A Computationally Efficient U-Net Architecture for Lung Segmentation in Chest Radiographs

Abstract

Lung segmentation plays a crucial role in computer-aided diagnosis using Chest Radiographs (CRs). We implement a U-Net architecture for lung segmentation in CRs across multiple publicly available datasets. We utilize a private dataset with 160 CRs provided by the Riverain Medical Group for training purposes. A publicly available dataset provided by the Japanese Radiological Scientific Technology (JRST) is used for testing. The active shape model-based results would serve as the ground truth for both these datasets. In addition, we also study the performance of our algorithm on a publicly available Shenzhen dataset which contains 566 CRs with manually segmented lungs (ground truth). Our overall performance in terms of pixel-based classification is about 98.3% and 95.6% for a set of 100 CRs in Shenzhen dataset and 140 CRs in JRST dataset. We also achieve an intersection over union value of 0.95 at a computation time of 8 seconds for the entire suite of Shenzhen testing cases.