Algorithm 1. Lung Field Segmentation

Input: Given a set of CXR images X and a set of ground truth masks Y. $\mathit{I}\in \mathit{X}$ and $\mathit{M}\in \mathit{Y}$.

Output: O, the segmentation results.

1  Decompose I into homogeneous matrix H of homogeneous regions and a boundary matrix B of the boundaries of superpixels using superpixel extraction.

2  Downsample I to obtain the downsampled image ${\mathit{I}}^D$ using Equation (14).

3  Downsample M to obtain the downsampled image ${\mathit{M}}^D$.

4  Store the superpixel label information for each pixel of I.

5  In training phase:

5.1   Input a set of ${\mathit{I}}^D$ and a set of ${\mathit{M}}^D$ to the encoder–decoder segmentation network to train the model.

6  In prediction phase:

6.1   Input ${\mathit{I}}^D$ to the encoder–decoder segmentation network to predict the low-resolution segmentation results ${\mathit{O}}^D.$

6.2   Upsample ${\mathit{O}}^D$ to obtain the high-resolution segmentation results O using Equation (16).

6.3   Run the post-processing procedure on O to correct the segmentation results.

6.3.1    Keep the two largest regions and discard other small regions.

6.3.2    Fill all the holes in the two largest regions.

7  Output the final result O.