Automatic Lung Segmentation in Chest X-Ray Images using Dilated Dense-ResUNet

Statistical evidence from the World Health Organization (WHO) has shown that lung diseases rank among the primary reasons for fatalities worldwide. With a lack of radiologists interpreting Chest X-Ray (CXR) images for common lung diseases, it becomes a challenge to diagnose lung diseases in a timely manner: Thus, Computer Assisted Diagnostic (CAD) tools can help expedite the procedure. Researchers have begun using deep learning techniques to efficiently analyze lungs in CXR images. To further improve the analysis of lungs, the lung regions can be segmented as a preliminary step. This paper aims to develop and test a lung segmentation model using a modified U-Net architecture. The architecture utilizes transfer learning with DenseNet201, dilated convolutions using a dilation rate of 2 and residual blocks. Numerous experiments were conducted to validate the efficacy of using ImageNet architectures and ImageNet weights on CXR images. Furthermore, different dilation rates were tested to determine the effects of configuring the receptive field. The model was trained using two datasets; Montgomery County (MC) CXR Dataset and Shenzhen Hospital (SH) CXR Dataset. The segmented lungs were evaluated using Jaccard Index (IoU) and Dice Similarity Coefficient (DSC). Compared with results drawn from some of the current state-of-art methods, the proposed model achieves competitive results that has slightly higher IoU and DSC scores.