Liver cancer ct image segmentation method based on channel attention

Abstract

The application relates to a liver cancer CT image segmentation method based on channel attention, a segmentation network for liver and tumor regions is constructed, residual modules are used to replace ordinary convolution layers on the basis of a U-Net encoding-decoding structure, the training speed and precision of the network are improved, an attention mechanism is added in the decoder stage, the importance and mutual dependence of different channels in a feature map are fully considered, feature maps output by each stage of the decoder are up-sampled to the size of an original image through a deep supervision module and are spliced with the feature map output in the last stage, feature loss is prevented, two same segmentation networks are used for segmenting a liver region of interest and segmenting tumor cells respectively. The method has good segmentation effect without significantly increasing the parameter amount; from the clinical point of view, liver and tumor regions in an abdominal CT image are accurately segmented, a good auxiliary role is provided for doctor diagnosis, and the diagnosis efficiency is improved.

Description

Technical Field

[0001] This invention relates to a medical image processing technique, and more particularly to a method for segmenting hepatocellular carcinoma CT images based on channel attention. Background Technology

[0002] The liver is the largest gland in the human body and one of the most important metabolic organs. Liver cancer is one of the most frequently diagnosed cancers worldwide, with poor prognosis and a high mortality rate. Hepatocellular carcinoma (HCC) is the most common form of liver cancer, accounting for over 90% of liver cancer cases in my country. Due to the complexity of liver diseases and the existence of various treatment options, it is crucial to develop a suitable clinical treatment plan tailored to the patient's specific condition. Treatment methods for liver cancer include surgical resection, radiofrequency ablation, interventional therapy, and radiation therapy. Surgical resection aims to achieve self-healing by completely removing cancerous tissue. Radiation therapy uses high-energy rays to destroy cancer cells in the lesion area. Radiofrequency ablation requires doctors to determine the number of ablation sessions, needle insertion points, and puncture depth based on the patient's computed tomography (CT) images. It is evident that each of these treatment options relies on precise information regarding the tumor's location, size, and shape; otherwise, incomplete removal of cancer cells or damage to normal tissue may occur. Therefore, accurately segmenting the liver and tumor cells from abdominal CT images is essential for the treatment of liver cancer patients.

[0003] Traditionally, the localization and segmentation of the liver and liver tumors are usually done manually by experienced radiologists, who segment the liver outline and tumor area layer by layer from abdominal CT images. However, this work is not easily reproducible, is labor-intensive, and the segmentation results are greatly affected by the doctor's experience and subjective judgment.

[0004] In recent years, with the continuous development of computer vision in the field of medical imaging, automatic segmentation of liver and tumor cells has become a key and challenging problem in computer-aided diagnosis. Before the widespread application of deep learning, some traditional machine learning methods were applied to liver and tumor segmentation. These included algorithms based on statistical shape models, grayscale values, and level sets. However, the liver, tumors, and surrounding organs have small grayscale differences and blurred edges, making it difficult for these algorithms to accurately segment the liver, tumors, and other organs, resulting in low segmentation accuracy and long computation time. Summary of the Invention

[0005] To address the challenge of clear segmentation of liver and tumor cells, a channel attention-based segmentation method for hepatocellular carcinoma CT images is proposed, enabling fully automated segmentation of the liver and tumor regions in abdominal CT images.

[0006] The technical solution of this invention is as follows: a segmentation network for liver and tumor regions, which replaces the original ordinary convolutional modules in U-Net with residual modules based on the U-Net encoding and decoding structure; a channel attention module is added at the skip connections in the decoder stage to mimic human visual cognition, increase the weights of important channels, and help establish interdependent relationships between channels; a deep supervision module is added in the decoder stage to upsample the feature maps output by each stage of the decoder to the original image size, and the feature maps output by the last stage are concatenated to fuse the convolutional information of each level into the output, thereby improving the edge information and tumor detail information in the feature maps.

[0007] Preferably, the residual module consists of a residual mapping part and a direct mapping part. The residual mapping part consists of two or three convolutions, and the input information is output after being passed through the convolutions in sequence. The direct mapping part passes the input information to the output through a bypass, thus protecting the integrity of the information.

[0008] Preferably, the channel attention module structure is as follows: First, global average pooling is used to compress the input feature map by feature channel to obtain global information Z; then, a one-dimensional convolution is used to map the global information Z onto the feature s, and then channel compression weights are calculated; finally, the compression weights of the feature channels are multiplied by the input feature map by channel to obtain the final output of the attention module.

[0009] A channel attention-based CT image segmentation method for hepatocellular carcinoma utilizes a cascaded segmentation network with two identical structures. An abdominal CT image is input into the first segmentation network, which segments the entire liver with lesions. The segmented result is then masked onto the initial input abdominal CT image and fed into the second segmentation network to segment the tumor cells within the segmented liver with lesions, ultimately obtaining a CT pathological image of hepatocellular carcinoma with segmented liver and tumor cells.

[0010] Furthermore, the two identical segmentation networks are trained separately using the same training method. After training, the segmentation networks are used to segment the region of interest in the liver and to segment tumor cells, respectively.

[0011] The beneficial effects of this invention are as follows: This invention provides a channel attention-based segmentation method for hepatocellular carcinoma CT images. Based on the U-Net encoding / decoding structure, this method replaces ordinary convolutional layers with residual modules to improve the network's training speed and accuracy. Furthermore, an attention mechanism is added in the decoder stage to fully consider the importance and interdependence of different channels in the feature map. Simultaneously, a deep supervision module upsamples the feature maps output from each stage of the decoder to the original image size and then concatenates them with the feature maps output from the final stage to prevent feature loss. Compared with existing methods, this method achieves better segmentation results without significantly increasing the number of parameters. From a clinical perspective, it enables accurate segmentation of the liver and tumor regions in abdominal CT images, providing excellent assistance to doctors in diagnosis and improving diagnostic efficiency. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the U-Net architecture for deep monitoring with residual and channel attention mechanisms in the method of this invention;

[0013] Figure 2 This is a schematic diagram of the residual module RB in the method of the present invention;

[0014] Figure 3 This is a schematic diagram of the channel attention module (CA) in the method of the present invention;

[0015] Figure 4 This is a flowchart of the hepatocellular carcinoma CT image segmentation method of the present invention;

[0016] Figure 5 Comparison of CT pathological image segmentation results for hepatocellular carcinoma using the method of this invention. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0018] U-Net has achieved excellent results in end-to-end learning, especially in the field of medical image segmentation, but it also has some unavoidable drawbacks. When the network has a large number of layers, gradient vanishing or exploding may occur during training, making training impossible. Therefore, this invention proposes... Figure 1 The diagram shown illustrates the U-Net architecture for deep monitoring with residual and channel attention mechanisms, used for segmentation. First, based on the U-Net encoding / decoding structure, the original ordinary convolutional modules in U-Net are replaced with residual modules. Second, the channel attention module (… Figure 1The CA module (in the code) is added to the skip connections in the decoder stage to mimic human visual cognition, increasing the weights of important channels and helping to establish interdependencies between channels. In addition, a deep supervision module is added in the decoder stage to supplement features lost during convolution, see [link to module]. Figure 1 The dashed lines in the middle section connect the parts, which allows the information from each level of convolution to be fused into the output, thereby improving the edge information and tumor detail information in the feature map.

[0019] In neural networks, the most common feature extraction tools are 1×1 and 3×3 convolutional filters. Figure 2 The diagram below illustrates the residual module (RB), where H×W is the size of the feature map, Cin and Cout are the number of channels in the input and output feature maps, respectively, and M is the number of convolutional kernels. The output of the residual module is represented as:

[0020] y = x + F(x, {W i})

[0021] Where x is the input feature map, y is the output feature map, and F(x,{W i}) is used to perform a convolution operation on the input feature map, W i This represents the parameters in the i-th convolutional layer.

[0022] The residual module consists of a residual mapping part and a direct mapping part. The residual mapping part is generally composed of two or three convolutions, as shown on the right side of Figure 2, which contains several convolutional parts. The input information is sequentially processed through convolutions and then output. The left side is the direct mapping part. Figure 2 (This is the case with the RB structure). The direct mapping part in the residual module can bypass the input information and pass it to the output, thus protecting the integrity of the information. At the same time, replacing ordinary convolution with residual modules can avoid the problems of vanishing or exploding gradients during training. In addition, the direct mapping part can also effectively reduce feature redundancy.

[0023] Channel attention, simply put, is a process of compression, mapping, and decompression. Compression yields global information for each channel; mapping aims to learn the interrelationships between channels, i.e., learning channel weights; and decompression highlights important channel information.

[0024] Figure 3 This is a schematic diagram of the channel attention module in the method. For a given input feature map... Where W, H, and C represent the width, height, and number of channels of the feature map, respectively. The first step is to compress the features along the spatial dimension, compressing each two-dimensional feature channel into a real number, which represents the global information of that channel. Global average pooling is used to compress the feature channels, obtaining the global information Z∈{z1, z2, z3, ..., z...}c}, the c-th element z in Z c The calculation formula is as follows:

[0025]

[0026] Where i and j represent that the c-th channel corresponds to the i-th and j-th rows, x c This represents the pixel values ​​at the i-th and j-th rows of the c-th channel in the input feature map X. Then, channel weights are learned, considering not only the importance of individual channels but also the correlation between adjacent channels. Therefore, a one-dimensional convolution maps the global information Z onto the feature s, as shown below. Figure 3 As shown, its channel compression weight can be calculated using the following formula: w d =σ{C1D k (Z)},

[0027] Among them, C1D k For a one-dimensional Same convolution, k is the kernel size. From the above formula, we know that the kernel size k determines the interaction range of the channels. An optimal interaction range can be manually adjusted in different feature maps. However, manually adjusting this range through cross-validation consumes a significant amount of computational resources. Therefore, considering that the interaction range of the channels should be related to the total number of channels in the feature map, i.e., there exists a linear mapping... This makes the total number of feature map channels As is well known, the number of channels C is generally a power of 2. Therefore, let:

[0028]

[0029] When C is determined, then:

[0030]

[0031] m odd Let m be the odd number closest to m. In this method, we take a = 2 and b = 1.

[0032] Finally, the decompression process involves multiplying the obtained weights by channel with the input feature map to obtain the final output of the attention module.

[0033] The flowchart of the channel attention-based CT image segmentation method for hepatocellular carcinoma is as follows: Figure 4 As shown, the method is implemented using two identical structures. Figure 1The cascaded segmentation networks are shown. An abdominal CT image is input into the first segmentation network, which segments the entire liver with lesions. The resulting segmentation is then masked over the initial input abdominal CT image and fed into the second segmentation network. This network segments the tumor cells within the segmented liver with lesions, ultimately obtaining a CT pathological image of hepatocellular carcinoma with segmented liver and tumor cells. Two identical segmentation networks are used for segmenting the liver region of interest model and the tumor cell model, respectively. They need to be trained separately but employ the same training strategy.

[0034] The network was built using PyTorch. The experimental hardware environment was configured as follows: four NVIDIA RTX 2080 GPUs and 16GB of RAM; the software environment consisted of Ubuntu 16.04, CUDA 9.2, Tensorflow 1.14.0, Keras 2.2.4, and OpenCV 4.5.0; the loss function was cross-entropy loss, the batch size was set to 8, and the training iterations were set to 150 epochs. The Adam optimizer was used to optimize the network, with an initial learning rate of 0.0001. Early stopping was used to control the model training process; training was stopped if the loss function value did not decrease within 10 epochs.

[0035] The experimental results are shown in the table below. Liver DICE and Tumor DICE represent the overlap coefficients for liver and tumor segmentation, respectively. Segmentation tests were conducted by comparing various network structures. According to the table below, the method of this invention achieved higher segmentation accuracy for both liver and tumor segmentation.

[0036]

[0037] The comparison of segmentation effects of hepatocellular carcinoma CT pathological images obtained using this invention is shown in the following figure. Figure 5 As shown, Figure 5 As can be seen from the results, the predicted segmentation results obtained by this invention are very close to the labeled images annotated by professional physicians in terms of both position and overall shape. The labeled lesion areas are quite accurate and the outlines are very fine.

[0038] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A channel attention-based CT image segmentation method for hepatocellular carcinoma, characterized in that, By cascading two segmentation networks with the same structure, the abdominal CT image is input into the first segmentation network, which segments the entire liver with lesions. The segmented result is then masked over the initial input abdominal CT image and sent to the second segmentation network to segment the tumor cells in the entire liver with lesions, and finally obtains a CT pathological image of hepatocellular carcinoma with liver and tumor cells segmented. The segmentation network is based on the U-Net encoding and decoding structure, where residual modules replace the original ordinary convolutional modules in U-Net; channel attention modules are added at the skip connections in the decoder stage to mimic human visual cognition, increase the weights of important channels, and help establish interdependent relationships between channels. A deep supervision module is added to the decoder stage. The deep supervision module upsamples the feature maps output by each stage of the decoder to the original image size and concatenates the feature maps output by the last stage, thus integrating the convolutional information of each level into the output and improving the edge information and tumor detail information in the feature maps. The residual module consists of a residual mapping part and a direct mapping part. The residual mapping part consists of two or three convolutions, and the input information is output after being passed through the convolutions in sequence. The direct mapping part passes the input information to the output through a bypass, thus protecting the integrity of the information. The channel attention module structure is as follows: First, global average pooling is used to compress the input feature map according to the feature channels to obtain global information Z; then, a one-dimensional convolution is used to process the global information. Z Mapping to features s Next, the channel compression weight is calculated using the following formula: ,in, For one-dimensional Same convolution, k It refers to the size of the convolution kernel; considering that the interaction range of the channels should be related to the total number of channels in the feature map, i.e., there exists a linear mapping. This makes the total number of feature map channels... Number of channels C It is a power of 2, therefore, let: When C is determined, then: , Indicates separation m The nearest odd number, take a =2, b =1; Finally, the compression weights of the feature channels are multiplied by the input feature map according to the channel to obtain the final output of the attention module.

2. The channel attention-based CT image segmentation method for hepatocellular carcinoma according to claim 1, characterized in that, The two identical segmentation networks were trained separately using the same training method. After training, the segmentation networks were used to segment the region of interest in the liver and to segment tumor cells, respectively.

Images