HFANet: hierarchical feature fusion attention network for classification of glomerular immunofluorescence images
HFANet: hierarchical feature fusion attention network for classification of glomerular immunofluorescence images
复制标题
DOI：
10.1007/s00521-022-07676-6
复制
发表时间：
2022-09
影响因子：
6
通讯作者：
Haoran Liu;Ping Zhang;Yongle Xie;Xifeng Li;Dongjie Bi;Yu-rong Zou;Lei Peng;Guisen Li
中科院分区：
文献类型：
--
作者：
Haoran Liu;Ping Zhang;Yongle Xie;Xifeng Li;Dongjie Bi;Yu-rong Zou;Lei Peng;Guisen Li
慢性肾脏病（CKD）伴随着永久性肾脏损害，已成为全球公共卫生的沉重负担。临床上，肾小球免疫荧光（IF）图像被广泛用于揭示CKD的发生概率和类型。在肾小球IF图像的组织病理学评估中，使用多个描述性指标从不同方面表征沉积物，这些沉积物提示相关肾脏病变。在本文中，我们设计了一个层次特征融合注意力网络（HFANet）分类两个主要的描述性指标，即荧光强度和分布模式。HFANet通过层次化特征融合关注度（HFA）模块，利用浅层纹理特征补充深层语义特征，最大限度地提高特征提取能力和信息融合效率。与直接添加或连接不同，HFANet加权连接来自不同层次的特征图，以突出更具鉴别力的区域。此外，通过将HFANet与建议的强度均衡（IE）算法，U-Net++和Grad-CAM相结合，构建了肾小球IF图像的计算机辅助诊断系统。该系统对荧光强度和分布模式的分类准确率分别达到90.48%和90.87%。广泛的对比实验和消融研究表明，HFANet优于其他通用骨干的帮助下，HFA模块，和设计的系统的分类性能与高级病理学家。该系统给出的热图与临床医生使用的分类证据相似，可用作病理学家的诊断参考和培训材料。系统演示视频可在补充材料中找到。
The chronic kidney disease (CKD) accompanied by permanent kidney damage, has become a heavy burden for worldwide public health. Clinically, glomerular immunofluorescence (IF) images are widely-used to reveal the occurrence probability and type of CKD. In histopathological assessment for glomerular IF image, multiple descriptive indicators are used to characterize deposits from different aspects, which suggest associated kidney lesions. In this paper, we design a hierarchical feature fusion attention network (HFANet) to classify two main descriptive indicators, namely fluorescence intensity and distribution pattern. Through the hierarchical feature fusion attention (HFA) module, HFANet supplements deep semantic features using shallow texture features to maximize its feature extraction capability and efficiency of information fusion. Different from directly adding or concatenating, HFANet weighted concatenates the feature maps from different hierarchies to highlight more discriminative regions. Further, by integrating HFANet with the proposed intensity equalization (IE) algorithm, U-Net++, and Grad-CAM, a computer-aided diagnostic system for glomerular IF images is constructed. With this system, the classification accuracy of the fluorescence intensity and distribution pattern reaches 90.48% and 90.87%, respectively. Extensive comparative experiments and ablation studies demonstrate that HFANet outperforms other universal backbones with the help of HFA module, and the classification performance of the devised system is comparable to senior pathologists. The heatmap given by the system, which is similar to the classification evidence used by the clinicians, can be used as diagnostic reference and training material for pathologists. The systematic demonstration video is available in the supplementary material.