病理学家通过分析肺腺癌低级别组织和癌旁组织来确定病灶切除范围,然而,两者间的细胞形态差异较小,分析时依赖病理学家的主观经验,耗时且易误诊.故提出一种结合多尺度特征融合和通道自注意力的胶囊网络(Multi-Scale Feature Fusion with Self-Channel Attention for Capsule Network,MSCNet),用于帮助医生高效诊断疾病,为患者提供更好的治疗方案.首先,设计了多尺度特征融合模块来提升胶囊网络以捕捉同源图像不同尺度间的语义信息,试图减少模型计算量以提高处理速度及分类准确性.其次,通道自注意力(Self-Channel Attention,SCA)模块作为MSCNet的另一重要组件,可以寻找到更具代表性的特征,辅助识别组织病理学图像中的细微特征,降低误诊风险.实验结果表明,在肺腺癌低级别组织与癌旁组织的二分类任务中,MSCNet实现了99.34％的分类准确率、97.65％的F1-Score值和97.57％的精确度.
A Capsule Network Combining Multi-Scale Feature Fusion and Attention Mechanism for Lung Adenocarcinoma Classification Task
Pathologists determine the extent of lesion resection by analyzing low-grade and paracancerous tissues of lung adenocarcinoma;however,the cellular morphology differences between the two are small,and the analysis relies on the pathologist's subjective experience,which is time-consuming and prone to misdiagnosis.Therefore,this study proposes a capsule network(Multi-Scale Feature Fusion with Self-Channel Attention for Capsule Net-work,MSCNet)that combines multi-scale feature fusion and attention mechanisms for helping physicians efficiently diagnose diseases and provide better treatment options for patients.Specifically,we first design a multi-scale fea-ture fusion module to enhance the capsule network to capture semantic information between different scales of homologous images in an attempt to reduce the amount of model computation in order to improve processing speed and classification accuracy.Second,the self-channel attention(SCA)module,as another important com-ponent of MSCNet,can find more representative features to assist in identifying subtle features in histopathology images and reduce the risk of misdiagnosis.Experimental results show that MSCNet achieves 99.34％classification accuracy,97.65％F1-Score value,and 97.57％precision in the task of binary classification of low-grade tissues and paracancerous tissues of lung adenocarcinoma.
万方数据
维普
