CN115409847A

CN115409847A - Lightweight segmentation method based on CT image

Info

Publication number: CN115409847A
Application number: CN202210867416.8A
Authority: CN
Inventors: 黄新; 林洁沁
Original assignee: Guilin University of Electronic Technology
Current assignee: Guilin University of Electronic Technology
Priority date: 2022-07-22
Filing date: 2022-07-22
Publication date: 2022-11-29

Abstract

The invention provides a light-weight segmentation method based on a CT image. First, the dataset with the lesion segmentation label is preprocessed. Next, a lightweight pneumonia segmentation network was constructed, and trained and tested using the above data set. The light pneumonia division network takes an encoder-decoder as a main structure. In order to improve the image segmentation efficiency and reduce the redundant connection of the model structure, pruning operation is carried out on the model. The loss function used by the invention is used for solving the problem of too slow convergence in training, and the overlapping degree loss function and the cross entropy loss function are combined to be used as a new loss function. The output layer uses the Sigmoid activation function and the other convolutional layers use the Swish activation function. Experimental data show that for network pruning operation, under the condition that the number of network parameters is reduced, model performance change before and after pruning is small, and operation speed and operation efficiency are obviously improved.

Description

Lightweight segmentation method based on CT image

Technical Field

The invention relates to the field of medical image processing, in particular to a light-weight segmentation method based on CT images.

Background

Computer-aided diagnosis is an important means for assisting doctors to make rapid and accurate diagnosis. The deep neural network benefits from the advantage of high precision, and gradually replaces the traditional method to become a mainstream segmentation mode. The existing segmentation network can better realize medical image segmentation. However, the improved models are complex, often at the cost of a large amount of computational effort, despite the better performance achieved. Meanwhile, complex networks often require the consumption of large amounts of memory and computer resources, and are difficult to deploy directly to embedded platforms and mobile terminals. When network parameters are reduced, most lightweight network models have the problems of sharply reduced segmentation performance, limited calculation capacity and the like. In order to enable deep learning semantic segmentation to be better developed and applied in the medical field, the problem of large network parameter quantity is solved. Narrowing the input training samples is the most straightforward and easy method. However, the disadvantage of this method is also obvious, and the reduction of the input size and data volume can cause the network to lose spatial information and edge detail information in the training process, which directly results in poor segmentation effect. The best mode is to design and modify a deep learning framework. Currently, the important difficult problem is to construct a lightweight network while ensuring that the network performance is basically unchanged, so as to reduce the parameter quantity and the requirement of computing resources. Therefore, designing a lightweight segmentation method based on CT images is a key to solve this problem.

Disclosure of Invention

The present invention is directed to provide a light-weight segmentation method based on CT images to solve the above-mentioned problems in the background art.

In order to achieve the purpose, the invention provides the following technical scheme:

and acquiring a data set with a focus segmentation label, and preprocessing the data set. The format of the processed pictures is fixed to be 512 multiplied by 512, and a training set and a testing set are constructed according to the proportion of 8:2. And constructing a lightweight segmentation network based on the CT image, and training and testing by using the data set. In order to improve the image segmentation efficiency, the model structure needs to be pruned, and the aim is to reduce the redundant connection of the model structure in the calculation process. The invention relates to a scheme for pruning an original trunk network by combining the semantic hierarchy characteristics of a pneumonia focus. The network can keep more information, maintain the accuracy of the original network and reduce the calculated amount and the storage amount of the network. Through experiments, a lightweight segmentation network based on CT images is designed, and an encoder-decoder is used as a main structure. The encoder portion includes 3 downsampling for obtaining local feature information of context in the CT image. The decoder part comprises 3 times of upsampling and is used for reading high-level semantic information of the CT image. The problem that convergence is too slow exists in network training of a single cross entropy Loss function, and the overlapping degree Loss function (Dice Loss) and the cross entropy Loss function (BCE Loss) are combined to serve as a new Loss function by combining characteristics of pneumonia images. In the network, an output layer uses a Sigmoid activation function to control the threshold value of output response within the range of [0,1] to obtain a mask image. Other convolutional layers use the Swish activation function.

As a further scheme of the invention: each sampling process includes 2 convolutional layers with a convolutional kernel size of 3*3, a max pooling operation with a convolutional kernel of 2*2. Each convolution is followed By Normalization (BN) and Swish activation function processing.

As a further scheme of the invention: the output layer uses the Sigmoid activation function and the other convolutional layers use the Swish activation function.

As a further scheme of the invention: the overlap Loss function (Dice Loss) is combined with the cross entropy Loss function (BCE Loss) as a new Loss function. The problem of slow convergence is solved while pixel-by-pixel pulling is carried out from the global angle, and the segmentation performance is further improved.

（1）

Equation (1) is a cross entropy loss function.

Is a one-hot vector.

（2）

Equation (2) is the Dice loss function.

In order to predict the result of the event,

is a sample label.

（3）

Equation (3) is a loss function used in the present invention.

Compared with the U-Net original network, the segmentation method has the advantages that the Accuracy (Accuracy) is improved by 0.01 percentage point, the class average Pixel Accuracy (MPA) is reduced by 0.43 percentage point, and the average Intersection over Unit (MIoU) is improved by 0.1 percentage point. The network operation time is reduced by 20.83 percentage points. Experimental data show that for the pruning operation of the network, under the condition that the network parameters are reduced, the performance change of the model before and after pruning is small, and the operation speed and the operation efficiency are obviously improved. The segmentation structure in the invention is relatively stable, and has a certain reference value compared with the existing structure.

Description of the drawings:

FIG. 1 is a diagram showing a structure of a lightweight segmented pneumonia network according to the present invention

FIG. 2 is an image of the Swish activation function used in the present invention

The method comprises the following specific implementation steps:

the present invention will be described in further detail with reference to the accompanying drawings.

Fig. 1 is a diagram showing a structure of a lightweight pneumonia division network according to the present invention. The network structure in the invention is composed of a coder-decoder as a main structure. The encoder portion includes 3 downsampling and the decoder portion includes 3 upsampling. Each sampling process includes 2 convolutional layers with a convolutional kernel size of 3*3, a max pooling operation with a convolutional kernel of 2*2. Each convolution is followed By Normalization (BN) and Swish activation function processing. The output layer uses the Sigmoid activation function and the other convolutional layers use the Swish activation function.

Figure 2 is an image of the Swish activation function used by the present invention. The output layer of the invention uses a Sigmoid activation function, and other convolution layers use a Swish activation function. Compared with ReLU, when x >0, the Swish activation function has no case of gradient disappearance; when x <0, the neuron will not be inactivated as in ReLU.

The above description is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention will be covered within the protection scope of the present invention.

Claims

1. The invention provides a light-weight segmentation method based on a CT image.

2. The method of claim 1, wherein each sampling process comprises 2 convolution layers with a convolution kernel size of 3*3 and a maximal pooling operation with a convolution kernel of 2*2, after each convolution, normalization (BN) and Swish activation function processing are performed, an output layer uses a Sigmoid activation function, and other convolution layers use a Swish activation function, so that the problem of too slow convergence is solved while the convolution layers are pulled in from a global perspective pixel by pixel, and a superposition Loss function (Dice Loss) and a cross entropy Loss function (BCE Loss) are combined to serve as a new Loss function.

3. The light-weight segmentation method based on the CT image according to claim 1 and claim 2, wherein the segmentation structure has an Accuracy (Accuracy) that is improved by 0.01 percentage point, a class average Pixel Accuracy (Mean Pixel Accuracy, MPA) that is reduced by 0.43 percentage point, an average cross over unit (MIoU) that is improved by 0.1 percentage point, and a network operation time that is reduced by 20.83 percentage point, compared with a U-Net original network, experimental data show that, for the pruning operation of the network, in the case of a reduction in the number of network parameters, the model performance before and after pruning is less changed, and the operation speed and the operation efficiency are significantly improved.