A Novel Approach for Recognition of Thoracic Pathologies in Chest X-Ray Radiography Using Deep Model

Abstract

A Deep learning models with pre-trained CNN architecture have made a major impact on chest radiology field such as Chest X-Ray radiography. Many of deep models used publicly available Chest X-ray14 dataset for training and testing purposes. This dataset has 14 different thoracic pathology distribution and is split into Image-wise and patient-wise. Existing researchers used resized Chest X-Ray14 images to avoid computational complexity, which resulted in loss of image resolution. Existing models are trained to concentrate on the whole image rather than more on pathologically significant regions of X-Ray. The Chest X-Ray14 dataset is imbalanced 14 pathology distribution, which affected model performance especially on minority pathology classes. Our proposed deep model is designed to address all these issues. A DenseNet-121 is used as a core block supported by Grad-CAM-based attention mechanism as a supporting block. This attention mechanism guides backbone network to concentrate more on abnormal regions of the image instead of entire resized images of above dataset. The proposed model is trained with Adaptive Augmentation training, which aims to address imbalance effect of the dataset. The performance comparison of our deep model is done with other existing models under both split of database. We achieved high average AUC of 84.55%, 90.2% respectively in both split of data.