Multi-label chest X-ray image classification based on long-range dependencies capture and label relationships learning

Diagnosing chest diseases from X-ray images using convolutional neural networks (CNNs) is an active area of research. However, existing methods mostly focus on extracting feature information from local regions for prediction, while ignoring the larger-scale image contextual information. Moreover, anatomical segmentation knowledge and co-occurrence relationships among labels, which are important for classification, are not fully utilized. To address the above problems, we proposed a method to capture long-range dependent information in chest X-ray images using a CNN with large kernel convolution. Furthermore, it captures the detailed features of the interest region through anatomical segmentation and builds the potential relationships of different diseases using a graph convolutional network (GCN). Firstly, we pre-trained UNet from a dataset with organ-level annotations for segmenting anatomical regions of interest in the images. Secondly, we build a four-stage backbone network using the large kernel attention (LKA) mechanism and superimpose anatomically segmented regions on the feature maps of each stage to obtain different scales of feature maps for the regions of interest. Thirdly, we utilized a GCN to obtain a co-occurrence matrix representing the potential relationships between all disease labels in the training dataset. Finally, we get the disease diagnosis by combining the label co-occurrence matrix and the visual feature maps. We experimentally show that our proposed method achieves excellent AUC scores of 91.5%, 84.5%, and 82.5% on three publicly available CXR datasets-NIH, Stanford CheXpert, and MIMIC-CXRJPG, respectively.