Automated 3D Lung Segmentation in Chest Scans Using Channel Attention Hybrid Dilated Convolution-Based U-Net Model

Accurate lung segmentation from 3D chest scans is a difficult problem due to the lung volume shape variations and less contrast between the lung and the neighbouring tissues. Traditional methods are computationally expensive and often result in false positives. Therefore, this paper develops a computationally efficient and robust 3D lung segmentation approach called Channel Attention Hybrid Dilated Convolution-based U-Net (CAHDC-UNET) model. This model is formulated by including Channel Attention and Hybrid Dilated Convolution in the U-Net architecture to increase deep semantic features' learning and segmentation efficiency. Initially, the pre-processing of the 3D scans is performed to extract the coarse lung contours, and then, the refinement is done using the CAHDC-UNET to obtain refined lung coarse segmentation. Experiments are conducted on a publically available COVID-19 database that contains 750 chest computed tomography (CT) scans. The CAHDC-UNET model attained better results with an accuracy of 99.8% and BF of 99.5% and provides reliable 3D lung segmentation that can be utilized as a clinical diagnosis tool to help doctors deliver precise additional estimations.