignee: Southwest Petroleum University
Priority date: 2020-12-23
Filing date: 2020-12-23
Publication date: 2021-04-09
Anticipated expiration: 2040-12-23
Also published as: CN112634285B

Abstract

The invention discloses a method for automatically segmenting an abdominal CT visceral fat area, which comprises the following steps: selecting a clinical abdomen CT image as a data set; preprocessing an image in the data set; extracting an image of a final visceral fat region; constructing an Attention U-net network; training an Attention U-net network; preprocessing an abdomen CT image needing to be predicted, inputting the preprocessed image into a trained Attention U-net network, wherein an image output by the Attention U-net network is a segmentation image. The invention can accelerate the speed of abdominal CT visceral fat area segmentation, simplify the segmentation steps, perform batch segmentation, greatly improve the segmentation efficiency, simplify the early operation of abdominal visceral fat amount calculation, and provide a better basis for the subsequent visceral fat amount calculation.

Description

Method for automatically segmenting abdominal CT visceral fat area

Technical Field

The invention belongs to the technical field of medical image processing, and particularly relates to a method for automatically segmenting an abdominal CT visceral fat area.

Background

An increase in the body fat in general and an increase in abdominal fat, particularly in the abdominal cavity, in the human body may cause cardiovascular diseases such as diabetes, hyperlipidemia, hypertension, insulin resistance and hyperuricemia. Therefore, the quantitative analysis of the abdominal fat content has important clinical value for preventing and treating related cardiovascular diseases. Currently, the most common abdominal fat measurement is performed by CT, which can accurately quantify the volume of the Adipose Tissue (AT) of the human body, particularly the area AT such as subcutaneous fat area (SA) of the abdominal wall or visceral fat area (VA) in the abdominal cavity. AT present, the application range of the CT measurement AT technology has been related to a plurality of fields such as clinical nutrition, geriatric medicine, epidemiology, genetics, especially endocrine metabolism and cardiovascular system.

In the process of measuring AT by CT, a region of interest (ROI) needs to be drawn when measuring SA and VA, but in many previous studies, segmentation is performed by manual or conventional segmentation methods. As used on siemens CT workstation (VE40D), is a watershed-based segmentation method. The watershed segmentation method is used for the segmentation of the visceral fat, and when the abdominal wall muscle layer is thin or is less continuous, the visceral fat and the abdominal wall fat layer can be segmented together. At this time, manual adjustment is needed to delineate the visceral fat area, so the degree of automation is not enough, which results in low efficiency and is not suitable for developing general investigation.

Deep learning is currently applied in the field of segmentation of medical images. Long et al first proposed a full convolutional neural network model (FCN) that can achieve end-to-end image segmentation, extending image-level classification to pixel-level classification, and replacing the convolutional layer with a full connection layer in a classification network framework. However, the FCN results are not fine enough, do not take into account the pixel-to-pixel correlation, and are not sensitive to the details of the image. Ronneberger et al propose that a U-net network fuses high-level information and shallow-level information of an image by using a cascade operation between a decoder and an encoder, thereby avoiding the loss of high-level semantic information and achieving a good effect on segmentation tasks of some medical images. And the U-net network can effectively utilize the training data set, and reduce the pressure on sample requirements. Millettari et al propose to use a voxel-based full-convolution V-net network for the segmentation of three-dimensional medical images and to use a framework of residual networks on the part of the encoder to prevent gradient disappearance or gradient explosion while the network is continuously deepened. Zhao et al propose a pyramid scene parsing network (PSPNet). The PSPNet network obtains more context information by utilizing feature fusion on the basis of the FCN network, and then improves the obtaining capability of the global information by aggregating the context information of different areas.

Disclosure of Invention

The invention mainly overcomes the defects in the prior art and provides a method for automatically segmenting an abdominal CT visceral fat area.

The invention solves the technical problems, and the provided technical scheme is as follows: a method of automatically segmenting abdominal CT visceral fat regions, comprising:

s100, selecting clinical abdomen CT images of different age groups, different abdomen positions and different slice thicknesses as a data set;

s200, preprocessing an image in a data set to obtain a preprocessed data set image;

step S300, manually dividing a mask image of the visceral fat area according to the preprocessed data set image, and performing AND operation on pixel points corresponding to the divided mask image and the preprocessed data set image to extract a final visceral fat area image;

s400, constructing an Attention U-net network, and inputting the preprocessed data set image and the final image of the visceral fat area into the constructed Attention U-net network as training and predicting data;

s500, training an Attention U-net network according to the value of the loss function and the result of the precision;

step S600, preprocessing the abdominal CT image needing to be predicted, and then inputting the preprocessed image into a trained Attention U-net network, wherein the image output by the Attention U-net network is a segmentation image.

The method has the further technical scheme that in the step S100, midriff images of young and strong people and old people are selected, and the slice width is 1 mm; and selecting images of the upper abdomen and the lower abdomen of young and middle-aged people and old people, wherein the slice width is 5 mm.

The further technical scheme is that the specific process of the preprocessing in the step S200 is as follows: extracting pixel information in the DICOM file to obtain an original CT image, and performing image binarization on the original CT image to distinguish an visceral fat region from a background region.

The further technical scheme is that the gray value between the pixel values of 874-974 in the image binarization is set to be 255, and the gray value of the pixel value greater than 974 or less than 874 is set to be 0.

The further technical scheme is that the Attention U-net network in the step S500 comprises two parts, an encoding part and a decoding part;

the coding part comprises a 5-layer structure, the first 4-layer structure comprises two convolutional layers and a maximum pooling layer, the 5 th-layer structure comprises two convolutional layers, the convolutional layers use relu as an activation function, the size of a convolutional kernel is 3 x 3, the number of output channels of the first-layer structure is 32, and then the number of output channels of each-layer structure is 2 times that of the last layer;

the decoding part comprises 5 layers of structures, each layer of structure comprises a bilinear interpolation up-sampling structure, an attention structure, three layers of convolution layers and a splicing structure, and the convolution layers all use relu as an activation function; the convolution kernel size is 3 x 3.

The further technical solution is that the process of step S600 is: extracting the pixel information in the DICOM file, then carrying out image binarization according to the preprocessing method in the step S200, inputting the image subjected to image binarization into a trained Attention U-net network, wherein the image output by the Attention U-net network is a segmentation image.

The invention has the following beneficial effects: the method can accelerate the speed of abdominal CT visceral fat area segmentation, simplify the segmentation steps, perform batch segmentation, and does not need manual adjustment, thereby greatly improving the segmentation efficiency, simplifying the early operation of abdominal visceral fat amount calculation, and providing a better basis for the subsequent visceral fat amount calculation.

Drawings

FIG. 1 is a flow chart of an embodiment of the present invention;

FIG. 2 is a diagram of an Attention U-net network model;

FIG. 3 is an attention mechanism module;

FIG. 4 is a graph comparing results of visceral fat segmentation;

fig. 5 is a process diagram of CT image processing.

Detailed Description

The technical solutions of the present invention will be described clearly and completely with reference to the accompanying drawings, and it should be understood that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

As shown in fig. 1, a method for automatically segmenting abdominal CT visceral fat region according to the present invention comprises:

s100, selecting midriff images of young and strong people (age is 30 +/-10, average value is +/-10) and old people (60 +/-10), wherein the slice width is 1 mm; selecting images of the upper abdomen and the lower abdomen of young and middle-aged people and old people, and taking the section width of 5mm as a data set;

the specific process of the pretreatment comprises the following steps: firstly, extracting pixel information in a DICOM file to obtain an original CT image, and performing image binarization on the original CT image to distinguish a visceral fat area and a background area as shown in FIG. 5a, wherein in order to ensure that a set threshold value can distinguish the visceral fat area and a tissue with higher density through a pixel value; the threshold value set by the invention is as follows: the gray scale value between 874-974 and 0 is set to be 255, the gray scale value of more than 974 or less than 874 is set to be 0, the image is divided into two parts by the threshold value, as shown in FIG. 5 b;

step S300, manually dividing a mask image of the visceral fat region according to the preprocessed data set image, as shown in fig. 5c, and then performing an and operation on pixel points corresponding to the divided mask image and the preprocessed data set image, to extract a final visceral fat region image, as shown in fig. 5 d;

the AttentionU-net network includes an encoding portion and a decoding portion;

the coding part comprises a 5-layer structure, the first 4-layer structure comprises two convolutional layers and a maximum pooling layer, the 5 th-layer structure comprises two convolutional layers, each convolutional layer uses relu as an activation function, the size of a convolutional kernel is 3 x 3, the number of output channels of the first-layer structure is 32, and then the number of output channels of each-layer structure is 2 times that of the last layer;

the decoding part comprises 5 layers of structures, each layer of structure comprises a bilinear interpolation up-sampling, an attention structure, three layers of convolutional layers and a splicing structure, and each layer of convolutional layer uses relu as an activation function. The convolution kernel size is 3 × 3, and its attention structure (AG) is shown in fig. 3: taking the output of a down-sampling layer and the output of an up-sampling layer as input, enabling the two inputs to pass through a convolution layer (convolution kernel 1) and a batch normalization layer respectively, then performing addition operation, rule activation function, convolution layer (convolution kernel 1), batch normalization and sigmoid activation function, and finally performing multiplication operation on the output of the sigmoid activation function and the input of the down-sampling layer;

in the training stage of the whole network, in order to prevent the over-fitting problem and improve the generalization capability of the Attention U-Net model, a Dropout layer is added after two convolutions before and after the fourth down-sampling; the Dropout layer randomly subtracts some neurons in the training of each batch, and can set the probability of how much neurons are removed by each Dropout layer, when the training of the first batch is carried out, a part of neurons are taken out according to the preset probability, then the training is started, and only the neurons which are not taken out and the corresponding weight parameters are updated and kept; after all the parameters are updated, a part of neurons are taken out again according to the corresponding probability, then training is carried out, and if the neuron which is used for training newly is trained, the parameters of the neuron are continuously updated; the neuron which is taken off for the second time, and the weight parameter which has been updated for the first time is kept, and is not modified until the nth time when the batch performs Dropout, the parameter is not deleted; dropout needs to be added in the training stage to prevent overfitting from improving the generalization capability of the model, and a Dropout layer is not added in the testing stage; through cross validation, the effect is best when the Dropout rate is 0.5, and the Dropout randomly generates the most network structures when the Dropout rate is 0.5;

in the whole network training process, the iteration times of the Attention U-net network training are 120 times, the batch size is 2, and the learning rate is set to be 1.0 e-5; the iteration times of the U-net network training are 120 times, the batch size is 4, and the learning rate is set to be 1.0 e-5; training the optimizer using Adam as a model, the tensor of the input data is 2018 × 1 × 256;

and S600, extracting pixel information in a DICOM file of the abdominal CT image needing to be predicted, performing image binarization according to the preprocessing method in the S200, inputting the image subjected to image binarization into a trained Attention U-net network, wherein the image output by the Attention U-net network is a segmented image.

Examples

FIG. 4 is a comparison of the manual segmentation of the doctor, the U-net network, and the Attention U-net network in the test set:

the segmentation Precision (SA), the under-segmentation rate (UR), the over-segmentation rate (OR), the Precision, and the Recall were calculated for the segmentation result and the golden standard image, and the calculation results are shown in table 1:

TABLE 1 comparison of visceral fat segmentation results

As can be seen from the calculation results in table 1, the accuracy of the two deep learning networks for the CT image segmentation of different populations and different abdomen positions is relatively high, and the over-segmentation rate and the under-segmentation rate are both low. Because the Attention gate structure is added into the Attention U-Net network, the segmentation effect of the model is effectively enhanced, and the accuracy rate is higher compared with the U-Net network. And from the segmentation effect of FIG. 4, the extension U-Net segments less redundant area than the U-Net network. The OR of the Attention U-Net model is 1.87 percent less than that of the U-Net model, and the two networks have good segmentation on the detailed structure of the visceral fat area, so that the segmentation rate can be seen from the under-segmentation rate. However, the division is performed for the purpose of calculating the visceral fat area, which is calculated by the number of pixel points in the divided image, and therefore, the result is more accurate as the unnecessary area is divided, but the present invention divides the unnecessary area, but the divided parts are within an acceptable error range compared with the manual division. The method of the present invention can meet the requirement of automatic segmentation of visceral fat regions.

Although the present invention has been described with reference to the above embodiments, it should be understood that the present invention is not limited to the above embodiments, and those skilled in the art can make various changes and modifications without departing from the scope of the present invention.

Claims

1. A method for automatically segmenting abdominal CT visceral fat regions, comprising:

2. The method for automatically segmenting the abdominal CT visceral fat area of claim 1, wherein the midriff images of young and old people with a slice width of 1mm are selected in step S100; and selecting images of the upper abdomen and the lower abdomen of young and middle-aged people and old people, wherein the slice width is 5 mm.

3. The method for automatically segmenting abdominal CT visceral fat area according to claim 1, wherein the preprocessing in step S200 comprises: extracting pixel information in the DICOM file to obtain an original CT image, and performing image binarization on the original CT image to distinguish an visceral fat region from a background region.

4. The method as claimed in claim 3, wherein the gray scale value between 874-974 pixel values in the image binarization is set as 255, and the gray scale value with pixel value greater than 974 or less than 874 is set as 0.

5. The method for automatically segmenting abdominal CT visceral fat area according to claim 1, wherein the Attention U-net network in step S500 comprises two parts, an encoding part and a decoding part;

6. The method for automatically segmenting abdominal CT visceral fat area according to claim 4, wherein the step S600 comprises the steps of: extracting the pixel information in the DICOM file, then carrying out image binarization according to the preprocessing method in the step S200, inputting the image subjected to image binarization into a trained Attention U-net network, wherein the image output by the Attention U-net network is a segmentation image.