Abstract: The automatic identification of thoracic diseases from chest X-rays is a significant area of research in computer-aided diagnosis. However, many existing methods for thoracic disease classification struggled to handle differences in lesion area sizes and often failed to accurately identify and localize the lesion areas. To address the above problems, a Cross-scale Attention Network (CANet) approach for thoracic disease classification was proposed. In this method, DenseNet-121 was employed as the feature extraction network, and three main modules were integrated: Self Aware Attention (SAA), Upward Focus Attention (UFA) and Downward Guidance Attention (DGA) modules. The SAA module was designed to refine spatial pathological features by extracting channel and region information relevant to thoracic diseases, while reducing interference from irrelevant areas. In order to achieve cross-scale interaction of spatial context information, image feature calibration was performed using the upward focus attention and downward guidance attention modules. Additionally, the Spatial Attention Pyramid Pooling (SAPP) module was proposed to fuse multi-scale features from different feature maps, thereby enhancing the detection performance for thoracic diseases. Experiments on the proposed method using the ChestX-ray14 and DR-Pneumonia datasets achieve Area Under Curve (AUC) values of 83.4% and 82.6%, respectively, outperforming the DualCheXNet, A3Net, and CheXGAT methods. Compared with the CheXGAT method, the AUC values of the proposed method improve by 0.7 and 0.1 percentage points, respectively. The experimental results show that the proposed method effectively identifies critical information in chest radiography, significantly improving the performance of thoracic disease classification.

Key words: thoracic disease classification, chest X-ray, attention mechanism, convolutional neural network, feature fusion

摘要： 从胸部X光片中自动识别胸部疾病是计算机辅助诊断的重要研究领域。然而，现有的许多胸部疾病分类方法在处理病变区域大小差异方面存在困难，同时也无法准确识别和定位不同疾病的病变区域。针对上述问题，提出了一种基于跨尺度注意力网络(CANet)的胸部疾病分类方法。该方法使用DenseNet-121作为特征提取网络，并集成三个主要模块：自感知注意力(SAA)、向上聚焦注意力(UFA)和向下引导注意力(DGA)模块。自感知注意力模块通过提取与胸部疾病相关的通道和异常区域信息，细化空间位置上的病理特征，并减少不相关区域的干扰。为了实现不同尺度空间上下文信息的跨尺度交互，使用向上聚焦注意力和向下引导注意力模块进行图像特征校准。此外，还提出了空间注意力金字塔池化(SAPP)模块，用于融合不同特征图的多尺度特征，以增强胸部疾病的检测性能。在ChestX-ray14和DR-Pneumonia数据集上对所提出的方法进行实验，受试者工作特征曲线下面积（AUC）值分别达到了83.4%和82.6%，优于DualCheXNet、A3Net和CheXGAT等方法。与CheXGAT方法相比，所提方法的AUC值分别提高了0.7和0.1个百分点。实验结果表明，所提方法可以识别胸部X光片中的重要信息，有效提高了胸部疾病分类的性能。

关键词: 胸部疾病分类, 胸部X光片, 注意力机制, 卷积神经网络, 特征融合