CN-121982040-A - Lung nodule image segmentation method, device, equipment, medium and product

CN121982040ACN 121982040 ACN121982040 ACN 121982040ACN-121982040-A

Abstract

The application discloses a lung nodule image segmentation method, a device, equipment, a medium and a product, and relates to the technical field of image processing, wherein the method comprises the steps of acquiring a three-dimensional lung CT image and inputting the three-dimensional lung CT image into an encoder, and performing image preprocessing through an initial convolution layer; the method comprises the steps of carrying out layered feature extraction through a multi-level residual block, enhancing response intensity of a feature map through a channel attention module and an edge guiding attention module, carrying out parallel void rate convolution calculation through a void space pyramid pooling module to obtain enhanced high-level features, carrying out step-by-step up sampling on the enhanced high-level features through a decoder, fusing jump connection features from corresponding levels of an encoder through an attention gating module at each level to obtain final fused features, and generating a segmentation probability map of lung nodules through segmentation output layers. The method and the device remarkably improve the accuracy and the robustness of the lung nodule segmentation result and improve the reliability and the application value in clinical application.

Inventors

Yang Tianshuai
LIU WEIPING
ZHANG GUANGYUN
ZHAO HUI
LI XIANG
HU JUNYI
JIN JING
SUN YINGLI
YANG LEI
LU ZHENLI
SU ZHEN
LI XIN
LI JING
GUO HONGWEI

Assignees

浙江泉林智能医疗科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260123

Claims (10)

1. A lung nodule image segmentation method, the lung nodule image segmentation method comprising: acquiring a three-dimensional lung CT image and inputting the three-dimensional lung CT image into an encoder for image preprocessing to obtain a feature map; Carrying out layered feature extraction on the feature map through a multi-level residual block to obtain layered features, wherein the shallowest layer features in the layered features enhance response strength through a channel attention module and an edge guiding attention module; Carrying out parallel void rate convolution calculation on the deepest layer feature in the layered features through a void space pyramid pooling module to obtain enhanced high-level features; Step-by-step up-sampling the enhanced high-level features through a decoder, and fusing jump connection features from corresponding levels of the encoder through an attention gating module at each level to obtain final fused features; and generating a segmentation probability map of the lung nodule through the segmentation output layer based on the final fusion feature.
2. The lung nodule image segmentation method according to claim 1, wherein the specific process of the image preprocessing comprises the steps of inputting the three-dimensional lung CT image into an initial convolution layer of an encoder, and performing three-dimensional convolution, batch normalization and nonlinear activation processing through the initial convolution layer to obtain a feature map.
3. The lung nodule image segmentation method according to claim 1, wherein the specific process of enhancing the response intensity of the feature map by the channel attention module and the edge guide attention module comprises: obtaining the output of a first-stage residual block in a multi-stage residual block, wherein the output of the first-stage residual block comprises main path output and residual branch output; carrying out global average pooling on the main path output through a channel attention module to generate a channel statistical vector; The channel statistical vector sequentially passes through a first full-connection layer, a ReLU activation function and a second full-connection layer to carry out nonlinear transformation, so that an intermediate vector of an original channel dimension is obtained; applying a Sigmoid function to the intermediate vector to generate a weight coefficient of each channel; multiplying the weight coefficient and the main path output channel by channel, and outputting the characteristics of weighted attention of the channels; residual mapping is carried out on the characteristics of the weighted channel attention, and after the mapping result is added with the residual branch output, the channel attention module output characteristics are obtained through ReLU activation function processing; The channel attention conversion is carried out on the channel attention module output characteristics through the edge guiding attention module, so that channel enhancement characteristics are obtained; averaging the channel enhancement features along the channel dimension to obtain a gray level map; Extracting an edge feature map from the gray map by a laplace edge detector; converting the edge feature map into edge features by an encoder; and splicing the edge features and the channel enhancement features, inputting the edge enhancement features into an edge guiding attention module, and generating the attention enhancement features of the fusion edge structure through the edge guiding attention module.
4. The lung nodule image segmentation method according to claim 1, wherein the specific process of performing parallel void-rate convolution calculation on the deepest feature of the layered features by the void space pyramid pooling module comprises: Simultaneously inputting the deepest features into N parallel hole convolution branches and a global context branch, wherein the N hole convolution branches respectively adopt different hole rates to carry out three-dimensional hole convolution and are subjected to batch normalization and ReLU activation function processing; Global average pooling is carried out on the deepest features through the global context branches, and global features are output after convolution and up-sampling processing; Splicing the outputs of the N cavity convolution branches with the global feature; carrying out convolution, batch normalization, nonlinear activation and 3D Dropout fusion treatment on the spliced features; And adding the fused features and the deepest features through residual connection to obtain enhanced high-level features.
5. The lung nodule image segmentation method according to claim 1, wherein the specific process of fusing the skip connection features from the corresponding levels of the encoder by the attention gating module comprises: Receiving, by the attention gating module, the skip connection feature and the current upsampling feature of the decoder; performing convolution, batch normalization and ReLU activation function processing on the jump connection features, and compressing to obtain first features; performing convolution and batch normalization on the current up-sampling feature of the decoder, and compressing to obtain a second feature; Adding the first feature and the second feature, and then processing by a ReLU activation function, convolution and Sigmoid function to generate a spatial attention weight; Multiplying the spatial attention weight with the jump connection feature to obtain a feature after gating screening; and splicing the features subjected to the gate control screening with the current up-sampling features of the decoder, and performing convolution, batch normalization and ReLU activation function processing to obtain the current up-sampling fusion features.
6. The method of claim 1, wherein the specific process of generating a segmentation probability map of a lung nodule by segmenting the output layer comprises: regularizing the final fusion feature to obtain a feature after random discarding; Mapping the channel dimension of the randomly discarded features into a category space through 1 multiplied by 1 convolution to obtain an original category score map; and carrying out probability normalization on the original category score map through an activation function to generate a segmentation probability map of the lung nodule.
7. A lung nodule image segmentation apparatus, the apparatus comprising: the encoder is used for extracting layered characteristics of an input three-dimensional lung CT image, comprises a plurality of cascaded residual blocks, and is integrated with a channel attention module and an edge guiding attention module at a first-stage residual block; the cavity space pyramid pooling module is connected to the output end of the encoder and used for fusing multi-scale context characteristics; The decoder is symmetrically arranged with the encoder and comprises a plurality of up-sampling stages, and each stage fuses jump connection characteristics from a corresponding level of the encoder through the attention gating module; And the segmentation output layer is connected with the output end of the decoder and is used for outputting a segmentation probability map of the lung nodule.
8. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the lung nodule image segmentation method of any of claims 1-6.
9. A computer readable storage medium having stored thereon a computer program, which when executed by a processor, implements the steps of the lung nodule image segmentation method of any of claims 1-6.
10. A computer program product comprising a computer program which, when executed by a processor, implements the steps of the lung nodule image segmentation method of any of claims 1-6.

Description

Lung nodule image segmentation method, device, equipment, medium and product Technical Field The present application relates to the field of image processing technologies, and in particular, to a method, an apparatus, a device, a medium, and a product for segmenting a lung nodule image. Background Cancer is one of the major challenges facing modern medicine, where the mortality rate of lung cancer is high in the first place of all cancers, and severely threatens human life health and quality of life. Statistics show that five-year survival rate of patients with advanced lung cancer is only 15%, while early-discovered and intervened lung cancer patients can be obviously improved to 40% -70%. It is worth noting that early lung cancer exists in the form of lung nodules, if the lung nodule can be effectively screened through low-dose Computed Tomography (CT), lung nodule lesions can be timely identified and located, and the five-year survival rate of patients is expected to be further improved by adopting targeted diagnosis and treatment measures. However, since lung nodules are often characterized by uneven distribution, irregular morphology, blurred boundaries, etc. in images, especially for tiny nodules or ground glass nodules (GGO), it is very likely to interfere with the subjective interpretation by radiologists. In recent years, along with the deep application of artificial intelligence technology in the field of medical image analysis, the development of an accurate and efficient automatic lung nodule segmentation algorithm has become a key link for promoting early screening and accurate diagnosis and treatment of lung cancer. Currently, the lung nodule segmentation technology has gradually evolved from the traditional image processing methods such as early-stage dependent threshold segmentation, morphological processing and the like to an intelligent analysis mode taking deep learning as a core. Particularly, based on a model of a Convolutional Neural Network (CNN), such as U-Net and various improved architectures (such as 3D U-Net, res U-Net and the like) thereof, the method combines with advanced technologies such as multi-scale feature fusion, attention mechanism and the like, has obvious breakthrough in segmentation precision and robustness, and can more effectively cope with complex scenes such as small nodules, blurred edges or adjacent pleura and the like. However, most lung nodule segmentation methods still use a two-stage process of "detection first and then segmentation", i.e. a candidate Region (ROI) is first located by a target detection model, and then each candidate region is finely segmented. Although the mode is clear in logic, the method has obvious defects that the detection stage is easy to miss (FALSE NEGATIVE) low-contrast or fine structures such as micro nodules, blood vessel ends and the like, is easy to be influenced by redundancy or noise channels, so that the distinguishing capability of targets and backgrounds is poor, space details are easy to lose, the adaptability to focuses with irregular shapes or large scale changes is poor, the risks of error accumulation and transmission exist, the accuracy of a segmentation result is poor, and the reliability and the application value of the method in clinical application are limited. Disclosure of Invention The application aims to provide a lung nodule image segmentation method, device, equipment, medium and product, which can effectively capture edge and texture information in an image, remarkably improve the attention of a network to low contrast or fine structures such as a tiny nodule, a blood vessel end and the like, further improve the segmentation integrity and boundary precision of a lung nodule small target, strengthen the response of a key feature channel, inhibit the influence of redundancy or noise channels, improve the discrimination of feature representation, more accurately distinguish the target from the background, effectively relieve the problem of space detail loss, improve the adaptability to focuses with larger irregular shapes or scale changes, effectively avoid the problem of noise transmission possibly caused by jump connection in the traditional U-Net, improve the accuracy and robustness of lung nodule segmentation results, and improve the reliability and application value in clinical application. In order to achieve the above object, the present application provides the following solutions: in a first aspect, the present application provides a method for segmenting a lung nodule image, comprising: acquiring a three-dimensional lung CT image and inputting the three-dimensional lung CT image into an encoder for image preprocessing to obtain a feature map; carrying out layered feature extraction on the feature map through a multi-level residual block to obtain layered features, wherein the shallowest layer feature in the layered features enhances response strength through a channel attention module and an edge