Label Co-occurrence Learning with Graph Convolutional Networks for Multi-label Chest X-ray Image Classification.
- 作者列表："Chen B","Li J","Lu G","Yu H","Zhang D
:Existing multi-label medical image learning tasks generally contain rich relationship information among pathologies such as label co-occurrence and interdependency, which is of great importance for assisting in clinical diagnosis and can be represented as the graph-structured data. However, most state-of-the-art works only focus on regression from the input to the binary labels, failing to make full use of such valuable graph-structured information due to the complexity of graph data. In this paper, we propose a novel label co-occurrence learning framework based on Graph Convolution Networks (GCNs) to explicitly explore the dependencies between pathologies for the multi-label chest X-ray (CXR) image classification task, which we term the "CheXGCN". Specifically, the proposed CheXGCN consists of two modules, i.e., the image feature embedding (IFE) module and label co-occurrence learning (LCL) module. Thanks to the LCL model, the relationship between pathologies is generalized into a set of classifier scores by introducing the word embedding of pathologies and multi-layer graph information propagation. During end-to-end training, it can be flexibly integrated into the IFE module and then adaptively recalibrate multi-label outputs with these scores. Extensive experiments on the ChestX-ray14 and CheXpert datasets have demonstrated the effectiveness of CheXGCN as compared with the state-of-the-art baselines.
: 现有的多标签医学影像学习任务一般包含标签共现、相互依存等病理间丰富的关系信息, 这对于辅助临床诊断非常重要，可以表示为图结构数据。然而，大多数最先进的作品只关注从输入到二进制标签的回归, 由于图数据的复杂性，未能充分利用这些有价值的图结构信息。在本文中，我们提出了一种新的基于图卷积网络 (GCNs) 的标签共生学习框架为了明确探索多标签胸部 X 线 (CXR) 图像分类任务的病理之间的相关性，我们称之为 “CheXGCN”。具体来说，提出的 CheXGCN 由两个模块组成，即图像特征嵌入 (IFE) 模块和标签共生学习 (LCL) 模块。由于 LCL 模型，通过引入病理学的词嵌入和多层图信息传播，将病理学之间的关系推广到一组分类器评分中。在端到端训练期间，它可以灵活地集成到 IFE 模块中，然后用这些分数自适应地重新校准多标签输出。在 ChestX-ray14 和 CheXpert 数据集上的大量实验证明了 CheXGCN 与最先进的基线相比的有效性。
METHODS:BACKGROUND AND PURPOSE:A critical role for sphingosine kinase/sphingosine-1-phosphate (S1P) pathway in the control of airway function has been demonstrated in respiratory diseases. Here, we address S1P contribution in a mouse model of mild chronic obstructive pulmonary disease (COPD). EXPERIMENTAL APPROACH:C57BL/6J mice have been exposed to room air or cigarette smoke up to 11 months and killed at different time points. Functional and molecular studies have been performed. KEY RESULTS:Cigarette smoke caused emphysematous changes throughout the lung parenchyma coupled to a progressive collagen deposition in both peribronchiolar and peribronchial areas. The high and low airways showed an increased reactivity to cholinergic stimulation and α-smooth muscle actin overexpression. Similarly, an increase in airway reactivity and lung resistances following S1P challenge occurred in smoking mice. A high expression of S1P, Sph-K2 , and S1P receptors (S1P2 and S1P3 ) has been detected in the lung of smoking mice. Sphingosine kinases inhibition reversed the increased cholinergic response in airways of smoking mice. CONCLUSIONS AND IMPLICATIONS:S1P signalling up-regulation follows the disease progression in smoking mice and is involved in the development of airway hyperresponsiveness. Our study defines a therapeutic potential for S1P inhibitors in management of airways hyperresponsiveness associated to emphysema in smokers with both asthma and COPD.
METHODS::The interim results from this 90-day multi-dose, inhalation toxicology study with life-time post-exposure observation has shown an important fundamental difference in persistence and pathological response in the lung between brake dust derived from brake-pads manufactured with chrysotile, TiO2 or chrysotile alone in comparison to the amphiboles, crocidolite and amosite asbestos. In the brake dust exposure groups no significant pathological response was observed at any time. Slight macrophage accumulation of particles was noted. Wagner-scores, were from 1 to 2 (1 = air-control group) and were similar to the TiO2 group. Chrysotile being biodegradable, shows a weakening of its matrix and breaking into short fibers & particles that can be cleared by alveolar macrophages and continued dissolution. In the chrysotile exposure groups, particle laden macrophage accumulation was noted leading to a slight interstitial inflammatory response (Wagner-score 1-3). There was no peribronchiolar inflammation and occasional very slight interstitial fibrosis. The histopathology and the confocal analyses clearly differentiate the pathological response from amphibole asbestos, crocidolite and amosite, compared to that from the brake dust and chrysotile. Both crocidolite and amosite induced persistent inflammation, microgranulomas, and fibrosis (Wagner-scores 4), which persisted through the post exposure period. The confocal microscopy of the lung and snap-frozen chestwalls quantified the extensive inflammatory response and collagen development in the lung and on the visceral and parietal surfaces. The interim results reported here, provide a clear basis for differentiating the effects from brake dust exposure from those following amphibole asbestos exposure. The subsequent results through life-time post-exposure will follow.
METHODS::The respiratory tract is lined by a pseudo-stratified epithelium from the nose to terminal bronchioles. This first line of defense of the lung against external stress includes five main cell types: basal, suprabasal, club, goblet and multiciliated cells, as well as rare cells such as ionocytes, neuroendocrine and tuft/brush cells. At homeostasis, this epithelium self-renews at low rate but is able of fast regeneration upon damage. Airway epithelial cell lineages during regeneration have been investigated in the mouse by genetic labeling, mainly after injuring the epithelium with noxious agents. From these approaches, basal cells have been identified as progenitors of club, goblet and multiciliated cells, but also of ionocytes and neuroendocrine cells. Single-cell RNA sequencing, coupled to lineage inference algorithms, has independently allowed the establishment of comprehensive pictures of cell lineage relationships in both mouse and human. In line with genetic tracing experiments in mouse trachea, studies using single-cell RNA sequencing (RNAseq) have shown that basal cells first differentiate into club cells, which in turn mature into goblet cells or differentiate into multiciliated cells. In the human airway epithelium, single-cell RNAseq has identified novel intermediate populations such as deuterosomal cells, 'hybrid' mucous-multiciliated cells and progenitors of rare cells. Novel differentiation dynamics, such as a transition from goblet to multiciliated cells have also been discovered. The future of cell lineage relationships in the respiratory tract now resides in the combination of genetic labeling approaches with single-cell RNAseq to establish, in a definitive manner, the hallmarks of cellular lineages in normal and pathological situations.