Self-Sequential Attention Layer Based DenseNet for Thoracic Diseases Detection

Abstract

Thoracic diseases are the various collections of diseases that are associate with the cavity of the thorax such as lungs, heart, oesophagus, chest wall, and diaphragm. The Chest X-Ray is the widely used radiological examination in the diagnosis of such diseases. Many existing methods used entire Chest X-Ray images for the process of training however, it suffered from few limitations. The misalignment of potential or exist of unrelated objects in the Chest XRay images resulted in irrelevant noises which limited the network performance. The pre-processing steps of existing methods during training the neural network resulted in lower resolution images. Due to loss of information, it is very difficult to find the tiny lesion regions from the images. To solve such issues, the Self-Sequential Attention Layer based DenseNet (SAL-DN) model is proposed to enhance thoracic disease prediction on Chest X-Ray. The SAL uses the co-relation among the class labels and abnormal pathology by analysing the known feature maps from DenseNet-121.The SAL-DN has the advantages of providing the best representation of images and the capacity to handle large imbalanced datasets like Chest X-Ray14. The performance of the proposed SAL-DN model is compared with existing methods in-terms of the achieved AUC score. The experimental result shows that the proposed SAL-DN outperforms the Thorax-Net method by obtaining an Average AUC score of 0.8715, whereas Thorax-Net obtained an Average AUC score of 0.7876 in patient-wise official split.