基于局部上下文关系与自适应感知学习的细粒度肺部分割  

Fine-Grained Lung Segmentation Based on Local Context Relation and Adaptive Perception Learning

作　　者：全福 张智高 王郝日钦 关强 王庆虎[1] QUAN Fu;ZHANG Zhigao;WANG Haoriqin;GUAN Qiang;WANG Qinghu(College of Computer Science and Technology,Inner Mongolia Minzu University,Tongliao 028000,China)

机构地区：[1]内蒙古民族大学计算机科学与技术学院,内蒙古通辽028000

出　　处：《软件导刊》2025年第1期115-122,共8页Software Guide

基　　金：内蒙古自然科学基金项目(2022MS06029);内蒙古自治区科技计划项目(2020GG0189);内蒙古民族大学信息技术赋能智慧城乡建设产学研用创新团队项目(2021)。

摘　　要：肺部CT图像分割是计算机辅助诊断系统识别肺癌等疾病的基础步骤,但肺部组织的多样性易造成CT图像的肺部区域出现局部噪点,且心脏等器官的介入也易导致肺部边缘模糊。为解决上述问题,提出一种结合局部上下文关系建模与自适应感知学习的U型卷积网络模型。针对肺部区域的局部噪点,模型通过周边信息对肺部特征构建多层次的上下文关系,增强网络对多样肺细胞的识别能力。针对边界难以区分的问题,设计了自适应感知学习模块:模块设计了混合注意力从通道层面与空间层面引导模型赋予肺部区域更多的关注;自适应感知模块之间由下至上的特征融合机制增强了模型识别肺部特征的鲁棒性。该方法在LUNA和SHCXR两个肺部分割数据集上取得了98.58%和97.68%分割精度,较其他分割方法平均提升了0.34%和0.25%,可为进一步的肺部疾病分析提供有力支持。Automatic lung image segmentation is a key step for computer-aided diagnosis systems to detect diseases such as lung cancer.How‐ever,the diversity of lung cells easily introduces local noise in the lung region of CT images,and organ interventions like the heart often blur the lung boundaries.To address these issues,this paper proposes a novel U-Net framework that combines local context relation learning mod‐ules and adaptive perception modules.To reduce the influence of local noises,this paper constructs a multi-level context relationship for the lung feature extraction by exploiting the surrounding information of specific lung region,which enhances the recognition ability of U-Net to di‐verse lung cells.To address the issue of blurry lung boundaries,this paper proposes an adaptive perception learning module working at the skipping connections.This module includes a mixed attention mechanism to guide the model for focusing more on the lung area at both the channel and spatial dimensions.In addition,the designed bottom-up feature fusion path can further enhance the robustness of the learned lung features.The proposed method achieves 98.58%and 97.68%accuracy on LUNA and SHCXR dataset,which are 0.34%and 0.25%high‐er than other segmentation methods on average.The proposed approach can provide help for further analysis of lung diseases.

关 键 词：图像分割 肺部图像分割 U型网络 上下文关系学习 自适应感知学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]