Abstract: Thoracic disease classification based on chest radiographs is important to improve diagnostic accuracy and reduce the pressure on the healthcare system.The huge variation in the size of the regions of different thoracic diseases is the main challenge in the chest radiograph-based classification of thoracic diseases.When classifying diseases with small onset regions,most of the regions in the image are noisy regions,and it is difficult for traditional methods to cope with the huge size differences among di-seases effectively.To address this problem,a mask construction method combining multi-scale features is proposed,using DenseNet-121 as the feature extractor,a two-branch network is constructed,in which the global network is used for the overall classification,and tiny lesion regions are fed into the local branches to mitigate the interference of noisy regions.The branch feature fusion module based on the attentional mechanism is used to fuse the classification features from the two branches' information adaptively.Comparison experiments,ablation experiments,and parameter sensitivity analyses are performed on the ChestX-ray14 dataset.The experimental results show that the average AUC of the proposed method for classifying 14 thoracic diseases is higher than that of the existing methods,which is effective and parameter insensitive.

Key words: Thoracic disease classification, Deep learning, Chest radiographs, Attention mechanism, Feature fusion