Multi-Scale DenseNet with Hybrid Shuffle-Spatial Attention for Robust Lung Disease Classification on Chest X-Rays

Abstract

Lung diseases such as pneumonia, tuberculosis, and the current coronavirus (COVID-19) are significant causes of morbidity and mortality in the world. Early diagnosis and accurate diagnosis based on chest X-rays (CXR) is very important for effective treatment, but manual interpretation takes time and can be prone to errors. It is well known that the conventional deep learning(DL) models are incomplete to capture not only the finegrained but also the global pathological features of the heterogeneous CXRs, which lead to a suboptimal classification performance. This paper proposes to build a robust automated framework for classification of lung diseases which enables multiscale feature extraction along with attention mechanism interpreted for lung disease. This study proposes a multi-scale DenseNetwork of a hybrid shuffle-spatial attention (HSSA) module, in order to suppress the non-important features of medical images and capture discriminative features at multiple resolutions and emphasize on clinically important regions. The model was trained and tested against the NIH ChestX-ray14 dataset using some preprocessing, augmentation, and end-to-end supervised learning. Experimental results show that it achieves better performance compared to 5 state-of-the-art models with 92.8% accuracy, 91.1% F1-score and 95.4% AUC-ROC with attention maps and visual interpretability. The proposed framework of MS-DenseNet + HSSA shows significant improvement in terms of automatic lung disease detection from CXRs, it provides a reliable, explainable, and clinically applicable method for the purpose of analysing CXRs for the radiologist community and also willing to be incorporated in real-world diagnostic processes.