SP-CPGCN:用于尘肺病分期的超像素先验因果感知图卷积网络
2025-08-03王月莹纪国华冯伟毅赵涓涓强彦马建芬施熠炜杨帆
SP-CPGCN : causality-perception graph convolutional network on superpixel prior for pneumoconiosis staging
Wang Yueyingla,Ji Guohuala,Feng Weiyi1a,Zhao Juanjuanla,Ibt,Qiang Yanla.2,MaJianfenla,Shi Yiwei 3 ,Yang Fan4 (1.a.Collegfeeceeece)odfrenUiesit 030024,Chna;2holfSfare,NUniersitfhnaTaianO3OhnaDeptofulmaryamp;ricalCaredicine ShanxiKeybsoeosliclUsi 030012,China;4Ocupioalisesereetioamp;ontroltefisnCalamp;Electricity(Gou)anna)
Abstract:Toaddress thelowaccuracyinpneumoconiosisstaging causedbythesmallandthinpneumoconiosis fociandthe influenceof non-causal features inexistingdeeplearning methods,thispaperproposedacausality-perceptiongraph convolutionalnetworkonsuperpixelpriors(SP-CPGCN).Themethodextractedlocal,subtlefeaturesbyperforming featureextractiononsuperpixelsratherthantheentirechestradiograph.The method constructedagraph network byfullyconsidering the spatial proximityandfeaturesimilarityamongnodes,anddesignedahierarchicalaggegatedgraphconvolutionalnetwork to enableinformationtransferandfeatureintegrationacrossdiffrentdepths.Aditionally,themethodemployedadaptivecausal inferenceongraphconvolutionalnetworks,usingacausalinterventionstrategythatcombinedinterventionlossandstabilityloss to avoidheinterfereneof non-causalfeatures.Itlsointroducedanintraclassconsistencylosstoalance individual-specific features with group-universal features.Thevalidationresultsonaclinicalpneumoconiosischestradiograph datasetshowthat SP-CPGCN achieved an accuracy of 82.4% ,a precision of 78.9% ,a sensitivity of 77.3% ,a specificity of 88.6% ,and an AUC of 90.3% ,outperforming other methods. The experimental results show that SP-CPGCN efectively improves the accuracyand stability of pneumoconiosis staging and provides areliablenew method for automated medical diagnosis.
'ey words:graph convolutional network(GNN);causal inference; superpixel; chest X-ray;pneumoconiosis stagi
0 引言
尘肺病是一种不可逆转的呼吸系统疾病。由于其高发性和难以治愈性,尘肺病已经成为一项重要的职业健康问题[1]。
胸片作为一种低剂量和经济的检查方法,可以提供肺部结构的纹理和密度变化,是尘肺诊断中最常用的影像学工具之一。在临床上,放射科医生通过观察胸片上影像学特征,将尘肺病分为正常、一期、二期、三期四个期别。但由于尘肺病变具有轻薄弥漫性和渐进演化性,导致相邻期别之间的可分难度增加,尤其是在尘肺早期诊断上。如图1所示,这就使得放射科医生进行尘肺定级时费时费力且主观差异大,并且对医生的临床经验和专业知识都提出了较高的要求。因此,迫切需要一种准确、高效和低成本的计算机辅助的尘肺分类诊断方法。

目前,已经有研究探索端到端的学习模型自动提取深层特征,在尘肺病诊断方向取得了不错的进展。例如,Zhang等人[2采用经典的ResNet模型,Wang等人[3]选用Inception-V3模型,Devnath等人[4运用CheXNet模型进行微调对尘肺胸片进行正常阶段和其他病变阶段的分期。为了进一步提高分期准确率,考虑尘肺病变的微薄弥漫性特征,Luo等人[5]建立了双分支暗通道差分卷积神经网络,Sun等人[67]提出了完全深度学习范式和深度专家咨询范式以克服尘肺病分期的噪声标签。这些研究共同验证了深度学习在解读尘肺胸片方面与专业放射科医生有着相媲美的表现。然而这些方法虽然在特定条件下能够较好地捕捉胸片的全局特征,但在处理尘肺病灶这种微小且弥漫的病变时仍然存在不足,尤其是受非因果特征的影响(如正常的肺部结构、拍摄角度及噪声等),导致诊断准确率不高。……
